You are on page 1of 6

Accurate and Efficient Crosstalk Analysis by

Full-wave Computations and System Identification


Carl Holmberg Thomas Rylander Jan Carlsson
Department of Electrical Engineering Department of Electrical Engineering Department of Electrical Engineering
Chalmers University of Technology Chalmers University of Technology Chalmers University of Technology
and Volvo Car Corporation Göteborg, Sweden and Provinn AB
Göteborg, Sweden rylander@chalmers.se Göteborg, Sweden
holcarl@chalmers.se jan.carlsson@provinn.se

Tomas McKelvey
Department of Electrical Engineering
Chalmers University of Technology
Göteborg, Sweden
tomas.mckelvey@chalmers.se

Abstract—We present a system-identification procedure ap- numerical methods applied to the 3D problem, at the expense
plied to a crosstalk problem, which can be accurately analyzed of computational time and other resources.
by an in-house full-wave solver based on the method of moments
(MoM). Given a relatively few frequency-samples of the crosstalk
current, we demonstrate that it is feasible to estimate a reduced- Some of the most popular and important numerical methods
order model of the continuous frequency dependence, where the for Maxwell’s equations are the finite-difference time-domain
frequency band of interest corresponds to an electrical size of scheme [1], [2], the finite element method [3], [4], and the
the system that ranges roughly from 0.005 to 2.7 wavelengths. method of moments (MoM) [5], [6]. In the context of EMC,
The estimated low-order model of the frequency response takes it is often necessary to analyze many different cases and, thus,
the form of a Padé approximant and it yields an accurate repre-
sentation when compared to the reference frequency-response, the reduction of computational time is attractive to enable use
as it is computed directly by the MoM solver. In addition, of full-wave numerical solvers in a viable manner.
the low-order model can be evaluated at an extremely low
computational cost and it provides useful information such as System identification (SI) is a powerful tool which is used
resonance frequencies and their quality factors. For the system to find mathematical models of dynamic systems by applying
analysis of a real-world device, the low-order model can serve as
a computationally powerful and accurate representation of the statistical methods to observed time, or frequency, domain
electromagnetic response for a subsystem that features full-wave data. In particular, SI can be used to estimate the full frequency
phenomena that otherwise could be too challenging to describe response of a linear time-invariant (LTI) system from fre-
at a system level. quency samples that may be computed by a numerical solver.
Index Terms—EMC, Crosstalk, Method of Moments, System The estimated frequency response also yields a reduced-order
Identification
model, which can be useful as it provides information of, e.g.,
damped resonances in terms of their complex poles.
I. I NTRODUCTION
Crosstalk problems occur frequently in electromagnetic In this article, we estimate the complete and continuous
devices and, thus, techniques to analyze the crosstalk are im- frequency response of an electromagnetic device based on
portant for assessing the electromagnetic compatibility (EMC) a relatively few discrete frequency samples of the device’s
of such devices. electromagnetic response computed by a full-wave solver.
Transmission line theory is a relatively simple but still very Here, we focus on a crosstalk test-case, which can be modelled
useful expedient for the analysis of crosstalk that occur in by the transmission line theory for sufficiently low frequencies.
many practical situations. However, such an approach can However, a correct physical description of the system at higher
not be used for cases where full-wave phenomena (due to, frequencies requires a full-wave solver. System identification
e.g., radiation or complicated geometry) are important. Full- yields a reduced-order model that is fast to evaluate, and
wave effects can in many cases be accurately examined by allows for rapid computation of the continuous frequency
response. Our tests for the crosstalk problem demonstrate that
This research has been carried out in the project Virtual Electric Propulsion
Systems (VEPS) with reference number 2017-05505. A project financed by the reduced-order model can give results that are in very good
The Swedish Governmental Agency for Innovation Systems (Vinnova), Volvo agreement with the full-wave solver. The approach can be
Cars, Chalmers, Provinn and Modelon. used with any method that computes the response at discrete
978-1-7281-5579-1/20/$31.00 ©2020 IEEE frequency points.
II. M ETHOD (modified) Green’s function for free space. Further, Ei (r) is
An overview of the proposed estimation process is shown the incident electric field, n̂(r) is the surface unit normal,
in Figure 1. In the data generation step, accurate samples j is the imaginary number, ω is the angular frequency, µ0
of the electromagnetic device’s response are computed by a is the permeability of free space and k0 is the free space
full-wave solver. Next, system identification is used to fit a wave number. The surface currents J are represented by Rao-
parameterized transfer function to the response data. This is Wilton-Glisson basis functions [7] on triangular elements.
repeated for models of increasing order until a suitable model Equation (1) is multiplied by a weighting function (that is
order is determined. If no satisfactory model can be estimated chosen from the set of basis functions) and integrated over the
for the existing samples of the response, the set of response surface S, which gives a system of linear equations Zj = e.
samples can be extended by generation of additional data and Contributions to the impedance matrix Z from integrals over
the system identification be repeated. This procedure is iterated pairs of testing and source elements that are well separated are
until a satisfactory model is produced. computed by adaptive Gaussian quadrature, while the contri-
butions from adjacent or overlapping pairs are computed by
the radial-angular singularity cancellation scheme of Khayat
Start and Wilton [8]. The presence of an infinite PEC ground plane
is handled by modification [9] of the Green’s function. Also,
modifications of the impedance matrix for incorporation of
Generate data lumped loads are treated in the same fashion as demonstrated
for thin wires in [6], where the necessary steps to implement
discrete voltage sources as delta-gap feeds are also presented.
System identification
B. System identification
Estimate model The continuous-time state-space representation of a linear
Model-order
time-invariant dynamic system of order n is
selection
Model validation ẋ(t) = Ax(t) + Bu(t), (2)
y(t) = Cx(t) + Du(t), (3)
p×1 m×1
where y(t) ∈ R is the output vector, u(t) ∈ R
No
Satisfactory model? the input vector and x(t) ∈ Rn×1 the vector of states. The
matrices A ∈ Rn×n , B ∈ Rn×m , C ∈ Rp×n and D ∈ Rp×m
Yes are often called the state-space parameters, which describe the
dynamics of the system. For most physical systems, the input
Stop can not instantaneously influence the output, which implies
that the direct feedthrough term, D can be set to zero [10].
For completeness, the matrix D is present in the formulas
Fig. 1. A schematic of the proposed process to estimate a satisfactory below. However, all models in this article are estimated under
frequency response function. the assumption of no direct feedthrough, i.e. D is set to zero
throughout the rest of the article.
A. Method of moments The state-space parameters and the transfer function are
In order to demonstrate the characteristic features of the related by
suggested estimation procedure, we use an in-house MoM G(s) = C(sI − A)−1 B + D, (4)
code that solves the electric field integral equation (EFIE) for
metal surfaces in the presence of a flat infinite perfect electric which results from the application of the Laplace transform
conductor (PEC) ground plane. Conducting surfaces shaped as to Equations (2) and (3). The frequency response function is
strips can be equipped with lumped loads and discrete voltage defined as G(jω), which is the transfer function restricted
sources. to evaluation along the imaginary axis. The elements of the
For a metal surface S, the electric field integral equation is frequency response function G(jω) take the form of so-
given by called Padé approximants [11], where a Padé approximant
j  is a rational function R(jω) = P (jω)/Q(jω) that involve
n̂(r) × Ei (r) =

