Professional Documents
Culture Documents
Wei Zeng1,2, Yue Yang1, Wen-Sheng Qiu1, Huan Xie3 and Su-Chao Xie1
Abstract
Asymmetrical rail grinding in sharp-radius curves could reduce the side wear of railheads and enhance curve capacity of
rail vehicles. The wheel/rail contact performance and curve capacity could be further improved by the optimization of
the asymmetrical rail grinding target profile. In order to modify the target profile smoothly, the nonuniform rational B-
spline curve with adjustable weight factors is used to establish a parameterized model of railhead curves in the asymme-
trical grinding region. The indices of contact performance and curve capacity for different weight factors are obtained
using experiment design and service performance simulation. Two Kriging surrogate models are proposed, in which the
design variables are the adjustable weight factors, and the response parameters are the indices of contact performance
and curve capacity, respectively. The multi-objective optimization model of the target profile is established, in which the
objective functions are the two Kriging surrogate models of contact performance and curve capacity. The optimized
weight factors are sought using a nondominated sorting genetic algorithm II, and the corresponding optimal target profile
is obtained. The wheel/rail service performance simulation before and after optimization indicates that the contact per-
formance and curve capacity are improved significantly.
Keywords
Asymmetrical rail grinding, Kriging, nonuniform rational B-spline, sharp-radius curve, target profile
Creative Commons CC-BY: This article is distributed under the terms of the Creative Commons Attribution 3.0 License
(http://www.creativecommons.org/licenses/by/3.0/) which permits any use, reproduction and distribution of the work without
further permission provided the original work is attributed as specified on the SAGE and Open Access pages (https://us.sagepub.com/en-us/nam/
open-access-at-sage).
2 Advances in Mechanical Engineering
P
n
Ni, p ðuÞwi Pi
From equations (1)–(4), the effective conicity of the i=0
C ðu Þ = Pn , aub ð5Þ
tread J and the lateral offset y* will increase with Dr Ni, p ðuÞwi
when the lateral offset y remains constant, and the i=0
creep force moment M will also increase. It is beneficial where {Pi} are the reference points, {wi} are the weight
for the wheelset to pass through the curve using the factors, and {Ni,p(u)} are the B-spline basis functions of
creep force moment guidance method, which means degree p that are defined in the aperiodic and nonuni-
that the wheel/rail correlation will be improved. form knot vector U (u0, u1,., um). The {Ni,p(u)} are
With decrease in the side wear, the asymmetrical rail defined in equation (6)
grinding could also reduce the shelling defects in gauge
corner. But this requires grinding the gauge corner of
1, ui u ui + 1
the outer rail, which will decrease the rolling radius of Ni, 0 =
0, others
the outer wheel; the guidance ability of wheelset will u ui ui + p + 1 u
also decline.9 Therefore, the gauge corner of the outer Ni, p = Ni, p1 ðuÞ + Ni + 1, p1 ðuÞ
ui + p ui ui + p + 1 ui + 1
rail is not the key grinding region,3 and it will not be
ground in this work. The key regions of asymmetrical ð6Þ
rail grinding are shown in Figure 2. where U is calculated in equation (3)
The contact point of the outer rail in a curve moves
8 9
from co to c0o , and the rolling radius increases from ro < =
to ro0 after asymmetrical grinding. The contact point of U = 0, . . . , 0 , u4 , . . . , um4 , 1, . . . , 1 ð7Þ
the inner rail in a curve also moves from co to c0o , and :|fflfflfflffl{zfflfflfflffl} |fflfflfflffl{zfflfflfflffl} ;
p+1 p+1
the rolling radius decreases from ro to ro0 . Therefore, the
radius difference between the outer rail and the inner
rail Dr increases from ro ri to ro0 ri0 , which is benefi- Parameterization of the railhead curve based on
cial for the train to pass through the curve via the creep cubic NURBS curve
force moment guidance method, meaning that the curve
An NURBS curve is defined by reference points {Pi},
capacity of the rail vehicle will be improved.
weight factors {wi}, and knot vector U, and the curve
shape will change with the modification of any {Pi},
Parameterization of the railhead curve of {wi}, or U. In order to reduce computational effort and
asymmetrical rail grinding ensure the continuity and smoothness of the adjustment
of the target profile, in this study, we adjust the local
In order to analyze the corresponding contact stress region of the railhead curve by modifying the weight
smax and the rolling radius difference of the inner/outer factors {wi}, after which the parameterized model of
wheel Dr in differential asymmetrical rail grinding tar- the railhead curve based on adjustable weight factors is
get profiles, multi-groups of rail grinding target profiles constructed.
