Fdident

Computation
Visualization
Programming
Frequency Domain
System Identification Toolbox
For Use with MATLAB
Istvn Kollr
How to Contact The MathWorks:
508-647-7000 Phone
508-647-7001 Fax
The MathWorks, Inc. Mail
24 Prime Park Way
Natick, MA 01760-1500
http://www.mathworks.com Web
ftp.mathworks.com Anonymous FTP server
comp.soft-sys.matlab Newsgroup
[email protected] Technical support
[email protected] Product enhancement suggestions
[email protected] Bug reports
[email protected] Documentation error reports
[email protected] Subscribing user registration
[email protected] Order status, license renewals, passcodes
[email protected] Sales, pricing, and general information
Frequency Domain System Identification Toolbox Users Guide
COPYRIGHT 1994 - 1997 by The MathWorks, Inc. All Rights Reserved.
The software described in this document is furnished under a license agreement. The software may be used
or copied only under the terms of the license agreement. No part of this manual may be photocopied or repro-
duced in any form without prior written consent from The MathWorks, Inc.
U.S. GOVERNMENT: If Licensee is acquiring the software on behalf of any unit or agency of the U. S.
Government, the following shall apply:
(a) for units of the Department of Defense:
RESTRICTED RIGHTS LEGEND: Use, duplication, or disclosure by the Government is subject to restric-
tions as set forth in subparagraph (c)(1)(ii) of the Rights in Technical Data and Computer Software Clause
at DFARS 252.227-7013.
(b) for any other unit or agency:
NOTICE - Notwithstanding any other lease or license agreement that may pertain to, or accompany the
delivery of, the computer software and accompanying documentation, the rights of the Government
regarding its use, reproduction and disclosure are as set forth in Clause 52.227-19(c)(2) of the FAR.
Contractor/manufacturer is The MathWorks Inc., 24 Prime Park Way, Natick, MA 01760-1500.
MATLAB, Simulink, Handle Graphics, and Real-Time Workshop are registered trademarks and Stateflow
and Target Language Compiler are trademarks of The MathWorks, Inc.
Other product or brand names are trademarks or registered trademarks of their respective holders.
Printing History: January 1994 First printing
August 1995 Reprint
FAX
u
@
i
Contents
1
Before You Begin
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-3
2
Tutorial
Frequency Domain Formulation and Solution . . . . . . . . . . . . 2-2
Basic Concepts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-2
Covariance of the Estimate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-7
Key Features of ELiS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-10
Imposing Constraints on the Estimates . . . . . . . . . . . . . . . . . . 2-12
Solutions for Some Special Cases . . . . . . . . . . . . . . . . . . . . . . . 2-14
Numerical Stability and Speed of the Procedures . . . . . . . . . . 2-15
Excitation Signals for Identification in the Frequency
Domain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-20
Multisine . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-21
Binary Excitation Signals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-22
Preprocessing of Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-24
Preprocessing in the Time Domain . . . . . . . . . . . . . . . . . . . . . 2-24
Transformation into the Frequency Domain . . . . . . . . . . . . . . 2-26
Preprocessing in the Frequency Domain . . . . . . . . . . . . . . . . . 2-28
Presentation of the Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-30
Model Validation and Simulations . . . . . . . . . . . . . . . . . . . . . . . 2-32
The First Quick Checks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-32
Stability and the Choice of the Proper Delay . . . . . . . . . . . . . . 2-32
Detection of Undermodeling and Overmodeling . . . . . . . . . . . 2-33
ii Contents
Study of the Residuals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-36
Simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-37
Cross Validation with Another Set of Measured Data . . . . . . . 2-40
Model Conversions from/to the System Identification
Toolbox . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-41
Conversion from ELiS to the theta-Format . . . . . . . . . . . . . . . 2-43
Conversion from the theta-Format to ELiS . . . . . . . . . . . . . . . 2-45
Data Formats and File Handling . . . . . . . . . . . . . . . . . . . . . . . . . 2-47
A Typical Identification Session . . . . . . . . . . . . . . . . . . . . . . . . . . 2-49
Investigation of the Time Domain Data . . . . . . . . . . . . . . . . . . 2-50
Examination of the Signal-to-Noise Ratios . . . . . . . . . . . . . . . 2-54
Conversion to Frequency Domain . . . . . . . . . . . . . . . . . . . . . . . 2-57
Variance Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-57
Identification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-62
Model Validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-69
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-77
3
Reference
Function Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-3
Excitation Signal Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-3
Preprocessing of Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-4
Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-4
Presentation of the Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-4
Model Validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-5
Model Conversions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-5
Data Vector and File Read/Write . . . . . . . . . . . . . . . . . . . . . . . . 3-6
Other . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-7
iii
A
Appendix
Description of the Data Vector and File Formats . . . . . . . . . . A-2
Common Conventions for Every Data Type . . . . . . . . . . . . . . . . A-2
Conventions for Individual Data and File Types . . . . . . . . . . . . A-4
Examples for ASCII Files . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A-10
Description of the Available Measurement Data Files . . . A-15
aluplate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A-15
bandpass, bp1, bandp8k . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A-16
cable1-6 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A-17
crankcas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A-24
emachine . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A-25
glassfib . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A-26
inpchan . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A-28
lowpaswb, lowpass1, lowpass2 . . . . . . . . . . . . . . . . . . . . . . . . . A-29
robotarm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A-33
iv Contents
v
About the Author
Istvn Kollr
Istvn Kollr received his M.S. degree in Electrical Engineering from the
Technical University of Budapest, Hungary, in 1977, and his Ph.D. degree from
the Hungarian Academy of Sciences, Budapest, in 1985. Currently he is a
Fulbright scholar at Stanford University. His permanent affiliation is as an
associate professor of Electrical Engineering at the Technical University of
Budapest.
Dr. Kollr is a Senior Member of the IEEE, scientific secretary of TC7 of
IMEKO and Editor-in-Chief of Periodica Polytechnica. He has published over
50 technical papers and a textbook in the areas of signal processing and system
identification. He is co-author of the book, Technology of Electrical
Measurements, Schnell, L. ed., Wiley, 1993.

vi

1
Before You Begin
1 Before You Begin
1-2
Modern system identification methods heavily use matrix calculations, usually
based on complex numbers. Therefore, algorithms for system identification can
be very effectively implemented in MATLAB. Its interactive environment and
the graphics possibilities offer an easy-to-use and flexible tool for specific
applications and for further development of the implemented methods.
This toolbox is a collection of frequency domain system identification
procedures, covering the whole identification process from excitation signal
design, through data preprocessing, parameter estimation, graphics
presentation of the results, to model verification.
Installation of the Toolbox
Installation instructions are in the computer-specific MATLAB installation
guides. The recommended directory name for the Frequency Domain System
Identification Toolbox is FDIDENT.
Demonstration script M-files can be run individually or through the main
demonstration M-file fdiddemo, showing the main features and possibilities of
the toolbox, and calculating several of the results given in the book of
Schoukens and Pintelon. To start any demo separately, execute it from inside
MATLAB. Type help fddemos for a list of the available demonstrations, and/or
run fdiddemo.
Introduction
1-3
Introduction
Identification means determining models of physical systems from noisy
measured data. Since the modeling of nature is the basis of our understanding
of the world, identification methods have applications in virtually every field of
sciences, especially technical ones.
One of the most often used models for dynamic physical systems is an ordinary
linear differential equation with constant coefficients. This model appears in
the same mathematical form in very different fields. Thus, the common
properties of these equations, and the measurement and estimation procedures
of their coefficients can be treated independently of the dimensions of the
physical quantities. This is done in the frame of linear system theory.
Linear dynamic systems have two equivalent descriptions: in the time domain
(differential equations), and in the frequency domain (transfer functions in the
f-domain or in the s-domain). Also the discrete-time descriptions exhibit this
duality: difference equations are equivalent to the zdomain transfer functions.
Time domain and frequency domain have different advantages and
disadvantages, they are in several respects complementary to each other.
Engineers often prefer frequency domain descriptions, because of the following
reasons:
While the solution of a differential equation needs convolution in the time
domain, this convolution is substituted by simple multiplication in the
frequency domain. This often makes it possible to explain system behavior in
a visual way.
It is often possible to decompose signals/noises into different frequency
bands. Signal-to-noise ratios can often be improved in the frequency domain
by choosing appropriate bands, and an occasional dc offset can also be easily
removed. Moreover, the energy (power) of periodic signals is concentrated to
discrete points in the frequency domain. Thus, frequency domain is very
selective with respect to periodic components.
By selecting the Fourier coefficients of the appropriate frequency band only,
significant data reduction can be achieved.
1 Before You Begin
1-4
It is often easier to calculate accurately a good model of a
(continuous-domain) physical system in the s-domain, using a digital
computer, than in the time domain. Continuous-time linear systems can be
modeled without systematic errors by difference equations, if the excitation
is piecewise constant (Ljung, 1987). If this assumption is not met, modeling
errors are introduced, resulting in a limited accuracy of the results.
The attainable dynamic range is usually significantly larger in the frequency
domain than in the time domain.
Slight nonlinearities are easier detected and measured by frequency domain
methods.
FFT is a very powerful tool for fast time domain to frequency domain
conversion, and thus, e. g., the evaluation of a correlation function is much
quicker via the frequency domain, than directly in the time domain.
On the other hand, time domain has also important advantages:
It is very natural to deal with time domain signals (although after some
practice, even the two-sided frequency axis becomes very visual, too).
Recursive methods often provide on-line calculation possibilities.
Time varying systems are easier modeled in the time domain.
The transient behavior of systems can be directly measured in the time
domain.
Time domain digital methods are not very sensitive to the type of the signal,
while frequency domain methods suffer from leakage effects when
non-periodic signals are processed.
Certain nonlinearities (clipping, slew rate etc.) are easier detected in the
time domain.
Although the two descriptions are basically equivalent to each other, the
formulation of the identification problem leads to very different methods in the
two domains. Thus, time domain identification methods and frequency domain
identification methods form two distinct groups.
Time domain methods are covered by several books, and a comprehensive
System Identification Toolbox is available for MATLAB. However, frequency
domain parametric methods are usually not treated, and these are not
incorporated in the System Identification Toolbox.
Introduction
1-5
This Frequency Domain System Identification Toolbox has been written to fill
this gap. It is based mostly on the book of Schoukens and Pintelon (1991), and
covers the whole identification procedure from excitation signal design through
data preprocessing and system parameter estimation to model validation. It is
also possible to convert identified parameters to the System Identification
Toolbox and vice versa.
The Tutorial chapter contains the essentials for the use of these methods, thus
the toolbox can be used in its own right. However, for a more profound
understanding of the methods, the book of Schoukens and Pintelon is highly
recommended.
MATLAB proved to be an ideal frame for all these methods, because most
algorithms are based on complex vector/matrix calculations and array
manipulations, and since interactive graphical checking of the results is often
essential. MATLAB also provides an easy-to-use environment for the
embedding of the prepared routines into a dedicated program, written for use
in a given field. Thus, the use of the toolbox may be twofold: it can be used as a
tool for the design and evaluation of experiments, and also as a frame to check
ideas before writing a lengthy special-purpose measurement and data
processing program.
The demonstrations accompanying the toolbox are not merely illustrations to
the use and the power of the functions, but also work with the measured data
used in the examples of Schoukens and Pintelon. Thus, the toolbox is an ideal
supplement to the book.
The author of this toolbox is very much indebted to Johan Schoukens and Rik
Pintelon for long and fruitful discussions on the implemented methods, and
also to the Vrije Universiteit Brussel and to the National Fund for Scientific
Research of Belgium for providing the conditions of this project.
Thanks are due also to Yves Rolain, Patrick Guillaume, Hugo Van hamme,
Bla Pataki, Tadeusz Dobrowiecki, Frank Louage, Franoise Renneboog and
Johan Top for their useful suggestions and remarks.
Budapest, Aug. 23, 1993.
Istvn Kollr
1 Before You Begin
1-6

2
Tutorial
2-2 Frequency Domain Formulation and Solution
2-2 Basic Concepts
2-8 Covariance of the Estimate
2-10 Key Features of ELiS
2-12 Imposing Constraints on the Estimates
2-14 Solutions for Some Special Cases
2-15 Numerical Stability and Speed of the Procedures
2-20 Excitation Signals for Identification in the Frequency Domain
2-21 Multisine
2-22 Binary Excitation Signals
2-24 Preprocessing of Data
2-24 Preprocessing in the Time Domain
2-26 Transformation into the Frequency Domain
2-28 Preprocessing in the Frequency Domain
2-30 Presentation of the Results
2-32 Model Validation and Simulations
2-32 The First Quick Checks
2-32 Stability and the Choice of the Proper Delay
2-33 Detection of Undermodeling and Overmodeling
2-36 Study of the Residuals
2-37 Simulations
2-40 Cross Validation with Another Set of Measured Data
2-41 Model Conversions from/to the System Identification Toolbox
2-43 Conversion from ELiS to the theta-Format
2-45 Conversion from the theta-Format to ELiS
2-47 Data Formats and File Handling
2-49 A Typical Identification Session
2-50 Investigation of the Time Domain Data
2-54 Examination of the Signal-to-Noise Ratios
2-57 Conversion to Frequency Domain
2-57 Variance Analysis
2-61 Identification
2-68 Model Validation
2-77 Bibliography
2 Tutorial
2-2
Frequency Domain Formulation and Solution
Basic Concepts
A general model used in the frequency domain identification of dynamic linear
systems is shown in Figure 2-1. The system is represented by its transfer
function H(), where = s = j = j2f in the Laplace-domain, or
= z
-1
= exp(jT
s
) in the zdomain, respectively, and H is a rational form,
eventually extended by a delay term, (see Equation (1).
Figure 2-1: The general model used in frequency domain system
identification.
(b)
Y
m
X
N
x
X
m
Y
(a)
x(t)
y(t)
h(t)
(
y
m
t
)
x
m
(t)
N
y
n
x
(t)
n
y
(t)
H()
2-3
(1)
The excitation signal has complex amplitudes X
k
at angular frequencies
k
, the
response of the system is Y
k
= H(
k
)X
k
. The measured input and output
complex amplitudes are both corrupted by noises N
x
and N
y

(errors-in-variables model), which are usually assumed to be Gaussian,
uncorrelated between input and output, and also uncorrelated between
different frequency points. Input-output correlation may also be considered,
see Equation (14) later in this chapter.
Measurements are made at angular frequencies
k
, k = 1...F, the measured
complex input and output amplitudes are X
mk
and Y
mk
, respectively. The
unknown parameters are those of the transfer function (vector P), and the
complex input and output amplitudes (vectors X and Y). The basic equations
can be written as
(2)
and
(3)
Assuming that the noise on the complex amplitudes is Gaussian and
uncorrelated, its joint probability density function can be written as
(4)
where N
Rxk
, N
Ixk
, N
Ryk
and N
Iyk
are the real and imaginary parts of the input
and the output noise samples, respectively. N is the complex conjugate of N,
xk

and
yk
are the corresponding standard deviations, N
x
and N
y
denote the
H ( ) e
j T
d
b
0
0
b
1
1
b
nn
nn
+ +
a
0
0
a
1
1
a
nd
nd
+ +
---------------------------------------------------------------- =
Y
k
H
k
P ( , )X
k
k 1 2F, , = , =
Y
mk
H
k
P ( , ) X
mk
N
xk
( ) N
yk
+ k 1 2F, , = , =
p N
x
N
y
( , ) =
1
2
xk
2
---------------
k 1 =
F
exp
N
Rxk
2
N
Ixk
2
+
2
xk
2
-----------------------------
,

_
1
2
yk
2
---------------
k 1 =
F
exp
N
Ryk
2
N
Iyk
2
+
2
yk
2
-----------------------------
,

_
=
1
2
xk
2
---------------
k 1 =
F
exp
N
xk
N
xk
2
xk
2
------------------
,

_
1
2
yk
2
---------------
k 1 =
F
exp
N
yk
N
yk
2
yk
2
----------------- -
,

_
=
2 Tutorial
2-4
vectors formed of N
xk
, N
yk
, respectively. Here the input and output noises are
assumed to be uncorrelated. Correlatedness will be considered later in this
section.
By expressing the noise variables in (2.3) by X
mk
and X
k
(N
xk
= X
mk
-X
k
, N
Rxk
=
X
Rmk
-X
Rk
etc.), and taking the logarithm, with the assumption that
x
and
y

are known from a preceding noise analysis, the log-likelihood function is
obtained:
(5)
X, Y and P are not independent of each other, since Equation (2) must be
fulfilled.
The maximization of Equation (5) is equivalent to the minimization of
(6)
subject to the constraints
(7)
The constraints can be substituted into Equation (6), to eliminate Y. The result
is a nonlinear weighted least squares problem. Since generally we are not
interested in X either, a better way of the minimization of Equation (5) with
the constraints Equation (7) is to use the Lagrange multiplier technique to
eliminate both X and Y. Fortunately, X, Y and the multipliers can really be
eliminated, and the following expression is obtained for minimization:
(8)
where N(,P) and D(,P) are the numerator and the denominator of the
transfer function, respectively.
ln L X Y P , , ( ) ( ) =
const
X
mk
X
k
( ) X
mk
X
k
( )
2
xk
2
-----------------------------------------------------
,

_
k 1 =
F
Y
mk
Y
k
( ) Y
mk
Y
k
( )
2
yk
2
----------------------------------------------------
,

_
k 1 =
F
=
C
LS
X Y P , , ( )
X
mk
X
k
( ) X
mk
X
k
( )
2
xk
2
-----------------------------------------------------
,

_
k 1 =
F
Y
mk
Y
k
( ) Y
mk
Y
k
( )
2
yk
2
----------------------------------------------------
,

_
k 1 =
F
+ =
Y
k
H
k
P ( , )X
k
= k 1 = 2F , ,
C P ( )
1
2
---
e
j
k
T
d
N
k
P ( , )X
mk
D
k
P ( , )Y
mk
yk
2
D
k
P ( , )
2
xk
2
N
k
P ( , )
2
+
----------------------------------------------------------------------------------------- -
k 1 =
F
=
2-5
The cost function in Equation (8) may look somewhat strange. However, a quite
simple explanation of the underlying idea can be given as follows. This is by no
means a proof of the maximum likelihood nature of the estimate, but can help
to understand its structure.
A natural way of developing an appropriate estimate of the transfer function
is to minimiz
(9)
This is a weighted least squares type cost function. The weights have to be
chosen equal to the variances of the terms whose absolute values are taken, in
order to have an approximately chi-squared cost function. (The chi-squared
cost function is the one usually obtained in maximum likelihood estimations for
Gaussian data.)
The problem is that the terms between the absolute value signs are not any
more Gaussian distributed because of the division by X
mk
, moreover, the
division makes the distribution asymmetric, introducing a bias. So this is not
a proper way to obtain a cost function that provides high accuracy.
An alternative formulation for avoiding division is to investigate the terms in
(10)
For zero noise, and a proper model, this expression equals zero. The terms are
Gaussian, and independent of each other, so if the proper weights W
k
can be
found to form a chi-squared cost function, minimization of this may provide a
good estimate.
Since
(11)
we are looking for the variances of the remaining terms
(12)
C
t f
P ( )
1
2
--- W
t f k
e
j
k
T
d
N
k
P ( , )
D
k
P ( , )
--------------------
Y
mk
X
mk
---------
2
k 1 =
F
=
C
WLS
P ( )
1
2
--- W
t f k
e
j
k
T
d
X
mk
N
k
P ( , ) Y
mk
D
k
P ( , )
2
k 1 =
F
=
e
j
k
T
d
X
k
N
k
P ( , ) Y
k
D
k
P ( , ) 0 =
e
k
e
j
k
T
d
N
xk
N
k
P ( , ) N
yk
D
k
P ( , ) =
2 Tutorial
2-6
The variances are equal to
(13)
for independent N
xk
and N
yk
. Setting W
k
=1/var{e
k
}, the cost function Equation
(8) of ELiS is obtained. For correlated input-output noise, a more general
weighting can be developed (see Equation (14)).
Since Equation (8) is a sum of quadratic terms (though nonlinear in P because
of the denominators), C(P) can be minimized using powerful numerical
techniques developed for nonlinear least squares problems (Newton-Gauss
method, Levenberg-Marquardt method). This can be done in complex terms,
paying attention to maintain that the elements of P are real, or alternatively,
Equation (8) can be written as a sum of squared real terms. Because both N and
D are linear in P, the cost function is insensitive to the multiplication of P by
a scalar. Therefore, an additional constraint has to be introduced in order to
obtain a well-defined solution: e. g., the norm of the vector P, or the value of at
least one nonzero parameter can be fixed.
In order to obtain starting values for the iteration, the sum of the numerators
in Equation (8) can be minimized. This is an ordinary linear least squares
problem, having a unique solution, and can be readily solved by standard
procedures. However, this LS step does not provide any information on the
delay T
d
, thus an initial value has to be given by the user.
In the following sections the above described estimator will be referred to as
ELiS (Estimator for Linear Systems).
Input-Output Correlation
An even more general approach is to assume that the input and output noises,
belonging to the same frequencies, may be mutually correlated. This
correlation can be the result of different sources:
The input signal of the system can be noisy. In this case the noise is no more
an observation noise, since it excites the system under test, and the output
noise is at least partly produced from the input noise. A preceding noise
analysis, assuming observation noises only, gives too large input and output
variances, and without introducing a correction, leads to an erroneous cost
function. Considering input/output noise covariance, this error source can be
corrected for.
var e
k
{ } E N
xk
N
xk
{ } N
k
P ( , )
2
E N
yk
N
yk
{ } D
k
P ( , )
2
+ =
2
xk
2
N
k
P ( , )
2
2
yk
2
D
k
P ( , )
2
+ =
2-7
When the system under test is inside a feedback loop, the process noise is at
least partly led back to the input, resulting in a noisy excitation
1
.
The excitation signal generator may produce slightly unstable amplitudes, a
phenomenon that can be considered by an additive noise at the input.
Slight synchronization imperfections of measurements may virtually
increase the instability of Fourier coefficients, an effect that can also be well
compensated for by considering the input/output correlation. (Strictly
speaking, this is a kind of phase noise that may violate the complex Gaussian
assumption. ELiS will not be a maximum likelihood estimate any more, but
since it is also a least squares estimator, robust with respect to the noise
distribution (see the Key Features of ELiS on page 2-10), it will still
perform well.)
It can be shown that for consideration of the covariances, the cost function
must be modified as follows:
(14)
with
(15)
and
(16)
1. R. Pintelon, P. Guillaume, Y. Rolain and F. Verbeyst, Identification of Linear Sys-
tems Captured in a Feedback Loop, IEEE Trans. on Instrumentation and Measure-
ment, Vol. 41, No. 6, pp. 747-754, Dec. 1992.
C P ( )
1
2
---
e
j
k
T
d
N
k
P ( , )X
mk
D
k
P ( , )Y
mk
yk
2
D
k
P ( , )
2
xk
2
N
k
P ( , )
2
2 CND
k
{ } real +
---------------------------------------------------------------------------------------------------------------------
k 1 =
F
=
CND
k
c
xyk
e
j
k
T
d
N
k
P ( , )D
k
P ( , ) =
c
xyk
0.5cov N
xk
N
yk
{ , } 0.5E N
xk
N
yk
{ , } = =
2 Tutorial
2-8
Covariance of the Estimate
The above estimator is a maximum likelihood one in the Gaussian case.
Practice shows that the covariance matrix is usually close to the corresponding
Cramr-Rao lower bound. Let us maintain the constraints Equation (7) during
the derivations. In this case the Cramr-Rao bound is
(17)
where T denotes the transpose of the row vector, and X, Y and P denote the
true (exact) values.
Equation (17) has an alternative form:
(18)
which gives the very same values as Equation (16). The lower bound is
asymptotically approximated when the number of observations is large, or
when the signal-to-noise ratio is large.
Since we are interested in the covariance matrix of P
ML
only, it is desirable to
bring Equation (17) or Equation (18) into a form that does not contain the
covariances of X
ML
. After a long derivation, the following expression can be
obtained:
(19)
An alternative form can be given as follows. The cost function can be expressed
as
(20)
with the elements of the error vector EScm being,
cov X
ML
P
ML
[ , ] { }
E
C
LS
X Y X P ( , ) P , , ( )
X P [ , ]
--------------------------------------------------
,
_
T
C
LS
X Y X P ( , ) P , , ( )
X P [ , ]
--------------------------------------------------
,
_

' ;

1
cov X
ML
P
ML
[ , ] { } E

2
C
LS
X Y X P ( , ) P , , ( )
X P [ , ]
2
----------------------------------------------------

' ;

1
cov P
ML
{ }

2
C P ( )
P
2
------------------
,

_
X
m
X = Y
m
Y = P P
true
= , ,
1
C P ( )
1
2
--- EScm k ( )
2
k 1 =
F
1
2
---EScm
T
EScm = =
2-9
(21)
see Equation (8) and Equation (14), and the second derivative is
(22)
With the substitutions X
m
=X, Y
m
=Y made, we will denote the noise-free
error vector as ESce. This still depends on the parameter vector P. For P=P
true

the elements of ESce are all zero, since the true values of the parameters
satisfy the system equation, so the second sum in Equation (22) disappears,
and Equation (19) becomes:
(23)
The problem with Equation (19) or Equation (23) is that the exact values P
true
,
X and Y are not known in practice. When the noise on the measured complex
amplitudes is small, these expressions can be well approximated by leaving X
m

and Y
m
alone, and substituting P
ML
for P.
The above statements are true when the system under test can be perfectly
modeled by the given model structure (rational form in s-domain or in
z-domain, with given orders, maybe with a given delay). If this is not true
(which is the case in a few practical cases when an approximate model is
identified, e. g., for the approximate description of a distributed system), the
estimate is not a maximum likelihood one any more, and the above variance
expressions cannot be used. In such a case the approximate covariance matrix
can be given as follows.
The uncertainty of the estimated parameters is due to the noise on the
measurements. Random deviations from the mean values are due to the fact
that the parameters are selected for each noise record to minimize the actual
cost function C. Therefore, the variations of the parameter values directly
depend on the variations of the complex amplitudes (through the cost function
C), caused by the noise. For large signal-to-noise ratio let us develop the cost
EScm k ( )
e
j
k
T
d
2
N
k
P ( , )X
mk
e
j
k
T
d
2
D
k
P ( , )Y
mk
yk
2
D
k
P ( , )
2
yk
2
N
k
P ( , )
2
2real CND
k
{ } +
------------------------------------------------------------------------------------------------------------------------- =
2
C P ( )
P
2
------------------
EScm
P
-------------------
,
_
T
EScm
P
-------------------
,
_
real

2
EScm k ( )
P
2
---------------------------- EScm k ( )

' ;

k 1 =
F
+ =
cov P
ML
{ }
ESce
P
---------------- -
,
_
T
ESce
P
-----------------
,
_
,
_
P P
true
=
1
2 Tutorial
2-10
function into Taylor series around the true values X, Y and the expected value
of P
ML
, in terms of the noise and in terms of the parameters, respectively. By
comparing the two series up to the second-order terms (the first series is
second-order, anyway), it is obtained that
(24)
with
(25)
The approximation of this expression is implemented in elis. Since the exact
values are not known, X
m
, Y
m
, and the estimate P
ML
are used instead. The
second term of Equation (25) will usually be small if no systematic modeling
errors are present, thus it will not introduce a serious error, especially since the
substitutions Y~Y
m
, etc., also mean approximations in the same order of
magnitude. A heuristic argument is that E{EScm} equals zero, and EScm is
small and more or less independent from the second derivative, so the
summation effectively averages out the random terms.
Serious model errors will cause a significant difference between Equation (24)
(Cp) and the approximation of Equation (19) (CR), this is why both quantities
are calculated in elis.
Key Features of ELiS
ELiS has been developed to handle the most important practical situations.
First of all, it takes into consideration both input and output observation noise.
These noises are present in every situation when both the input and the output
signals are measured, e. g., quantization noise can be taken into consideration
by ELiS very easily. Disregarding the input noise a common practice in
identification can lead to severe modeling errors.
However, if the input observation noise is for some reason negligible, this can
also be taken into consideration by setting
xk
= 0.
The use of measured input and output signals has the advantage that only
relative calibration of the measurement channels is necessary: as long as the
transfer functions of the two channels are close to each other, their
imperfections do not cause additional errors.
cov P
ML
{ } Q
1 ESce
P
-----------------
,
_
T
ESce
P
---------------- -
,
_
Q
1
,
_
P E P
ML
{ } =
Q
ESce
P
---------------- -
,
_
T
ESce
P
-----------------
,
_
real

2
ESce k ( )
P
2
---------------------------ESce k ( )

' ;

k 1 =
F
+ =
2-11
In the maximum likelihood approach it was assumed that the frequency
domain noise is Gaussian. This is not a severe restriction, because signals are
usually measured in the time domain, and it is easy to see that the DFT, which
is the commonly used procedure to obtain complex amplitudes, leads to
approximately Gaussian frequency domain noise, even if the time domain noise
is not Gaussian. Moreover, as shown in Pintelon and Schoukens (1991), the
properties of the estimate are robust with respect to the noise distribution
2
.
ELiS as an estimate has further attractive properties. The estimator is
asymptotically normally distributed, and it converges very well even for rather
small signal-to-noise values. An important factor of the good convergence
behavior is the well chosen initial LS step.
As a frequency domain method, ELiS is based on band-limited measurements
of the input and the output signals, and can directly identify s-domain transfer
functions. These measurements are quite easy to do with commercial
measurement devices, and by avoiding the need of intermediate discrete-time
identification, the systematic errors can be kept at a low level
3
.
An important cause of small estimation errors is the improved signal-to-noise
ratio in selected bands. There is however one important requirement: the
frequency domain data must not exhibit systematic errors, otherwise the
estimated transfer function can be biased. An important source of such
distortions is leakage: because of the always limited time record length, the
calculated digital spectrum is smeared, unless the signal consists of integer
periods of sine waves, or it is limited in time (transient signal). However,
transient signals have usually worse signal-to-noise ratio, and the Fourier
transform is subject to aliasing (if the signal is of limited duration, the
spectrum theoretically cannot be band-limited). Because of this fact, ELiS
should be used whenever possible with periodic excitation signals, in order to
exploit its accuracy. Nevertheless, aperiodic signals can also be used for
excitation, but the transfer function will be accurate in the order of leakage and
aliasing only. With sufficiently high sampling frequency and record length, or
by using advanced signal processing techniques (interpolated FFT and so on)
both effects can be reduced to an acceptable level.
2. See also: I. Kollr, On Frequency Domain Identification of Linear Systems, IEEE
Trans. on Instrumentation and Measurement, Vol. 42, No. 1, pp. 2-6, Feb. 1993.
3. J. Schoukens and R. Pintelon, Identification Why do we need it, how to use it?
Conference Record of the Instrumentation and Measurement Technology Conference
IMTC/93, 93CH3292-0, Irvine, Orange County, CA, May 18-20, 1993. pp. 246-251.
2 Tutorial
2-12
The Excitation Signals for Identification in the Frequency Domain on page
2-20 discusses the possibilities to design optimal excitation signals for the
frequency domain identification procedure.
Imposing Constraints on the Estimates
Sometimes certain properties of the transfer function are known. In these cases
you may want to impose these properties as constraints to the estimate. In this
section some possibilities for this will be discussed, as:
Fixing some parameters
Fixing some of the poles/zeros
Maintaining known partial transfer function
Fixing the value of the transfer function at certain frequencies
Looking for special forms, like allpass or linear phase
Fixing Some Parameters
In a few situations, some of the parameters can be set a priori to certain values.
For example, in the s-domain, when the bandpass or highpass nature of the
transfer function is known, some of the trailing coefficients of the numerator
(belonging to s
0
, s
1
etc.) should be set equal to zero. (The leading coefficients
need not be set equal to zero for the bandpass/lowpass case, because the order
on the numerator can simply be chosen lower.) In elis, the value of selected
parameters can be fixed, using the input argument fixp.
Fixing Some Poles/Zeros
This is a somewhat more complicated case, since poles/zeros can be fixed by
means of the coefficients only if the whole numerator or denominator can be
fixed. Therefore, the only solution is to precompensate the complex amplitudes to
be fitted (and also the variance vectors) by the term formed of the fixed poles/
zeros: instead of the equations
(26)
the modified model
(27)
Y
mk
H
fi x

