Computational Aids

in Aeroservoelastic Analysis
Using MATLABTM

Changho Nam • Youdan Kim • Terrence A. Weisshaar

Chapter 1
Introduction to MATLAB

1.1 General MATLAB Information

MATLAB is a high-performance language for technical computing, and is an

interactive system whose basic data element is a matrix. These days, MATLAB is
widely used for (i) mathematics and computation, (ii) algorithm development, (iii)
modeling, simulation, and prototyping, data analysis, exploration, and visualization,
(iv) scientific and engineering graphics, and application development including
graphic user interface building.

These days, many specialized toolboxes have been developed: Control Systems,
Signal Processing, Optimization, Neural Network, System Identification, and so on.
Those toolboxes provide MATLAB very powerful tools for various and practical
applications.

The purpose of this chapter is to introduce a number of MATLAB functions that are
widely used for dealing with engineering problems. Because there exist a lot of
MATLAB functions, only selected MATLAB functions will be provided in the
following sections. Especially, some basic MATLAB functions used in this book will
be briefly summarized.

© 2001 by C. Nam, Y. Kim 1

 1.2 Introduction to MATLAB

When you get into the MATLAB, MATLAB displays a prompt ( >> ) indicating that
MATLAB is ready to receive a command. When you type help command, you can
obtain on-screen information about MATLAB commands which are useful for users.
And when you type quit at the MATLAB prompt, you can exit MATLAB anytime. A
brief summary of some of MATLAB commands is provided in the following sections
for readers who are not familiar with MATLAB. Especially, some MATALB
commands for matrix operation and solution as well as plotting commands for
visualization of the numerical results are explained.

1.2.a Vectors and Matrices

Row vectors are entered by typing a list of numbers in brackets as follows:

>> x = [ 0 3 2]

Note that elements are separated using either spaces or commas. To create column
vectors, transpose the vector using the apostrophe (’) key. When a complex vector is
transposed, the conjugate transpose of the vector will be obtained. To obtain the
unconjugated transpose, type a dot before the transpose symbol (.’).

Matrices are entered row by row, and each row is separated by semicolon. For
example, to enter 3-by-3 matrix, type the following command.

>> A = [ 0 3 2; 5 1 2; 9 -6 7]

Note that a semicolon is used to indicate the end of each row.

The workspace is the area of memory used by the MATLAB. Two commands, who
and whos, provide the information of current variables in the workspace. The
command who lists all the variables in memory, while whos displays the variables,
sizes and storage information. The command size(A) returns the number of rows and

© 2001 by C. Nam, Y. Kim 2

columns of a matrix A, while the command length(x) returns the dimension of a
vector x.

The element in row i and column j of a matrix A is denoted by A(i,j). For example,
A(2,3) is the number in the second row and third column. The colon operator (:)
represents all the rows or all the columns of a matrix, and therefore the colon operator
can be used to define vectors form a matrix. Consider the following expressions.

>> x=A(3, :)
>> y=A(:, 2)

By typing the above commands, x gives a row vector of third row of matrix A, and y
gives a column vector of second column of matrix A, respectively. Also, A(i:j, k:l)
returns rows i through j and columns k through l of matrix A, and therefore the
following expression yields a two-by-two submatrix of A.

>> A(1:2, 1:2)

When output variable is not specified, MATLAB uses the variable ans to store the
results of a calculation, and ans keeps the variable value of most recent answer

Rows and columns can be deleted from a matrix using just a pair of square brackets.
To delete the first column of matrix A, type the following command.

>> A(:,1) = [ ]

Concatenation is the process of joining small matrices to make bigger ones. For the
appropriately dimensioned matrices A, B, C, and D, the following command can be
used to define a new matrix.

>> E = [A B; C D]

The command clear clears specified variables from the workspace. For example, to
clear variables x and y from the workspace, type the following command.

© 2001 by C. Nam, Y. Kim 3

 >> clear x y

Typing clear by itself clears the entire workspace.

1.2.b Basic Scalar and Matrix Computations

Conventional decimal notation is used in MATLAB, and scientific notation an also be

used by using the letter e to specify a power-of-ten scale factor. Several predefined
functions provide special and useful constant values as follows:

pi The value of π , (= 3.141592 …)

i or j Imaginary unit, (= − 1 )
eps Floating-point relative precision for the computer being used
Inf Infinity
NaN Not-a-Number

MATLAB provides several functions that generate special matrices. The zeros(m,n)
function generates an m-by-n matrix containing all zeros, the ones(m,n) function
generates an m-by-n matrix containing all ones, the eye(n) function generates n-by-n
identity matrix, respectively. Also, rand(m,n) function generates an m-by-n matrix
with uniformly distributed random elements.

MATLAB supports basic algebraic operations: addition (+), subtraction (−),

multiplication (∗), division (/), left division (\), and exponentiation or power (^).
MATLAB also supports the elementary math functions including standard
trigonometric functions, hyperbolic functions, and computing the absolute value and
the square root. For example,

>> a1 = sin(x)
>> a2 = cosh(x)
>> a3 = asin(x)
>> a4 = abs(x)
>> a5 = sqrt(x)

© 2001 by C. Nam, Y. Kim 4

 >> a6 = exp(x)
>> a7 = log(x)

The above commands perform the well-known math functions, and yield the sine
value, the cosine hyperbolic value, the arcsine or inverse sine value, the absolute
value, the square root value, the exponential value, and the natural logarithm value of
the variable x, respectively.

Complex variable x can be defined by the following command.

>> x = 1 + 2 ∗ i

MATLAB supports several functions for treating complex numbers.

real(x) This function computes the real part of x

imag(x) This function computes the imaginary part of x
abs(x) This function computes the magnitude of x
angle(x) This function computes the phase angle of x
conj(x) This function computes the complex conjugate of x

Matrix addition, subtraction, and multiplication are performed according to linear

algebra rules. For example,

>> C = A + B
>> C = A − B
>> C = A ∗ B

The above operations define a new matrix C as sum of two matrices, subtraction
matrix B from matrix A, and matrix multiplication, respectively.

Other useful matrix commands are listed as follows with a brief description:

sum(A) This function performs the sum over each column

A' This function performs the conjugate transpose of matrix A

© 2001 by C. Nam, Y. Kim 5

 diag(A) This function picks off the diagonal of A
det(A) This function produces the determinant of matrix A
trace(A) This function produces the sum of diagonal elements of matrix A
rank(A) This function produces the rank of matrix

Matrix inversion is defined by

>> A\B
>> A/B
>> inv(A)

In the above expressions, A\B is equivalent to inv(A)∗B, and A/B is equivalent to

B∗inv(A), respectively.

Other useful matrix functions such as eig (eigenvalue), cond (condition number),
expm (matrix exponential), and norm are discussed in detail in the subsequent section.

1.2.c Format Commands and File Manipulation

All computations in MATLAB are preformed in double precision. However, the

display format of data can be modified using the format command. For example, the
displayed output can be controlled using the following commands.

>> format short

>> format long
>> format short e
>> format long e

The command format long displays the output in 15 digit scaled fixed point, while the
command format long e displays in scientific notation (15 digit floating point). Note
that the chosen format remains in effect until changed.

All variables in MATLAB workspace can be saved in the binary MATLAB format by
using the following save command.

© 2001 by C. Nam, Y. Kim 6

 >> save filename

If no extension is specified, MATALB adds the extension .MAT to the filename. Only
selected variables can be saved by performing a partial save.

>> save filename x y

where variables x and y are saved in a file name filename.mat. Variables can be saved
in a standard text format.

>> save data.dat x y /ascii /double

The above command save variables x and y in a file name data.dat in 16-digit ASCII
format. Without double option data is saved in 8-dgit ASCII format.

To restore the variables in the file that was generated by save command, use the load
command.

>> load filename

Only specified variables can be loaded by using the following command.

>> load filename x y

The command what shows the MATLAB files and the MAT data files under the
current directory, and the command dir shows all the files under the current directory.
The command delete deletes files, the command cd changes directory, and the
command type lists the contents of the file, respectively.

The command diary is used to create all the input commands and displayed outputs in
a MATLAB session. To create the MATLAB session in a file with a different name,
use

© 2001 by C. Nam, Y. Kim 7

 >> diary filename

To close diary file, use

>> diary off

Now, all the command and answer lines that are displayed on the screen are saved
under a given name in ASCII format, and the file can be edited later.

1.2.d Matrix Functions

MATLAB supports a number of matrix functions that manipulates the contents of a

matrix, which is one of the features of the MATLAB.

Basic Matrix Functions

Matrix exponential, logarithm, and square root can be obtained by the following
commands.

>> expm(A)
>> logm(A)
>> sqrtm(A)

Note that exp(A) (that is without the m) gives exponential of matrix A element-by-
element.

Norms
The norm of a matrix is a scalar that gives some measure of the size of the matrix.
Several different definitions are commonly used.

norm(A) Largest singular value of A, max(svd(A))

norm(A,2) Same as norm(A)
norm(A,1) 1-norm of A, the largest column sum, max(sum(abs((A))))
norm(A,inf) Infinity norm of A, the largest row sum, max(sum(abs((A')))).
norm(A,'fro') Frobenius norm, sqrt(sum(diag(A'*A))).

© 2001 by C. Nam, Y. Kim 8

Note that the command norm is also used to compute the norm of vector. For a vector
x, the command norm(x) produces the Euclidean norm of vector x.

Condition Number
The command cond returns the condition number (the ratio of the largest singular
value of matrix A to the smallest of matrix).

>> cond(A)

The condition number represents a quantitative measure of ill-

conditioning(singlularity) of a matrix. Large condition numbers indicate a nearly
singular matrix, and therefore the condition number of a singular matrix is infinity. If
you try to compute the inverse of singular or nearly singular matrix, then you will get
a warning message telling you that the computed results may be inaccurate.

Solution of Linear Equations

Consider a set of simultaneous linear algebraic equations in matrix notation.

y=Ax

where x R n and y R m . The above problem has three possibilities: (i) exactly one
solution, (ii) an infinity of exact solutions, and (iii) no exact solution. For a
determined case [m=n=rank(A)], the solution is obtained by the following command.

>> x = inv(A) ∗ y
or
>> x = A \ y

The second and third cases are under-determined case and over-determined case,
respectively, and the following command yields the minimum norm solution and least
square solution, respectively.

>> x = pinv(A) ∗ y

© 2001 by C. Nam, Y. Kim 9

The Eigenvalue Problem
Consider an eigenvalue problem of a matrix

Ax=λx

where λ is the eigenvalue of matrix A, and x is the associated eigenvector. The

eigenvalue problem can be solved using the eig command.

>> lambda = eig(A)

>> [V, D] = eig(A)

The command eig produces a vector lambda including the eigenvalues of matrix A, a
diagonal matrix D of eigenvalues and a full matrix V whose columns are the
corresponding eigenvectors, respectively.

For the square matrices A and B, the generalized eigenvalue problem is defined as
follows.

Ax=λBx

Then, the following command returns a diagonal matrix D of generalized eigenvalues

and a square matrix X whose columns are corresponding eigenvectors.

>> [V, D] = eig(A, B)

Singluar Value Decomposition

Any rectangular matrix can be decomposed as follows.

A = U Σ VH

where supscript H denotes the Hermitian or conjugate transpose of matrix, matrices U

and V are unitary matrices, and the matrix Σ is a diagonal matrix containing as its

© 2001 by C. Nam, Y. Kim 10

elements the real (non-negative) singular value of matrix A. The command svd is used
for singular value decomposition of a matrix.

>> [U, S, V] = svd(A)

>> sigma = svd(A)

Note that the diagonal matrix S and the vector sigma include the singular values in
decreasing order.

MATLAB supports various matrix decomposition methods such as QR

decomposition (command qr), Cholesky decomposition (command chol), LU
decomposition (command lu), and so on, which will not be discussed in this section.
More information can be found in MATLAB manual or using the on-line help
commands.

1.2.e Data analysis

MATLAB uses column-oriented analysis for multivariable statistical data, and

MATLAB has data analysis functions for a vector or a column of a matrix.

The following commands can be used to determine maximum and minimum, and
their locations.

>> y = max(x)
>> y = min(x)
>> [y, k] = max(x)
>> [y, k] = min(x)

If an input variable is a vector, then the function returns the largest or smallest value
of the input vector. However, if an input variable is a matrix, then the function returns
a row vector containing the maximum or minimum element from each column. And k
is an index of the maximum value in the vector.

To sort a vector in the ascending order, the command sort can be used.

© 2001 by C. Nam, Y. Kim 11

 >> [y, i] = sort(x)

The sorted values are returned to the vector y and the reordered indices are stored in
the vector i. When x is complex, the elements are sorted by the magnitude of x.

The mean value and the standard deviation of a vector can be calculated by the
following commands.

>> mean(x)
>> std(x)

1.2.f Polynomials

In MATLAB, polynomials are represented as vectors containing the polynomial

coefficients in descending order. For example, x 3 + 3 x 2 + 3 x + 1 can be represented as
follows:

>> y = [1 3 3 1]

The roots of the polynomial equation x 3 + 3 x 2 + 3 x + 1 = 0 are solved using the

command roots.

>> y = [1 3 3 1]; r=roots(y)

The roots of the polynomial equation may be reassembled back into the original
polynomial equation with the command poly.

>> poly(r)

The command poly returns the characteristic equation of matrix, when the input
variable is a matrix.

>> poly(A)

© 2001 by C. Nam, Y. Kim 12

A polynomial can be evaluated at a point using the command polyval, and the
command polyval evaluates polynomials in the matrix sense.

>> polyval(y, val)

>> polyvalm(y, A)

where y is a polynomial and val is the point at which the polynomial is to be

evaluated.

The product of the polynomials, the convolution of the coefficients, can be performed
by the command conv, while a polynomial can be decomposed by the deconvolution
command.

>> conv(x, y)
>> [m, r] = deconv(a, b)

where m is the quotient and r is the remainder, that is a ( s ) = b( s )m( s ) + r ( s ) .

The partial fraction expansion can be performed by the command residue.

>> [r, p, k] = residue(num, den)

The above command finds the residues, poles and direct term of a partial fraction
expansion of the ratio of two polynomials num(s)/den(s). If there are no multiple
roots, the residues are returned in the column vector r, the pole locations in column
vector p, and the direct terms in row vector k. The direct term coefficient vector is
empty if the function num(s)/den(s) is strictly proper, or length(num) < length(den).
The residue command can also be used to form the (num, den) polynomials from its
partial fraction expansion:

>> [num, den] = residue(r, p, k)

© 2001 by C. Nam, Y. Kim 13

To generate a polynomial curve that fits a given set of data, the command polyfit can
be used.

>> p = polyfit(x, y, n)

where x and y are vectors of the given data set in (x, y) form. The above command
finds the coefficients of a polynomial p of degree n that fits the data in a least squares
sense, that is the polynomial is obtained by minimizing the error between the
polynomial and the given data set.

1.2.g Plotting and Graphics

MATLAB provides a powerful graphic tools that display vectors and matrices in a
variety of formats: x-y plots, polar plots, bar graphs, contour plots, and 3-D plots. In
this section, only a few of graphics functions will be described.

The command plot plots the vector y versus the vector x.

>> plot(x,y)

Note that the vectors must have the same length. To display the graphs in a
logarithmic scale, following commands are used.

semilogx(x, y) x-axis (logarithmic), y-axis (linear)

semilogy(x, y) x-axis (linear), y-axis (logarithmic)
loglog(x, y) both axes (logarithmic)

To open a new figure window, the command figure is used. The command plot erases
the previous plots and reset all axis properties before drawing new plots. The
command hold on holds the current plot and all axis properties so that subsequent
graphing commands add to the existing graph, and the command hold off returns to
the default mode. The sample plotting command is shown below.

>> t = 0 : pi/100 : 2∗pi;

© 2001 by C. Nam, Y. Kim 14

 >> y = sin(t);
>> plot(t, y)
>>
>> figure
>> y2 = sin(t - 0.25);
>> plot(t, y2)
>> hold on
>> plot(t, y+y2)

In the above example, t=0:pi/100:2∗pi represents a vector t which starts from 0 and
ends at (2∗pi) with an interval of (pi/100).

Multiple curves can be plotted against the same axis.

>> plot(x1, y1, x2, y2, x3, y3)

The variables (x1, y1), (x2, y2), and so on, are pairs of vectors, and therefore the
above command plots y1 versus x1, y2 versus x2, and y3 versus x3. In this case, x1,
x2, and x3 can have different length. However, xi and yi must have the same length
(i.e., same number of data points).

The command plot generates a line plot that connects the points with line segments.
Different line type can be used by line option, and also a point plot can be plotted
instead of line plot. For example, the following command plots a red dashed line with
a x-mark at each data point;

>> plot(x,y,'rx--:')

And the following command plots green circle at each data point but does not draw
any line.

>> plot(x,y,'go')

© 2001 by C. Nam, Y. Kim 15

Various line types, plot symbols and colors may be obtained with the following
command

>> plot(x1, y1, 's1', x2, y2, 's2’, ...)

where s1 and s2 are characters string made from one element from any or all the
following options.

Colors: indicator color indicator color

r red g green
b blue w white
y yellow k black
m magenta c cyan

Lines: indicator line type indicator line type

- solid : dotted
-. dashdot -- dashed

Points: indicator point type indicator point type

. point o circle
x x-mark + plus
v triangle (down) d diamond
^ triangle (up) s square
< triangle (left) * star

After plotting a graph on the screen, grid lines, title, x-axis and y-axis label may be
inserted. The following commands are used for the plot output.

>> grid on
>> title(‘Title of graph’)
>> xlabel(‘Time (sec)’)
>> ylabel(‘ydata’)

© 2001 by C. Nam, Y. Kim 16

The command grid off takes grid off, and the command grid, by itself, toggles the grid
state. Also, a text can be written at the specified point on the graphic screen using the
following command.

>> text(x,y,‘text contents’)

where (x, y) are the coordinates using the axes from the current plot, and text is
written beginning at a point (x,y).

>> text(x,y,‘text contents’, ‘sc’)

The above command inserts the text on the graphic screen at the point (x,y) assuming
that the lower-left corner is (0,0) and the upper-right corner is (1,1).

The command gtext writes the text at the position on the graphic screen indicated by
the mouse or arrow keys. That is,

>> gtext(‘text contents’)

MATLAB automatically scales the axes to fit the data values. The following axis
command sets the axes to the specified setting.

>> axis( [ xmin xmax ymin ymax ] )

The command axis, by itself, freezes the current setting for subsequent plots, and a
second execution of the axis command resumes auto-scaling. There are more options
in the axis command as follows.

axis squre Makes x-axes and y-axes the same length

axis equal Makes the individual tick mark increments on
the x-axes and y-axes the same length
axis auto Returns the axis scaling to its default
axis on Turns on axis labeling and tick marks
axis off Turns off axis labeling and tick marks

© 2001 by C. Nam, Y. Kim 17

The above commands are useful when comparing curves from different plots.

The command subplot is used to split the graph window into subwindows, and
therefore multiple plots can be displayed in the same window.

>> subplot(m, n, p)

or

>> subplot(mnp)

The above command breaks the figure window into an m-by-n matrix of small
subplots and selects the p-th axes for the current plot. The subplots are numbered
along first the top row of the figure window, then the second row, and so on. The
sample plotting command for plotting data in four different subwindows is shown
below.

>> t = 0 : pi/100 : 2∗pi;

>> y1 = sin(t);
>> y2 = cos(t);
>> y3 = sinh(t);
>> y4 = cosh(t);
>> subplot(2, 2, 1); plot(t, y1); title(‘sin(x)’)
>> subplot(2, 2, 2); plot(t, y2); title(‘cos(x)’)
>> subplot(2, 2, 3); plot(t, y3); title(‘sinh(x)’)
>> subplot(2, 2, 4); plot(t, y4); title(‘cosh(x)’)

There are several commends for screen control. The commend pause is useful when
we display multiple graphic windows subsequently. The pause command causes a
procedure to stop and wait for the user to strike any key before continuing. And the
command pause(n) pauses for n seconds before continuing. The commend clf clears
the current figure window, and the command cla clears the current plot.

© 2001 by C. Nam, Y. Kim 18

MATLAB supports extensive graphic facilities. In this section, only some commends
for plotting 2-D (x-y) graphs are discussed. Contour plots, 3-D plots, mesh plots with
different color scales can be displayed in MATLAB, and detailed descriptions can be
found in the Help menu as well as the MATLAB user manual.

1.3 Programming in MATLAB

1.3.a M-Files

MATLAB can execute sequences of commands that are stored in files that are called
M-files. There are two kinds of M-files: Script M-files and function M-files that cam
be created by a text editor. Script M-files are useful for automating a series of
commands you want to execute many times. Script M-files operates on data in the
workspace, and do not accept input arguments nor return output arguments. Function
M-files are useful for adding more subroutines to MATLAB for specific applications,
and therefore accept input arguments and return output arguments. M-files can be
called from the command window, or from within another M-file.

Consider the example of script M-files that is saved under the name of test1.m.

% finding roots for second order algebraic equation

a = 1; b = 3; c=2;
D=b∗b – 4∗a∗ c;
root(1) = (-b +sqrt(D))/(2∗a);
root(2) = (-b -sqrt(D))/(2∗a);

The symbol % is used to add comments to the program, and everything to its right on
that line is ignored and is not executed. To perform the series of commands in the
test1.m file, the command test1 is used as follows

>> test1

The above script M-file can be written as a function M-file under the name of test2.m
as follows.

© 2001 by C. Nam, Y. Kim 19

 function root = test2(a,b,c)
% finding roots for second order algebraic equation
D=b∗b – 4∗a∗ c;
root(1) = (-b +sqrt(D))/(2∗a);
root(2) = (-b -sqrt(D))/(2∗a);

The work of the script M-file test1.m can also be done by the function M-file test2.m
as follows:

>> a = 1; b = 3; c=2;
>> root = test2(a,b,c);

MATLAB has five flow control statements (if statements, switch statements, for
loops, while loops, and break statements) that support some basic programming
structures that allow looping and conditioning commands along with relational and
logical operators. These features are briefly summarized in the subsequent sections.

1.3.b Relational and Logical Operators

MATLAB has six relational operators for comparing two scalrars (or matrices of
equal size, element by element) as follows:

Relational Operator Interpretation

< less than
<= less than or equal
> greater than
>= greater than or equal
== equal
~= not equal

For complex data, “==” and “~=” compare both the real and imaginary parts; other
operators compare only the real part.

© 2001 by C. Nam, Y. Kim 20

Two logical expressions can also be combined using the logical operators. There are
three logical operators available in MATLAB; they are

Logical Operator Symbol

and &
or |
not ~

1.3.c If Statements

The if statement evaluates a logical expression and executes a series of statements

when the expression is true. The optional elseif and else statements provide for the
execution of alternate groups of statements. The syntex is shown below:

If logical expression #1
statement group #1
elseif logical expression #2
statement group #2
else
statement group #3
end

If logical expression #1 is true, the first statement group is executed, if logical

expression #2 is true, then the second statement group is executed; otherwise, the
third statement group is executed. Note that the end statement is mandatory. The if
statement can also be nested; if statement can be included in the statement group.

Consider the following example.

>> n = 3.14;
>> if n > 0
>> disp('positive number')
>> elseif n < 0
>> disp('negative number')

© 2001 by C. Nam, Y. Kim 21

 >> else
>> disp('zero')
>> end

In the above example, the command disp displays the given string or variable.

Several functions are very useful with if statement: these functions are any, all,
isequal, isempty and so on, which will not be discussed in this book. The interested
readers may refer to the user manual or on-line help command.

1.3.d For Loops

The for statement is used to execute a series of statements a predetermined number of

times. The syntax is shown below:

for index = expression

statement group
end

Note that the index of the for loop must be a variable, and the index contains the last
value used after completing the for loop. There may be also multiple loops. The colon
operator is used to define the index expression using the following format:

For iter = ni : nd : nf

Where iter is a loop index which starts from ni and ends at nf with an increment of nd.
If nd is not defined, the default value of nd=1 is used. The value of iter would be ni,
then ni+nd, then ni+nd+nd, and so on until the final value of nf.

Consider the following example including multiple loops.

>> A=rand(3,4); B=rand(4,5);

>> C=zeros(3,5);
>> for i = 1: 3

© 2001 by C. Nam, Y. Kim 22

 >> for j = 1:5
>> for k = 1:4
>> C(i,j) = C(i,j) + A(i,k) ∗ B(k,j)
>> end
>> end
>> end

The break statement can be used to exit a loop before it has been completed. This
statement is useful if an error condition is detected within the loop. Therefore, break
statement is usually used with if statement, because if statement may be used to check
the escape condition. In nested loops, break exits from the innermost loop only.

1.3.e While Loops

The while loop executes a group of statements repeatedly as long as the controlling
expression is true. The syntax is shown below.

while expression
statement group
end

Consider the following example

>> n = 1;
>> while n <= 100
>> n = n + 1;
>> end

The break statement can also be used in the while loop to exit a loop before it has
been completed.

1.3.f Switch Statement

© 2001 by C. Nam, Y. Kim 23

The switch statement executes a group of statements based on the value of a variable
or expression. The syntax is shown below.

switch switch_expression (scalar or string)

case case_expression1
statement group
case {case_expression2, case_expression3}
statement group
otherwise
statement group
end

The statement group following the first case where the switch_expression matches the
case_expression is executed. When the case expression is a cell array (as in the
second case above), the case_expression matches if any of the elements of the cell
array match the switch expression. If none of the case expressions match the switch
expression then the otherwise case is executed (if it exists). Only one case is executed
and execution resumes with the statement after the end. Consider the following
example.

>> switch var

>> case 0
disp('0')
>> case {-1, +1}
>> disp('-1 or +1')
>> otherwise
>> disp('something else')
>> end

1.3.g Other Useful Commands

return
Normally functions return when the end of the function is reached. A return statement
can be used to force an early return.

© 2001 by C. Nam, Y. Kim 24

global
Each MATLAB function has its own local variables separated from those of other
functions, and from those of the base workspace and non-function scripts. By
declaring the variable as global in all the functions, several functions can share a
single copy of that variable. Any assignment to that variable, in any function, is
available to all the other functions declaring it global. Consider the following
example.

>> function [x,y] = projectile(v, gamma, t)

>> global gravity
>> x = v ∗ t ∗ cos(gamma);
>> y = v ∗ t ∗ sin(gamma) – 0.5 ∗ gravity * t^2;

1.3.h Basic Input-Output Functions

input
The value of variables can be entered through the keyboard using the input command
that displays a text then waits for input. The values entered by user are stored in the
variable specified. Consider the following example.

>> var1 = input(‘Enter values for var1’);

When the above command is executed, the text ‘Enter values for var1’ is displayed on
the screen. Then the user can enter an expression that specifies values for var1. For
character string input string input, use the following option.

>> R = input('What is your name' , 's')

disp
The disp command displays a string of text or numerical values on the screen.
Consider the following example.

>> disp(‘This is a MATLAB displacement command test.’)

© 2001 by C. Nam, Y. Kim 25

fprintf
The fprint command displays a text and values on the screen with specified format.
The syntax is shown below.

fprintf(format, variable)

where the format contains the text and format specifications. Within the text, %e and
%f are used for exponential notation or fixed point notation, respectively. If the string
\n appears in the format, the text statements before \n is displayed, and the rest of the
information will be displayed on the next line. A simple example is shown below.

>> temp=18;
>> fprint(‘The temperature is %f degrees C \n’, temp)

The fprintf command can also write formatted data to file using the file open and
close commands fopen and fclose which will not be discussed in this book. Again, the
interested readers may refer to the user manual or on-line help command.

1.4 Numerical Analysis using MATLAB

MATLAB supports lots of numerical techniques that are useful for solving many
engineering problems. MATLAB functions for solving ordinary differential equation,
nonlinear algebraic equation, and numerical integration will be briefly summarized in
this section.

1.4.a Ordinary Differential Equations

Linear and nonlinear differential equations can be solved using MATLAB. Higher-
order differential equation can be written as a system of coupled first-order
differential equations. A first-order ordinary differential equation (ODE) can be
written in the following form.
dx
x& = = f (t , x)
dt

To integrate the above ODE, the function representing the ODE should be defined as
the function M-file, and the integrating function calls the function M-file. The most

© 2001 by C. Nam, Y. Kim 26

popular methods for integrating a first-order differential equation are Runge-Kutta
methods. The syntax of Runge-Kutta fourth and fifth order algorithm is shown below.

>> [t, x] = ode23(‘funct’, t0, tf, x0)

where ‘funct’ is a function containing the derivative information, t0 and tf are the
initial and final time, respectively, and x0 is an initial conditions. For example, let us
consider the van der Pol equation.

&x& − 3(1 − x 2 ) x& + x = 0

The above equation can be written as

x&1 = x2
x& 2 = 3(1 − x12 ) x2 − x1

where x1 = x and x2 = x& . To integrate the above equations, function M-file vandP.m
should be first defined as follows.

function dy = vandP(t,x)
dy(1) = x(2);
dy(2) = 3(1-x(1)∗x(1)) ∗x(2) - x(1);

The following statements integrate the van der Pol equation over the interval [0,10]
with zero initial conditions.

>> x0=zeros(2,1);
>> [t,x] = ode45(‘vandP’, 0, 10, x0);

1.4.b Nonlinear Algebraic Equations

MATLAB provides a several functions that perform optimization-related works. In

this section, MATLAB functions solving the minimizing functions and finding zeros
of functions are discussed.

Minimization of a function
The MATLAB command fmin minimizes a function by finding over a value which
minimizes the given function. The syntax is shown below.

sol = fmin(‘funct’, x1, x2)

where ‘funct’ is the name of a function to be minimized, x1 and x2 are lower limit and

© 2001 by C. Nam, Y. Kim 27

upper limit of the interval of the function argument, respectively. For example, let us
consider the following problem: Find the solution that minimizes the following
performance index.

( x1 x2 − 3) 2
J = x12 +
x2

For the given example, function perf.m should be first defined as follows.

function J = perf(x)
J = x(1)∗x(1) + ( x(1)∗x(2))^2/x(2);

The minimum solution between –3 and 3 can be obtained by the following command.

>> sol = fmin(‘perf’, -3, 3)

The fmins function is similar to fmin except that given function is a function of several
variables, and an initial vector x0 should be specified rather than an interval. The
syntax is shown below.

Sol = fmins(‘funct’, x0)

Note that the fmins produces a local miminum solution.

Finding zeros of a function

The MATLAB command fzero tries to find a zero of one variable. The syntax is
shown below.

