Received May 22, 2021, accepted June 3, 2021, date of publication June 10, 2021, date of current version

June 17, 2021.

Digital Object Identifier 10.1109/ACCESS.2021.3088285

Crosstalk Prediction in Twisted-Wire Pairs Based

on Beetle Swarm Optimization Algorithm
BAOMING XIAO1,2 , JIANMING ZHOU 3, XINGFA LIU4 , WEI YAN 3, (Member, IEEE),
YI CAO3 , AND YANG ZHAO3
1 State Grid Electric Power Research Institute Company Ltd., Nanjing 210046, China
2 State Key Laboratory of Smart Grid Protection and Operation Control, Nanjing 210046, China
3 School of Electrical and Automation Engineering, Nanjing Normal University, Nanjing 210046, China
4 State Key Laboratory of Power Grid Environmental Protection, Wuhan Branch of China Electric Power Research Institute Company Ltd., Nanjing 210046, China

Corresponding author: Wei Yan ([email protected])

This work was supported in part by the Key Social Development Project of Jiangsu Province under Grant BE2019716, in part by the
Nanjing International Industrial Technology Research and Development Cooperation Project under Grant 201911021, and in part by the
Open Fund Project of State Key Laboratory of Power Grid Environmental Protection under Grant GYW51202001558.

ABSTRACT In this paper, a new algorithm for predicting per unit length (p.u.l) parasitic parameters of
transmission line is proposed. In the twisted-wire pair (TWP), different rotation degrees correspond to
different parasitic parameters, which brings difficulties to the solution of telegraph equation. At present,
the mainstream method is to divide TWP and use the cascade theory to solve each segment as a parallel
transmission line. Therefore, we propose to use the beetle swarm optimization (BSO) algorithm to optimize
the weights of the back propagation neural network, use a certain sample to train the network. This
algorithm is used to predict the p.u.l parameters of uniform and non-uniform TWP, and the finite-difference
time-domain (FDTD) method is used to solve telegraph equation. Among them, the crosstalk value of the
uniform TWP is compared with the simulation value in the CST to verify the effectiveness of the proposed
algorithm; the non-uniform TWP performs 500 random generation calculations to give the maximum
envelope value of the crosstalk.

INDEX TERMS Beetle swarm optimization (BSO), crosstalk prediction, finite-difference time-domain
(FDTD), multiconductor transmission lines (MTL).

I. INTRODUCTION generally to divide it as a cascade of parallel transmission

With the increasing frequency and integration of current lines [7]. This method has the advantages of convenient pro-
power equipment, the problem of crosstalk between trans- cessing and high precision, so it has been widely used later.
mission lines cannot be ignored [1], [2]. Crosstalk is an The research on TWP mainly focuses on its non-uniform
unintentional coupling phenomenon, which means that the characteristics and crosstalk solving methods. In [8], a new
electromagnetic field generated by the voltage and current on non-uniform TWP modeling method is proposed to simulate
the excitation line will affect the adjacent transmission lines the change of the center position of the TWP. Spadacini et al.
and generate induced signals at their endpoints [3]. proposed a prediction model of low-frequency non-uniform
As a common type of transmission line, TWP is of TWP crosstalk under plane wave excitation [1], [2]. The
great significance to the analysis of its crosstalk [4]. equivalent cable bundle method (ECBM) is used with
For parallel transmission lines, its crosstalk value can be some modifications to predict crosstalk within a bundle of
achieved by directly solving the telegraph equation. However, TWP [11].
the cross-section position in TWP is random and unknown, In addition, methods of using neural networks for crosstalk
which brings different p.u.l parameter, so it is difficult to prediction have also been widely proposed. Cannas et al.
predict crosstalk directly by the traditional method [5], [6]. proposed the use of cascading method to segment TWP,
At present, the processing method for TWP crosstalk is some of the p.u.l parameters are used as training samples,
and the rest are predicted by back propagation (BP) neural
The associate editor coordinating the review of this manuscript and network [12], [13]. Similar ideas have also been reported
approving it for publication was Flavia Grassi . in [14], [15].