k
P ( , )H
est

k
P ( , ) X
mk
N
xk
( ) N
yk
+ =
Y
mk
H
f i x

k
P ( , )
-------------------------- H
est

k
P ( , ) X
mk
N
xk
( )
N
yk
H
fi x

k
P ( , )
-------------------------- + =
2-13
can be identified. The routine modifyfv generates the modified complex
amplitudes and variances.
Maintaining Known Partial Transfer Function
Sometimes a multiplicative term of the transfer function is a priori known,
perhaps from earlier identification, or because it represents a well-designed
building block of the system. In such cases the effect of this term can be
removed from the measured data by using modifyfv. This is essentially the
same case as fixing some poles and zeros.
Fixing the Value of the Transfer Function at Certain
Frequencies
The simplest case is when the lowpass, highpass or bandpass character of an
s-domain transfer function is known: see Fixing Some Parameters. However,
in the z-domain this is not a solution any more. There may be other situations
as well, when one would like to fix, e. g., the transfer function at dc to a given
value. Using elis, there is a very simple solution for this: artificially set
complex amplitude(s) are to be added to the Fourier vector, with the input and
output variances set equal to a small value (e. g., eps or even smaller). If any
of the variance values is set to a larger value, the constraint will be
approximately followed; thus, for example, the maximum absolute value of the
transfer function in a non-measured band can be controlled by introducing
artificial zero transfer function points with large variance.
Looking for Special Forms
Sometimes the transfer function is sought in a special form, like allpass, linear
phase, imaginary (differentiator/integrator) etc. This happens when elis is
used for filter design, as in the allpass design.
The allpass constraint in the z-domain is a built-in service of the basic elis
routine. However, the other cases enumerated above will usually be well
approximated, if the amplitude vector is given in the appropriate form. For
example, linear phase FIR filters, fitted in LS sense, can be designed by setting
the delay to (order/2), and defining real amplitudes for fitting.
2 Tutorial
2-14
Solutions for Some Special Cases
It may happen that the available measurement setup does not completely
correspond to the assumptions of ELiS, or you experience difficulties fitting
your data. The purpose of this section is to provide some hints for such cases.
Dealing with Data from a Network Analyzer (Dynamic Signal
Analyzer)
When input-output measurements are made by a digitizer, the input-output
Fourier amplitudes can be determined and used for estimation. However, you
may already have an analyzer that provides the measured transfer function
points (Y
mk
/X
mk
) at given frequencies, usually on a linear grid from zero to
about half of the sampling frequency, and a parametric model of the system is
sought.
Theoretically, if you have to rely on the above transfer function estimate points
only, the noise on X
mk
may introduce an annoying bias through the division.
Variance can be reduced by simple averaging, but this will not remove this kind
of bias. You may consider taking the complex geometric mean of estimates (see
gmean), which has a smaller bias.
Having assured a possibly small bias of the transfer function points, you may
try to use elis, accepting limited accuracy. The input amplitudes can be all set
to ones, and the output ones can be set to the complex transfer function points.
A more or less acceptable result can be achieved by setting zero input and
uniform output variances, but without a correct absolute variance value the
actual value of the cost function will provide no information about the quality
of the fit and about the bias actually introduced. A plot of the complex residuals
(see rdueelis) can give some idea about the trends in the fitting error. Strongly
correlated complex error is an indication of modeling errors.
However, you can do somewhat better. If repeated measurements can be done,
you should consider variance analysis using varanal, even if only a few data
sets are available. In such a case the cost function and the mean model error
can be used for characterization of the bias.
If repeated measurements are not available, but you have some knowledge
about the input and output noises, stdtfm can be used. For this, you should
know not only the forms of the spectra of the input and output noises, which
are often white if quantization noise dominates, but also their absolute levels
or at least their levels relative to each other, to be able to combine them into
2-15
stdAm. Proper noise variances may improve the accuracy of your estimates
(with the above limitations due to the bias of the transfer function points).
Extensive bias can cause that elis tends to concentrate poles/zeros to given
frequency bands. Quite often, people incline to increase emphasis to other
frequency parts by repeating transfer function points or by artificially increase
weighting in these bands, decreasing the variance values passed to the
estimation algorithm. This is not a good practice. The general recipe is to make
better measurements by designing optimized excitation signals, improvement
of the signal-to-noise ratio, by provision of input-output Fourier amplitudes
instead of transfer function points, and so on. Or, if all these are not practicable
for some reason, partial subband fits can be tried out, as described below.
Wide-Band Model Fitting
When a system has to be modeled in a very wide band (the frequency range
stretching to several orders of magnitude), the cost function can have a very
complicated surface, and the iteration algorithms may iterate to unacceptable
local minima. In such cases it is often possible to select subbands of the
frequency range where fitting by a lower-order subsystem seems to be
reasonable. Such subfits have to be made with care, because the effect of such
subsystems can be quite intensive in neighboring bands, so the combinations
of independent subsystems can provide a very poor overall fit. A better strategy
is to use modifyfv after each fit, removing the effect of the subsystems already
obtained. By this strategy a more or less acceptable compound system can be
obtained with poles/zeros in every band of importance. Finally, the subsystems
can be refitted cyclically using the residuals of the other fits, or the whole
compound transfer function can be polished using the combination as a
starting value, and using Levenberg-Marquardt iteration in order to assure
decreasing steps of the cost function during iterations. In very wide bands,
however, the evaluation of the transfer functions, based on polyval, may
become inaccurate, so all the results should be used with precaution.
Numerical Stability and Speed of the Procedures
When performing simple calculations, MATLAB has virtually infinite precision.
However, for larger equation sets, as for the frequency domain identification of
systems of order 15-20 or higher, finite precision may play an important role,
and the results may become unreliable if elis is not applied with proper care.
2 Tutorial
2-16
The speed of calculations has been dramatically increased with the
development of modern CPU-s. However, the calculation time of certain
iterations is still not negligible. By proper use of the built-in possibilities of the
functions of this toolbox, the speed can be remarkably increased.
This subsection discusses the above aspects of the usage of the toolbox.
Numerical Stability
In elis, often a large set of linear equations is solved in every iteration step, in
order to find the necessary correction of the parameter vector. In the s-domain
the parameters are coefficients of polynomials of the variable s. It is easy to see
that when having roots around 10
3
Hz, the coefficients of higher powers of s will
be very small, typically in the order of 10
-3n
for the coefficient of s
n
. From the
numerical point of view, to handle parameters of different orders of magnitude
in parallel is very disadvantageous.
A measure of the numerical behavior of the equations is the condition number
of the matrix in the basic equation to be solved. The condition number is the
ratio of the maximum and minimum singular values of this matrix, and it
should not reach the order of magnitude of 1/eps on the given machine.
In the toolbox functions, and especially in elis, an attempt is made to avoid
numerical problems. If this cannot be maintained, warning messages are sent,
describing the concrete cause. It is therefore advisable to take each warning
message seriously, and try to find a way of better conditioning. Some hints are
given below.
In some cases elis cannot catch the problems in time, and the internal
routines of MATLAB will send messages like
Warning: Matrix is close to singular or badly scaled.
Results may be inaccurate. RCOND = 1.617346e-17
or
Warning: Divide by zero
These messages usually warn about bad conditioning.
In order to avoid numerical problems, scaling is introduced in elis for s-domain
calculations. According to experience, the equations behave the best when the
majority of the poles/zeros have their magnitudes around 1.
2-17
However, when starting identification, no usable information is at hand about
poles/zeros. This is why in elis a simple rule of thumb is used, based on the
excitation band, which is available from the Fourier data. The center radian
frequency of the excitation band:
(28)
is applied, supposing that a reasonable excitation will inject the energy in the
band of interest, where most of poles and zeros are located.
After having made a more or less acceptable fit, a better guess of the optimum
scaling frequency can be made. elis will display a suggestion for a new scaling
frequency after the last iteration, if the ratio of the actual one and the
suggestion is larger than 2 or smaller than 0.5. In the case of a system of order
20, a change of factor 2 in the scaling frequency may mean a change of 2
20
10
6

in the condition number! However, the decision is left to the user; the new
scaling frequency can be set for elis in a repeated run. The suggestion is made
either on the basis of the mean of the absolute values of all the poles and zeros,
or if these are not calculated in elis, on the basis of the ad hoc formula in
exppar (see in the Reference chapter). This latter value is accessible in the
workspace after the run of elis as pvect(2).
Bad conditioning can result from a few other causes, too. Its basic meaning is
that for the algorithm no sufficient information is provided about the
parameters, and/or the machine precision is not sufficient for the treatment of
the given problem. This can mean, for example, that the measurement
bandwidth is not wide enough, and we try to let elis extrapolate the behavior
of the system far beyond the band limit of our measurement. In such a case, no
better advice can be given than repeating the measurements in a wider band.
The treatment of high-order, wide-band systems may sometimes also require
high machine precision, even with proper measurement design.
Bad signal-to-noise ratio also results in bad conditioning: if this is the cause,
averaging can be suggested, or the redesign of the whole experiment.
A further cause can be overmodeling: the algorithm has no information for the
placement of the nonsense poles/zeros, and this fact is reflected in bad
conditioning. Decrease of the order may immediately improve the condition
number. Even undermodeling might cause bad conditioning, if the model is far
f
s
min

max
+
2
------------------------------ =
2 Tutorial
2-18
too simple even to roughly describe the system. Improper fixing of nonzero
parameters can also increase the condition number: the default setting (no
fixed parameters) is usually optimal with respect to the condition number.
elis provides three condition numbers in the vector fit.
The most relevant information about the conditioning of the solution is
provided as the condition number of the matrix actually decomposed or
inverted in the last iteration. The description is somewhat general, but this is
the way to have a number relevant for all the algorithms. For reliable results,
this number should be well below 1/eps.
However, the above condition number may depend on the applied iteration
algorithm. Newton-Raphson usually has the worst condition number, so if bad
conditioning is a danger, it should be avoided. The other extreme is
Levenberg-Marquardt (without svd): if lambda is large enough (the value 0.1 is
sufficient in general), the equations are well conditioned, but the iteration may
be painfully slow.
Most of the iteration schemes operate on the Jacobian J of the error vector. The
condition number of J is therefore relevant to the problem itself. (It should be
mentioned, however, that also scaling and measurement band selection
influence conditioning.) Should the condition number of J be too large, it is
better to rescale or to make new measurements.
The so-called approximate covariance matrix (Cp) is obtained by inverting the
approximate Hessian (see Equation (24)). Since this matrix is usually close to
J
T
J, its condition number equals approximately the square of the condition
number of J. As a consequence, the approximate covariance matrix may be
unreliable even if the solution is good. In such cases, when no or small modeling
error is anticipated, the approximate Cramr-Rao bound can be used instead
(CR), because this is calculated by singular value decomposition of J.
Speed of the Procedures
The speed of the calculations depends on two important factors: the number of
iterations performed, and the time spent on each iteration.
The number of iterations can be influenced by the starting values, the iteration
method, and the stop criteria, but depends also on the investigated problem
and the selected model order.
2-19
Most iterating functions in this toolbox have a good default setting of the
starting values, thus these should be changed with care. An example for such a
justified decision is when a more or less reasonable estimate is at hand, and
iteration is used primarily for polishing these values. This can happen when a
very wide-band fit (covering several orders of magnitude) is being done. In such
cases it can be extremely difficult to find the global minimum (or a reasonable
local minimum) of the cost function. A strategy could be to fit certain bands
separately with low-order subsystems, modify the Fourier and variance data
after each partial fit using modifyfv, and at the end, polish the compound
system in a global fit.
There is a wide choice of iteration methods in elis. As a general rule it can be
stated that the Newton-Gauss type iteration schemes (Newton-Gauss and
singular value decomposition) have about the same rate of convergence.
Newton-Gauss is quicker, but numerically less robust than svd. Near to the
optimum, Newton-Raphson is usually the quickest, but far from the optimum
it behaves worse than other algorithms. The Levenberg-Marquardt algorithms
are the most robust, but they converge rather slowly. With proper setting of
rpalg an attempt can be made to set lambda to zero (switch to the
Newton-Gauss algorithm) after a few successful iteration steps, which can
accelerate convergence. If the Newton-Gauss step is not successful, the
Levenberg-Marquardt iteration continues.
The setting of the stop criteria can also be decisive. According to experience, the
default setting (relative change of the cost function is smaller than 10
-6
) is
appropriate for most cases. A too small prescribed value may prevent stopping,
since because of numerical roundoff the derivatives will never exactly equal
zero.
The selection of the model order can also seriously influence the number of
necessary iteration steps. A too high model order means nonsense poles or
zeros, for which the measured data do not provide information. The iteration
often wanders around rather flat portions of the cost function surface.
The cycle time is also an important factor of calculation speed. In identification,
much depends on your skill, and we made an attempt to provide you with as
much graphics information as possible. But the sophisticated graphics
procedures may cost a lot of time. In most iterative algorithms the plots can be
made more rarely than in every cycle, or can even be completely suppressed.
This can dramatically improve the speed of elis, dibs, msinclip and
optexcit, but total elimination of plots can be extremely dangerous, since this
will mean blind acceptance of the results of numerical methods that might
sometimes even diverge.
2 Tutorial
2-20
Excitation Signals for Identification in the Frequency
Domain
As mentioned previously, frequency domain methods work best if periodic
excitation is applied. Nevertheless, there is still a lot of freedom left. One can
select the harmonic frequencies where the system is to be excited, and also the
amplitude and the phase of the sinusoids. Thus, criteria are necessary for the
correct decision.
The final aim is clear: the parameters of the transfer function are to be
estimated with as small error as possible. However, practical limitations may
be present: an arbitrary waveform generator is at hand or not, the actuator in
the system may allow continuous amplitude changes, or it may be an on-off
relay, the measurement equipment usually has limited amplitude resolution,
etc.
Roughly speaking, as much energy is to be injected at the interesting
frequencies, as it is possible under the given restrictions. Thus, the first task is
to select these frequencies, and determine the optimal power distribution
4
,
5
,
6
.
If at least a rough model of the system is at hand, the volume of the information
matrix can be maximized, by distributing the energy among the frequency
points. This can be done by using the routine optexcit. However, this
maximization is to be done with care, since after some iteration steps the
information only slightly increases, but the spectrum becomes spiky, and the
possibility of crest factor minimization (see the following Multisine on page
2-21) becomes limited.
A less systematic approach is to try to localize the sensitive frequency bands
(where the transfer function has a large value, or it changes rapidly, or it has
its important poles/zeros). Here users are left to their own skill and knowledge
about the system.
4. J. Schoukens and R. Pintelon, Identification of Linear Systems: a Practical Guideline
for Accurate Modeling, London, Pergamon Press, 1991.
5. K. R. Godfrey, ed.: Perturbation Signals for System Identification. Englewood Cliffs,
Prentice-Hall, 1993.
6. F. Delbaen, Optimizing the Determinant of a Positive Definite Matrix, Bulletin So-
cit Mathmatique de Belgique Tijdschrift Belgisch Wiskundig Genootschap, Vol. 42,
No. 3, pp. 333-346.
Excitation Signals for Identification in the Frequency Domain
2-21
While in MATLAB it is easy to define a linear grid, is not that simple to design
a quasi-logarithmic one (near to logarithmic, on the linear grid of the FFT). The
routines lin2qlog and log2qlog serve this purpose.
The following subsections deal with the two most important cases: the first one
is when there is no restriction imposed on the time domain signal form, i. e., a
general multisine can be used, the other one is the case of a binary excitation
signal.
Multisine
Having determined the desired power distribution, users can still choose the
phases. The phases of the components of a multisine have important influence
on the time domain signal shape: by proper choice of the phases the maximal
peak can be significantly compressed, allowing larger energy to be injected for
the given input range of the measurement device, or keeping the system in the
linear working region. This procedure is called crest factor minimization. The crest
factor is defined as the ratio of the maximal absolute peak value to the effective
value of the signal in the frequency band of interest:
(29)
The achievable decrease in the crest factor is a factor of 2-3 for nearly uniform
spectra, compared to the crest factor of a random phase sine wave.
Crest factor minimization can be performed on a single signal, or on two signals
related to each other by a linear system (input-output optimization). Both tasks
are implemented in the routine msinclip. This routine is based on consecutive
transforms from the time domain to the frequency domain and vice versa,
applying a mild clipping in the time domain, and combining the new phases
with the desired amplitudes in the frequency domain
7
.
The optimized signal can be applied to the system, using an arbitrary
waveform generator. This usually works with a zero-order hold (a D/A
converter), which introduces some amplitude distortion:
7. E. van der Ouderaa, J. Schoukens and J. Renneboog, Peak Factor Minimization, Us-
ing TimeFrequency Domain Swapping Algorithm, IEEE Trans. on Instrumentation
and Measurement, 1988, Vol. 37, No. 1, pp. 144-147.
cr
max abs x t ( ) ( ) { }
V
RMSeff
--------------------------------------- -
max abs x t ( ) ( ) { }
V
RMS
in-band power
total power
----------------------------------
----------------------------------------------------- = =
2 Tutorial
2-22
(30)
where t is the reciprocal of the clock frequency. This distortion increases
towards the half of the clock frequency: here the value of the transfer function
is approximately 2/0.6366, instead of 1 (see Figure 2-2).
Figure 2-2: The amplitude distortion introduced by a zero-order hold
The routine msinprep, which generates a time series for the arbitrary signal
generator, can perform a precompensation for this distortion. This prepares the
data vector to be downloaded in to the arbitrary signal generator, producing
the desired waveform (if no other amplitude or phase distortion is introduced
in the chain).
Binary Excitation Signals
There are situations when only binary excitation signals can be used.
Theoretically the state transition instants can be anywhere within the signal
period, but in practice these signals are produced using a high-frequency clock.
H
zoh
f ( )
f t ( ) sin
f t ( )
----------------------- =
Excitation Signals for Identification in the Frequency Domain
2-23
Thus, users can choose the clock periods where the state is inverted (discrete
interval binary sequence). This does not achieve exactly the desired power
distribution, but the approximation is generally usable.
In order to approximate the desired spectrum, a similar algorithm,
implemented in msinclip, can be applied. The difference is that instead of
clipping, a comparator is used in the time domain, resulting in a binary signal
8
.
This algorithm, implemented in the routine dibs, converges rapidly, but in the
result there may be frequencies where only a fraction of the desired power is
present. To improve this property, a systematic search can be performed using
the routine dibsimpr, which tries to maximize the minimum relative power
(calculated with respect to the desired power at a given frequency)
9
.
If an approximately white spectrum is acceptable, the so-called maximum
length binary sequences (or pseudorandom binary sequences, PRBS)
10
can be
also used (routine mlbs). These sequences are produced using binary shift
registers with appropriate feedback. Since mlbs can be considered as a
pseudorandom sequence applied to a zero-order hold, it is easy to see that the
amplitude spectrum is modified again by the transfer function depicted in
Figure 2-2.
This spectrum is valid for the continuous case; however, because of the rich
overharmonic contents, the use of an anti-aliasing filter in the measurement
setup is advisable to avoid undesirable spectral distortions due to aliasing.
The length of the sequence is N = 2
n
-1, where n is the register length, thus with
synchronized sampling the length of the time record is not a power of two.
Consequently, instead of standard base-2 FFTs, some special DFT algorithm
has to be used (see the section Transformation into the Frequency Domain).
However, there is also another possibility: by careful adjustment of the
sampling frequency, the number of samples in a period of the excitation signal
can be set to a power of two, thus a standard FFT can be applied.
8. A. van den Bos and R. G. Krol, Synthesis of Discrete-Interval Binary Signals with
Specified Fourier Amplitude Spectra, International Journal of Control, 1979, Vol. 30,
No. 5, pp. 871-884.
9. K.-D. Paehlike and H. Rake, Binary Multifrequency Signals Synthesis and Appli-
cation, Proc. 5th IFAC Symposium on Identification and System Parameter Estimation,
Darmstadt, FRG, Sept. 24-28, 1979. Vol. 1, pp. 589-596.
10. K. R. Godfrey, ed.: Perturbation Signals for System Identification. Englewood Cliffs,
2 Tutorial
2-24
Preprocessing of Data
Before starting the estimation procedure, the measured data have to be
prepared. Though ELiS works well for rather small signal-to-noise ratios, the
quality of the estimate can be improved by decreasing the measurement noise.
Thus, for precise estimations averaging is usually recommended. Moreover,
the noise often has to be analyzed in order to determine the necessary variance
values; variance analysis can be done in parallel with averaging.
According to experiences with ELiS, the measurement of a few periods is often
sufficient for obtaining a usable estimate. By segmenting the records into
equal-length parts which contain full periods of the periodic excitation signal,
an approximate noise analysis can be performed, and then the estimation can
be performed on the averaged data.
Preprocessing in the Time Domain
Averaging can be done directly in the time domain. For this, the measurements
have to be started in synchronization with the excitation signal. To achieve
perfect synchronization, the clock of the measurement equipment and that of
the signal generator have to be synchronized (it is not enough to have very
stable clocks, because they may slip slowly with respect to each other). With no
proper synchronization, time domain averaging and noise analysis are
hopeless; however, as shown in the next subsection, in the frequency domain
there is a quite effective way to synchronize the measurement records even if
the measurements themselves were not synchronized.
The time domain variance can be determined in parallel with the averaging.
Since the noise on the signals is assumed to be ergodic, the following formula
can be applied:
(31)
where M is the number of measurements, N is the number of measured points
in each measurement, and x
av
denotes the result of averaging over different
xt
2 1
N
--- -
1
M 1
------------- -
n 1 =
N
x
m
n t ( ) x
av
n t ( ) ( )
2
m 1 =
M
=
1
N
--- -
1
M 1
------------- -
n 1 =
N
x
m
2
n t ( ) Mx
av
2
n t ( ) ( )
m 1 =
M
=
2-25
measurements. This estimate is c2 distributed, with r = N(M-1) degrees of
freedom. The confidence interval for the confidence limit a can be obtained
using the corresponding c2-values (see also Table 2-1):
(32)
The
2
-values can be taken from statistical tables.
For r = N(M-1) > 30 and small , a useful approximation
11
for the terms in
Equation (32) is:
(33)
where = /2 or = 1-/2, and u
is the corresponding abscissa of the standard

normal distribution.
11. G. A. Korn and T. M. Korn, Mathematical Handbook for Scientists and Engineers:
Definitions, Theorems and Formulas for Reference and Review, 2nd ed, London,
McGraw-Hill, 1968.
Table 2-1: Confidence intervals for the variances (multipliers of the empirical
variance)
r 90% 95% 98%
10 0.55 2.54 0.48 3.12 0.43 3.91
20 0.64 1.84 0.58 2.10 0.53 2.42
30 0.69 1.62 0.63 1.80 0.59 2.01
40 0.72 1.51 0.67 1.64 0.63 1.81
50 0.74 1.44 0.70 1.55 0.66 1.69
100 0.80 1.28 0.77 1.35 0.74 1.43
200 0.85 1.89 0.83 1.23 0.80 1.28
P
N M 1 ( )
2
N M 1 ( ) 2 ;
------------------------------------------
xt
2
xt
2 N M 1 ( )
2
N M 1 ( ) 1 2 ;
---------------------------------------------------
xt
2
< <
,
_
=
r
r ;
2
---------
1
1
2
9r
----- u
2
9r
----- +
,
_
3
--------------------------------------------- =
2 Tutorial
2-26
If you assume that the noise spectrum is approximately white, the time domain
variance data can be directly used for the calculation of the variances of the
real and of the imaginary parts of the complex amplitudes:
(34)
However, when the noise spectrum is to be determined, and/or the
input-output noise covariances have to be calculated, the auto- and
cross-correlation functions have to be estimated. This is usually done via the
frequency domain, using FFTs; in this case it is better to do averaging and
noise analysis in the frequency domain.
Transformation into the Frequency Domain
The standard procedure of time domain to frequency domain conversion is the
fast Fourier transform (FFT). In the case of periodic signals this will not
introduce spectral leakage if the time record contains integer numbers of
periods of each component. Leakage can be detected by throwing a glance at the
frequency spectrum, if the signal contains a few spectral lines only: it appears
as skirts around spectral lines (Figure 2-3).
N = 100; t = [0:N-1]; %N = 100 points
%two sine waves:
xt = sin(2*pi*10/N*t)+0.5*sin(2*pi*20/(N-1)*t);
Xfa = abs(fft(xt)); %spectrum
Xsh = [zeros(1,N);Xfa;zeros(1,N)];
Xsh = [0;Xsh(:);0]; %prepare for bar plot
tsh = [t;t;t]/N; tsh = [0;tsh(:);1]; plot(tsh,Xsh)
500 0.90 1.11 0.89 1.14 0.87 1.16
100
0
0.93 1.08 0.92 1.09 0.90 1.11
Table 2-1: Confidence intervals for the variances (multipliers of the empirical
variance)
r 90% 95% 98%
x
2 N
2
--- -
xt
2
=
2-27
Figure 2-3: Absolute value of the DFT of two sine waves
The first sine wave (f
1
= 10/100 = 0.1) exhibits no leakage because exactly ten
periods have been measured, while the second one (f
2
= 20/99 0.2) shows
considerable leakage.
The FFT is base N, where N is the record length. If N is not a power of two,
special techniques (like the chirp z-transform
12
) have to be applied. MATLABs
fft routine can effectively transform sequences of arbitrary length. The
applied formula is
12. L. R. Rabiner and B. Gold, Theory and Application of Digital Signal Processing, En-
glewood Cliffs, NJ, Prentice-Hall, 1975, pp. 393-399.
X
k
x
i
e
j 2
ki
N
----
i 1 =
N 1
=
2 Tutorial
2-28
Preprocessing in the Frequency Domain
When the measured input and output signals have been transformed into the
frequency domain, averaging and noise analysis (including the estimation of
the input/output covariances) can be performed using the routine varanal.
This routine provides the results in the form necessary for elis, and even
calculates the 95% confidence intervals of the variances. As a rule of thumb,
input-output covariances should be calculated and passed to elis. If they are
small, they make no harm; if they are large, they have to be considered.
Post-Measurement Synchronization
The routine varanal can, if required, make an attempt to synchronize the
measurements to the first one, by looking for the delay which results in the
smallest phase differences. The routine works well even for as small as 3 dB
frequency domain signal-to-noise ratios. If the synchronization attempt fails, it
sends an error message.
The synchronization algorithm in varanal is based on an approximation of the
maximum likelihood estimation of the delay
13
. It is worth mentioning that for
the maximum likelihood estimation of the delay between records elis can also
be used, identifying a hypothetical pure delay transfer function between the
two records, considering them as input and output.
A rough guess of the shape of the transfer function can be quickly obtained by
averaging the transfer function estimates, calculated as the ratio of the
measured output and input complex amplitudes. Since the bias is much
smaller for the geometric mean of the transfer function estimates than for the
simple arithmetic one
14,15
, the geometric mean is to be used for averaging. The
routine gmean calculates this mean value for complex numbers, without
suffering from phase wrapping problems.
13. I. Kollr, Signal Enhancement of Non-Synchronized Measurements for Frequency
Domain System Identification, IEEE Trans. on Instrumentation and Measurement,
Vol. 41, No. 1, pp. 156-159, Feb. 1992.
14. J. Schoukens and R. Pintelon, Measurement of Frequency Response Functions in
Noisy Environments, IEEE Trans. on Instrumentation and Measurement, Vol. 39, No.
6, pp. 905-909, Dec. 1990.
15. P. Guillaume, R. Pintelon and J. Schoukens, Nonparametric Frequency Response
Function Estimates Based on Nonlinear Averaging Techniques, IEEE Trans. on In-
strumentation and Measurement, Vol. 41, No. 6, pp. 739-746, Dec. 1990.
2-29
A Priori Known Partial Transfer Function
Sometimes a term of the transfer function is already known (e. g., from earlier
identification or because it represents a block which was carefully installed to
meet strict specifications). It is not worthwhile to identify this term again,
since this may deteriorate the variance of the estimates of other parameters.
The routine modifyfv can be used to precompensate Fourier and variance files.
2 Tutorial
2-30
Presentation of the Results
Identification is a science and an art at the same time. Much depends on the
skill and imagination of the person, who identifies a given system. This toolbox
seeks to provide as much illustrative information as it is possible during the
whole procedure.
Therefore, during the most important procedures of the Frequency Domain
System Identification Toolbox, it is possible to follow the iterations on graphs.
Though this movie can be switched off, to accelerate the run, we recommend
that the user follows the iterations, and intervenes if necessary. Nothing is
more dangerous than to blindly believe the results of programs, even if they are
well tested and reliable.
In addition to the graphs, there are two plotting programs for the illustration
of the result of elis.
ploteltf plots magnitudes and phases of parameter files. It can plot the
transfer function of two parameter sets and, in addition, the data given in a
Fourier file, in order to facilitate comparisons. Moreover, on one of the transfer
functions confidence intervals can also be plotted, to give an insight into the
reliability of the estimation results.
The uncertainty of the estimated transfer function is usually much smaller
than the scattering of the Y
mk
/X
mk
points. This has to be so, since the estimated
transfer function is based on all measured data, and like that, its variance is
roughly inversely proportional to the number of measurement points, while the
variance of the points Y
mk
/X
mk
depends only on the input and output
amplitudes and measurement noises. On the other hand, almost all of the
confidence intervals of the nonparametric transfer function estimates Y
mk
/X
mk

should cover the parametric transfer function estimates calculated from pdat1
or pdat2.
The confidence interval plots are made by internally using the results of the
routines stdtf and stdtfm. When nonstandard forms of plots are preferred,
these routines can be directly used for the calculation of the standard
deviations of the magnitudes and phases.
If possible, we suggest you plot not just one amplitude set, but the result of
averages. If for some reason synchronization failed even with the routine
varanal, it is recommended to average the ratios of output and input
amplitudes via the routine gmean (see Section 4.3).
2-31
plotelpz plots pole/zero patterns of identification results, along with the
uncertainty ellipses. This is an often studied plot of dynamic systems.
However, this plot does not provide information on the coupling of poles and
zeros. The standard deviations and covariances of the poles and zeros can be
directly calculated by the routine stdpz for the preparation of special-form
plots. Coupling between poles and zeros can be detected from high-value
cross-correlation terms in the output argument rzp.
Both plotting routines have a number of optional arguments. Besides the usual
purposes, they can be used as sophisticated statements in script files,
preparing graphs for documentation, archiving and illustration of papers.
Only a small part of the demands can be covered by standard routines. Since
the Frequency Domain System Identification Toolbox uses the standard
MATLAB representations of s- and z-domain polynomials, script files can be
easily written to show other possible plots, as Nyquist, Nichols diagrams etc.
To obtain high-quality illustrations for scientific papers, etc., the plots made in
MATLAB can be saved in PostScript form. Data vectors can be stored into
ASCII files, using the routine expvect, for export to sophisticated graphing
programs.
2 Tutorial
2-32
Model Validation and Simulations
A basic rule for scientists and engineers using powerful computers is to never
believe the results of any numerical procedure, unless these can be checked in
an independent way. This is especially true for identification, where even the
right order of the system model is unknown. Therefore, model validation is
crucial. In this section the validation possibilities will be summarized, in the
Frequency Domain System Identification Toolbox and in the environment of
MATLAB.
The First Quick Checks
The best thing one can do is to observe the internal iteration steps of the
procedures on the presented graphs, and to plot the transfer function and the
pole/zero pattern of the result with the uncertainties (see the Presentation of
the Results on page 2-30). From these plots rough errors (instability, roughly
bad fit, outliers, etc.) can be easily detected.
Stability and the Choice of the Proper Delay
Instability of the fitted models is often a problem. The studied physical systems
are usually stable (otherwise measurements cannot be done, except if the
unstable system is stabilized by a feedback loop), thus for our purpose unstable
models are usually not acceptable.
Unstable identified model of a stable system can be due to several causes, like
Overmodeling or undermodeling (too large or too small model orders)
Improperly chosen value of the delay
Outliers
Too small signal-to-noise ratio
Local minimum of the cost function
Overmodeling and undermodeling are going to be discussed in the next section.
Improper choice of the delay is a common source of instability. The reason is
that a too large or a too small delay value drastically changes the phase
behavior of the identified rational form, and thus can easily deteriorate the
fitting.
2-33
One possibility is to let also the delay be estimated. This is an attractive idea,
however, especially for high-order systems, the convergence region of the delay
is narrow, thus the estimated value will strongly depend on the starting one.
One can make fits with different delay values around the a priori guess.
Usually the fits become better when approaching the true value.
Another possibility is the study of the estimation results. In the zdomain too
small delay values usually result in small values of the leading coefficients of
the numerator
16
. In both domains, a too small delay value often gives unstable
models. The gradual increase of the delay can lead to the good estimate.
The rest of the causes can usually be recognized by looking at the transfer
function or pole/zero plots, estimation with different (eventually more
effectively averaged) measurements, or eventually by estimations on
simulated data.
Detection of Undermodeling and Overmodeling
Overmodeling means that the model order is too high, while undermodeling
means that it is too small. Both should be avoided. However, in practical
situations two important limitations should be considered. First, a linear
transfer function is often an approximate model only of the physical system:
some small nonlinearities, some distributed parameters, etc., may always be
present. Therefore, we usually have to content ourselves with a limited
modeling error of our identification result. Second, it cannot be expected to
decrease modeling errors significantly below the random errors, unless new
measurements with smaller noise (more averaging, etc.) are done.
Undermodeling is treated in detail in Chapter 5 of the book of Schoukens and
Pintelon (1991). Here only the most important statements will be summarized.
The best indicator of a bad model is the too large value of the cost function. As
it was established in Section 2, in the case of a good model the double of the cost
function is a random variable of
2
-distribution, with 2Fn
p
degrees of
freedom. Consequently, the standard
2
-test can be applied: if the value is too
large, modeling errors are present.
However, not only undermodeling can result in a large value of the cost
function. A wrong value of the delay, or incorrectly small variance values can
also increase the value of it.
16. L. Ljung, System Identification Toolbox for Use with Matlab. User's Guide, July,
1991. The MathWorks, Inc. p. 1-64.
2 Tutorial
2-34
If modeling errors are present, it is often desirable to obtain some information
about the extent of these errors. The so-called mean model error may provide a
rough measure. For an approximately constant transfer function value H
e
(the
subscript e refers to the exact values), approximately constant excitation signal
amplitudes, and approximately flat noise spectra with variances of the real and
the imaginary parts and
the mean model error is
(35)
where h = H - H
e
is the complex modeling error of the transfer function, C is
the value of the cost function, F is the number of frequencies, and n
p
is the
number of estimated parameters. For non-uniform H
e
, the mean modeling
error can be averaged over the frequencies used in the estimation, as it is done
in elis, but the result has to be interpreted with precaution.
Using h
mean
, special care has to be taken to use correct variance values, since
these directly scale the value of C. Too large variance values lead to a too small
cost function (and maybe to an imaginary h
mean
), too small variance values
result in a too large cost function and a large mean model error.
A confidence interval plot of the magnitude and phase of the estimated transfer
function may provide important information. The standard deviations of the
magnitude and phase are also accessible using stdtf.
When overmodeling occurs, the estimation procedure is forced to find additional
poles/zeros which are not relevant to the measured data. This will increase the
condition number of the normal equations (the condition number is accessible
as an output argument of the routine elis). Moreover, the variance of the
estimated parameters will also increase, which can be detected easily if
additional estimations are done on different measurement data.
For a too large number of parameters, much more iteration steps are necessary
to reach the minimum of the cost function.
It is even easier to detect overmodeling using the pole/zero plot, preferably with
the uncertainty ellipses. The superfluous poles and zeros will have a large
uncertainty, poles and zeros often appear in pairs (nearly canceling each
other), and at different locations in identifications of different measurement
data. They are often located outside the measured band. (Note that sometimes
out-of-the-band poles and zeros do belong to the desired model, far from the
x
2
y
2
h
mean
2
2C 2F n
p
( )
F X
2
------------------------------------ H
e
2
x
2
y
2
2real c
xy
H
e
{ } + ( )
2-35
imaginary axis in the Laplace-domain, or far from the unit circle in the
z-domain.)
The standard deviations of the poles and zeros are also accessible using stdpz.
Unnecessary pole/zero pairs can also be detected by studying the correlation
coefficients between poles and zeros (see the output argument rpz of stdpz).
Pole/zero uncertainties can be studied by calculating several sets of roots using
random perturbations of the parameters, according to the covariance matrix,
and by plotting all the pole/zero sets using plotelpz.
A further possibility is to evaluate the Akaike criterion for identification results
of different orders, and find the one with the smallest AIC value:
(36)
When two or more independent data sets are available, cross validation can
indicate overmodeling (see also Cross Validation with Another Set of
Measured Data on page 2-40). A too high-order model tends to follow the noise
patterns. The cost function for this model, evaluated with an independent data
set, is therefore significantly larger than with the data used for modeling.
Finally, the so-called F-test (Sderstrm and Stoica, 1989) can also be used. If
two identified models are available with cost function values C
1
and C
2
, and
numbers of estimated parameters n
p1
< n
p2
, where second model contains the
first one as a special case, and both models are good, the expression
(37)
is F-distributed (this has nothing to do with the number of frequencies F !), with
the degrees of freedom (n
p2
-n
p1
,2F-n
p2
), and can be tested using standard
statistical tables. If the hypothesis can be accepted, the lower-order model is to
be chosen.
For large values of F the relative variance of C
2
becomes negligible, thus
C
2
2F-n
p2
2F, and the variable
(38)
is approximately
2
-distributed with n
p2
-n
p1
degrees of freedom, allowing the
use of the
2
-test.
AIC = 2(C+n
p
)
C
1
C
2
C
2
-------------------
2F n
p2
n
p2
n
p1
---------------------- =

x 2F
C
1
C
2
C
2
------------------- =
2 Tutorial
2-36
Study of the Residuals
As shown in statistics, whenever a good model of the system is obtained from
the maximum likelihood formulation for Gaussian noise, the residuals (the
deviations of the data from the estimated values) will exhibit some
characteristic properties. They will be approximately independent and
normally distributed with zero mean and the given variances. These properties
can be tested.
The routine rdueelis calculates the residuals for the model given in Equation
(3). However, care should be taken, because estimates of the so-called nuisance
parameters (X and Y) are not consistent: with the increase of the number of
measured points, their variance remains finite. Their estimation error can only
be decreased by processing several data sets (several experiments) in parallel.
As an effect, though the input and output residuals will have the above
approximate properties, they will be correlated with each other at each point.
Therefore, the routine calculates also the equation error vector:
(39)
The two terms in Equation (39) are positively correlated, since a deviation of
Y
m
/X
m
from its theoretical value will pull the estimates in the direction of
the deviation. It can be shown that
(40)
The second variance in Equation (40) is usually much smaller than the first
one, since the estimated parametric transfer function is a kind of average over
the noisy Fourier amplitudes. It is calculated by stdtf.
When the above two variances are in the same order of magnitude, the variance
of the residual may become very small. This small variance should be
interpreted with care: it usually means that the model follows the noise, so the
averaging effect of the model fitting does not work at the given frequency. The
best check of this is to compare the variance with the result of stdtf and
stdtfm. Theoretically, Equation (40) is non-negative, if evaluated with the
same variance values as used for the estimation. However, because of the
approximations, sometimes small negative values may appear in Equation
(40).
r
eqk
Y
mk
X
mk
--------- H
k
( ) =
var r
eqk
{ } var
Y
mk
X
mk
---------