Sol = fzero(‘funct’, x0)

where ‘funct’ is the name of a function and x0 is an initial value of the scalar variable.
The fzero function first searches for an interval around the given initial value where
the function changes sign. The solution returned by fzero is near a point where ‘funct’
changes sign, or NaN if the search fails. Note that x0 can be a vector of length 2. In
this case, x0 is an interval where the sign of funct(x0(1)) differs from the sign of
funct(x0(2)). An error occurs if this is not true. Calling fzero with an interval
guarantees fzero will return a value near a point where the function changes sign.

1.4.c Numerical Integration (Quadrature)

© 2001 by C. Nam, Y. Kim 28

The integral of a function f(x) over the interval [a,b] is defined to be the area under the
curve f(x) between a and b, which can be determined by numerically integrating f(x), a
process referred to as quadrature. MATLAB has two quadrature functions for
performing numerical function integration as follows:

quad Adaptive Simpson’s rule

quad8 Adaptive Newton Cotes 8 panel rule

The syntax is shown below.

>> q = quad(‘funct’, a, b)

where ‘funct’ is the name of a function, a and b are interval to be integrated.

© 2001 by C. Nam, Y. Kim 29

Chapter 2
Linear Control Systems

Before designing a control system for a given system, the modeling procedure of the
given system is required to analyze the dynamic characteristics of the system.
Usually, a mathematical model of a dynamic system is defined as a set of nonlinear
partial differential equations. The mathematical model includes unmodeled dynamics
as well as uncertainties, i.e., the physical system cannot be completely modeled. By
discretizing the spatial coordinates, the partial differential equations can be
approximated by the ordinary differential equations. And by linearizing the system
around the operating point or equilibrium state, linear modeling can be obtained. Even
though the behavior physical systems are in general nonlinear with respect to inputs,
most of physical systems can be approximated by linear systems as long as the system
is operating around the equilibrium. A system is called linear if the principle of
superposition applies. The linear characteristics make us analyze the system and
design control systems very easily. MATLAB supports a lot of analysis tools as well
as control system design tools for linear systems. In this chapter, the basic theories of
linear systems will be briefly summarized, covering the stability analysis and control
design methods.

2.1 Linear System Analysis

2.1.a Mathematical Representation

Consider a linear system represented by the state-space form:

x& = Ax + Bu
(1)
y = Cx + Du

where x ∈R n is state variables, u ∈ Rm is control variables, and y ∈ Rl is output

variables. The state variable representation can be converted to the transfer function
form:
 Y (s)
= C ( sI − A) −1 B + D (2)
U (s)

where s denotes the Laplace variable.

MATLAB has useful commands to transform a mathematical model of linear system

to another model. The command tf2ss converts the system in the transfer function
form to the state-space form as follows:

>> [A,B,C,D]=tf2ss(num,den)

where num and den are numerator and denominator of the transfer function,
respectively. If the system has one input and one output, then the command ss2tf
converts the system in state space to the transfer function as follows:

>> [num, den]=ss2tf(A,B,C,D)

If the system has more than one input, the following command is used.

>> [num, den]=ss2tf(A,B,C,D,iu)

where iu is an input index for transfer function.

2.1.b Stability of the system

Consider an autonomous system described by the linear vector differential equation

x& = Ax + Bu (3)

The above system is said to be stable, if the solution of Eq. (3) tends toward zero
(which is obviously the only equilibrium state if A is of full rank) as for an arbitrary
initial condition. It is well known that the solution of Eq. (3) is given by

t
x(t) = eAt x(0) + ∫0 e A(t − τ ) Bu(τ )dτ (4)
If all eigenvalues of A are distinct, the response of system ensuing from a given
initial condition x0 with zero control input can be written as

x(t) = eAt x(0) = Φe −∆ tΦ −1x(0) (5)

where ∆ = diag (λ1 , L , λn ) , λi are the eigenvalues of A , futher, Φ denotes the

modal matrices of A . For the repeated eigenvalue case, the situation is much more
complicated (i.e., we should solve for the generalized eigenvectors of A ). The
generalization of Eq. (5) for the case of generalized eigenvectors has a similar form
but will not be discussed here

From Eq. (5), we can see that the system will be asymptotically stable if and only if
all the eigenvalues of A have negative real parts, i.e.,

ℜ{λi (A)} < 0 (6)

where ℜ{⋅} denotes the real part of {⋅}. Thus, the stability of a linear system can be
completely characterized by the eigenvalues of the system. This eigenvalue approach
to stability yields both necessary and sufficient conditions.

2.1.c Controllability and Observability

The purpose of this section is to introduce concepts of controllability and

observability. The concepts of controllability and observability are very important in
the study of dynamical system control and estimation problems. The concept of
controllability is closely related to the existence of a feasible control law, since it is
not always possible to find a control law of a given form that makes the closed-loop
system stable with respect to the desired state or trajectory. Similarly, the concept of
observability is related to the existence of a feasible algorithm for estimating the state
variables from measurements. Controllability measures the particular actuator input
configuration's ability to control all system states, whereas observability measures the
particular sensor output configuration's ability to obtain the information needed to
estimate all system states. Following are the conventional definitions for linear
dynamical systems.

Consider a linear system

x& (t ) = Ax(t ) + Bu (t )
(7)
y (t ) = Cx(t )

Definition: Controllability
A linear system is said to be controllable at time t0 if there exists a control u (t ) that
will transfer the initial state x(t0 ) to the origin at some finite time t1 > t 0 . If this is
true for all initial times t 0 and all initial states x(t0 ) , the system is said to be
completely controllable.

There are several necessary and sufficient conditions (or criteria) for Controllability.

Theorem: Criteria for Controllability

A system is completely controllable if and only if any one of the following equivalent
conditions is satisfied:
(1) The n × nm controllability matrix C has rank n .

C ≡ [B AB A2 B L An B] (8)

(2) The controllability grammian Wc is full-rank (in fact, positive definite).

∞
Wc ≡ ∫0 e Aτ BBT e A τ dτ
T
(9)

There are other criteria for checking the controllability of the linear system, known as
PBH(Popov, Belevitch, and Hautus) rank test and PBH eigenvector test that will not
be discussed in this book.

Note that MATLAB operators for computing controllability matrix C and the
controllability grammian Wc are as follows:
 >> C = ctrb(A,B)
>> Wc = gram(A,B)

Definition: Observability
A linear system is said to be observable at t0 , if x(t0 ) can be determined from the
knowledge of u (τ ) , y (τ ) for t 0 ≤ τ ≤ t1 . If this is true for all t0 and all x(t0 ) , the
system is said to be completely observable.

Theorem: Criteria for Observability

A system is completely observable if and only if any one of the following equivalent
conditions is satisfied:

(1) The nl × n observability matrix O has rank n .

 C 
 
 CA 
O ≡ CA 2  (10)
 M 
 
 CAn 

(2) The observability grammian Wo is full-rank (in fact, positive definite).

∞
Wo ≡ ∫ 0 e A τ CT CeA τ dτ
T
(11)

There are also PBH tests for checking the observability of the linear system that will
not be discussed in this book. The above criteria show the duality of the observability
and controllability. It has been proven that the controllability and observability
grammian matrices satisfy the following Lyapunov equations:

AWc + Wc A T + BBT = 0
(12)
A TWo + Wo A + CT C = 0

Note that MATLAB operators for computing the observability matrix O and the
observability grammian Wo are as follows:
 >> O = obsv(A,B)
>> Wo = gram(A',C')

2.1.d Eigenvalue Problem

Let us introduce the generalized eigenvalue problem. Consider the system of

differential equations

Bx& = Ax (13)

Note that overdots denote time differentiation throughout this text. The right and left
eigenvalue problems associated with x = φe λt solutions of Eq. (13) are

right : λ i Bφ i = Aφi ,
i = 1,2, L , n (14)
left : λi BTψ i = A Tψ i ,

with the usual normalization of the biorthogonality conditions

φiT Bφi = 1, i = 1,2,L,n

(15)
ψ Tj Bφ i = δ ij , i, j = 1,2,L,n
so that
ψ Tj Aφi = λiδij , i, j = 1,2,L,n (16)

The MATLAB statement to solve this eigenvalue problem is

>> [Phi, Lambda] = eig(A, B)

>> [Psi, Lambda] = eig(A',B')

where Phi and Psi are right and left modal matrices, and Lambda is a matrix with the
eigenvalues on the diagonal.

Note that every eigenvalue solver (computer software package) uses its own special
normalization equation for eigenvectors. In MATLAB, eigenvectors are normalized
so that norm of each eigenvector becomes 1. In order to obtain eigenvectors that
satisfy some other normalization conditions [e.g., Eq. (15)], we need our own special
normalization routine. The following code solves right and left eigenvalue problems
and normalizes eigenvectors according to Eq. (15). The MATLAB operator to
compute this normalized eigensolution is

>> [Lambda, Phi, Psi]=eign(A,B)

Observe that the above code cannot be used for the repeated eigenvalue case since
eigenvectors satisfying Eqs. (15) and (16) do not generally exist. The source code for
eign is shown in Table 2.1.

There are many eigenvalue problem-related applications, such as system analysis and
synthesis techniques that use modal frequencies and mode shapes, as well as stability
analysis and controller design methods such as eigenstructure assignment techniques.
We will consider these in subsequent chapters of this text.

Now let us introduce spectral decomposition. Defining the modal matrices of right
and left eigenvectors as

Φ = [φ1 , L φn ]
(17)
Ψ = [ψ1 , L ψ n]

For the case that B matrix is an identity matrix, Φ −1 = ΨT from Eq. (15). Therefore,
we can obtain spectral decomposition of matrix A as follows:

A = Φ∆ΨT = Φ∆Φ−1 (18)

The diagonalization of A and existence of a nonsingular Φ depend upon matrix A

having n linearly independent (but not necessarily unique) eigenvectors. If the
eigenvalues of A are distinct, A can be diagonalized, otherwise we generally must
solve a more complicated generalized eigenvalue problem for repeated eigenvalues
for obtaining the modal matrix Φ , and ∆ takes on a form that generally has entries
only on the diagonal and the superdiagonal.

2.1.e Eigenvalue Sensitivities

The usefulness of modal sensitivities for analysis and design of engineering systems is
well known. Some specific applications include identification of dynamical systems,
optimum design of structural systems, and design of control systems by pole
placement. In the above algorithms, eigenvalue and eigenvector derivatives with
respect to design parameters are often useful. In this section, for the nondefective
linear systems, analytic expressions for eigenvalue and derivatives are derived using a
modal expansion approach.

Consider the nondefective system

Bx˙ = Ax (19)

The right and left eigenvalue problems of the above system are represented as
follows.

λi Bφ i = Aφi , i = 1,2,L,n (20)

λi B ψ i = ATψ i , i = 1,2,L,n
T
(21)

where the conventional normalization of the biorthogonality conditions for the

eigenvectors are adopted as in Eqs. (15) and (16).

Theorem: Eigenvalue Sensitivity

Partial derivative of eigenvalues of the system (19) is given by the following
equation:

∂λi  ∂A ∂B 
= ψ iT  − λ i φi (22)
∂ρ  ∂ρ ∂ρ 

where ρ is some arbitrary parameter upon which A and B depend.

Proof:
Let us derive the eigenvalue derivatives with respect to a scalar parameter ρ . First,
we take the partial derivative of Eq. (20) to get

∂λi ∂B ∂φ ∂A ∂φ
Bφi + λi φi + λ i B i = φi + A i (23)
∂ρ ∂ρ ∂ρ ∂ρ ∂ρ

We then premultiply the above equation by ψ Tj to obtain

∂λi T ∂φ  ∂A ∂B 
ψ j Bφi + (λiψ Tj B −ψ Tj A) i = ψ Tj  − λ i  φi (24)
∂ρ ∂ρ  ∂ρ ∂ρ 

Using Eqs. (15) and (21) in the above equation gives the analytical expressions of
eigenvalue sensitivities as

∂λi  ∂A ∂B 
= ψ iT  − λ i φi (25)
∂ρ  ∂ρ ∂ρ 

2.1.f State Covariance Matrix

Consider a linear system

x& (t ) = Ax(t ) + Bu (t ) (26)

In general, the state response may describe the detailed information about system
performance. And the state covariance matrix gives useful information to study the
properties of linear system, because certain information about the linear system is
embodied in the state covariance matrix. The state covariance matrix due to impulsive
inputs in u is defined as follows:

∞
X≡ ∫0
x(t)x T (t)dt (27)

Using Eq. (4), the response of system, i.e., the solution of Eq. (26), ensuing from
impulsive inputs with zero initial condition can be written as

x(t) = eAt B (28)

Substitution of Eq. (28) into Eq. (27) yields the state covariance matrix as follows:

∞
∫
T
X≡ e At BBT e A tdt (29)
0

The state covariance matrix X can be computed by solving Lyapunov equation

introduced in the following theorem.

Theorem: State Covariance Matrix for Linear Systems

The state covariance matrix satisfies the following Lyapunov equation if and only if a
linear system is asymptotically stable.

AX + XAT + BB T = 0 (30)

Proof:
(Necessity)
The integrand is an infinite sum of terms of the form t k eλt , where λ is an eigenvalue
of A and, since ℜ{λi (A)} < 0, the integrand will exist for all values of t . Next, by
direct substitution of Eq. (29) into the left hand side of Eq. (30), we find that the
combined integrand is a perfect differential, permitting us to directly establish

∞ ∞
AX + XAT = ∫0 AeAt BBT e A t dt + ∫0 eAt BBT e A t AT dt
T T

( )
∞
d At T AT t ∞
∫
T AT t
= = At
e BB e dt e BB e (31)
0 dt 0

= −BB T

T
Note that e At (and e A t ) →0 as t → ∞, since all eigenvalues of A have negative real
T
parts, and since e At (and e A t ) equal 1 at t = 0.

(Sufficiency)
The sufficiency part can be proved by Lyapnov (direct) stability theorem that will not
be covered in this book.
Note that the Lyapunov equation is equivalent to a set of n(n + 1) / 2 linear equations
in n(n + 1) / 2 unknowns for an n th order system. The Lyapunov equation can be
solved by using numerical algorithms utilizing QR factorization; Schur
decomposition, or spectral decomposition, however, we believe that the most efficient
and robust algorithm utilizes the QR factorization. Note that the MATLAB operators
for computing the solution of Lyapunov equations are

>> X = lyap(A, B∗B')

2.2 Linear Optimal Control

Among the most fundamental results in modern control theory are the two sets of
analytical developments that underlie the linear-quadratic regulator (LQR) and
eigenstructure assignment regulator (EAR). Design of control systems for
aeroservoelasticity problems has been accomplished using both design techniques.
The central aspects of these two sets of approaches are presented in this section.

In the LQR approach, the central feature is the minimization of a quadratic

performance index, subject to satisfaction of a linearized system differential equation
model. The integral quadratic performance index appeals to the intuitive energy-based
notion that we should seek a weighted compromise between the closed-loop output
error energy and a measure of the control effort required to regulate the system.
However, the fundamental practical difficulties in applying the LQR approach arise
because it is difficult to capture many practical performance, robustness, and
admissible design inequality constraint issues in a single scalar performance measure,
simply through a judicious a priori choice of weight matrices.

Since the complete response of a given linear system can be fully determined by the
closed-loop eigensolution, another attractive approach is to find a control law that
satisfies prescribed constraints on the closed-loop eigenvalues and eigenvectors. It is
also possible to determine many of the candidate measures of performance and
stability robustness (insensitivity with respect to system model errors and external
disturbances) as simple functions of the closed-loop eigenvalues and eigenvectors.
The class of methods that directly impose prescribed constraints on the eigenvalues
and eigenvectors are known as eigenstructure assignment regulator design methods.
While the EAR approaches allow some immediate insight that narrows the search for
practical designs, some additional difficulties arise. For controllable systems, we will
see that the closed-loop eigenvalues can be placed exactly at an infinity of locations
while, generally, the closed-loop eigenvectors can be placed only in certain feasible
subspaces.

The quadratic regulator and eigenstructure assignment approaches are among the most
important representatives of modern control theory, which is essentially the set of
approaches based on the time-domain state space description of linear dynamical
systems. As will be evident in the discussion below, the EAR formulations are
actually a diverse family of recently developed concepts and the associated gain
design algorithms that subsume and greatly expand upon the methodology historically
known as pole placement.

2.2.a Linear Quadratic Regulator

If a dynamical system is completely controllable, these exists at least one control that
will transfer the system from any/all given initial states to any/all desired states.
Consider a completely controllable and observable dynamical system whose behavior
is modeled by a nonlinear system of differential equations in the first-order state
variable form

x˙ (t) = Ax(t) + Bu(t) (32)

where x(t) ∈ Rn is the state vector, and u(t) ∈R m is the control vector.

The objective of optimal control theory is to determine the control law that satisfies
the differential equation constraints, i.e., all trajectories are consistent with our model
of the system dynamics [Eq. (32)], and at the same time minimizes (or maximizes)
some performance criterion. The most general performance index functional to be
minimized (in the current discusion) is defined as

( ) ∫
J = h x( t f ),t f +
tf
g( x(t),u(t),t ) dt (33)
t0
where h and g are given functions, t0 is the initial time, and t f is the final time. The
objective is to find the optimal control u (t ) that drive the solution of the differential
equations of Eq. (32) from the given initial conditions to the final time such that the
performance index of Eq. (33) is minimized.

The derivation of the necessary conditions for optimal control is relatively lengthy
and straightforward and is not presented here. The linear regulator problem is the
most obvious special case of a linear system with the feedback control law being
parameterized as a linear function of the instantaneous state. The object is to
determine an optimal feedback control gain K for a linear law of the form

u(t) = −Kx(t) (34)

which minimizes the quadratic performance index

x (t f )Sx(t f ) + ∫t {x T (t)Qx(t) + uT (t)Ru(t)} dt

1 T 1 tf
J = (35)
2 2 0

where S and Q are real symmetric positive semidefinite matrices, and R is a real
symmetric positive definite matrix. It is assumed that the states and controls are not
bounded. Moreover, we assume that all elements of the final state vector x(tf ) are
free and the final time t f is fixed.

Note that the physical and geometric motivation for this performance measure is as
follows: It is desired to maintain the time-varying state close to the origin without
incurring an excessive expenditure of control effort. The balance between small errors
and small control effort is achieved through judicious assignment of the weight
matrices.

By applying the principles of the calculus of variations to minimization of the

performance index, the necessary conditions for the linear quadratic regulator
problem are obtained: the matrix differential Riccati equation.

dP(t)
= − AT P(t) − P(t)A + P(t)BR− 1B T P(t) − Q (36)
dt
The solution P(t) (an n × n positive semi-definite symmetric matrix), called the
Riccati matrix, is a symmetric positive definite matrix. The optimal gain matrix is
related to the Riccati gains in the following manner:

K(t) = R−1B T P(t) (37)

Note that, even though the plant matrices are constant, the optimal feedback gain
matrix K is time-varying for the finite time control case. Also, note that
measurements (or estimates) of all of the state variables must be available to
implement the optimal control law. For obvious reasons, such a law is known as a full
state feedback control law.

The most important special case of the linear regulator problem is the one for which:
(i) The plant is completely controllable, (ii) The plant dynamics are time invariant,
i.e., A, B, Q, and R are constant matrices, (iii) The final state error weight matrix is
zero, S=0 in Eq. (2.35), and (iv) The final time t f is infinite. In other words, the
problem is to find the optimal control minimizing the performance index

{x (t)Qx(t) + u T (t)Ru(t )} dt
1 ∞ T
J = ∫
2 0
t
(38)

In this case, it can be established that the optimal control law for minimizing the
quadratic index over an infinite time duration is stationary, and the feedback gain
matrix is a constant,,

u(t) = −Kx(t)
(39)
K = R−1 BT P

determined by the constant solution P of the matrix Riccati equation, Eq. (36),
reduced to a set of quadratically nonlinear algebraic equations,

A T P + PA − PBR−1B T P + Q = 0 (40)
which is called the algebraic Riccati equation (ARE). The fact that P approaches a
constant can be established from the general time-varying P(t) solution using
numerical method and letting t f → ∞.

Note that P in Eq. (40) is an n × n matrix, carrying out the algebra to obtain the
scalar elements of Eq. (40) generates a coupled system of n 2 first-order differential
equations. Notice that if P satisfies Eq. (40), PT also satisfies it, because the
transpose of Eq. (40) yields exactly Eq. (40). Since P is symmetric, there are not n 2
but rather n(n + 2 ) / 2equations that must be solved.

Note that the MATLAB (Control System Toolbox) operator for the steady-state LQR
problem has the following syntax:

>> [K,P] = lqr(A, B, Q, R)

which returns the constant optimal gain K and the ARE solution P.

2.2.b Linear Quadratic Regulator with Output Feedback

In the previous section, full state optimal control law was discussed. In real situation,
only some of the states are available as measured outputs. Therefore, in order to
utilize the full state feedback, observer or estimator should be implemented with
controller to estimate the state variables using the mathematical model and output
measurements. This makes the control system very complicated. In this section, the
LQR with output feedback is briefly introduced. Because the estimator is not
involved, the whole control system using output feedback control law will become
very simple. However, the control system performance will be degraded because only
limited information is used for output feedback purpose, and the design procedure
will be more complicated than the full state LQR.

Consider a linear time-invariant system

x˙ = Ax + Bu
(41)
y = Cx
The control law is restricted to a linear time-invariant output feedback of the form

u = −Ky (42)

where K is an output feedback gain matrix to be determined by the design procedure.

The design objective for flutter suppression may be obtained by selecting the control
input to minimize a quadratic performance index of the type

{y (t)Qy(t) + u T (t)Ru(t)} dt
1 ∞ T
J = ∫
2 0
t
(43)

where Q ≥ 0 and R > 0 are symmetric weighting matrices.

The performance index may be expressed in terms of K as

x (C QC + C T K T RKC) x dt
1 ∞ T T
J =
2 ∫0
(44)

The design problem is to select the gain K so that J is minimized subject to

following the system dynamic constraints

x˙ = (A − BKC)x ≡ Ac x (45)

The dynamical optimization problem may be converted into an equivalent static

optimization problem that is easier to solve. Assume that there exists a constant,
symmetric, positive semidefinite matrix P that satisfies the following equation.

d T
dt
( x Px) = −x T ( CT QC + CT K T RKC) x (46)

By assuming that the closed-loop system is asymptotically stable so that x(t )

vanishes with time, the performance index can be represented as

1 T 1
J = x (0)Px(0) = trace{Px(0)x T (0)} (47)
2 2
where the operator trace(⋅ ) denotes the trace of a square matrix. If P satisfies Eq.
(46), then we may use Eq. (45) to see that

−x T (CT QC + CT K T RKC ) x = ( x Px ) = x˙T Px + x T Px˙

d T
dt (48)
= x ( A P + PAc) x
T T
c

The above equation should satisfy for all initial conditions, and therefore the
following condition should be satisfied for all state trajectories.

AcT P + PAC + CT K T RKC + CT QC = 0 (49)

Now we can state that, if there exists a constant symmetric positive semidefinite
matrix P that satisfies Eq. (49), then the performance index J for the closed-loop
system can be represented in terms of P by Eq. (47). Since the initial state x(0) may
not be known, we adopt Athans and Levine's procedure that minimizes the expected
value of performance index

1
J = E{J} = trace(PX ) (50)
2

where the operator E(⋅) denotes the expected value over all possible initial conditions,
and the n × n symmetric matrix X is defined by

X ≡ x(0)xT (0) (51)

The usual assumption is that the initial states are uniformly distributed on the unit
sphere; then X becomes identity matrix. By applying optimality conditions for the
above problem, the optimal control gain matrix K can be obtained by solving the
following three coupled nonlinear algebraic matrix equations simultaneously.

AcT P + PAC + CT K T RKC + CT QC = 0 (52)

ACS + SAcT + X = 0 (53)
RKCSC T − BT PSC T = 0 (54)
Note that P , S , and K are matrices to be determined from the above three equations.
There exists several numerical approaches to solve Eqs. (52)-(54). Gradient-based
routines can be used with various optimization algorithms. Moerder and Calise
proposed an efficient iterative algorithm, which converges to a local minimum if the
following conditions hold: (i) there exists an output gain matrix K that Ac is stable,
(ii) the output matrix C has full row rank, (iii) control weighting matrix R is positive
definite, and (iv) output state weighting matrix Q is positive semidefinite and should
be chosen so that all unstable states are weighted in the performance index. The
iterative algorithm is outlined as follows:

Step 1: Choose K such that the closed-loop system matrix Ac has all eigenvalues
with negative real parts.

Step 2: With the given K and the according Ac , solve the Lyapunov Eqs. (52) and
(53) for P and S , and compute the performance index value J .

Step 3: Evaluate the gain update direction

∆K = R −1B T PSC T (CSC T ) − K

−1
(55)

Note that for the case that the matrix CSC T in Eq. (55) is singular, Moore-Penrose
generalized inverse utilizing singular value decomposition may be useful to inverse
the matrix CSC T to deal with the rank deficiency problem

Step 4: Update the gain by

Knew = K + α∆K (56)

where α is chosen so that the closed-loop system Ac is asymptotically stable, and the
performance index of the updated system is less than that of the current system. To
obtain the appropriate α , set the initial step size α 0 (for example, 0.3), and proceed if
the above two conditions are satisfied. If not, decrease the step size using the
following simple rule till the above conditions are satisfied.
 αnew = καold , 0<κ <1 (57)

Step 5: Iterate Steps 2 through 4 until converged solution is obtained.

Any minimization algorithm requires the selection of an initial stabilizing gain,

however such a gain may or may not be easy to find. One possible way to find the
initial gain matrix is to use root locus techniques by closing one loop at a time in the
control system. For the initial gain matrix for the iterative procedure, we first solve
the full state LQR problem with performance index Eq. (2.42), and construct an initial
stabilizing output gain matrix by choosing the corresponding gain parameters to be
used in the output feedback out of the resulting LQR gain matrix elements. Let us
explain this procedure in detail. The following relation can be obtained by considering
the full state feedback and output feedback control scheme

u = −K F x = −KCx (58)

where K F is the full state gain matrix resulting from LQR solver. The initial output
feedback gain matrix can be computed using the following equation.

KF = KC (59)

In general, the matrix C is not square, and therefore the Moore-Penrose pseudo-
inverse should be used as follows:

K = KF C + (60)

where the pseudo-inverse of the matrix C can be expressed as

C + = V1 Σ −1U 1T (61)

where U 1 , V1 are unitary matrices obtained by singular values decomposition of the

matrix C , Σ = diag{σ 1,L,σ r } contains as its elements the real singular values of C ,
and r denotes the rank of C .
 2.3 Pole Placement Technique

2.3.a Sylvester Approach

The eigenstructure assignment regulator theory is employed to regulate the system

with the smallest control effort. It is assumed that all states are available for feedback.
The gain matrix of the completely controllable and observable linear dynamic is
expressed as follows

u = −Kx (62)

The right and left eigenvalue problems of the closed-loop system can be written,
respectively, as

(A − BK )φ ic = λci φic (63)

where φ ic is the right eigenvectors of the closed-loop system corresponding to the

closed-loop eigenvalue λci . The pole placement algorithm based on Sylvester equation
utilizes the parameter vector defined by

hi = Kφic (64)

Substituting Eq. (64) into Eq. (63) yields the following Sylvester equation

(A − λciI)φic = Bhi (65)

or in matrix form

AΦ c − Φ cΛ D = BH (66)

where
Φ c = [φ1c , φ 2c, ...,φ nc ]
Λ D = diag(λc1, λc2, ...,λcn ) (67)
H = [h1, h2 , . . . ,hn ]
with Λ D containing the desired closed-loop eigenvalues. From Eq. (65), it can be
seen that

φic = (A − λci I)−1 Bhi (67)

If λci are distinct from their open-loop positions, the columns of H generate the
corresponding closed-loop eigenvectors. For the given parameter matrix H , the
closed-loop modal matrix Φ c can be determined by solving Sylvester equation. The
gain matrix is determined using the relation Eq. (64)

K = HΦ−c 1 (68)

2.3.b Nonlinear Programming (NLP) Approach

The eigenvalues of the system can be assigned to the desired values by using a pole
placement technique. Our problem can be stated as the nonlinear parameter
optimization problem in which we seek to impose specified eigenspace equality
constraints, and the elements of output feedback gain matrix are parameters to be
determined. We introduce the output feedback control law

u = −Ky (69)

where K denotes the output feedback gain and y includes only the measurement
information. The eigenvalue problem of the closed-loop system can be written as

(A − BKC)φic = λciφ ic (70)

where φ ic is the eigenvectors of the closed-loop system corresponding to the closed-

loop eigenvalue λci . Let us define p as our parameter vector that consists of the
elements of output feedback gain K . Now, we can formulate our nonlinear
programming problem as follows:
Find output feedback gain parameter p satisfying the following equality constraint
equation that is related to the closed-loop eigenvalue assignment

fi ( p) = ( p) = 0, i = 1, 2 , . . r,
d
λi − λ i (71)

d
where r is the number of the structural modes, and λi denotes the desired
eigenvalues of the structural modes. The above problem can be solved by using any
gradient-based nonlinear programming algorithm. In this section, homotopic
nonlinear programming with minimum norm correction algorithm is used. In order to
enhance the convergence, the linear homotopy map is generated: original eigenvalue
assignment problem is replaced by the one-parameter α family as follows:

fi ( p) = {αλdi }
+ (1 − α )λ i ( pstart ) − λi ( p ) = 0, 0 ≤ α ≤1 (72)

where pstart is the initial guess of parameter vector. Sweeping α at a suitably small
increment generates a sequence of neighboring problems. We solve this sequence of
problems using the neighboring converged solutions to generate starting iteratives for
each subsequent problem. If we obtain a solution for α = 1 , then we have the desired
solution. The local design corrections for each α are performed by linearizing the
neighboring problem (nonlinear constraint equations) about the local solution and
computing the minimum norm differential correction that satisfies the linearized
constraint equation. This algorithm is a generalized Newton process for solving a
system of underdetermined nonlinear equations. On each iteration, the norm of the
correction vector is minimized while satisfying the following linearized constraint
equation.