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
84588 VOLUME 9, 2021
B. Xiao et al.: Crosstalk Prediction in TWP Based on BSO Algorithm

In view of the large prediction error of the existing BP the height of the twisted center from the ground is h, and the
neural network, it is easy to fall into the local minimum during length of the TWP is l. The line is twisted along the z-axis.
the training process, and the result does not change. This The positions of the center points of the two wires are
paper proposes to use the beetle swarm optimization (BSO) El1 and El2 , which are expressed as follows:
algorithm to optimize the weights of the BP neural network, 
d1

d1

that is, the input of the optimized algorithm is the weight and 
 El1 (x1 , y1 , z) = cos θ aEx + h + sin θ aEy + zEaz
2 2


the output is the network error.
El2 (x2 , y2 , z) = d1 cos θ + π aEx

  
On the basis of obtaining the p.u.l parameter at the segment (1)
2 2
π

position, the traditional method to solve the crosstalk is chain

d1

 
sin θ +

parameter matrix method [7], [16]. But the accuracy of this

 + h+ aEy + zEaz
2 2
method is not high, and the number of segments is related
to the constructed TWP model. In [17]–[19], the FDTD where, aEx aEy and aEz represent unit vectors of the x, y, and
method is used to solve the time-domain and frequency- z axes, respectively. θ is the cross section rotation degree at
domain crosstalk of uniform and non-uniform TWP. each position z, which is related to different pitch lengths.
This paper is organized as follows. In Section II, the basic θ = f (z) (2)
assumptions, non-uniform TWP modeling method and MTL
distributed parameter circuit are given. In the section III, From Fig. 1, si is the length of the corresponding segment
the process of optimizing the weights of the BP neural net- in the axial direction, and pi is the type of the corresponding
work by the BSO algorithm and solving the telegraph equa- pitch length. The cross-section rotation degree θ in each
tion by the FDTD method are introduced. The prediction section of Si is:
result is compared with the previously mentioned method. i−1
P
!
In section IV, We give the crosstalk values of uniform and i−1
2π z − sj
non-uniform TWP under three methods, and verify the effec- X 2πsj j=1
θ= + i ≥ 2, z ∈ Si (3)
tiveness of the algorithm by comparing with the simulation pj pi
j=1
value of CST software.
When i = 1 corresponds to uniform TWP.
II. MODELING OF NON-UNIFORM TWP AND
DISTRIBUTED CIRCUIT B. MTL EQUIVALENT DISTRIBUTED PARAMETER CIRCUIT
Before building the model, let us assume some conditions: In the analysis of MTL, we believe that the electromagnetic
1) The bundle is above an infinite metallic ground plane. field surrounding the transmission line has a transverse elec-
2) The geometric shape of the cross section can be consid- tromagnetic (TEM) field structure. In other words, the spatial
ered as a circular outline. electric and magnetic fields are transverse or orthogonal to
3) The structure and material of each line are the same. the axis.
4) The twist length is much shorter than the wavelength. In the TEM propagation mode, MTL can be represented by
5) Weak coupling between transmission lines. a distributed parameter circuit. The micro-element conduc-
Generally speaking, TWP refers to a double transmission tion model of MTL per unit length is shown in Fig. 2. rij , lij , cij
line structure that is evenly twisted. However, in the actual and gij respectively represent the elements in the resistance R,
production process, the twisting distance of each segment inductance L, capacitance C, and conductance G parameter
will produce random fluctuations, which adds difficulty to the matrices, here: i, j = 1, 2, . . . , n. In this paper, the p.u.l
prediction of crosstalk, so we will first introduce a modeling parameter is regarded as not changing with frequency.
method of non-uniform TWP [20].
A. MODELING OF NON-UNIFORM TWP
At present, the modeling method of non-uniform TWP has
been reported in [21], [22]. Taking into account that the
method used in this paper needs to segment the TWP,
the non-uniformity is expressed as: the ratio of the length
of a certain section of the TWP to the pitch length of the
corresponding type is an indefinite value, as shown in Fig. 1.

FIGURE 2. The equivalent circuit model of micro-element.

