01 0302031 L Norgren PDF
01 0302031 L Norgren PDF
01 0302031 L Norgren PDF
J. Lundstedt
Ericsson Radio Systems AB
SE-164 80 Stockholm, Sweden
M. Norgren
Division of Electromagnetic Theory
Alfven Laboratory
Kungliga Tekniska Hogskolan
SE-100 44 Stockholm, Sweden
1 Introduction
2 Problem Formulation
3 Dispersion Model and Formulation of the Inverse
Problem
3.1 The Dispersion Model in the Frequency Domain
3.2 The Dispersion Model in the Time Domain
3.3 The Values of the Electrical Parameters
4 The Direct Problem in the Frequency Domain
4.1 Wave Splitting and the Riccati Equation
4.2 Stray Capacitances at the Connections
5 The Direct Problem in the Time Domain
5.1 Wave Splitting
5.2 Greens Functions
5.3 The Direct Problem
6 The Inverse Problem in the Frequency Domain
7 The Inverse Problem in the Time Domain
7.1 De-Embedding of the Transient Response
7.2 Reconstruction Procedure
8 Numerical Results
8.1 Comparisons of the Solutions Obtained to the Direct
Problem Using Frequency and Time domain Direct
Solvers, Respectively
8.2 Reconstructions Using Noise Contaminated Data Gen-
erated in the Frequency Domain
8.3 Reconstructions Using Noise Contaminated Data Gen-
erated in the Time Domain
8.4 Reconstructions Using Frequency Domain Data Gener-
ated with Stray Capacitances at the Endpoints of the
Band Cable
9 Discussion and Conclusions
References
Frequency domain and time domain methods 3
1. INTRODUCTION
2. PROBLEM FORMULATION
NWA / TDR
V i V1i Vt
V r V1r
0 l x
- + -
Figure 2. Cross section of the at band cable in sand.
is the upper bound for the relative permittivity of dry sand. If possible,
sand should be determined from a measurement on dry sand, instead
of using (10).
At frequencies below 60 GHz, the electric susceptibility of water
can be described with the Debye model [6]
s
water () = + 1, (11)
1 + jd
in which s is the relative permittivity for static elds, d is the Debye
relaxation time, and is the optical response [5]; the contribution
from the fast processes in the medium to the permittivity at moderate
frequencies.
The total shunt capacitance between the inner conductor and the
two outer conductors (even mode) in an insulated band cable embedded
in moist sand (see Figure 2) is estimated with the following formula:
C2 r (x, ) C3
C (x, ) = C1 + . (12)
C2 + r (x, ) C3
C1
Using (11) and (9) in (12), it follows that the total capacitance is
described by the Debye model
Cs (x) C (x)
C (x, ) = C (x) + , (13)
1 + je (x)
8 Lundstedt and Norgren
where
C2 C3 (sand + q (x) ( 1))
C (x) = C1 + , (14)
C2 + C3 (sand + q (x) ( 1))
C2 C3 (sand + q (x) (s 1))
Cs (x) = C1 + , (15)
C2 + C3 (sand + q (x) (s 1))
C2 + C3 (sand + q (x) ( 1))
e (x) = d . (16)
C2 + C3 (sand + q (x) (s 1))
C is the optical response of the capacitance, Cs is the static
capacitance, and e is the eective relaxation time. Note that all
parameters in (13) depend on the local value of the moisture parameter
q (x).
If we allow the parameters in equations (1) and (2) to be complex
valued, there will be a seeming ambiguity in the separation into
dissipative and reactive parameters. For example, if all electrical losses
are attributed to the imaginary part of the complex valued capacitance
C (x), we have in the view of equation (2) that
In the present paper we will use (17), since it yields a shorter notation
in the FD analysis and is more appropriate in a problem involving
dielectric losses only; (18) is appropriate when the losses are dominated
by a static conductivity. The series resistance R is considered to be
negligible, i.e.,
R (x, ) = 0. (19)
obtain
Cs (x) C (x) t
C (x, t) = C (x) (t) + H (t) exp , (20)
e (x) e (x)
L (x, t) = Lb (t) , (21)
R (x, t) = G (x, t) = 0, (22)
where (t) is Diracs delta function and H (t) is Heavisides step
function. With G (x, t) = 0 it follows that all electric losses are
included in the dispersion kernel C (x, t) for the capacitance. In
conformance with the ambiguity in the frequency domain, one can
transform from C (x, t) to another capacitance kernel and a nonzero
conductance kernel; see also the discussion in Section 5.