 ∂f 
 
 ∂p  ∆p = − f ( p)
(73)
p

where f (p) = [ f1(p), f2 (p),..., fr ( p)] is a constraint vector. The solution of the above
T

equation can be obtained by pseudo-inverse of matrix {∂f /∂p} p utilizing singular

value decomposition. The eigenvalue sensitivities required in the above equation for
iterative direction search can be evaluated by using the following analytic
formulation.
 ∂λci  ∂(A − BKC)  c
= ψ iT   φi (74)
∂pk  ∂pk 

where φ ic and ψ ic are normalized right and left eigenvectors of the closed-loop system
c
corresponding to the eigenvalue λi , respectively. Figure shows the flow chart of this
algorithm.
 Set α = 0
Select initial
p = p0

Update
homotopy parameter
α = α + ∆α

Set portable desired eigenvalues

(λdi ) = αλdi + (1 − α)λ i ( p0 )
α

Compute gain K using p

Compute closed - loop eigenvalues
λ i ( p) = λ i (A − BKC)

f i ( p ) = ( λ i ) α − λ i ( p)
d

∂λ i ( p)
Compute sensitivity
∂p j

+
 ∂f 
∆p = −   f (p)
 ∂p 

f ≤ ε1 ? NO
∆p ≤ ε 2 ? p = p + ∆p

YES
NO
α =1 ?

YES

Compute gain matrix K

Chapter 3
Aeroservoelastic Analysis of the Airfoil

3.1. Unsteady Aerodynamic Forces of the Typical Section Model

The unsteady aerodynamic forces are calculated based on the linearized thin-airfoil
theory. In this section, Theodorsen’s approach (Ref: T. Theodorsen, "General Theory of
Aerodynamic Instability and the Mechanism of Flutter", NACA TR 496, 1935.) will be
summarized and the flutter analysis will be conducted based on his approach.
In his approach, aerodynamic surfaces were modeled by flat plates. Theodorsen assumed
that the flat airfoil is oscillating about the shear center (elastic axis) and unsteady flow is
composed of two flow components, (a) non-circulatory flow which can be expressed
through the sources and sinks and (b) circulatory flow related to the flat vorticity surface
extending from trailing edge to the infinity. For each flow component, he obtained the
velocity potential, then calculate the pressure using Bernoulli theory.

y
V h (x1,y1)
b=1
α ab x x

b= -1 b= 1 (x1,-y1)
elastic axis
Circle
Flat Airfoil
Figure . Flat airfoil and the transformed circle.

3.1.a The Non-circulatory Flow

By Joukowski’s conformal transformation, the airfoil can be mapped onto a circle. The
velocity potential of a source ε on circle ( x1 , y1 ) can be expressed as the follow.

ϕ=
ε
4π
[
Ln ( x − x1 ) + ( y − y1 )
2 2
] (1)

© 2001 by C. Nam, Y. Kim 1

Similarly, the velocity potential due to a source 2ε at on circle ( x1 , y1 ) and a sink − 2ε
at on circle ( x1 ,− y1 )

ε  ( x − x1 )2 + ( y − y1 )2  (2)
ϕ= Ln  2 
2π  (x − x1 ) + ( y + y1 ) 
2

Since y = 1 − x 2 , the velocity potential is a function of x only.

The downward displacement of the airfoil can be written as

z = h + α (x − ab) (3)
Then, the upwash will be
 ∂z ∂z 
wa ( x, t ) = − + V  = −[h& + α& ( x − ab)] − Vα (4)
 ∂t ∂x 
Therefore, the velocity potential due to pitch angle, α will be

− Vα  ( x − x1 ) + ( y − y1 ) 
2 2
1 1
ϕ α = b ∫ ϕ dx = b ∫ Ln  2 
dx1 = Vαb 1 − x 2 (5)
 (x − x1 ) + ( y + y1 ) 
−1 −1 2π 2

Similarly, velocity potentials due to plunge motion, h& and angular velocity, α& are
respectively expressed as

ϕ h& = h&b 1 − x 2 (6)

α&b 2 x
ϕ α& = π ( x + 2) 1 − x 2 − α&b 2 (1 + a) 1 − x 2 = α&b 2 ( − a) 1 − x 2 (7)
2π 2
The total velocity potential due to noncirculatory flow becomes
ϕ NC = ϕ α + ϕ h& + ϕ α&
x (8)
= Vαb 1 − x 2 + h&b 1 − x 2 + α&b 2 ( − a ) 1 − x 2
2
By Bernoulli theorem, the pressure is obtained
 ∂ϕ ∂ϕ  dϕ
∆p = −2 ρ  +V  = −2 ρ = −2 ρϕ& (9)
 ∂t ∂x  dt

© 2001 by C. Nam, Y. Kim 2

And the force (positive downward) and the pitching moment (positive nose-up) about the
elastic axis will be expressed as
1 1
FNC = b ∫ ∆pdx = −2 ρb ∫ ϕ& dx
−1 −1 (10)
(
≡ −πρb h&& + Vα& − baα&&
2
)
1 dϕ1
M NC = b ∫ ∆p ( x − a )bdx = −2 ρb 2 ∫ ( x − a )dx
−1 dt
−1

 1   (11)
≡ πρb 2 Vh& + bah&& + Vα& 2 − b 2  + a 2 α&& 
 8  

3.1.b The Circulatory Flow

To satisfy the Kutta condition, Theodorsen employed a bound vortex distribution over the
airfoil, and a vortex distribution over the airfoil wake.

(X,Y)
1/X0 X0
b=1
x
∆Γ −∆Γ

Figure Circle in circulatory flow

In order to consider wake, let’s assume a bound vortex ∆Γ = γdx at 1 / X 0 , and a shed
vortex − ∆Γ at X 0 .
Then, the velocity potential due to a vortex
∆Γ  −1 Y Y 
∆ϕ Γ = −  tan − tan −1 
2π  X − X0 X − 1/ X 0 
(12)
∆Γ  ( X 0 − 1 / X 0 )Y 
=− tan −1  2 
2π  X − X ( X 0 + 1 / X 0 ) + Y + 1
2

© 2001 by C. Nam, Y. Kim 3

Let’s define X 0 + 1 / X 0 = 2 x 0 , and X = x, Y = 1 − x 2
Then,

2
X 0 = x0 + x0 − 1
1 2 (13)
1/ X 0 = = x0 − x0 − 1
2
x0 + x0 − 1

The velocity potential can be expressed as

∆ϕΓ = −
∆Γ −1 
tan  2

(
1 − x 2 2 x0 2 − 1 


)
2π  x − x (2x 0 ) + (1 − x ) + 1
2

(14)
 2 2 
∆Γ −1  1− x x 0 − 1 
=− tan
2π  x 2 − xx 0 
 
where
−1 ≤ x ≤ 1, 1 ≤ x0 ≤ ∞

It has to noted that the vortex is moving away from the airfoil with velocity of V.
Therefore, by Bernoulli theorem, the pressure due to the vortex is
 ∂∆ ϕ Γ ∂∆ϕ Γ 
∆ p = −2 ρ  +V  (15)
 ∂t ∂x 
where
  1 − x 2 x 2 − 1 
2π ∂∆ϕ Γ ∂
= −  tan −1  0

∆Γ ∂x ∂x   1 − xx0 
   (16)
2
x0 − 1 1
=
1− x 2 x0 − x

© 2001 by C. Nam, Y. Kim 4

 2π ∂∆ϕ Γ 1− x2 1
= (17)
∆Γ ∂x 0 x0
2
− 1 x0 − x

The pressure at x due to the vortex at x0

∆Γ  x 0 − 1 1− x2  1
2

∆p Γ = −2 ρ V + 
2π  1 − x 2 x 0 − 1  x0 − x
2

∆Γ   1
2
x0 − x 2
= − ρV   (18)
2π  1 − x 2 x 2 − 1  x0 − x
 0 
∆Γ x0 + x
= − ρV
2π 1 − x 2 x 2 − 1
0

The force on the whole airfoil due to a vortex at x0 will be

1
∆FΓ = b ∫ ∆pΓ dx
−1

x0 (19)
= − ρVb ∆Γ,1 ≤ x0 ≤ ∞
2
x0 − 1

The total force can be calculated by integrating with respect to x0

FΓ = ∫ ∆FΓ
∞ x0 (20)
= −ρVb ∫1 2
γdx0
x0 − 1
Similarly,

∆M Γ = b 2 ∫ ∆pΓ ( x − a )dx,1 ≤ x0 ≤ ∞
1
(21)
−1

M Γ = ∫ ∆M Γ

∞ 1 x0 + 1  1  x0  (22)
= − ρVb 2 ∫  − a +  γ dx
1 2 x0 − 1  2  x 2 −1
 0 

© 2001 by C. Nam, Y. Kim 5

It has to be noted that the force and moment are function of vortex strength γ. By
applying Kutta condition at trailing edge we can determine the vortex strength. First,
let’s get the total velocity potential. The total velocity potential is

ϕ total = ϕ Γ + ϕ α + ϕ h& + ϕ α&

x (23)
= ϕ Γ + Vαb 1 − x 2 + h&b 1 − x 2 + α&b 2 ( − a) 1 − x 2
2
By applying the Kutta condition, we can get the following equation
∂ϕ Γ Vαb(− x ) h&b(− x ) 1 2
+ + + α&b 1 − x 2
∂x 1− x2 1− x2 2
x (24)
α&b 2 ( − a )(− x )
+ 2 = finite at x = 1
1− x2

Therefore,
 2 ∂ϕ Γ   & 2 1 
 1 − x ∂x  + − Vαb − hb − α&b ( 2 − a ) = 0 (25)
  x =1  

Since
2
∂∆ϕ Γ ∆Γ x0 − 1 1
= (26)
∂x 2π 1− x 2 x0 − x
we can get the following expression from the above equation.
2
 2 ∂ϕ Γ  ∆Γ x0 − 1
∂x  x =1 ∫ 2π x0 − 1
 1 − x =

2
b ∞ x0 − 1
=
2π ∫
1 x0 − 1
γdx0 (27)

1 
= Vαb + h&b + α&b 2  − a 
2 
Let’s define

© 2001 by C. Nam, Y. Kim 6

 2
1 ∞ x0 − 1 1 
2π ∫
1 x0 − 1
γdx0 = Vα + h& + α&b − a  ≡ Q
2 
(28)

Then, the total force and moment on the airfoil will be

∞ x0
FΓ = −ρVb ∫ γdx0
1 x0 2 − 1
∞ x0
∫
1
x 02 − 1
γdx0
= −2πρVbQ (29)
∞ x0 + 1
∫ 1 x0 − 1
γdx0

= −2πρVbCQ
1 x +1  1 x0 
MΓ = − ρVb ∫  0
∞
− a + 
2  γdx
 x0 2 − 1 
 2 x0 − 1  2 0
1
(30)
1  1 
= −2πρVb 2Q  − C a +  
2  2 
where C is the Theodorsen function, and defined as
∞ x0
∫ 1
x 02 − 1
γ dx 0
C= (31)
∞ x0 + 1
∫ 1 x0 − 1
γ dx 0

Let’s assume that the airfoil has a simple harmonic motion

 s  
i  k  − x 0  +ϕ 
= γ 0e [
 b   i ωt − kx 0 +ϕ ]
γ = γ 0e
ωb ks
where s = Vt, k = , = ωt
V b

Then, Theodorsen function will be expressed as

© 2001 by C. Nam, Y. Kim 7

 ∞ x0
∫
1 x0 − 1
2
γ 0 e iωt e − ikx e iϕ dx 0
0

C=
∞ x0 + 1
∫ 1 x0 − 1
γ 0 e iωt e − ikx0 e iϕ dx 0

∞ x0
∫
1 x0 − 1
2
e −ikx 0 dx0
=
∞ x 0 + 1 −ikx 0
∫ 1 x0 − 1
e dx0
(32)
= F(k) + iG(k)

The Theodorsen function can be expressed in terms of the following Bessel functions
π π
−
J1 + i Y1
C(k) = 2 2
 π π   π π 
 − J1 − Y0  +  i Y1 − J 0  (33)
 2 2   2 2 
− J1 + iY1
=
− (J1 + Y0 ) + i (Y1 − J0 )
where J1, Y1 are the first and second kind Bessel functions. Also, the Theodorsen function
can be expressed as
H1(2) (k)
C(k) = (2) (34)
H1 (k) + iH0(2) (k)
where H1, H0 are the Hankel functions.

for i=1:5001,ri=(i-1)*0.001;yr(i)=theod(ri);xr(i)=ri;end
plot(xr,real(yr)),ylabel('F'),xlabel('k'),axis([0 5 0 1.2]),grid
plot(xr,imag(yr)),ylabel('G'),xlabel('k'),axis([0 5 -0.3 0]),grid

function thd = theod(rk)

% Theodorsen Function Calculation using modified bessel function
if(rk == 0), thd=1;
else
i=sqrt(-1);thd=BESSELK(1,i*rk)/(BESSELK(0,i*rk)+BESSELK(1,i*rk));
end

© 2001 by C. Nam, Y. Kim 8

 1 .2 0

1 -0 .0 5

0 .8 -0 .1

0 .6 -0 .1 5
F

G
0 .4 -0 .2

0 .2 -0 .2 5

0
0 0 .5 1 1 .5 2 2 .5 3 3 .5 4 4 .5 5 0 0 .5 1 1 .5 2 2 .5 3 3 .5 4 4 .5 5
k k

Figure Real and Imaginary part of the Theodorsen function

The Theodorsen function is frequentlyreplaced by simple algebraic approximations such

as
0.0075 0.10055
C(k) = 0.5 + +
jk + 0.0455 jk + 0.3

Figure Comparison of the Theodorsen function.

The total force and moment resulting from the noncirculatory and circulatory flows are
expressed as
[
F = FNonCir + FΓ = −πρb 2 h&& + Vα& − baα&& − 2πρVbQC (k ) ] (35)

© 2001 by C. Nam, Y. Kim 9

  1  1    1
M = πρb 2 bah&& − Vb − a α& − b 2  + a 2 α&& + 2πρVb 2  a + QC (k ) (36)
 2  8    2
where
1 
Q = Vα + h& + α&b − a 
2 

If we assume the quasi-steady aerodynamics (The aerodynamic characteristics of an

airfoil whose motion consists of variable linear and angular motions are equal, at any
instant of time, to the characteristics of the same airfoil moving with constant linear and
angular velocities equal to the actual instantaneous values.), then C(k) becomes one, and
the force and moment will be

[ ]  1 
FQS = −πρb 2 h&& + Vα& − baα&& − 2πρVb Vα + h& + α&b − a  (37)
 2 
 1  1  
M QS = πρb 2 bah&& − Vb − a α& − b 2  + a 2 α&&
 2  8  
(38)
 1  1 
+ 2πρVb 2  a +  Vα + h& + α&b − a 
 2  2 

3.2. Flutter Equation of the Typical Section Model

3.2.a Equation of Motion

Consider the typical section shown in Figure.

© 2001 by C. Nam, Y. Kim 10

The model has a translational spring with stiffness Kh and a torsion spring, with stiffness
KT. These springs are attached to the airfoil at the shear center. Therefore, it is two
degrees of freedom model (h, α). And h is measured at the shear center (elastic axis).
The downward displacement of any other point on the airfoil is

z = h + xα (39)
where x is a distance measured from the shear center.
The strain energy and the kinetic energy are respectively given by

1 1
U= KTα 2 + K h h 2 (40)
2 2
1 (41)
T=
2 ∫ ρ z& 2 dx

where ρ be the mass per unit length of the airfoil.

T=
2
(
1 &2
h ∫ ρdx + 2h&α& ∫ ρxdx + α& 2 ∫ ρx 2 dx ) (42)

Let’s define the followings.

Mass m = ∫ ρdx

The second moment of inertia of the airfoil about shear center, I θ = ∫ ρx 2 dx = mrθ2

The first moment of inertia of the airfoil about shear center, Sθ = ∫ ρxdx = mxθ

where rθ is the radius of gyration and xθ is a distance from the coordinate to the mass
center.
Then, the kinetic energy can be rewritten as follows:

1 &2 1
T= mh + mxθ h&α& + I θ α& 2 (43)
2 2
The virtual work due to the unsteady aerodynamic forces is

© 2001 by C. Nam, Y. Kim 11

 δWa = ∫ ∆pδz dx = ∫ ∆p{δh + xδα }dx = Qh δh + Qα δα (44)

where the force Qh is positive downward, and moment Qα is positive nose-up.

Lagrange's equations provide the equations of motion of the airfoil.

d  ∂ (T − U )  ∂ (T − U )
 − = Qq , where q=h, α (45)
dt  ∂q&  ∂q

 m mxθ   h&&  K h 0   h  Qh   F 

mx 2  +  = = 
 θ mrθ  α&&  0 K T  α  Qθ  M 

1 xθ  h&& / b ω h2 0  h / b  F / mb 
x 2  +  =  (46)
 θ rθ   α&&   0 ω θ rθ 2   α  M / mb 2 
2

where ω 2h = Kh / m , ωθ2 = KT / Iθ , xθ = xθ / b , rθ = rθ / b .

iωt
If we assume the harmonic motion h = h0 e ,α = α 0 e iωt , the equation of motion will be

1 xθ  h / b ω h2 0  h / b  F / mb 
−ω2  2  +   =  (47)
 xθ rθ   α   0 ω θ2 rθ 2   α  M / mb 2 

The unsteady aerodynamic force and moment will be

[ ]  1  
F = −πρb 2 h&& + Vα& − baα&& − 2πρVbC (k ) Vα + h& + b − a α& 
 2  
h  1  1 1 1  C  (48)
= πρb 3ω 2  1 − i 2C  + α  − a − i − 2C 2 − i 2 − a  
b  k  k k 2  k 
h  1  
= πρb 3ω 2  Lh + α  Lα −  + a  Lh 
b  2  
ωb 1 1 1 + 2C 2C
where the reduced frequency is k = , and Lh = 1 − i 2C , Lα = − i − 2 .
V k 2 k k

Similarily,

© 2001 by C. Nam, Y. Kim 12

  1   h  1  1 
2
 
M = πρb ω M h −  + a  Lh  + M α −  + a (Lα + M h ) +  + a  Lh α 
4 2

 2   b  2  2   
(49)
1 3 1
where M h = , Mα = − i
2 8 k

Then, the equation motion can be rewritten as

1 xθ  h / b ω h2 0  h / b 
−ω2  2  +   
 xθ rθ   α   0 ω θ2 rθ 2   α 
 1   (50)
Lh Lα −  + a  Lh
ω 2
2 
 h / b 

= 2 
µ  1  1  1   α 
 M h −  2 + a  Lh ( )
M α −  + a  Lα + M h +  + a  Lh 
   2  2  
m
where the mass ration is defined as µ = . m is the airfoil mass per length, l.
πρb 2 l

ω2 ωh2
Let’s define Ω 2 = and R 2
= , then
ωθ2 ωθ2

1 xθ  h / b  R 2 0  h / b 
−Ω  2
2  + 2  
 xθ rθ   α   0 rθ   α 
(51)
 1  
2 
Lh Lα −  + a  Lh  h / b 
Ω  2 
= 2
 
µ  1  1  1   α 
 M h −  2 + a  Lh M α −  + a (Lα + M h ) +  + a  Lh 
   2  2  

3.2.b V-g method for flutter analysis

Let’s express the above flutter equation in the following matrix form.

© 2001 by C. Nam, Y. Kim 13

 [K ]hα/ b = Ω [A 2
]h / b 
+ M ij   (52)
α 
ij ij
 
where Kij is the stiffness matrix, Mij mass matrix, and Aij is the aerodynamic matrix. Note
that the aerodynamic matrix is funtion of the reduced frequency, k.
V-g method assumes first the artifial structural damping, g.
[ K ij ] = (1 + ig )[ K ij ] (53)

This artificial damping indicates the required damping for the harmonic motion. The
eigenvalues of the equation of motion represent a point on the flutter boundary if the
corresponding value of g equals the assumed value of g.

ωb
For give reduced frequency, k = , it is a complex eigenvalue problem.
V
(1 + ig ) h / b  h / b 
[ K ij ]  = [ Aij + M ij ]  (54)
Ω α  α 
2

The eigenvalue is
1 + ig
λ= (55)
Ω2
From this eigenvalue we have
1 ω i2
= 2 (56)
λ Re ωθ
λ Im
g= (57)
λ Re
The complex eigenvalue problem is solved beginning with large values of k and then
decreasing k until a flutter velocity is found. If there is no actual damping in the system,
when the artificial damping, g, first becomes positive, flutter will occur.
Example
Calculate the flutter speed of the following typical section model.
xθ =0.2, rθ =0.5, µ=20, a=-0.1, b=1, and R=0.3

% 2 dof system
% xth ; x _ theta

© 2001 by C. Nam, Y. Kim 14

% rth2 ; ( r _theta)^2
% R2 ; (omega h /omega alpha)^2
xth =0.2; R2 =(0.3)^2;
rth2=(0.5)^2; mu =20.0 ; a =-0.1 ;
%
i=sqrt(-1);
result1=zeros(2,400); result2=zeros(2,400); veloci=zeros(2,400);
for kk=400:-1:1; rk=kk*0.01;
%
% Theodorsen Function Calculation
C=0.5+0.0075/(i*rk+0.0455)+0.10055/(i*rk+0.3);
% Mass
ms=[ 1 xth
xth rth2];
%
% stiffness matrix
ks=[ R2 0
0 rth2];
%
lh=1-i*2*C/rk; la=.5-i*(1+2*C)/rk-2*C/rk/rk;
mh=0.5; ma=3/8-i/rk;
aero=[ lh la-(.5-a)*lh
mh-(.5+a)*lh ma-(.5+a)*(la+mh)+(.5+a)*(.5+a)*lh]/mu;

ddd=eig(inv(ks)*(aero+ms)); rrr=real(ddd); iii= imag(ddd);

rst1(:,kk)=sqrt(1./rrr); rst2(:,kk)=iii./rrr;
vel(:,kk)=sqrt(1./rrr)/rk;
end

figure(1);
plot(vel(1,:),rst1(1,:),'.r',vel(2,:),rst1(2,:),'.r'),
axis([0.0 5.00 0 1.60]),xlabel('Velocity'),ylabel('Frequency Ratio'),grid;
figure(2);
plot(vel(1,:),rst2(1,:),'.r',vel(2,:),rst2(2,:),'.r'),
axis([0.0 5.00 -1.2 1.0]),xlabel('Velocity'),ylabel('g'),grid;

1.6 1

0.8
1.4
0.6
1.2
0.4
Frequency Ratio (ω / ω )
θ

1 0.2

0
0.8
g

-0.2

0.6 -0.4

-0.6
0.4
-0.8
0.2
-1

0
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
Velocity (V/ ω b) Velocity (V/ ω b)
θ θ

© 2001 by C. Nam, Y. Kim 15

 ωi V
Figure Plots of and g versus
ωθ ωθ b
VFlutter ω
Flutter speed of the model is = 1.99 , frequency ration Flutter = 0.58 . From V-g
ωθ b ωθ
plot results, Can you estimate the divergence speed of this airfoil ?

3.3. Typical Section with the Trailing Edge Flap

Let’s put the trailing edge flap at the airfoil that we discussed in the previous section.
The trailing edge flap is hinged at x=bc. Again the airfoil has a translational spring as
well as the torsional spring at the shear center. The same way that we discussed at the
previous section, we can calculate the unsteady aerodynamic forces.

Figure Typical Section with the Trailing Edge Control Surface

The displacement of the airfoil at x will be

z = h + α ( x − ab) + β (x − bc)U( x − bc) (58)
where x is the distance measured from the midchord and U(x-bc) is the unit step function.
Then, the upwash will be
 ∂z ∂z 
wa ( x, t ) = − + V 
 ∂t ∂x  (59)
= −[h& + α& ( x − ab)] − Vα − β& ( x − bc)U ( x − bc) − Vβ U ( x − bc)
The total force and moment resulting from the noncirculatory and circulatory flows are
expressed as

© 2001 by C. Nam, Y. Kim 16

  V b 
F = −πρb 2 h&& + Vα& − baα&& − T4 β& − T1 β&& − 2πρVbQC (k ) (60)
 π π 
 1  1  V2
M = πρb 2 bah&& − Vb − a α& − b 2  + a 2 α&& − (T4 + T10 ) β
 2  8  π
Vb  1  & b2 
+ − T1 + T8 + (c − a)T4 − T11  β + {T7 + (c − a)T1 }β&& (61)
π  2  π 
 1
+ 2πρVb 2  a + QC (k )
 2
b Vb  1 
M β = πρb 2  T1 h&& + 2T9 + T1 − (a − )T4 α&
π π  2 
2
2b 2 V 
− T13α&& −   (T5 − T4T10 ) β (62)
π π 
2

+
Vb
T T & +  b  T β&& − ρVb 2T QC (k )
β
4 11
π 
3 12
2π 2 
where
1  V b
Q = Vα + h& + α&b − a  + T10 β + T11 β& (63)
2  π 2π
T functions are defined as follows.
2 + c2
T1 = - 1 − c 2 + ccos −1 c
3
1-c 2 1 1
T3 = - ( 5c 2 + 4 ) + c( 7 + 2c 2 ) 1 − c 2 cos −1 c-( + c 2 )(cos −1 c) 2
8 4 8
2 1
T4 = c 1− c -cos− c
T5 = -(1-c2 )-(cos −1 c) 2 + 2c 1 − c2 cos −1 c
2
7 + 2c 2 1 2 1
T7 = c 1 − c -( + c )cos− c
8 8
1
T8 = - (1 + 2c 2 ) 1 − c 2 + c cos−1 c
3

© 2001 by C. Nam, Y. Kim 17

 1  1 − c 2 (1-c 2 ) 
T9 = + aT4 
2  3 


T10 = 1 − c2 + cos−1 c

T11 = (2-c) 1− c2 + (1-2c)cos −1 c

T12 = (2 + c) 1 − c 2 -(1+ 2c)cos−1 c
1
T13 = - [T + (c-a)T1]
2 7
T15 = T4 + T10
1
T16 = T1-T8 -(c-a)T4 + T11
2
1
T17 = -2T9 -T1 + (a- )T4
2
T18 = T5-T4 T10
1
T19 = - T4 T11
2

Note that Q can be rewritten as the following form by using S1 and S2

Q = V S1 {x s } + b S 2 {x& s } (64)
where
 T 
S1 = 0 1 10 
 π 
 T 
S 2 = 1 (0.5 − a) 11 
 2π 
h / b 
 
xs =  α 
 β 
 

In matrix form, the aerodynamic forces and moments will be

© 2001 by C. Nam, Y. Kim 18

  F /b 
 2   2b
 M / b  = q 2C (k ){R}S1 {x s } + C (k ){R}S 2 {x& s }
M / b 2   V
 β  (65)
2b 2 2b 
+ 2 [ M nc ]{&x&s } + [ Bnc ]{x& s } + 2[ K nc ]{x s }
V V 
where
 − 2π 
 
R = 2π (a + 0.5)
 −T 
 12 
The subscript nc indicates the non-circulatory part.

 
- π πa T1 
 1 
M nc = πa − π ( + a 2 ) − 2T13 
 8 
T T3 
− 2T13
 1 π 
 
0 −π T4 
 1 
Bnc = 0 π (a − ) − T16 
 2 
0 T19 
− T17 −
 π 
 
0 0 0 
K nc = 0 0 − T15 
 
0 0 − T18 
 π 

The equation of motion of the airfoil are derived from Lagrange’s equations.
d  ∂ (T − U )  ∂ (T − U )
 − = Qq , where q=h, α, β (66)
dt  ∂q&  ∂q
The kinetic energy is

© 2001 by C. Nam, Y. Kim 19

 1 2
2∫
T= z& ρ dx (67)
where airfoil deflection is
z = −h − xα − ( x − (bc − ba) )β U ( x − (bc − ba ) ) (68)
It has to be noted that the coordinate is placed at the shear center.

T = mh& 2 + mxθ h&α& + I θ α& 2 + I β β& 2 + mx β h&β& + (I β + mx β (bc − ba) )α&β& (69)
1 1 1
2 2 2
The potential energy U is stored in springs attached at the control surface hinge lines and
at the elastic axis.
1
U = (K h h 2 + KTα 2 + K β β 2 ) (70)
2
 m mxθ mx β   h&& 
  
mrθ + mx β (bc − ba) α&& 
2 2
 mxθ mrθ
mx β mrθ 2 + mx β (bc − ba) mrβ
2
  β&&
  
(71)
K h 0 0  h   F 
    
+  0 KT 0  α  =  M 
0
 0 K β   β  M β 

© 2001 by C. Nam, Y. Kim 20

 1 xθ xβ  h&& / b ω h2 0 0  h / b 
     
rθ + x β (c − a)  α&&  +  0 ω θ rθ 0  α 
2 2 2 2
 xθ rθ
 xβ 2
rθ + x β (c − a) rβ
2
  β&&   0 0 ω β2 rβ 2   β 
   
  − 2π  h / b 
q   T10   
= 2C (k )2π (a + 0.5)0 1  α 
m  − T  π  
   12   β 
 − 2π  h& / b
2b   T11   
+ 2π (a + 0.5)1 (0.5 − a )   α& 
V   2π   & 
 − T12   β 
 
- π πa T1  h&& / b
2b 2  1  
+ 2 πa − π ( + a 2 ) − 2T13   α&& 
V  8  
T T3   β&& 
− 2T
 1 13
π 
 
0 −π T4  h& / b
2b  1  
+ 0 π (a − ) − T16   α& 
V  2  
0 T19   β& 
− T17 − (72)
 π 
  
0 0 0  h / b 
 
+ 2 0 0 − T15   α 
 
0 0 − T18   β 
 π  

Or

© 2001 by C. Nam, Y. Kim 21

 q 2b
[ M ]{&x&s } + [ K ]{x s } =  2C (k ){R}S1 {x s } + C (k ){R}S 2 {x& s }
m V
(73)
2b 2 2b 
+ 2 [ M nc ]{&x&s } + [ Bnc ]{x& s } + 2[ K nc ]{x s }
V V 
V-g method for flutter analysis
Let’s assume the harmonic motion of the system, x s = x s 0 e iωt
The equation of motion will be expressed in the following matrix form.
q 2b
− ω 2 [ M ]{x s } + [ K ]{x s } =  2C (k ){R}S1 {x s } + (iω ) C (k ){R}S 2 {x s }
m V
(74)
2b 2 2b 
+ (−ω ) 2 [ M nc ]{x s } + (iω ) [ Bnc ]{x s } + 2[ K nc ]{x s }
2

V V 

1 1 1 1
[ K ]{x s } = [ M ]{x s } +  2 C (k ){R}S1 {x s } + i C (k ){R}S 2 {x s }
ω 2
µ k k
(75)
1 1 
− [ M nc ]{x s } + i [ Bnc ]{x s } + 2 [ K nc ]{x s }
k k 

where µ is the mass ratio, k is the reduced frequency.