− polynomials P (jω) and Q(jω) in the variable jω.
ωµ0
ZZ   (1) Given some, often noise corrupted, samples of the frequency
1
= n̂(r) × 1 + 2 ∇∇· J(r0 )G(r, r0 ) dr0 , response function G(jωi ), the goal of the frequency-domain
S k0 state-space system identification is to estimate a set of state-
where r is the field point on the surface S, r0 is the source space parameters, {Â, B̂, Ĉ, D̂} such that the corresponding
point, J(r0 ) is the induced surface current and G(r, r0 ) is the estimated frequency response function Ĝn (jω) of model or-
der n gives an optimal fit to the samples G(jωi ) in some given most common and it is a standard choice for measuring the
measure. goodness of fit to the validation and test data sets.
The model order n is the number of internal states of the
system or, equivalently, the number of poles of the corre- III. R ESULTS
sponding transfer function. Typically, the model order n is The test case consists of two parallel flat strips, termi-
also an unknown parameter. To determine the proper model nated to a common infinite PEC ground plane by lumped
order, a set of models with different orders are estimated impedances. Conductor 1 is fed by a delta-gap voltage feed
and their performance compared to find an appropriate model between the one end and the ground plane, and terminated
with regards to the trade-off between model performance and with a resistive load of 1 kΩ at the other end. Conductor
complexity. 2 is connected to the ground plane by 1 kΩ resistive loads
For the model estimation, the samples of the frequency re- at both ends. The quantity of interest is the current through
sponse function are often divided into two sets. The estimation resistor Rf of conductor 2, which serves as a measure of
of models are performed using an estimation data set the far-end crosstalk (FEXT). The length of the strips are
160 mm and the strip width, separation and height over the
Ze = G(jω1e ), . . . , G(jωne e ) ,