FIGURE 1. Non-uniform TWP. The satisfied telegraph equation are:

The TWP is cylindrical line with an insulation layer, ∂V(z, t) ∂I(z, t)
the line radius is r, the distance between the lines is d1 , + R(z)I(z, t) + L(z) =0 (4)
∂z ∂t
VOLUME 9, 2021 84589
 B. Xiao et al.: Crosstalk Prediction in TWP Based on BSO Algorithm

∂I(z, t) ∂V(z, t)
+ G(z)V(z, t) + C(z) =0 (5) parasitic parameter matrix are ranked in a column refer to (6)
∂z ∂t as output.
Here, V(z,t) and I(z,t) are voltage and current vectors at
O = [tri(R), tri(L), tri(C), tri(G)]T (6)
different locations and at different times on the transmis-
sion line. R(z), L(z), C(z), and G(z) are variables about Next, we give specific steps to optimize the weights. Since
the position z of the transmission line, both of which are the model simulation in the next section really selects two sets
n-dimensional matrix. The cross-section model of the trans- of TWP, the input of the BP neural network is defined as two
mission line is established in the insert 2D Extractor Design degrees: θ1 and θ2 .
module of ANSYS software to obtain the p.u.l RLCG param- The hidden layer uses the sigmoid function f (x), and output
eter matrices. layer uses the linear function g(x). They are as follows:
After segmenting the TWP, each segment can be regarded 1
as a parallel transmission line structure, which satisfies (4-5). f (x) = (7)
1 + e−x
Therefore, if the parasitic parameter matrix on each segment g (x) = x (8)
of the transmission line is obtained, the FDTD method can
be used to solve the partial differential equation to obtain the Then, the output of neural network is:
crosstalk value. nh
X w2ij
yj = (9)
1 + e−w1i θ
1
III. PREDICT AND SOLVE PARAMETERS i=1
We can get from the section II that the key point of crosstalk For M sets of training samples, the mean square error E
analysis is the acquisition of p.u.l parameters and the solution between the neural network output value yj and the ANSYS
of telegraph equations. Therefore, in this section, the steps of software extracted actual value y0j is:
BSO algorithm to optimize BP neural network weights and
M nh 
FDTD method to solve telegraph equations will be given. 1 XX 2
f (w) = E w1i , wij = yj − y0j


(10)
2M
A. BSO OPTIMIZE WEIGHTS i=1 j=1

In the previous research, we proposed a method to optimize where f (w) is the objective function optimized by the BSO
the weights of the BP neural network using the beetle anten- algorithm, that is, the training error value E, and the optimal
nae search (BAS) algorithm [23], [24]. BAS is a biological value of objective function is set fbest .
heuristic optimization algorithm. It is assumed that there is a From Fig. 3 we can see a basic process for optimizing
beetle in the space, and its left and right antennae are able to weights.
receive the ‘‘smell’’ emitted by the target respectively. If the
odor received by the left antennae is stronger, the beetle shifts
to the left, and vice versa. Ultimately achieve the purpose of
optimization..
However, during our continuous research process,
we found that the BAS algorithm is not ideal when dealing
with high-dimensional data, and the final iteration result
depends to a large extent on the initial position of the beetle.
In other words, the selection of the initial beetle position
affects the effectiveness of the entire optimization process.
Inspired by the particle swarm algorithm, we extend the
beetle individual in the BAS algorithm to the beetle group,
which is the BSO algorithm.
In this algorithm, each beetle represents a potential opti-
mization solution. Similar to the particle swarm algorithm, FIGURE 3. BSO algorithm to optimize BP weights.
the beetles can share information with each other. However,
the specific distance and direction of each beetle depends Here, the position of each beetle is the BP weight, and the
on the odor intensity, which is the fitness function. fitness function is the BP error.
In this section, we need to optimize the weights of the BP Step 1: Generate a population of n beetles: B =
neural network. In other words, the position of each beetle is (B1 , B2 , . . . , Bn ) in an S-dimensional search space, where
represented by a set of weights, and the fitness function is the the i th beetle represents a vector: Bi = (bi1 , bi2 , biS ), this
error of the neural network. After the final iteration is com- represents a potential solution to the optimization problem.
pleted, the position of the beetle corresponding to the smallest The speed of the i th is expressed as: Vi = (Vi1 , Vi2 , . . . ViS ).
error value is found. In order to improve the calculation speed, we have stipulated
In the BP neural network, we take the rotation degrees the threshold range for the individual and the whole beetle:
of TWP as input, and the upper triangular elements of the Ui = (Ui1 , Ui2 , . . . UiS ) and Ug = (Ug1 , Ug2 , . . . UgS ).