In V-g method, the artifial structural damping, g is assumed.
1 + ig 1 1 1
[ K ]{x s } = [ M ]{x s } +  2 C (k ){R}S1 {x s } + i C (k ){R}S 2 {x s }
ω 2
µ k k
(76)
1 1 
− [ M nc ]{x s } + i [ Bnc ]{x s } + 2 [ K nc ]{x s }
k k 
ωb
Again, it’s a complex eigenvalue problem. For give reduced frequency, k = ,
V
1+ ig
the eigenvalue, λ = is obtained, and we will have
ω2
1
= ω i2
λ Re

© 2001 by C. Nam, Y. Kim 22

 λ Im
g=
λ Re

% 3 dof system
wh=50.0; wa=100.0; wb=300.0;
a=-0.4; c=0.6; b=1;
xa=0.2; xb=0.0125;
ra=sqrt(0.25); rb=sqrt(0.00625);
mu=40;

nn=3;i=sqrt(-1);rho=0.002378;
mas=pi*rho*mu*b*b; damb=0.0;
sqr=sqrt(1-c*c);
arc=acos(c);
%
%
% final version of t function generating the aero matrix
t1=-(2+c*c)/3*sqr+c*arc;
t3=-(1-c*c)/8*(5.*c*c+4)+.25*c*(7+2*c*c)*sqr*arc-(1./8.+c*c)*arc*arc;
t4=c*sqr-arc;
t5=-(1-c*c)-arc*arc+2*c*sqr*arc;
t7=c*(7+2*c*c)/8*sqr-(1/8+c*c)*arc;
t8=-1/3*(1+2*c*c)*sqr+c*arc;
t9=.5*(sqr*(1-c*c)/3+a*t4);
t10=sqr+arc;
t11=(2-c)*sqr+(1-2*c)*arc;
t12=(2+c)*sqr-(1+2*c)*arc;
t13=-.5*(t7+(c-a)*t1);
t15=t4+t10;
t16=t1-t8-(c-a)*t4+0.5*t11;
t17=-2*t9-t1+(a-.5)*t4;
t18=t5-t4*t10;
t19=-.5*t4*t11;
%
iden=zeros(nn,nn);ks=zeros(nn,nn);knc=zeros(nn,nn);bs=zeros(nn,nn);
for ii=1:nn
iden(ii,ii)=1.0;
end
% mass matrix
ms(1,1)=1; ms(1,2)=xa; ms(1,3)=xb;
ms(2,1)=xa; ms(2,2)=ra*ra; ms(2,3)=rb*rb+xb*(c-a);
ms(3,1)=xb; ms(3,2)=rb*rb+xb*(c-a); ms(3,3)=rb*rb;
%
%
mnc(1,1)=-pi; mnc(1,2)=pi*a; mnc(1,3)=t1;
mnc(2,1)=pi*a; mnc(2,2)=-pi*(1./8.+a*a); mnc(2,3)=-2*t13;
mnc(3,1)=t1; mnc(3,2)=-2*t13; mnc(3,3)=t3/pi;

© 2001 by C. Nam, Y. Kim 23

%
%
% stiffness matrix
ks(1,1)=wh*wh;
ks(2,2)=ra*ra*wa*wa;
ks(3,3)=rb*rb*wb*wb;
%
%
knc(2,3)=-t15;
knc(3,3)=-t18/pi;
%
%
%
bnc(1,1)=0; bnc(1,2)=-pi; bnc(1,3)=t4;
bnc(2,1)=0; bnc(2,2)=pi*(a-.5); bnc(2,3)=-t16;
bnc(3,1)=0; bnc(3,2)=-t17; bnc(3,3)=-t19/pi;
%
%
%
r1=[ -2.*pi
2*pi*(a+0.5)
-t12];
%
s1=[0 1 t10/pi];
s2=[1 (.5-a) t11/(2*pi)];
%
% v-g method

m=200;
rst1=zeros(3,m); rst2=zeros(3,m); vel=zeros(3,m);
for kk=m:-1:1;
rk=kk*0.02;
[f,g]=faero(rk,b,mnc,bnc,knc,r1,s1,s2,mu);
aero=f+i*g;
ddd=eig(inv(ks)*(ms+aero));
rrr=abs(real(ddd)); iii= imag(ddd);
rst1(:,kk)=sqrt(1./rrr); rst2(:,kk)=iii./rrr;
vel(:,kk)=sqrt(1./rrr)*b/rk;
end

xxx=[vel(1,:); rst1(1,:); rst2(1,:); vel(2,:); rst1(2,:); rst2(2,:);

vel(3,:); rst1(3,:); rst2(3,:)]';
figure(1);
plot(vel(1,:),rst1(1,:),'*r',vel(2,:),rst1(2,:),'*r',vel(3,:),rst1(3,:)
,'*r'),
axis([0.0 500 0 500]),xlabel('Velocity'),ylabel('Frequency'),grid;
figure(2);
plot(vel(1,:),rst2(1,:),'*r',vel(2,:),rst2(2,:),'*r',vel(3,:),rst2(3,:)
,'*r'),
axis([0.0 500 -0.5 0.5]),xlabel('Velocity'),ylabel('g'),grid;

© 2001 by C. Nam, Y. Kim 24

 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% file : faero.m
% aero forces in frequency domain
function [f,g]=faero(rk,b,mnc,bnc,knc,r1,s1,s2,mu)
i=sqrt(-1);cs=theod(rk);
z=-mnc+i*(bnc+cs*r1*s2)/rk+(knc+cs*r1*s1)/(rk^2);
z=z/(mu*pi);f=real(z);g=imag(z);
%

Example

Calculate the flutter speed of the following typical section model.

ωh = 50.0 rad/ sec , ωθ = 100.0 rad/ sec , ω β = 300.0rad/ sec

a = -0.4, c = 0.6, b = 1, xθ = 0.2, x β = 0.0125, rθ2 = 0.25, rβ2 = 0.00625, µ = 40

500 0.4

0.3
400 0.2
0.1
300 0
Frequency

-0.1
g

200 -0.2
-0.3
100 -0.4

-0.5
0
0 100 200 300 400 500 0 100 200 300 400 500
Velocity Velocity
Frequency versue Velocity g versus Velocity
Figure . Flutter Result

3.4. Rational Function Approximation of the Unsteady Aerodynamic Forces

For the aeroservoelastic analysis and design, it is necessary to transform the equations of
motion into the state space form. This requires to approximate the frequency domain
unsteady aerodynamic forces in terms of rational functions of the Laplace variable.
There are several methods for the rational function approximation (RFA) such as Roger's
method, Matrix Pade method, and Minimum State method. However, when the RFA is

© 2001 by C. Nam, Y. Kim 25

conducted, it causes an increase in the total number of states due to the number of
augmented aerodynamic states to represent unsteady aerodynamic forces accurately.

In this section, we will discuss about Roger’s method since Roger’s method is very
simple and accurately transforms the unsteady aerodynamic forces from frequency
domain into time domain. The minimum state method will be discussed in the next
chapter,

First, let’s express the unsteady aerodynamic forces in the frequency domain as follows;

 F /b 
 2   2b
 M / b  = q 2C (k ){R}S1 {xs } + C (k ){R}S 2 {x& s }
M / b 2  V
 β 
2b 2 2b 
+ 2 [ M nc ]{&x&s } + [ Bnc ]{x& s } + 2[ K nc ]{xs }
V V 
[
= q 2C (k ){R}S1  + i 2kC (k ){R}S 2  − 2k 2 [ M nc ] + i 2k[ Bnc ] + 2[ K nc ] {xs } (77) ]
= q[2C (k ){R}S  + 2 s ' C (k ){R}S  + 2 s '
1 2
2
[ M nc ] + 2 s ' [ Bnc ] + 2[ K nc ]]{x }
s

= qA( s ' ){xs }

where s’ is the nondimensionalized Laplace variable s' = ik = sb / V

Roger’s method approximates the unsteady aerodynamic forces in the following form.

[A ] = [P ] + [P ]s'+[P ]s' +∑ s['+Pγ]s'

N
2 j
ap 0 1 2 (78)
j =3 j −2

γ j −2 is the aerodynamic poles which are usually preselected in the range of reduced
frequencies of interest.

Let’s define the calculated aerodynamic forces as [A(s')] = [F(s')] + i[G(s')] .

The real and imaginary part of the approximated aerodynamic matrix will be

© 2001 by C. Nam, Y. Kim 26

 [ ]
Real part of Aap = [P0 ] + [P2 ]s ' + ∑ 2
N [P ](− s'
j
2
)
(79)
j =3 (− s ' ) + γ
2 2
j −2

[ ]
Imaginary part of Aap = [P1 ]s'+ ∑
N [P ]γ
j j −2 s'
(80)
j =3 (− s' ) + γ 2j − 2
2

Note that the nondimensionalized Laplace variable is s’=ik.

The real matrices [Pj] are determined using least square technique for a term by term
fitting of the aerodynamic matrix, [A(s’)].

∑[ Aaero] [ Aaero]{ Xp} = ∑[ Aaero] {Baero}

i
T

i
T
(81)

where

 ki2 
1 0 − ki2 ... 
ki2 + γ 2j − 2 
[Aaero] =  γ j − 2 ki 

 0 ki 0 ...
ki2 + γ 2j − 2 


 F ( ki ) 
{Baero} =  
G (ki ) mn

 P0 
P 
 1 
{Xp} =  P2 
: 
 
PN mn

The augmented aerodynamic state is defined as follows.

s' s
{xaj } = {xs } = {xs } (82)
s'+γ j − 2 V
s + γ j −2
b

© 2001 by C. Nam, Y. Kim 27

 V
{x&aj } = {x&s } − γ j − 2{xaj } (83)
b

Then, the state space equation of motion is expressed in the following form.

 0 I 0 .. 0 
 x&s  − M −1K x 
 &x&   −M B −1 −1
qM P3 .. qM PN   s 
−1

 xs 
&
 s   0 V  
I −  γ 1I .. 0
 x&a 3  =  b   xa 3  (84)
 ..   .. .. .. .. ..   .. 
    
 0 V 
 x&aN  I 0 0 −  γ N − 2 I   xaN 
 b 

where
2
b
M = M − qP2  
V 

b
B = B − qP1  
V 

K = K − qP0

Example

Using the state space equation, calculate the flutter speed of the following typical section
model. Assume four aerodynamic poles for the rational function approximation using
Roger’s method. (γ=0.2, 0.4, 0.6, 0.8)
ω h = 50.0 rad/sec, ω θ = 100.0 rad/sec, ω β = 300.0rad/sec
a = -0.4, c = 0.6,b = 1, xθ = 0.2, x β = 0.0125,rθ2 = 0.25, rβ2 = 0.00625, µ = 40

© 2001 by C. Nam, Y. Kim 28

 0 5 1

-2 0
0
-4
-5
Imag of A(1,1)

Imag of A(1,2)

Imag of A(1,3)
-1
-6
-10
-8
-2
-15
-10
-3
-12 -20

-14 -25 -4
-5 0 5 10 15 20 25 -15 -10 -5 0 5 -6.5 -6 -5.5 -5 -4.5 -4 -3.5 -3
Real of A(1,1) Real of A(1,2) Real of A(1,3)

1.4 0 0

1.2 -2 -0.5

1
-4 -1
Imag of A(2,1)

Imag of A(2,2)

Imag of A(2,3)
0.8
-6 -1.5
0.6
-8 -2
0.4

0.2 -10 -2.5

0 -12 -3
0 2 4 6 8 10 12 1 2 3 4 5 6 7 8 -2.2 -2.1 -2 -1.9 -1.8 -1.7
Real of A(2,1) Real of A(2,2) Real of A(2,3)

0 0 0

-0.1 -0.05
-0.02
-0.2 -0.1
Imag of A(3,2)
Imag of A(3,1)

-0.04 Imag of A(3,3)

-0.3 -0.15

-0.4 -0.2
-0.06

-0.5 -0.25
-0.08
-0.6 -0.3

-0.1 -0.7 -0.35

-0.1 0 0.1 0.2 0.3 0.4 0.5 0.6 -0.1 0 0.1 0.2 0.3 0.4 0.5 0.6 -0.16 -0.14 -0.12 -0.1 -0.08 -0.06
Real of A(3,1) Real of A(3,2) Real of A(3,3)

Figure Rational function approximation results

© 2001 by C. Nam, Y. Kim 29

 400
Imag. Part(rad/sec)

Open Loop Flutter Analysis

300 3 Degree of Freedom
Roger Approx.
200

100

0
-100 -50 0
Real Part
20 400

Imag. Part(rad/sec)
0
300
Real Part

-20
200
-40
100
-60

-80 0
0 100 200 300 400 0 100 200 300 400
Velocity Velocity

The result shows that the model has a flutter speed of about 300 ft/sec, and the flutter
frequency is around 78 rad/sec, which is the pitching mode dominant.

© 2001 by C. Nam, Y. Kim 30

% function [result]= roger3(wh,wa,wb,a,c,xa,xb)
% version 1992 1 6
% 3 dof system
wh=50.0; wa=100.0; wb=300.0;
a=-0.4; c=0.6; b=1;
xa=0.2; xb=0.0125;
ra=sqrt(0.25); rb=sqrt(0.00625);
mu=40;

nn=3;i=sqrt(-1);rho=0.002378;
mas=pi*rho*mu*b*b; damb=0.0;
sqr=sqrt(1-c*c);
arc=acos(c);
%
%
% final version of t function generating the aero matrix
t1=-(2+c*c)/3*sqr+c*arc;
t3=-(1-c*c)/8*(5.*c*c+4)+.25*c*(7+2*c*c)*sqr*arc-(1./8.+c*c)*arc*arc;
t4=c*sqr-arc;
t5=-(1-c*c)-arc*arc+2*c*sqr*arc;
t7=c*(7+2*c*c)/8*sqr-(1/8+c*c)*arc;
t8=-1/3*(1+2*c*c)*sqr+c*arc;
t9=.5*(sqr*(1-c*c)/3+a*t4);
t10=sqr+arc;
t11=(2-c)*sqr+(1-2*c)*arc;
t12=(2+c)*sqr-(1+2*c)*arc;
t13=-.5*(t7+(c-a)*t1);
t15=t4+t10;
t16=t1-t8-(c-a)*t4+0.5*t11;
t17=-2*t9-t1+(a-.5)*t4;
t18=t5-t4*t10;
t19=-.5*t4*t11;
%
iden=zeros(nn,nn);ks=zeros(nn,nn);knc=zeros(nn,nn);bs=zeros(nn,nn);
for ii=1:nn
iden(ii,ii)=1.0;
end
% mass matrix
ms(1,1)=1; ms(1,2)=xa; ms(1,3)=xb;
ms(2,1)=xa; ms(2,2)=ra*ra; ms(2,3)=rb*rb+xb*(c-a);
ms(3,1)=xb; ms(3,2)=rb*rb+xb*(c-a); ms(3,3)=rb*rb;
ms=ms*mas
%
%
mnc(1,1)=-pi; mnc(1,2)=pi*a; mnc(1,3)=t1;
mnc(2,1)=pi*a; mnc(2,2)=-pi*(1./8.+a*a); mnc(2,3)=-2*t13;
mnc(3,1)=t1; mnc(3,2)=-2*t13; mnc(3,3)=t3/pi;
%

© 2001 by C. Nam, Y. Kim 31

%
% stiffness matrix
ks(1,1)=wh*wh;
ks(2,2)=ra*ra*wa*wa;
ks(3,3)=rb*rb*wb*wb;
ks=ks*mas
%
%
knc(2,3)=-t15;
knc(3,3)=-t18/pi;
%
%
bs(3,3)=2*rb*rb*wb*damb;
bs=bs*mas
%
%
bnc(1,1)=0; bnc(1,2)=-pi; bnc(1,3)=t4;
bnc(2,1)=0; bnc(2,2)=pi*(a-.5); bnc(2,3)=-t16;
bnc(3,1)=0; bnc(3,2)=-t17; bnc(3,3)=-t19/pi;
%
%
%
r1=[ -2.*pi
2*pi*(a+0.5)
-t12];
%
s1=[0 1 t10/pi];
s2=[1 (.5-a) t11/(2*pi)];
%
%
% Roger approximation of aero forces
rka=[0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.6 0.7 0.8 0.9 1.0
1.2 1.4 1.6 1.8 2.0]';
[m,n]=size(rka);
%
gam1=0.2; gam2=0.4; gam3=0.6; gam4=0.8;

for ii=1:nn
for jj=1:nn
tempa=0; tempb=0;
for jjrk=1:m
rk=rka(jjrk);
[fl,gl]=aero(rk,b,mnc,bnc,knc,r1,s1,s2);
flij=fl(ii,jj); glij=gl(ii,jj);
al=[ 1 0
0 rk
-rk*rk 0
rk*rk/(rk*rk+gam1*gam1) rk*gam1/(rk*rk+gam1*gam1)
rk*rk/(rk*rk+gam2*gam2) rk*gam2/(rk*rk+gam2*gam2)

© 2001 by C. Nam, Y. Kim 32

 rk*rk/(rk*rk+gam3*gam3) rk*gam3/(rk*rk+gam3*gam3)
rk*rk/(rk*rk+gam4*gam4) rk*gam4/(rk*rk+gam4*gam4) ]';

tempa=tempa+al'*al;
tempb=tempb+al'*[flij glij]';
end
xe=inv(tempa)*tempb;
p0(ii,jj)=xe(1);
p1(ii,jj)=xe(2);
p2(ii,jj)=xe(3);
p3(ii,jj)=xe(4);
p4(ii,jj)=xe(5);
p5(ii,jj)=xe(6);
p6(ii,jj)=xe(7);

end
end

% For ploting approximated Aero. Matrix

for jj=1:m
rk=rka(jj);
irk=i*rk;
rkg(jj)=rka(jj);
[fl,gl]=aero(rk,b,mnc,bnc,knc,r1,s1,s2);
zap=p0+p1*irk+p2*irk*irk+p3*irk/(irk+gam1)+p4*irk/(irk+gam2)+p5*irk/(ir
k+gam3)+p6*irk/(irk+gam4);
y1f(jj)=fl(1,2);
y1g(jj)=gl(1,2);
y2f(jj)=real(zap(1,2));
y2g(jj)=imag(zap(1,2));
end
%subplot(221),xlabel('Reduced Frequency')
%subplot(221),ylabel('Real of A(1,2)')
%subplot(221),plot(rkg,y1f,'.r',rkg,y2f,'.g')
%subplot(222),xlabel('Reduced Frequency')
%subplot(222),ylabel('Imag. of A(1,2)')
%subplot(222),plot(rkg,y1g,'.r',rkg,y2g,'.g')
%subplot(223),xlabel('Real of A')
%subplot(223),ylabel('Imag of A')
%subplot(223),plot(y1f,y1g,'.r',y2f,y2g,'.g')
%
%pause
%clg

iter=0;
ivf=0;

for kk=1:40;
state=0;

© 2001 by C. Nam, Y. Kim 33

vel=10.*(kk-1)+10.0;
q=.5*rho*vel*vel;
invms=inv(ms-.5*rho*b*b*p2);

ans1=-invms*(ks-q*p0);
ans2=-invms*(bs-.5*rho*b*vel*p1);
ans3=q*invms*p3;
ans4=q*invms*p4;
ans5=q*invms*p5;
ans6=q*invms*p6;
for ii=1:nn
for jj=1:nn
state(ii,jj+nn)=iden(ii,jj);
state(ii+2*nn,jj+nn)=iden(ii,jj);
state(ii+3*nn,jj+nn)=iden(ii,jj);
state(ii+4*nn,jj+nn)=iden(ii,jj);
state(ii+5*nn,jj+nn)=iden(ii,jj);
state(ii+nn,jj)=ans1(ii,jj);
state(ii+nn,jj+nn)=ans2(ii,jj);
state(ii+nn,jj+2*nn)=ans3(ii,jj);
state(ii+nn,jj+3*nn)=ans4(ii,jj);
state(ii+nn,jj+4*nn)=ans5(ii,jj);
state(ii+nn,jj+5*nn)=ans6(ii,jj);
state(ii+2*nn,jj+2*nn)=-(vel/b)*gam1*iden(ii,jj);
state(ii+3*nn,jj+3*nn)=-(vel/b)*gam2*iden(ii,jj);
state(ii+4*nn,jj+4*nn)=-(vel/b)*gam3*iden(ii,jj);
state(ii+5*nn,jj+5*nn)=-(vel/b)*gam4*iden(ii,jj);
end
end

eig(state);
for ii=1:6*nn;
if imag(ans(ii)) >= 0.0,
iter=iter+1;
vsol(iter)=vel;
xsol(iter)=real(ans(ii));
ysol(iter)=imag(ans(ii));

if(xsol(iter) > 0.0),

if(ivf==0),
ivf=1;
velf=vel;
end
end

end
end
end
vsol=vsol';

© 2001 by C. Nam, Y. Kim 34

xsol=xsol';
ysol=ysol';
m=size(xsol);
for i=1:m;
result(i,1)=vsol(i);
result(i,2)=xsol(i);
result(i,3)=ysol(i);
end
save open3.out result /ascii ;
figure(1);
plot(xsol,ysol,'or'),xlabel('Real Part'),ylabel('Imag. Part'),grid;
figure(2);
subplot(221),plot(xsol,ysol,'.r'),xlabel('Real Part'),
ylabel('Imag. Part'),grid;
subplot(222);axis('off');
text(0.0,0.85,'Open Loop Flutter Analysis','sc'),
text(0.0,0.75,'3 Degree of Freedom','sc'),
text(0.0,0.65,'Roger Approx.','sc');
subplot(223),plot(vsol,xsol,'.r'),xlabel('Velocity'),
ylabel('Real Part'),grid;
subplot(224),plot(vsol,ysol,'.r'),xlabel('Velocity'),
ylabel('Imag. Part'),grid;
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% file : aero.m
% approximation of aero forces
function [f,g]=aero(rk,b,mnc,bnc,knc,r1,s1,s2)
i=sqrt(-1);
cs=theod(rk);
z=(mnc*(i*rk)*(i*rk)+(bnc+cs*r1*s2)*(i*rk)+knc+cs*r1*s1);
z=2*z;
f=real(z);g=imag(z);
%

© 2001 by C. Nam, Y. Kim 35

3.5. Aeroservoelastic Design using LQR/LQG

Using rational function approximation described in the previous section, the aeroelastic
equation of motion is successfully transformed into the linear time invariant state space
equation. The state equation can be expressed into the following form;

 0 I 0 .. 0 
 x& s   x   0 
 &x&  − M K
−1
−M B −1 −1
qM P3 .. qM PN   s   −1
−1

 xs   M B ACT 
&
 s   V  
0 I −  γ 1 I .. 0 β
 x& a 3  =  b   xa 3  +  0
 ..   ..   ..   
.. .. .. ..  0 
   V 
 
 x& aN   0 I 0 0   
−  γ N −2 I  xaN    0 
 b
  

(85)
or
x& = Ax + Bu (86)
where

 0 
 
{B ACT } =  0 
ω 2 r 2 
 θ θ 
Let’s design flutter suppression system using the following airfoil data.
ω h = 50.0 rad/sec, ω θ = 100.0 rad/sec, ω β = 300.0rad/sec
a = -0.4, c = 0.6,b = 1, xθ = 0.2, x β = 0.0125,rθ2 = 0.25, rβ2 = 0.00625, µ = 40

The flutter speed of this model is about 300 ft/sec. Control system design in the state
space requires to fix the design airspeed. Usually the design airspeed is fixed as
arbitrarily. Let’s fix the design airspeed as 320 ft/sec, which is higher than the open loop
flutter speed.
At the design airspeed, the system eigenvalues are

-14.3870 ± i 339.6737
5.0715 ± i 70.9743
-25.5913 ± i 74.9236

There are several control algorithms are available. Here, let’s use the classical LQR
(Linear Quadratic Regulator) theory to design a control system.

© 2001 by C. Nam, Y. Kim 36

Consider the linear time invariant state space equation
x& = Ax + Bu
(87)
y = Cx
where A is the system matrix, B is the control matrix, and C is the output matrix. LQR
theory determines the optimal gain matrix K such that the state-feedback law u=-Kx
subject to minimize the quadratic cost function
1 ∞
J = ∫ ( x T Qx + u T Ru )dt (88)
2 0
where Q and R are the weighting matrices. The corresponding optimal control is given
by
u = − Kx = − R −1 B T Px (89)
where K is optimal feedback gain matrix, which can be obtained from the Riccati matrix,
P. The Riccati matrix is determined by the solution of the following steady state Riccati
equation.

PA − AT P + C T QC − PBR −1 B T P = 0 (90)

The closed loop state equation can be written as

x& = ( A − BK ) x (91)
The controller gives the flowing closed loop eigenvalues.

-107.2088 ± i 314.2080
-5.1270 ± i 73.0405
-49.1710 ± i 64.1661

It can be seen that the controller stabilizes the unstable pitching mode. It is assumed that
Q and R are identity matrices, and C is defined to feedback the plunge and pitching
modes only. Figure shows the closed loop system eigenvalues for different airspeed.

© 2001 by C. Nam, Y. Kim 37

 400
Closed Loop Flutter Analysis
300 3 Degree of Freedom
Imag. Part

Roger Approx.
200

100

0
-150 -100 -50 0 50
Real Part

20 400

0
300
Imag. Part
Real Part

-20
200
-40
100
-60

-80 0
0 100 200 300 400 0 100 200 300 400
Velocity Velocity
Figure Closed Loop Flutter Analysis Results.

close all
clear
% function [result2]= close3(wh,wa,wb,a,c,xa,xb)
% version 1992 1 6
% By nam, changho
% 3 dof system ;roger approx.
wh=50.0; wa=100.0; wb=300.0;
a=-0.4; c=0.5; b=1;
xa=0.2; xb=0.0125;
ra=sqrt(0.25); rb=sqrt(0.00625);
mu=40;

nn=3;i=sqrt(-1);rho=0.002378;
mas=pi*rho*mu*b*b; damb=0.0;
sqr=sqrt(1-c*c);
arc=acos(c);
%
%

© 2001 by C. Nam, Y. Kim 38

%
% final version of t function generating the aero matrix
t1=-(2+c*c)/3*sqr+c*arc;
t3=-(1-c*c)/8*(5.*c*c+4)+.25*c*(7+2*c*c)*sqr*arc-(1./8.+c*c)*arc*arc;
t4=c*sqr-arc;
t5=-(1-c*c)-arc*arc+2*c*sqr*arc;
t7=c*(7+2*c*c)/8*sqr-(1/8+c*c)*arc;
t8=-1/3*(1+2*c*c)*sqr+c*arc;
t9=.5*(sqr*(1-c*c)/3+a*t4);
t10=sqr+arc;
t11=(2-c)*sqr+(1-2*c)*arc;
t12=(2+c)*sqr-(1+2*c)*arc;
t13=-.5*(t7+(c-a)*t1);
t15=t4+t10;
t16=t1-t8-(c-a)*t4+0.5*t11;
t17=-2*t9-t1+(a-.5)*t4;
t18=t5-t4*t10;
t19=-.5*t4*t11;
%
%
%
iden=zeros(nn,nn);ks=zeros(nn,nn);knc=zeros(nn,nn);bs=zeros(nn,nn);
for ii=1:nn
iden(ii,ii)=1.0;
end
% mass matrix
ms(1,1)=1;
ms(1,2)=xa;
ms(1,3)=xb;
ms(2,1)=xa;
ms(2,2)=ra*ra;
ms(2,3)=rb*rb+xb*(c-a);
ms(3,1)=xb;
ms(3,2)=rb*rb+xb*(c-a);
ms(3,3)=rb*rb;
ms=ms*mas
%
%
mnc(1,1)=-pi;
mnc(1,2)=pi*a;
mnc(1,3)=t1;
mnc(2,1)=pi*a;
mnc(2,2)=-pi*(1./8.+a*a);
mnc(2,3)=-2*t13;
mnc(3,1)=t1;
mnc(3,2)=-2*t13;
mnc(3,3)=t3/pi;
%
%
% stiffness matrix

© 2001 by C. Nam, Y. Kim 39

ks(1,1)=wh*wh;
ks(2,2)=ra*ra*wa*wa;
ks(3,3)=rb*rb*wb*wb;
ks=ks*mas
%
%
knc(2,3)=-t15;
knc(3,3)=-t18/pi;
%
%
bs(3,3)=2*rb*rb*wb*damb;
bs=bs*mas
%
%
bnc(1,1)=0;
bnc(1,2)=-pi;
bnc(1,3)=t4;
bnc(2,1)=0;
bnc(2,2)=pi*(a-.5);
bnc(2,3)=-t16;
bnc(3,1)=0;
bnc(3,2)=-t17;
bnc(3,3)=-t19/pi;
%
%
g=[ 0.0
0.0
rb*rb*wb*wb];
g=g*mas;
%
%
r1=[ -2.*pi
2*pi*(a+0.5)
-t12];
%
%
s1=[ 0
1
t10/pi]' ;
%
%
s2=[ 1
(.5-a)
t11/(2*pi) ]' ;
%
%
%
% Roger approximation of aero forces
rka=[0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.6 0.7 0.8 0.9 1.0
1.2 1.4 1.6 1.8 2.0]';

© 2001 by C. Nam, Y. Kim 40

[m,n]=size(rka);
%
gam1=0.2;
gam2=0.4;
gam3=0.6;
gam4=0.8;

for ii=1:nn
for jj=1:nn
tempa=0;
tempb=0;
for jjrk=1:m
rk=rka(jjrk);
[fl,gl]=aero(rk,b,mnc,bnc,knc,r1,s1,s2);
flij=fl(ii,jj);
glij=gl(ii,jj);
al=[ 1 0
0 rk
-rk*rk 0
rk*rk/(rk*rk+gam1*gam1) rk*gam1/(rk*rk+gam1*gam1)
rk*rk/(rk*rk+gam2*gam2) rk*gam2/(rk*rk+gam2*gam2)
rk*rk/(rk*rk+gam3*gam3) rk*gam3/(rk*rk+gam3*gam3)
rk*rk/(rk*rk+gam4*gam4) rk*gam4/(rk*rk+gam4*gam4) ]';

tempa=tempa+al'*al;
tempb=tempb+al'*[flij glij]';
end
xe=inv(tempa)*tempb;

p0(ii,jj)=xe(1);
p1(ii,jj)=xe(2);
p2(ii,jj)=xe(3);
p3(ii,jj)=xe(4);
p4(ii,jj)=xe(5);
p5(ii,jj)=xe(6);
p6(ii,jj)=xe(7);

end
end

% For ploting approximated Aero. Matrix

for jj=1:m
rk=rka(jj);
irk=i*rk;
rkg(jj)=rka(jj);
[fl,gl]=aero(rk,b,mnc,bnc,knc,r1,s1,s2);