with different railhead curves should be designed. As The railhead curve of the 60N used in Chinese trunk
the asymmetrical rail grinding is aimed at a local region routes is taken as the object to construct the NURBS
of a railhead curve, the parameterized model of the rail- parameterized curve of the asymmetrical rail grinding
head curve must have the function to adjust the curve target profile. The curve segment that could contact
shape locally. An NURBS curve is a unified mathemat- with the tread, as shown in Figure 3(a), is selected to be
ical form that is used to describe standard analytical parameterized. Polar coordinates are used, using the
curves and free curves. It could also modify the local center of the railhead curve as the origin. Then, 23
shape of a curve by moving reference points or adjust- points are selected uniformly in the parameterized rail-
ing weight factors.10 Therefore, it is widely used in the head curve as the data points, considering the adjust-
curve parameterization of engineering structures. The ment accuracy and the number of the adjustment
4 Advances in Mechanical Engineering
Figure 3. Parameterized model of the asymmetrical rail grinding target profile based on cubic NURBS curve: (a) the railhead
silhouette of 60N and (b) the parametric railhead curve.
Figure 6. Simulation process of the wheel/rail contact performance and curve passage performance.
Huge computational and low optimization efficiency where b^ is the regression coefficient of variables, f(x) is
will appear if the simulation method, as shown in the base function, and n is the number of training
Figure 6, is taken to calculate the evaluation indices of samples.
the repeated target profiles. A KS model is a kind of Z(X) are the errors of random distribution, and their
interpolation method with a smoothing effect and mini- mean value is 0. Additionally, the variance is s2Z , and
mum variance estimation, considering spatial correla- the covariance is given by
tion of variables sufficiently.16 But additional time will
be spent to construct a KS model, which would increase Cov Z ðxi Þ, Z xj = s2 R R xi , xj ð11Þ
the optimization time, especially when the sampling
points are huge. As it can forecast the dynamic trans- where xi and xj are any two points of the training sam-
formation trend of variables based on the spatial corre- ple and ½R(xi , xj ) is the correlation function, which is
lation analysis of sample data, a KS model is used to fit used to describe the relationship between xi and xj. As
the nonlinear characteristics of the wheel/rail contact the Gaussian correlation function is usually used in
stress and the curve capacity in this study. engineering applications, the correlation function could
be written as
2!
Construction of KS models X
h
R xi , xj = exp uk xik xjk ð12Þ
Basic form of a KS model. A typical KS model is com- k =1
posed of two parts, a polynomial and a random distri-
bution,16 as in equation (9) where h is the number of design variables and xik and xjk
are the corresponding kth components of xi and xj in
^ X + Z ðX Þ
yðX Þ = F b, ð9Þ the training sample.
The response value ^y(x) and an estimated value of
where X is the surrogate model variable space and variance s^ at any point x can be calculated using equa-
^ X ) is the regression model determined by known
F(b, tions (13) and (14)
functions X that are calculated in equation (10)
^ + rð xÞT R1 y f b
^yð xÞ = b ^ ð13Þ
^ x =b
F b, ^ 2 f 2 ð xÞ + + b
^ 1 f 1 ð xÞ + b ^ n f n ð xÞ
1
^ T R1 y f b
^
=b^ ½ f1 ð xÞf2 ð xÞ + + fn ð xÞT ð10Þ ^2 =
s yfb ð14Þ
n
^ T ð xÞ
= bf
Zeng et al. 7
No. Adjustable weight factors Current value Upper limit Lower limit
1 w16 1 2 0
2 w17 1 2 0
3 w18 1 2 0
where rT is a relevance vector of length n, f is a unit col- Several groups points of weight factors are sampled
umn vector of length n, and b^ = ( f T R1 y)1 f T R1 y. in the experimental design method after the value
uk is the value of maximum likelihood estimation ranges of the design variables as w16, w17, and w18 are
(MLE), which is calculated from equation (15) determined. The indices of wheel/rail contact perfor-
mance s and curve capacity Dr under different combi-
1 2 nations of adjustable weight factors are simulated in
MLE = max h ln s^ + lnj Rj ð15Þ
uk .0 2 the analysis process, as shown in Figure 6. The corre-
sponding indices are taken as the training samples to
where the variance s ^ is calculated by equation (14). construct KS models of wheel/rail contact performance
The relevance vector rT between an unknown point and curve capacity. And an error test method is taken
and known sample data points could be calculated to guarantee the fitting accuracy of the KS model.
after the correlation parameter uk is obtained. Then,
the response value of the wheel/rail contact stress s
and rolling radius difference Dr are calculated using Experiment design
equation (9).