' ;

var H
k
( )

' ;

=
2-37
The first variance in Equation (40), or rather its half, that is, the variances of
the real and imaginary parts of Y
m
/X
m
, can be calculated as follows:
(41)
using the substitution
(42)
which usually has a smaller error than the estimated value of Y
k
/X
k
.
These variance values result from the linear approximation of the division, and
Equation (41) uses the inconsistent estimates of Y
k
and X
k
, but for large
signal-to-noise ratios the approximation is good.
A very powerful method is cross validation (see Cross Validation with Another
Set of Measured Data on page 2-40). Overmodeling is indicated by large
residuals. Undermodeling means modeling errors that appear in the residuals
for any data set. Statistical analysis of residuals of a series of data sets can
quickly reveal such patterns.
Simulations
An important tool for the study of the properties of the estimates and of the
behavior of the whole procedure is the simulation of data vectors. The
Frequency Domain System Identification Toolbox provides two possibilities for
this purpose.
simfou generates simulated frequency domain data from the system model and
the given variance vectors. This means that the preprocessed data are
simulated.
It is also possible to simulate time domain data, using the routine simtime. This
simulation has the advantage that the whole data processing chain is checked,
from the actual measurements to the estimates. Moreover, by using this
facility, frequency domain identification can be compared directly to time
domain methods. simtime can generate both the transient response and the
steady-state response of the system.
var Re r
eqk
( ) { } var Im r
eqk
( ) { } H

k
( )
2

xk
2
X
k
2
-----------

xy
2
Y
k
2
----------- 2real
c
xy
X
kY
k
------------

' ;

+
,

_

H
k
( )
Y
k
X
k
----- -
2 Tutorial
2-38
There is just one difficult point in this simulation. elis uses a nonparametric
representation of the noise variances, defined only at the frequencies of
interest. However, for the time domain simulation full information on the
correlation function or the power spectral density of the noise is required. The
routine simtime will read in the parameters of a filter, which produces the
colored system noise from a white input sequence.
If the noise is approximately white, there is no such problem, the transfer
function of this filter is constant. However, for colored noise a solution still
needs to be found.
In ELiS, the noise characteristics are hidden in the variance vectors. In
principle, an additional identification step would be necessary to obtain the
values of the parameters of the noise shaping filters. The difficulty is that
amplitude-only information is available, thus this identification cannot be
directly done by using elis. A correct solution is to construct an estimate for
this purpose, but this is rather involved
17,18
. Fortunately, the required
accuracy allows the use of approximate methods, since the exact modeling of
the noise spectrum is usually not critical in the evaluation of the identification
procedures.
The knowledge of the physical system may often help. If the noise source is
located in the system under test, and you have some idea about the path of the
noise to the measurement point, a rough approximation can be given.
The phase of the transfer function can be freely chosen. Thus, filter design
methods which approximate a given magnitude response can be used. In
MATLAB, the Signal Processing Toolbox routine yulewalk can be used for IIR
filter approximation with orders n/n, or the routine remez for linear phase FIR
approximation. However, care should be taken with both methods, because the
variance values usually have stochastic nature, and this can lead to gross
errors. Furthermore, yulewalk fits the autocorrelation function in the time
domain, and this may result in unexpected bias, especially because the built-in
Hamming window
19
. For these reasons, the above possibilities should be
considered as rough approximations of the noise spectrum.
17. Y. Rolain, R. Pintelon and J. Schoukens, Amplitude-Only versus Amplitude-Phase
Estimation, IEEE Trans. on Instrumentation and Measurement, Vol. IM-39, No. 6, pp.
818-823, Dec. 1990.
18. Y. Rolain, Identification of Linear Systems from Amplitude Information Only, Ph.D
Thesis, Vrije Universiteit Brussel, Dienst ELEC, Brussels, Belgium, 1993.
19. The built-in Hamming window can be switched off, as it has been done in the routine
ywalk.
2-39
Figure 2-4 shows a simulation example of the approximation of a second-order
noise variance shape, where the variance values were estimated from 20
measurements, using the expression of the empirical variance. The simulation
has been performed using the following statements:
num = 1; r = 0.7; phi = pi/4;
den = real(poly(r*[exp(j*phi),exp(-j*phi)])); %denominator
F = 50; fv = [0:F]'/F/2; %frequency points
%Exact transfer function:
ea = j*fv*2*pi; N = 20;
tf = polyval([0,0,num],exp(ea))./polyval(den,exp(ea));
%simulated random variables:
tfn = abs(tf).*sum(randn(2*N-2,length(fv)).^2)'/(2*N-2);
%transfer function by no windowing version of yulewalk:
[num2,den2] = ywalk(2,2*fv,tfn);
tf2 = polyval([0,0,num2],exp(ea))./polyval(den2,exp(ea));
plot(fv,abs(tf),':w',fv,abs(tf2),'-w',fv,tfn,'+w')
title('Magnitude')
Figure 2-4: Result of the approximation of the noise spectrum, using ywalk
2 Tutorial
2-40
The + signs denote the measurements, the continuous line shows the result
of ywalk, while the dotted line marks the exact shape.
Simulations can even be performed starting from pole-zero models. The
function fdcovpzp converts these models with their uncertainties into transfer
function data, used by simfou and simtime.
Cross Validation with Another Set of Measured Data
A final test of the quality of the identified model is cross validation. This means
that the model is evaluated using a different measured data set. The residuals
and the cost function of a model can be evaluated for different data sets (see
Detection of Undermodeling and Overmodeling on page 2-33 and the Study
of the Residuals on page 2-36). A further possibility is to make a new fit, with
the LS starting values or to start from the previous model. The identified
parameters should be equal to the previously obtained ones within the
uncertainty bounds; large deviations may indicate overmodeling.
Model Conversions from/to the System Identification Toolbox
2-41
Model Conversions from/to the System Identification
Toolbox
The basic aim of the Frequency Domain System Identification Toolbox is the
same as of the System Identification Toolbox: to identify linear systems from
measured data. The main difference is that the Frequency Domain System
Identification Toolbox works in the frequency domain, while the System
Identification Toolbox works in the time domain. Thus, they are
complementary to each other, and both can sometimes be used to solve a given
problem. In this case, their results should be directly compared. This
comparison can be easily done if the identified models are in the same format,
because the results (pole/zero pattern, transfer function etc.) can be presented
in a similar form. This section discusses the possibilities of conversion between
the two toolboxes.
In the Frequency Domain System Identification Toolbox the investigated
system model is as follows:
(43)
where X
k
, Y
k
(k = 1,2...F) are the complex input and output amplitudes, and
is a delay operator:
(44)
in the s- or the z-domain, respectively, and T
d
is the delay, N(
k
) and D(
k
) are
polynomials of
k
, and
k
= s
k
= j
k
in the s-domain, or
k
= z
k
= exp(j
k
T) in
the z-domain. The observation equations are
(45)
where X
mk
and Y
mk
are the measured complex amplitudes, while N
xk
and N
yk

are the measurement noises.
Y
k

N
k
( )
D
k
( )
----------------X
k
=
e
s T
d
= or e
j 2f
s
T
d
f
s
T
d
= =
Y
mk
Y
k
N
yk
+ = and X
mk
X
k
= N
xk
+
2 Tutorial
2-42
Substituting Equation (45) into Equation (43), the following equation is
obtained:
(46)
In Equation (46) the statistical properties of the noises are assumed to be
known at the beginning of the identification of N(
k
), D(
k
) and eventually .
Thus, the noise analysis is to be done separately, using the available measured
data.
In the time domain System Identification Toolbox, identification is performed
in the discrete time domain. The general model of the system (with the
exception of the nonparametric methods) is in z-notations:
(47)
where A(z), B(z), C(z), D(z) and F(z) are polynomials of z, Y(z) is the measured
signal, nk is the delay (an integer value), U(z) is the input signal, and E(z) is
white noise. The orders of the polynomials are na, nb and so on. The methods
included into the System Identification Toolbox treat special cases of the
general model:
ARX: nc = nd = nf = 0 (instrumental variables can be used)
ARMAX: nd = nf = 0
ARARX: nc = nf = 0 (generalized least squares model)
ARARMAX: nf = 0 (extended matrix model)
OE: na = nc = nd = 0 (output error model)
BJ: na = 0 (Box-Jenkins model)
The methods usually minimize the prediction error; for the ARX case the
instrumental variable method can also be used.
The parameters of the model Equation (47) are stored in the so-called
theta-format. In what follows, this model will be referred to as theta.
Y
mk

N
k
( )
D
k
( )
---------------- X
mk
N
xk
( ) N
yk
+ =
A z ( )Y z ( )
B z ( )
F z ( )
----------- z
nk
U z ( )
C z ( )
D z ( )
-----------E z ( ) + =
2-43
When comparing the two models given in Equation (46) and Equation (47), the
following significant differences can be noticed.
Because of the above important differences, the models can only be converted
into each other with restrictions. Also, there are some less important
differences: the theta-format can handle integer delays only, and the leading
coefficients of A(z), C(z), D(z) and F(z) must be ones, while in the Frequency
Domain System Identification Toolbox there are no such restrictions.
Conversions are facilitated by the routines elis2tha and tha2elis in this
toolbox.
Conversion from ELiS to the theta-Format
It is clear from Equation (46) and Equation (47) that for z-domain models the
conversion is more or less straightforward, while s-domain models need special
considerations.
Conversion of Discrete Time Models
The main model parameters have direct equivalents:
(48)
With na set equal to 0, the only remaining task is to define the equivalents of
the noises. Here troubles arise again, since on the one hand ELiS does not use
a parametric model of the noise, and on the other hand, the theta-format does
not model the input noise.
Table 2-2: Comparison of system model
ELiS theta
s- and z-domain z-domain
input and output noise one noise source (output noise)
nonparametric noise model parametric noise model
also fractional delay delay integer multiple of 1/f
s
single input (presently) also multiple input
N z ( ) B z ( ) D z ( ) F z ( ) f
s
nk , ,
2 Tutorial
2-44
Let us assume first that the input noise is negligible, and the variance of the output
noise is approximately constant along the frequency axis. In this case, since na = 0,
{N
yk
} will correspond to a white time domain noise, thus simply nc = nd = 0 can
be chosen in Equation (47) (C(z) = 1, D(z) = 1), and the variance value can be
passed.
For the comparison of identification results, the properties of the estimates of
the transfer function itself are important. This is the suggested procedure even if
the above assumptions do not hold.
Approximations for Colored Noise and input/output Noise. If the results of
ELiS are to be used for simulations of time domain data, and the variance is
significantly different at different frequencies, the above described solution for
model conversion may not be acceptable. In this case, the user has to provide a
parametric description of the noise spectrum. However, rather than trying to
convert the model into the theta-format, we suggest using the simtime
simulation routine, which generates time domain data that can be used by both
toolboxes. (Be aware of the fact that simtime also requires a parametric
description of the noise spectrum.)
For obtaining a usable (but approximate) parametric model of the noise,
solutions are suggested in the Simulations section.
If the input noise cannot be neglected, the situation is even more complex.
simtime can simulate both input and output noises, thus this routine can be
used to get realistic simulation data.
Here the question arises, what will the time domain methods identify in the
case of input-output noises. Since they will consider the input data as the exact
input sequence, it is easy to see that in the result they will reduce the input noise
to the output of the system. The resulting noise they identify in the model is in the
prediction error method
(49)
Since the frequency domain noises are assumed to have a rotationally
symmetric two-dimensional probability distribution at each frequency:
(50)
the variance of the combined noise will be
N
red
z ( ) H z ( )N
x
z ( ) N
y
z ( ) + =
var Re N
x
( ) { } var Im N
x
( ) { } = cov Re N
x
( ) Im N
x
( ) { , } 0 = ,
2-45
(51)
with
C(z)/D(z) will approximate the spectrum of this reduced noise.
Conversion of the Covariances. The routine elis2tha will also accept the
covariance matrix of the estimated parameters, and pass it to the theta-format.
However, since the leading coefficient of F(z) must be equal to 1 in the
theta-format, the polynomial needs to be scaled by the first coefficient. If this
coefficient has been estimated (i. e., its variance is not zero), the scaling will
change the variances of the denominator coefficients. elis2tha will make an
attempt to calculate the new covariance matrix by linear approximation. If the
linear approximation cannot be applied, it sends a warning message, and
passes a covariance matrix full of zeros.
Conversion of Continuous-Time Models
Continuous-time and discrete-time systems should be transformed into each
other with extreme care
20
. Since ELiS is based on band-limited measurements,
the best recipe is to repeat the identification in the z-domain with the same
data. Those who like adventures can also try the standard sdomain to z
domain mapping methods (bilinear transform, impulse invariant transform
etc.
21
), but generally the results will be much worse.
Conversion from the theta-Format to ELiS
In this direction there are fewer difficulties. Let us divide Equation (47) by A(z):
(52)
20. J. Schoukens and R. Pintelon, Identification Why do we need it, how to use it?
Conference Record of the Instrumentation and Measurement Technology Conference
IMTC/93, 93CH3292-0, Irvine, Orange County, CA, May 18-20, 1993. pp. 246-251.
21. see e.g. L. R. Rabiner and B. Gold, Theory and Application of Digital Signal Process-
ing, Englewood Cliffs, NJ, Prentice-Hall, 1975.
var N
red
{ } H z ( )
2
var N
x
{ } var N
y
{ } 2H z ( )cov N
x
N
y
{ , } + =
cov N
x
N
y
{ , } 0.5E N
x
N
y
{ , } =
Y z ( )
B z ( )
A z ( )F z ( )
----------------------z
nk
U z ( )
C z ( )
A z ( )D z ( )
----------------------- E z ( ) + =
2 Tutorial
2-46
Most of the parameters in Equation (46) can be immediately recognized:
(53)
The input noise is equal to zero. The variance vector can be easily generated
from the coefficients of E(z) in Equation (52). If the model in the theta-format
is for multiple inputs, in ELiS an individual model will correspond to each
input.
Conversion of the Covariances
The routine tha2elis will try to pass the covariance matrix of the estimated
parameters to ELiS. However, since the denominator of the transfer function
is a product of A(z) and F(z), if the coefficients of A(z) have been estimated, i. e.,
their variance is not zero, the covariance matrix of the result of this
multiplication will be a nonlinear function of the original covariances.
tha2elis will make an attempt to calculate the new covariance matrix by
linear approximation. If the linear approximation cannot be applied, it sends a
warning message, and passes a covariance matrix full of zeros.
B z ( ) N z ( ) A z ( )F z ( ) D z ( ) nk f
s
, ,
Data Formats and File Handling
2-47
Data Formats and File Handling
For the Frequency Domain System Identification Toolbox, standard data and
file formats are defined. The functions of the toolbox exchange data with each
other via variables in the workspace, and these variables generally have a
straightforward definition. There are some compound variables like the ones
comprising the parameters of an estimated transfer function, the delay, the
sampling frequency and the domain (s or z), to be used by different functions.
Such variables must have a standard form. Moreover, since the purpose of the
toolbox is working on real measurement problems, we attempted to facilitate
data exchange between it and other programs, like the ones for the control of the
measurement setup, data logging and preprocessing, etc., by providing a
standard archive format. The data and file formats cover
Time vectors and files containing measured or simulated time domain data
Fourier vectors and files containing measured, simulated or calculated
frequency domain complex amplitudes or input-output point pairs
Variance vectors and files containing variances of the real parts (that is, also
of the imaginary parts) of the measured frequency amplitudes
Parameter vectors and files containing parameters of the transfer functions
Covariance vectors and files, containing the covariance matrix of the
estimated parameters
There are also so-called report files, containing all relevant information about a
run, in textual format. These files are ASCII files, and they are not
standardized.
Data vectors usually contain different kinds of data, belonging to the same
measurement or estimation procedure, in a long vector form. The data vectors
can be taken apart by using the imptim, impfou, impvar, imppar, and impcov
functions. This allows direct access to the numerator and denominator
coefficients, so that the user does not need to worry about the internal
structure of the vector. The data vector formats are defined in Appendix A1.
Each file type listed above (but not the report file) has an ASCII and a MATLAB
binary version as described below. The ASCII and binary versions can be
generated from MATLAB using the exptim, expfou, expvar, exppar, and
expcov functions, respectively. Both the ASCII files and the binary files can be
read into MATLAB using the imptim, impfou, impvar, imppar, and impcov
functions.
2 Tutorial
2-48
The binary files can also be directly loaded into (or saved from) MATLAB, if the
conventions are followed (see Appendix A). This allows much quicker data
transfer. However, this must be done properly, because the variable names are
fixed in direct load/save, and the checkings otherwise performed by the export/
import routines are bypassed.
The full ASCII files can also be read in MATLAB, using the above import
functions, but this may be a rather slow procedure. The read-in is based on the
function loadasc, which ignores comments (preceded by a % mark in a line) in
an ASCII file, and generates a vector with a user-defined name, containing all
the numbers present in the file.
The different file formats can be converted to each other using the file
conversion function elisfcnv.
The file types and formats are recognized by their extensions. The first
character of the extension refers to the contents of the file (Fourier, variance,
etc.), the second and third to the file format: nt is reserved for ASCII files with
no text (flat ASCII files), and bn is reserved for binary files. The standard
extensions are:
The complete definition of the file formats is given in Appendix A.
An extra facility, incorporated into the toolbox, allows the export of vectors into
flat ASCII files (expvect), for data transfer to graphing programs.
Time files: .tim, .tnt, .tbn
Fourier files: .fou, .fnt, .fbn
Variance files: .var, .vnt, .vbn
Parameter files: .par, .pnt, .pbn
Covariance files: .cov, .cnt, .cbn
A Typical Identification Session
2-49
This section illustrates the frequency domain identification procedure for the
identification of a flexible robot arm. The procedures described below include a
few typical steps, applicable under a variety of circumstances. However, every
identification task has its own flavor. For systems other than this robot arm,
other procedures may be adequate. Therefore, while this section serves as an
example, and suggests a series of good solutions, it should not be followed
blindly for other systems.
The procedures described below are implemented in the M-file rarmdemo. The
code fragments do not necessarily constitute a full program; their purpose is
rather to illustrate a simple way of calculation.
The behavior of a flexible robot arm was measured by applying controlled
torque to the vertical axis at one end of the arm, and measuring the tangential
acceleration of the other end. The excitation signal was a multisine, generated
with frequency components at [1:2:199]*df, with df = 500/4096 0.122 Hz;
that is, the frequency range was about 0.12 Hz 24 Hz. The originally flat
multisine was distorted by the nonlinear behavior of the actuator. The odd
harmonic frequencies provided that components produced by a squaring
nonlinearity would not disturb the identification. The input and output signals
were sampled with sampling frequency fs = 500 Hz. Sampling was
synchronized to the excitation signal so that 4096 samples were taken from
each period. The data records contain 40960 points; that is, 10 periods were
measured. The data are available in the file robotarm.mat
22
. The time series
are scaled to 16-bit integers in order to reduce the file size.
22. This measurement was made at the Department of Mechanical Engineering, Cath-
olic University of Leuven (KUL), in cooperation with Department ELEC, Vrije Univer-
siteit Brussel (VUB), as a part of the Belgian program "Interuniversity Attraction Poles
(IUAP50: Robotics and Industrial Automation)" initiated by the Belgian State, Prime
Minister's Office, Science Policy Programming. These data belong to the public domain
and can be freely used by anyone.
2 Tutorial
2-50
Investigation of the Time Domain Data
First the time domain data and the autocorrelation function can be
investigated. Let us have a look at the time domain data.
load robotarm.mat
%xt = scaled input time record, 40960x1
%yt = scaled output time record, 40960x1
%fs = sampling frequency, 500 Hz
%ascale = scaling factor of the time records
%N = number of points in a period (4096)
%freqind = index numbers of sine waves in DFT of
%a period, [1:2:199]'
xt = xt*ascale; yt = yt*ascale; dt = 1/fs; df = fs/N;
Nl = length(xt); expno = Nl/N;
T = Nl*dt; timevtot = [1:Nl]'*dt; freqindtot = freqind*expno;
clf, hold off
subplot(2,1,1), plot(timevtot,xt,'-w')
title(sprintf(['Input data (torque), number of points:',...
' .0f'],Nl))
xlabel('Time, s')
subplot(2,1,2), plot(timevtot,yt,'-w')
title(sprintf(['Output data (acceleration), ',...
'number of points: .0f'],Nl))
xlabel('Time, s')
2-51
Figure 2-5: Time domain data
Not much can be stated on the basis of the time functions; not even the period
length can be read. The input signal apparently has a smaller crest factor than
the output one, but that's about all we can see. More can be determined from
the autocovariance function. We will evaluate the so-called circular correlation,
which is the inverse Fourier transform of the periodogram, (1/Nl)*abs(X).^2.
We will suppress the dc component to get the autocovariance. In order to have
immediate information about the periodicity, we will connect every 4096th
point by a dotted line.
2 Tutorial
2-52
%Calculations:
X = fft(xt); dcx = X(1); X(1) = 0; Sx = (1/Nl)*abs(X).^2;
Cx = real(ifft(Sx));
%Plotting:
clg
subplot(2,1,1)
axv = [0,max(timevtot),1.1*min(Cx),1.1*max(Cx)];
plot(timevtot,Cx,'-',timevtot([1:4096:Nl,Nl]),...
Cx([1:4096:Nl,Nl]),':')
axis(axv)
title('Circular autocovariance function'), xlabel('time, s')
From the autocovariance function (see below) several conclusions can be
drawn. First of all, there is indeed a periodicity of 4096*dt = 8.192 s. The
autocovariance function corresponds to a bandlimited white spectrum; the
negative peaks verify the use of odd harmonics at [1:2:199]*df. The signal
was oversampled by a factor of 2048/199 = 10, that is, from the sin(x)/x
shaped main lobes of the autocovariance function, about 20 points are sampled.
p1h = axes('Position',[0.1300,0.1100,0.3175,0.3375]);
pv = 1*4096+1+[-15:15];
axv = [min(timevtot(pv))-dt,max(timevtot(pv))+dt,...
1.2*min(Cx(pv)),1.1*max(Cx(pv))];
plot(timevtot(pv),Cx(pv),'o',timevtot(pv),Cx(pv),':')
axis(axv)
mpv = median(pv); hold on
plot(timevtot(mpv)*[1,1],[0,Cx(mpv)],':w',...
axv(1:2),[0,0],':w')
title('Lag No. 1')
hold off
p2h = axes('Position',[0.5825,0.1100,0.3175,0.3375]);
pv = 5*4096+1+[-15:15];
axv = [min(timevtot(pv))-dt,max(timevtot(pv))+dt,...
1.2*min(Cx(pv)),1.1*max(Cx(pv))];
plot(timevtot(pv),Cx(pv),'o',timevtot(pv),Cx(pv),':')
axis(axv)
mpv = median(pv); hold on
plot(timevtot(mpv)*[1,1],[0,Cx(mpv)],':w',...
axv(1:2),[0,0],':w')
title('Lag No. 5'), hold off
2-53
Figure 2-6: The autocovariance function
It is obvious from the enlarged peaks that synchronization is very good
between the excitation signal and the sampling clock. This can also be verified
in the frequency domain. If there was a slip, it could not be more than about
0.003% (dt/2 in a time of 4*4096*dt).
2 Tutorial
2-54
Examination of the Signal-to-Noise Ratios
The autocovariance function can be used for approximate determination of the
signal-to-noise ratio; the power of the periodic components is given by one of
the peaks at nonzero lag, the total power is given by the covariance value at
zero.
PtotxC = Cx(1); PperxC = mean(Cx([1:9]*N+1));
%
fprintf(['Cx(0) = %.3g, Cx(k*Tp) = %.3g, SNRx = %.1f',...
' dB\n'],Cx(1),Cx(N+1),10*log10(PperxC/(PtotxC-PperxC)) )
Cx(0) = 0.0293, Cx(k*Tp) = 0.0288, SNRx = 17.3 dB
However, this is not exactly what we need. The useful signal has power at the
given frequencies only; the rest are spurious components, produced by the
nonlinearities.
%Calculations:
freqvtot = [0:Nl/2-1]'/Nl*fs;
Y = fft(yt); dcy = Y(1); Y(1) = 0;
Sy = (1/Nl)*abs(Y).^2; clear Y
%Plotting:
clg, hold off, subplot(2,1,1)
plot(freqvtot,abs(X(1:Nl/2)) )
title('Input amplitudes'), xlabel('Frequency, Hz')
subplot(2,1,2), plot(freqvtot,abs(Y(1:Nl/2)) )
title('Output amplitudes'), xlabel('Frequency, Hz')
2-55
Figure 2-7: Input and output Fourier amplitudes
The input amplitude spectrum is not really flat. The cause is most probably the
nonflat transfer function of the system composed of the actuator and the device
under test. The two dips at 7.2 Hz and 15.7 Hz in the input spectrum
correspond to resonance points of the system, where the actuator is not capable
of maintaining the signal level. This is not a serious problem since the
frequency range of interest is sufficiently covered by nonzero excitation
amplitudes. The powers of the useful signal, of the harmonic components, and
of the noise, can be calculated for both the input and the output signals.
2 Tutorial
2-56
%Calculations:
Pux = 2*sum(Sx(freqindtot+1))/Nl; Ptotx = sum(Sx)/Nl;
Pperx = real(exp(j*2*pi*N*[0:Nl-1]/Nl)*Sx)/Nl;
Puy = 2*sum(Sy(freqindtot+1))/Nl; Ptoty = sum(Sy)/Nl;
Ppery = real(exp(j*2*pi*N*[0:Nl-1]/Nl)*Sy)/Nl;
%Display results:
fprintf(' Input signal:\n')
fprintf([' Total power: %.3g, useful power: %.3g, ',...
'noise power: %.3g\n'],Ptotx,Pux,Ptotx-Pperx)
fprintf([' Power of spurious periodic components: ',...
'%.3g\n'],Pperx-Pux)
fprintf([' SNR: %.1f dB, for useful components only:',...
' %.1f dB\n'],10*log10(Pperx/(Ptotx-Pperx)),...
10*log10(Pux/(Ptotx-Pux)) )
fprintf(' Output signal:\n')
fprintf([' Total power: %.3g, useful power: %.3g,',...
' noise power: %.3g\n'],Ptoty,Puy,Ptoty-Ppery)
fprintf([' Power of spurious periodic components: ',...
'%.3g\n'],Ppery-Puy)
fprintf([' SNR: %.1f dB, for useful components only: ',...
'%.1f dB\n'],10*log10(Ppery/(Ptoty-Ppery)),...
10*log10(Puy/(Ptoty-Puy)) )
Input signal:
Total power: 0.0293
useful power: 0.0286, noise power: 0.000508
Power of spurious periodic components: 0.000218
SNR: 17.5 dB, for useful components only: 16.0 dB
Output signal:
Total power: 0.0828
useful power: 0.0797, noise power: 0.00139
Power of spurious periodic components: 0.00171
SNR: 17.7 dB, for useful components only: 14.1 dB
The signal-to-noise ratios are quite good, although the nonlinearities produce
significant components. Since the noise is larger than the nonlinearity
products, these will hopefully not deteriorate the identification significantly.
Let us select the excitation lines [1:2:199] only, so the SNR will be improved
by a factor of about 20 (13 dB).
2-57
Conversion to Frequency Domain
Now periodwise conversion to frequency domain follows. Ten periods were
measured. We will treat each period as a separate experiment, thus noise
analysis can be performed. The time vector of a period (one experiment) is
shorter than the measurement data vectors xt and yt. Therefore, the easiest
way for passing the data to tim2fou is to use exptim.
timevect = [1:N]'*dt; freqv = freqind*df;
F = length(freqind);
[x,y] = tim2fou(exptim(timevect,xt,yt),freqv);
Variance Analysis
Frequency domain noise analysis can be performed using varanal. In parallel
with noise analysis, a test of synchronization could be performed, making use
of the post-measurement synchronization possibility. This would take
considerable time. We already know that the measurements were made with
good synchronization, so testing of synchronization will not be done.
%Variance analysis:
[vx,vy,cxy,mx,my,Na,Np,cfl,dv,sd] = ...
varanal(expfou(freqv,x,y));
%Calculation of the SNRs:
PNx = 0; Px = 0; PNy = 0; Py = 0;
Px = 2*sum(abs(mx).^2/N)/N; Py = 2*sum(abs(my).^2/N)/N;
PNx = 2*sum(2*vx/N)/N; PNy = 2*sum(2*vy/N)/N;
%Plotting:
subplot(2,2,1)
plot(freqv,vx,'+'), title('Input variances')
xlabel('Frequency, Hz')
ylabel(sprintf('SNR = %.1f dB',10*log10(Px/PNx)))
subplot(2,2,2)
plot(freqv,vy,'+'), title('Output variances')
ylabel(sprintf('SNR = %.1f dB',10*log10(Py/PNy)))
subplot(2,2,3)
plot(freqv,abs(cxy),'+'), title('I/O covariances'),
subplot(2,2,4)
plot(freqv,abs(cxy)./sqrt(vx.*vy),'+')
title('I/O corr. coefficients'), xlabel('Frequency, Hz')
2 Tutorial
2-58
Figure 2-8: The results of noise analysis
The total SNR is increased by the selection of the points of interest:
fprintf('SNRinp = %.1f dB, SNRoutp = %.1f dB\n',...
10*log10(Px/PNx),10*log10(Py/PNy))
SNRinp = 22.6 dB, SNRoutp = 23.7 dB
The increase of about 5-6 dB is due to the oversampling and selection of points
of interest only. It is less than expected, probably because the noise has less
power at higher frequencies than in the lower frequency band.
2-59
The input-output covariances are quite large so they can not be ignored. This
could mean that a part of the noise goes through the system, that is, the
estimation is corrupted by less noise than calculated above. This can be verified
by plotting cxy./vx. If an important part of the noise goes through the system,
this plot will have a shape similar to the transfer function:
We will make this plot when the approximate shape of the transfer function is
plotted.
A rough guess about the gain in SNR can be obtained by calculating the SNR
of the nonparametric estimate of the transfer function. The SNR of this
nonparametric estimate is smaller than those above, because the division
ym./xm amplifies the noise significantly where xm is small. In elis, this division
is not used, as it can be seen from the cost function.
[tfm,stdAm] = stdtfm([freqv,mx,my],[vx,vy]);
[tfmc,stdAmc] = stdtfm([freqv,mx,my],[vx,vy,cxy]);
SNRtfm = 10*log10(sum(abs(tfm).^2)/sum(2*stdAm.^2));
SNRtfmc = 10*log10(sum(abs(tfmc).^2)/sum(2*stdAmc.^2));
fprintf(['SNRtfm without covariance: %.1f dB, with ',...
'covariance: %.1f dB\n'],SNRtfm,SNRtfmc)
SNRtfm without covariance: 13.6 dB, with covariance: 14.2 dB
The SNR can also be studied at the selected points by calculating it both for the
input and the output. However, it is much quicker to plot the nonparametric
transfer function estimates with uncertainties using ploteltf.
ploteltf('','',[freqv,x(1:F),y(1:F)],'','',[vx,vy,cxy])
cxy 0.5E N
x
N
y
{ } E N
x
N
x
tf { } vx
tf = =
2 Tutorial
2-60
Figure 2-9: Uncertainties of the nonparametric transfer function estimate
The SNR is quite good indeed. Now cxy./vx will be plotted to check that a
significant part of the noise goes through the system.
ploteltf('','',[freqv,ones(100,1),cxy./vx])
2-61
Figure 2-10: The shape of cyx./vx, showing a pattern similar to the transfer
function
cxy./vx has a shape very similar to the transfer function. A large part of the
noise goes through the system.
Identification
We can now proceed with identification. Since the experiments are well
synchronized, the average of the complex amplitudes, mx and my, can be used.
Since these averaged quantities have smaller variances, vx, vy , and cxy have
to be divided by the number of averaged experiments, Na. For the run of elis,
the numerator and denominator orders of the transfer function have to be
given. From the nonparametric plot it is obvious that at least two complex pole
pairs and two complex zero pairs will be necessary. So, let us start with a
system 4/4.
Fdat = [freqv,mx,my]; vdat = [vx,vy,cxy]/Na;
[pv,fit,Cp] = elis(Fdat,vdat,['s',4,4],[],'',10);
2 Tutorial
2-62
Figure 2-11: Result of the 4/4 fit
The fit is quite good, but the cost function is still large, and there is an apparent
mismatch at the higher frequency band. It seems reasonable to increase the
orders. Let us try a 6/6 system.
[pv66,fit66,Cp66] = elis(Fdat,vdat,['s',6,6],[],'',10);
2-63
The fit is much better. The cost function got quite close to the theoretical value,
however, it is still larger than the theoretical value by about a factor of 2.5, so
there are probably still small modeling errors. A model of order 8/8 can still be
tried.
[pv,fit88,Cp] = ...
elis(Fdat,vdat,['s',8,8],[],[NaN,NaN,100],25);
2 Tutorial
2-64
The 8/8 model is not much better than the 6/6 one. The cost function decreased
from 220.5 to 207.6 only and the theoretical values are 91 and 93, respectively.
The large number of necessary iterations is also an indicator of probable
overmodeling. Now other attempts can be made with different numerator and
denominator orders, but none of them is successful finding a better fitting
stable model than the 6/6 one. The modeling error is probably due to
nonlinearities. The order need not be further increased.
However, there is still a chance that a lower-order system can be as good as the
6/6 one. Let us make a pole-zero uncertainty plot of the 6/6 model.
plotelpz(pv66,[],Cp66,2)
2-65
Figure 2-14: Uncertainty ellipses of the 6/6 fit
The confidence ellipses are quite small, so they are of no use in this case. What
can be seen is that the one zero pair and one pole pair have larger variance than
the rest. However, we can speculate that the real zero pair far from the
imaginary axis plays no important role, so it is reasonable to decrease the
numerator order by 2. This will also allow the transfer function to decrease for
higher frequencies, the usual behavior of physical systems. The two poles may
correspond to the resonance around 42 Hz, shown in the complex output
amplitude plot. At this frequency there was no excitation applied, however, the
nonlinearities produced enough overharmonics to show this resonance. For a
proper identification of it, the complex input/output amplitudes around this
frequency should also be used, and a broader excitation signal should have
been applied. We are not going to specifically deal with this resonance, but will
proceed with the above used data. But before making the 4/6 fit, let us have a
closer look at the uncertainties of the important poles and zeros.
2 Tutorial
2-66
plotelpz(pv66,[-0.8,0,-200,200],Cp66)
Figure 2-15: Uncertainties of the important zeros and poles
The dominant error is in the damping of the poles and zeros; their frequencies
are well determined.
2-67
Let us do now a 4/6 fit. For a thorough study of this fit, a report file can be
generated.
[pv46,fit46,Cp46] = elis(Fdat,vdat,['s',4,6],[],'',10,[],...
'robotarm.rep');
2 Tutorial
2-68
Model Validation
type robotarm.rep
FREQUENCY DOMAIN SYSTEM IDENTIFICATION TOOLBOX FOR MATLAB
ELiS RUN PROTOCOL, date and time: 13-Dec-93, 17:54:48
Report file: robotarm.rep
Default run parameters, modified in command line
Fourier data given in command line
Experiment: 1, number of frequencies: 100
Variance data given
Input-output covariances are taken into account
Fit in s-domain, frequencies normalized internally
by omegasc = 76.699 rad*Hz
suggestion: omegasc = 109.48 rad*Hz
Orders: 4/6
No fixed nonzero parameters (norm=1 solution)
Fixed value of the delay: 0 s
Algorithm: Singular value decomposition
Initial value setting: WLS, by singular value decomposition
Allowed maximum number of iterations: 50,
iterations performed: 4
Total run time: 0.43 min, time used for plots: 0.28 min
time used for pole/zero calculations: 0.00 min
Stop if relative change of cost function is smaller than
1.00e-06
last relative variation: -2.03e-11
Stop if maximum relative change of parameters is smaller than
0.00e+00
last max. rel. variation: +1.09e-06
Condition number of the decomposed or inverted matrix:
70.001, reciprocal: 0.014285
Condition number of J: 70.001, reciprocal: 0.014285
Condition number of Q=d^2C/dp^2: 4496.6,
reciprocal: 0.00022239
eps = 2.220e-16
Value of the cost function: 247.49, its double: 494.99
Theoretical value of the cost function: 94.0
Degrees of freedom of the chi-square value (2*cfth): 188
2-69
5%-95% points of the theoretical distribution of the cf:
75.959, 113.93
Approximate mean model error: 0.2232
Mean absolute value of the transfer function: 1.3037
Approximate relative mean model error: 0.17121
Akaike criterion: 518.99
Number of free parameters: 12
Values of the parameters (with standard deviations, calculated
from the approximate covariance matrix):
Numerator:
s^0 2.573756993444797e-02 std: 2.5441e-05 (0.098848%)
s^1 9.551589736680904e-06 std: 4.4310e-07 (4.639%)
s^2 4.165357328394967e-05 std: 1.7078e-08 (0.041%)
s^3 2.261458387371986e-09 std: 4.5396e-11 (2.0074%)
s^4 2.026311165334144e-09 std: 1.3721e-12 (0.067713%)
Denominator:
s^0 -2.311031465095138e-01 std: 2.0162e-04 (0.087241%)
s^1 -1.209269285803730e-04 std: 2.4367e-06 (2.0151%)
s^2 -1.379252065522331e-04 std: 1.5389e-07 (0.11158%)
s^3 -5.572203126063595e-08 std: 1.0717e-09 (1.9233%)
s^4 -1.355803200968540e-08 std: 2.7230e-11 (0.20084%)
s^5 -3.740349830798937e-12 std: 7.8835e-14 (2.1077%)
s^6 -2.006123724224281e-13 std: 1.3813e-15 (0.68853%)
Delay:
0 s fixed
Zeros (rad*Hz):
-4.5424e-01 -1.4113e+02*j
-4.5424e-01 +1.4113e+02*j
-1.0378e-01 -2.5252e+01*j
-1.0378e-01 +2.5252e+01*j
Poles (rad*Hz):
-8.7516e+00 -2.3556e+02*j
-8.7516e+00 +2.3556e+02*j
-4.4728e-01 -9.9527e+01*j
-4.4728e-01 +9.9527e+01*j
-1.2343e-01 -4.5749e+01*j
-1.2343e-01 +4.5749e+01*j
2 Tutorial
2-70
First nonzero numerator coefficient is b(4)=2.0263e-09
First nonzero denominator coefficient is a(6)=-2.0061e-13
b(4)/a(6)=-1.0101e+04
Static gain: -1.1137e-01, -19.1 dB
Parameter file to save data: -
Covariance file to save data: -
%%%%% End of report file robotarm.rep %%%%%
The fit seems to be very reasonable. We have a good identified model.
The 4/6 model can be verified using the standard techniques of the toolbox. We
will not do all the possible tests, but will do some of the typical ones. One of the
most important indicators of the quality of the fit is the value of the cost
function, already discussed above. It is also important to examine visually the
quality of the fit on the plot of elis. In this case the error is too small to be
easily detected on the plots. The phase errors at the zeros are not really
important, since here the phase information of the measurements is small. The
confidence interval plots using ploteltf could also be used, but the confidence
intervals have to be magnified for visual checking.
ploteltf(pv46,[],expfou(freqv,x(1:F),y(1:F)),'','',Cp46,100)
2-71
Figure 2-17: Magnified confidence bounds of the 4/6 fit
2 Tutorial
2-72
Even with a bound of 100*sigma, not much can be seen. It is better to look for
other tests. We think that there are modeling errors, so let us check the
approximate mean model error.
fprintf('Approximate mean model errors:\n')
fprintf('4/6 model 6/6 model 8/8 model\n')
fprintf(' %.2f %.2f %.2f\n',...
fit46(10),fit66(10),fit88(10))
fprintf('Mean absolute value of the transfer function:\n')
fprintf(' %.2f %.2f %.2f\n',...
fit46(11),fit66(11),fit88(11))
Approximate mean model errors:
4/6 model 6/6 model 8/8 model
0.22 0.21 0.20
Mean absolute value of the transfer function:
1.30 1.31 1.31
These values illustrate that the modeling error is not negligible, and is in the
same order of magnitude for all three fits. Because of the modeling error, the
Akaike criterion cannot be used. For closer investigation of the quality of the
fit, the residuals can be calculated.
[rx,ry,ryx,vryx,xe,ye] = ...
rdueelis(pv46,Cp46,expfou(freqv,x,y),[vx,vy,cxy]);
rx, ry and ryx can be studied for normal distribution and whiteness. We will
do a few tests for ryx. First, it has to be standardized, dividing the residuals at
each frequency point by the standard deviation.
il = [1:F]'; il = il(:,ones(1,expno)); il = il(:);
ryxn = ryx./sqrt(vryx(il));
If the fit is good, the standardized residuals have to exhibit circular standard
normal distribution at each point, and they have to be independent. These
properties will be checked by simple tests. First, let us draw the histograms of
the real and imaginary parts. The dotted lines show the standard normal
probability density function, scaled up to the histogram, which is made of 1000
points, with dx = 0.2.
2-73
%Calculations:
dx = 0.2; X = [-3.8:dx:3.8];
Nhr = hist(real(ryxn),X); Nhi = hist(imag(ryxn),X);
fX = 1/sqrt(2*pi)*exp(-X.^2/2);
np = F*expno;
%Plotting:
clg
subplot(121), bar(X,Nhr), hold on, plot(X,fX*np*dx,':g'), hold
off
title('Histogram of real part'), xlabel('real(ryxn)')
subplot(122), bar(X,Nhi), hold on, plot(X,fX*np*dx,':g'), hold
off
title('Histogram of imaginary part'), xlabel('imag(ryxn)')
Figure 2-18: Histograms of the real and imaginary parts of the complex
residual of y./x
2 Tutorial
2-74
The fit is good. The chi-squared value can also be evaluated for both
histograms, simply by using the approximate probabilities.
i = find((X >= -2)&(X <= 2)); NPiv = fX(i)*dx*np;
chir = sum(((Nhr(i)-NPiv).^2)./NPiv);
chii = sum(((Nhi(i)-NPiv).^2)./NPiv);
fprintf(['E{chi^2} = %.0f, chi^2_real: %.1f,',...
' chi^2_imag: %.1f\n'],length(i)-1,chir,chii)
E{chi^2} = 20, chi^2_real: 22.5, chi^2_imag: 10.4
The test shows no significant deviation from the standard normal distribution.
2-75
As a last test, let us plot the autocorrelation function of the frequency domain
residual series.
Cf = real(fft(1/np*abs(ifft(ryxn)).^2));
clg, plot(Cf), title('Frequency domain correlation of ryxn')
xlabel('Indices (through all experiments)')
Figure 2-19: Frequency domain autocorrelation of the residual of the non-
parametric estimate y./x
The autocorrelation function has a dominant peak at zero, an indication of the
approximate uncorrelatedness of the residuals. The repeated smaller peaks are
at lag distances of experiment lengths each, which indicates a small modeling
error again, since it corresponds to a repetitive pattern in the residuals.
2 Tutorial
2-76
A last thing we will try in this session is to make a fit using the input and
output variances, but not the covariances, in order to explore what happens in
this case.
[pv,fit,Cp] = elis(Fdat,[vx,vy]/Na,['s',4,6],[],'',10);
Figure 2-20: A 4/6 fit without taking input-output covariance into account
The fit seems to be as good as with the covariances and the cost function is even
smaller. But this small cost function is wrong, and has to be avoided. This is
one reason why it is advisable to use the covariance values whenever possible;
the cost function will only have a reasonable value by application of a correct
noise model. But the most important reason for using the covariances is to
utilize the available information correctly, with more emphasis on the bands
where the amplitudes are measured with smaller error.
Bibliography
2-77
Bibliography
The basic source of the algorithms and ideas used in the Frequency Domain
System Identification Toolbox is the book of Schoukens and Pintelon. This book
provides a more comprehensive treatment of the methods than this brief
tutorial. On general questions of identification and especially time domain
identification the books of Eykhoff, Godfrey, Goodwin and Payne, Ljung,
Norton, Sderstrm and Stoica are recommended. For the numerical methods,
the book of Press et al. is excellent.
P. Eykhoff, System Identification, London, John Wiley and Sons, 1974.
K. R. Godfrey, ed., Perturbation Signals for System Identification. Englewood
Cliffs, Prentice-Hall, 1993.
G. C. Goodwin and R. L. Payne, Dynamic System Identification: Experiment
Design and Data Analysis, New York, Academic Press, 1977.
L. Ljung, System Identification: Theory for the User, Englewood Cliffs,
J. P. Norton, An Introduction to Identification, London, Academic Press, 1986.
W. H. Press, B. P. Flannery, S. A. Teukolsky and W. T. Vetterling, Numerical
Recipes: The Art of Scientific Computing, Cambridge, Cambridge University
Press, 1986.
J. Schoukens and R. Pintelon, Identification of Linear Systems: A Practical
Guideline for Accurate Modeling, London, Pergamon Press, 1991.
T. Sderstrm and P. Stoica, System Identification, Englewood Cliffs,
2 Tutorial
2-78