© 2001 by C. Nam, Y. Kim 41

zap=p0+p1*irk+p2*irk*irk+p3*irk/(irk+gam1)+p4*irk/(irk+gam2)+p5*irk/(ir
k+gam3)+p6*irk/(irk+gam4);
r1f(jj) =fl(1,1); r1g(jj)= gl(1,1);
a1f(jj)=real(zap(1,1));a1g(jj)=imag(zap(1,1));
r2f(jj) =fl(1,2); r2g(jj)= gl(1,2);
a2f(jj)=real(zap(1,2));a2g(jj)=imag(zap(1,2));
r3f(jj) =fl(1,3); r3g(jj)= gl(1,3);
a3f(jj)=real(zap(1,3));a3g(jj)=imag(zap(1,3));
r4f(jj) =fl(2,1); r4g(jj)= gl(2,1);
a4f(jj)=real(zap(2,1));a4g(jj)=imag(zap(2,1));
r5f(jj) =fl(2,2); r5g(jj)= gl(2,2);
a5f(jj)=real(zap(2,2));a5g(jj)=imag(zap(2,2));
r6f(jj) =fl(2,3); r6g(jj)= gl(2,3);
a6f(jj)=real(zap(2,3));a6g(jj)=imag(zap(2,3));
r7f(jj) =fl(3,1); r7g(jj)= gl(3,1);
a7f(jj)=real(zap(3,1));a7g(jj)=imag(zap(3,1));
r8f(jj) =fl(3,2); r8g(jj)= gl(3,2);
a8f(jj)=real(zap(3,2));a8g(jj)=imag(zap(3,2));
r9f(jj) =fl(3,3); r9g(jj)= gl(3,3);
a9f(jj)=real(zap(3,3));a9g(jj)=imag(zap(3,3));
end
figure(11)
plot(r1f,r1g,'*r',a1f,a1g,'og'),xlabel('Real of A(1,1)'),
ylabel('Imag of A(1,1)'),grid on
figure(12)
plot(r2f,r2g,'*r',a2f,a2g,'og'),xlabel('Real of A(1,2)'),
ylabel('Imag of A(1,2)'),grid on
figure(13)
plot(r3f,r3g,'*r',a3f,a3g,'og'),xlabel('Real of A(1,3)'),
ylabel('Imag of A(1,3)'),grid on
figure(14)
plot(r4f,r4g,'*r',a4f,a4g,'og'),xlabel('Real of A(2,1)'),
ylabel('Imag of A(2,1)'),grid on
figure(15)
plot(r5f,r5g,'*r',a5f,a5g,'og'),xlabel('Real of A(2,2)'),
ylabel('Imag of A(2,2)'),grid on
figure(16)
plot(r6f,r6g,'*r',a6f,a6g,'og'),xlabel('Real of A(2,3)'),
ylabel('Imag of A(2,3)'),grid on
figure(17)
plot(r7f,r7g,'*r',a7f,a7g,'og'),xlabel('Real of A(3,1)'),
ylabel('Imag of A(3,1)'),grid on
figure(18)
plot(r8f,r8g,'*r',a8f,a8g,'og'),xlabel('Real of A(3,2)'),
ylabel('Imag of A(3,2)'),grid on
figure(19)
plot(r9f,r9g,'*r',a9f,a9g,'og'),xlabel('Real of A(3,3)'),
ylabel('Imag of A(3,3)'),grid on

© 2001 by C. Nam, Y. Kim 42

state=0.0;
temp=inv(ms)*g;
% Let's set design airspeed
vel=320.
q=.5*rho*vel*vel;
invms=inv(ms-.5*rho*b*b*p2);
ans1=-invms*(ks-q*p0);
ans2=-invms*(bs-.5*rho*b*vel*p1);
ans3=q*invms*p3;
ans4=q*invms*p4;
ans5=q*invms*p5;
ans6=q*invms*p6;

for ii=1:nn
for jj=1:nn
state(ii,jj+nn)=iden(ii,jj);
state(ii+2*nn,jj+nn)=iden(ii,jj);
state(ii+3*nn,jj+nn)=iden(ii,jj);
state(ii+4*nn,jj+nn)=iden(ii,jj);
state(ii+5*nn,jj+nn)=iden(ii,jj);
state(ii+nn,jj)=ans1(ii,jj);
state(ii+nn,jj+nn)=ans2(ii,jj);
state(ii+nn,jj+2*nn)=ans3(ii,jj);
state(ii+nn,jj+3*nn)=ans4(ii,jj);
state(ii+nn,jj+4*nn)=ans5(ii,jj);
state(ii+nn,jj+5*nn)=ans6(ii,jj);
state(ii+2*nn,jj+2*nn)=-(vel/b)*gam1*iden(ii,jj);
state(ii+3*nn,jj+3*nn)=-(vel/b)*gam2*iden(ii,jj);
state(ii+4*nn,jj+4*nn)=-(vel/b)*gam3*iden(ii,jj);
state(ii+5*nn,jj+5*nn)=-(vel/b)*gam4*iden(ii,jj);
end
g1(ii)=0.0;
g1(ii+nn)=temp(ii);
g1(ii+2*nn)=0.0;
g1(ii+3*nn)=0.0;
g1(ii+4*nn)=0.0;
g1(ii+5*nn)=0.0;
end
g1=g1';
state=state;
%
%
%
iter=0;
eig(state);
for ii=1:6*nn;
if imag(ans(ii)) >= 1.0e-4,
if abs(real(ans(ii))) >= 1.0e-4,
if real(ans(ii)) <= 50.0,
if real(ans(ii)) >= -500.0,

© 2001 by C. Nam, Y. Kim 43

 iter=iter+1;
xopen(iter)=real(ans(ii));
yopen(iter)=imag(ans(ii));
end
end
end
end
end
xopen=xopen'; yopen=yopen';
%
%
% dx/dt = [state] x + [g1] u
% y = C x
%

for i=1:4
for j=1:6*nn
c(i,j)=0.0;
end
end
c(1,1)=1.0;
c(2,2)=1.0;
c(3,4)=1.0;
c(4,5)=1.0;
%
qq=c'*c/10000.0; rr=0.1;
%qq=qq/10000;rr=rr/10000;
[gain,ricca]=lqr2(state,g1,qq,rr);
% u = -Kx minimizes the cost function:
temp= -g1*gain;
iter=0;
eig(state+temp);
for ii=1:6*nn;
if imag(ans(ii)) >= 1.0e-4,
if abs(real(ans(ii))) >= 1.0e-4,
if real(ans(ii)) <= 50.0,
if real(ans(ii)) >= -500.0,
iter=iter+1;
xclse(iter)=real(ans(ii));
yclse(iter)=imag(ans(ii));
end
end
end
end
end
xclse=xclse';
yclse=yclse';
%
break
%

© 2001 by C. Nam, Y. Kim 44

iter=0;

for kk=1:41;
vel=10.*(kk-1)+50;
q=.5*rho*vel*vel;
state=0;
invms=inv(ms-.5*rho*b*b*p2);

ans1=-invms*(ks-q*p0);
ans2=-invms*(bs-.5*rho*b*vel*p1);
ans3=q*invms*p3;
ans4=q*invms*p4;
ans5=q*invms*p5;
ans6=q*invms*p6;
for ii=1:nn
for jj=1:nn
state(ii,jj+nn)=iden(ii,jj);
state(ii+2*nn,jj+nn)=iden(ii,jj);
state(ii+3*nn,jj+nn)=iden(ii,jj);
state(ii+4*nn,jj+nn)=iden(ii,jj);
state(ii+5*nn,jj+nn)=iden(ii,jj);
state(ii+nn,jj)=ans1(ii,jj);
state(ii+nn,jj+nn)=ans2(ii,jj);
state(ii+nn,jj+2*nn)=ans3(ii,jj);
state(ii+nn,jj+3*nn)=ans4(ii,jj);
state(ii+nn,jj+4*nn)=ans5(ii,jj);
state(ii+nn,jj+5*nn)=ans6(ii,jj);
state(ii+2*nn,jj+2*nn)=-(vel/b)*gam1*iden(ii,jj);
state(ii+3*nn,jj+3*nn)=-(vel/b)*gam2*iden(ii,jj);
state(ii+4*nn,jj+4*nn)=-(vel/b)*gam3*iden(ii,jj);
state(ii+5*nn,jj+5*nn)=-(vel/b)*gam4*iden(ii,jj);
end
end

state=state+temp;

eig(state);
for ii=1:6*nn;
if imag(ans(ii)) >= -1.0e-4,
if abs(real(ans(ii))) >= 1.0e-4,
if real(ans(ii)) <= 50.0,
if real(ans(ii)) >= -500.0,
iter=iter+1;
vsol(iter)=vel;
xsol(iter)=real(ans(ii));
ysol(iter)=imag(ans(ii));
end
end

© 2001 by C. Nam, Y. Kim 45

 end
end
end
%
end
%
vsol=vsol';
xsol=xsol';
ysol=ysol';
m=size(xsol);
for i=1:m;
result2(i,1)=vsol(i);
result2(i,2)=xsol(i);
result2(i,3)=ysol(i);
end
save close3.out result2 /ascii;
load open3.out
[mopn,nopn]=size(open3);
for i=1:mopn;
vopn(i)=open3(i,1);
xopn(i)=open3(i,2);
yopn(i)=open3(i,3);
end

figure(3);
plot(xsol,ysol,'.r',xopn,yopn,'.g'),
xlabel('Real Part'),ylabel('Imag. Part'),
title('3 Degree of Freedom, Open/Closed Loop System'),grid

%clg;
figure(4);
plot(xsol,ysol,'+r',xopen,yopen,'og',xclse,yclse,'*b'),
axis([-300 50 0 400])
xlabel('Real Part'),ylabel('Imag. Part'),
title('3 Degree of Freedom, Closed Loop System'),grid
axis;
%pause
figure(5);
subplot(221),plot(xopen,yopen,'og',xclse,yclse,'*b',xsol,ysol,'.r'),
axis([-150 50 0 400]),xlabel('Real Part'),ylabel('Imag. Part'),grid;

subplot(222),axis('off');
text(0.0,0.85,'Closed Loop Flutter Analysis','sc'),
text(0.0,0.75,'3 Degree of Freedom','sc'),
text(0.0,0.65,'Roger Approx.','sc');

subplot(223),plot(vsol,xsol,'.r'),
axis([0 400 -80 20]),xlabel('Velocity'),ylabel('Real Part'),grid;

subplot(224),plot(vsol,ysol,'.r'),

© 2001 by C. Nam, Y. Kim 46

axis([0 400 0 400]),xlabel('Velocity'),ylabel('Imag. Part'),grid;
%pause

3.6. A New Flutter Analysis Technique

Using Unsteady Aerodynamic Eigen Formulation
The aeroelastic analysis of the wing relies on an accurate representation of the unsteady
aerodynamic forces. Since the analysis is faced with the problem of coupling the
unsteady aerodynamic representation into the structural response model, this
representation also must be in a convenient computational form. The unsteady
aerodynamic model is divided into two groups which are frequency domain and time
domain analyses. The aeroelastic stability analysis is usually carried out in the frequency
domain, such as V-g method and p-k method. However, in modern aircraft design, the
interaction among the structure, aerodynamics and control system can not be neglected.
The control system exhibits sufficient gain and phase margins. A difficulty in using
modern control theory for the design of systems to control aeroelastic behavior is the
requirement of transforming the unsteady aerodynamic forces, normally provided in the
frequency domain, into the time domain. The usual procedure for the time invariant
state-space form is to approximate the unsteady aerodynamic force represented in the
frequency domain as a ration function in the Laplace domain. However, this requires
quite amount of CPU time and results in increase in the state due to the augmented
aerodynamic states.

In this section, a formulation for the unsteady aerodynamics in time domain is discussed.
In structural dynamic analysis, the dynamic behavior of the complex structures is often
reduced to a few degrees of freedom using normal mode analysis technique. Recently,
researchers expected that these techniques could be applied to unsteady aerodynamic
model. With the reduced order models, it might be possible to predict the unsteady
aerodynamic response of the system over a wide reduced frequency range. In this
section, we will develop an aeroelastic model using an aerodynamic eigen formulation.
A reduced-order model will be obtained by expanding the unsteady vortex into a sum of
first few aerodynamic eigenmodes. To account for the effects of truncated higher modes,
a static correction is included in the model. To get the system matrix with the real

© 2001 by C. Nam, Y. Kim 47

constant, a modal decomposition technique which transforms the aerodynamic matrix
into a canonical modal form is introduced.

3.6.a Basic Theory – Time Domain Unsteady Aerodynamics

We divide the bound vortex representing the airfoil into M elements, divide the free
vortex representing the traveling wake into (N-M) elements. The total number of vortex
elements is N. For an isolated flat airfoil in two-dimensional incompressible flow, the
kernel function is

1
K ij = (92.a)
2 π( x i 3 / 4 − ξ j )

where xi3/4 is the location of the i-th collocation, i.e., the three-quarter chord of each
element, and ξj is the location of the j-th point vortex, i.e., the quarter chord of each
element.

ξi (Point Vortex)
❊
Z xi (Collocation Point)
3/4

ξj (Point Vortex)
❊ ❊ ❊ ❊ .... ..... X
-b b
airfoil wake

Figure Discrete Vortex Model for 2D Airfoil

© 2001 by C. Nam, Y. Kim 48

For kernel function representation, we express the downwash in the discrete-space
domain,

{Γ }
{W3 / 4 } = [[ K 1 ] [ K 2 ]] 1  (92.b)
{Γ2 }

where W3 / 4 is a (M×1) vector containing downwashes at three-quarter points of vortex

elements of the airfoil. Γ1 ,Γ2 denote the vortex strength on the airfoil, wake,
respectively. And K1 , K2 are the corresponding kernel functions, which are sized M×M
and M×(N-M)

For conservation of vorticity,

d b
dt ∫−b
γ ( x, t )dx = −Vγ TE

where b is the semichord, V is the velocity and γ is the vorticity distribution on the
airfoil. In the discrete-space domain,

Γ
− [ S1 ]{Γ&1} = M +1 V (93)
∆x

with [S1 ] = [1 1 ... 1] (1×M).

For the special case of ∆ x = V∆ t , we can write

M
ΓMn++11 + ΓMn +1 = −2∑ (Γin+1 − Γin )
i =1

For convection of free wakes,

∂Γ Γ − Γi +1
Γ& i = −V i ≅ −V i −1 , (i = M + 2,..., N − 2)
∂x 2∆x

© 2001 by C. Nam, Y. Kim 49

 Γ − ΓN −1
Γ& N −1 ≅ −V N −2 (94)
∆x

Γ + ( w − 1)ΓN
Γ& N ≅ −V N −1
∆x

However, the free wake convection is more conveniently described in discrete time
domain as follows.

Γin+1 ≅ Γin−1 , (i = M + 2,..., N − 1)

ΓNn+1 ≅ wΓNn + wΓNn −1 , (0.95 < w < 1) (95)

The weighting factor w in Eqs. (94) and (95) is introduced to prevent sudden change in
the induced downwash due to the finite length of the wake vortex sheet. By assigning a
value near unity, one can accumulate all of the vortex strengths that would exist between
the last cutting point and infinity, thereby simulating the effect of these truncated
vorticities on the induced downwash. It can be shown that any w value equal to or larger
than unity would overestimate this effect. For example, w=1 will render all of the
truncated vortex strengths to be accumulated in the last vortex sheet element, i=N. Since
the location of this last vortex ξN is closer to a collocation point on the airfoil than any
other vortex locations behind, the resulting kernel function will yield an induced
downwash with a magnitude greater than that of true one. Hall suggests 0.95 < w < 1.

Putting Eq(92) (93) and (94) together in matrix form,

 0 0  & [ K1 ] [ K 2 ]   Γ  W 
 − S1  0 0    Γ1  =  V  1  −  3/ 4  (96)
   
 Γ2 &  0 [ B ]
a  Γ2   0 
  0  0 I N −M −1      ∆x 

1
where [I N −M −1 ] is the (N-M-1) × (N-M-1) identity matrix, and [B ] matrix
∆x a
approximates the partial derivatives of free vortex elements in x, given in Eq. (94).

Solving for Γ1 from the top of Eq.(96) gives

© 2001 by C. Nam, Y. Kim 50

 {Γ1} = −[ K1−1 K 2 ]{Γ2 } + [ K1 ]−1{W3 / 4 }

Hence,

 − S1  −1 0 0  & V S 
 [ K1 K 2 ] +   {Γ2 } = [ Ba ]{Γ2 } +  1 [ K 1 ] −1{W& 3 / 4 } (97)
 0  0 I N − M −1   ∆x 0

On the other hand, a discrete version of Eq.(97) can be written as

 − 2S1  −1 
 [ K 1 K 2 ] + [ I N − M ]{Γ2n +1 } =
 0  
(98)
 − 2S1  −1   − 2 S1  2 S 
 [ K 1 K 2 ] + [C ]{Γ2n } +  [ K 1 ] −1{W3n/+41 } +  1 [ K 1 ] −1{W3n/ 4 }
 0    0   0 

where

− 1 0 0 ... 0 
1 0 0 ... 0 

C=0 1 0 ... 0 
 
 .. .. .. ... .. 
 0 0 ... 1 w

3.6.b Determination of Ba Matrix

Ba can be obtained either from the original continuous model given by Eqs. (93) and (94),
or from the discrete model given by Eq. (95). Since the discrete model given by Eq. (95)
describes the behavior of trailing wake adequately, it is recommended that the discrete
model be converted to a continuous one in the following manner. First, write the
continuous model given by Eq. (97) in the form

V
[ Aa ]{Γ& 2 } = [ Ba ]{Γ2 } + [C a ]{W& 3 / 4 } (99)
∆x

© 2001 by C. Nam, Y. Kim 51

and the discrete one given by Eq.(98) as

[ Aad ]{Γ2n +1 } = [ Bad ]{Γ2n } + [C ad ]{W& 3n/+41 } (100)

V
where the time step in this discrete model is ∆t = .
∆x

It should be emphasized that the downwash velocity vector {W& 3 / 4 } acts as an input to the
vortex systems. Inverting the homogeneous part of the continuous equation yields,

V
{Γ& 2 } = [ Aa ] −1 [ Ba ]{Γ2 } (101)
∆x

By integrating the Eq. (101) from tn to tn+1, one can relate the discrete model to the
continuous model as

 1 
{Γ2n +1 } = [ Aad ] −1 [ Bad ]{Γ2n } = exp( [ Aa ] −1 [ Ba ])∆t {Γ2n } (102)
 ∆t 

from which one obtains

[ Ba ] = [ Aa ] ln([ Aad ] −1 [ Bad ]) (103)

Once Ba is determined, the unsteady lift and moment can be obtained as follows:

M M j
L = ρV ∑ Γ j + ρ ∑∑ Γ& k ∆x j (104)
j =1 j =1 k =1

M M j
M = − ρV ∑ (ξ j − ab)Γ j − ρ ∑ (ξ j − ab)∑ Γ& k ∆x j (105)
j =1 j =1 k =1

where ρ, ab denote the air density, distance from the midchord to the elastic axis,
respectively.

© 2001 by C. Nam, Y. Kim 52

3.6.c Eigen Analysis of Aerodynamic Model

The homogeneous part of Eq. (99) (or Eq. (101)) allows an eigen analysis of the unsteady
vortex equation. Assuming a harmonic solution of the vortex in the from Γ2 = xe λt yields

V
λ[ Aa ]{x} = [ Ba ]{x} (106)
∆x

The solutions of Eq. (106) for λ are always stable, i.e., the real parts of the eigenvalues
are negative. This confirms the fact that the creation of wake vortices through unsteady
motion of the airfoil should be dissipative; the wake vortices are formed as a result of
bound vortices being shed at the trailing edge through viscous activity in the boundary
layer around the airfoil. It is seen that the eigenvalues λ are inversely proportional to the
factor ∆x/V, the time for a free vortex element to travel the distance ∆x. Thus, if λ i and
xi are an eigenvalue and the corresponding right eigenvector of the system

λ i [ Aa ]{xi } = [ Ba ]{xi } (107)

V
then the corresponding eigenvalue and the eigenvector of the system (106) are λ i and
∆x
xi , respectively. More generally, we can write the right eigenvector problem as

[ Aa ][ X ][ Λ ] = [ Ba ][ X ]

where Λ is a diagonal matrix containing the eigenvalue of the generalized problem given
by (107), and X is a matrix whose columns contain the corresponding right eigenvectors.
Similarly, for the left eigenvector problem,

T T
[ Aa ] [Y ][Λ ] = [Ba ] [Y ]

The eigenvectors satisfy the orthogonality conditions

[Y ]T [ Aa ][ X ] = [ I ]

© 2001 by C. Nam, Y. Kim 53

 [Y ]T [Ba ][X] = [Λ ]

It can be shown that the eigenvalues of the continuous-time system λ i are related to the
~
eigenvalues of the discrete-time system, λ i as

~
λi = ln(λi ) (108)

while the eigenvectors of two systems remain the same. Thus, for our purpose, the
calculation of [Ba ] matrix given by Eq. (103) and the eigenvalues, eigenvectors of the
continuous-time system are all facilitated by obtaining the eigenvalues and eigenvectors
of the discrete-time system.

3.6.d System Equations for Aeroservoelastic Analysis

The equations of motion for classical 2-DOF wing model are used to conduct
aeroservoelastic analysis.

© 2001 by C. Nam, Y. Kim 54

 Figure Typical Section Model

For a typical section, the equation of motion is

 m mxθ   d 2 h / dt 2   K h 0   h  − L 
mx 2  +  =  (109)
 θ mrθ  d 2α / dt 2   0 K α  α   M 

 1 / rθ2 xθ / rθ2  h&&   R 2 / rθ2 0 h  − Lb / K α 

   +    =  
 xθ / rθ
2
1  α&&  0 1 α   M / K α 

h&&  h  Γ 
[ M ]  + [ K ]  = [V ] 1  (110)
α&& α  Γ2 

2 2 ωh
where ω2h = K h / m , ωθ = Kα / mrθ , xθ = xθ / b , rθ = rθ / b , h = h / b , R = . (•) now
ωθ
denotes a derivative with respect to the nondimensional time τ = ωθt

And

− Vθ F1 − F2 
= [Vθ V1 V2 ]
1
V = V G
πρrθ 2  θ 1 G 2 

[ F1 ] = [1 1 .... 1]
[ F2 ] = ∆x [M M − 1 ... 1]

[
[G1 ] = − (ξ 1 − a ) (ξ 2 − a ) .... (ξ M − a ) ]
[G2 ] = − ∆x [h1 h2 ... hM ]

M
hk = ∑ ξ j − ( M − k + 1)ab
j =k

© 2001 by C. Nam, Y. Kim 55

where the mass ratio is defined as µ , Vθ is the reduced velocity and Γ is Γ / b2 ωθ

For an airfoil undergoing plunging and pitching motion, the nondimensional downwash
vector is related to the plunging and pitching degrees of freedom via

h&  h 
{W3 / 4 } = [ E1 ]  + Vθ [ E 2 ]  (111)
α&  α 

E1 = [1 ( x3 / 4 − e )],E 2 = [0 1] (112)

Here x3 / 4 and a a are the vectors containing nondimensional locations of the three-
quarter points of vortex elements and elastic axis of the airfoil aft of midchord.
Combining Eqs. (92), (94), and (110), and eliminating Γ1 from the equation, we can get
the following contracted equations:

Γ& 2  Γ 
 &   Vθ [ B ] [0] [0]  h
 2 
 [ Aa ] Vθ [ Ac12 ] [ Ac13 ]  h   ∆x a  
 [0]      α 
[ I ] [ 0 ]  α  = [ 0 ] [ 0 ] [ I ]
 2  &  2
 

[ Ac 31 ] Vθ [ Ac 32 ] [ Ac 33 ] h &&  V [ B ] − [ K ] + Vθ [ Bc 32 ] Vθ [ Bc 33 ]  h& 
2
   θ c 31   
 α&&    α& 
(113)

where

© 2001 by C. Nam, Y. Kim 56

 − S 
[ Ac12 ] =  1 [ K 1 ] −1 [ E 2 ],
 0 
− S 
[ Ac13 ] =  1 [ K 1 ] −1 [ E1 ]
 0 
[ Ac 31 ] = [V2 ][ K 1 ] −1 [ K 2 ]
[ Ac 32 ] = −[V2 ][ K 1 ] −1 [ E 2 ]
[ Ac 33 ] = [ M ] − [V2 ][ K 1 ] −1 [ E1 ]
[ Bc 31 ] = [V1 ][ K 1 ] −1 [ K 2 ]
[ Bc 32 ] = [V1 ][ K 1 ] −1 [ E 2 ]
[ Bc 33 ] = [V1 ][ K 1 ] −1 [ E1 ]

3.6.e Construction of Reduced-Order Model

We decompose the unsteady solution of vortex into two parts; a modal expansion part
which is the sum of a few of the aerodynamic eigenmodes, and a static correction part:

{Γ2} ≅ [X R ]{c} + {Γs} (114)

Here X R is a (N-M)×NR matrix where NR << (N-M) whose columns are the NR right
eigenvectors corresponding to the first NR eigenvalues of Eq. (106) that are nearest to the
origin, c is the generalized coordinate vector. The modal expansion part approximates
the solution by spanning an NR-dimensional subspace of the complete (N-M)-dimensional
space, and therefore does not include the effects of all the remaining (N-M-NR)
aerodynamic modes. The role of the static correction part Γs is to account for the
missing effects by providing at least the quasi-static contribution from the truncated
higher modes. It can be shown that the static correction is obtained as

{Γs } = {Γ2 s } − [ X R ]{c s } (115)

Although the reduced-order model can be used to predict the open-loop flutter and
divergence by solving the complex generalized eigenvalue problem, it is not appropriate

© 2001 by C. Nam, Y. Kim 57

for a closed-loop control purpose. This is because the some of the entries in the matrix
equation are not real but complex due to the existence of the complex eigenvector, X R .

To overcome such difficulty and make all elements of the matrix equation real, we
introduce a modal decomposition technique which transforms the aerodynamic matrix
into a canonical modal form where a pair of complex conjugate eigenvalues appear in a
(2×2) block on the diagonal and the real eigenvalues appear on the diagonal. For the
V
aerodynamic system Eq. (101) (without the factor of θ ) that has m pairs of complex
∆x
conjugate eigenvalues − σ i ± jωi and n real eigenvalues -λi (N-M=2m+n), the modal
matrix is

[ Acan ] = [Q]T ([ Aa ]−1 [ Ba ])[ P]

(116)
[Q]T [P ] = [ IN − M ]

For the reduction of order, we decompose the unsteady vortex into

{Γ2} ≅ [PR ]{q} + {Γs} (117)

where PR is a (N-M)×NR matrix whose columns are the NR columns of P corresponding

to the first NR eigenvalues that are nearest to the origin, q is the new generalized
coordinate vector.

Therefore, the static correction is obtained as

{Γs} = {Γ2s} − [PR ]{q} (118)

The first term is the solution of the following system:

Vα
0= [ Ba ]{Γ2 s } + [C a ]{W& 3 / 4 } (119)
∆x

The quasi-static solution of q is

© 2001 by C. Nam, Y. Kim 58

 {q s } = −
Vα
∆x
{ }
[QR ]T [ Aa−1 Ba ][ PR ] [QR ]T [ Aa−1C a ]{W& 3 / 4 }
−1

(120)
V
= − α [ AcanR ] −1 [QR ]T [ Aa−1C a ]{W& 3 / 4 }
∆x

where AcanR is a NR×NR submatrix whose nonzero entries are those of Acan corresponding
to the NR retained eigenvalues, and QRT is a matrix whose rows contain the corresponding
NR left eigenvectors.

Therefore,

∆x
{Γs } =
Vθ
{ }
− [ Ba ] −1 + [ PR ][ AcanR ] −1 [QR ]T [ Aa ] −1 [C a ]{W& 3 / 4 } (121)

With the static correction as obtained above, a new reduced-order aeroservoelastic model
can be obtained:

 q&  q 
 &   Vθ  
 [ I NR ] Vθ [ Ac12 ] [ Ac13 ] h  [ AcanR ] [0] Vθ [ Bc13 ] h
     ∆x  
 α 
[ 0 ] [ I ] [ 0 ] α
   = [ 0 ] [ 0 ] [ I ]
 
 2 & 2
[ Ac 31 ] Vθ [ Ac 32 ] [ Ac 33 ] h&&   Vθ [ Bc 31 ] − [ K ] + Vθ2 [ Bc 32 ] Vθ [ Bc 33 ] h& 
  
α&&    &  (122)
α 
[Q R ]T [ K s ][ Ac13 ]
∆x   &h&&
+ [ 0 ]  
Vθ 
[ Ac 31 ][ K s ][ Ac13 ]  α
&&&
 

where

© 2001 by C. Nam, Y. Kim 59

 [ Ac12 ] = [QR ]T [ Aa ] −1 [ Ac12 ]
[ Ac13 ] = [QR ]T ([ Aa ] −1 [ Ac13 ] + [ Aa ] −1 [ Ba ][ K s ][ Ac13 ] − ∆x[ K s ][ Ac12 ])
[ Ac 31 ] = [ Ac 31 ][ PR ],[ Ac 32 ] = [ Ac 32 ]
[ Ac 33 ] = [ Ac 33 ] − ∆x ([ Ac 31 ][ K s ][ Ac12 ] − [ Bc 31 ][ K s ][ Ac13 ])
[ Bc13 ] = −[QR ]T [ Aa ] −1 [ Ba ][ K s ][ Ac12 ]
[ Bc 31 ] = [ Bc 31 ][ PR ],[ Bc 32 ] = [ Bc 32 ]
[ Bc 33 ] = [ Bc 33 ] − ∆x ([ Bc 31 ][ K s ][ Ac12 ])
[ K s ] = −[ Ba ] −1 + [ PR ][ AcanR
−1
][QR ]T [ Aa ] −1

As discussed in this section, static correction is adopted during the order reduction in
order to provide the quasistatic contribution from the truncated higher modes. Therefore,
the terms containing (&h&& α
&&&) are appeared in the reduced order model. When we design
the controller, these terms will not be accounted; they might be assumed as the structural
uncertainty. If the magnitude of the structural uncertainty is small, the robust stability of
the closed loop system is not degraded even though the controller is designed with the
model that neglects the uncertainty. As a matter of fact, it was found that the effects of
those terms are negligibly small, these terms are not considered in the control system
design. Note that the reduced modes of the aerodynamic model are retained in the
reduced order model, and both models (full order model and reduced order model) use
same physical coordinates for the structural modes.