for which the inverse transformation from split voltages to voltage and
current reads
V (x) 1 1 V + (x) V + (x)
S01 , (29)
I (x) Y0 Y0 V (x) V (x)
where Y0 = Z01 . Using (28) and (29) in (27), we obtain the ODE for
the split voltages:
V + (x) +
1 V (x) a (x) b (x) V + (x)
x S 0 D (x) S = , (30)
V (x) 0
V (x) b (x) a (x) V (x)
where
1
a (x) = j C (x) Z0 + L (x) Y0 + G (x) Z0 + R (x) Y0 , (31)
2
1
b (x) = j C (x) Z0 L (x) Y0 + G (x) Z0 R (x) Y0 . (32)
2
Frequency domain and time domain methods 11
where
L (x) = Lf (x, t) dt. (39)
0
From (41) and (42) we notice that R (x), the direct response of
the series resistance, and L (x, 0), the initial value of the inductance
dispersion kernel, are equivalent from a signal propagation point
Frequency domain and time domain methods 13
of view. The same holds for G (x) and C (x, 0), R (x, t) and
Lt (x, t), G (x, t) and Ct (x, t). Thus, we make the following variable
substitutions:
V (x, t) V (x, t)
= A (x) + B (x) . (47)
I (x, t) I (x, t)
The left hand side of equation (47) together with the matrix operator
A (x) determine the characteristics of (47), i.e., the wavefront velocity.
The matrix operator B (x) contains the dissipative and dispersive
terms. Note that B (x) is zero on a uniform, lossless transmission
line with nondispersive parameters. Thus, the B (x) term is zero
on the connected transmission lines at x < 0 and x > l. In
terms of signal propagation the transmission line is characterized by
four nondispersive parameters; L(x), C(x), r(x) and g(x), and two
dispersive parameters; r(x, t) and g(x, t). It is important to notice that
this means that the dispersion in the inductance and series resistance
cannot be distinguished in any measurements. The same holds for the
dispersion of the capacitance and shunt conductance.
For our specic problem, with the transmission line model given
by equations (20)(22), the parameters in (47) become
L (x) = Lb , (48)
(x) C2 C3
C (x) = C (x) = C1 + , (49)
C2 + C3 (x)
14 Lundstedt and Norgren
C2 + C3 s (x) 1
= C2 C3 s (x) (x) ,
C2 + C3 (x) d (C2 + C3 (x))
(51)
Cs (x) C (x) t
g (x, t) = 2 H (t) exp
e (x) e (x)
C2 C3 (C2 + C3 s (x))
= H (t)
(d (C2 + C3 (x)))2
V (x, t) 1 1 V + (x, t)
V + (x, t) 1
S (x) ,
I (x, t) Y (x) Y (x) V (x, t)
V (x, t)
(57)
where Z(x) and Y (x) are the time domain characteristic impedance
and admittance, respectively:
1 L (x)
Z (x) = = . (58)
Y (x) C (x)
The PDE for the split components is given by (47) and (56)(57) as
+
+
V+ 1 V+ (x) (x) V+
x + t =
V c (x) V (x) (x) V
(x, t) (x, t) V+
+ , (60)
(x, t) (x, t) V
where
1
c (x) = , (61)
L (x) C (x)
is the wavefront velocity and the nondispersive parameters are given
by
1
(x) = (Zx (x) Y (x) g (x) Z (x) r (x) Y (x)) , (62)
2
1
(x) = (Zx (x) Y (x) g (x) Z (x) + r (x) Y (x)) , (63)
2
1
(x) = (Zx (x) Y (x) + g (x) Z (x) r (x) Y (x)) , (64)
2
1
(x) = (Zx (x) Y (x) + g (x) Z (x) + r (x) Y (x)) , (65)
2
16 Lundstedt and Norgren
cd (x, t) V1 (t) ,
V + (x, t + (0, x)) = a+ (x) V1i (t) + G+ i
(68)
V (x, t + (0, x)) = a (x) V1 (t 2 x, l)+Gcd (x, t) V1 (t) , (69)
i i
where V1i (t) = V + (0+ , t) is the incident signal from the left on the
nonuniform transmission line, (x1 , x2 ) is the wavefront travel time
from x1 to x2 :
x2
dx
(x1 , x2 ) = . (70)
x1 c (x)
a+ (x) and a (x) describe the attenuation of the wavefront traveling
to the right and left, respectively. The subscript cd on the Greens
Frequency domain and time domain methods 17
G+ 2 0 (x) (x) G+
x cd
t = cd
G
cd c (x) Gcd (x) (x) G
cd
(x, t) (x, t) G+
+ cd
(x, t) (x, t) G
cd
+a+ (x)
(x, t)
+a (x) (x, t 2 (x, l)) , (71)
(x, t) (x, t 2 (x, l))
where the attenuation factors are given by
x
+
a (x) = exp x dx , (72)
0
x
+
a (x) = rl a (l) exp x dx , (73)
l
Likewise, equation (77) is derived from (68) and (69) and the fact
that V (l+ , t) = 0 (there is no incident wave from the right), using
continuity of the voltage and the current.