3
Reference
3-3 Function Tables
3-3 Excitation Signal Design
3-4 Preprocessing of Data
3-4 Estimation
3-4 Presentation of the Results
3-5 Model Validation
3-5 Model Conversions
3-6 Data Vector and File Read/Write
3-7 Other
3 Reference
3-2
This section contains detailed descriptions of all Frequency Domain System
Identification Toolbox functions. It begins with a list of functions grouped by
subject area and continues with the reference entries in alphabetical order.
Information is also available through the online Help facility.
Function Tables
3-3
Function Tables
Most functions have several default arguments. Under the Syntax heading, the
function is listed first with all necessary input arguments, then with all
possible input arguments. The functions can also be used with fewer input
arguments. Missing trailing arguments are given default values. Default
values are also obtained by entering the arguments as empty arrays or strings.
In MATLAB, all output arguments do not need to be specified. Unspecified
output arguments are not returned.
Excitation Signal Design
Function Purpose
dibs Discrete interval binary sequence design
dibsimpr Discrete interval binary sequence improvement
lin2qlog Quasi-logarithmic frequency set from linear grid
log2qlog Quasi-logarithmic frequency set from log grid
mlbs Maximum length binary sequence (PRBS)
msinclip Crest factor minimization of multisine
msinprep Time domain multisine for download
optexcit Excitation signal with optimum power spectrum
3 Reference
3-4
Estimation
Function Purpose
modifyfv Data prefiltering by inverse of known partial tf
tim2fou Time domain to frequency domain conversion
varanal Variance analysis and averaging of data
Function Purpose
elis Iterative minimization of the LS cost function
elisqa Generate run parameter settings for elis
elrpf2v List run parameter file or convert to run parameter
vectors
elrpv2f List run parameter vectors or convert to run parameter
file
gmean Geometric mean of complex vectors and numbers
Function Purpose
expvect Export vectors to ASCII files for plotting
ploteltf Plot transfer functions and confidence intervals
plotelpz Plot pole/zero patterns with uncertainty ellipses
Function Tables
3-5
Model Validation
Model Conversions
Function Purpose
Function Purpose
fdcovpzp Pole-zero model to transfer function conversion
rdueelis Calculate residuals after identification
simfou Generate simulated frequency domain data
simtime Generate simulated time domain data
stdpz Calculate zero/pole uncertainties
stdtf Calculate transfer function uncertainties
stdtfm Calculate uncertainties of Y
m
/X
m
points
Function Purpose
elis2tha ELiS to theta format conversion
tha2elis Theta format to ELiS conversion
3 Reference
3-6
Data Vector and File Read/Write
Function Purpose
elisfcnv Conversion between different file formats
expcov Write data to covariance vector or file
expfou Write data to Fourier vector or file
exppar Write data to parameter vector or file
exptim Write data to time domain data vector or file
expvar Write data to variance vector or file
expvect Export vectors to ASCII files for plotting etc.
impcov Read data from covariance vector or file
impfou Read data from Fourier vector or file
imppar Read data from parameter vector or file
imptim Read data from time domain data vector or file
impvar Read data from variance vector or file
loadasc Load contents of ASCII file into variable
Function Tables
3-7
Other
Function Purpose
fdiddemo Demonstrations for the toolbox
fnamanal Analysis of filenames
loadvar Load value of single variable from MAT-file
pairs Find closest point pairs in two complex vectors
savevar Save variable into existing MAT-file
yesinput Intelligent input function with default value
ywalk yulewalk without windowing
dibs, dibsimpr
3-8
dibs, dibsimpr
Purpose Discrete interval binary sequence design.
Syntax bitser = dibs(N,dt,freqv,ampv)
[bitser,ampopt,Puf,Ptot] = dibs(N,dt,freqv,ampv,trialno,graphmod)
[bitser,ampoptn] = dibsimpr(bits0,dt,freqv,ampv)
[bitser,ampoptn,Puf,Ptot] = ...
dibsimpr(bits0,dt,freqv,ampv,itno,graphmod)
Description dibs generates a zero-mean discrete interval binary sequence of length N, with
interval size dt, approximating the power spectrum given in ampv for the
frequency points freqv. The algorithm is started trialno times from random
starting values. The iteration can be followed on the screen, unless graphmod is
given with a value 'nograph'.
The bit sequence (values t1) is returned in bitser, and the complex amplitudes
of the generated sequence at the points freqv in ampopt. The complex
amplitudes are scaled in such a way that the total power of the designed signal
equals the total power Ptot, defined by ampv.
In order to have a measure of the quality of the design, a so-called equivalent
crest factor is calculated (it is shown in the plots). The basic idea is as follows.
The crest factor of a zero-mean binary signal is 1. However, in our case the
spectrum is not equal to the desired one, it only approximates it. If the smallest
relative amplitude is increased to 1, by amplification of the binary signal, in
order to assure that the system is excited at each frequency at least at the
desired level, the peak value is multiplied by the reciprocal of the smallest
relative amplitude: eqcr = max(abs(ampv./ampopt)).
Puf gives the useful power (the sum of the power at the desired lines), as a
fraction of the total signal power, that is, the theoretical maximum of Puf is 1.
Ptot is the total signal power, calculated from ampv.
dibsimpr attempts to improve the properties of a discrete interval binary
sequence given in bits0: it maximizes the smallest relative amplitude of the
actual amplitude vector, normalized by ampv. The bit series is searched for
improvements by changing the sign of a pair of bits: this search is started by
itno times.
dibsimpr may take quite a long time. To make it possible to follow how it
proceeds, each already processed bit is marked by a dot on the screen.
dibs, dibsimpr
3-9
A typical example of the plot of dibs is shown in the next figure. trial is the
serial number of the actual trial (or at the end of the iterations the serial
number of the one in which the optimum was found), iter is the number of the
iteration cycles in the given trial, trials is the total number of the trials (given
in trialno), eq. cr is the "equivalent crest factor" (see above), N is the length of
the bit series, and dt is the length of the sampling interval.
In the frequency domain plot the desired amplitudes are given by dotted lines,
the actual ones by solid lines. Puf is the part of the total signal power, which is
concentrated at the given frequencies. The minimum and maximum values of
the relative amplitudes (actual amplitudes vs. the desired ones) are also given
in percents. For these numbers the actual amplitudes are scaled to have the
same total signal power as prescribed by ampv.
The frequency axis is scaled in frequency indices, that is, the unit is df, the
reciprocal of the period length.
The labels of the plots of dibsimpr are similar.
dibs, dibsimpr
3-10
Default
Argument
Values
ampv = ones(size(freqv)), trialno = 25, graphmod = 'graph', itno = 1.
Examples Let us assume that a system is to be excited at uniformly distributed frequency
points between 500 Hz and 2 kHz, with 100 Hz resolution. A good choice for the
sampling frequency is four times the highest harmonic defined in freqv. A
possible program is as follows:
freqv = [500:100:2e3];
fs = 42e3; N = round(fs/100); %N = 80
bits0 = dibs(N,1/fs,freqv,[],10);
bitser = dibsimpr(bits0,1/fs,freqv);
bitser can be directly used for the control of a relay, a thyristor, etc.
Diagnostics The sizes of freqv and ampv must be the same, otherwise an error message is
generated:
freqv and ampv must have the same size
Since the discrete interval binary sequence is periodic, the frequency
components must be at the points k/T, where T is the period length. If N, dt and
freqv are inconsistent, the error message is:
a df value in freqv is smaller than 1/(Ndt)
The condition of the sampling theorem must be fulfilled, further the elements
of freqv must be non-negative, otherwise an error message is sent:
freqv is out of range
dibsimpr iterates until the maximum iteration number is reached, or no
further improvement is found. In this latter case an information message is
sent:
dibsimpr cannot further improve signal
dibs, dibsimpr
3-11
Algorithm dibs is based on [1], with the modification that the returned signal is the one
with the largest minimum relative amplitude. The algorithm generates a
multisine with the amplitudes in ampv, takes the sign of the time function,
combines the obtained phases with the given amplitudes, generates a new
multisine, and so on. The mean value of the binary signal will be kept equal to
zero if N is even, or will be equal to 1/N if N is odd.
dibsimpr changes the sign of two intervals of length dt at a time, observing the
change in the minimum relative amplitude [2].
See Also mlbs
References [1] A. van den Bos and R. G. Krol, Synthesis of Discrete-Interval Binary
Signals with Specified Fourier Amplitude Spectra, International Journal of
Control, 1979, Vol. 30, No. 5, pp. 871-884.
[2] K.-D. Paehlike and H. Rake, Binary Multifrequency Signals Synthesis
and Application, Proc. 5th IFAC Symposium on Identification and System Parameter
Estimation, Darmstadt, FRG, Sept. 24-28, 1979. Vol. 1, pp. 589-596.
[3] K. R. Godfrey, ed.: Perturbation Signals for System Identification. Englewood
Cliffs, Prentice-Hall, 1993.
elis, elisqa, elrpf2v, elrpv2f
3-12
Purpose Basic iteration routine to calculate parametric estimate of linear transfer
functions (elis); treat run parameters (elisqa, elrpf2v, elrpv2f).
Syntax elis(Fdat,vdat,rppar)
[pvect,fit,Cp,CR,cfv] = ...
elis(Fdat,vdat,rppar,fixp,rpalg,rppl,initp,rpfs)
elisqa
rpfout = elisqa(rpf,defaults)
elrpf2v(rpfile)
[rppar,fixp,rpalg,rppl,initp,rpfs] = elrpf2v(rpfile)
elrpv2f(rpfile,rppar)
elrpv2f(rpfile,rppar,fixp,rpalg,rppl,initp,rpfs)
Description elis is the routine which performs the desired nonlinear least squares
iteration to obtain the parameter estimates. elisqa is an optional question/
answer routine. elis is rather complex, and the possibilities may generally
need some explanation; in elisqa such explanations are given, and run
parameters may be set. elrpf2v and elrpv2f perform conversions between run
parameter vectors and a run parameter file.
The run parameters of elis can be set by its input arguments, or they can be
passed through a so-called run parameter file, generated by elisqa. Most run
parameters have default values, thus it is not necessary to give the values of
all parameters for every run.
Fdat contains the input and output Fourier amplitudes: it may be defined as
[freqv,x,y] (N-by-3 array), where freqv is the frequency vector, x is the
complex input vector, and y is the complex output vector; or it may be a
compound Fourier vector, generated by expfou; or it may be the name of a
Fourier file.
vdat sets the variance values. It may be an N-by-2 or N-by-3 array,
[varx,vary] or [varx,vary,covxy]; an 1-by-2 or 1-by-3 row vector, containing
the constant variances; a variance vector (see expvar); or it may be a string
with the name of a variance file.
Instead of either Fdat or vdat, the name of a run parameter file may also be
given: this must be a string which ends by '.ebn'. It may not contain any other
3-13
period, and the given run parameter file must define the Fourier or the
variance data, respectively.
rppar, rpalg, rppl, rpfs are run parameter vectors, each composed of a set of
run parameters. These prescribe the running of the algorithm, by elements
that are either letters or numbers. It does not matter whether the vectors are
given as strings or a numerical vectors, elis will take care of the proper
meanings. The vectors may be shorter than their maximum length, or may
contain NaN elements; in such a case the default values will be assigned to the
corresponding run parameters. Some parameters may be influenced in
different ways (e. g., the setting of starting values by rpalg(2) and by initp);
in such cases the last setting will be valid, e. g., the one defined by initp in the
above case.
rppar is the vector of the most often changed run parameters, associated with
the model structure.
rppar(1) = domain: 's' or 'z'
rppar(2) = numord: order of the numerator
rppar(3) = denomord: order of the denominator
rppar(4) = fs: scaling angular frequency (in s-domain) or sampling frequency
(in z-domain)
rppar(5) = 'a' for allpass design.
In the absence of a run parameter file, at least rppar(1:3) must be given.
fixp defines the fixed parameters. If it is given as an M-by-2 array, the first
column must contain the serial numbers of the fixed parameters in the vector
[num,denom,delay]', and the second column the values. For example, if the
numerator order is 4 and the denominator order is 8, the parameter vector has
(4+1+8+1+1=15) elements. The first denominator coefficient and the delay can
be fixed by fixp = [6,1;15,0]. The delay alone can be fixed by
fixp = [15,0].
If fixp is empty, previously set fixed parameters (by default or by a run
parameter file) will not be changed.
A special form is when fixp is just a one-character string: 'n' for no fixed
parameters at all (the delay is variable, too); 'f' for fixed delay, or 'v' for
3-14
variable delay (for these two values the fixing of numerator and denominator
coefficients remains untouched). Another special form is fixp = '0', for which
the delay is fixed (to zero, or to the value defined by initp) and the zero-order
coefficient of the denominator is set to 1.
rpalg provides control over the iteration possibilities.
rpalg(1) is the type of the iteration algorithm
'g' (Newton-Gauss),
'l' (Levenberg-Marquardt),
'm' (LM with svd),
's' (singular value decomposition),
'r' (Newton-Raphson)
rpalg(2) = initset, way of setting the initial values:
'l' (ordinary least squares),
'w' (weighted least squares),
's' (WLS with svd),
'f' (file or parameter vector, see initp),
'e' (equation error method of the Signal Processing Toolbox)
rpalg(3) = itmax: maximum number of iterations
rpalg(4) = rcostvar: stop if relative variation of cost function is smaller than
this value.
rpalg(5) = rparvar: stop if relative variation of all parameters is smaller than
this value.
Levenberg-Marquardt settings:
rpalg(6) = lambdadecr: number of consecutive decreases of the cost function
before trial with lambda = 0
rpalg(7) = lambda: initial value of lambda
rpalg(8) = lambdalim: minimum value of lambda which allows stopping of
iteration for small variations of the cost function or of the parameters.
rppl controls the plots during the iterations:
3-15
rppl(1) = plotdens: make plots after every plotdens cycle and after the
initial setting and the last cycle if plotdens is finite; if it is inf, no plots will be
made at all; if it is negative, the procedure is the same as for a positive number,
except that the result of the initial setting will not be plotted.
rppl(2:5) = [fmin,fmax,amin,amax]: axis vector for the magnitude plots
rppl(6) = plotmode: type of frequency axis, 'i' for linear, 'o' for logarithmic
rppl(7) = calcrnum: calculate zeros from numerator polynomial, 'c' or 'n'
rppl(8) = calcrdenom: calculate poles from denominator polynomial, 'c' or
'n'
rppl(9:12): axis vector for zero/pole plots. Special meanings (the strings are
always 4-character long, with appropriately set trailing spaces): [n,n,n,n]
(four identical numbers): limit plotted zeros/poles by n2max(freqv) in the
s-domain, or by just n in the z-domain; 'a' to show all zeros and poles; 'p' to
show all zeros and poles, with the same scaling on the two axes.
rppl(13) = pzfollown: follow movement of zeros and poles by plotting several
zero/pole sets, the number given by pzfollown.
initp may have different meanings: it may be the parameter vector of the
initial parameter values; or a parameter filename; if it is a scalar number, it is
the new value of the delay.
rpfs may contain the names of files to be generated, in the form of a string
array, with trailing spaces if the filenames have different lengths. Possibilities:
.rep for elis a report file, .par, .pbn or .pnt for generation of a parameter
file, and .cov, .cbn or .cnt for a covariance file. Other extensions are not
allowed in rpfs.
If no argument is given for elis, elisqa will be started, with the name
elisrpar.ebn.
elisqa offers default answers to the questions, these can be accepted by simply
pressing Enter or Return.
The output arguments of elis are as follows.
pvect is the calculated parameter vector (see exppar).
3-16
fit is a vector containing information relevant to the results of the fit:
fit(1) = cost function
fit(2) = theoretical value of the cost function (c2/2)
fit(3) = 95% point of the theoretical distribution of the cf
fit(4) = number of frequencies
fit(5) = number of free parameters
fit(6) = performed iterations
fit(7) = last change of the cost function
fit(8) = last maximum of the relative changes of parameters
fit(9) = last lambda (for Levenberg-Marquardt, otherwise NaN)
fit(10) = approximate mean model error
fit(11) = mean absolute value of the transfer function
fit(12) = Akaike criterion
fit(13) = condition number of the matrix actually decomposed
or inverted in the last iteration step
fit(14) = condition number of J
fit(15) = condition number of Q=d^2C/dP^2, inverted when
calculating the approximate covariance matrix
(see Eq. (2.23)).
fit(16) = scaling frequency for internal calculations
The meaning and significance of the above condition numbers is explained in
Numerical Stability and Speed of the Procedures on page 2-15.
The approximate mean model error is calculated according to Detection of
Undermodeling and Overmodeling on page 2-33. The value of h
mean
is given,
which will be imaginary if the cost function is too small (e. g., because the
variances were overestimated, or simply because the cost function may be
somewhat smaller than its expected value). This can be best checked from the
information given in the report file. H
ek
and X
k
are usually not even nearly
constant. In these cases the square of the mean model error is averaged over
all the frequencies given.
The mean model error can be compared to the mean absolute value of the
transfer function, computed from the values at the same frequency points as
above.
Cp is the approximate covariance matrix of the parameters (see Covariance of
the Estimate on page 2-8). CR is the approximate Cramr-Rao lower bound for
the covariance matrix of the estimates. Cp and CR are calculated from the
3-17
Jacobian of the last iteration (therefore at least one iteration step is necessary
to calculate them). The covariances are usable only if the algorithm has
converged.
cfv is the vector of the values of the cost function in each cycle (the initial cycle
included). In the case of the Levenberg-Marquardt method cfv has a second
column: the values of lambda are also given.
elis is a rather complex function, and the results often need careful
documentation and studying. Therefore, a so-called report file can be requested
with extensive textual information on the run (see rpfs above).
elisqa can be run separately, in order to set run parameters in a file for elis.
The default run parameters are taken from the file defaults, or if this is not
given, from the file rpf, or if neither of them is given, from an internal table
(see below). If rpf is not given, the name of the file to be generated will be
elisrpar.ebn.
The run parameter files are MATLAB binary files, with the extension '.ebn'.
filenames given without extension will be extended in elisqa by '.ebn'. In
principle, these run parameter files might be modified directly (e. g., by the
routine savevar), but this is not recommended, since it is easy to make a
mistake when doing this. elisqa offers a safe and easy way for such
modifications.
The user of the routines usually need not bother about the internal names and
values of the run parameters. However, the internal run parameters are listed
below with their default values.
Sometimes it may be useful to use the same parameters in vector form, as given
in a run parameter file, or vice versa. elrpf2v and elrpv2f serve this purpose.
The meanings of the input and output arguments are explained above.
When elrpf2v has no output argument, or rpfile is empty when invoking
elrpv2f, the values of the run parameters will be displayed on the screen.
The following paragraphs give a short description of the possibilities and
solutions of elis, providing more detailed information than was possible in the
description of the run parameter vectors.
The measurement data are supposed to be given by Fdat, or maybe in a Fourier
file. For the internal calculations the frequency vector is scaled by the sampling
3-18
frequency in the z-domain, or by a scaling frequency in the s-domain. The
default setting for the scaling frequency is (
min
+
max
)/2.
When defining the model to be fitted, the domain must be specified (s or z), the
orders of the numerator and the denominator must to be given, and the fixed
parameters have to be defined (in the transfer function something must be
fixed, since multiplication of each coefficient by the same constant gives the
same transfer function). If no fixed nonzero parameters are given, elis will set
the norm (the square root of the sum of squares) of the scaled coefficients (but
not the delay) to 1. Another possibility is to set at least one coefficient of either
the numerator or the denominator to a fixed value.
The allpass filter is treated as a special case of parameter fixing (the
parameters of the denominator are the same as those of the numerator, but in
reverse order), however, in this case at least one parameter of the numerator
has to be fixed.
You can also specify whether the roots of the numerator and the denominator
are to be calculated. When dealing with high order systems (>30), the iteration
speed can be increased by sacrificing the pole/zero plot. However, the most
significant acceleration can be achieved by completely suppressing plots
(plotdens set to inf), but this has the risk of missing something that could be
seen from the plot. A possible compromise is to set plotdens to a high value; in
this case the starting values and the result of the last cycle will be plotted
(independently of the actual serial number of the last cycle).
In the transfer function an extra delay term can be present. The value of this
delay can be either fixed or estimated in the iteration procedure. However, a
guess of the delay has to be given even if it is to be estimated, since the initial
value may seriously influence the convergence properties.
The numerical methods solving the nonlinear LS problem (see Basic
Concepts on page 2-2) are standard methods in numerical analysis, including
Newton-Gauss, Newton-Raphson, singular value decomposition (for the
Newton-Gauss formulation), Levenberg-Marquardt and Levenberg-Marquardt
with singular value decomposition. In the Levenberg-Marquardt algorithm an
identity matrix, multiplied by lambda and by the Frobenius norm of J
T
J, is
added to J
T
J before inversion; if a better fit is found, lambda is divided by 2,
otherwise it is multiplied by 10 and the previous parameter values are
restored. After lambdadecr successive divisions, an attempt is made with
lambda = 0. (This corresponds to a Newton-Gauss step).
3-19
The singular value decomposition is always applied to J, even if it is combined
with Levenberg-Marquardt, in order to make full use of the power of the
singular value decomposition algorithm.
There are a few different possibilities to set the initial values for iteration. The
first one is the standard initial setting procedure of ELiS: the linear least
squares fitting. This can be modified to a weighted LS problem (weighting by
the variances of the complex output amplitudes), and can also be solved by
singular value decomposition. There are two more possibilities: the initial
values can be taken from a parameter vector or file (e. g., the result of an earlier
fit or of another method can be used), and also the equation error method of the
Signal Processing Toolbox can be applied for initial value determination for
elis (see invfreqz or invfreqs). In this latter case the input noise is
transformed to the output as if a compound output noise was present.
For the control of the iterations, the maximum number of iteration cycles, the
minimum relative variations of the cost function and of the estimated
parameters can be set. The iteration will be terminated when the maximum
number of iteration cycles is reached, or when any of the absolute values of the
above variations is less than the corresponding minimum value. In the case of
the Levenberg-Marquardt method the value of lambda is also considered: the
iteration is terminated because of small variations only if the value of lambda
is also smaller than lambdalim. The default value (1e-10) corresponds to 30
halvings of the default starting value 0.1.
3-20
A typical plot of the function elis is shown in the figure.
The left-hand side illustrates the fitting of the transfer function. The diagram
consists of two parts. The upper part shows the absolute values of the ratios of
the complex output and input amplitudes (+ marks), along with the magnitude
of the estimated transfer function. The lower part shows the phase errors
between the ratios of the measured complex amplitudes and the estimated
transfer function. The plot is scaled vertically to (180,180).
As a default, the pole/zero plot is also displayed at the right-hand side. The
numbers of poles and zeros, the numbers of non-minimal phase zeros and
unstable poles are also given, along with the number of zeros/poles occasionally
not shown in the plot.
3-21
When elis is terminated, the current axes is the pole/zero plot. This can be
manually rescaled if desired, but the text about unstable poles etc., has to be
deleted first:
h = get(gca,'ch'); delete(h(1))
A somewhat simpler solution is to reapply plotelpz with a given axis vector:
plotelpz(pvect,axv,'','','nomsg')
If desired, even the uncertainty ellipses can be added:
plotelpz(pvect,axv,Cp,2,'nomsg')
The screen contains some further information concerning the run: the value of
the cost function, the last change of the cost function, the theoretical expected
value of the cost function (number of frequencies minus half of the number of
free parameters), cycle number, iteration algorithm and way of setting the
initial conditions, and the value of the delay.
3-22
A report file, that can be requested when running elis, will contain something
like this:
FREQUENCY DOMAIN SYSTEM IDENTIFICATION TOOLBOX FOR MATLAB
ELiS RUN PROTOCOL, date and time: 1-Nov-93, 18:55:1
Report file: elisrpar.rep
Default run parameters, modified in command line
Fourier data given in command line
Experiment: 1, number of frequencies: 30
Input and output variances:
5e-07 1.4754e-07
Input-output covariances are not given
Fit in s-domain, frequencies normalized internally
by omegasc = 2.0452 rad*Hz
suggestion: omegasc = 1.4408 rad*Hz
Orders: 1/3
No fixed nonzero parameters (norm=1 solution)
Fixed value of the delay: 0 s
Algorithm: Singular value decomposition
Initial value setting: WLS, by singular value decomposition
Allowed maximum number of iterations: 50,
iterations performed: 4
Total run time: 0.56 min, time used for plots: 0.51 min
time used for pole/zero calculations: 0.00 min
Stop if relative change of cost function is smaller than
1.00e-06, last relative variation: -1.59e-14
Stop if maximum relative change of parameters is smaller
than 0.00e+00, last max. rel. variation: +2.33e-11
Condition number of the decomposed or inverted matrix:
54.176, reciprocal: 0.018459
Condition number of J: 54.176, reciprocal: 0.018459
Condition number of Q=d^2C/dp^2:
2932.8, reciprocal: 0.00034097
eps = 2.220e-16
Value of the cost function: 30.785, its double: 61.57
Theoretical value of the cost function: 27.0
Degrees of freedom of the chi-square value (2*cfth): 54
5%-95% points of the theoretical distribution of the cf:
3-23
17.789, 38.098
Approximate mean model error: 0.00027292
Mean absolute value of the transfer function: 0.40298
Approximate relative mean model error: 0.00067726
Akaike criterion: 73.57
Number of free parameters: 6
Values of the parameters (with standard deviations,
calculated from the approximate covariance matrix):
Numerator:
s^0 -7.035487766480046e-02 std: 1.7580e-04 (0.24987%)
s^1 -7.038203864830019e-02 std: 5.6617e-05 (0.080443%)
Denominator:
s^0 -2.813969892740070e-01 std: 5.4478e-04 (0.1936%)
s^1 -2.110973116935604e-01 std: 1.6580e-04 (0.07854%)
s^2 -1.407034079272249e-01 std: 1.8950e-04 (0.13468%)
s^3 -7.037131510238287e-02 std: 9.1545e-05 (0.13009%)
Delay:
0 s fixed
Zeros (rad*Hz):
-9.9961e-01
Poles (rad*Hz):
-1.6501e+00
-1.7467e-01 -1.5469e+00*j
-1.7467e-01 +1.5469e+00*j
First nonzero numerator coefficient is b(1)=-7.0382e-02
First nonzero denominator coefficient is a(3)=-7.0371e-02
b(1)/a(3)=1.0002e+00
Static gain: 2.5002e-01, -12 dB
Parameter file to save data: -
Covariance file to save data: -
%%%%% End of report file elisrpar.rep %%%%%
3-24
Default
Argument
Values
Here is the list of the run parameters and their default values (the possible
answers are given between parentheses):
Ffile = ''; %name of Fourier file
vfile = ''; %name of variance file
pfile = ''; %parameter filename (maybe empty)
cfile = ''; %name of the covariance file to be generated
rfile = ''; %name of the report file to be generated
initpfile = ''; %name of param file, with initial values
algtype = 'NG'; %iteration algorithm (NG,LM,LMsvd,NR,svd)
amin = NaN; %axisvect(3) for magnitude plot
amax = NaN; %axisvect(4) for magnitude plot
calcrnum = 'c'; %calculate roots of numerator (c,n)
calcrdenom = 'c'; %calculate roots of denominator (c,n)
covxy = []; %input-output covariance
delay = 0; %(initial) value of the delay
delaytreat = 'f'; %delay fix (f) or variable (v)
denomord = 2; %order of the denominator
denomfix = []; %fixed denominator coefficients
denomfixind = []; %fixed denominator coefficient indices
domain = 's'; %domain of the model (s,z)
expi = 1; %serial number of the experiment to be used
fmin = 0; %lower bound of displayed frequencies
fmax = NaN; %max. displayed frequency in the s-domain
fs = NaN; %sampling (normalizing) frequency
initset = 'l'; %initial value setting (l,w,s,f,e)
itmax = 50; %maximum number of iteration cycles
lambda0 = .1; %starting lambda value for Levenberg-Marquardt
lambdadecr = 10; %after 10 consecutive decreases, 0 is tried
lambdalim = 1e-10; %iteration may stop below this value
numord = 1; %order of the numerator
numfix = []; %fixed numerator coefficients
numfixind = []; %fixed numerator coefficient indices
paramtreat = 'n'; %No fixed params (n), some are fixed:
%(0, d, r) or allpass filter is designed (a)
plotdens = 1; %cycles of iteration to plot (1,2...inf)
pzfollown = 1; %pole/zero sets to be plotted on same plot
pzlimit = 'y'; %limit poles/zeros on plot to
%pzlimitv2fmax (s) or 2 (z)
pzlimitv = 10; %limit poles/zeros on plot (see pzlimit)
3-25
rcostvar = 1e-6; %minimum relative variation of the cf
rparvar = 0; %minimum relative variation of parameters
vardef = 'u'; %variances: numbers or take from file (u,f)
varx = 1; %uniform variance value of input coefficients
vary = 1; %uniform variance value of output coefficients
Default values of the run parameter vectors:
rppar = [NaN,NaN,NaN,NaN];
rpalg = ['g','w',50,1e-6,0,10,0.1,1e-10];
rppl = [1,0,NaN,NaN,NaN,'i','c','c','a ',1];
rpfs = '';
The NaN values mean that the actual values will be controlled by the Fourier
data.
Examples [freqv,x,y] = impfou('bandpass.fbn',1);
elis([freqv,x,y],3.4e-4[1,1],['s',4,6]);
elis('inpchans.ebn');
elis('inpchanz.ebn',[],['z',16,16],[35,0]); %fixing the delay
elisqa('elisqtst.ebn','inpchans.ebn');
[rppar,fixp,rpalg,rppl,initp,prfs] = elrpf2v('inpchans.ebn');
elrpf2v('inpchans.ebn')
elrpv2f('newfile.ebn',['z',12,12],[0],'r');
elrpv2f('',['z',12,12],'0','r');
Diagnostics The error and warning messages are self-explanatory. For the messages about
condition numbers, see Numerical Stability and Speed of the Procedures on
page 2-15.
Algorithm The algorithm and the main expressions are briefly described in Chapter 2, or
in detail in [1].
References [1] J. Schoukens and R. Pintelon, Identification of Linear Systems: a Practical
elisfcnv
3-26
elisfcnv
Purpose Convert elis data files between different formats (binary, ASCII, flat ASCII).
Syntax elisfcnv
elisfcnv(fromfile,tofile)
Description Using the data file read/write routines (impfou etc.), elisfcnv reads the
contents of the file fromfile, and writes it to the file tofile.
Default
Argument
Values
There is no predefined default value; the routine prompts for the filenames if
they are not given.
Examples elisfcnv('inpchans.pbn','inpchans.par')
See Also expcov, expfou, exppar, expvar, exptim, impcov, impfou, imppar, impvar,
imptim.
elis2tha, tha2elis
3-27
elis2tha, tha2elis
Purpose Model conversion from/to the theta-format of the System Identification Toolbox
Syntax theta = elis2tha(pdat,cdat,varet)
theta = elis2tha(pdat,cdat,varet,numn,denomn)
[num,denom,delay,fs] = tha2elis(theta)
[num,denom,delay,fs,vary,ccovar] = tha2elis(theta,N,freqv)
Description These are the two routines which convert models of the Frequency Domain
System Identification Toolbox from/to the System Identification Toolbox. pdat
is the parameter vector in ELiS (see exppar), or the name of a parameter file.
cdat is the covariance matrix of the parameters if it is an array (the rows and
columns of the fixed parameters should contain zeros), or the covariance vector
(see expcov), or the name of the covariance file. cdat may be empty, if the
covariances are not available.
varet is the variance of the time domain noise, reduced to the output of the
system, without noise shaping.
If the input noise is zero, and the variances of the real and of the imaginary
parts of the complex amplitudes are uniformly equal to vary in the N-point
spectrum, varet should be calculated as: varet = 2/Nvary, and numn and
denomn need not be given.
numn and denomn are the numerator and the denominator of the observation
noise shaping filter (see Model Conversions from/to the System Identification
Toolbox on page 2-41). If the frequency domain noise shape is fitted by the
noise shaping filter:
and C corresponds to numn, D to denomn, the value of varet should be 2/N. The
sampling frequency for the noise shaping filters is the same as that of the
parameter vector.
When transforming models from the theta-format of the System Identification
Toolbox, theta is the array of the theta-format to be converted, and N is the
number of FFT points (for the calculation of vary only). vary is a scalar if
C(z)D(z)1, and a vector if the noise is not white. freqv is the vector of
C z
k
( )
C z
k
( )
-------------
2
vary f
k
( ) z
k
, e
j2f
k
f
s
=
elis2tha, tha2elis
3-28
frequencies where the variances are to be calculated. If vary is not required, N
and freqv are not necessary.
The output arguments of tha2elis are as follows: num is the numerator, denom
is the denominator of the transfer function, delay is the additional delay in the
model, fs is the sampling frequency, vary is the vector of frequency domain
variances of the output complex amplitudes at the frequencies in freqv (or just
the frequency domain variance if this is constant), and ccovar is the covariance
matrix of the vector [num,denom,delay].
Default
Argument
Values
numn = 1, denomn = 1
Examples theta = elis2tha('inpchanz.pbn','inpchanz.cbn',2/2561e-9);
[num,denom,delay,fs] = tha2elis(theta);
Diagnostics As it is discussed in Model Conversions from/to the System Identification
Toolbox on page 2-41, the covariances can usually be converted only by using
linear approximation of a ratio. This is done automatically by the conversion
routines, but if approximation was applied, a warning message is sent:
WARNING: variance of denom(1) is not zero in elis2tha
or
WARNING: ccovar will be approximated in tha2elis
If in elis2tha the variance of denom(1) is too large, that is, larger than
0.2denom(1)^2, the calculated variances are useless. In this case an error
message is sent:
var(denom(1)) is too large
It is checked in tha2elis whether the approximated covariances are plausible.
If not, the warning message is:
WARNING: In tha2elis the approximated covariances are too large
elis2tha calls poly2th or mktheta, and tha2EliS calls th2poly or polyform,
thus the System Identification Toolbox must be installed.
See Also System Identification Toolbox
expcov, impcov
3-29
expcov, impcov
Purpose Read/write ELiS covariance vectors and/or ELiS covariance files; furthermore,
extend covariance matrix of free parameters by zero rows/columns of fixed
parameters, or delete zero rows and columns, belonging to fixed parameters.
Syntax cvect = expcov(coeffcovar,fixpind)
[cvect,Cp] = expcov(coeffcovar,fixpind,... filename,comments,fdate)
coeffcovar = impcov(cdat)
[coeffcovar,fixpind,comments,fdate] = impcov(cdat,nofixp)
Description expcov and impcov perform conversions of different representations of the
covariance matrix of parameters. The structure of the ELiS covariance vector
and the file format are described in Appendix A1.
cvect is the vector of covariances, Cp is the covariance matrix of all the
parameters (both estimated and fixed ones).
coeffcovar is an n-by-n array, the covariance matrix of the coefficients. If
fixpind is not given, n is equal to the number of the numerator coefficients plus
the number of the denominator coefficients plus one (for the delay). If fixpind
is given, coeffcovar should either not contain the rows and columns belonging
to fixed coefficients, or these rows and columns should consist of zeros. fixpind
contains the indices of fixed parameters in the total parameter vector, defined
as [num,denom,delay]', where the coefficients are in descending order of
powers of s in the s-domain, and in ascending order of the powers of z
-1
in the
z-domain.
If fixpind is given, and in coeffcovar there are zero rows and columns, the
two notations must correspond to each other.
filename is the name of the covariance file. If filename is missing or empty,
no file will be generated.
If the generation of a file is requested, the file will be created in the active
subdirectory or folder. If the name has no extension, expcov extends it by
'.cbn'. If the extension is .cbn, the result will be a binary file, otherwise an
ASCII file. If the extension is '.cnt', no text is sent to the ASCII file, only data.
If a file is written or read, the most important values will be displayed on the
screen, unless a global variable expimpmessages with value 'no' is defined.
expcov, impcov
3-30
Covariance values will be written to ASCII files by expcov in floating-point
form, using 8 digits in the mantissas.
comments is a string with comments (optional), and fdate is the date (and time)
string (also optional). If fdate is missing or empty, a date string will be
generated.
impcov is the inverse of expcov. The data vector or the name of the file is cdat;
the data vector contains the data in the same order as a .cnt file. If nofixp is
given with the value 'nofixp', the zero rows and columns in coeffcovar will
be deleted. The default extension is .cbn.
Default
Argument
Values
nofixp = ''
Examples coeffcovar = eye(5); expcov(coeffcovar,[],'data.cbn');
[coeffcov,fixpind] = impcov('inpchans.cbn','nofixp');
Diagnostics expcov checks whether coeffcovar is quadratic, real, finite and symmetric.
Also the validity of fixpind is checked.
If a file already exists with the same name, expcov tries to delete it. This will
be unsuccessful if the file is not in the active subdirectory/folder. In this case
the error message is:
Cannot delete existing file ...
impcov checks the length of the covariance vector form: if it is not valid, the
error message is sent:
Number of data is not n(n+1)/2
See Also Description of the Data Vector and File Formats on page A-2
expfou, impfou
3-31
expfou, impfou
Purpose Read/write ELiS Fourier vectors and/or ELiS Fourier files.
Syntax Fvect = expfou(freqvect,x,y)
Fvect = expfou(freqvect,x,y,i,expno,...
filename,comments,fdate,digitnum)
[freqvect,x,y] = impfou(Fdat)
[freqvect,x,y,expno,comments,fdate] = impfou(Fdat,expi)
Description expfou exports data of one experiment (or several experiments) to a vector in
the workspace, and/or to a (perhaps already existing) Fourier file.
The output argument Fvect is the vector containing the same data as would be
sent to a .fnt file.
freqvect is the vector of frequency points, x is the complex input amplitude
vector (or array for multiple inputs), and y is the complex output amplitude
vector (or array for multiple outputs). Each amplitude vector must be a column
vector. The amplitudes belonging to different experiments have to be put under
each other. x or y may be empty, but both of them have to be given.
expno is the total number of experiments. i contains the number(s) of the
actual experiment(s). If several experiments are given, they must be denoted
by successive numbers. If i is empty, the default is i = [1:length(x(:,1))/
length(freqvect)].
The string filename contains the name of the output file. If this name has no
extension, expfou extends it by '.fbn'. If the extension is .fbn, the result will
be a binary file, else an ASCII file. If the extension is .fnt, no text is sent to the
ASCII file, only data.
The file will be created in the active subdirectory or folder. If filename is
empty, no file will be generated.
If a file is read or written, the most important values will be displayed on the
When exporting the data of the first experiment, any file with the same name
will be deleted.
comments is a string with comments (optional), and fdate is the date (and time)
string (also optional). If fdate is missing or empty, a date string will be
generated.
expfou, impfou
3-32
Amplitude values will be written to ASCII files by expfou in floating-point
form. digitnum is the number of digits of mantissas, sent to ASCII files, with
the limits 1 digitnum 16, and default value 7.
impfou reads complex amplitudes from Fourier vectors or files. The file may be
an ASCII file with comments (usual extension: .fou), a so-called flat ASCII file
without comments (.fnt), or a binary file (.fbn). The file has to be somewhere
within the path of MATLAB, or the path is to be explicitly given. The default
extension is .fbn.
The input argument Fdat of impfou is the data vector, or the name of a file, or
may be an array of size F-by-3 in order to be able to process directly any output
format of other M-files. The data vector contains the data in the same order as
a .fnt file, and the array is [freqv,x,y].
expi contains the number(s) of the experiment(s) to be read (integer vector).
expi is optional; if it is omitted or empty, all the experiments will be read.
Default
Argument
Values
digitnum = 7, i = [1:length(x(:,1))/length(freqvect)],
expno = max(i)
Examples expfou([1:20],ones(100,1),0.1ones(100,1),...
[1:5],5,'data.fbn');
[freqvect,x,y,expno] = impfou('data.fbn');
Send results of experiments in two steps:
Fvect = expfou([1:20],ones(60,1),0.1ones(60,1),[1:3],5);
Fvect = [Fvect;expfou([1:20],...
ones(40,1),0.1ones(40,1),[4,5],5)];
[freqvect,x,y,expno] = impfou(Fvect);
Diagnostics If not experiment No. 1 is being exported to a file, expfou looks for an already
existing file. If this is not found, the error message is
File has to exist already when exporting ... experiment
expfou, impfou
3-33
If a file already exists with the same name, and the first experiment is being
exported, expfou tries to delete it. This will be unsuccessful if the file is not in
the active subdirectory/folder. In this case the error message is:
impfou checks the validity of the contents of the vector (file). If it is not
consistent, an error message is sent:
Not enough data in file ...
or if the vector is too long, a warning message is displayed:
Number of data is incorrect in file ...
exppar, imppar
3-34
exppar, imppar
Purpose Read/write ELiS parameter vectors and/or ELiS parameter files.
Syntax pvect = exppar(domain,num,denom)
pvect = exppar(domain,num,denom,delay,fs,...
filename,comments,fdate)
[domain,num,denom,delay,fs] = imppar(pdat)
[domain,num,denom,delay,fs,comments,fdate] = imppar(pdat,fsc)
Description exppar writes parameters of the transfer function to a vector in the workspace,
and/or to parameter files (used by elis).
pvect is the vector containing the same data as the .pnt file.
domain may be 's' or 'z', depending on the domain; num is the numerator
vector, z-domain coefficients in ascending order of powers of z
-1
, or s-domain
coefficients in descending order of powers of s. denom is the denominator vector,
similarly to num. delay is the additional delay, fs is the sampling frequency in
the z-domain, or the scaling frequency between the internal representation and
the s-domain parameter vector or file, which is written in standard SI units.
The string variable filename is the name of the file. If the name has no
extension, this function extends it by '.pnt'. If the extension is .pbn, the result
will be a binary file, else an ASCII text file. If the extension is .pnt, no
comment is sent to the ASCII file, only data.
The file will be created in the active subdirectory or folder. If a file is written or
read, the most important values will be displayed on the screen, unless a global
variable expimpmessages with value 'no' is defined. If filename is empty, no
file will be written.
The parameters will be written to ASCII files in floating-point form by exppar,
with 16-digit accuracy of the mantissas.
comments is a string with eventual comments (optional), and fdate is a date
(and time) string (also optional). If fdate is missing or empty, an actual date
string will be generated.
In an s-domain parameter vector or file, the saved fs value is the so-called
suggested scaling frequency, usable in a later import. The algorithm uses a
exppar, imppar
3-35
simple ad hoc formula to find a scaling frequency, which brings the roots of the
numerator and denominator polynomials possibly close to 1:
When the numerator or the denominator is degenerate, the corresponding term
is set to 1; if the formula would still give an unusable result, like inf or NaN, fs
is set to 1.
imppar reads parameters from parameter vectors or files (used by elis). The
file may be an ASCII file with comments (usual extension: .par), a flat ASCII
file without comments (.pnt) or a binary file (.pbn). The file has to be
somewhere within the path of MATLAB, or the path is to be explicitly given. The
default extension is .pnt.
As an output argument of imppar, fs is the sampling frequency in the
z-domain, while in the s-domain it is the scaling frequency between the internal
representation and the parameter vector or file which is written in standard SI
units. pdat is the data vector or the name of the file; the data vector contains
the data in the same order as a .pnt file.
The scaling frequency between the s-domain internal representation and the
parameter file can be set by fsc. If fsc is not given, no scaling will be performed
(fsc = 1), if it is empty (fsc = []), the previously saved value in the parameter
vector or file is used. In the case of z-domain files, fsc is ignored.
Default
Argument
Values
delay = 0, fs = 1, fsc = 1
Examples pvect = exppar('s',[1,1],[1,2,3,4]);
exppar('z',[1,1],[4,3,2,1],0,1,'filter.pnt',...
'First trial',date);
[domain,num,denom,delay,fs] = imppar('filter.pnt');
[domain,num,denom,delay,fsc] = imppar('inpchans.pbn',[]);
fs
2
3
---
1
nn nd +
-------------------
b
i
b
nn
--------
1
nn i
--------------
i 0 =
nn 1
a
i
a
nd
--------
1
nd i
--------------
i 0 =
nd 1
+
,