It has to be noted that the nondimensional divergence speed of the system, VθDiv , can be
obtained by solving the eigenvalue problem of the following equation;

h 
[−[K ] Vθ2Div [Bc 32 ]]  = {0}
α 
(123)

3.6.f Example - Flutter Analysis Using Eigen Formulation

The typical section model is used in this study as shown in Figure. A model has two
degrees of freedom which are plunge and pitch. Linear and torsion springs at the elastic

© 2001 by C. Nam, Y. Kim 60

axis act to restrain motion in plunge and twist. The mass ratio µ is 20, the static
imbalance xθ =0.2, the radius of gyration rθ is 0.5, and the location of the elastic axis aft
of midchord a is -0.1. The frequency ratio R of the uncoupled modes is 0.3.

Figure show the discrete time eigenvalues and continuous time eigenvalues for the typical
section model, respectively. For this analysis, ten vortex elements are used to model the
airfoil. The wake was modelled using 100 vortex elements, and the length of the wake
was taken to be 10 chord lengths. As shown in Figure, the magnitude of all the discrete
time eigenvalues is less than 1. These eigenvalues are mapped into the left in the
continuous time domain, which means that the system is stable.

discrete model continous model

1 4

0.8
3
0.6
2
0.4
1
Imaginary Part
Imaginary Part

0.2

0 0

-0.2
-1

-0.4
-2
-0.6
-3
-0.8

-1 -4
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1 -0.5 -0.45 -0.4 -0.35 -0.3 -0.25 -0.2 -0.15 -0.1 -0.05 0
Real Part Real Part

Figure (a) Discrete Time Eigenvalues, (b) Continuous Time Eigenvalues of the
Unsteady Aerodynamics

The eigenmode information is used in order to compute the flutter speed in the typical
section model. For model reduction, total 20 eigenmodes from 100, which have the
nearest continuous time eigenvalues to the origin, are used in the reduced order model
along with the static correction.

© 2001 by C. Nam, Y. Kim 61

 Eigen Formulation Eigen Formulation

0.2 2

1.8
0
1.6

1.4

( /ω )
θ
-0.2
1.2

Frequency Ratio ω
Damping

-0.4 1

0.8
-0.6
0.6

0.4
-0.8
0.2

-1 0
0 0.5 1 1.5 2 2.5 0 0.5 1 1.5 2 2.5
Nondimension Velocity (V/ ω θ b) Nondimension Velocity (V/ ω θ b)

V-g Method V-g Method

0.2 2

1.8
0
1.6

1.4
( /ω )

-0.2
θ

1.2
Frequency Ratio ω

-0.4 1
g

0.8
-0.6
0.6

0.4
-0.8
0.2

-1 0
0 0.5 1 1.5 2 2.5 0 0.5 1 1.5 2 2.5
Velocity (V/ ωθ b) Velocity (V/ ωθ b)

Figure Frequency and Damping Plots of the Typical Section Model with the
Variation of the Reduced Velocity

Figure show the frequency( ω/ ωθ ), damping of the reduced order model against the
reduced velocity( V / ωθb ). For comparison, the results of V-g method are also plotted.

© 2001 by C. Nam, Y. Kim 62

In V-g method, the unsteady aerodynamics formulated by Theodorsen are used. The
flutter speed calculated using V-g method is about 2.0. Note that the frequency and g
from V-g method do not have the physical meaning except at the flutter speed. The value
of g in V-g method denotes the damping required for the harmonic motion. On the other
hand, the value of g obtained from the current method denotes the real damping of the
system at the specified airspeed. That is identical with the flutter speed using the current
formulation.

It should be noted that his system model is shown in the discrete time domain and has the
complex coefficient in the matrix equation. However, in this paper, the equations of
motion were derived on the continuous time domain, and the reduced order model with
the real constant was proposed using a modal decomposition technique. Therefore, the
current formulation can be directly used for the aeroelastic control system design. The
predicted divergence speed from V-g method is about 2.5.

Comparing the V-g method with given typical section model, the current method has a
disadvantage in the computational time for aerodynamic eigenvalues. However, the
situation will be different when the more realistic wing model is used for the aeroelastic
analysis. Currently, this method is being extended to analize the aeroservoelastic
characteristics of the 3D wing model using the unsteady vortex lattice method. It is
expected that there will be substantial CPU time savings compared to the existing
method, such as frequency domain analysis with the rational function approximation.

Matlab List

Test.m
close all
clear

e=-0.1;
xalpha=0.2;ralpha=0.5;
xmu=20.0;omegah=0.3; alpha=0.996;
% e=-0.2;
% xalpha=0.1;ralpha=0.5;
% xmu=20.0;omegah=0.4; alpha=0.996;

© 2001 by C. Nam, Y. Kim 63

disp('Aeroservoelastic Design Study Using Eigen Formulation')
iopen=input('Open loop snalysis ? (y/n)=(1/0) ');
if(iopen == 1),
ui=input('initial vel '); % nondimensionalized w.r.t (omega_theta
b)
uf=input('final vel ');
delu=input('increment vel ');
else
disp('Fix the design airspeed in order to design a control system')
u=input('design velocity ');
end
n=110;m=10;nm=100;nms=104;
nmode=20;nmd2=24;
numcon=2;numout=4;
xm=m;xn=n;
%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
vortex1
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%
%=====Invert xk1 matrix=====
%
xk1=inv(xk1);
aaa1=zeros(n-m,m);
aaa1(1,1:m)= -s1(1,1:m);
bad=aaa1*xk1*xk2;
%
aad(1:n-m,1:n-m)=bad(1:n-m,1:n-m);
bad(1:n-m,1:n-m)=bad(1:n-m,1:n-m)+xc(1:n-m,1:n-m);

for i=1:n-m
aad(i,i)=aad(i,i)+1.0d00;
end
%
aad=inv(aad);
ad=aad*bad;
bd(1,1)=1.0;
% %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
d2cmr
msize=nrdc+4;
%
% Set desired speed for designing controller
%
if(iopen == 1)
opnfltr;
disp('Open Loop Analysis is Done !')
break
end
%

© 2001 by C. Nam, Y. Kim 64

Vortex1.m
%===============================================
% m: number of vortex elements for airfoil.
% n: total number of vortex elements for airfoil & free wake.
% x(i): locations of 3/4 chord points of vortex elements.
% xi(i): locations of 1/4 chord points of vortex elements.
% xk: K = {K1 K2] matrix (mxn).
% xk1: K1 matrix (mxm).
% xk2: K2 matrix (mx(n-m)).
% s1: S1 matrix (1xm).
% xc: C matrix (mx(n-m)).
%===============================================
%
dx=2./m;dx1=dx/4.;dx2=3.*dx1;
%
for i=1:n
x(i)=dx*(i-1)+dx2-1.0;xi(i)=dx*(i-1)+dx1-1.0;
end
%
for i=1:m
s1(1,i)=2.0;
for j=1:n
xk(i,j)=1./(x(i)-xi(j))/2./pi;
end
end
xk1=xk(1:m,1:m); xk2(1:m,1:n-m)=xk(1:m,1+m:n);
%
for i=1:n-m
for j=1:n-m
xc(i,j)=0.d0;
end
end
for i=1:n-m-1
xc(i+1,i)=1.d0;
end
xc(1,1)=-1.0;xc(n-m,n-m)=alpha;

D2cmr.m
npar=nm; nmpar=nm;
nmode=20;
%
% converting routine by zero-order-hold approximation
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
ts=1.0;err1=1.d-7 ;err2=1.d-10 ;
%
% construct system matrix for d2c
%

© 2001 by C. Nam, Y. Kim 65

 rad=zeros(nmode,nmode);sysmat=zeros(nm,nm);
sysmat(1:nm,1:nm)=ad(1:nm,1:nm);
lneig=zeros(nm,nm);
cw1=zeros(nm,nm);cw2=zeros(nm,nm);
%
[ceig,phi,psi]=eign(sysmat,eye(nm));
figure (1) ;
plot(real(ceig), imag(ceig),'r*'),
xlabel('Real Part'),ylabel('Imaginary Part'),
title('discrete model');grid on
%
%
for i=1:nm
lneig(i,i)=log(ceig(i)) ;
end
% compute log(sysmat)=phi*log(lneig)*transpose(psi)
%
cw2=phi*lneig*conj(psi');
%
%
% error check by observing imaginary parts (should be zero)
%
errsum=0.d0 ;
for i=1:nm
for j=1:nm
chkimag=imag(cw2(i,j));
errsum=errsum+chkimag*chkimag/(ts*ts) ;
end
end

errchk=sqrt(errsum);
if errchk > err1,
disp('warning: accuracy of d2c conversion may be poor.')
end

a=real(cw2)/ts;
wk1=a;
[ceig2,phi,psi]=eign(wk1,eye(nm));
figure (2) ;
plot(real(ceig2), imag(ceig2),'r*'),
xlabel('Real Part'),ylabel('Imaginary Part'),
title('continous model');grid on;axis([-0.5 0 -4 4])

wk1=zeros(nm,nm);
ii=1;
for kk=1:nm
wr(ii)= real(ceig2(ii));wi(ii)=imag(ceig2(ii));

if(abs(wi(ii)) > 0.d0),

wk1(ii,ii) = wr(ii);wk1(ii+1,ii+1)= wr(ii);

© 2001 by C. Nam, Y. Kim 66

 wk1(ii,ii+1) = wi(ii);wk1(ii+1,ii) =-wi(ii);
ii=ii+2;
else
wk1(ii,ii)=real(ceig2(ii));
ii=ii+1;
end
if(ii > nm), break; end
end
rad=zeros(nmode,nmode); rad=wk1(1:nmode,1:nmode);
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

[vecr,vecl]= rmcomm2(npar,nm,ceig,phi,psi);
nrdc=nmode;
rvecr(1:nm,1:nmode)=vecr(1:nm,1:nmode);
rvecl(1:nm,1:nmode)=vecl(1:nm,1:nmode);

Opnfltr.m
icheck=0;
jcount=1;
disp('Working ......')
for u=ui:delu:uf
if(icheck == 0)
fluttrr;
ac0(1:msize,1:msize)=ac(1:msize,1:msize);
bc0(1:msize,1:msize)=bc(1:msize,1:msize);
icheck=1;
else
flutteru;
end
ac1(1:msize,1:msize)=ac(1:msize,1:msize);
bc1(1:msize,1:msize)=bc(1:msize,1:msize);
alf=eign(bc1,ac1);

for i=1:msize
if( (imag(alf(i)) > 0.0) & (imag(alf(i)) < 5.0)),
xxx=[u, real(alf(i)) imag(alf(i))];
sol(jcount,1:3)=xxx ;jcount=jcount+1;
end
end
end
figure(3)
plot(sol(:,1),sol(:,2),'b*'),grid,axis([0 2.5 -1.0 0.2]),
xlabel('Nondimension Velocity (V/\omega _\theta b)'),
ylabel('Damping'),
figure(4)
plot(sol(:,1),sol(:,3),'b*'),grid,axis([0 2.5 0 2.0])
xlabel('Nondimension Velocity (V/\omega _\theta b)'),
ylabel('Frequency Ratio (\omega / \omega _\theta)'),

© 2001 by C. Nam, Y. Kim 67

 figure(5)
plot(sol(:,2),sol(:,3),'b*'),grid,

© 2001 by C. Nam, Y. Kim 68

Chapter 4
Composite Box Beam Modeling

In this chapter, aeroelastic analysis will be discussed of the high aspect ratio wing. Wing
can be modeled as a box beam with an assumption that the chordwise segments of the
wing are rigid. Wing skin is covered with a composite laminated plate. In this chapter,
we assume that the readers are familiar with the composite materials and the laminated
composite.

4.1 Deformation of the High Aspect Ratio Composite Wing

The wing structure is idealized as a box beam whose flexibility is due entirely to the
presence of upper and lower laminated cover sheets. It is also assumed that the
chordwise sections are rigid so that the wing deformation is a function of the spanwise
coordinate. Since the lamina are constrained to act as a unit, in a manner prescribed by
the Euler-Bernoulli deformation assumptions, the equivalent bending and torsion stiffness
may be computed.

The displacement of the box beam model can be represented by a bending deflection h
(positive, upward) along the spanwise axis and a rotation, α (postive, nose-up) about that
axis.

w(x, y) = h(y) − x α (y) (1)

The reference axis is placed at the midchord of the airfoil section along the span. And it
is assumed that the upper and lower cover sheets are oriented symmetrically with respect
to the middle surface. In the case of slender swept composite beam of high aspect ration,
an effective root can be assumed as shown in figure. The y-axis is placed on the
midchord along the spanwise direction.

© 2001 by C. Nam, Y. Kim 1

 Effective root

y’

Actual root y
x
x’

Figure High Aspect Ratio Wing Model

The constitutive relations of the composite lamina can be written as,

 σ  Q Q Q16   ε x 
 x   11 12
 
 σ y  = Q12 Q22 Q26   ε y  (2)
    
σ xy  Q16 Q26 Q66  ε xy 

where Qij is the stiffness matrix of the lamina. Note that the stiffness matrix is function
of the ply orientation of lamina.

The moment resultants are given as the sum of the stress times the area over which they
act multiplied by the moment arm with respect to the midplane, M ij = ∫ σ ij zdz

Due to the symmetry with respect to the midplane, the moment equation can be reduced
as

 M  D D12 D16  κ x 
 x   11   
 M y  =  D12 D22 D26  κ y  (3)
    
 Mxy   D16 D26 D66  κ xy 

where Dij is the bending stiffness of the laminated composite. It has to be noted that Mij
is the moment per unit width of the box beam along the midplane.

The curvature, κ is expressed in terms of the displacement of box beam as

© 2001 by C. Nam, Y. Kim 2

  ∂2 w 
 
κ   ∂x 2 
 x   ∂2 w 
 κ y  = − 2  (4)
   ∂y 
κ xy   ∂2 w 
 
 ∂x∂y 

The bending stiffness Dij is define as

N
1 3
Dij = ∫ Qij z 2 dz = ∑ Qij( p ) (t p + 3t p z 2p + 3 z p t 2p ) (5)
p =1 3

where tp is the thickness, zp is the coordinate of lower surface for the each lamina. Since
we are modeling a box beam, Mx=0 and the wing bending moment, M (positve upward),
and the torsion moment, T (positive nose-up), will have the following relations.

M = −cM y
(6)
T = 2cM xy

where c is the chord of the box beam segment. And the curvature will be expressed in
terms of the following variables.

∂2 w
κy = − = − h ′′
∂y 2
(7)
∂2 w
κ xy = −2 = 2α ′
∂x∂y

The equation for the bending and torsion moments of a composite box beam model will
be

− M / c   D22 D26  − h ′′

 =   (8)
 T / 2c   D26 D66   2α ′ 

M   cD22 − 2cD26  h ′′

 =  
 T  − 2cD26 4cD66  α ′
(9)
 EI − K  h ′′
=  
− K GJ  α ′

© 2001 by C. Nam, Y. Kim 3

where EI is the bending stiffness, GJ is torsional stiffness and K is the coupling between
beinding and torsion stiffnesses. The K parameter provides the elastic coupling between
bending and torsion not found in metallic beam analysis. Since K depends upon both the
stacking sequence of the laminate and the lamina fiber orientation angles, and
opportunity to tailor the wing structure to provide advantageous bending-torsion coupling
is present.

By taking the inverse of the above equation, the flexibility relations are obtained.

−1
h ′′  EI − K  M 
 =  
α ′ − K GJ   T 
1 GJ K  M 
= 2    (10)
EI ⋅ GJ − K  K EI   T 
 1 g 
 EI (1 − kg ) EI (1 − kg )  M 
=  
 g 1  T 
 EI (1 − kg ) GJ (1 − kg ) 

where k=K/EI, g=K/GJ.

4.2 Finite Element Formulation of Box Beam Model

In this chapter, the finite element method will be used to develop an equation of motion
for aeroservoelastic analysis. The total strain energy of a laminated beam may be written
as follows;

1 L
U=
2 ∫0
[ EI (h′′) 2 + GJ (α ′) 2 − 2 Kh ′′α ′]dy (11)

The static equilibrium equation of the box beam are derived from the energy equation.
The coupled differential equations of the composite box beam model is expressed as

[ EI ⋅ h ′′ − K ⋅ α ′]′′ = 0
(12)
[− K ⋅ h ′′ − GJ ⋅ α ′]′ = 0

© 2001 by C. Nam, Y. Kim 4

 q1 q4
q3 q2 q5
q6
l

Figure Beam Finite Element Model (6DOF/element)

Now, it is necessary to express displacement, h and rotation, α in terms of the nodal

displacements. The bending displacement and rotation of the beam element can be
obtained by applying the boundary conditions at each end of the beam element. The
boundary conditions are as follows;

At y=0,

q1=h@ y=0, q2=h’@ y=0, q3=α@ y=0

At y=l,

q4=h@ y=l, q5=h’@ y=l, q6=α@ y=l

By applying the boundary conditions, the bending displacement, h and rotation,α can be
expressed in terms of the known nodal displacements, q1,…,q6.

6
h ( y ) = ∑ψ i ( y ) ⋅ q i
i =1
6
(13)
α ( y) = ∑ ϕ i ( y) ⋅ qi
i =1

where ψ i , ϕ i are called as shape functions, and defined as

© 2001 by C. Nam, Y. Kim 5

   y
2
 y 
3

 1 − 3   + 2   
 l l 
 2y 2
y3 
 y − +

l l 2 
0 
ψi =  2 3  (14)
 3 y  − 2 y  
 l l 
 y 2
y3 
 − + 2 
 l l 
 0 

  y y2  
 6 g  − 2 + 3  
  l l 
  y  y  2 
3 g  − +   
  l  l  
 y 
 1− 
ϕi =  l (15)

 6 g  y − y  
2

  l2 l3  
 
 2 
3 g  − y +  y  

  l  l  
  
 y 
 l 

The beam element has six degrees of freedom where qi are the nodal displacements.
With these displacement and rotation functions, the shear force, V bending moment, M
and torsional moment, T at each end of the beam element also can be expressed in terms
of the nodal displacements.

V y =0 = EI ⋅ h ′y′′=0 − K ⋅ α ′y′=0
M y =0 = EI ⋅ h ′y′=0 − K ⋅ α ′y =0
T y =0 = GJ ⋅ α ′y =0 − K ⋅ h ′y′=0
(16)
V y =l = EI ⋅ h ′y′′=l − K ⋅ α ′y′=l
M y =l = EI ⋅ h ′y′=l − K ⋅ α ′y =l
T y =l = GJ ⋅ α ′y =l − K ⋅ h ′y′=l

© 2001 by C. Nam, Y. Kim 6

In matrix form , the static equilibrium equation can be written as

[K]{q} = {Q} (17)

where {Q} is the element force vector, [K] is the element stiffness matrix and {q} is the
nodal displacement vector.

The components of stiffness matrix are given as;

ke(1,1)= 12.0*EI*(1.0-k*g)/el3;
ke(1,2)= 6.0*EI*(1.0-k*g)/el2;
ke(1,3)= 0.0; ke(1,4)=-ke(1,1);
ke(1,5)= ke(1,2); ke(1,6)= 0.0;

ke(2,1)= ke(1,2);
ke(2,2)= EI*(1.0+3.0*(1.0-k*g))/el;
ke(2,3)=-K/el; ke(2,4)=-ke(1,2);
ke(2,5)=-EI*(1.0-3.0*(1.0-k*g))/el; ke(2,6)=-ke(2,3);

ke(3,1)= ke(1,3); ke(3,2)= ke(2,3); ke(3,3)= GJ/el;

ke(3,4)= 0.0; ke(3,5)=-ke(2,3); ke(3,6)=-ke(3,3);

ke(4,1)= ke(1,4); ke(4,2)= ke(2,4); ke(4,3)= ke(3,4);

ke(4,4)= ke(1,1); ke(4,5)=-ke(1,2); ke(4,6)= 0.0;

ke(5,1)= ke(1,5); ke(5,2)= ke(2,5); ke(5,3)= ke(3,5);

ke(5,4)= ke(4,5); ke(5,5)= ke(2,2); ke(5,6)= ke(2,3);

ke(6,1)= ke(1,6); ke(6,2)= ke(2,6); ke(6,3)= ke(3,6);

ke(6,4)= ke(4,6); ke(6,5)= ke(5,6); ke(6,6)= ke(3,3);
Note : el is element length

The kinetic energy of the composite wing will be written as

1
2 ∫∫∫
T= ρ w& 2 dxdydz
1 l 1 l (18)
= ∫ m (h& − xα α& ) 2 dy + ∫ I oα& 2 dy
2 0 2 0

where m is the mass of the wing cross section per span, xα is the distance from the
reference axis to the mass center of the wing cross section, and Io is the mass moment of
inertia of the wing cross section about the center of mass.

The kinetic energy can also be expressed in terms of the nodal displacements.

© 2001 by C. Nam, Y. Kim 7

 2 2
1 l  6 6
 1 l  6 
T=
2 ∫0
m  ∑
 i =1
ψ i ( y ) ⋅ q
& i − xα∑ i
i =1
ϕ ( y ) ⋅ q
& i 

dy +
2 ∫0
I o  ∑ ϕ i ( y ) ⋅ q& i  dy
 i =1 
1
q& ∫0 m {ψ i ( y ) − xα ϕ i ( y )}⋅ ψ j ( y ) − xα ϕ j ( y )dy{q&}
l
= (19)
2
1
+ q& ∫ I o {ϕ i ( y )}⋅ ϕ i ( y )dy{q&}
l

2 0

The moment of inertia about mass center, Io can be written as

I o = I α − m xα2 (20)

where Iα = m rα2 is the mass moment of inertia about the reference axis and rα2 is the
radius of gyration.

The kinetic energy equation can be rewritten as

1
T= q& [ M ]{q&} (21)
2

where [M] is the element mass matrix.

M ij = ∫ m {ψ i ( y ) − xα ϕ i ( y )}⋅ ψ j ( y ) − xα ϕ j ( y ) dy
l

+ ∫ I o {ϕ i ( y )}⋅ ϕ i ( y )dy
l

0
(22)
= ∫ m {ψ i ( y ) − xα ϕ i ( y )}⋅ ψ j ( y ) − xα ϕ j ( y ) dy
l

+ ∫ m (rα2 − xα2 ){ϕ i ( y )}⋅ ϕ i ( y )dy

l

V z
xα
α x

mass center

Figure wing cross section

The element of the mass matrix is expressed as

ml
M ij = M ij (23)
420

© 2001 by C. Nam, Y. Kim 8

The component of the M ij is as follows;

me(1,1)= 156.0+rg*(420.0*xa/el+504.0*rg*ra2/el2);
me(1,2)= 22.0*el+rg*(147.0*xa+252.0*rg*ra2/el);
me(1,3)=-147.0*xa-210.0*rg*ra2/el; me(1,4)=54.0-504.0*rg2*ra2/el2;
me(1,5)=-13.0*el+rg*(63.0*xa+252.0*rg*ra2/el);
me(1,6)=-63.0*xa-210.0*rg*ra2/el;

me(2,1)= me(1,2); me(2,2)=4.0*el2+rg*(42.0*xa*el+126.0*rg*ra2);

me(2,3)=-21.0*xa*el-105.0*rg*ra2;
me(2,4)= 13.0*el+rg*(63.0*xa-252.0*rg*xa2/el);
me(2,5)=-3.0*el2+126.0*rg2*ra2; me(2,6)=-14.0*xa*el-105.0*rg*ra2;

me(3,1)= me(1,3); me(3,2)=me(2,3); me(3,3)=140.0*ra2;

2
me(3,4)=-63.0*xa+210.0*rg*ra /el; me(3,5)=14.0*xa*el-105.0*rg*ra2;
2
me(3,6)= 70.0*ra ;

me(4,1)= me(1,4); me(4,2)=me(2,4); me(4,3)=me(3,4);

me(4,4)= 156.0+rg*(-420.0*xa/el+504.0*rg*ra2/el2);
me(4,5)=-22.0*el+rg*(147.0*xa-252.0*rg*ra2/el);
me(4,6)=-147.0*xa+210.0*rg*ra2/el;

me(5,1)= me(1,5); me(5,2)=me(2,5); me(5,3)=me(3,5);

me(5,4)= me(4,5);
me(5,5)= 4.0*el2+rg*(-42.0*xa*el+126.0*rg*ra2);
me(5,6)= 21.0*xa*el-105.0*rg*ra2;

me(6,1)=me(1,6); me(6,2)=me(2,6); me(6,3)=me(3,6);

me(6,4)=me(4,6); me(6,5)=me(5,6); me(6,6)=140.0*ra2;

The equation of motion of the composite box beam for the vibration analysis is expressed
in the following matrix form:

[ M ]{q&&} + [ K ]{q} = {0}

(24)
− ω 2 [ M ]{q} + [ K ]{q} = {0}

First, the free vibration analysis will be conducted and the equations of motion for flutter
analysis will be expressed in terms of the modal coordinates, {ξ}.

[ M ]{ξ&&} + [ K ]{ξ } = {0} (24)

where

[M ] = [Φ]T [M][Φ]
(25)
[K ] = [Φ]T [K][Φ]

© 2001 by C. Nam, Y. Kim 9

[Φ ] is the eigen matrix.

4.3 Unsteady Aerodynamic Forces

In order to calculate the unsteady aerodynamic coefficients for the three dimensional
wing model, incompressible 2 dimensional unsteady strip theory can be used. The strip
theory defines lift and moment forces along the midchord and acting on sections
perpendicular to the mid-chord line. As discussed in Chapter 3, the flow field is
represented by a noncirculatory component and a circulatory component. The vertical
displacement of the wing segment is expressed as

w(x, y,t) = h(y,t ) − xα (y,t)

where h is the displacement of the midchord (positive up), α is the rotation of a wing
section measured perpendicular to the midchord (positive nose up), and x is the distance
from the reference axis (aft positive). The lift and moment are written in terms of a
circulatory part and non-circulatory part. Using modified strip theory, outlined by Yates,
the lift (positive up) per unit length is given as

L = −πρb 2 {h&& + Vnα& + Vnσ& tan Λ − ba (α&& + Vnτ& tan Λ )}nc − {C lα ,n ρVn bC (k )Q}c (26)

dh dα
where σ = , τ=
dy dy

The subscript nc indicates the circulatory part and c indicates the circulatory part of the
unsteady aerodynamic forces. And a is location of the reference axis with respect to the
midchord, as a fraction of the semi-chord b and ρ is the density, b is the semi chord, Vn
(=VcosΛ) is the airspeed measured perpendicular to the midchord, and Λ is the sweep
angle. C lα ,n is the section lift coefficient and Λ is the sweep angle. And C(k) is the
Theoderson function. Q is the downwash distribution defined by

C
Q = {h& + Vn α + Vn σ tan Λ − b( lα ,n + a cn − a )(α& + Vn τ tan Λ )}
2π

where acn is the location of the aerodynamic center with respect to the mid-chord as a
fraction of the semichord, b.

© 2001 by C. Nam, Y. Kim 10

{Ref: Yates, E.C., “Calculation of Flutter Characteristics for Finite-Span Swept or
Unswept Wings at Subsonic and Supersonic Speeds by a Modified Strip Analysis,”
NACA RM L57110, March 1958.}

Similiary, the pitching moment (positive nose down) per unit length about the midchord
is

1
T = −πρb 4 {( + a 2 )(α&& + Vnτ& tan Λ ) + πρb 2Vn (h& + Vnσ tan Λ ) + πρb 3 a (h&& + Vnσ& tan Λ )
8
(27)
1 C lα ,n
+ πρb Vn (α − abτ tan Λ)}nc − 2πρb Vn [{ − (a − a cn )C (k )
2 2 2
}Q]c
2 2π

The virtual work due to a distributed lift and moment located at the reference axis is
written as

l c l
δW = ∫ ( ∫ pδwdx)dy = ∫ ( Lδh − Tδα )dy = ∑ Q j δq j (28)
0 0 0
j

where p is the pressure distribution, Qj is the generalized aerodynamic forces, and qj is

the generalized displacement.

For calculation of subsonic, compressible unsteady aerodynamic forces on harmonically

oscillating lifting surfaces, the most generally used scheme is the Doublet Latice Method
(DLM). The DLM has proven useful for both planar and nonplanar configurations of
multiple interacting lifting surfaces of fairly general geometry. In DLM, the surfaces are
represented by a grid of boxes of traperzoidal shape. At the quarter chord of each box is
located a line of pressure doublets with a constant strength which is the unsteady lift per
unit length of the line. The normalwash at the control points for all boxes in the gird is
computed using the kernel function representation of Watkins et. al.{Ref: Watkins, C.,
Runyan, H., Woolston, D., “On the Kernel Function of the Integral Equation Relating the
Lift and Downwash Distributions of Oscillating Finite Wings in Subsonic Flow,” NACA
Report 1234, 1955.} The control points are at the 3/4 chord point, mid-span for each box.

It is assumed that the flow is irrotational and inviscid so that the flow meets the Kutta

© 2001 by C. Nam, Y. Kim 11

conditions (i.e. finite velocity at the edge) to get around d’Alembert’s paradox. Also
there is no seperation.

The upwash normal to an oscillating lifting surface is expressed as {ref : Albano, E. and
Rodden, W.P., “A Doublet Lattice Method for Calculating Lift Distributions on
Oscillating Surfaces in Subsonic Flows,” AIAA Journal, Vol 7, Feb. 1969, pp279-285.}

w( x c , y c ) 1
U∞
=
8π ∫∫ K ( x c − x 0 , y c − y 0 , k , M )∆C p ( x 0 , y 0 )dx 0 dy 0 (29.a)

where k is the reduced frequency (=ωb/U∞) and M is the Mach number. K is called
Kernel function, which is actually the velocity field at (xc,yc) due to an elemental normal
force at (x0, y0).

The above integral equation can be approximated as follows;

w( x c , y c ) 1
U∞
=
8π
∑ (∆C
n
p )n ∫∫ K ( x c − x 0 ,y c − y 0 , k , M )dx 0 dy 0 (29.b)
element , n

The integration of Kernel function in the streamwise direction is done simply by lumping
the effect into a loaded line at the ¼ chord line of the element. The results is an unsteady
horseshoe vortex whose bound position lies along the ¼ chord line of the element. There
is one control or receiving point per element and the surface normalwash boundary
condition is satisfied at each of these points. The control point is centered spanwise on
the ¾ chord line of the element. The upwash equation can be written in matrix form;

{w}3 / 4 = [ AIC ]{∆C p } (30)

In the matlab code attached at this book, DLM is used to calculated the unsteady
aerodynamic coefficient matrix, [AIC] and the result is saved for the further reduction
into the modal coordinate system.