Equation (71) together with (75)(78) can accurately and
eciently be solved numerically with the method of characteristics.
The jump condition (76) has to be used in the numerical program
since it constitutes a discontinuity in G+ cd along the characteristic
line of G+ cd . There are also discontinuities in G+ cd across the
characteristic line (x, 2 (0, l)), and in Gcd along the characteristic line
(x, 2 (0, l) + 2 (x, l)). However, since these discontinuities occur along
the respective characteristic lines, it is not necessary to treat them
separately in the numerical program. But, to achieve better accuracy
one should include the analytical solution of these discontinuities. By
integrating equation (15) along these characteristic lines, using (72)
(78), we obtain
1 +
G+ cd (x, 2 (0, l))= a (x) c (0) a (0) (0) , (79)
2
a (x) 1
G
cd (x, 2 (0, l)+2 (x, l))= rl a+ (l) c(0)a (0)(0) . (80)
a (l) 2
where = (0, l) is the one-way travel time, V1r (t) = V (0+ , t) and
V1i (t) is the incident wave at x = 0+ , as shown in Figure 1. The
reection factor b , and the reection kernel Rcd (t), are given by
b = a (0) , (82)
Rcd (t) = G +
cd 0 , t . (83)
Frequency domain and time domain methods 19
This is readily derived from equations (68) and (69). The transient
response including the impedance mismatch at x = 0 is given by
V r (t) = rk+ V i (t k 2 ) + Rdd (t) V i (t) , (84)
k=0
V r (t) r0 1 r0 V i (t)
, (85)
V1i (t) 1 + r0 r0 V1r (t)
The relation from Rcd (t) to Rdd (t) is found from equations (81) and
(84) by substituting (V i (t), V r (t)) for (V1i (t), V1r (t)) ((87) in (81)), and
using the fact that V i (t) is arbitrary. For the reection data we obtain
r0+ = r0 , (88)
2 k1
rk+ = 1 (r0 ) b r0 b , k 1, (89)
Rdd (t) + r0 Rcd (t) Rdd (t) = 1 (r0 )2 Rcd (t) r0 b Rdd (t 2 )
r0 rk+ Rcd (t k 2 ) . (90)
k=1
where, from (17), (31) and (32), we have (L and R do not depend on
q)
a (x, ) = j Z0 C
(x, ) = b (x, ) , (93)
2
at x = l is
From (35), it follows that the boundary condition for r
(l, ) = 0.
r (94)
Frequency domain and time domain methods 21
we have from (99) and (98) that the perturbation in the objective
functional becomes
l
=max
J (q) = dxRe jZ0 u (x, ) (1+ r (x, ))2 C
(x, ) . (100)
0 =min
J
=max
C
(x) = Re jZ0 u (x, ) (1 + r (x, ))2 (x, ) . (103)
q =min q
This is because the hard reections may contain more energy than
the useful signals (continuously scattered) do. As each hard reection
introduces numerical errors in the algorithm, it is clear that one should
use data containing as few hard reections as possible, i.e. only use
one round trip of data. In the frequency domain this would mean using
1
frequency intervals of 2 , where is the one round trip propagation
time on the transmission line.