_
=
exppar, imppar
3-36
Diagnostics If a file already exists with the same name, exppar tries to delete it. This will
imppar checks the contents of the vector (file); if this is inconsistent, an error
message is sent:
exptim, imptim
3-37
exptim, imptim
Purpose Read/write ELiS time domain data vectors and/or ELiS time domain files.
Syntax tvect = exptim(timevect,xt,yt)
tvect exptim(timevect,xt,yt,i,expno,...
filename,comments,fdate,digitnum)
[timevect,xt,yt] = imptim(tdat)
[timevect,xt,yt,expno,comments,fdate] = imptim(tdat,expi)
Description exptim exports data of one experiment (or several experiments) to a vector in
the workspace, and/or to a (perhaps already existing) time domain data file.
The output argument tvect is the vector containing the same data as would be
sent to a .tnt file.
timevect is the vector of sampling time instants, xt is the input vector (or array
for multiple inputs), and yt is the output vector (or array for multiple outputs).
Each vector must be a column vector. The data belonging to different
experiments have to be put under each other. xt or yt may be empty, but both
of them have to be given.
expno is the total number of experiments, i contains the number(s) of the
actual experiment(s). If several experiments are given, they must be denoted
by successive numbers. If i is empty, the default is i = [1:length(xt(:,1))/
length(timevect)].
The string filename contains the name of the output file. If this name has no
extension, exptim extends it by '.tbn'. If the extension is .tbn, the result will
be a binary file, else an ASCII file. If the extension is '.tnt', no text is sent to
the ASCII file, only data.
The file will be created in the active subdirectory or folder. If filename is
empty, no file will be written.
If a file is read or written, the most important values will be displayed on the
When exporting the data of experiment No. 1, any file with the same name will
be deleted.
comments is a string with eventual comments (optional), and fdate is the date
(and time) string (also optional). If fdate is missing or empty, an actual date
string will be generated.
exptim, imptim
3-38
Amplitude values will be written to ASCII files by exptim in floating-point
form. digitnum is the number of digits of the mantissas, sent to ASCII files,
with the limits 1 digitnum 16, and default value 7.
imptim reads time records from time domain data vectors or files (used by
elis). The file may be an ASCII file with comments (usual extension: .tim), a
so-called flat ASCII file without comments (.tnt), or a binary file (.tbn). The
file has to be somewhere within the path of MATLAB, or the path is to be
explicitly given. The default extension is .tbn.
The input argument tdat of imptim is the data vector or the name of the file,
or maybe an array of size tl3, [timevl,xt,yt]; the data vector contains the
data in the same order as a .tnt file.
expi contains the number(s) of the experiment(s) to be read (integer vector).
expi is optional; if it is omitted or empty, all the experiments will be read.
Default
Argument
Values
digitnum = 6, i = [1:length(xt(:,1))/length(timevect)],
expno = max(i)
Examples tvect = exptim([1:10],ones(50,1),0.1ones(50,1),...
[1:5],5,'data.tbn');
[timevect,xt,yt,expno] = imptim('data.tbn');
Diagnostics If not experiment No. 1 is being exported to a file, exptim looks for an already
existing file. If this is not found, the error message is
File has to exist already when exporting ... . experiment
If a file already exists with the same name, and the first experiment is being
exported, exptim tries to delete it. This will be unsuccessful if the file is not in
the active subdirectory/folder. In this case the error message is:
imptim checks the validity of the contents of the vector (file). If it is not
consistent, an error message is sent:
exptim, imptim
3-39
or if the read vector is too long, a warning message is displayed:
Number of data is incorrect in file ...
expvar, impvar
3-40
expvar, impvar
Purpose Read/write ELiS variance vectors and/or ELiS variance files.
Syntax vvect = expvar(varx,vary)
vvect = expvar(varx,vary,covxy,filename,comments,fdate,Ffile)
[varx,vary] = impvar(vdat)
[varx,vary,covxy,comments,fdate] = impvar(vdat)
Description expvar writes variance data to variance vectors or files.
The output argument vvect is the vector containing the same data as the .pnt
file.
varx is a column vector containing the variances of the real part (that is, also
of the imaginary part) of the input Fourier coefficients; varx is an array for
multiple inputs. vary is the column vector containing the variances of the real
part (that is, also of the imaginary part) of the output Fourier coefficients; vary
is an array for multiple outputs.
covxy is the column vector of input-output covariances, covxy =
E{conj(Nx)Ny}. For multiple inputs or outputs, covxy contains just the
covariances between input(1) and output(1), or all the covariances beside each
other, as [ci1o1,ci1o2,...ci2o1...]. covxy may be empty.
filename is the name of the output file. If the name has no extension, expvar
extends it by '.vbn'. If the extension is '.vbn', the result will be a binary file,
else an ASCII file. If the extension is '.vnt', no text is sent to the ASCII file,
but data.
The file will be created in the active subdirectory or folder. If a file is written or
read, the most important values will be displayed on the screen, unless a global
variable expimpmessages with value 'no' is defined. If filename is empty, no
file will be written.
Variance values will be written to ASCII files by expvar in floating-point form,
using 4 digits in the mantissas.
comments is a string with eventual comments (optional), fdate is the date (and
time) string (also optional). If fdate is missing or empty, an actual date string
will be generated.
Ffile is the associated Fourier vector or file (optional, for cross-checking).
expvar, impvar
3-41
impvar reads variances from variance vectors or files.
vdat is a vector, or the name of the file, or an array of size F-by-2 or F-by-3,
[varx,vary] or [varx,vary,covxy]; the vector contains the data in the same
order as a .vnt file.
The file may be an ASCII file with comments (usual extension: .var), an ASCII
file without comments (.vnt), or a binary file (.vbn). The file has to be
somewhere within the path of MATLAB, or the path is to be explicitly given. The
default extension is .vbn.
Examples expvar(ones(10,1),0.1ones(10,1),zeros(10,1),'data.vbn');
[varx,vary] = impvar('data.vbn');
Diagnostics If a file already exists with the same name, expvar tries to delete it. This will
If Ffile is given, impvar compares the lengths of the frequency vector and the
input and output variance vectors. If they are inconsistent, the error message
is:
Data incompatible with Fourier file
expvar and impvar also check the variance vectors for negative, complex,
infinite elements.
expvect
3-42
expvect
Purpose Export MATLAB vectors to ASCII files for plotting.
Syntax expvect(file,v1)
expvect(file,v1,digitnum)
expvect(file,v1,v2,v3,v4,v5,v6,v7,v8,v9,v10,v11,v12,digitnum)
Description The contents of the real vectors v1, v2, ... will be exported into a flat ASCII file
(file), for further processing by plotting/graphing programs, or for transfer to
other computers, or for sending by electronic mail, etc. The vectors will be
written into an array, where the first line will contain the first elements of all
the vectors and so on. The delimiter between the numbers in a row is tab. The
last character in each line is a CR/LF or a CR.
digitnum defines the number of digits of the mantissa.
Default
Argument
Values
digitnum = 7
Examples t = [1:100]; x = 100ones(100,2)+rand(100,2)+jrand(100,2);
expvect('data.txt',t,real(x(:,2)),imag(x(:,2)),8)
Diagnostics expvect checks the vectors for complex, infinite or NaN elements, and sends an
error message if any element of this kind is found.
The length of the vectors must be the same, otherwise an error message is
generated.
If a file with the same name exists, the routine attempts to delete it. This can
only be done if this file is in the active subdirectory/folder. If this is not the case,
an error message is sent:
See Also save
fdcovpzp
3-43
fdcovpzp
Purpose Calculate transfer function model and its covariance matrix from pole-zero
model, including the standard deviations and covariances of poles and zeros
and gain values.
Syntax [pvect,Cp] = fdcovpzp(zv,stdz,pv,stdp)
[pvect,Cp] = fdcovpzp(zv,stdz,pv,stdp,rzp,g,stdg)
[pvect,Cp] = fdcovpzp(zv,stdz,pv,stdp,rzp,g,stdg,...
domain,fs,da,dzp)
Description It may be desirable to simulate or examine systems of which a pole-zero model
is available. fdcovpzp converts such a model into the transfer function model
required by this toolbox. It is the inverse of stdpz.
The output argument pvect is the parameter vector of the system (see exppar)
and Cp is the corresponding covariance matrix.
zv and pv are the column vectors of the zeros and poles, respectively.
stdz and stdp contain the corresponding uncertainties. In each row there are
three elements: the standard deviation of the real part, the standard deviation
of the imaginary part, and the correlation coefficient between the real part and
the imaginary part of the corresponding zero or pole.
For a full description of the interrelations of different poles and zeros, a full
correlation coefficient matrix of the real and imaginary parts of all zeros and
poles and two gain factors can be given in rzp. The order of the random
variables in the corresponding random vector is: real part of the first zero,
imaginary part of the first zero, real part of the second zero, etc., then the real
part of the first pole, etc. The last two variables are the leading coefficients of
the numerator and the denominator, respectively.
g contains two gain factors: the leading coefficients of the numerator and the
denominator, respectively. If g is given as a scalar, g(2) is set to 1. The
standard deviations of the gain values and their correlation coefficient are to be
given in stdg.
By default, an s-domain model is assumed. For z-domain models, the domain
(domain = 'z') and the sampling frequency (fs) can be given.
fdcovpzp allows the use of analytical or numerical differentiation in the
calculation of the sensitivity matrix. By numerical differentiation it can be
fdcovpzp
3-44
checked whether variations of the poles and zeros, made in the order of
magnitude of the standard deviations (in the directions of the eigenvalues of
the appropriate covariance matrix), cause the same changes in the transfer
function parameters, as calculated from the analytical sensitivity calculations.
Numerical differentiation can be requested by setting da to the value 'num'.
The amount of perturbations can be influenced by the variable dzp. By default,
its value is 1, which means perturbations equal the eigenvalues of the
covariance matrix of the zeros or poles, in the directions of the eigenvectors.
These step lengths are multiplied by the value of dzp.
Default
Argument
Values
g = [1;1]; stdg = [0;0;0]; domain = 's'; fs = 1; da = 'num';
dzp = 1.
Examples [zv,stdz,pv,stdp,rzp,g,stdg] = ...
stdpz('inpchans.pbn','inpchans.cbn');
[pvectn,Cpn] = fdcovpzp(zv,stdz,pv,stdp,rzp,g,stdg);
subplot(121)
plotelpz('inpchans.pbn',[-6,2,-4,4]*1e5,'inpchans.cbn',...
10,'nomsg')
subplot(122)
plotelpz(pvectn,[-6,2,-4,4]*1e5,Cpn,10,'nomsg')
Diagnostics The validity of the given correlation coefficients are checked, and an error
message is sent if they are not consistent.
Algorithm In analytical calculations first the sensitivity matrix of zeros and poles on the
coefficients of the corresponding polynomials is calculated using
where r(i) is the ith root of the polynomial f(r(i)). The sensitivity of zeros and
poles is then calculated as the pseudo inverse of this matrix. The covariance
matrix of the parameters is calculated multiplying the covariance matrix of
the real and imaginary parts of the poles/zeros and the gains by the sensitivity
matrix from both sides.
S i j ( , )
r i ( )
j
df r i ( ) ( )
dr i ( )
-------------------
-------------------, =
fdcovpzp
3-45
For numerical derivation, perturbations of the zeros and poles are introduced
in the direction of each eigenvector of the corresponding covariance matrix,
with step sizes equal to the corresponding eigenvalue, multiplied by dzp. The
so determined sensitivities are then used in the covariance calculations.
See Also stdpz
References [1] P. Guillaume, J. Schoukens and R. Pintelon, Sensitivity of roots to errors
in the coefficients of polynomials obtained by frequency-domain estimation
methods, IEEE Trans. on Instrumentation and Measurement, Vol. 38, No. 6,
pp. 1050-1056, Dec. 1989.
fdiddemo
3-46
fdiddemo
Purpose Demonstrate the Frequency Domain System Identification Toolbox.
Syntax fdiddemo
Description Invoking fdiddemo starts a question and answer procedure, offering the choice
among different demonstrations.
Many of the demonstrations work on measured data. These data are mostly
results of measurements done by Department ELEC, Vrije Universiteit
Brussel, Belgium (see Appendix A2). The data are public domain; they may be
freely used by anybody.
Diagnostics Some of the demonstrations work on the data files used in the book of
Schoukens and Pintelon. For these demonstrations the appropriate data files
(aluplate.fbn, bandpass.fbn, etc.), must be available in the search path of
MATLAB. (See Description of the Available Measurement Data Files on page
A-15 for more information.)
Many of the demonstrations, using measured data, need more memory than
available on small machines. These demonstrations will run on larger
computers (with a minimum memory of 4 MBytes) only.
References [1] J. Schoukens and R. Pintelon, Identification of Linear Systems: A Practical
fnamanal
3-47
fnamanal
Purpose Syntactic analysis of filenames, splitting filenames to names and extensions,
assigning default extension to filenames. This routine is used internally by
other functions of the toolbox.
Syntax nfilename = fnamanal(filename)
[nfilename,fnamsh,ext] = fnamanal(filename,stext)
Description fnamanal analyzes the filename given in filename, puts the filename without
extension into fnamsh, and the extension to ext. If stext is given (standard
extension), and the filename has no extension, the standard one will be
appended (preceded by a period) to the filename. The complete filename is
given in nfilename.
Examples fnam = 'inpchan'; [fnam,fnshort,ext] = fnamanal(fnam,'fbn')
fnam =
inpchan.fbn
fnshort =
inpchan
ext =
fbn
gmean
3-48
gmean
Purpose Geometric mean of complex numbers.
Syntax gm = gmean(X)
Description gmean makes an attempt to eliminate the effect of phase wrapping when
calculating the geometric mean of complex numbers. For matrices, gmean(X) is
a row vector containing the complex geometric mean value of each column.
The complex geometric mean has important applications in the averaging of
nonparametric estimates of complex transfer function values (see [1], [2], and
Solutions for Some Special Cases on page 2-14).
Example x = -ones(100,2) + 0.1randn(100,2) + j0.01randn(100,2);
mfalse1 = prod(x).^(1/100), mfalse2 = prod(x.^(1/100)),
mx = gmean(x)
Algorithm The absolute value of the geometric mean is calculated from the arithmetic
mean of the logarithms of the absolute values; the phase is the phase of the
arithmetic mean value of the complex numbers.
References [1] .J. Schoukens and R. Pintelon, Measurement of Frequency Response
Functions in Noisy Environments, IEEE Trans. on Instrumentation and
Measurement, Vol. 39, No. 6, pp. 905-909, Dec. 1990.
[2] R. Pintelon, J. Schoukens and J. Renneboog, The Geometric Mean of Power
(Amplitude) Spectra Has a Much Smaller Bias than the Classical Arithmetic
Averaging, IEEE Trans. on Instrumentation and Measurement, Vol. 37, No. 2,
pp. 213-218, June 1988.
lin2qlog, log2qlog
3-49
lin2qlog, log2qlog
Purpose Generate quasi-logarithmic frequency set for use with FFT and for the
generation of periodic excitation signals.
Syntax fqlog = lin2qlog(freqv,rf)
[fqlog,df,cdmax,freqind] = lin2qlog(freqv,rf)
fqlog = log2qlog(freqv,mhno)
[fqlog,df,cdmax,freqind] = log2qlog(freqv,mhno)
Description Both routines calculate a subset of the linear frequency grid, where the
members of this subset are more or less logarithmically distributed.
lin2qlog starts from a linear frequency grid, given in freqv, and selects a
quasi-logarithmic set, providing that the ratio of successive frequencies is
about rf (or larger, if the frequency vector is not dense enough).
log2qlog starts from the given logarithmic frequency set freqv, and rounds
these frequencies to values of the linear (DFT) frequency grid. The harmonic
number mhno will be associated to the highest frequency. Multiple frequency
points are eliminated. log2qlog also works with a non-logarithmic input
vector, and produces a close equivalent on the DFT grid.
fqlog is the quasi-logarithmic frequency vector, and df is the size of the
frequency step in the corresponding full linear grid. The period length of the
corresponding multisine is usually 1/df, but if the harmonic numbers have a
common divider, it may be smaller. The largest common divider is given by
cdmax, thus the period length is in general 1/(cdmaxdf).
The harmonic numbers in fqlog can be calculated as
harmno = round(fqlog/df);
freqind is a column vector, containing indices of the selected frequency points
of freqv.
Examples fqlog = lin2qlog([1:128],sqrt(2));
[fqlog,df] = ...
log2qlog(logspace(log10(1),log10(256),17),256);
Diagnostics freqv must be real, non-negative and strictly increasing; mhno must be a
positive integer.
lin2qlog, log2qlog
3-50
Algorithm lin2qlog: starting from the first non-zero frequency point fqlog(1), the next
point fqlog(2) will be the one in freqv, closest to fqlog(1)rf, and larger
than fqlog(1). The points of freqv between fqlog(1) and fqlog(2) are
deleted. This is repeated until the end of the file: the last point will be taken
only if the last frequency is larger than or equal to fqlog(n-1)rf.
log2qlog: the points of fqlog will be chosen from the points of the grid
max(freqv)[1:mhno]/mhno. A point will be selected if it is closer to a point in
freqv than any other point of the linear grid.
See Also logspace
loadasc
3-51
loadasc
Purpose Load data from ASCII files into a variable.
Syntax data = loadasc(filename)
data = loadasc(filename,ASCIItype)
Description loadasc loads the contents of the ASCII file filename into data. In some sense,
this is an intelligent version of load, where
the generated variable need not have the same name as the file,
the ASCII file may contain comments, not only numbers,
the lines in the ASCII file need not contain the same number of numbers.
ASCIItype may have the value 'flat', in this case the load command is
directly used, or 'text', which means that the file contains some textual
information which has to be deleted during the loading, or the number of
numbers in a line may vary from line to line. The filtering of the file is done
using the command sscanf, available in MATLAB Version 4.0 or higher.
Everything after a % character in a line is ignored, and the rest is searched for
numbers. loadasc will read the file properly into a vector even if it is not
provided the same number of elements in each line if used with the option
'text'.
The resulting variable data is a column vector, unless ASCIItype is set to
'flat' and each line of the flat ASCII file contains more than one number (but
the number of elements in each row must be the same in a flat ASCII file).
Default
Argument
Values
ASCIItype = 'text'
Examples fprintf('lasctest.txt',...
'%.0f %%N\n%.4e, %.4e %%amp1\n',1,5,6)
type lasctest.txt, v3 = loadasc('lasctest.txt')
See Also load
loadvar, savevar
3-52
loadvar, savevar
Purpose Load single variable from MAT-file, and save variable to existing MAT-file.
Syntax variable = loadvar(matfile,varname)
savevar(matfile,varname)
savevar(matfile,varname,varvalue)
Description loadvar loads a single variable from a MAT-file and assigns its value to a
variable. The name of the file is given in the string matfile, the name of the
variable in the string varname.
savevar saves a single variable to an already existing MAT-file. The value of
the variable is given in varvalue. varvalue may be any valid expression in
MATLAB. If varvalue is not given, the variable will be deleted from the
MAT-file.
If no extension is given, the default extension (.mat) is appended to the
filename.
The MAT-file should be in the active subdirectory/folder if savevar is used, or
the path has to be explicitly given.
Note: using savevar, the variable donotusethisnamemf will also appear in the
MAT-file, with the name of the file as string value.
Examples The value of a variable can be checked easily without destroying the complete
workspace by loading the complete binary file
itlimit = loadvar('inpchans.ebn','itmax')
The value of a variable in a previously saved workspace can be changed,
and/or a new variable can be added to the saved workspace:
x = 1; save savevtst.mat x
savevar('savevtst.mat','x',10)
savevar('savevtst.mat','y',20)
Algorithm Both loadvar and savevar are function M-files, which assign the input
arguments to hopefully not used long variable names (donotusethisnamev,
donotusethisnamevv), load the MATLAB binary file, and create the desired
variable. savevar also saves the MAT-file again. loadvar and savevar will
behave in an unpredictable way when either of these two long variable names
are used in the MAT-file.
mlbs
3-53
mlbs
Purpose Generate maximum length binary sequence (pseudo-random binary sequence).
Syntax bitseries = mlbs(log2N)
[bitseries,nextstnum] = mlbs(log2N,bitno,startnum)
Description mlbs generates a maximum length binary sequence (column vector bitseries),
using a shift register of length log2N. The minimum value of the argument
log2N is 2, the maximum value is 30. The length of the generated sequence is
given by bitno, that is, a partial sequence can also be generated. The default
value of bitno is 2^log2N-1, that is, the full length of the PRBS with the given
register length.
The generation is based on a binary shift register with modulo 2 feedback. The
starting value of the register is startnum. The register contents can be obtained
via nextstnum, which can be used for the continuation of the sequence
generation.
Default
Argument
Values
bitno = 2^log2N-1, startnum = 2^log2N-1
Examples bitseries = mlbs(10); %Length: N = 1023 = 2^10-1
Diagnostics The register length must be an integer number between the allowed minimum
and maximum, otherwise an error message is sent:
log2N is not integer
or
log2N = ... is not allowed
startnum must be between 1 and 2
log2N
-1, otherwise the error message is
startnum out of range
Algorithm The feedback shift register is implemented in a MATLAB vector, according to
the definition (see [1]).
See Also dibs
References [1] K. R. Godfrey, ed.: Perturbation Signals for System Identification.
Englewood Cliffs, Prentice-Hall, 1993.
modifyfv
3-54
modifyfv
Purpose modifyfv prefilters measured data by the inverse of a known part of the
transfer function.
Syntax Fdatm = modifyfv(pdat,Fdat)
[Fdatm,vdatm] = modifyfv(pdat,Fdat,vdat,plotmode)
Description If a multiplicative term of the transfer function is known, it is disadvantageous
to estimate this part from the measured data, since this increases the variance
of the other term, too. A possible solution is to prefilter the input data by the
inverse of this term.
The parameter vector is given in pdat (a string if the filename is given).
Fdat contains the Fourier data: it is an array: [freqv,x,y], or a Fourier vector
(see expfou), or the name of the Fourier file.
The filtered Fourier data are given in Fdatm, as an array if Fdat is an array, or
as a Fourier vector if Fdat is a vector or a filename.
If the known partial transfer function has significant dynamics in the band of
interest, the variance data also have to be changed, since the modification
preserves the relative variance of the Fourier amplitudes, and in ELiS the
absolute variances are given.
vdat is a variance array, [varx,vary] or [varx,vary,covxy]; or a variance
vector (see expvar); or the name of the variance file. The filtered variance data
are given in vdatm, as an array if vdat is an array, or as a variance vector if
vdat is a vector or a filename.
plotmode defines the form of the plots of the results:
'lin', 'linpb' stands for linear frequency scale,
'log', 'logpb' stands for logarithmic frequency scale,
where pb requests plot in the passband only (the points where the transfer
function is zero will be excluded from the plot).
If plotmode has any other value, no plot will be shown.
modifyfv
3-55
Default
Argument
Values
plotmode = 'linpb'
Examples Fdatm = modifyfv('inpchanz.pbn','inpchan.fbn');
Algorithm The output Fourier amplitudes are divided by the transfer function values of
the known term, and output variances by their absolute square values; the
covariances are divided again by the transfer function values.
msinclip
3-56
msinclip
Purpose Minimize crest factor of multisine using the clipping algorithm.
Syntax cx = msinclip(freqv,ampv)
[cx,crx,crxmax,cry,crymax] = ...
msinclip(freqv,ampv,tf,gmod,itno,ovs,N,cl0)
Description msinclip iterates towards an optimum set of phases for which the crest factor
of the multisine with the given amplitudes (or the larger of the crest factors of
the two multisines at the input and the output of a linear system) is minimum.
freqv is the vector of frequencies where the nonzero amplitudes are given. The
elements must be integer multiples of a df value, and the minimum number of
sines is 2. freqv must be monotonously increasing.
ampv contains the absolute values of the desired nonzero complex amplitudes
at the corresponding frequencies (halves of the real amplitudes). If any element
of ampv is complex, the phases of ampv will be used as starting values, otherwise
the Schroeder multisine is used.
tf contains the complex transfer function values at the given frequencies. If tf
is given, input-output optimization will be performed.
If gmod is given with the value 'nograph', iteration results will not be plotted,
if with the value 'lastgraph', the result of the last iteration only, if with the
value 'graph10' or 'graph100', the result of every 10
th
or 100
th
iteration, if
with 'graph', the result of every iteration will be plotted.
When gmod contains the string 'ZOH' (e. g., its value is 'graphZOH', or just
'ZOH'), a zero-order hold multisine will be designed, instead of the
band-limited one. This means that a stepwise function will be designed
(prepared for a D/A converter), the number of samples will not be rounded up
to the next power of two, and the maximum overshoots between samples will
not be calculated.
itno contains the maximum number of iteration cycles. If itno = 0, the
starting values will be returned (the Schroeder multisine or the one externally
given).
msinclip
3-57
ovs determines the minimum resolution of the peak factor calculation; the
sampling frequency will chosen according to fs ovs2fmax. If the resolution
is small, and a band-limited design is requested (that is, not a zero-order hold
design), the grid is often not dense enough to catch the maximum values with
certainty. The default value of ovs = 16 usually provides errors less than 1%.
N offers a direct control of the length of the time series used for crest factor
calculation, often necessary in ZOH design. The point number in the time series
will be chosen equal to N in the ZOH design if the above condition for fs can be
fulfilled when generating just one period, otherwise the condition will
determine the point number. In the BL design the point number will be
rounded up to the next power of 2 to use a base-2 FFT, and the condition has to
be fulfilled again.
N must be an even number for msinclip.
If N is not given, it will be chosen to provide that each frequency be an integer
multiple of fs/N.
The last input argument, cl0, lets the user set the initial clipping level of the
algorithm between (0,1).
The output arguments are as follows.
The vector cx contains the complex amplitudes of the multisine (their absolute
values are equal to the halves of the real amplitudes).
crx is the crest factor of the generated multisine, calculated with the given
oversampling factor. However, the true value of the crest factor may be
somewhat larger because the true peak value is usually slightly larger than the
peak value on the grid.
msinclip
3-58
crxmax is the upper limit of the crest factor of the multisine, calculated for the
maximum possible peak value. This calculation is based on the determination
of the maximum curvature of the time function:
cry is the crest factor of the multisine, calculated with the given oversampling
factor, at the output of the linear system. crymax contains the worst-case crest
factor of the output multisine.
x t ( ) a
i
2f
i
t ( ) cos
i 1 =
N
=
d
2
x t ( )
t
2
d
--------------- a
i
2f
i
( )
2
2f
i
t ( ) cos
i 1 =
N
a
i
2f
i
( )
2
i 1 =
N
=
x
max
2f
i
( )
2 t
2
8
-------
i 1 =
N
msinclip
3-59
Typical plots of the results of msinclip are shown in the figures. The first one
illustrates a usual multisine design, the second one a zero-order hold one.
msinclip
3-60
crf denotes the calculated crest factor of the multisine, followed by the worst
case value. The actual clipping level is given next, and is shown by two dotted
lines on the plot. In cycle 0 or if the iteration has converged, the clipping level
is set equal to 1. The next number gives the number of performed iterations,
and opt. cyc. denotes the cycle in which the smallest crest factor was found
(the last plot of a run shows this multisine).
msinclip
3-61
Default
Argument
Values
ampv = ones(length(freqv),1), tf = [],
gmod = 'graph10' %plot every 10th result
The number of iterations: itno = 250.
ovs = 16 for band-limited design, ovs = 1 for zero-order hold design. In the
case of band-limited design, the sampling frequency will be
2ovs2fmax > fs ovs2fmax
or higher, when the given value of N prescribes it.
The clipping level will be chosen depending on the crest factor: for small crest
factors (around 1.5) as 0.9, and smaller if the crest factor is larger.
Examples Multisine design:
[cx,crestx] = msinclip([1:15]',ones(15,1));
Multisine design, iteration started from random phases:
cx = msinclip(4:15,ones(1,12).exp(j2pirand(1,12)));
Calculation of a Schroeder multisine:
[cx,crx] = msinclip([1:15]',[],[],'',0);
Calculation of a zero-order hold multisine:
[cx,crx] = msinclip([0.2:0.01:0.4]',ones(21,1)/sqrt(42),...
[],'ZOH',250,1,100);
Diagnostics ovs must be at least 1, and the frequency vector must increase strictly
monotonously. If a dc value is also given (freqv(1) = 0), the corresponding
amplitude must not be zero.
The frequency resolution (df) is also calculated from freqv (the smallest
common divider). If the maximum harmonic number is found to be larger than
1023 (which may well mean that the frequency vector was given erroneously),
a warning message is sent:
WARNING: maximum harmonic index found in 'msinclip' is ...
msinclip
3-62
Algorithm The algorithm is based on swapping between time domain and frequency
domain: the time domain waveform is clipped, then transformed to the
frequency domain, and the amplitudes are restored to the desired values. The
clipping level is slowly adjusted according to the evolution of the crest factor:
when the result is improved, the clipping level is decreased, otherwise it is
increased.
If input-output minimization is performed, the worse crest factor is minimized.
See Also msinprep, dibs
References [1] E. van der Ouderaa, J. Schoukens and J. Renneboog, Peak Factor
Minimization, Using TimeFrequency Domain Swapping Algorithm, IEEE
Trans. on Instrumentation and Measurement, 1988, Vol. 37, No. 1, pp. 144-147.
[2] K. R. Godfrey, ed.: Perturbation Signals for System Identification.
msinprep
3-63
msinprep
Purpose Generation of time domain multisine from complex amplitudes, and
preparation for downloading into an arbitrary waveform generator.
Syntax xtim = msinprep(freqv,cx)
[xtim,df] = msinprep(freqv,cx,N,fs,dev)
Description msinprep generates time series from a set of complex amplitudes (multisine).
The complex amplitudes are usually produced by msinclip. It is assumed that
the frequencies are harmoniously related and an integer number of periods is
to be generated. The algorithm can introduce a predistortion for a zero-order
hold.
xtim is the generated time series, df is the calculated common divider of the
given frequencies. freqv is the frequency vector, where the complex amplitudes
are given, cx is the vector of complex amplitudes, and N is the length of the time
series. If N is not given or is empty, it will be defined as N = fs/df, where df is
the maximal common divider of the frequencies in freqv.
fs is the sampling frequency. If it is not given, it will be chosen as fs = Ndf.
If N is not given, either, fs will be chosen as fs = df2(max(freqv)/df+1).
dev defines the device for which the series is prepared. Use dev = 'screen' for
plotting, etc., with no modification of the Fourier series, or use dev = 'DAC' for
D/A converter. In the latter case, the amplitudes will be multiplied by the
inverse transfer function of the zero-order hold.
Default
Argument
Values
dev = 'DAC'; fs and N as defined above.
Examples Generate a multisine and prepare it for downloading:
[cx,crx,crxmax] = msinclip([1:15]'/256,ones(15,1));
arbitgen = msinprep([1:15]'/256,cx,512,1,'DAC');
msinprep
3-64
Diagnostics freqv must be monotonously increasing, and must not exceed the half of the
given or above defined fs. N must also be large enough to have at least one
period of each sine. If the length of the time series is smaller than the period
length, the error message will be sent:
N (...) must be at least ... for one period
If the length of the time series is not equal to an integer multiple of the period
length, a warning message will be sent:
WARNING: the N (...) samples cover ... periods, this is not an
integer
Algorithm The amplitudes are divided by the transfer function of the zero-order hold. The
time function is calculated by inverse FFT.
See Also msinclip, optexcit
optexcit
3-65
optexcit
Purpose Generate the optimum input power spectrum for transfer function
measurements.
Syntax X = optexcit(pdat,freqv)
[X,CR,fsv,vXwdev,Fiw] = ...
optexcit(pdat,freqv,vdat,fixpind,X0,Ncyc,Fiw,pd)
Description optexcit iterates towards the optimum power spectrum of the input signal in
the sense that it minimizes the volume of the uncertainty ellipsoid of the
estimated parameters of the linear system. The transfer function is given by
pdat (the vector of all the parameters, see imppar; or a filename), the
frequencies where the optimum spectrum is looked for are given in the vector
freqv. The input and output variance vectors are given by vdat. If this is an
1-by-2 or 1-by-3 vector, its elements are taken as constant input and output
variances, and may be the input-output covariance. If this is an N-by-2 or
N-by-3 array, the variance vectors are formed from the first two columns, and
the covariance vector from the third one; if this is a vector, the variance vectors
and the covariance vector are obtained using impvar; if it is a string, the
variance file is looked for.
fixpind defines the fixed parameters in the following way: if the parameters
(numerator, denominator and the delay) are put together into a vector as:
[num,denom,delay]', the elements of fixpind are the indices of the fixed
parameters in this vector. (Here num and denom are row vectors defined in the
usual way: in descending order of powers of s in the s-domain, or in ascending
order of powers of z
-1
in the z-domain.) When fixpind is given as just fixpind
= 0, the zero-order coefficient of the denominator and the delay will be fixed.
fixpind = 'n' means that there are no fixed parameters.
X0 contains the starting values of the amplitudes, and Ncyc gives the number
of the iteration cycles. If Ncyc = 0, a partial run is performed, and the