The virtual work due to aerodynamic forces can be rewritten as follow;

© 2001 by C. Nam, Y. Kim 12

δW A = q d δw c / 4 [ S ]{∆C p }
(31)
= q d δw c / 4 [ S ][ AIC ] −1{w}3c / 4

qd is the dynamic pressure, a diagonal matrix [S] denotes the panel area. And δwc / 4 ,
{w}3c / 4 denote the virtual displacement at 1/4 chord, the downwash at 3/4 chord of the
panel, respectively. With harmonic motion assumption, the downwash is expressed as
the following;

ω
{w}3c / 4 = [α ij ]3c / 4 {q} − i [Wij ]3c / 4 {q} (32)
V

where

 ∂Z j 
[α ]
ij 3c / 4 = − 
 ∂x  i ( 3c / 4 ) (33)
[W ]
ij 3c / 4 [ ]
= Zj i ( 3c / 4 )

Zj is the displacement function in z-direction and {q} is the generalized coordinate.

Similarly, the virtual displacement is written as

{w}c / 4 = [ Z j ]i ( c / 4 ) {q} (34)

Then, the virtual work becomes

 ω 
δW A = q d {δq}Z j Ti ( c / 4) [ S ][ AIC ] −1 [α ij ]3c / 4 − i [Wij ]3c / 4 {q}
 V  (35)
= q d {δq}[ A]{q}

It has to be noted that the aerodynamic forces, [A] is function of the reduced frequency
and Mach number.

An executable code, unsdlm.exe is provided for the calculation of the unsteady

aerodynamic coefficient. This code will need the input data file, unsdlm.dat.

List of unsdlm.dat : This input data is used in example. For input variables, refer
figure.
%%% Title : text
Composite wing with I/O Control Surfaces(DLM-GEN)
%%% warea,swep,taper,aspct,y1bar,y2bar,l1bar,l2bar,ftc1,ftc2

© 2001 by C. Nam, Y. Kim 13

1800.0 25.0 0.8 2.48 0.15 0.45 0.25 0.3 0.25 0.25
%%% mach,nk,nctotal,nstotal
0.0 20 6 15
%%% (freq(i), i=1,nk)
0.8 0.7 0.65 0.6 0.55 0.5
0.45 0.4 0.35 0.3 0.275 0.25
0.225 0.2 0.175 0.15 0.125 0.1
0.05 0.0
% End of DATA
NOTE:
warea : wing area
swep : sweep angle(Deg)
taper : taper ratio
aspct : aspect ratio
y1bar : (inboard location/span) of the 1st control surface
l1bar : (length/span) of the 1st control surface
ftc1 : control surface-to-chord ratio of the 1st control surface
y2bar : (inboard location/span) of the 2nd control surface
l2bar : (length/span) of the 2nd control surface
ftc2 : control surface-to-chord ratio of the 2nd control surface
mach : mach number
nk : no. of redeuced frequencies
nctotal : no. of panels in chordwise direction
nstotal : no. of panels in spanwise direction
freq(1~nk) : reduced frequencies

If model does not have control surfaces set ftc1, ftc2 to 0.2 !!

© 2001 by C. Nam, Y. Kim 14

 Y

V
Lo

Control
Λ Surfaces

Yo
Li

Yi θ X

Figure . Composite wing model with control surfaces.

Matlab file, aero.m, will read the aerodynamic coefficient for various reduced
frequencies, and calculate the generalized aerodynamic forces in the modal coordinate
system using the procedure that we discussed in this section.

List of aero.m
%
function [ictrl,jnk,b0,rka,raero] = aero(igust,ndof,need)
% reads the unsteady aerodynamic influence coefficient matrix
% supplied by a doublet lattice method
%
% computes the generalized force matrix in modal coordinates
% using doublet lattice aerodynamics
%
%
fid_16= fopen('aeroshp.out','w');
file_20 = fopen(['dlmaero.mat']);
deg2rad=360.0d0/(2.0d0*pi);
hhc=zeros(100,10);
jk=sqrt(-1);

© 2001 by C. Nam, Y. Kim 15

ictrl = fscanf(file_20,'%d',1);

msize=need;
if(igust == 0), msize1=msize+ictrl; end
if(igust == 1), msize1=msize+ictrl+1; end

if(ictrl ~= 0),
for i=1:ictrl,
ipnl(i) = fscanf(file_20,'%d',1);
npnl=ipnl(i);
tmp= fscanf(file_20,'%d',npnl);
ipanel(i,1:npnl)=tmp';
%
% xhng, yhng : hinge locations of the control surfaces
%
% x1: x-coord of leading edge/root of the wing.
% x2: x-coord of trailing edge/root of the wing.
% x3: x-coord of leading edge/tip of the wing.
% x4: x-coord of trailing edge/tip of the wing.
% y1: y-coord of the wing root.
% y2: y-coord of the wing tip.
tmp= fscanf(file_20,'%g',6);
x1=tmp(1);x2=tmp(2);x3=tmp(3);x4=tmp(4);y1=tmp(5);y2=tmp(6);
tmp= fscanf(file_20,'%g',2);
xhng(i)=tmp(1);yhng(i)=tmp(2);
%
%
swpl=atan((x3-x1)/(y2-y1));
swpt=atan((x4-x2)/(y2-y1));
aswp(i)=(swpl+(swpt-swpl)*xhng(i))*deg2rad;
xl=x1+(x3-x1)*yhng(i);
xt=x2+(x4-x2)*yhng(i);
xhng(i)=xl+(xt-xl)*xhng(i);
yhng(i)=y1+(y2-y1)*yhng(i);
end
end
%
rkk2 = 1.;
tmp= fscanf(file_20,'%g',2);
nchrd=tmp(1);nspan=tmp(2);
tmp= fscanf(file_20,'%g',3);
ncll=tmp(1);jnk=tmp(2);b0=tmp(3);
b0 = b0/2.;
%
%
tmp= fscanf(file_20,'%g',2*ncll);
for m=1:ncll, xb(m)=tmp(2*m-1);yb(m)=tmp(2*m);end
% upwash & nosedown twist (h,r) w.r.t. 3/4 chord
%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
[h,r]=tranm(fid_16,msize,ncll,yb,xb);
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
for i = 1:ncll
for j = 1:msize
tpr(i,j)=r(i,j)+jk*h(i,j)/b0 ;
end

© 2001 by C. Nam, Y. Kim 16

 hh(i)=xb(i)/b0;
end
%
% for each control surface find the upward displacement
%
if(ictrl ~= 0),
for ic1=1:ictrl
npnl=ipnl(ic1);
cswp=cos(aswp(ic1)/deg2rad); sswp=sin(aswp(ic1)/deg2rad);
for ic2=1:npnl
icp=ipanel(ic1,ic2);
hhc(icp,ic1)=sswp*(xb(icp)-xhng(ic1))-cswp*(yb(icp)-yhng(ic1));
end
end
end
tmp= fscanf(file_20,'%g',ncll);
for m=1:ncll, xb(m)=tmp(m);end
%
% upwash & nosedown twist (h,work) w.r.t. 1/4 chord
%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
[h,work]=tranm(fid_16,msize,ncll,yb,xb);
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
tmp= fscanf(file_20,'%g',ncll);
for m=1:ncll, ar(m)=tmp(m); end
%
for nt=1:jnk
rk2= fscanf(file_20,'%g',1);
rka(nt)=rk2;
tpr=real(tpr)+jk*imag(tpr)*(rk2/rkk2);
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
fprintf(fid_16,'reduced frequency %g \n',rk2);
fprintf(fid_16,'the upwash due to elastic mode\n');
fprintf(fid_16,'%10.5e %10.5e %10.5e %10.5e %10.5e %10.5e %10.5e
%10.5e %10.5e %10.5e \n',tpr);
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%
% control mode
%
if(ictrl ~= 0),
for ic1=1:ictrl,
cswp=cos(aswp(ic1)/deg2rad); npnl=ipnl(ic1);
for ic2=1:npnl,
icp=ipanel(ic1,ic2); rr1=-cswp;aa1=rk2/b0*hhc(icp,ic1);
tpr(icp,msize+ic1)=(rr1+jk*aa1);
end

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
fprintf(fid_16,'reduced frequency %g \n',rk2);
fprintf(fid_16,'the upwash due to the control surface no %g \n',ic1);
fprintf(fid_16,'%10.5e %10.5e %10.5e %10.5e %10.5e %10.5e %10.5e
%10.5e %10.5e %10.5e\n',tpr(:,msize+ic1));
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

end
end
%

© 2001 by C. Nam, Y. Kim 17

% gust
% ***********************************
% tpr is the upwash due to the gust
% ***********************************
if(igust == 1),
for k =1: ncll,
gcos=cos((rk2)*hh(k));gsin=sin((rk2)*hh(k));
tpr(k,msize1)=gcos-jk*gsin;
end
%%%%%%%%%%%%%%%%%%%%
fprintf(fid_16,'reduced frequency %g \n',rk2);
fprintf(fid_16,'the upwash due to the gust\n');
fprintf(fid_16,'%10.5e %10.5e %10.5e %10.5e %10.5e %10.5e
%10.5e %10.5e %10.5e %10.5e\n',tpr(:,msize1));
end
%
for i = 1: ncll,

tmp= fscanf(file_20,'%g',2*ncll);
for m=1:ncll, a(m)=tmp(2*m-1)+jk*tmp(2*m);end

for j =1:msize1,
hsb(i,j) = 0;
for k=1:ncll, hsb(i,j)=hsb(i,j)+a(k)*tpr(k,j);end
end
for j=1:msize1, hsb(i,j)=hsb(i,j)*ar(i);end
end
%
for i=1:msize,
for j=1:msize1,
hat(j) =0;
for k =1:ncll, hat(j)=hat(j)+h(k,i)*hsb(k,j);end
end
for j=1:msize1,
cain(i,j)=-hat(j);
raero(nt,i,j)=cain(i,j);
end
end
%
%
rkk2 = rk2;
end
fclose(file_20);
fclose(fid_16);
%

4.4 Mathematical Aeroservoelastic Model

The goal of this chapter is to conduct an aeroservoelastic wing planform design by using
the control surfaces to suppress flutter and to reduce the gust response of the wing.

The equations of motion for aeroservoelastic analysis can be written as

© 2001 by C. Nam, Y. Kim 18

 { }
[ M ]{q&&} + [C ]{q&} + [K ]{q} + [M c ] δ&&c = [Qa ]{q} + [Qc ]{δ c } + [QG ]
wG
V
 w 
= q d  [Aa ]{q} + [Ac ]{δ c } + [ AG ] G  (36)
 V 

where {q&&} are the generalized modal coordinates, q d is the dynamic pressure. The
matrices [Qa ] , [Qc ] and [QG ] are the generalized aerodynamic forces due to flexible
modes, control surfaces deflection and gust, respectively. After vibration analysis, a
modal reduction is performed using the selected elastic modes.

The aerodynamic forces are approximated as the transfer functions of the Laplace
variable by a least square curve fit approximation in order to define the aeroservoelastic
equations of motion in a linear time invariant state-space form. As discussed in the
Chapter 3, Roger’s method approximates the unsteady aerodynamic forces in the
following form.

[A ] = [A A A ] = [P ]+ [P ]s'+[P ]s' +∑ s['+Pγ ]s'

N
2 j
ap ξ c g 0 1 2 (37)
j =3 j −2

[ ]
where P = Pξ Pc Pg , s ' = ik = iωb / V = sb / V and s is the Laplace variable, k is the
reduced frequency, b is the semi-chord and V is the airspeed. The subscripts ξ, c and g
indicate elastic, control surface and gust modes, respectively and γ are the aerodynamic
poles that are usually preselected in the range of reduced frequencies of interest. In this
study, four terms of the aerodynamic poles are included for the rational functions
approximation. The augmented aerodynamic state is defined as follows.

ξ 
{x ja } = s'
s '+γ j −2
[  
]
Pjξ Pjc Pjg  δ c  , j=3,4,5,6. (38)
w 
 g

It has to be noted that P2 g is set to be zero to avoid w

&& g term in the state equation.
Therefore,

[A ] = [P ] + [P ]s'+∑ s['P+γ ]s'

N
js
g 0s 1s (39)
j =3 j −2

© 2001 by C. Nam, Y. Kim 19

Using the rational function approximation, the system equations of motion is written in
the following form.

{x& s } = [As ]{x s } + [Bs ]{u s } + [BG ]{wG } (40)

where

 0 I 0 .. 0 
~ −1
− M ~ −1 ~ −1 ~ −1 
 K −M B qd M I .. qd M I 
 V  
[ As ] =  0 P3ξ −  γ 1 I .. 0
b 
 .. .. .. .. 0 
 
 0 V 
PNξ 0 0 −  γ N − 2 I 
 b 

 0 0 0 
 ~ −1 ~ −1  b  ~ −1  b  2 
q d M P0 c q d M P1c   q d M P2 c   
V  V  
[ Bs ] =  0 P3c 0
 
 .. .. .. 
 
 0 PNc 0 

 0 0 0 
 ~ −1 ~ −1  b  ~ −1  b  2 
q d M P0 g q d M P1g   q d M P2 g   
V  V  
[ BG ] =  0 P3 g 0 
 .. .. .. 
 
 0 PNg 0 

2
~ b
M = M − q d P2ξ  
V 
~ b
B = B − q d P1ξ  
V 
~
K = K − q d P0ξ

{x s } = ξ T ξ& T xaT T

© 2001 by C. Nam, Y. Kim 20

 {u s } = δ c1 δ&c1 δ&&c1 δ c 2 δ&c 2 δ&&c 2 T

{wG } = wg ,w& g T

In the above equations, xa denotes the aerodynamic states and δ c is the control surface
deflection. The control surfaces actuator transfer functions can be expressed in a state
space form as follows.

{x& c } = [Ac ]{xc } + {Bc }{δ aileron }

(41)
{u s } = [C c ]{xc }

The actuator transfer function for each of the control surfaces is fixed as follows.

{δ c } a a0
= (42)
{δ aileron } s + a s + a1 s + a 0
2

The gust state space model is included for random gust response calculations. The
vertical gust is modeled by a second order Dryden model;

3V  V 
w
s + 
g L  L 3
= σ wg
w (43)
 V
2

s + 
 L 

where σ wg is the RMS value of the gust velocity, L is the characteristic gust length and V
is the airspeed. When the low pass filter (a/(s+a)) is included, the state space equation of
the gust is expressed as follows.

 x& g1   0 1 0   x g1   0 
      
a   x g 2  +  0 w
−2 −1
 x& g 2  = − τ g − 2τ g
3
 x&   0 0 − a   x g 3  τ g 2 σ wg 
−
 g3  
x 
α   1 3τ g 0   g1 
{wG } =  & g  =  −1
α g  − 3τ g 1 − 2 3 ( ) x 
3τ g a   g 2 
xg3  (44)
L
where τ g = .
V
Or

© 2001 by C. Nam, Y. Kim 21

 {x& } = [A ]{x }+ {B }w
g g g g
(45)
{w } = [C ]{x }
G g g

By including the gust dynamics system and the actuator system for the control surface,
the following state space aeroservoelastic model is obtained.

{x&} = [A]{x} + [B]{u} + {Bw }w (46)

{y} = [C ]{x}

where {x} = xsT xcT x Tg  . The resulting state vector is consists of elastic modes,
T

aerodynamic states, actuator states and 3 gust states including low pass fiter.

4.5 Controller Design - Pole placement technique

For a given design airspeed, a controller for active flutter suppression is designed. The
performance index value (control effort) is used as a measure of control performance.
The following output feedback control law is introduced.

{u} = −[K ]{y}G

(47)

where [ ] denotes output feedback gain and {y} includes only structural information.
K
G

The eigenvalue problem of the closed-loop system can be written as

([A] − [B][K ][C ]){φ } = −λ {φ }

G
c
i
c
i
c
i (48)

where {φc }i is the eigenvector of the closed-loop system corresponding to the eigenvalue
c
λi . The eigenvalues of the structural modes can be assigned to desired values by using a
pole placement technique.

The problem can be stated as a nonlinear parameter optimization problem in which it is

necessary to impose specified eigenspace equality constraints and the elements of output
feedback gain matrix are the parameters to be determined. Defining {p} as the parameter
vector that consists of the elements of output feedback gain [ ], the nonlinear
K
G

programming problem can be formulated as follows.

Determine output feedback gain parameter {p}
subject to

© 2001 by C. Nam, Y. Kim 22

 f i ({p}) = λdi − λi ({p}) = 0,i = 1,2,.., Ns
where f i ({p}) is the equality constraint equation that is related to the closed-loop
d
eigenvalue assignment, Ns is the number of structural modes and λi denotes the desired
eigenvalues of these modes.

The above problem can be solved by using any gradient-based nonlinear programming
algorithm. The homotopic nonlinear programming with minimum norm correction
algorithm is used. To enhance convergence, the linear homotopy map is generated. The
original eigenvalue assignment problem is replaced by the one-parameter α family as
follows.

f i ({p}) = αλdi + (1 − α ) λ i ({pstart }) − λ i ({p})

= 0,
0 ≤ α ≤1 (49)

where {pstart } is the initial starting value of the parameter vector. Sweeping α using a
suitably small increment generates a sequence of neighboring problems. These sequence
of problems are solved using the neighboring converged solutions to generate starting
iteratives for each subsequent problem. The desired solution is reached once solution for
α=1 is obtained. More details about this algorithm can be found in chapter 2.

The control performance canbe investigated by comparing a quadratic type energy

function

J =∫
0
∞
{{y} [Q]{y} + {u} [R]{u}}dt
T T
(50)

where the positive definite weight matrices [Q] and [R] should be related to make the
integrands correspond to physical energy measures. Note that the first term in J is related
to the error energy of the structural modes and the second term is related to the control
energy. This objective function is obtained as follows

J = tr ([P ][ X 0 ]) (51)

where [X0] is an initial autocorrelation of the state and [P] is the solution of the following
Lyapunov equation.

[P][Ac ] + [Ac ]T [P] + [C ]T ([Q] + [K G ]T [R][K G ])[C ] = 0 (52)

© 2001 by C. Nam, Y. Kim 23

where

[Ac ] = [A] − [B][K G ][C ] (53)

The solution of the Lyapunov equation can be obtained by solving eigenvalue problem of
the closed-loop system matrix. However, it is unstable when the eigensystem is ill-
conditioned. Since aeroservoelastic system dealt with in this study is ill-conditioned,
Lyapunov solver, utilizing Schur decomposition, is used. This algorithm transforms the
closed-loop system matrix to Schur form by orthogonal similarity transformation,
computes the solution of the resulting triangular system and transforms this solution
back.

Another purpose of the control system is to prevent performance degradation due to

external disturbances such as gust. For the gust response reduction system, the root mean
square (RMS) values of gust response for the different elastic modes can be calculated for
performance comparison. The state covariance matrix [X] of the closed-loop system is
the solution of a Lyapunov equation in the form.

[Ac ][X ] + [X ][Ac ]T + {Bw }[Qw ]{Bw }T =0 (54)

where [Qw] is the intensity matrix of the white noise. The square of the RMS of the
system outputs(the modes of the system) is computed as follows.

[
σ i2 = [C ][X ][C ]
T
] ,i = 1,2,.., Ns
ii
(55)

4.6 Example : Gust Response Reduction System Design

A composite wing model with inboard/outboard control surfaces is used as an example

model for aeroelastic control system design. The model has 25 degrees of sweep angle
and a taper ratio and aspect ration of the wing are 0.8 and 2,48, respectively. The wing
area is set to be 1800 in2. The wing skin is composed of 14 symmetric composite layers,
which are [-90/+45/-45/-40/+45/-45/-90]s. Each layer has uniform thickness, 0.008, 0.01,
0.01, 0.104, 0.01, 0.01 and 0.008 inches, respectively. The inboard control surface is
located at 15% of span. The spanwise size of the control surface is 25% of span (i.e.,
Yi=15%, Li=25% of the span). The outboard control surface is located at 45% of span.

© 2001 by C. Nam, Y. Kim 24

The spanwise size of the control surface is 30% of span (i.e., Yo=45%, Lo=30% of the
span). The chordwise size of the each control surface is set to be 25% of chord.

The actuator dynamics for each control surface is defined as

{δ c } 20 1.6 × 10 5
=
{δ aileron } s + 20 s 2 + 400s + 1.6 × 10 5

The modal reduction is conducted using first 6 vibration modes. The first 6 natural
frequencies of the mode are 6.1, 15.0, 26.8, 38.7, 53.6, 72.3 Hz, respectively. For RFA
using Roger’s method, 4 aerodynamic poles are defined. These are 0.2, 0.4, 0.6, and 0.8.
Therefore, the resulting aeroservoelatic state space model is 45th order including 6 elastic
modes, 24 aerodynamic states, 6 actuator states and 3 gust states.

For this wing model, open loop flutter analysis is conducted. Flutter occurs at 1850
feet/sec(fps) and is due to the second mode (13.5 Hz).

1 500

450
0.5
400
0
350
-0.5
300
Imaginary Part
Real Part

-1 250

200
-1.5

150
-2
100
-2.5
50

-3 0
0 500 1000 1500 2000 2500 0 500 1000 1500 2000 2500
Velocity Velocity

200

180

160

140

120
Imaginary Part

100

80

60

40

20

0
-3 -2.5 -2 -1.5 -1 -0.5 0 0.5 1
Real Part

Figure Open Loop Flutter Analysis Result

© 2001 by C. Nam, Y. Kim 25

The control system is designed at the design speed of 2000 fps. Table shows the open
loop system eigenvalues at design speed. As shown in Table, 2nd mode becomes unstable.
The desired eigenvalues are defined in order to design a control system using pole
placement technique.

Table : Open Loop Eignvalues and Desired Closed Loop Eigen Values.

Open Loop Eigenvalues Desired Closed Loop Eigenvalues

of Six Elastic Modes
-6.5677e+000 +5.9379e+001i -7.2245e+000 +5.9379e+001i
3.2996e-001 +8.7866e+001i -4.6195e-001 +8.7866e+001i
-1.7126e+000 +1.7084e+002i -1.7126e+000 +1.7084e+002i
-1.2166e+000 +2.3888e+002i -1.2166e+000 +2.3888e+002i
-5.2780e-001 +3.3429e+002i -5.2780e-001 +3.3429e+002i
-7.8903e-001 +4.5397e+002i -7.8903e-001 +4.5397e+002i

The gain matrix is determined as

Columns 1 through 6
7.6267e-003 -4.6367e-003 -8.8729e-003 -5.0137e-002 5.8575e-004 1.0342e-003
3.1175e-002 -2.4089e-002 -4.6352e-003 -3.8585e-002 1.2787e-003 -5.7431e-005

Columns 7 through 12
8.1452e-002 8.0865e-002 -1.3649e-002 -3.6692e-002 -2.3506e-003 1.2538e-003
-5.2469e-002 -7.4833e-002 2.1801e-002 5.8818e-002 1.3415e-002 -5.4627e-005

The RMS value of the gust response for the flexible modes can be used as a metric to
compare the control performance.

List of Test.m
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%%
%
clear all
close all

% NOTE : If you want to change one of the following,

% you need to change input data for unsteady aero code.
% check unsdlm.dat
swep =25.0; taper =0.8; aspct =2.48;
y1bar =0.15; y2bar =0.6;

© 2001 by C. Nam, Y. Kim 26

 l1bar =0.25; l2bar =0.3;
ftc1=0.25; ftc2=0.25;
%*wing area
warea=1800.0;
% tc = wing thickness to chord (constant across span)
tc=0.38;
% nele : number of elements
nele=10;

save wingdvs swep taper aspct y1bar y2bar l1bar l2bar ftc1 ftc2
%
% composite laminated plate
% nl is the number of layer
nl=14;
%
layerdata =[ 1.0000 0.9920 -90.0000 0.0080
2.0000 0.9820 45.0000 0.0100
3.0000 0.9720 -45.0000 0.0100
4.0000 0.8680 -40.0000 0.1040
5.0000 0.8580 -45.0000 0.0100
6.0000 0.8480 45.0000 0.0100
7.0000 0.8400 -90.0000 0.0800
8.0000 -0.8480 -90.0000 0.0080
9.0000 -0.8580 45.0000 0.0100
10.0000 -0.8680 -45.0000 0.0100
11.0000 -0.9720 -40.0000 0.1040
12.0000 -0.9820 -45.0000 0.0100
13.0000 -0.9920 45.0000 0.0100
14.0000 -1.0000 -90.0000 0.0800 ];
%
% wti= non-structural weight of the wing per unit span
% wti is unit of slugs/in
wti=[0.01345
0.01111
0.01099
0.01172
0.01134
0.01051
0.00681
0.00502
0.00502
0.00502];
%
save dataprop nl layerdata wti
%
%
getinput
%
% create the mass and stiffness matrices
%
model
%
% solve eigenvalue problem
%
modes
%
% ================================

© 2001 by C. Nam, Y. Kim 27

% number of modes in reduced model
% ================================
%
nmodes=6;
rmass =eivec(:,1:nmodes)'*gmass *eivec(:,1:nmodes);
rstiff=eivec(:,1:nmodes)'*gstiff*eivec(:,1:nmodes);
%
%
% calculate the unsteady aerodynamic forces
%
disp('Reading Unsteady Aerodynamic Forces ...')
iopen=input('Do you need to run aero code ? (y/n)=(1/0) ');
%
if (iopen == 1),
!unsdlm
!prepost
end
%
%
% gust mode
disp('Let us include the gust mode. please type 1')
igust=input('Do you want to include gust ? (y/n)=(1/0) ');
%
% ictrl : no. of control surfaces
need=nmodes;
save tranm eivec nele yb1 yb2 xroot yroot sweep
disp('Converting aerodynamic force matrix in modal coordinate....')
[ictrl,jnk,b0,rka,raero] = aero(igust,ndof,need) ;

if(igust==0),
need1=need+ictrl;
elseif(igust==1),
need1=need+ictrl+1;
end
%
% rational function approximation: roger's method
%
xxx(1)=0.2;xxx(2)=0.4;xxx(3)=0.6;xxx(4)=0.8;
%
disp( 'writing rfa results .....')
[p0,p1,p2,p3,p4,p5,p6]=roger(igust,ictrl,need,jnk,xxx,rka,raero);
disp( 'rfa.==>done, assemble system matrix...')
%
% icls :0 for open loop , 1 for closed loop analysis
disp('icls :0 for open loop , 1 for closed loop analysis ')
icls=input('Do you want to design closed loop system ? (y/n)=(1/0)
');
%
%
% i0vel : initial velocity (ft/sec)
% j0vel : final velocity (ft/sec)
% istep : increament in velocity
%
if (icls == 0 ),
% open loop analysis
i0vel=100; j0vel=2500; istep=50;
elseif (icls == 1 ),

© 2001 by C. Nam, Y. Kim 28

% set design airspeed
ivel=input('set design airspeed = (2000)');
i0vel=ivel; j0vel=ivel; istep=100;
end
%
%
% low pass filter for gust mode
% input chlng,afrq,sigwg
% chlng : characteristic length(inch)
% afrq (rad/sec) : a of the low pass filter a/(s+a)
% sigwg : sigma_wg
%
if(igust ~= 0),
sigwg=1; chlng=50; afrq=160;
end
% input actuator dynamics
% 'numerator',xnctrln(i)
% 'denominator'
%
if(ictrl ~= 0),
for i=1:ictrl
xnctrln(i)=32.0e06;
dnctrl3(i)=1; dnctrl2(i)=420; dnctrl1(i)=16.08e6; dnctrl0(i)=32.0e6;
end
disp('actuator dynamics')
[xnctrln']
[dnctrl3' dnctrl2' dnctrl1' dnctrl0']
end

disp( 'Open Loop System Analysis ...')

system

if (icls ==1),
save openloop ivel a_sys b_sys bw c_sys eivec gsystem
end

figure(1)
plot(solopn(:,1),solopn(:,2),'*r'),grid on,axis([0 2500 -3 1]),
xlabel('Velocity'),ylabel('Real Part')

figure(2)
plot(solopn(:,1),solopn(:,3),'*r'),grid on,axis([0 2500 0 500]),
xlabel('Velocity'),ylabel('Imaginary Part')

figure(3)
plot(solopn(:,2),solopn(:,3),'*r'),grid on,axis([-3 1 0 300]),
xlabel('Real Part'),ylabel('Imaginary Part')

if (icls ==1),
disp('open loop system data is creadted at ivel !')
disp('If you want to design a control system, just run clsdsn !')
end

List of Test.m

© 2001 by C. Nam, Y. Kim 29

% Pole placement technique;
% CHeck eigsrt.m file and main file before you design a control system
!!

load openloop ivel a_sys b_sys bw c_sys eivec gsystem

A=a_sys;B=b_sys;C=c_sys;
n=max(size(A));
% number of dsesired mode = elastic modes
neigd=6;
eigd=zeros(n,1);
[eigop,phi0,psi0]=eignsort(A);
[eigop,phi0,psi0]=select(eigop,phi0,psi0,neigd);
%
% set desired closed loop eigenvalues
%
for i=1:neigd;
eigd(2*i-1)=-abs(real(eigop(2*i-1)))+sqrt(-1)*imag(eigop(2*i-1));
if(i==1);
eigd(2*i-1)=-1.1*abs(real(eigop(2*i-1)))+sqrt(-1)*imag(eigop(2*i-1));
end;
if(i==2);
eigd(2*i-1)=-1.4*abs(real(eigop(2*i-1)))+sqrt(-1)*imag(eigop(2*i-1));
end;
eigd(2*i)=conj(eigd(2*i-1));
end
%
% design controller using nonlinear programming

G=polenlps(A,B,C,eigd,neigd);

[eigcl,phicl,psicl]=eignsort(A-B*G*C);
[eigcl,phicl,psicl]=select(eigcl,phicl,psicl,neigd);
%disp('open loop eigenvalues')
%disp([eigop(1:2*neigd)])
disp('desired and closed eigenvalues')
disp([eigd(1:2*neigd) eigcl(1:2*neigd)])

%[eigop eigcl]
disp('gain matrix is')
disp([G])

% Closed Loop System at Design Airspeed

Ac=A-B*G*C;

disp('open loop/closed loop eigenvalues')

[eignsort(A) eignsort(Ac)]

sigma =rms(A ,B,C,bw);

sigmac=rms(Ac,B,C,bw);
disp('RMS values of the open/closed loop system due to disturbance')
[sigma sigmac]

© 2001 by C. Nam, Y. Kim 30

Chapter 5
Aeroservoelastic Design of Composite Plate-Wing
Using Piezoelectric Materials

To prevent the static and dynamic instability in a composite wing, we can passively or
actively suppress the instability by a proper choice of ply orientation and actuating the
control surface. Recently, the application of piezoelectric materials to structural control
may provide a new dimension in design to eliminate the possibility of instability by
changing wing configuration to cause lift distribution variation. There has been a
considerable amount of research activity in using piezo sensor/actuators for the vibration
control of flexible structures. Since the piezoelectric materials exhibit elastic
deformation in proportion to the magnitude of an applied electric field and transfer
forces to the structure, piezoelectric materials that are bonded at proper locations of a
base structure can be used as actuators.