(5)
ground plane are all 2 mm, as shown in Figure 2. For all
where ωie , i = 1, . . . , ne , denotes ne distinct but not nec-
essarily ordered or uniformly spaced frequency points. This
allows for frequency dependent resolution, where additional
data points can be added in regions of particular interest.
To prevent overfitting to (possible) noise in the estimation
data set, the performance of the estimated model is measured
on a validation data set
Zv = G(jω1v ), . . . , G(jωnv v ) .

(6)
Fig. 2. The geometry of the crosstalk test-case with two parallel metal strips
This set consist of the frequency response at frequencies ωiv , above an infinite ground plane. The metal strip to the left is terminated by one
i = 1, . . . , nv , which are different from the frequency points voltage source and one resistor and the metal strip to the right is terminated
by two resistors.
used in the estimation data set. The validation data set allows
for an unbiased evaluation of the model fit when deciding on
MoM computations in the frequency band of interest, the
the appropriate model order.
geometry is discretized by a fixed structured triangular mesh
To confirm the performance of the model chosen in the such that λ/h ≥ 10, where h is the maximum edge length. In
estimation steps, we also define a test data set Figure 3, the current through Rf is shown for the frequency
Zt = G(jω1t ), . . . , G(jωnt t ) , range from 10 MHz to 5 GHz: solid curve – the in-house

(7)
MoM code outlined in Section II-A; and dash-dotted curve –
where ωit , i = 1, . . . , nt , are different from the frequencies the transmission line solver EMEC [15]. For low frequencies,
used in the estimation and validation data sets. The test data we see a good agreement between the two solvers. However,
set allows for measuring the actual performance of the model as the frequency increases, the responses begin to differ due
for unseen data. A better fit to the training data than to the to the increased influence of factors such as radiation losses
test data set shows a lack of generalizability of the model, and and end effects, which are not captured by the transmission
is usually a sign of overfitting. line solver. Even for the quite simple geometry of the test
This paper employ a two-step estimation procedure [12]. case, it is evident from the data in Figure 3 that solvers based
First, a non-iterative subspace method N4SID [13] is used on normal transmission line theory are not viable options
to compute an initial estimate Ĝn (s), where n denotes the for high-frequency cases. To maintain the accuracy at higher
model order. This initial estimate is then refined by successive frequencies, we must use full-wave methods such as the
iterations of the prediction error method (PEM), [14], which MoM at the expense of an increased computational cost. This
aim to minimize the cost function computational cost is however greatly reduced by the system-
Xne identification procedure presented in this article, which allows
V = ||Ĝn (jωie ) − G(jωie )||2F (8) for low-cost generation of an accurate frequency response over
i=1 the desired range of frequencies.
using numerical optimization. If all state-space parameters are
simultaneously optimized, this is a non-linear least-squares A. Construction of estimation, validation and test data sets
problem. Problems of this type can be solved using standard To emulate a situation where no information of the system
methods such as steepest descent or the Levenberg-Marquardt frequency response is known a priori, we construct the esti-
algorithm among many others [14]. mation data set Ze containing 35 uniformly spaced frequency
The model quality can be measured in a large number of points in the frequency interval from 10 MHz to 5 GHz. To
different metrics. The mean squared error (MSE) is one of the avoid the construction of an unnecessarily large estimation
-40
B. Model order selection
MoM
EMEC During the estimation, a set of models with different model
Magnitude [dBA]

-60
order are produced. Increasing the model order leads to a more
-80 flexible model, which can be more easily fitted to a given data
set. However, there is little gain in adding states which have
-100
comparably little or no effect on the model response. Also,
-120 there is a bias-variance trade-off to take into account. When the
0 number of parameters increases, the model bias decreases due
to the added model flexibility. However, the model variance
instead increases, since the extra flexibility can be used for
Phase [Deg]

-1000

fitting to sample-specific noise.