returned amplitude vector equals X0, but the Cramr-Rao bound CR is properly
calculated. Fiw is a large array of partial information matrices, exported in a
previous run for the same system for acceleration of the subsequent runs.
pd is the density of plots: plotting occurs if rem(cycle,pd) = 0, and in the last
cycle. pd = inf will totally suppress plotting.
optexcit
3-66
X is the generated amplitude vector. CR is the Cramr-Rao lower bound of the
covariance matrix of the estimated parameters. For higher orders this matrix
is usually very badly scaled in the s-domain, thus for the calculation of the
determinant, scaling is advisable. fsv is the suggested scaling vector
(CRscaled = CR.(fsvfsv')).
vXwdev is an 1-by-2 vector, containing the minimum and maximum deviations
of the dispersion function from its limiting value, that is, the number of the free
parameters.
Fiw is the above mentioned array. If Fiw is too large to be stored on the
computer, it will be returned as empty array, and recalculated in every
iteration cycle.
Important: The delay may also be estimated. Consequently, if it is fixed, it is to
be given in fixpind.
A typical plot of optexcit is shown in the figure.
optexcit
3-67
The figure consists of four subplots. The first one shows the absolute value of
the transfer function (linear scale), the one below it represents the previous
amplitudes. The previous dispersion function shows the dispersion function
calculated from the previous amplitudes. The result of the actual calculation
cycle, the new amplitude set, is shown in the fourth subplot.
Below the dispersion function the number of free parameters is displayed (also
represented by the dotted horizontal line in the plot). This is the theoretical
final value of the dispersion function. The determinant of the covariance
matrix, and the determinant of its scaled version are also shown.
Default
Argument
Values
vdat = [1,1].
Index vector of the fixed parameters: if fixpind is not given, in the s-domain
fixpind = [nn+nd,nn+nd+1],
that is, the coefficient of s
0
in the denominator and the delay are fixed, while in
the z-domain
fixpind = [nn+1,nn+nd+1]
the coefficient of z
0
in the denominator and the delay are fixed. nn is the number
of parameters in the numerator, nd is the same in the denominator. Note that
fixpind = [] means that no fixed parameters are given (even the delay is
variable); the default value can be given with fixpind = 0.
X0 = ones(F,1)/sqrt(F)
where F is the length of the frequency vector,
Ncyc = 1.
Examples (see optexdem)
1 Let us reproduce the results given in [1], Subsection 4.3.5, for a bandpass fil-
ter. After each power spectrum optimization step the crest factor is mini-
mized, and the volume of CR is multiplied by an appropriate power of the
crest factor.
optexcit
3-68
Some remarks concerning the program: optexcit is performed consecutive-
ly, but some rearrangement of the results was necessary because the routine
produces the new amplitudes and the previous covariance matrix. Thus,
crest factor minimization is performed on Xold. Moreover, msinclip is used
in two steps: experience shows that it is advantageous to optimize first the
input crest factor, and then use this result for input-output optimization.
num = [3.2010e-17,5.5155e-12,8.973e-10,0,0];
denom = [1.0131e-21,2.5351e-18,3.6031e-14,...
5.5550e-11,3.5869e-7,2.5017e-4,1];
F = 50; fv = 20[1:F]'; pd = exppar('s',num,denom);
fixpar = [4;5;12;13]; np = 9; Fiw = '';
Xold = ones(F,1)/sqrt(F); Nold = 0;
for N = [1,2,3,4,10,11,100,101]
[X,CR,fsv,vXwdev,Fiw] = ...
optexcit(pd,fv,[1,1],fixpar,Xold,N-Nold,Fiw);
tf = polyval(num,sqrt(-1)freqv2pi)./...
polyval(denom,sqrt(-1)freqv2pi);
[cx,crx,crxmax] = msinclip(fv,Xold, [],'lastgraph');
[cx,crx,crxmax,cry,crymax] = ...
msinclip(fv,cx,tf,'lastgraph');
X = Xold; N = Nold;
cyc = N-1, dCR = det(CR), crx, dCRs = dCRcrx^(2np)
end %for N
2 Let us check the theoretical results of the example given in [1] (Section 4.1,
Example 2). A first order system is excited at just one frequency. The
Cramr-Rao lower bound can be calculated in closed form (CRtheor), thus it
can be compared to the covariance matrix given by the routine optexcit.
b0 = 1; b1 = 1; num = [1]; denom = [b1,b0]; f = 2;
parvect = exppar('s',num,denom);
cf = b0^2+(2pifb1)^2;
CRtheor = [cf(1+cf)/(2pif)^{2,0;0,cf(1+cf)];
[X,CR] = optexcit(parvect,f,[1,1],[1,4],1,1);
format long e, CR, CRtheor
Diagnostics The sizes of X0 and freqv must be equal, otherwise an error message is sent:
X0 has not the same size as freqv
optexcit
3-69
If fixpind is given with value 0, a warning message is generated:
The denominator coefficient of s^0 and the delay will be fixed in
optexcit
or
WARNING: the coefficient of z^0 and the delay will be fixed in
optexcit
The variance vectors given by vdat must have the same length as fvect,
otherwise an error message is sent:
The length of varx is ... instead of ...
Fiw is generated only if the computer can store it in full size. If this is not the
case, Fiw will not be generated (this results in a longer run time. If Fiw is
requested by defining it as an output argument, a warning message is sent:
WARNING! Fiw would be too large, it cannot be generated
and the output argument will be returned as an empty variable.
Algorithm The algorithm is the one described in [1], [2] and [3]. First the partial
information matrices are generated in each cycle for all the frequencies, then
the dispersion function is determined, and the new power distribution is
calculated.
See Also msinclip, dibs, msinprep
References [1] J. Schoukens and R. Pintelon, Identification of Linear Systems: a Practical
[2] F. Delbaen, Optimizing the Determinant of a Positive Definite Matrix,
Bulletin Socit Mathmatique de Belgique Tijdschrift Belgisch Wiskundig
Genootschap, Vol. 42, No. 3, pp. 333-346.
[3] K. R. Godfrey, ed.: Perturbation Signals for System Identification.
pairs
3-70
pairs
Purpose Find the closest point pairs in two complex vectors.
Syntax indab = pairs(a,b)
[indab,cycle,digits] = pairs(a,b,p,maxcycle,digitsreq,D)
Description pairs looks for a permutation of the elements of the complex vector b to
minimize
or if the array D is given,
The appropriate indices of b are returned in indab. The vector a may not be
longer than b. If each element of a is paired to an element of b, indab will
contain positive indices only; if not (the algorithm did not converge in the
allowed number of iterations), indab will contain at least one zero.
cycle gives the number of iterations of the Hungarian method. If it is returned
with the value zero, the simple nearest neighbors search gave the optimum.
The Hungarian method may converge very slowly in some rare cases.
Therefore, the maximum number of allowed iterations may be given in
maxcycle. If maxcycle is given with a finite value, the internal cost values of
the Hungarian method (the complements of the powers of distances, with
respect to the maximum value) will be rounded to a reasonable number of
digits (convergence is assured for integer numbers). Such a rounding will be
indicated by a value of digits, smaller than floor(log10(1/eps)), which is 15
on most platforms. The number of used digits can also be adjusted at the
beginning of the iteration (digitsreq).
Default
Argument
Values
p = 2, maxcycle = inf, digitsreq = floor(log10(1/eps))
a
i
b
i
perm
p
i 1 =
n
a
D i i
perm
, ( )
i 1 =
n
a
pairs
3-71
Examples indab = pairs(roots([1,2,3,4]),roots([1.01,2,3,4]),2);
pind = pairs([-0.1-1.5j,-0.1+1.5j],roots([1,2,3,4]));
If cycle == 0, it may be assumed that the new positions of the roots of the
perturbed polynomial p1 have been found. If cycle>0, and all(indab)>0, the
optimal indices are found, but there is a chance that the roots are not paired
properly, since the nearest neighbors cannot all be paired to each other.
Diagnostics The vectors are checked for infinite or NaN elements. If a or b is empty or is an
array, an error message will be sent. p is also checked for positivity.
If the closest pairs do not give the best permutation, a warning message is sent:
WARNING: nearest neighbors do not give optimum in PAIRS
The algorithm may converge very slowly. In this case the reciprocals of the
distances will be rounded, and a warning message will be sent:
Number of digits is set to ... in PAIRS
The algorithm may not converge if the number of allowed iterations is small. If
this happens, a warning message is sent:
WARNING: PAIRS did not converge
and indab will contain at least one element that is equal to zero.
Algorithm The so-called Hungarian method ([1], [2]) is used: it is based on looking for
alternating paths in bipartite graphs. The convergence of the method is
assured for rational distances only; the algorithm implemented here makes an
attempt to find a solution without rounding; when it fails, the complements of
the distances to the maximum distance are rounded. By this the smallest
distances, which are the most interesting, will be distorted the least.
See Also plotelpz
References [1] H. W. Kuhn, The Hungarian Method for the Assignment Problem, Naval
Res. Logist. Quart., Vol. 2, 1955, pp. 83-97.
[2] B. Andrsfai, Graph Theory: Flows, Matrices, Budapest, Akadmiai Kiad;
Bristol, UK, Adam Hilger, 1991.
plotelpz
3-72
plotelpz
Purpose Plot pole/zero pattern (maybe with confidence ellipses) of estimated transfer
function.
Syntax plotelpz(pdat)
[rnum,rdenom] =...
plotelpz(pdat,axv,cdat,Pc,ntx,parr,zarr,domain,da,dp,plm)
Description plotelpz plots pole/zero pattern of the transfer function defined by the
parameter vector pdat or by the named file if pdat is a string.
The scaling may be modified by axv: if this is a four-element vector, it will be
passed through axis; if this is 'z', the statement axis([-2,2,-2,2]) will be
executed; and if this is 'p', the plot will show all poles and zeros, and the
vertical and the horizontal scaling will be the same. If axv is empty or missing,
the plot will be scaled to show every pole and zero.
cdat is the covariance matrix (array), or the covariance vector (see expcov), or
the name of the file containing the covariance matrix of the estimated
parameters. It is used for the plot of the uncertainty ellipses.
The standard deviations of poles and zeros are calculated using stdpz.
The routine plots the one- contours as uncertainty ellipses. This can be
modified by Pc which determine the multiplier of for the contours. It is easy
to see that, supposing two-dimensional nondegenerate normal distribution, the
confidence limit for the event that the pole (zero) falls inside the contour is
p 1 e
s c
2
2
=
plotelpz
3-73
with sc being the multiplier of . This expression gives the following values:
if 0 < Pc < 1, Pc will be interpreted as the confidence level p. If Pc > 1, its
value will be assigned to sc.
In the degenerate cases (one-dimensional distribution, as for real poles and
zeros, or for certain constraints) the ellipses reduce to straight lines, and are
represented by narrow strips on the plots. In such cases the probabilities can
be calculated from the normal distribution (see third column).
The plot of uncertainty ellipses does not provide information about coupling of
poles and zeros. Such couplings can be explored using the routine stdpz.
For multiple zeros/poles the analytical sensitivity calculations give infinite
standard deviations: in such cases dotted rectangulars are plotted instead of
ellipses around the multiple zeros/poles.
By the help of the string argument ntx, the plot style can be modified. If it is
given with the value 'notext', no text at all will be written onto the plot; if
with the value 'nomsg', the warning messages will be suppressed only.
A set of poles and zeros each can also be given in the arrays parr and zarr,
accompanied by the domain ('s' or 'z'). These sets can also be plotted in
addition to or instead of the zeros and poles given in pdat.
plotelpz can pass some arguments to stdpz: da, dp and plm can influence the
calculation mode (see stdpz).
sc p degenerate p
1.0 0.39 0.68
1.5 0.68 0.866
2.0 0.86 0.954
2.5 0.96 0.988
3.0 0.989 0.9973
3.5 0.9978 0.99953
4.0 0.99966 0.999937
plotelpz
3-74
The output arguments rnum and rdenom are the row vectors of the poles and
zeros, calculated from pdat. They are exported from plotelpz to avoid the
necessity of repeated use of roots, if the order is large.
plotelpz can be used as a building block of complex M-files, especially with the
'notext' option. It does not even change the state of the graphics window,
previously set by the subplot or axes statement, if ntxt is given as 'notext'
or 'nomsg'.
A typical plot of plotelpz is shown in the figure.
The numbers of poles and zeros, the numbers of non-minimal phase zeros and
unstable poles are given on the plot, along with the number of zeros/poles not
shown because of the axis scaling applied. If any of the uncertainty ellipses is
so large that less than 10 points of the dotted line can be shown on the plot, a
warning message appears at the lower left corner:
WARNING: fully or partly not visible ellipse
plotelpz
3-75
Default
Argument
Values
axv = [], Pc = 0.39 (sc = 1), ntx = '', da = 'anal', dp = 1,
plm = ''.
Examples plotelpz('inpchanz.pbn','z');
[rnum,rdenom] = plotelpz('inpchans.pbn',[],'inpchans.cbn');
Algorithm The calculation of the poles/zeros is done by roots. The determination of the
uncertainty ellipses is rather involved. The covariances of the real and
imaginary parts of the poles/zeros are calculated by linear transformation from
the covariance matrix of the parameters, using the sensitivity matrix (see
stdpz).
See Also stdpz, roots, ploteltf
References [1] P. Guillaume, J. Schoukens and R. Pintelon, Sensitivity of Roots to Errors
in the Coefficient of Polynomials Obtained by Frequency-Domain Estimation
Methods, IEEE Trans. on Instrumentation and Measurement, Vol. 38, No. 6,
pp. 1050-1056, Dec. 1989.
ploteltf
3-76
ploteltf
Purpose Plot transfer function (may be with confidence bounds), and/or ratio of output
and input Fourier amplitudes.
Syntax ploteltf(pdat1)
ploteltf(Fdat)
[ha1,ha2,fsc] = ploteltf(pdat1,pdat2,Fdat,fscale,msc,...
cdat1,cc,cn,rec,expi,ntx)
Description ploteltf plots transfer functions. pdat1 and pdat2 may be parameter vectors
(see exppar), or parameter filenames. The first transfer function is plotted by
mark '', the second one by '.', at 256 equally distributed frequency points
(equal distribution is understood here in either linear or logarithmic scaling).
Also the ratio of output and input Fourier amplitudes can be plotted with mark
'+': these can be given in the Fourier vector Fdat, or in a file given by this
string.
fscale defines the scaling of the frequency axis. Possible values:
'lin': linear, from 0 to either the half of the sampling frequency in any of the
parameter sets, or the maximal frequency in Fourier frequency vector
(whichever is larger). If no Fourier data were given, and s-domain parameter
set(s) are investigated, the maximum frequency will be the double of the
suggested scaling frequency, calculated by exppar for the given transfer
function.
'linF': linear, from 0 to maximal frequency in the Fourier data.
'lin',f1,f2]: linear from f1 to f2.
'log': logarithmic, between 1e-3fs/2 to fs/2 (fs is the higher sampling
frequency in the parameter sets, if there is any), or from 1e-3maxf to maxf
(maximal frequency in the Fourier data). For s-domain parameter sets with no
Fourier file, plot from 1e-2fscale to 10fscale.
'logF': logarithmic, from minimal nonzero frequency to maximal frequency in
the Fourier data.
'log',f1,f2]: logarithmic from f1 to f2.
The scaling of the amplitude plot can be influenced by msc. Its possible values
are as follows: 'full' to scale to all points, 'passb' to scale to passband only
(defined by the minimum and maximum of frequencies, for which y is not zero
ploteltf
3-77
in Fdat). If a '' is appended to this parameter, only the phase is plotted, if a
'+', only the amplitude.
For confidence interval plots of pdat1, the covariance matrix of the estimated
parameters of pdat1 can be given by cdat1 as an array, or as a covariance
vector (see expcov), or as the name of the covariance file. The t2 bounds are
plotted, usually at 32 points of the estimated magnitude and phase values, or
at each frequency in Fdat. The number of points can be changed by cn between
1 and 256. The bounds can be changed by a multiplying factor cc. Its default
value is 2.
The value of cc and the approximate confidence limit (Pc) are displayed under
the plot, assuming normal distribution.
It is also possible to add confidence bounds to the nonparametric transfer
function estimate points, defined by Fdat. For this purpose, the variance data
have to be given in the place of cdat1, as a variance array [vx,vy] or
[vx,vy,cxy], or as a variance vector (see expvar), or as a variance filename.
ploteltf will recognize the type of the covariance/variance data; in an
ambiguous case (long vector form) a covariance matrix of pdat1 will be assumed
as given.
One is interested sometimes in the reciprocal of the transfer function, or in the
reciprocal of the measured transfer function in the Fourier data (equalization).
You can plot the reciprocal of any of the functions. rec = 'abc', where the
letters refer to pdat1, pdat2, Fdat, respectively: 's' means straight (no
reciprocal building is necessary), 'r' means reciprocal before plotting.
expi is the number of the experiment(s) in Fdat to be plotted. If expi is empty,
all the experiments in Fdat will be plotted.
ploteltf puts some textual information to the plots. If ntx is given with the
value 'notext', the filenames will not be shown in the plot.
For the purpose of further plots on the screen, the handles of the magnitude
and phase plots (the axis vectors in MATLAB 3.5) are exported to ha1 and ha2.
To avoid unwanted exponents below the frequency axis, the frequency vector is
often scaled internally (to kHz, MHz or mHz). The scaling frequency is exported
in fsc.
The magnitude plots are scaled in dB. To avoid annoying downscaling, the
zeros of the transfer function are substituted by -100 dB values.
ploteltf
3-78
A typical plot is shown in the figure.
Default
Argument
Values
Fdat = '', fscale = 'lin', msc = 'full', cdat1 = '', cc = 2,
cn = 32, rec = 'sss', expi = [], ntx = ''.
If only Fdat is given, this must be a Fourier filename, otherwise the program
cannot distinguish it from a parameter vector.
Examples ploteltf('inpchans.pbn','','inpchan.fbn','linF','full+');
num = [1.1,1]; denom = [4,3,2,1];
[ha1,ha2,fsc] = ploteltf(exppar('z',num,denom,0));
Algorithm The confidence bounds are calculated using stdtf or stdtfm.
See Also stdtf, stdtfm
rdueelis
3-79
rdueelis
Purpose Calculate residuals of an ELiS fit.
Syntax [rx,ry] = rdueelis(pdat,cdat,Fdat,vdat,expi)
[rx,ry,ryx,vryx,xe,ye] = ...
rdueelis(pdat,cdat,Fdat,vdat,expi,inp,outp)
Description rdueelis calculates the complex residuals of a fit of elis (see Study of the
Residuals on page 2-36). pdat is the parameter vector of the fit (see exppar),
or the name of the parameter file. cdat is the covariance array, or the
covariance vector, or the name of the covariance file of pdat. If cdat is empty,
the variance of the parametric estimate of the transfer function will not be
considered.
Fdat is the Fourier vector, or the array [freqvect,x,y], or the name of the
Fourier file. expi contains the number(s) of the experiment(s) in Fdat, for
which the residuals are to be calculated. vdat is a 1-by-2 vector if the variances
are constant (or a 1-by-3 vector if the covariances are also given), or an N-by-2
array if the variances are given point by point (or an N-by-3 array if the
covariances are also given), or a long variance vector if the variances are put
together by expvar, or a string if the variance file is referred to.
inp and outp select the serial numbers of the input and of the output port in
the Fourier file.
The residuals are given in rx (complex residuals of the input amplitude vector),
ry (complex residuals of the output amplitude vector), ryx (complex residuals
of ym./xm vs. the estimated transfer function).
vryx contains the variance vector of the real and of the imaginary parts of ryx.
If the distributions of the complex ryx values are circularly symmetric, this
equals the halves of the variances of ryx.
vryx is the difference of two quantities (see Study of the Residuals on page
2-36). Theoretically, this is never negative if the proper data are given (the
same variance vector as used in elis; pdat and cdat that belong to this run of
elis). However, because of the approximations, small negative values may
occur: these usually indicate very small variances.
xe and ye are the estimated complex input and output amplitude vectors. These
estimates are not consistent for one experiment. However, processing of several
experiments at the same time does decrease their estimation error.
rdueelis
3-80
Default
Argument
Values
expi = [], inp = 1, outp = 1.
Examples [rx,ry,ryx,vryx] = rdueelis('inpchans.pbn','inpchans.cbn',...
'inpchan.fbn',[9.61e-12,9.61e-10]);
Diagnostics The validity of the variance values and the common length of the variance and
frequency vectors is checked. If negative variance values are obtained, a
warning message is sent.
Algorithm The complex input and output amplitudes are estimated via weighted LS
fitting, having the estimated transfer function parameters fixed.
References [1] I. Kollr, On Frequency Domain Identification of Linear Systems, IEEE
Trans. on Instrumentation and Measurement, Vol. 42, No. 1, pp. 26, Feb.
1993.
simfou
3-81
simfou
Purpose Generate simulated Fourier amplitudes from parameters and variances.
Syntax [x,y] = simfou(pdat,freqv,x0,vdat)
[x,y] = simfou(pdat,freqv,x0,vdat,expno)
Description simfou generates simulated Fourier amplitudes (and perhaps also a Fourier
file) for elis. The parameters of the transfer function are given in the vector
pdat (see exppar), or in a file if pdat is a string. The frequencies are given in
the vector freqv. The input amplitudes are given in the vector x0.
The variances are defined by vdat: if this is an 1-by-2 or 1-by-3 vector, its
elements are taken as constant input and output variances (and perhaps the
input-output covariance). If this is an N-by-2 or N-by-3 array, the variance
vectors (and perhaps the covariance vector) are formed from the two (or three)
columns. If this is a vector, the variance and covariance vectors are obtained
using impvar. If it is a string, the variance file is looked for.
expno is the number of experiments to be generated.
Default
Argument
Values
expno = 1
Examples The convergence properties of elis can be checked by running it several times
on data simulated using a known transfer function:
freqv = [100:50:1000]';
num = 1; denom = [1e-6,1e-3,1];
pdat = exppar('s',num,denom);
vdat = [0.01,0.001];
[x,y] = simfou(pdat,freqv,[],vdat);
ploteltf(pdat,'',[freqv,x,y])
for k = 1:5
[x,y] = simfou(pdat,freqv,[],vdat);
Fdat = [freqv,x,y];
[pvect,fit,Cp] = elis(Fdat,vdat,['s',0,2],[],'',100);
pause
end
simfou
3-82
Diagnostics The routine checks the validity of the variance values and their compatibility
with the frequency vector, and sends an error message when incompatibility is
found.
When the system is unstable (the real part of a root is 0 in the s-domain or the
absolute value of a root is 1 in the z-domain), a warning message is sent:
WARNING! The given system is unstable in simfou
but the simulated values will be calculated in the frequency domain.
Algorithm simfou calculates the input and output amplitude vectors, and adds zero-mean
complex Gaussian noise with the given variances and properly set covariance
to them.
See Also simtime
simtime
3-83
simtime
Purpose Generate simulated input and output time series from a transfer function.
Syntax [xt,yt] = simtime(pdat,u,numi,denomi,numo,denomo)
[xt,yt] = simtime(pdat,u,numi,denomi,numo,denomo,typ)
Description simtime generates simulated input and output time series for elis from a
transfer function parameter set. The parameters of the transfer function are
given in the vector pdat (see exppar, imppar), or in a file if pdat is a string. The
excitation time series is given in u. The additive observation noise is generated
from white Gaussian noise of variance 1 by the noise shaping filters, defined by
numi, denomi and numo, denomo.
The input and output noises are independent. For the generation of correlated
noise, the inverse Fourier transform of the outputs of simfou can be used, after
having been called with a frequency vector [0:N/2-1]df.
The resulting time series are returned in xt and yt.
typ chooses between two basic possibilities. If its value is periodic, u is
considered as just one period of a periodic excitation, and one period of the
steady-state system response is calculated, while the value 'transient' makes
simtime produce the transient response, starting from energyless state.
If pdat defines an s-domain model, only steady-state simulation is allowed. In
such cases, typ must contain the sampling frequency in Hz.
Default
Argument
Values
typ = 'periodic'
Examples f = 400[1:49];
u = msinprep(f,msinclip(f,[],[],'',0),256,51200);
[xt,yt] = simtime('inpchanz.pbn',u,3e-5,1,3e-5,1);
[xtr,ytr] = simtime('inpchanz.pbn',u,...
3e-5,1,3e-5,1,'transient');
Diagnostics The leading coefficient of the denominator should not be close to zero (this
would approximate a predictor). If a too small leading coefficient is detected (its
simtime
3-84
absolute value is smaller than 10
-10
times the largest coefficient), a warning
message is sent:
WARNING: the leading coefficient of the denominator is very small
in simtime (maxdenom/denom(1) = ...)
In the transient case, the delay is realized by time shifts. The value must not
be negative (this would mean a predictor). If the value of the delay is negative,
an error message is sent:
The delay must not be negative (predictor cannot be simulated)
For the time shifting, the value of the delay is rounded to the nearest integer.
When the value of the delay is changed, a warning message is sent:
WARNING: the delay has been rounded in simtime
The system defined by pdat must be stable. If this is not true, an error message
is sent:
System is not stable
The noise shaping filters must be stable. If the absolute value of any of their
poles is larger than 1-10
-10
, an error message is sent:
Input noise shaping filter is not sufficiently stable
or
Output noise shaping filter is not sufficiently stable
Algorithm The transient response is calculated via the function filter of MATLAB, the
steady-state one by inverse Fourier transform of the frequency domain
response.
See Also simfou
stdpz
3-85
stdpz
Purpose Calculate standard deviations and covariances of poles and zeros of an
identified transfer function, using the covariance matrix of the parameters.
Syntax [zv,stdz,pv,stdp] = stdpz(pdat,cdat)
[zv,stdz,pv,stdp,rzp] = stdpz(pdat,cdat,zv0,pv0)
[zv,stdz,pv,stdp,rzp,g,stdg,dps] = ...
stdpz(pdat,cdat,zv0,pv0,da,plm,dp,axv)
Description If the covariance matrix of a set of estimated parameters is given, the
uncertainties of the zeros and poles can be calculated. pdat is the parameter
vector, usually generated by elis, or the name of a parameter file (see exppar).
cdat is the covariance matrix, corresponding to pdat, in the form of an array,
as generated by elis, or in the form of a vector (see expcov), or the name of a
covariance file.
zv and pv are the column vectors of the zeros and poles, respectively, and stdz
and stdp contain the corresponding uncertainties. In each row there are three
elements: the standard deviation of the real part, the standard deviation of the
imaginary part, and the correlation coefficient between the real part and the
imaginary part of the corresponding zero or pole.
For multiple zeros or poles the standard deviations cannot be determined by
analytical differentiation; in such a case the corresponding standard deviations
are given as NaN.
For the investigation of the interrelations of different poles and zeros, a
correlation coefficient matrix of the real and imaginary parts of all zeros and
poles and the two gain values can also be obtained in rzp. The order of the
random variables in the corresponding random vector is: real part of the first
zero, imaginary part of the first zero, real part of the second zero, etc., then the
real part of the first pole, etc. The last two elements are the leading coefficients
of the numerator and the denominator.
g contains the leading coefficients of the numerator and the denominator. The
standard deviations of the gain values and their correlation coefficient are
given in stdg.
Sometimes it is desirable to prescribe the order of the poles and zeros in zv and
pv. This can be done by giving zv0 and pv0. Any of these input arguments may
be given as an empty variable, if the order need not be prescribed.
stdpz
3-86
stdpz allows the use of numerical differentiation in the calculation of the zero/
pole sensitivities to variations of parameters. It can check whether variations
of the parameters, in the order of the standard deviations, cause the same
changes in the positions of zeros and poles, as calculated from the analytical
sensitivity calculations. The numerical differentiation can be requested by
giving da with the value 'num'.
The amount of perturbations can be influenced by dp. By default, its value is 1,
which means perturbations equal the eigenvalues of the covariance matrix, in
the directions of the eigenvectors. These steps can be multiplied by a given
value dp. When the perturbations are too large, the zeros and poles cannot all
be paired to their nearest neighbors in the perturbed sets, and a suggested
value of dp, for which all zeros and poles can be paired to their nearest
neighbors, will be given in dps; otherwise dps will be equal to the actual value
of dp.
The perturbed sets of zeros and poles can be plotted on the screen. If plm is
given with value 'mc', the perturbed sets and the pairing will be shown on the
screen, one after the other. For the value 'mp', a statement pause will be
executed after each pairing. The axis vector for these plots may be given in axv.
Default
Argument
Values
zv0 = [], pv0 = [], da = 'anal', dp = 1, plm = ''.
Examples [pvect,fit,Cp] = elis('inpchans.ebn');
[zv,stdz,pv,stdp,rzp] = stdpz(pvect,Cp);
Diagnostics The standard deviations cannot be calculated by analytical derivation for
multiple zeros or poles; in such cases the corresponding standard deviations
will be given as NaN.
Algorithm The sensitivity matrix of zeros and poles on the coefficients of the
corresponding polynomials is calculated using
S i j ( , )
r i ( )
j
df r i ( ) ( )
dr i ( )
-------------------
-------------------, =
stdpz
3-87
where r(i) is the ith root of the polynomial f(r(i)). The covariance matrix of the
real and imaginary parts of the poles/zeros is calculated multiplying the
covariance matrix of the parameters by the sensitivity matrix from both sides.
For numerical derivation, perturbations of the parameter vector are introduced
in the direction of each eigenvector of the covariance matrix, with step sizes
equal to the corresponding eigenvalue, multiplied by dp. The original and
perturbed sets are paired to each other by using pairs, with p = 1. This means
that even for the multiple zeros or poles a standard deviation will be calculated,
which will be in the order of the actual movements caused by the uncertainty
of the parameters.
See Also plotelpz
References [1] P. Guillaume, J. Schoukens and R. Pintelon, Sensitivity of roots to errors
in the coefficients of polynomials obtained by frequency-domain estimation
methods, IEEE Trans. on Instrumentation and Measurement, Vol. 38, No. 6,
pp. 1050-1056, Dec. 1989.
stdtf
3-88
stdtf
Purpose Calculate standard deviations of amplitude and phase values of an identified
transfer function, using the covariance matrix of the parameters.
Syntax [tf,stda,stdph,stdri,rtf] = stdtf(freqv,pdat,cdat)
Description The approximate standard deviations of the amplitude and phase values of
estimated transfer functions are calculated from the covariance matrix of the
estimated parameters. freqv is the vector of frequencies at which the standard
deviations are to be evaluated. pdat is the parameter vector, usually generated
by elis, or the name of a parameter file (see exppar). cdat is the covariance
matrix, corresponding to pdat, in the form of an array, as generated by elis,
or in the form of a vector (see expcov), or the name of a covariance file.
tf is the column vector of the complex values of the transfer function,
calculated from pdat at the given frequencies, stda is the column vector of the
standard deviations of the amplitudes, and stdph is the column vector of the
standard deviations of the phases, in radians.
If the distribution of the complex error of tf is circularly symmetric, stda also
equals the standard deviations of the real and imaginary parts, but these can
also be obtained in stdri, which has three columns: standard deviations of the
real parts and of the imaginary parts of tf, furthermore the correlation
coefficients of the real and imaginary parts of the corresponding tf value.
rtf is the correlation coefficient matrix of the vector of the real and imaginary
parts of all the points: the real and imaginary parts of the first point are the
first two elements, those of the second point are the third and the fourth, and
so on.
The variance of the transfer function can be calculated as
vartf = stda^2 + (abs(tf).*stdph).^2
or
vartf = stdri(:,1).^2 + stdri(:,2).^2
Examples [pvect,fit,Cp] = elis('inpchans.ebn');
[tf,stda,stdph,stdri,rtf] = stdtf([400:400:19600],pvect,Cp);
stdtf
3-89
Algorithm The sensitivity matrices of the amplitudes and phases on the estimated
parameters, furthermore of the real and imaginary parts are calculated by
direct derivation of the corresponding expressions. The appropriate covariance
matrices of the amplitudes and phases are calculated multiplying the
covariance matrix of the parameters by the sensitivity matrices from both
sides.
See Also fdcovpzp, ploteltf
stdtfm
3-90
stdtfm
Purpose Calculate empirical standard deviations of nonparametric transfer function
estimates.
Syntax [tfm,stdAm,stdphm] = stdtfm(Fdat,vdat)
Description For model validation purposes you may wish to calculate the nonparametric
transfer function estimate tfm = ym./xm, and check the uncertainties of these
values.
Fdat contains the Fourier data. It is either an array: [freqv,xm,ym], or a
Fourier vector (see expfou), or the name of the Fourier file. The variances are
defined by vdat. If this is an 1-by-2 or 1-by-3 vector, its elements are taken as
constant input and output variances (and perhaps the input-output
covariance). If this is an N-by-2 or N-by-3 array, the variance vectors (and
perhaps the covariance vector) are formed from the two (or three) columns. If
this is a vector, the variance and covariance vectors are obtained using impvar.
If it is a string, the variance file is looked for.
tfm is the nonparametric estimate of the transfer function: tfm = ym./xm.
stdAm contains the approximate standard deviations of the absolute values of
the transfer function. These are approximately equal to the standard
deviations of the real and of the imaginary parts of tfm, if the distribution of
the complex noises is circularly symmetric, as it is in the usual case.
stdphm returns the standard deviations of the phases in radians.
Examples [tfm,stdAm,stdphm] = stdtfm('emachine.fbn','emachine.vbn');
Diagnostics The routine checks the validity of the variance values and their compatibility
with the frequency vector, and sends an error message when incompatibility is
found.
stdtfm
3-91
Algorithm stdtfm calculates the standard deviations by evaluating the formula
which gives the standard deviations of the real and of the imaginary parts of
tfm. The output argument stdphm is obtained as stdAm./abs(tfm).
If the standard deviations of the estimate tfav = my./mx are sought, using the
results of varanal, these can be obtained by vx/Na, vy/Na, and cxy/Na,
respectively.
See Also ploteltf
xk
2 Y
mk
X
mk
---------
2
yk
2
2real c
xyk
Y
mk
X
mk
---------
,