84590 VOLUME 9, 2021

B. Xiao et al.: Crosstalk Prediction in TWP Based on BSO Algorithm

Step 2: Give the iterative formula of beetle position and

speed.

Bk+1
is = Bkis + λVisk + (1 − λ)δisk (11)

where: s = 1, 2, . . . S; i = 1, 2, . . . n; k is the current number

of iterations. δis is increase in beetle position movement. λ is
a constant.
   
Visk+1 = ωVisk + c1 r1 Uisk − Bkis + c2 r2 Ugs k
− Bkgs (12)

where: ω is inertia weight, c1 , c2 are constants, r1 and r2 are

FIGURE 5. Spatial-temporal dispersion of voltage and current.
two random functions in the range [0,1].
Step 3: Give the distance and direction of the beetle.
Equation (1-2) can be discretized according to the
    
δisk+1 = εk ∗Visk ∗ sign f Bkrs − f Bkls (13)
Implicit-Wendroff format, as shown below:
In this step, ε is step size. The search functions of the left Vnj+1 − Vnj + Vn−1 n−1
j+1 − Vj
and right antennae of the beetle are respectively expressed as:
21z
Bk+1 = Bkrs + Visk ∗ d/2 (14) R(j1z)(Inj+1 + Inj + In−1 n−1
j+1 + Ij )
rs +
Bk+1
ls = Bkls − Visk ∗ d/2 (15) 4
L(j1z)(Inj − In−1
j
n−1
+ Inj+1 − Ij+1 )
d represents the distance between two antennae. After the + =0 (17)
21t
iterative loop is over, return a minimum error value, and Inj+1 − Inj + In−1 − In−1
j+1 j
the corresponding beetle position, that is, the weights. For
equations of BP neural network construction, please refer 21z
to [23], [24]. G(j1z)(Vnj+1 + Vnj + Vn−1 n−1
j+1 + Vj )
+
4
B. FDTD SOLVE CROSSTALK C(j1z)(Vnj − Vn−1
j + Vnj+1 − Vn−1
j+1 )
After using the trained neural network to predict p.u.l param- + =0 (18)
21t
eter matrix, the next step is to use FDTD to solve the telegraph
equation to get crosstalk. The derivative in the telegraph equa- Further simplified to

tion is discretized and approximated by different finite differ- Vnj − Vnj+1 + AVj (Inj + Inj+1 ) = −Vn−1 + Vn−1
j j+1


ences. As shown in Fig. 4, the transmission line is divided 
+ B (In−1 + In−1 )


vj
into NDZ sections, each length is 1z; the total solution time j j+1
is divided into NDT sections, each length is 1t. 

 AIj (Vnj + Vnj+1 ) + Inj − Inj+1 = BIj (Vn−1
j + Vn−1 n−1
j+1 ) − Ij
 n−1
 +I j+1
(19)
where the coefficient matrix are:
AVj = −(R(j1z)/2 + L(j1z)/1t)1z
BVj = (R(j1z)/2 − L(j1z)/1t)1z (20a)
AIj = −(G(j1z)/2 + C(j1z)/1t)1z
BIj = (G(j1z)/2 − C(j1z)/1t)1z (20b)
Combining (19) with the Thevenin condition of the trans-
FIGURE 4. Discretization of transmission line length in FDTD analysis. mission line, the near-end voltage V(0,t) and the far-end
voltage V(L, t), and the voltage and current at any position
At the same time, in order to ensure the stability of the dis- and time can be obtained. As for the near-end crosstalk
cretization and the accuracy of the second-order calculation, (NEXT) and the far-end crosstalk (FEXT), the voltage can be
the voltage at NDZ + 1 point and the current at NDZ point transformed from the time domain to the frequency domain
are interleaved, as shown in Fig. 5. and are expressed as:
Where:
NEXT = 20 log10 (V2 (0, f )/Vs ) , (21a)
Vnj = V(j1z, n1t), Inj = I(j1z, n1t) (16) FEXT = 20 log10 (V2 (L, f )/Vs ) . (21b)

VOLUME 9, 2021 84591

 B. Xiao et al.: Crosstalk Prediction in TWP Based on BSO Algorithm

where: VS is supply voltage, V2 (0, f ) represents the near-end TABLE 2. Twist lay length.
induced voltage of victim line, V2 (l, f ) represents the far-end
induced voltage of victim line, f is frequency.