-2000 In Figure 5, we show the MSE for models with a model
order that ranges from 15 to 26, where the estimation is
-3000 based on the estimation data set Ze . There is a significant
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
Frequency [GHz]
reduction in the MSE for model order n = 23 as compared to
models with n < 23, which indicates that we have found the
Fig. 3. The frequency response of the current through resistor Rf : solid minimum number of states needed to represent the majority
curve – the in-house MoM solver; and dash-dotted curve – the transmission of the dynamics of the data. We can also note that the MSE
line solver EMEC [15].
calculated for the estimation data is consistently smaller than
that calculated on the validation data set.
data set, one could use an iterative approach where additional
data points are added to the estimation data set as the model 10-7
order is increased. Such an approach also makes it possible Estimation data
Validation data
to use information from previously estimated models when
10-8
deciding which frequency points to add to the estimation set.
Since the response computed by the MoM solver is essentially
without noise, the validation data set can be chosen small. 10-9
MSE [A 2 ]

Here, we choose to construct the validation data set using 34


data points centered between the frequencies chosen in the
10-10
estimation data set. The estimation and validation data sets
are shown in Figure 4.
10-11

-50 Estimation data


Validation data
-60 10-12
Magnitude [dBA]

15 16 17 18 19 20 21 22 23 24 25 26
-70 n [-]

-80
Fig. 5. The mean squared error for models as a function of the model order n:
-90 solid circles – estimation data set; and hollow squares – validation data set.
-100
0 In Figure 5, we note that both the estimation and validation
data sets lack the sample-specific noise that is normally
present, should the data be generated by physical measure-
Phase [Deg]

-1000

ments instead of numerical computations. The lack of sample-


-2000 specific noise makes it possible to avoid the usage of a
validation data set and, here, we include it only to demonstrate
-3000 that the MSE with respect to the validation data set does not
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
Frequency [GHz]
increase as the model order n increase beyond 23.

Fig. 4. The far-end crosstalk current through resistor Rf for the frequencies
C. Estimated model and its error relative the test data set
in the estimation and testing data sets. In Figure 6, we show the frequency response Ĝ23 (jω) of
the estimated model of order 23 together with the estimation
The test data set Zt contain 512 uniformly distributed and validation data sets. The estimated frequency response is
frequency response data points, which are not contained in Ze in good agreement with both of these data sets. A comparison
or Zv and give a good resolution of the frequency response with the complete response computed by the MoM, shown in
shown in Figure 3. Figure 3, demonstrates that the model is also able to capture
rapid variations in the response, corresponding to system chosen too high. Should there be pairs of poles and zeros that
resonances, even when no points in the estimation data set coincide, such pairs result in pole-zero cancellations, which is
are close to the resonances. a sign that the model order may be reduced. Thus, the pole-
zero plot is an important tool also in the model-order selection
-40 Estimated model stage.
Estimation data
Validation data
Magnitude [dBA]

-60

-80
6
Poles
Zeros
-100
4

Imaginary Axis [GHz]


-120
0 2
Phase [Deg]