_
,

_
+
X
mk
---------------------------------------------------------------------------------------------
Y
mk
X
mk
---------
xk
2
X
mk
2
---------------

yk
2
Y
mk
2
--------------- 2real
c
xyk
X
mk
Y
mk
-------------------
,

_
+ =
tim2fou
3-92
tim2fou
Purpose Convert time domain data to frequency domain data for elis.
Syntax [x,y] = tim2fou(tdat,freqv)
[x,y,fv] = tim2fou(tdat,freqv,expi)
Description The time domain data vector is given in tdat (see exptim); or this is an array
[tvl,xt,yt], where the column vector tvl contains the time instants,
repeated with the same values for each experiment; or this is the name of the
time domain data file, with obligatory extension .tbn, .tim, or .tnt.
tdat can also be the name of an M-file. For example, if tdat = 'gettim', the
call [tv,xt,yt] = gettim(i) must return the results of experiment i for the
values contained in expi.
freqv is the desired frequency vector; possibly each element of freqv should be
a divider of the sampling frequency, but the algorithm works even if this is not
true. If freqv is empty or missing, all possible frequencies in the FFT grid will
be used from 0 to fs/2.
The generated input and output vectors (or arrays for multiple inputs or
outputs) of complex amplitudes are x and y. Multiple experiments can also be
processed. The actually used frequency vector is returned in fv.
Default
Argument
Values
freqv = [0:N/2-1]'/Nfs, where N is the length of the time vectors, and fs =
1/dt is the sampling frequency.
Examples [x,y,fv] = tim2fou([[0:63]',cos([0:63]'/642pi3)]);
Diagnostics The time vector should consist of equidistant points, otherwise an error
message is sent:
Samples are not equidistant
fs = 1/dt should be larger than the maximum of freqv, otherwise an error
message is generated:
Maximum of freqv is larger than fs/2
tim2fou
3-93
Every element of freqv has to be a divider of the sampling frequency, otherwise
a warning message is sent:
'WARNING: frequencies are not FFT points, leakage will appear
Algorithm The usual expression of the DFT is evaluated:
If the frequency points in freqv are on the FFT grid, the DFT of the time
domain vectors is calculated via FFT, and the corresponding complex
amplitudes are selected. Otherwise the DFT is evaluated for the given
frequency points.
See Also exptim
X
k
x
i
e
j 2
ki
N
----
i 1 =
N 1
=
varanal
3-94
varanal
Purpose Averaging and variance analysis of multiple experiments.
Syntax [vx,vy,cxy] = varanal(Fdat,expi)
[vx,vy,cxy,mx,my,Na,Np,cfl,dv,sd] = ...
varanal(Fdat,expi,synch,T,inp,outp)
Description varanal averages frequency domain amplitudes as results of experiments, and
calculates the empirical variances and input-output covariances of complex
amplitudes. The results of the individual experiments are given either in just
one Fourier vector Fdat (see expfou); or an array [freqvlong,x,y], where
freqvlong contains the frequency vector, repeated as many times as the
number of experiments (the first frequency may not be repeatedly present in
the short frequency vector); or in a file with name given in Fdat. A further
possibility is when the string Fdat is the name of a user-defined function M-file,
which can be called using eval. If the serial number of a given experiment is i,
and Fdat is 'getfou', then [freqv,x,y] = getfou(i) must return the data of
the i
th
experiment.
expi is the vector of the serial numbers of the experiments to be processed.
Averaging can be easily performed if the measurements were made in a
synchronized way. However, when this is not the case, the routine can also be
asked to try to synchronize the input vectors (assuming that they resulted
from FFTs of the same length), minimizing the weighted phase differences of
the complex amplitudes in the frequency domain by introducing a delay. Such
synchronization can be requested by giving the argument synch with the value
'synch'. For this synchronization attempt, the maximum value of the delay
(the period length for periodic signals) is advisable to be given in T, although
varanal makes an attempt to find a reasonable period length if T is not given.
Since T specifies the range of rough search, a value of T slightly larger than the
period length is preferred to a smaller one.
For synchronization, the elements of the frequency vector have to be given in
Fdat with sufficient accuracy. The best solution is to generate them in MATLAB
from the harmonic numbers. Inaccurate frequency values can be prohibitive for
appropriate phase fitting.
If there are several input and/or output ports in the measurement, the user
may choose from among them by defining the scalars (or vectors) inp and outp.
varanal
3-95
If any of them is not given, or it is given as an empty vector, each port present
in the Fourier data will be used.
The variances are returned in the vectors (arrays) vx and vy, the input/output
covariances
in cxy. In the case of MIMO data, cxy will be an array, containing the
covariances beside each other, as [cx1y1,cx1y2,...cx2y1...].
The averaged input and output amplitudes are returned in mx and my.
The variances and covariances of the averaged complex amplitudes mx and my
can be obtained as vx/Na, vy/Na, cxy/Na, respectively.
If the nonparametric estimate Ym./Xm is calculated, its approximate standard
deviation can be determined from vx, vy, and cxy, using stdtfm. The standard
deviations are to be scaled by 1/sqrt(Na) when my/mx is used.
The number of averaged experiments is returned in Na, while the total number
of processed experiments is Np. The two numbers may differ because when a
synchronization attempt fails, the given experiment is not averaged to the
others.
The multiplicative factors to obtain the lower and higher bounds of the 95%
confidence intervals of the variances in the Gaussian case are returned in the
12 vector cfl. The confidence limits of the real and imaginary parts of the
covariance values can be obtained by calculating
(cfl-1)sqrt(varx(k)vary(k)).
If synchronization is requested, the vector of all the determined delay values is
returned in dv. For the values of the delay the inequalities -T/2<delay<=T/2
hold. If the synchronization failed, the corresponding element of dv is NaN.
The delays obtained by varanal can be used for restoring synchronization: if
the value dv(i-1) is obtained for the complex amplitude set Xi,
exp(j2pifreqvdv(i-1)).Xi will give the amplitudes which correspond in
phase to the reference set.
0.5E N
x
N
x
{ }
varanal
3-96
sd is the Cramr-Rao bound of the delay values. It is calculated from vx and mx
(see [1]):
Using the value of sd, a 95% significance level test is performed in order to
determine if the estimated delays can all be zero (synchronized experiments).
If there is no significant deviation from the hypothesis that the experiments
are synchronized, a warning message is sent:
The estimated delays are small. With this SNR,
the statistical test shows no significant desynchronization.
If the signal-to-noise ratio is small, the synchronization attempt may fail. This
is detected by observing for the best fit a phase deviation higher than /2 at
least at one frequency. When this happens, the experiment will not be averaged
to the others.
Default
Argument
Values
synch = '', inp = [], outp = []
Examples [vx,vy,cxy,mx,my] = varanal('bandpass.fbn',[1:6],'synch');
[vx,vy,cxy] = varanal('lowpass1.fbn',[1:5],'synch');
Diagnostics Synchronization can only be done if the Fourier amplitudes turn around by
integer multiples of 2 with a time shift of T. If this condition is not met, a
warning message is sent:
WARNING! Not all frequency vector elements are integer multiples
of 1/T
Maximum relative deviation is larger than 1e-6 in varanal.
sd
2
mx
k i i ,
2
i i 1 =
i npno
vx
k
---------------------------------- 2f
k
( )
2
,

_
k 1 =
F
---------------------------------------------------------------------- =
varanal
3-97
varanal makes an attempt to find the period length of the signals from the
frequency vector, and compares this to the given value T. if there is a deviation,
one of the following warning messages is sent:
WARNING! T does not cover full period length found by varanal: T
= ..., Tp = ...
The proper delay may not be found.
or
WARNING! T is larger than period length found by varanal:
T = ..., Tp = ...
Search time may be unnecessarily long.
The synchronization procedure assumes that the signal-to-noise ratio is not
very small, that is, in synchronized state the phase differences of the complex
amplitudes in the different experiments, at the frequencies given in freqv, are
smaller than /2. Otherwise the experiment is skipped (and the delay value NaN
appears in dv), and a warning message is sent:
Synchrnization is not successful for experiment ...
Algorithm Averaging of every quantity is done in a recursive way.
The confidence limit factors are calculated from the approximation of the
2

distribution:
mx
k
k 1
k
----------- mx
k 1
1
k
---x
k
+ = mx
1
x
1
= ,
vx
k
s
xk
*2 k 2
k 1
-----------vx
k 1
k
2 k 1 ( )
2
---------------------- x
k
mx
k
2
+ = = vx
1
0 = ,
cxy
k
k 2
k 1
-----------cxy
k 1
k
2 k 1 ( )
2
---------------------- x
k
mx
k
( ) y
k
my
k
( ) + cxy
1
0 = , =
cfl
2 N
a
1 ( )
2N
a
2
1
,
2
----------------------------------
1
1
2
9 2 N
a
1 ( ) ( )
------------------------------- 1.96
2
9 2 N
a
1 ( ) ( )
------------------------------- t
,
_
3
--------------------------------------------------------------------------------------------------- = =
varanal
3-98
The degrees of freedom equal 2(Na-1), because both the real and imaginary
parts are used in the estimation.
The synchronization attempt starts with a scan through the possible values of
the delay in steps of T
pmin
/6, where T
pmin
is the smaller value of T and the
period length of the largest frequency in the vector freqv. Using the best delay
value, an additional, appropriately weighted LS fit of the phases is performed
(see [1]). The resulting delay estimate is a maximum likelihood one, if the
noises are Gaussian.
See Also tim2fou
References [1]I. Kollr, Signal Enhancement Using Non-synchronized Measurements,
IEEE Trans. on Instrumentation and Measurement, Vol. 41, No. 1, pp. 156-159,
Feb. 1992.
yesinput
3-99
yesinput
Purpose Intelligent input routine.
Syntax answer = yesinput(question,default)
answer = yesinput(question,default,possib)
Description yesinput is an intelligent version of the statement input. It displays a
message as input usually does (string question), offers a default answer
(defined by default) which can be accepted by pressing Return or Enter, and
checks the validity of the answer using the optional argument possib. The type
of the returned answer (string or number) is determined by the type of default.
If the type of the desired answer is string, possib may be either a string array,
where the rows contain the acceptable answers, or a string containing the
acceptable answers, separated by ('|') characters. If a number is desired,
possib may be an 12 vector, containing the lower and higher limits for the
input:
possib(1) answer possib(2).
possib(1) may be -inf, possib(2) may be inf.
If a number is to be typed in, any valid MATLAB expression may be given, e. g.,
2pi/128.
If the answer is not acceptable, the user is prompted again for a new answer.
For testing purposes, yesinput can be forced to accept the default answers and
not to wait for the keyboard, by defining the global variable yesinpacceptdef
with value 'yes'.
Examples order = yesinput('Order of the filter',10,[0,12]);
color = yesinput('Color of plot','red','red|blue|green');
See Also input
ywalk
3-100
ywalk
Purpose Fit absolute values of frequency response points by a linear z-domain transfer
function.
Syntax function [B,A] = ywalk(na,ff,aa,npt,lap)
Description ywalk is a slightly modified version of the routine yulewalk of the Signal
Processing Toolbox; the built-in Hamming window is switched off. For more
details, see yulewalk.