IV. SIMULATION RESULTS AND DISCUSSIONS

In the section IV, we selected two sets of TWPs as simulation
verification models. Firstly, the influence of TWP twisting
degree and relative position on the crosstalk value is given by
CST software; secondly, the parameter matrix is predicted by TABLE 3. Relative distance between two TWPs axial starting points.
the BSO algorithm proposed in Section III and the telegraph
equation is solved by FDTD, the training samples of neural
network are extracted by ANSYS Q3D EXTRACTOR mod-
ule. The results of uniform and non-uniform TWP are both
given.

A. VALIDATION MODEL
In this paper, we have established two sets of TWPs verifica-
tion models, as shown in Fig. 6. The rotation degree of the two
TWPs corresponds to the input of the neural network in Fig. 3.

FIGURE 6. Establishment of TWPs verification model.

The excitation source in the model is a sinusoidal signal

with an amplitude of 1V and an internal resistance of 100.
The port of each line is connected with 50 resistor, the spe-
cific parameters are shown in Table 1:

TABLE 1. Model parameters.

FIGURE 7. The effect of twisting tightness on crosstalk.

Since the result values of line 3 and line 4 are the same,
we only give one set of values.
From Fig. 7 we can observe that as the TWP twists tighter,
the greater the impact on another group of TWP. But this
B. INFLUENCE FACTORS OF TWP CROSSTALK trend of change is not significant enough, especially in high
In the process of studying the crosstalk of multiple pairs of frequency bands. At the same time, we noticed that if the two
TWPs, we found that the tightness of the twist and the relative sets of TWP adopt different twisting methods, the influence
position of the twist have a significant impact on the size of crosstalk can be effectively suppressed. This is because
of the crosstalk. We simulated the two influencing factors each group of TWP can actually be regarded as an unshielded
separately, and the control parameters are shown in Table 2 capacitor. The signal will send a certain amount of elec-
and Table 3, where: Table 2 is for synchronous twisting tromagnetic waves to the outside during the transmission
(0 distance), and Table 3 is for 25mm-25mm twisting. process. The electromagnetic waves emitted by different lay
The above situations were simulated in CST, near-end lengths have different frequencies, and they can interfere with
crosstalk and far-end crosstalk were separated and compared. each other smaller.

84592 VOLUME 9, 2021

B. Xiao et al.: Crosstalk Prediction in TWP Based on BSO Algorithm

FIGURE 10. Crosstalk comparison under different algorithms.

FIGURE 8. The effect of twisting relative distance on crosstalk.

the combination of particle swarm and BAS proposed in this

paper means that the BSO algorithm has better iterative abil-
ity and accuracy, which also makes the subsequent solution
of crosstalk more accurate.
In Fig. 10, we give the comparison of the near-end and
far-end crosstalk values, which obtained by CST simulation
and under three algorithms of BP, BAS-BP, and BSO-BP
respectively.
We observe from the Fig. 10 that the predicted value of the
BP neural network optimized by BSO has the best fit with the
simulation value, while the un-optimized BP neural network
FIGURE 9. BSO and BAS iterative fitness function. has the worst fit.
The simulation frequency band is divided into three sec-
From Fig. 8 we observe that the relative position of the tions, namely low frequency band: 5 ∼ 100MHz, middle
twist between TWPs has a certain effect on crosstalk. And frequency: band 100 ∼ 500MHz and high frequency band:
when there is a difference of 1/4 lay distance (10mm), 600 ∼ 900MHz. The average error values of the three meth-
the crosstalk value is the smallest. These two influencing ods for each interval are given in Table 4.
factors should be considered in the design process such as
network cables. TABLE 4. Average error (%) of crosstalk (N: Next; F: Fext).