An analytic model for an induced strain actuator coupled with beams and plates has been
developed by many authors. The static and dynamic aeroelastic behavior of a wing
structure with piezo actuators has been studied. Ehlers and Weisshaar showed the effects
of the piezoelectric actuators on the static aeroelastic behavior such as lift effectiveness,
divergence, roll effectiveness. Lazarus, Crawley and Bohlmann used the typical section
to represent a wing model and performed an analytic study to find the optimal placement
and thickness of a piezo actuator for camber or twist control. Lazarus, Crawley and Lin
showed the ability of both articulated control surfaces and piezoelectric actuators to
control the dynamic aeroelastic systems. Heeg conducted an experimental study for
flutter suppression of a beam model and demonstrated the application of piezoelectric
materials for flutter suppression, which is affixed to a flexible mount system.

In this chapter, an aeroservoelastic system will be designed using the piezoelectric

materials as actuators. The purpose of this chapter is to design an active control system
for flutter suppression of a composite plate-wing with segmented piezoelectric actuators.
The analysis of the laminated composite wing with segmented piezoelectric actuators is
conducted by Rayleigh-Ritz method. The active control system design for flutter

© 2001 by C. Nam, Y. Kim 1

suppression requires the equations of motion to be expressed in a linear time-invariant
state-space form. Therefore, it is necessary to approximate the unsteady aerodynamic
forces in terms of rational functions of the Laplace variables. Minimum state method will
be introduced to approximate the frequency domain unsteady aerodynamic forces as the
transfer functions of the Laplace variable.

© 2001 by C. Nam, Y. Kim 2

5.1 Modeling of Wing Structures with Piezoelectric Elements

V
L
Y

2 4
C
3
θ
1

X
Z Piezoelectric Actuators
Y

Figure Composite Plate-Wing with Pizeoelectric Actuatrors

Based on the classical laminated plate theory, the equations of motion of a model with
piezo actuators are formulated. The linear coupled electromechanical constitutive
relations of a piezoelectric materials can be written as;

{D} = [d ]T {T } + [ε t ]{e}
(1)
{S} = [ s e ]{T } + [d ]{e}

{S} is the strain, {T} is the stress, {D} is the electric displacement, and {e} is the electric
field intensity. [d],[εt],[se] denote the piezoelectric constant, dielectric constant, and
elastic compliance matrix, respectively. It is assumed that the "3" axis is associated with
the direction of poling and the mechanical property of the piezoelectric materials is
modeled as an isotropic material. For the thin plate shape of the piezoelectric materials,
we can rewrite Eq. (1).

© 2001 by C. Nam, Y. Kim 3

 0σ 1 σ 2 σ 12  + ε 33
T
D3 = d 31 d 32 t
e3
 ε1  S p S p12 0   σ 1   d 31 
   11     (2)
ε2  =  S p12 S p 22 0   σ 2  + d 32 e3
γ   0 0 S p 66  σ 12   0 
 12  

And, the stress-strain relations of a thin piezoelectric layer are

 σ1  Q p Q p12 0   ε1  d 31  
   11      
σ 2  = Q p12 Q p 22 0   ε 2  − d 32 e3  (3.a)
σ   0 0 Q p 66  γ 12   0  
 12  

or

{σ } p = [Q p ]{{ε } − {Λ}} (3.b)

[Qp] is the stiffness matrix of piezoelectric layer, {dij} is the piezoelectric strain
coefficient, and e3 is the applied electric field. This equation is similar to stress-strain
equation with thermal effects considering the fact that the piezoelectric strain Λ has the
same form as thermal strain α∆Τ.

Inplane forces and moments of laminated plate including the loads of the piezo actuators
are obtained by integrating the stresses over the ply thickness, and can be expressed as
follows,

{N } = ∫ σdz = [ A]{ε } + [ B]{κ } − {N Λ } (4.a)

{M } = ∫ σzdz = [ B]{ε } + [ D]{κ } − {M Λ } (4.b)

where ε, κ are the midplane strain and curvature, respectively. [A], [D], and [B] are
extension, bending, and extension/bending coupling stiffness matrices, respectively.
These matrices are influenced by not only the ply orientations of the laminate but also the

© 2001 by C. Nam, Y. Kim 4

geometry of the piezo actuators. Inplane forces and moments due to actuator strain are
given by

{N Λ } = ∫ [Q p ]{Λ}dz (5)
zp

{M Λ } = ∫ [Q p ]{Λ}zdz (6)
zp

where zp is the coordinate through thickness in the piezoelectric material.

The strain energy of plate with piezoelectric material can be expressed as

1  A B  ε  ε 
U= ∫∫
2 A, A p
ε κ  B D  κ dxdy − ∫∫ N Λ M Λ κ dxdy
   Ap  
(7)

The kinetic energy is

1
T= ∫∫
2 A, A p
ρ (u& 2 + v& 2 + w& 2 )dxdy (8)

where A, Ap represent the area of the composite plate and piezoelectric material,
respectively, u,v,w are displacements in x, y, z direction. These energies are the
functions of the ply orientations of the laminate and the piezoelectric parameters, which
are the actuator thickness, area, and locations.

It is impossible to obtain closed form solutions to the laminated plate due to the
complexity of equation, the arbitrary boundary conditions, and external forces such as
unsteady aerodynamic forces. In this study, the Rayleigh-Ritz method which is faster and
has the same accuracy as the finite element method is adopted for the structural analysis.
To apply the Rayleigh-Ritz method, we introduce displacement functions in generalized
coordinate system to represent displacements u, v, w

© 2001 by C. Nam, Y. Kim 5

 l
u ( x, y, t ) = ∑ X i ( x, y )q i (t )
i =1
l +m
v ( x, y , t ) = ∑ Y ( x, y ) q
j =l +1
j j (t ) (9)

l +m+n
w( x, y, t ) = ∑Z
k = l + m +1
k ( x, y )q k (t )

where X(x,y), Y(x,y), Z(x,y) are shape functions which must be properly chosen so as to
satisfy the geometric boundary conditions of the wing structures. Using these
displacement expressions, equations for strain energy and kinetic energy are written in
matrix form,

1
U = {q}T [ K s ]{q} − QΛ {q}
2 (10)
1
T = {q&}T [ M s ]{q&}
2

where QΛ is the control force vector due to the actuation strain.

For the design of suppression system of a symmetric laminated plate model with
piezoelectric actuators, the model is assumed to have the surface bonded piezoelectric
actuators on the opposite side of the plate at the same location. It is also assumed that the
same magnitude but the opposite direction of electric field is applied to the actuator in
order to create a pure bending moment for flutter suppression. With these assumptions,
the displacements in the x and y directions can be neglected. In order to express the
displacement in the z direction, we used the free-free beam vibration modes in the x
direction(chordwise direction) and the cantilever beam vibration modes in the y
direction(spanwise direction) as follows;

Z k ( x, y ) = Φ i ( x ) Ψ j ( y ) (11)

where

© 2001 by C. Nam, Y. Kim 6

Φ i ( x) = 1,i = 1
x
Φ i ( x) = 3 (1 − 2 ),i = 2 (12)
c
x x x x
Φ i ( x) = (cosα i + cosh α i ) − β i (sin α i + sinh α i ),i = 3,4,5..
c c c c

y y y y
Ψ j (y) = ( − cosα j + cosh α j ) − β i (sin α j + sinh α j ), i = 1,2,3.. (13)
l l l l

In Eq.(10), if we express {QΛ} in terms of the applied voltage {u}, it becomes as

follows;

T
np
  d 31 
1 ts t   
{QΛ } = 2∑ {( + t p ) 2 − ( s ) 2 }∫∫ [Q p ]d 32  {Z ' ' }dxdyV (i )
i =1 t p 2 2 Ap   0  (14)
  
= [ F p ]{u}

where

∂ 2 Z ∂ 2 Z ∂ 2 Z 
{Z ' '}T = −  2 , 2 ,2 
 ∂x ∂y ∂x∂y  (15)
{u}T = V (1),V (2),...,V (np) 

ts, tp are the thicknesses of the laminate and piezoelectric layer, respectively. [Fp] is the
control force per unit voltage, np is the number of surface bonded piezoelectric actuators,
V(i) is the applied voltage to the i-th piezoelectric actuator.

The virtual work due to aerodynamic forces can be rewritten as follow;

δW A = q δw c / 4 [ S ]{∆C p } = q δw c / 4 [ S ][ AIC ] −1{w}3c / 4 (16)

© 2001 by C. Nam, Y. Kim 7

q is the dynamic pressure, a diagonal matrix [S] denotes the panel area, and ∆Cp is the
pressure coefficient. And δwc / 4 , {w}3c / 4 denote the virtual displacement at 1/4 chord,
the downwash at 3/4 chord of the panel, respectively. With harmonic motion assumption,
the downwash is expressed as the following;

ω
{w}3c / 4 = [α ij ]3c / 4 {q} − i [Wij ]3c / 4 {q} (17)
V

where

 ∂Z j 
[α ]
ij 3c / 4 = − 
 ∂x  i (3c / 4) (18)
[W ]ij 3c / 4 [ ]
= Zj i ( 3c / 4 )

Zj is the displacement function in z-direction and {q} is the generalized coordinate.

Similarly, the virtual displacement is written as

{w}c / 4 = [ Z j ]i ( c / 4 ) {q}

Then, the virtual work becomes

 ω 
δW A = q{δq}Z j i ( c / 4) [ S ][ AIC ] −1 [α ij ]3c / 4 − i [Wij ]3c / 4 {q}
T

 V  (19)
= q{δq}[ A]{q}

It has to be noted that the aerodynamic forces, [A] is function of the reduced frequency
and Mach number.

Lagrange's equations result in a set of ordinary differential equations of motion as

follows;

[ M s ]{q&&} + [ K s ]{q} = [ F p ]{u} + q [ A]{q} (20)

© 2001 by C. Nam, Y. Kim 8

where q is the dynamic pressure, [A] is the unsteady aerodynamic force matrix. The
aerodynamic force matrix can be calculated using modified strip theory or doublet-lattice
method as discussed in the previous chapter. [Ms] and [Ks] are respectively the
generalized mass and stiffness matrices including the effect of the piezoelectric actuators
placement, [Fp] is the control force matrix due to unit applied electric field.

After the vibration analysis, a model reduction can be performed to obtain a set of
equations of motion in the modal coordinates using first rm vibration modes.

[ M s ]{ξ&&} + [ K s ]{ξ } = [ F p ]{u} + q [ A ]{ξ } (21)

where

[ M s ] = [Φ ]T [ M s ][Φ ]
[ K s ] = [Φ ]T [ K s ][Φ ]
(22)
[ A ] = [Φ ]T [ A][Φ ]
[ F p ] = [ Φ ]T [ F p ]

[Φ ] is the eigen matrix.

5.2 Frequency Domain Unsteady Aerodynamics and Rational Function

Approximation

The classical aeroelastic analysis such as the V-g method and p-k method is performed
in the frequency domain using unsteady aerodynamic forces computed for simple
harmonic motion. As discussed in the previous section, in frequency domain, unsteady
aerodynamic forces can be obtained from the following equation by DLM.

 ω 
δW A = q{δq}Z j i ( c / 4 ) [ S ][ AIC ] −1 [α ij ]3c / 4 − i [Wij ]3c / 4 {q}
T

 V  (23)
= q{δq}[ A]{q}

© 2001 by C. Nam, Y. Kim 9

However, for the aeroservoelastic analysis and design, it is necessary to transform the
equations of motion into the state space form. In this chapter, the Minimum State method
will be introduced for the rational function approximation. It is known that in minimum
state method, the increase in the size of the augmented aerodynamic state is smaller than
any other methods.

Minimum State Method

Minimum State method approximates the aerodynamic force matrix by

[ Aap (s )] = [A0 ] + [A1 ]s + [A2 ]s 2 + [D' ][s I − R' ]−1 [E]s (24)

where s is the nondimensionalized Laplace variable(s = sb / V ). V is the airspeed, b is

the semi chord, and s is the Laplace variable. The components of diagonal matrix [R'] are
negative constants which are selected arbitrarily. For a given [R'] matrix, [ A0 ], [ A1 ],
[ A2 ] , [D'], and [E] are determined by using repeated least-square fit.

First, let’s define the aerodynamic states, {xa}

[ ]
s {xa } = [E ]s {ξ } + R ′ {x a } (25)

Therefore,

{xa } = [s[ I ] − [R ′ ]]−1 [E ]s {ξ } (26)

Then, aeroservoelastic equation will be

[M s ]{ξ&&} + [C s ]{ξ&} + [K s ]{ξ } = q [[A0 ] + [A1 ]s + [A2 ]s 2 ]{ξ }

+ q [D ′ ][s [ I ] − [R ′ ]] [E ]s {ξ } + [ F p ]{u}
−1

(27)

[M s ]{ξ&&} + [C s ]{ξ&} + [K s ]{ξ } = q [[A0 ] + [A1 ]s + [A2 ]s 2 ]{ξ } + q [D ′ ]{x a } + [ Fp ]{u}

© 2001 by C. Nam, Y. Kim 10

After rearranging the equation, the system equations can be expressed

  b   &&
2
  b 
[ ]
 M s − q A2   {ξ } = − C s − q A1  {ξ&} − [K s − q A0 ]{ξ } + q D ′ {x a } + [ F p ]{u} (28)
  V     V 

[M~ ]{ } [C~ ]{ } [K~ ]{ }

ξ&& = − ξ& − [ ]
ξ + q D ′ {x a } (29)

where

[M~ ] = M 
2
b
s − q A2   
 V  

[]
~   b 
C = C s − q A1  
  V 
(30)

[]
~
K = [K s − q A0 ]

[ ] [ ]
Since s {xa } = E s {ξ } + R ′ {x a } , the aerodynamic state is expressed as

{x& a } = [E ]{ξ&}+ [R′ ] Vb {x } a (31)

 

Using the state vector {x}, the linear time-invariant state space equations of motion
which include the effects of piezoelectric control forces are expressed as follows

{x&} = [ F ]{x} + [G ]{u} (32)

where

© 2001 by C. Nam, Y. Kim 11

  
 0 [I ] 0 
[ ] [ ] [ ] [ ] [M~ ] [D′ ]
 ~ −1 ~
[ F ] = − M K − M C
~ −1 ~
qd
−1 

 
 0 [E ] [R′ ] Vb  
   
 0 
[ ][
~ −1
[G ] =  M F p 
 ]
 0 

The state vector, {x} is defined as {x} = ξ ξ& x a  . If [R'] is set to be a (m × m)

T

matrix, the total number of states is (2*rm × m). It should be noted that the system
matrix [F] if function of airspeed.

5.3 Optimal Output Feedback Design

Since the system matrix is a function of the dynamic pressure, it is necessary to choose
the design air speed for designing control system. By choosing the design air
speed(VDesign), the linear quadratic regulator(LQR) with output feedback is applied for
design of the flutter suppression system. The design speed, which is chosen arbitrarily is
usually larger than the open loop flutter speed.

The design procedure for LQR with output feedback is briefly summarized below.

Consider a linear time-invariant system

{x&} = [ F ]{x} + [G ]{u}

(33)
{ y} = [ H ]{x}

The output matrix [H] is specified such that the aerodynamic states are excluded from the
state for feedback purpose. The admissible controls are output feedback of the form

© 2001 by C. Nam, Y. Kim 12

{u} = −[ K G ]{ y} (34)

where [KG] is an output feedback gain matrix, which consists of constant coefficients to
be determined by the design procedure. The performance index to be minimized is
selected as follows:

∞
J = ∫ [{ y}T [Q]{ y} + {u}T [ R]{u}]dt (35)
0

where [Q]≥0 and [R]>0 are symmetric weighting matrices. Since the initial state {x0}
may not be known, we adopt Athans and Levine's procedure that minimizes the expected
value of performance index

J = E{J } = trace{[ P][ X 0 ]} (36)

where the operator E{•} denotes the expected value over all possible initial conditions.
The symmetric matrices [P] and [X0] are defined by

[ X 0 ] = {x 0 }{x 0 }T (37)

∞
{ }
[ P] = ∫ e [ Fc ] t [ H ]T [Q] + [ K G ]T [ R ][ K G ] [ H ]e [ Fc ]t dt
0
T
(38)

where

[ Fc ] = [ F ] − [G ][ K G ][ H ] (39)

The usual assumption is that the initial states are uniformly distributed on the unit sphere;
then [X0] becomes identity matrix. It is well known that matrix [P] can be obtained by
solving the following Lyapunov equation

[Σ] = [ Fc ]T [ P] + [ P][ Fc ] + [ H ]T [Q][ H ] + [ H ]T [ K G ]T [ R][ K G ][ H ] = [0] (40)

Note that Lyapunov equation is a symmetric linear matrix equation. The solution of the
Lyapunov equation can be obtained directly by solving simultaneous linear equations,
however this approach takes lots of computation time and storage. One of the most

© 2001 by C. Nam, Y. Kim 13

efficient and robust algorithms to solve Lyapunov equations uses QR factorization. In
this algorithm, the Lyapunov equation is transformed into a simpler equation by use of
Schur transformations. Now, our problem is to determine the optimal output feedback
gain matrix [KG] at the design speed, by minimizing the performance index subject to the
algebraic matrix constraint Eq.(40).

To solve this problem with a constraint equation, a Lagrange multiplier approach is

widely used. The Hamiltonian is defined by adjoining the constraint Eq.(40) to the
performance index and expressed as follows.

ℜ = trace{[ P][ X 0 ]} + trace{[Σ][ S ]} (41)

[S] is a symmetric matrix of Lagrange multipliers which still needs to be determined.

The following conditions are necessary for [P], [S], and [KG] to be the extrema:

∂ℜ ∂ℜ ∂ℜ
= [0], = [0], = [0] (42)
∂ [ P] ∂ [S ] ∂ [KG ]

Applying Eq.(42) to Eq.(41), necessary conditions for optimality are obtained as follows:

[ Fc ]T [ P ] + [ P][ Fc ] + [ H ]T [Q][ H ] + [ H ]T [ K G ]T [ R][ K G ][ H ] = [0]

[ Fc ][ S ] + [ S ][ Fc ]T + [ X 0 ] = [0] (43)
[ R][ K G ][ H ][ S ][ H ]T − [G ]T [ P][ S ][ H ]T = [0]

There exist several numerical approaches to solve Eqs. (43). Gradient-based routines can
be used with various optimization algorithms. Moerder and Calise proposed an efficient
iterative algorithm, which converges to a local minimum if the following conditions hold:

(i) there exists an output gain matrix [KG] such that [Fc] is stable,

(ii) the output matrix [H] has full row rank,

(iii) control weighting matrix [R] is positive definite, and

© 2001 by C. Nam, Y. Kim 14

 (iv) output state weighting matrix [Q] is positive semidefinite and should be chosen
so that all unstable states are weighted in the performance index.

The iterative algorithm for finding solution is outlined as follows:

Step 1: Choose [KG] such that [Fc] in Eq.(39) has all the eigenvalues with negative real
parts.

Step 2: With given [KG] and the corresponding [Fc], solve the Lyapunov Eqs.(43) for [P]
and [S], and compute current performance index value.

Step 3: Evaluate the gain update direction

∆[ K G ] = [ R] −1 [G ]T [ P][ S ][ H ]T ([ H ][ S ][ H ]T ) −1 − [ K G ] (44)

Step 4: Update the gain by

[KG ]new = [KG ] + α∆[KG ] (45)

where α is chosen so that the closed-loop system [Fc] is asymptotically stable,

and the performance index of the updated system is less than that of the current
system. α is usually set to be 0.3.

Step 5: Iterate Steps 2 through 4 until the converged solution is obtained.

Any minimization algorithm requires the selection of an initial stabilizing gain.

However, such a gain may or may not be easy to find. One possible way to find the
initial gain matrix is to use root locus techniques by closing one loop at a time in the
control system. Here, first we can solve the full state LQR problem with performance
index Eq.(35), and construct initial stabilizing output gain matrix by choosing the

© 2001 by C. Nam, Y. Kim 15

corresponding gain parameters to be used in the output feedback from the resulting LQR
gain matrix elements.

5.4 Aeroservoelastic Design Example

The plate model containing four sets of piezoelectric actuators is used as a model for the
control system design. It is assumed that a set of actuators is bonded on both top and
bottom surfaces of the laminated plate in order to generate a pure bending force. The size
of each piezo actuator is 1.5 × 3.0 inch. The material properties used in this study are
given in Table 1.

Table 1. Material Properties

Composite Materials Piezoelectric Materials

E1=14.21×106 psi Ep=9.137×106 psi

E2=1.146×106 psi ρp=0.28 lb/in3

G12=0.8122×106 psi νp=0.3

ρ=0.05491 lb/in3 d31=6.5×10-9 in/V

ν12=0.28 d32=6.5×10-9 in/V

ν12=0.28

The laminate has six symmetric layers, [105/ ±45]s. Each layer has uniform thickness,
which is 0.02 inch, respectively. The coordinates of the piezo actuators are pzx(i)=
{0.50, 0.50, 3.50, 3.50} inch and pzy(i)={1.00, 7.00, 1.00, 7.00} inch.

© 2001 by C. Nam, Y. Kim 16

After vibration analysis, first six modes are used for the modal reduction. The first six
natural frequencies are 34.9, 122.3, 220.8, 415.3, 523.5, 624.1 Hz, respectively. Figure
?? shows the first mode shapes used for the modal reduction.

1st mode 2nd mode

3rd mode 4th mode

© 2001 by C. Nam, Y. Kim 17

 5th mode 6th mode

Figure Vibration mode shapes of the laminated Composite Plate-Wing with

Piezoelectric Actuators.

Unsteady aerodynamics are calculated using DLM for Mach 0.8. For calculation of the
pressure distribution on an oscillating plate-wing undergoing simple harmonic motion,
the plate-wing model is uniformly divided into total 100 panels arranged in 10 spanwise
direction and 10 chordwise direction.

The open loop flutter analysis of [105/±45]s laminate is conducted with the unsteady
aerodynamics transformed into Laplace domain through Minimum State method. The
result is shown in Figure. For the rational function approximation, a total of 6
components of the [R'] matrix are used for the RFA. Therefore, the dimension of the state
vector is 18.

The diagonal components of [R'] matrix are {-0.0072, -0.048, -0.197, -0.49, -0.67, -
0.98}, respectively. As in Figure, in [105/±45]s model without a control system, flutter
occurs at 620 fps by the 2nd mode, which is the 1st torsion mode dominant and the flute
frequency is about 59 Hz.

© 2001 by C. Nam, Y. Kim 18

 20 0 70 0

10 0
60 0
0

-10 0 50 0

-20 0
40 0
-30 0
30 0
-40 0

-50 0
20 0
-60 0
10 0
-70 0

-80 0 0
0 100 200 300 400 500 600 7 00 0 100 200 300 400 500 600 7 00

60 0

50 0
Open Loop Analysis Results
40 0

[105/ ±45]s laminated composite with 4

30 0

PZT actuators.
20 0

10 0 Flutter speed = 620 ft/sec

0
-200 -150 -100 -50 0 50 Flutter frequency = 59 Hz

Figure Open Loop Flutter Analysis Result of [105/ ±45]s Composite Plate-Wing
with Piezoelectric Actuators.

For the active control design, the design speed is set to be 800 fps, which is chosen
arbitrarily. At the design speed, the open loop system is unstable in the first torsion
mode, yielding unstable eigenvalues at (162.9 ± I 294.2). The open loop matrix equation
is supplemented with control term. The control input {u} is determined by LQR with
output feedback subject to minimize the performance index at the design speed. In order
to select the initial gain matrix, first we solve the full state LQR problem with
performance index, and construct initial stabilizing output gain matrix by choosing the
corresponding gain parameters to be used in the output feedback from the resulting LQR

© 2001 by C. Nam, Y. Kim 19

gain matrix elements. For the full state LQR problem, the weighting matrices [Q] and [R]
in Eq.(31) are set to be 500[I] and [I], respectively.

The loci of the closed loop aeroelastic roots with increasing airspeed is plotted in Figure.
Above the design speed, the control law provides flutter mode control. The controller
push the unstable eigenalvues to (-124.8 ± I 315.5). And the model is stable until a
divergence occurs at about 900 fps.

20 0 70 0

10 0
60 0
0

-10 0 50 0

-20 0
40 0
-30 0
30 0
-40 0

-50 0 20 0

-60 0
10 0
-70 0

-80 0 0
0 100 200 300 400 5 00 60 0 700 800 900 1000 0 100 200 300 400 5 00 60 0 700 800 900 1000

80 0 Closed Loop Analysis Results

70 0

60 0
[105/ ±45]s laminated composite with 4
50 0 PZT actuators.
40 0

30 0
Divergence speed = 900 ft/sec
20 0

10 0

0
-600 -500 -400 -300 -200 -100 0 1 00

Figure Closed Loop Flutter Analysis Result of [105/ ±45]s Composite Plate-Wing
with Piezoelectric Actuators.

© 2001 by C. Nam, Y. Kim 20

5.5 Further Parametric Study

In this chapter, a control system is designed is to suppress the flutter of a composite wing
using segmented piezoelectric actuators. The control system design, especially using
optimal output feedback design, is performed for the active flutter suppression. An
example model shows that the PZT actuators are capable of aeroelastic control. For the
aeroelastic control, the piezoelectric actuator must create substantial changes in surface
shape to generate the aerodynamic loads. The piezoelectric actuator location, size, and
thickness should be considered in the control system design since these parameters may
affect the structural properties as well as the control characteristics. The optimization
technique can be applied to find the best geometry of actuators for flutter control subject
to minimize the control performance index.

The closed loop damping ratios can be imposed in order that the control system has more
stability margin, therefore higher critical aeroelastic airspeed. However, heavy
constraints on the damping ratio may result in higher gain matrix and much control
efforts required for flutter suppression. As the optimization progresses, a trade-off
between the reduction of performance index and satisfaction of constraints on the closed
damping ratios is needed in order to determine the best geometry of the piezo actuators.
Also, the electric power requirement should be addressed.

© 2001 by C. Nam, Y. Kim 21

References
1. Crawley, E.F., deLuis, J., ”Use of Piezo-Ceramics as Distributed Actuators in Large
Space Structures,” AIAA Paper No. 85-0626, Proceedings of the 26th Structures,
Structural Dynamics and Material Conference, NY, April 1985.
2. Crawley, E.F., deLuis, J., ”Use of Piezoelectric Actuators as Elements of Intelligent
Structures,” AIAA Journal, Vol.25, No.10, Oct. 1987, pp.1373-1385.

3. Crawley, E.F., Anderson, E.H., ”Detailed Models of Piezoceramic Actuation of

Beams,” AIAA Paper No. 89-1388, Proceedings of the 30th Structures, Structural
Dynamics and Material Conference, AL, April 1989.

4. Ehlers, S.M., Weisshaar, T.A., ”Static Aeroelastic Behavior of an Adaptive Laminated

Piezoelectric Composite Wing,” AIAA Paper No. 90-1078, Proceedings of the 31th
Structures, Structural Dynamics and Material Conference, CA, April 1990.

5. Ehlers, S.M., Weisshaar, T.A., ”Effects of Adaptive Material Properties on Static

Aeroelastic Control,” AIAA Paper No. 92-2526, Proceedings of the 33th Structures,
Structural Dynamics and Material Conference, TX, April 1992.

6. Lazarus, K.B., Crawley, E.F., Lin, C.Y., ”Fundamental Mechanisms of Aeroelastic

Control with Control Surface and Strain Actuation,” AIAA Paper No. 91-0985,
Proceedings of the 32th Structures, Structural Dynamics and Material Conference, MD,
April 1991.

7. Heeg, J., ” An Analytical and Experimental Investigation of Flutter Suppression Via

Piezoelectric Actuators,” AIAA Paper No. 92-2106, Proceedings of the AIAA Dynamics
Specialist Conference, April 1992.

7. Jones, R.M. , Mechanics of Composite Materials, Scripta Book Co., Washington,

D.C.,1975.

8. Whitney, J.M., Structural Analysis of Laminated Anisotropic Plates, Technomic

Publishing Co. INC., Lancaster, Pennsylvania, 1987.

© 2001 by C. Nam, Y. Kim 22

9. Albano, E., Rodden, W.P., ”A Doublet-Lattice Method for Calculating Lift
Distributions on Oscillating Surfaces in Subsonic Flows,” AIAA Journal, Vol. 7, No. 2,
Feb. 1969, pp. 279-285.

10. Karpel, M., ”Design for Active Flutter Suppression and Gust Alleviation Using State-
Space Aeroelastic Modeling,” Journal of Aircraft, Vol.19, No.3, March 1982, pp.221-
227.

11. Hoadley, S.T., Karpel, M., ”Application of Aeroservoelastic Modeling Using

Minimum-state Unsteady Aerodynamic Approximations,” Journal of Guidance, Control,
and Dy-namics, Vol.14, No.6, November-December 1991, pp. 1267-1276.

12. Levine, W.S., and Athans, M., ”On the Determination of the Optimal Constant
Output Feedback Gains for Linear Multivariable Systems,” IEEE Transactions on
Automatic Control, Vol. AC-15, No. 1, February 1970, pp. 44-48.

13. Moerder, D.D., and Calise, A.J., ”Convergence of a Numerical Algorithm for
Calculating Optimal Output Feedback Gains,” IEEE Transactions on Automatic Control,
Vol. AC-30, No. 9, September 1985, pp. 900-903.

14. Lewis, F.L., Applied Optimal Control and Estimation, Prentice-Hall, New Jersey,
1992.

15. Mueller, G.S., and Adeniyi, V.O., ”Optimal Output Feedback by Gradient Methods
with Optimal Stepsize Adjustment,” Proceedings of IEE, Control & Science, Vol. 126,
No. 10, October 1979.

16. Kailath, T., Linear Systems, Prentice-Hall, New Jersey, 1980.

17. Chen, C.F., and Shieh, L.S., ”A Note on Expanding PA + A T P = Q,” IEEE
Transactions on Automatic Control, Vol. AC-13, No. 2, February 1968, pp. 122-123.
18. Bartels, R.H., and Stewart, G.W., ”Algorithm 432: Solution of Matrix Equation AX +
XB = C,” Comm. of the ACM, Vol. 15, No. 9, September 1972, pp. 820-826.

19. Kirk, D.E., Optimal Control Theory, Prentice-Hall, New Jersey, 1970.

© 2001 by C. Nam, Y. Kim 23