A
Appendix
A-2 Description of the Data Vector and File Formats
A-2 Common Conventions for Every Data Type
A-4 Conventions for Individual Data and File Types
A-10 Examples for ASCII Files
A-15 Description of the Available Measurement Data Files
A-15 aluplate
A-16 bandpass, bp1, bandp8k
A-17 cable1-6
A-24 crankcas
A-25 emachine
A-26 glassfib
A-28 inpchan
A-29 lowpaswb, lowpass1, lowpass2
A-33 robotarm
A Appendix
A-2
Description of the Data Vector and File Formats
This section describes the following data vectors and file types:
time vectors and files containing measured or simulated time domain data
Fourier vectors and files containing measured, simulated or calculated
frequency amplitudes or input-output point pairs
variance vectors and files containing variances of the real parts (that is, also
of the imaginary parts) of the measured frequency amplitudes, sometimes
along with complex input/output covariances
parameter vectors and files containing parameters of the transfer functions
covariance vectors and files, containing the covariance matrix of the
estimated parameters
report files, containing textual documentation of the estimation procedure
The data vectors contain the same data in the same order as the flat ASCII files
described below. Thus, data vectors will not be treated separately.
Common Conventions for Every Data Type
Vectors and ASCII Files
ASCII files may be full files (with comments), or flat files (containing numbers
only).
The individual lines of full ASCII files may contain
comments (lines beginning with %)
a number, sometimes followed by explanation after the percent sign
several numbers, sometimes followed by explanation after the percent sign
For the purpose of easier interfacing to MATLAB, the ASCII files may be filtered
(ASCII flat file format): in this case they contain no text, the comment lines are
deleted, and numbers that were originally in the same row are put into
separate rows. For example, the rows
1.0000 2.0000 3.0000 4.0000
1.1000 2.1000
A-3
may be transformed into the form
1.0000
2.0000
3.0000
4.0000
1.1000
2.1000
This flat format corresponds to the data vectors used by the toolbox in the
workspace.
The data given in a file are identified by their position among the numbers in
the file; the comments and explanations are for readability only. However, we
strongly recommend you to use the standard explanations in full ASCII files,
because this makes possible to verify the contents of the files. The file structure
is also slightly redundant, thus programs can also verify to some extent
the correctness of files.
Full ASCII files can be directly read by the functions impfou, imppar, impvar,
impcov, and imptim, or just loaded into a MATLAB vector by using the function
loadasc, which is based on the C routine sscanf, accessible from MATLAB 4.0.
Further Conventions for the ASCII Files. A number in a line may be preceded
by spaces or tabs, but by no other character; separators between numbers are
spaces or tabs, sometimes commas. The numbers may be integers, or they may
be given in scientific notation, with the usual restrictions adopted (they may
not begin with a period the valid beginning is a digit, maybe preceded by a
sign, e. g., 1.0e-5, -1.2, etc). The number of digits is determined by the
individual programs, taking into account the relevant features of the data
(noise, uncertainty, etc.). Do not use much more digits than necessary to keep
the information, because this may decrease the speed of loading data from the
file.
The first lines of each file should be comment lines. The first comment line
should contain the filename (for printouts), file type with a version number of
the format, and the date (and perhaps also the time). Further comment lines
may contain text necessary for identification of the data.
The filenames consist of a name (names not longer than 8 characters are
preferred, to provide the easy transfer to MS-DOS), and an extension preceded
by a period (.). The standard extensions are for full ASCII files: .tim, .fou,
A Appendix
A-4
.var, .par, .cov, and .rep, but the programs usually can handle files with
other extensions. The names themselves must not contain periods or commas.
The extensions of the names of the filtered (flat ASCII) files end in .nt (which
stands for no text): the standard extensions are .tnt, .fnt, .vnt, .pnt, and
.cnt.
Files containing no text can be filled in with the appropriate texts; this can
currently be done by using the toolbox function elisfcnv.
Binary Files
Data read/write is much quicker through binary files. To maintain easy access
from MATLAB, use the .mat binary format (see the online MATLAB Reference
Guide, load and save statements). In binary files, each number or group of
numbers is identified by a standard name. The string variables comments,
fdate, and ftype may contain the necessary data for identification. The
comments string need not contain % marks (but the % mark is allowed in the
string). It may contain cr (Macintosh), lf (UNIX) or cr/lf (IBM-PC)
characters marking multiple lines of comments, but this is not recommended if
the transfer of binary files between different platforms is foreseen.
The standard extensions of the names of binary files are: .tbn, .fbn, .vbn,
.pbn, and .cbn.
An ASCII file can be converted into binary format or vice versa by using the
elisfcnv function of the toolbox.
Conventions for Individual Data and File Types
Examples for the individual file types are given in the next section.
Time Vector and File
The sample ASCII file shown in this Appendix is self-explanatory. However,
some remarks can be made concerning special problems:
1 The start time and the stop time are redundant data beside the time vector.
They are given for ease of manual check only, and are not used (though may
be checked) in programs. Usually the time vector is not important, either,
since sampling is usually equidistant; however, this convention was adopted
to be able to handle non-uniformly sampled data.
2 The time instants given must be the same for each experiment in a file.
A-5
The binary file (.tbn) contains the following standard variables:
comments eventual identification string (optional)
ftype 'time Vx.x'
fdate date (and time) string (optional)
expno number of experiments
timevect column vector of time instants
xt column vector of input amplitudes, or input amplitude array which
has as many columns as the number of inputs; the results of different
experiments are just put below each other.
yt column vector of output amplitudes, or output amplitude array
which has as many columns as the number of outputs; the results of
different experiments are just put below each other.
Fourier Vector and File
The sample ASCII file shown in this Appendix is self-explanatory. However,
some remarks can be made concerning special problems:
1 The start frequency and the stop frequency are redundant data beside the
frequency vector, they are given for ease of manual check only, and are not
used (though may be checked) in programs.
2 The frequencies given must be the same for each experiment in a file.
3 If the data come from amplitude-only measurements, the imaginary parts
are set to zero, which is a clear indication of the nature of the data.
A Appendix
A-6
The binary file (.fbn) contains the following standard variables:
ftype 'Fourier Vx.x'
expno number of experiments
freqvect column vector of frequency points
x column vector of complex input amplitudes, or input amplitude array
which has as many columns as the number of inputs; the results of different
experiments are just put below each other.
y column vector of complex output amplitudes, or output amplitude array
which has as many columns as the number of outputs; the results of
different experiments are just put below each other.
Variance Vector and File
These vectors and files contain the variances and perhaps covariances
associated with complex amplitudes given in Fourier files. The variances are
those of the real parts (that is, also of the imaginary parts) of the measured
frequency amplitudes, in other words, these variances are the halves of the
variances of the complex amplitudes. The covariances are defined as follows:
Since variances and covariances have a meaning with the corresponding
Fourier file only, you should give the same name to both files. The variance files
contain 2 or 4 or (2+2inputnooutputno) columns, the first two being the
variances of the real (and of the imaginary) part of the corresponding complex
amplitude vector, respectively, and the last columns being the real and
imaginary parts of the complex covariances. For MIMO systems, the order of
covariances is: ci1o1, ci1o2,...,ci2o1,...
c
xyz
0.5cov N
xk
N
yk
{ , } 0.5E N
xk
N
yk
{ } = =
A-7
The binary file (.vbn) contains the following standard variables:
ftype 'variance Vx.x'
varx column vector (or array for multiple inputs) of the input variances
vary column vector (or array for multiple outputs) of the output variances
covxy column vector (or array) of complex covariances: for MIMO
systems, it may contain the covariances between just input1 and output1, or
all the covariances beside each other, as [c11,c12,...,c21...]
Parameter Vector and File
The protocol of runs of elis may be rather complicated, since a lot of data
relevant to the run may be required. Thus, the protocol is provided in the report
file. However, there is a standard basic set of data that should be provided in
each case. In the following, this general block is defined.
The standard set of data is as follows:
Domain
Sampling frequency (in z-domain) or suggested scaling frequency (in
s-domain)
Order of the numerator
Order of the denominator
Coefficients of the numerator
Coefficients of the denominator
Delay
Notice that in the sample files along with the textual definition of the domain,
a number is also given (0 for the s-domain and 1 for the z-domain). This
provides that the domain information is contained also by flat ASCII files.
The above data may be followed by additional information, usually not used by
programs. However, more information about the run of elis is provided in the
vector fit, and in the report file.
A Appendix
A-8
The binary file (.pbn) contains the following standard variables:
ftype 'parameter Vx.x'
domain 's' or 'z', domain of the fit
fs sampling frequency (in z-domain); suggested scaling frequency (in
s-domain)
num coefficients of the numerator (in ascending order of powers of z
-1
or in
descending order in powers of s)
denom coefficients of the denominator (in ascending order of powers of z
-1

or in descending order in powers of s)
delay time delay associated to the transfer function (usually zero or a
negative number)
Covariance Vector and File
This file gives the covariance matrix of the coefficients (that is, coefficients of
the numerator, coefficients of the denominator, and the delay). For the fixed
coefficients, the rows and columns contain just zeros. The data set given in the
ASCII file is just the upper right triangle of the covariance matrix. The order
of the coefficients is numerator, denominator, delay. Numerator and
denominator coefficients are ordered as in the coefficient vector, that is, in
ascending order of powers of z
-1
, or in descending order of powers of s. Each row
of the triangle is given in a separate line; if the lines were too long, the
continuations of a line may be put into a separate line, preceded by five spaces.
The binary file (.cbn) contains the following standard variables:
ftype 'covariance Vx.x'
coeffcovar covariance matrix of the parameters (zeros for the fixed
parameters)
A-9
Report File
This ASCII file gives the information relevant to a run of elis, the format is for
direct use as text. The contents are:
Date, time, comments
Domain
Sampling frequency (z-domain) or suggested scaling frequency (sdomain)
Order of the numerator
Order of the denominator
Coefficients of the numerator
Coefficients of the denominator
Delay
Fixed numerator coefficients
Fixed denominator coefficients
Iteration algorithm, number of iterations
Way of setting the start values
Condition numbers
Total run time
Value of the cost function, and last variation
Degree of freedom
Theoretical cost function value and the associated confidence limit
Maximum of absolute deviations
Poles and zeros
etc.
A sample report file is given in the Reference chapter, in the description of elis.
A Appendix
A-10
Examples for ASCII Files
Sample Time File
%filename: data.tim, ftype: time V1.0, 06Nov1993, 11:51
%XH2 Synchronous machine identification
2 %Number of inputs
1 %Number of outputs
2 %Number of experiments
5 %Number of samples
0 %Start time (s)
1000 %Stop time (s)
%
% Measurement results
% Time input output
1 %Experiment no.
0.00 1.00 2.00 0.0125
250.00 1.00 3.00 0.2345
500.00 1.00 4.00 2.456
750.00 1.00 5.00 15.63
1000.0 1.00 6.00 5.678
%
2 %Experiment no.
0.00 1.00 2.00 0.0125
250.00 1.00 3.00 0.2545
500.00 1.00 4.00 2.876
750.00 1.00 5.00 16.63
1000.0 1.00 6.00 4.678
%
%End of time file
A-11
Sample Fourier File
%filename: data.fou, ftype: Fourier V1.0, 06Nov1993, 11:51
1 %Number of inputs
2 %Number of experiments
5 %Number of frequencies
0 %Start frequency (Hz)
1000 %Stop frequency (Hz)
%
% Measurement results
%Frequency input output
1 %Experiment no.
0.00 1.00 0.00 0.0125 1.03e4
250.00 1.00 0.00 0.2345 0.05412
500.00 1.00 0.00 2.456 0.8455
750.00 1.00 0.00 15.63 1.684
1000.0 1.00 0.00 5.678 0.3566
%
2 %Experiment no.
0.00 1.00 0.00 0.0123 1.08e3
250.00 1.00 0.00 0.2873 0.03712
500.00 1.00 0.00 2.736 0.8733
750.00 1.00 0.00 13.68 1.687
1000.0 1.00 0.00 3.678 0.8366
%
%End of Fourier file
A Appendix
A-12
Sample Variance File
%filename: data.var, ftype: variance V1.0, 06Nov1993, 11:51
1 %Number of inputs
1 %Number of covariances (0, 1 or inputno outputno)
%Variances
% Input Output Covariance: real, imag
1.0e5 5.2e4 3.9e5 2.6e4
1.0e5 5.2e4 3.9e5 2.6e4
1.0e5 5.2e4 3.9e5 2.6e4
1.0e5 5.2e4 3.9e5 2.6e4
%End of variance file
A-13
Sample Parameter Files
%filename: data.par, ftype: parameter V1.0, 06Nov1993, 11:51
%
0 %sdomain
1e3 %Suggested frequency scaling factor
0 %Order of the numerator
2 %Order of the denominator
%Coefficients of the numerator
%Powers of s coefficients
0 1.324678532345477e2
%
%Coefficients of the denominator
%Powers of s coefficients
0 1.000000000000000
1 1.596753424395043e+3
2 1.896563012344548e+6
0.0000 %Delay
%
%End of parameter file
%filename: data.par, ftype: parameter V1.0, 06Nov1993, 11:51
%
1 %zdomain
1e3 %Sampling frequency
0 %Order of the numerator
2 %Order of the denominator
%Coefficients of the numerator
%Powers of z^1 coefficients
0 1.324678532345477e2
%
%Coefficients of the denominator
%Powers of z^1 coefficients
0 1.000000000000000
1 1.596753424395043e+3
2 1.896563012344548e+6
%
0.0000 %Delay, normalized to 1/fs
%
%End of parameter file
A Appendix
A-14
Sample Covariance File
%filename: data.cov, ftype: covariance V1.0, 06Nov1993, 11:51
%Secondorder mechanical system
%
1.18e2 1.62e3 1.93e3
1.24e2 1.456e3
2.19e2
%End of covariance file
Description of the Available Measurement Data Files
A-15
The data files described here contain public domain data measured mostly by
members of Department ELEC, Vrije Universiteit Brussel, Belgium. The data
are public domain; they can be freely used by anyone.
aluplate
The measured mechanical test structure consisted of two rectangular
aluminum plates glued together. The plates were hung on two nylon threads,
to assure freedom for movement. The shaker was glued by beeswax to the plate
surface. The excitation force was measured as input signal, and the
acceleration at another point of the plate surface was measured as output
signal. The excitation signal was a multisine composed of 149 frequencies in
the band between 9.76 Hz and 298.8 Hz.
The input and output errors n
x
and n
y
are approximated by white noise sources:
Xk
= 5.6 V and
Yk
= 56 V (k = 1149). One experiment is available on the
file aluplate.fbn.
A Appendix
A-16
bandpass, bp1, bandp8k
Passive sixth order octave bandpass filters were measured. There are three
similar data files.
1 In the case of the file bandpass.fbn the DUT was a passive sixth order
octave bandpass filter with a center frequency of 500 Hz. The filter was
excited by a multisine consisting of 16 frequencies in the frequency band
5...2048.828 Hz, that is, 244.1 Hz ... 976.6 Hz. The time signals were mea-
sured with an 8-bit, 1024-point digitizer. The quantization noise dominates:
Xk
=
Yk
= 0.019 V. The results of 25 experiments are available in the file.
2 bp1.fbn is a synchronized version of bandpass.fbn (having used varanal).
3 In the case of the file bandp8k.fbn, the DUT was a passive sixth order octave
bandpass filter with a center frequency of 8 kHz. The filter was excited by a
multisine consisting of 29 frequencies in the frequency band 29...57195.3
Hz, that is, 5.664 kHz ... 11.132 kHz. The time signals were measured with
a 12-bit, 2048-point digitizer. The environmental noise and the DUT noise
A-17
dominate:
Xk
=
Yk
= 1.7 V. The results of 30 experiments are available in
the file.
cable1-6
Six reflectograms of 50 coaxial cables (RG 58 C/U), terminated by different
types of loads, have been recorded using a reflectometer (internal impedance:
75 ). The excitation impulse and the first reflection are cut out of the time
domain reflectogram (which means multiplication with a rectangular window)
and stored in file. The reflections usually had a much smaller amplitude,
therefore they were amplified before recording
this is the explanation of the

reflections in some cases apparently larger than the input pulses. From
preliminary noise measurements it was found that the power spectra of the
disturbing noise sources are flat within the frequency band of interest.
A Appendix
A-18
Description of the data files (T
s
= sampling rate, z = length of the cable, t
w
=
delay between the window of the excitation and reflection):
cable1.tbn: open ended, T
s
= 4 ns, z = 101.80 m, t
w
= 259 T
s
,
x
= 0.289 and
y
= 0.697 .
A-19
cable2.tbn: open ended, T
s
= 15.5 ns, z = 506.14 m, t
w
= 333 T
s
,
x
= 0.289
and
y
= 0.289 .
A Appendix
A-20
cable3.tbn: 75 terminated, T
s
= 5.5 ns, z = 102.55 m, t
w
= 188 T
s
,
x
=
0.289 and
y
= 0.289 .
A-21
cable4.tbn: 28 terminated, T
s
= 4 ns, z = 102.64 m, t
w
= 260 T
s
,
x
= 0.289
and
y
= 0.734 .
A Appendix
A-22
cable5.tbn short ended, T
s
= 5 ns, z = 102.55 m, t
w
= 205 T
s
,
x
= 0.289 and
y
= 0.289 .
A-23
cable6.tbn: 1.2 nF terminated, T
s
= 5 ns, z = 102.63 m, t
w
= 210 T
s
,
x
=
0.289 and
y
= 0.899 .
A Appendix
A-24
crankcas
A crankcase was excited with a multisine composed of 308 frequencies in the
band between 99.609 Hz and 399.414 Hz. From prior measurements it was
found that
Xk
= 70 V and
Yk
= 210 V. Thirteen independent measurements
are available on the file crankcas.fbn.
A-25
emachine
The stator impedance of an synchronous electrical machine was measured, in
standstill frequency response measurement arrangement. This means that the
rotor was kept mechanically at the same position during the whole
measurement. The data file corresponds to the so-called d-axis characteristics
of the machine. A current waveform was applied as excitation signal to the
stator, and the current and the voltage were measured in the stator windings.
The measurements were carried out with three signals: a low-frequency
multisine in the band (10 mHz, 990 mHz), a medium frequency multisine in the
band (1 Hz, 99 Hz), and a high-frequency multisine in the band (20 Hz, 240 Hz).
A total of 88 frequencies have been picked out of the 3 frequency bands, and are
given in the file emachine.fbn. The input and output noise variances are
available on the file emachine.vbn.
A Appendix
A-26
A-27
glassfib
The measurement arrangement was practically the same as for the aluminum
plates (see aluplate on page -15), except that the measured object was a
rectangular glass fiber plate. It was excited with a multisine composed of 200
frequencies in the band between 195.3 Hz and 389.6 Hz. The input and output
errors n
x

and n
y
are approximated by white noise sources:
Xk
= 31 V and
Yk
= 96 V. One experiment is available on the file glassfib.fbn.
A Appendix
A-28
inpchan
The flat Schroeder multisine used in the measurements of the data acquisition
channel (including an anti-aliasing filter of approximately 20 kHz cutoff
frequency) consisted of 63 components in the band between 400 Hz and 25.2
kHz (f = 400 Hz). The first 49 components of the excitation signal were
selected for the identification of the transfer function. The standard deviations
of the real and imaginary parts of the input and output spectra were found to
be 3.1 V and 31 V, respectively. The results of one experiment are available
on the file inpchan.fbn.
A-29
lowpaswb, lowpass1, lowpass2
The DUT is an active analog sixth order Chebyshev lowpass filter with a cutoff
frequency of 1 kHz. It was measured three times.
The first measurement was done in a relatively wide band, 48.8 Hz - 3 kHz,
with a flat Schroeder multisine with components [1:61]48.83 Hz, peak value
2 V. The measured input and output amplitudes of one experiment are
available in the file lowpaswb.fbn.
The second and third measurements were made with a flat Schroeder
multisine containing 20 components from 97.65625 Hz to 1025.391 Hz, that is,
2...2048.828 kHz.
A Appendix
A-30
In the second measurement the peak value of the multisine was 2 V, and the
noise power spectra were flat in the frequency band of interest:
Xk
= 22.3 V
and
Yk
= 25.6 V. 30 experiments are available on the file lowpass1.fbn.
A-31
In the third measurement the peak value of the multisine was reduced to 150
mV in this case. The noise power spectra were frequency dependent. 120
experiments are available on the file lowpass2.fbn. The variances of the
disturbing noise sources are available on the file lowpass2.vbn.
A Appendix
A-32
A-33
robotarm
The behavior of a flexible robot arm was measured by applying controlled
torque to the vertical axis at one end, and measuring the tangential
acceleration of the other end of the arm. The excitation signal was a multisine,
generated with frequency components at [1:2:199]*f, with f = 500/4096
0.125 Hz, that is, the frequency range was about 0.125 Hz - 25 Hz. The
originally flat multisine was distorted by the nonlinear behavior of the
actuator. The odd harmonic frequencies provided that components produced by
a squaring nonlinearity would not disturb the identification. The input and
output signals were sampled with sampling frequency 500 Hz. Sampling was
synchronized to the excitation signal so that 4096 samples were taken from
each period. The data records contain 40960 points, that is, 10 periods were
measured.
In the data file robotarm.mat, the following variables are provided:
xt = scaled input time record, 409601
yt = scaled output time record, 409601
fs = sampling frequency, 500 (Hz)
ascale = scaling factor of the time records
N = number of points in a period (4096)
freqind = index numbers of sine waves in the DFT of a period ([1:2:199]')
Amplitude scaling was necessary in order to have 16-bit integer numbers in the
time records, thus reducing the storage requirements by a factor of 4, to about
165 kBytes.
A Appendix
A-34
This measurement was made at the Department of Mechanical Engineering,
Catholic University of Leuven (KUL), in cooperation with Department ELEC,
Vrije Universiteit Brussel (VUB), as a part of the Belgian program
Interuniversity Attraction Poles (IUAP50: Robotics and Industrial
Automation), initiated by the Belgian State, Prime Minister's Office, Science
Policy Programming. These data, just like the other ones described in this
Appendix, belong to the public domain and can be freely used by anyone.
I-1
Index
A
Akaike criterion 1-35
Allpass design 2-13, 3-18
Aluminum plates A-15
aluplate.fbn A-15
Andrsfai, B. 3-72
Approximate covariance matrix 2-9, 3-16
Approximate mean model error 1-34, 3-16
Arbitrary waveform generator 2-20, 2-21, 3-64
Archive format 1-47
ASCII export 1-31, 3-42
ASCII file 1-47, 3-42, 3-51, A-2
Averaging 1-24, 1-28, 3-95
B
bandp8k.fbn A-16
Bandpass filter A-16
bandpass.fbn A-16
Binary excitation signal 2-22, 3-8, 3-53
Binary file 1-48, A-4
bp1.fbn A-16
C
cable1.tbn A-18
cable2.tbn A-19
cable3.tbn A-20
cable4.tbn A-21
cable5.tbn A-22
cable6.tbn A-23
Calculation speed 2-15, 2-18
Calibration of measurement channels 2-10
Chi-squared cost function 2-5
Choice of proper delay 1-32
Clipping 3-57
Complex geometric mean 2-14, 1-30, 3-48
Complex residuals 2-14, 3-80
Condition number 2-16, 2-18, 3-16
Confidence ellipse 1-31, 3-73
Confidence interval 1-30, 3-78
Constraint 2-4, 2-12
Correlation between input and output 2-6, 3-95
Covariance file 3-29, A-8
Covariance matrix 2-9, 3-16, 3-29, A-8
Covariance of the estimate 2-7
Covariance vector 3-29, A-8
Cramr-Rao bound 2-8, 2-18, 3-16, 3-66, 3-97
crankcas.fbn A-24
Crest factor 2-21, 3-8, 3-57
Crest factor minimization 2-21, 3-57
Cross validation 1-40
D
D/A converter 2-21, 3-57, 3-64
Data exchange 1-47
Data file
aluplate A-15
bandpass A-16
cable A-17
crankcas A-24
emachine A-25
glassfib A-26
inpchan A-28
lowpass A-29
robotarm A-33
Data formats 1-47, A-2
Default answer 3-100
Delay 2-2, 1-32, 1-41, 3-95
Delbaen, F. 2-20, 3-70
Demonstrations 1-2, 3-46
DFT 2-23, 3-49, 3-94
Index
I-2
dibs 2-23, 3-8
dibsimpr 2-23, 3-8
Differentiator 2-13
Discrete interval binary sequence 2-22, 3-8
Dispersion function 3-67
Dynamic Signal Analyzer 2-14
E
ELiS 2-6
elis 3-12
elis2tha 1-45, 3-27
elisfcnv 1-48, 3-26
elisqa 3-12
elrpf2v 3-12
elrpv2f 3-12
emachine.fbn A-25
emachine.vbn A-25
Empirical standard deviation 3-91
Equation error 1-36
Equation error method 3-14, 3-19
Errors-in-variables model 2-3
Estimator for Linear Systems 2-6
Excitation signal 2-20, 3-8, 3-53, 3-57, 3-64, 3-66
expcov 3-29
expfou 3-31
exppar 3-34
exptim 3-37
expvar 3-40
expvect 1-31, 1-48, 3-42
Eykhoff, P. 1-77
F
fdcovpzp 1-40, 3-43
fddemos 1-2
fdiddemo 1-2, 3-46
Feedback loop 2-7
FFT 2-23, 1-27, 3-93
File
ASCII 1-47, 3-51, A-2
binary 1-48, A-4
File conversion 3-26
File handling 1-47
Fixed parameters 2-12, 2-17, 3-18, 3-66
Fixing some poles/zeros 2-12
Fixing transfer function value 2-13
Flannery, B. P. 1-77
fnamanal 3-47
Fourier file A-5
Fourier vector A-5
Frequency Domain Formulation 2-2
Frequency grid 2-20, 3-49
F-test 1-35
G
Geometric mean 1-30, 3-48
glassfib.fbn A-26
gmean 2-14, 1-28, 1-30, 3-48
Godfrey, K. R. 2-20, 2-23, 1-77, 3-11, 3-54, 3-63, 3-70
Gold, B. 1-27, 1-45
Goodwin, G. C. 1-77
Guillaume, P. 2-7, 1-28, 3-45, 3-76, 3-88
H
Hessian 2-18
Hungarian method 3-71, 3-72
Index
I-3
I
Identification session 1-49
impcov 3-29
impfou 3-31
Imposing constraints 2-12
imppar 3-34
imptim 3-37
impvar 3-40
Information matrix 2-20, 3-66
Initial LS step 2-6, 2-11
Initial values 2-6, 2-11, 2-18, 3-19
inpchan.fbn A-28
Input and output noise 2-3
Input-output correlation 2-6, 3-95
Integrator 2-13
invfreqs 3-19
invfreqz 3-19
Iteration algorithm 3-14
Levenberg-Marquardt 2-6, 2-15, 2-18, 2-19, 3-14,
3-17, 3-18, 3-19
Newton-Gauss 2-6, 2-19, 3-14, 3-18
Newton-Raphson 2-18, 2-19, 3-14, 3-18
J
Jacobian 2-18
K
Key features of ELiS 2-10
Kollr, I. 2-11, 1-28, 3-81, 3-99
Korn, G. A. 1-25
Korn, T. M. 1-25
Krol, R. G. 2-23, 3-11
Kuhn, H. W. 3-72
L
Leakage 1-4, 2-11, 1-26
Least squares 2-4, 2-6, 2-7, 1-42, 3-12, 3-14, 3-19
Levenberg-Marquardt algorithm 2-6, 2-15, 2-18,
2-19, 3-14, 3-17, 3-18, 3-19
lin2qlog 2-20, 3-49
Linear phase 2-13
Ljung, L. 1-4, 1-77
loadasc 3-51
loadvar 3-52
log2qlog 2-20, 3-49
Logarithmic frequency set 3-49
Log-likelihood function 2-4
lowpass1.fbn A-30
lowpass2.fbn A-31
lowpass2.vbn A-31
lowpaswb.fbn A-29
LS, see Least squares
M
Machine precision 2-17
Magnitude plot 3-77
Magnitude uncertainty 3-77, 3-89
Maximum length binary sequence 2-23, 3-53
Mean model error 1-34, 3-16
mlbs 2-23, 3-53
Model conversion 1-40, 1-41, 3-27
Model error 1-34, 3-16
Model validation 1-32, 1-69
modifyfv 2-13, 2-15, 2-19, 1-29, 3-55
msinclip 2-21, 3-57
msinprep 2-22, 3-64
Multisine 2-21, 3-57, 3-64
Index
I-4
N
Nearest neighbors 3-71
Network analyzer 2-14
Newton-Gauss algorithm 2-6, 2-19, 3-14, 3-18
Newton-Raphson algorithm 2-18, 2-19, 3-14, 3-18
Noise analysis 1-28
Nonlinear weighted least squares 2-4
Nonparametric transfer function estimate 3-91
Norton, J. P. 1-77
Nuisance parameters 1-36
Numerical stability 2-15
O
optexcit 2-20, 3-66
Optimum power spectrum 2-20, 3-66
Overmodeling 2-17, 2-19, 1-33
P
Paehlike, K.-D. 2-23, 3-11
pairs 3-71
Parameter file 3-34, A-7
Parameter fixing, see Fixed parameters
Parameter vector 3-34, A-7
Partial transfer function 2-13, 1-29, 3-55
Payne, R. L. 1-77
Phase plot 3-77
Phase uncertainty 3-77, 3-89
Phase wrapping 3-48
Pintelon, R. 2-7, 2-11, 2-20, 1-28, 1-33, 1-45, 1-77,
3-25, 3-45, 3-46, 3-48, 3-70, 3-76, 3-88
Plot confidence ellipses 3-73
Plot confidence intervals 3-77
Plot magnitudes 3-77
Plot phases 3-77
Plot poles and zeros 3-73
plotelpz 1-31, 1-35, 3-73
ploteltf 1-30, 3-77
Pole-zero model 1-40, 3-43
Pole-zero plot 3-73
Pole-zero uncertainty 3-43, 3-73, 3-86
Post-Measurement Synchronization 1-28
PRBS 2-23, 3-53
Prediction error method 1-44
Preprocessing in the frequency domain 1-28
Preprocessing in the time domain 1-24
Preprocessing of data 1-24
Presentation of results 1-30
Press, W. H. 1-77
Pseudo-random binary sequence 2-23, 3-53
Q
Quasi-logarithmic frequency vector 2-20, 3-49
R
Rabiner, L. R. 1-27, 1-45
Rake, H. 2-23, 3-11
rdueelis 2-14, 1-36, 3-80
Reflectogram A-17
Renneboog, J. 2-21, 3-48, 3-63
Report file 1-47, 3-22, A-9
Residuals 1-36
robotarm.mat A-33
robotarm.mat? 1-49
Rolain, Y. 2-7
Run parameter file 3-12
Run parameter vector 3-12
S
savevar 3-52
Index
I-5
Scaling frequency 2-17, 3-18
Schoukens, J. 2-11, 2-20, 2-21, 1-28, 1-33, 1-45, 1-77,
3-25, 3-45, 3-46, 3-48, 3-63, 3-70, 3-76, 3-88
Signal Analyzer 2-14
Signal generator 2-20, 2-21, 3-64
Signal Processing Toolbox 1-39, 3-14, 3-19, 3-101
Signal-to-Noise Ratio 1-54
simfou 1-38, 3-82
simtime 1-38, 1-44, 3-84
Simulation 1-32, 1-37, 3-82, 3-84
Singular value decomposition 2-18, 2-19, 3-14, 3-18
Sderstrm, T. 1-35, 1-77
Special transfer function forms 2-13
Stability of system models 1-32
Start values 2-6, 2-11, 2-18, 3-19
stdpz 1-31, 1-35, 3-86
stdtf 1-30, 1-34, 1-36, 3-89
stdtfm 2-14, 1-30, 3-91
Stoica, P. 1-35, 1-77
Stop criterion 2-19, 3-14, 3-19
Subband fit 2-15
svd, see Singular value decomposition
Synchronization 1-28, 3-95
System Identification Toolbox 1-4, 1-41, 3-27
T
Teukolsky, S. A. 1-77
tha2elis 1-46, 3-27
Theta-format 1-42
tim2fou 3-93
Time domain data file 3-37
Time domain data vector 3-37
Time file A-4
Time vector A-4
Transfer function plot 3-77
Transform to frequency domain 1-26
U
Uncertainty ellipse 1-31, 3-73
Uncertainty of poles and zeros 3-43, 3-86
Uncertainty of transfer function 3-89
Undermodeling 2-17, 1-33
V
van den Bos, A. 2-23, 3-11
van der Ouderaa, E. 2-21, 3-63
varanal 2-14, 1-28, 3-95
Variance analysis 1-57, 3-95
Variance file 3-40, A-6
Variance vector 3-40, A-6
Verbeyst, F. 2-7
Vetterling, W. T. 1-77
W
Weighted least squares 2-4, 2-5, 3-14
Wide-band model fitting 2-15
WLS, see Weighted least squares
Y
yesinput 3-100
yulewalk 3-101
ywalk 1-39, 3-101
Z
Zero-order hold 2-21, 3-57, 3-64
Zero-pole model, see Pole-zero model
Index
I-6

Fdident

Uploaded by

Copyright:

Available Formats

Fdident

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Fdident

Uploaded by

Copyright:

Available Formats

Computation

is the corresponding abscissa of the standard

this is the explanation of the

You might also like