-1000 0

-2000 -2

-3000 -4
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
Frequency [GHz]
-6
-6 -4 -2 0 2 4 6 8 10
Fig. 6. The frequency response, Ĝ23 (jω), of the estimated model of order
23 compared to the estimation and validation data. Real Axis [ns-1 ]
Fig. 8. The poles and zeros of the estimated transfer function Ĝ23 (s) in the
In Figure 7 the absolute error of Ĝ23 (jω) is shown together complex plane.
with the test data set. Markers are added to indicate the
frequency points of the estimation and validation data sets. In Let the r-th complex pole be given by βr = −γr + jωr
the figure, we can see that the largest errors can be found at where γr is the damping and ωr = 2πfr is the angular
frequencies in-between points of the estimation data set, which frequency associated with the pole. Then, we can define the
indicates that the chosen sampling strategy for the validation quality factor of the r-th pole as
p
data set is indeed appropriate. γr2 + ωr2
Qr = . (9)
2γr
Test data
-50 Absolute error Table I show the quality factors for the complex-conjugate
Estimation data
Validation data pole pairs of Ĝ23 (s).
-60

-70 TABLE I
T HE FREQUENCY, DAMPING AND QUALITY FACTORS FOR THE
-80 COMPLEX - CONJUGATE POLE PAIRS OF THE ESTIMATED TRANSFER
Magnitude [dBA]

FUNCTION BASED ON THE M O M DATA .


-90

-100 r fr [GHz] γr [ns−1 ] Qr [-]


-110 1 0.45 0.25 5.7
2 0.90 0.49 5.7
-120
3 1.4 0.25 17
-130 4 1.8 0.45 13
-140
5 2.3 0.25 28
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
6 2.7 0.47 18
Frequency [GHz]
7 3.2 0.25 41
Fig. 7. The magnitude of the error for the model of order 23 compared to the 8 3.6 0.44 25
magnitude of the frequency response in the test data set. Different markers 9 4.1 0.26 50
have been added to the error at frequencies contained in the estimation or 10 4.5 0.46 31
validation data sets.
11 5.0 0.18 87