The rst step in the reconstruction procedure is to deconvolve
the reection data with the incident pulse to obtain the reection
impulse response. However, since the impedance mismatch at x = 0
only hides the useful signals, we want to de-embed this mismatch
from the reection data, to obtain the reection impulse response as
if the impedance was continuous at x = 0. In Subsection, 7.1 we
outline how to deconvolve and de-embed the reection data by means
of optimization. In Subsection 7.2, we then outline the procedure
to reconstruct the moisture parameter from the reection impulse
response. The input data to the inverse method is the de-embedded
and deconvolved data obtained from the procedure described in
Subsection 7.1.
Rcd (t) , b , r0 and are determined from (107) in three steps. First,
r0 is determined by matching the very initial time traces of V r and
V i , which correspond to the hard reection at x = 0. Then b and
are determined from the rst hard reection that arrives from x = l,
i.e., by matching the signals in a short time interval after t = 2 .
Finally, Rcd (t) is determined by deconvolving equation (107) with
24 Lundstedt and Norgren
where T is the time period for which Rcd (t) is to be determined, and
where
The optimization is easily done by using the exact expression for the
gradient of J r with respect to Rcd (t):
T
J r
JRr cd t = Rcd (t) V2i (t) VRm (t) V2i t t dt. (111)
Rcd t
8. NUMERICAL RESULTS
-0.2
0 5 10 15 20 25 30 35 40
Time (ns)
Reflected voltage (V)
0.4
(b) Relaxation time = 1 ns Relaxation
0.2
-0.2
0 5 10 15 20 25 30 35 40
Time (ns)
Reflected voltage (V)
0.4
(c) Relaxation time = 10 ns Relaxation
0.2
0 5 10 15 20 25 30 35 40
Time (ns)
Reflected voltage (V)
0.4
(d) Relaxation time = 1 s Relaxation
0.2
-0.2
-0.4
0 5 10 15 20 25 30 35 40
Time (ns)
0.4 0.4
0.2 0.2
0 0
0 0.5 1 0 0.5 1
Position x (m) Position x (m)
Relaxation time = 10 ns Relaxation time = 1 s
0.8 0.8
Moisture parameter q(x)
0.4 0.4
0.2 0.2
0 0
0 0.5 1 0 0.5 1
Position x (m) Position x (m)
0.4 0.4
0.2 0.2
0 0
0 0.5 1 0 0.5 1
Position x (m) Position x (m)
Relaxation time = 10 ns Relaxation time = 1 s
0.8 0.8
Moisture parameter q(x)
0.4 0.4
0.2 0.2
0 0
0 0.5 1 0 0.5 1
Position x (m) Position x (m)
0.4 0.4
0.2 0.2
0 0
0 0.5 1 0 0.5 1
Position x (m) Position x (m)
Relaxation time = 10 ns Relaxation time = 1 s
0.8 0.8
Moisture parameter q(x)
0.4 0.4
0.2 0.2
0 0
0 0.5 1 0 0.5 1
Position x (m) Position x (m)
reproduce the sharp slope in the prole around x = 0.75 m (cf. the
results in Figure 5 from the previous subsection). The reconstructions
obtained using the FD algorithm on TD generated data that has been
transformed with FFT exhibit slightly more oscillations superimposed
on the true prole, but are otherwise of the same quality as the ones
obtained when using the FD algorithm on FD generated data.
Frequency domain and time domain methods 33
0.4 0.4
0.2 0.2
0 0
0 0.5 1 0 0.5 1
Position x (m) Position x (m)
Relaxation time = 10 ns Relaxation time = 1 s
0.8 0.8
Moisture parameter q(x)
0.4 0.4
0.2 0.2
0 0
0 0.5 1 0 0.5 1
Position x (m) Position x (m)
0.4
0.4
0.2
0.2
0
0
-0.2
0 0.5 1 0 0.5 1
Position x (m) Position x (m)
Relaxation time = 10 ns Relaxation time = 1 s
Moisture parameter q(x)
0.4 0.4
0.2 0.2
0 0
0 0.5 1 0 0.5 1
Position x (m) Position x (m)
Relaxation time = 10 ns Relaxation time = 1 s
0.8 0.8
Moisture parameter q(x)
0.4 0.4
0.2 0.2
0 0
0 0.5 1 0 0.5 1
Position x (m) Position x (m)
REFERENCES