Multiple groups of sample points of adjustable weight
Construction of KS models. Taking the adjustable weight factors need to be obtained during the modeling of KS
factors w16, w17, and w18 as the design variables of the models of wheel/rail contact performance and curve
KS model based on the parameterized model of the capacity. The Latin hyper cubic experiment method is
asymmetrical rail grinding target profile shown in used to obtain sample points of adjustable weight fac-
Figure 3. The KS models of wheel/rail contact perfor- tors, as it could sample uniformly, thus avoiding the
mance and curve capacity are established in the con- data missing problem.17 The value ranges of the experi-
struction, as shown in Figure 7. ment design are shown in Table 1.
8 Advances in Mechanical Engineering
The number of training samples is usually three shown in equation (15). The concrete values of the cor-
times the quadratic polynomial coefficient k in the relation parameters are as follows
modeling of a KS model, and the corresponding coeffi-
cient k = (n + 1)(n + 2)/2, n is the number of vari- uk s = ½0:336 0:462 1:065
ables.18 As the number of design variables in this study uk Dr = ½2:000 2:000 1:999
is 3, so the number of training samples of the two KS
models is 30. The relevance vectors rT are calculated after the cor-
A total of 30 groups of adjustable weight factors are relation parameters uk are determined, and the response
obtained using the Latin hyper cubic experiment design value of any unknown design point could be calculated.
method, and the corresponding wheel/rail contact stress Then, the KS models of contact stress s and rolling
s and rolling radius difference Dr are simulated in the radius difference Dr are constructed as per Figure 8.
wheel/rail contact performance and curve capacity The modeling time is 2.3 s, and the memory consump-
analysis process as shown in Figure 6. The simulation tion is 216.3 M.
results are shown in Table 2.
Error test
KS models of wheel/rail contact performance and To guarantee the fitting accuracy of the constructed KS
curve capacity models of wheel/rail contact stress s and rolling radius
difference Dr, the root mean square error (RMSE)19 is
The correlation matrix R and unit column vector f are
used to estimate the fitting accuracy of the two KS
calculated based on the training sample of the KS mod-
models.
els, as shown in Table 2. The correlation coefficients of
^ s and b^ Dr are calculated as follows From equation (13), the mean square error (MSE) of
the two KS models b
any response value of the KS model is
h i
^ s = f T s R1 s ys 1 f T s R1 s ys = 0:395
b
MSE = E ^yð xÞ yð xÞ2
^ Dr = f T Dr R1 Dr yDr 1 f T Dr R1 Dr yDr = 0:005
b 1 ð16Þ
^2
=s 1 + mT f T R1 f m rT R1 r
^ is calculated using R, y, f, and b
Then, R1 (y f b) ^ pffiffiffiffiffiffiffiffiffiffiffi
RMSE = MSE ð17Þ
1 ^ s = ½2:556, 2:462, . . . , 0:307, 0:247
R ys f b
1
s
where m = fTR21r 2 fuk. To guarantee the accuracy of
R yDr f b ^ Dr = ½0:483, 1:131, . . . , 0:057, 1:831
Dr the KS model, the RMSE must be a minimum and .0.
In the error testing, 10 groups of adjustable weight fac-
The estimated value of unbiased variance s ^ 2 of KS tors in Table 2 are selected randomly to plug into the
models is calculated using equation (14). Then, the cor- KS models to calculate the fitting values of s and Dr.
relation parameters of contact stress and rolling radius Then, the fitting values of s and Dr are compared to
difference uks and ukDr are calculated using an MLE, as the corresponding values simulated by SIMPACK in
Zeng et al. 9
Figure 8. KS models of wheel/rail contact performance and curve passage capacity: (a) KS model of s and (b) KS model of Dr.
Figure 12. Contrast of wheel/rail contact performance and curve passage capacity between original and optimal versions: (a) time-
domain graph of contact stress and (b) time-domain graph of rolling radius difference.
1. The parameterized model of the asymmetrical the control system of the rail grinding vehicle,
rail grinding target profile in sharp-radius curves should be carried out.
is constructed based on an NURBS curve. The
weight factors of the NURBS curve are taken as
the design variables, to adjust the railhead curve Declaration of conflicting interests
in the region of the asymmetrical rail grinding, The author(s) declared no potential conflicts of interest with
and then the railhead curve could be adjusted respect to the research, authorship, and/or publication of this
continuously and smoothly. article.