C. COMPARATIVE RESULTS
In section IV-C, we use the algorithm proposed in the
section III to predict and solve the crosstalk, and compare
the results with the previous BP and BAS-BP neural network.
First of all, what we verified is the uniform twisted TWP,
as shown in Fig. 6 and Table 1.
We use BSO and BAS algorithms to iteratively optimize
the weights of the BP neural network, and the fitness curve is After verifying the effectiveness of the method proposed in
shown in Fig. 9. Compared with the previous BAS algorithm, this paper through the above model, we predict the crosstalk

VOLUME 9, 2021 84593

 B. Xiao et al.: Crosstalk Prediction in TWP Based on BSO Algorithm

cyclically, and give the maximum crosstalk threshold to pro-

vide a certain reference for subsequent EMC-related designs.

REFERENCES
[1] S. Chabane, P. Besnier, and M. Klingler, ‘‘A modified enhanced transmis-
sion line theory applied to multiconductor transmission lines,’’ IEEE Trans.
Electromagn. Compat., vol. 59, no. 2, pp. 518–528, Apr. 2017.
[2] Y. Wang, Y. S. Cao, D. Liu, R. W. Kautz, N. Altunyurt, and J. Fan, ‘‘A gen-
eralized multiple-scattering method for modeling a cable harness with
ground connections to a nearby metal surface,’’ IEEE Trans. Electromagn.
Compat., vol. 61, no. 1, pp. 261–270, Feb. 2019.
[3] C. R. Paul, Analysis of Multiconductor Transmission Lines, 1st ed.
Hoboken, NJ, USA: Wiley, 1994.
[4] F. Distler, G. Gold, K. Thurn, J. Schur, and M. Vossiek, ‘‘Crosstalk simula-
tion of multiple insulated twisted pairs based on transmission line theory,’’
in Proc. IEEE 21st Workshop Signal Power Integrity (SPI), Baveno, Italy,
May 2017, pp. 1–4, doi: 10.1109/SaPIW.2017.7944033.
[5] G. Spadacini and S. A. Pignari, ‘‘Radiated susceptibility of a twisted-
wire pair illuminated by a random plane-wave spectrum,’’ IEICE Trans.
Commun., vol. E93-B, no. 7, pp. 1781–1787, Jul. 2010.
[6] A. A. Ilumoka, ‘‘Efficient prediction of crosstalk in VLSI interconnections
using neural networks,’’ in Proc. IEEE 9th Topical Meeting Electr. Per-
form. Electron. Packag., Scottsdale, AZ, USA, Oct. 2000, pp. 87–90, doi:
10.1109/EPEP.2000.895499.
[7] C. Paul and J. McKnight, ‘‘Prediction of crosstalk involving twisted pairs
of wires—Part I: A transmission-line model for twisted-wire pairs,’’ IEEE
Trans. Electromagn. Compat., vol. EMC-21, no. 2, pp. 92–105, May 1979,
doi: 10.1109/TEMC.1979.303751.
[8] X. Song, J. Wang, B. Li, and D. Su, ‘‘Crosstalk model for shielded bundles
of random twisted-wire pairs,’’ in Proc. Asia–Pacific Int. Symp. Electro-
FIGURE 11. Non-uniform TWP crosstalk prediction chart. magn. Compat. (APEMC), Shenzhen, China, May 2016, pp. 766–769, doi:
10.1109/APEMC.2016.7522860.
[9] G. Spadacini, D. Bellan, and S. A. Pignari, ‘‘Impact of twist non-uniformity
value of the non-uniform TWP in the section II. In order on crosstalk in twisted-wire pairs,’’ in Proc. IEEE Symp. Electromagn.
Compat. Symp. Rec., Boston, MA, USA, vol. 2, Aug. 2003, pp. 483–488,
to make the results more general, we randomly generated doi: 10.1109/ISEMC.2003.1236649.
200 sets of non-uniform TWP, hoping to find the upper [10] G. Spadacini and S. A. Pignari, ‘‘Numerical assessment of radiated sus-
envelope value. ceptibility of twisted-wire pairs with random nonuniform twisting,’’ IEEE
Trans. Electromagn. Compat., vol. 55, no. 5, pp. 956–964, Oct. 2013, doi:
In Fig. 11, the upper envelope values of the near-end and
10.1109/TEMC.2012.2235446.
far-end crosstalk of the non-uniform TWP are given, respec- [11] S. Belkhelfa, M. Lefouili, and K. E. K. Drissi, ‘‘Frequency domain analysis
tively. We hope to generate and predict the crosstalk value of EM crosstalk problem in a quad by the equivalent cable bundle method
under non-uniform conditions through multiple generations, among twisted-wire pairs cable bundle,’’ IEEE Trans. Magn., vol. 51,
no. 11, pp. 1–4, Nov. 2015, doi: 10.1109/TMAG.2015.2434819.
which to provide a certain reference for the subsequent EMC [12] B. Cannas, A. Fanni, and F. Maradei, ‘‘A neural network approach to
design phase of electronic equipment. predict the crosstalk in non-uniform multiconductor transmission lines,’’
in Proc. IEEE Int. Symp. Circuits Syst., Scottsdale, AZ, USA, May 2002,
p. 1, doi: 10.1109/ISCAS.2002.1009905.
V. CONCLUSION [13] B. Cannas, A. Fanni, and F. Maradei, ‘‘Crosstalk prediction in twisted
In this paper, a modeling method of non-uniform TWP is bundles by a neural approach,’’ in Proc. 3rd Int. Symp. Electromagn.
Compat., Beijing, China, 2002, pp. 638–641, doi: 10.1109/ELMAGC.
introduced, which is regarded as the accumulation of several 2002.1177512.
segments of different pitch types. In addition, the simula- [14] J. Dai, F. Dai, and X. Zhao, ‘‘Output prediction about crosstalk coupling
tion analysis of different twist densities and relative dis- of representative multi-cable bundle on aircraft platform based on neural
tances under two uniform TWPs is done, and found that network,’’ in Proc. 3rd IEEE Int. Symp. Microw., Antenna, Propag. EMC
Technol. Wireless Commun., Beijing, China, Oct. 2009, pp. 381–384, doi:
the same TWPs had the worst anti-electromagnetic interfer- 10.1109/MAPE.2009.5355734.
ence ability. In other words, the non-uniformity improves the [15] F. Dai, G. Bao, and D. Su, ‘‘Crosstalk prediction in non-uniform cable
anti-interference ability. bundles based on neural network,’’ in Proc. 9th Int. Symp. Antennas,
Propag. EM Theory, Guangzhou, China, Nov. 2010, pp. 1043–1046, doi:
Finally, the difference in p.u.l parameters due to the twist 10.1109/ISAPE.2010.5696654.
degree of the TWP brings difficulties to the solution of [16] Y. Sun, J. Wang, W. Song, and R. Xue, ‘‘Frequency domain analy-
crosstalk. Therefore, this paper proposes the BSO-BP neural sis of lossy and non-uniform twisted wire pair,’’ IEEE Access, vol. 7,
pp. 52640–52649, Apr. 2019.
network algorithm to predict the parameter matrix, and its
[17] A. Tatematsu, F. Rachidi, and M. Rubinstein, ‘‘A technique for calculating
optimization and iteration ability is greatly improved com- voltages induced on twisted-wire pairs using the FDTD method,’’ IEEE
pared with the previous BAS algorithm. In the final crosstalk Trans. Electromagn. Compat., vol. 59, no. 1, pp. 301–304, Feb. 2017.
prediction, the accuracy of the BSO algorithm is about 3 times [18] V. R. Kumar, B. K. Kaushik, and A. Patnaik, ‘‘An accurate FDTD
model for crosstalk analysis of CMOS-gate-driven coupled RLC intercon-
that of BAS and 150 times that of BP. At the same time, nects,’’ IEEE Trans. Electromagn. Compat., vol. 56, no. 5, pp. 1185–1193,
we also use this algorithm to predict the non-uniform TWP Oct. 2014.