D. Poles and zeros of the estimated model This can be compared to the corresponding results for the
Figure 8 shows the poles and zeros of the transfer function transmission line model, which are shown in Table II. We note
Ĝ23 (s). Here, we note that all the poles and zeros are well that the resonance frequencies of the transmission line model
separated, which is an indication that the model order is not agree rather well with the MoM results. However, the quality
factors based on the MoM data are consistently smaller than low-order model P̂ (jω)/Q̂(jω) for the continuous frequency
those of the transmission line method, which indicates that response from 10 MHz to 5 GHz, which corresponds to an
the MoM model has more damping. Also, the damping of the electrical size of the system that ranges approximately from
highest-frequency pole in both tables is significantly smaller 0.005 to 2.7 wavelengths. The low-order model features 11
than for the other poles. This can be suspected to be caused pairs of complex-conjugate poles and one real-valued pole,
by the truncation of the frequency response at 5 GHz. Model which sums up to the model order n = 23. The low-
estimation for an extended frequency range verify that this is order model reproduces the correct frequency-response of the
indeed the case. crosstalk current (as computed directly by the MoM) with a
relative error of about 3% or less, which applies to frequencies
TABLE II below 4.5 GHz except where the crosstalk current is very
T HE FREQUENCY, DAMPING AND QUALITY FACTORS FOR THE
COMPLEX - CONJUGATE POLE PAIRS BASED ON THE TRANSMISSION LINE
close to zero. (Here, the estimated response towards the upper
MODEL . end of the frequency range, i.e. 4.5 GHz < f < 5.0 GHz,
suffers from end effects associated with poles and zeros that
r fr [GHz] γr [ns−1 ] Qr [-] are located away for the frequency-samples used for the system
identification.)
1 0.47 0.22 6.6
Finally, we compute the (non-zero) resonance frequen-
2 0.94 0.43 6.9
cies fr and the corresponding quality factors Qr , based on the
3 1.4 0.22 20
estimated complex-conjugate poles βr . Such information that
4 1.9 0.41 14
describes the damped system-resonances is often very useful in
5 2.3 0.22 33
the context of the electromagnetic compatibility assessment of
6 2.8 0.43 21
a device. In addition, the low-order model is extremely fast to
7 3.3 0.22 46
evaluate and, thus, it provides a very powerful representation
8 3.7 0.41 29
of a subsystem that features full-wave phenomena, which then
9 4.2 0.22 59
could be incorporated for efficient and accurate computations
10 4.7 0.41 36
in the context of a system model for a real-world device.
11 5.1 0.14 120
R EFERENCES
[1] K. Yee, “Numerical solution of initial boundary value problems in-
volving maxwell's equations in isotropic media,” IEEE Transactions on
IV. C ONCLUSIONS Antennas and Propagation, vol. 14, no. 3, pp. 302–307, may 1966.
We present a system-identification procedure applied to the [2] A. Taflove, Computational Electrodynamics the Finite-Difference Time-
Domain Method. Artech House Publishers, 1995.
frequency samples of the electromagnetic response computed [3] J. C. Nedelec, “Mixed finite elements in R3 ,” Numerische Mathematik,
by a numerical solver for Maxwell’s equations. We consider a vol. 35, no. 3, pp. 315–341, sep 1980.
crosstalk test-case that consists of two parallel metal strips of [4] J.-M. Jin, The Finite Element Method in Electromagnetics. John Wiley
& Sons, 2014.
length 160 mm that are placed (electrically speaking) very [5] R. F. Harrington, Field Computation by Moment Methods. Macmillan,
close to each other and equally close to a perfect electric 1968.
conducting ground plane. One of the metal strips is excited [6] W. C. Gibson, The Method of Moments in Electromagnetics. Chapman
and Hall/CRC, 2007.
by a voltage source at one end and terminated by a resistor [7] S. Rao, D. Wilton, and A. Glisson, “Electromagnetic scattering by
at the other end. The other strip is terminated by resistors surfaces of arbitrary shape,” IEEE Transactions on Antennas and Prop-
at both ends and we compute the induced far-end crosstalk agation, vol. 30, no. 3, pp. 409–418, may 1982.
[8] M. Khayat, D. Wilton, and P. Fink, “An improved transformation and
current. The electromagnetic response associated with the far- optimized sampling scheme for the numerical evaluation of singular
end crosstalk current is computed by an in-house method of and near-singular potentials,” IEEE Antennas and Wireless Propagation
moments (MoM) solver for the frequency range from 10 MHz Letters, vol. 7, pp. 377–380, 2008.
[9] C. C. Yıldırım, “Scattering and radiation problems of arbitrarily shaped
to 5 GHz, which provides an accurate solution to the full- conducting bodies above the ground plane,” Master’s thesis, Bilkent
wave problem. The test case allows for the application of the University, 2002.
transmission line method and the two models agree very well [10] T. McKelvey, “Frequency domain identification,” IFAC Proceedings
Volumes, vol. 33, no. 15, pp. 7–18, jun 2000.
below 1 GHz. [11] W. H. Press, S. A. Teukolsky, W. T. Vetterling, and B. P. Flannery,
Given the system-identification procedure, we demonstrate Numerical Recipes in C: The Art of Scientific Computing (2nd Edition).
that a relatively few frequency samples of the crosstalk current New York, NY: Cambridge University Press, 1992.
[12] S. Gumussoy, A. A. Ozdemir, T. McKelvey, L. Ljung, M. Gibanica,
can be used to estimate a low-order model that may be and R. Singh, “Improving linear state-space models with additional
expressed in terms of a Padé approximant P̂ (jω)/Q̂(jω), iterations,” IFAC-PapersOnLine, vol. 51, no. 15, pp. 341–346, 2018.
where P̂ (jω) is a polynomial of degree n − 1 and Q̂(jω) is a [13] T. McKelvey, H. Akcay, and L. Ljung, “Subspace-based multivariable
system identification from frequency response data,” IEEE Transactions
polynomial of degree n. Here, n denotes the so-called model on Automatic Control, vol. 41, no. 7, pp. 960–979, jul 1996.
order and the roots to the polynomial Q̂(jω) are estimates of [14] L. Ljung, System Identification: Theory for the User. Prentice Hall,
the complex poles βr = −γr + jωr of the system with reso- 1999.
[15] J. Carlsson, T. Karlsson, and G. Unden, “EMEC—an EM simulator
nances r = 1, 2, . . . , n. For our test case, we use 35 frequency- based on topology,” IEEE Transactions on Electromagnetic Compatibil-
sample points (that are uniformly distributed) to estimate a ity, vol. 46, no. 3, pp. 353–358, aug 2004.

You might also like