2. The indices of wheel/rail contact performance
and curve capacity in differential combinations Funding
of the adjustable weight factors are simulated The author(s) disclosed receipt of the following financial sup-
using the model in SIMPACK. On this basis, port for the research, authorship, and/or publication of this
two KS models of wheel/rail contact perfor- article: This research was undertaken at the CAD/CAM insti-
mance and curve capacity are constructed, tute, Central South University, China. This work was sup-
which provide a high-fitting surrogate analysis ported by the Natural Science Foundation of Hunan Province
model for the optimization of the target profile. (grant no. 2015JJ2168), the scientific research projects of
3. The multi-objective optimization model of the Shanxi Province Education Department in China (grant no.
symmetrical rail grinding target profile is pro- 16JK2242), and the project (grant no. 51405516) of the
posed, in which the objective functions are the National Natural Science Foundation in China.
two KS models of wheel/rail contact perfor-
mance and curve capacity. The NSGA-II is References
employed in order to optimize the adjustable 1. Cuervo PA, Santa JF and Toro A. Correlations between
weight factors to obtain the optimum Pareto wear mechanisms and rail grinding operations in a com-
solution of the weight factors. Then, the opti- mercial railroad. Tribol Int 2015; 82: 265–273.
mum target profile of the asymmetrical rail 2. Craven N, Bewes OG, Fenech BA, et al. Responding to
grinding is obtained by plugging the optimum the Environmental Noise Directive by demonstrating the
adjustable weight factors into the parameterized benefits of rail grinding on the Great Britain’s* railway
railhead curve model. The optimal results indi- network. Proc IMechE, Part F: J Rail and Rapid Transit
cate that the wheel/rail contact stress value of 2013; 227: 668–676.
the guidance wheel decreased significantly, and 3. Xiao J and Liu X. Design method of rail asymmetric sil-
the rolling radius difference also increased, houette. J Southwest Jiaotong Univ 2010; 45: 361–365.
4. Sroba P, Magel E and Prahl F. RT and S: improved pro-
which demonstrates the viability of this method.
files need special grinding. Railw Track Struct 2004; 100:
4. It must be admitted that the proposed approach
26–28.
is only validated by simulation, which cannot 5. Jia J and Si D. Target profile of rail grinding for small
validate the effectiveness of the approach. And radius curve of Shuohuang railway. China Railw Sci
the actual physical experiment, combining with 2014; 35: 15–21.
12 Advances in Mechanical Engineering
6. Doi H. A study on relationship between rail grinding 15. Wang K. The track of wheel contact points and the cal-
and flange climbing of railway vehicle at low speed in culation of wheel/rail geometric contact parameters. J
curve. Trans Jpn Soc Mech Eng C 2011; 77: 3325–3336. Southwest Jiaotong Univ 1984; 19: 89–99.
7. Zhou J, Liu L and Peng W. Optimization of rail profile 16. Lee J and Kim J. Kriging-based approximate optimiza-
in curved section of railway passenger dedicated line. tion of high-speed train nose shape for reducing micro-
Railw Stand Des 2014; 58: 20–23. pressure wave. Proc IMechE, Part F: J Rail and Rapid
8. Zhou Q, Yinhua Z, Changhai T, et al. Profile design and Transit 2007; 221: 263–270.
test study of 60N rail. China Railw Sci 2014; 35: 128–135. 17. Roshanian J and Ebrahimi M. Latin hypercube sampling
9. Lian S, Sun Q and Wang W. The rail curve rail abrasion applied to reliability-based multidisciplinary design opti-
and its mitigation measures. Beijing, China: China Rail- mization of a launch vehicle. Aerosp Sci Technol 2013; 28:
way Publishing House, 2001, p.126. 297–304.
10. Piegl L and Tiller W. The NURBS book. Beijing, China: 18. Yang Y, Zeng W, Wen-Sheng Q, et al. Optimization of
Tsinghua University Press, 2010, p.86. the suspension parameters of a rail vehicle based on a vir-
11. Espath LFR, Braun AL, Awruch AM, et al. NURBS- tual prototype Kriging surrogate model. Proc IMechE,
based three-dimensional analysis of geometrically nonlinear Part F: J Rail and Rapid Transit. Epub ahead of print 19
elastic structures. Eur J Mech A: Solid 2014; 47: 373–390. January 2016. DOI: 10.1177/0954409715617213.
12. Chen S. Interpolation and application for cubic NURBS 19. Wei Z, Yue Y, Huan X, et al. CF-Kriging surrogate
curves. Mech Sci Techonl 2001; 20: 692–693. model based on the combination forecasting method.
13. Adachi M and Matsumoto A. Improvement of curving Proc IMechE, Part C: J Mechanical Engineering Science.
performance by expansion of gauge widening and addi- Epub ahead of print 5 October 2015. DOI: 10.1177/
tional measures. Proc IMechE, Part F: J Rail and Rapid 0954406215610149.
Transit 2012; 226: 203–215. 20. Khalkhali A. Best compromising crashworthiness design
14. Ashofteh RS. Calculating the contact stress resulting of automotive S-rail using TOPSIS and modified
from lateral movement of the wheel on rail by applying NSGAII. J Cent South Univ 2015; 22: 121–133.
hertz theory. Int J Railw 2013; 6: 148–154.