84594 VOLUME 9, 2021

B. Xiao et al.: Crosstalk Prediction in TWP Based on BSO Algorithm

[19] P. Zhang, X. Du, J. Zou, J. Yuan, and S. Huang, ‘‘Iterative solution of MTL XINGFA LIU received the B.S. degree in elec-
based on the spatial decomposition and the second-order FDTD,’’ IEEE trical engineering from the Naval University of
Trans. Magn., vol. 54, no. 3, pp. 1185–1193, Mar. 2018. Engineering, Wuhan, China, in 2002, and the
[20] A. Umek, ‘‘Modeling the structural return loss in twisted pair cables,’’ in M.S. degree from the Huazhong University of
Proc. 10th Medit. Electrotech. Conf. Inf. Technol. Electrotechnol. Medit. Science and Technology, Wuhan, in 2007. He is
Countries MeleCon, Lemesos, Cyprus, vol. 1, 2000, pp. 173–176, doi: currently working with the China Electric Power
10.1109/MELCON.2000.880395. Research Institute Company Ltd., Beijing, China.
[21] G. Spadacini, F. Grassi, and S. A. Pignari, ‘‘A closed-form low-
His current research interests include electromag-
frequency statistical model for field-to-wire coupling in twisted-wire pairs
netic environments, electromagnetic compatibil-
with non-ideal twisting,’’ in Proc. 31st URSI Gen. Assem. Sci. Symp.
(URSI GASS), Beijing, China, Aug. 2014, pp. 1–4, doi: 10.1109/URSI- ity of power systems, and biological effects of
GASS.2014.6929551. electromagnetic fields.
[22] O. Gassab, L. Zhou, and W.-Y. Yin, ‘‘Stochastic analysis of multi-twisted
bundle of twisted-wire pairs (MTB-TWP) above ground plane with random
non-uniform twisting,’’ in Proc. Int. Symp. Electromagn. Compat. (EMC
Eur.), Barcelona, Spain, Sep. 2019, pp. 421–426, doi: 10.1109/EMCEu-
WEI YAN (Member, IEEE) received the M.S.
rope.2019.8871803.
degree in electrical engineering and the Ph.D.
[23] C. Huang, Y. Zhao, W. Yan, Q. Liu, and J. Zhou, ‘‘A new method for
predicting crosstalk of random cable bundle based on BAS-BP neural degree in physics and electronics from Nanjing
network algorithm,’’ IEEE Access, vol. 8, pp. 20224–20232, 2020, doi: Normal University, Nanjing, China, in 2011 and
10.1109/ACCESS.2020.2969221. 2014, respectively. Since 2014, he has been with
[24] Q. Liu, Y. Zhao, W. Yan, C. Huang, A. Mueed, and Z. Meng, the Jiangsu Electrical Equipment EMC Engi-
‘‘A novel crosstalk estimation method for twist non-uniformity in neering Laboratory, Nanjing Normal University,
twisted-wire pairs,’’ IEEE Access, vol. 8, pp. 38318–38326, 2020, doi: where he is currently working as an Associate
10.1109/ACCESS.2020.2976136. Professor. His current research interests include
integrated circuit electromagnetic compatibility
testing, bio-electromagnetic technology, and automotive electromagnetic
compatibility design.

BAOMING XIAO was born in July 1974.

He received the master’s degree in power sys-
tem and automation from North China Electric YI CAO was born in Ningguo, Anhui, in 1971.
Power University. He is currently the Director of She received the bachelor’s degree in electrical
the State Grid Corporation’s Automation Equip- engineering from the Hefei University of Tech-
ment Electromagnetic Compatibility Laboratory nology, in 1992, and the master’s degree from the
of the State Grid Corporation. His research inter- Nanjing University of Aeronautics and Astronau-
ests include the study of electromagnetic compati- tics, in 1999. She is mainly engaged in the research
bility mechanism and electromagnetic interference of control theory and computer control.
performance of power automation equipment test-
ing, certification, and related standardization work.

JIANMING ZHOU was born in Jiangsu, China. YANG ZHAO received the B.E., M.E., and Ph.D.
He received the B.S. degree from the School of degrees in power electronic technology from the
Electrical Engineering and Automation, Suzhou Nanjing University of Aeronautics and Astronau-
University of Science and Technology, Suzhou, tics, Nanjing, China, in 1989, 1992, and 1995,
China, in 2019. He is currently pursuing the mas- respectively. He is currently a Professor with
ter’s degree in electrical engineering with Nanjing Nanjing Normal University. His research interests
Normal University, Nanjing, China. His major include electromagnetic compatibility, power elec-
research interest includes new technology of elec- tronics, and automotive electronics.
trical engineering.

VOLUME 9, 2021 84595