MULTI-OBJECTIVE DESIGN METHOD OF RAIL GRINDING TARGET

PROFILE IN HIGH-SPEED TURNOUT AREA CONSIDERING LONG-TERM

SERVICE PERFORMANCE MAINTENANCE

Abstract: Aiming at the problems of fast wear rate and when the train passes the switch. At the same time, it can
serious damage of rails in the switch area of high-speed also reduce the dynamic response of the wheel-rail
turnouts, a design method of rail grinding target profile contact and slow down the wear and tear, which is
that can consider the long-term service performance of unparalleled for improving the service life of the rail and
turnouts is proposed by combining the deep neural ensuring the safety of train operation important role.
network model and the multi-objective optimization Existing studies are usually aimed at solving the existing
algorithm. The optimized grinding profile not only problems in the field, such as fatigue cracks (Fig. 1),
significantly improves the dynamic performance of the double light bands (Fig. 2), rolling cars, etc. At the same
vehicle passing the turnout in the short term, but also time, most of them only focus on the rail performance in
reduces the accumulated wear of the rail after the long- the short term after rail grinding. In fact, the long-term
term operation of the turnout, which is beneficial to retention capability of the rail profile is also extremely
prolong the service life of the turnout. important. Therefore, it is necessary to consider the
service performance changes of the turnout under long-
Keywords: High-speed turnout; dynamic performance; term operating conditions when designing the target
wheel-rail contact damage; long-term service profile for rail grinding.
performance; multi-objective optimization.

1. Introduction
The complex structure and many diseases of turnouts
have always been a difficult point in the maintenance and
repair of high-speed railway lines, and it is also a key
infrastructure that affects the safety and stability of
driving. The variable-section rails of turnouts are
relatively weak, and are prone to wear during long-term
operation, resulting in changes in the combined profile of Figure 1 The fatigue crack
the turnout, and then the problem of poor wheel-rail
matching, resulting in the continuous reduction of the
service performance of the turnout. Rail grinding is an
economical and effective maintenance measure for
turnouts. In the early days, most of the damaged rails
were ground by empirical methods. In recent years,
experts and scholars have carried out a lot of work on the
optimization of rail profiles. For example, Magel [1]
designed the rail grinding target profile based on the
theory of contact mechanics, which can effectively Figure 2 Double light strip
reduce the contact stress of the wheel and rail. Persson [2] This paper takes a group of No. 18 high-speed turnouts in
used Genetic Algorithm (GA) to optimize the profile of service at Linfen West Station as the research object, and
BV50 rail and UIC60 rail to reduce wheel-rail rolling proposes a design method for the target profile design of
contact fatigue. Pålsson [3] used the B-spline curve to rail grinding in the switch area based on long-term service
describe the profile of the variable-section rail in the performance prediction. The Nonuniform rational B-
switch area. Taking the wheel-rail contact stress and splines (NURBS) method is used to establish a
energy loss as the objective function, the GA algorithm parametric model of the rail space profile in the switch
was used to solve the target profile with better area, and the grinding amount of each control point is
performance. Wan [4] formulated the optimization used as the design variable. In order to improve the
problem as a weighted multi-objective problem, solved it stability of the vehicle over the turnout, and at the same
with different weight coefficient sets, optimized the time consider the long-term service performance of the
profile of the key section of the fixed frog, and reduced turnout, the lateral acceleration of the axle box and the
the wheel-rail rolling contact fatigue and wear. Wang [5] rail wear after a certain gross weight are selected as the
predicted the rail wear in the curved section of the heavy- objective functions. In order to improve the
haul railway based on the response surface method, and computational efficiency, the explicit relationship
used the genetic algorithm to optimize the rail profile to between the rail grinding amount and the objective
prolong the service life of the rail. function is constructed by combining with the deep
neural network. According to the corresponding
A good wheel-rail matching relationship is conducive to constraints, the multi-objective optimization
the improvement of the stability and comfort of the train mathematical model of rail grinding for turnout was
constructed, and the NSGA-II algorithm was used to 𝑁𝑖,𝑝
solve it, and the turnout profile with better long-term 𝑢 − 𝑢𝑖
= 𝑁 (𝑢)
service performance was obtained. Fig. 3 shows the 𝑢𝑖+𝑝 − 𝑢𝑖 𝑖,𝑝−1 (3)
optimization process of the rail combination profile in the 𝑢𝑖+𝑝−1 − 𝑢
switch area. + 𝑁 (𝑢)
𝑢𝑖+𝑝+1 − 𝑢𝑖+1 𝑖+1,𝑝−1
Start
The node vector U is:
Establish the spatial parametric
𝑼
model of high-speed turnout

= {0,
⏟. . . ,0 , 𝑢𝑝+1 , . . . , 𝑢𝑟−𝑝−1 , 1,
⏟. . . ,1} (𝑟 (4)
Generate training 𝑝+1 𝑝+1
samples(x1,x2, xn)
Generate combined profile = 𝑛 + 𝑝 + 1)
of turnout area
The above is the basic definition and representation
Establish the vehicle-turnout
system dynamic model method of the two-dimensional NURBS curve. Since the
section of the switch point rail changes continuously
Vehicle-turnout dynamic Wheel-rail Wheel-rail contact along the longitudinal direction of the line, it can be
model simulation interaction characteristics
considered to use the three-dimensional NURBS surface
No to parameterize it, which can maintain good continuity
and convex hull. A set of NURBS surface equations can
Gross weight>=10 Mt
Rail profile update
Wear amount be expressed as:
Objective functions:
Dynamic performance; Yes
𝑆(𝑢, 𝑣) =
∑𝑛 𝑚
Wheel-rail contact damage
𝑖=0 ∑𝑗=0 𝑁𝑖,𝑝 (𝑢)𝑁𝑗,𝑞 (𝑣)𝑊𝑖,𝑗 𝑷𝑖,𝑗
Establish deep neural network
∑𝑛 𝑚 (0 ≤ (5)
Hyperparameter
𝑖=0 ∑𝑗=0 𝑁𝑖,𝑝 (𝑢)𝑁𝑗,𝑞 (𝑣)𝑊𝑖,𝑗
optimization
𝑢, 𝑣 ≤ 1)
No
Prediction accuracy check where {Pi,j} is the position vector of the control vertex of
Yes
the NURBS surface, which is used to form the control
Multi-objective
grid and approximate the shape of the switch rail, {Wi,j}
optimization:NSGA-II is the weight factor, {Ni,p(u)} and {Nj,q(v)}are the p-spline
basis functions in the u and v directions, respectively. In
Output optimal target profile this paper, the basic rail is simplified as a constant cross-
section, and the two-dimensional NURBS curve
Stop (Equation 1) is used for parameterization. At the same
Figure 3 Flowchart of the optimal design of the rail profile time, the variable cross-section characteristics of the tip
rail are considered, and the three-dimensional NURBS
2. Parametric Model of Spatial Profile of High- surface (Equation 5) is used for the parameterization.
speed Turnout
In this paper, based on the NURBS theory, a three- 2.2. Parameterization of rail profile in switch area
dimensional parametric model of the rail profile in the In this section, a parametric model is established by
switch area is constructed. Only a few control points are taking the profile of the straight tip rail in the switch area
needed to achieve high-precision fitting of the rail profile of the No. 18 high-speed turnout as an example. The top
in the switch area. widths of the tip rails of 3 mm, 10 mm, 15 mm, 20 mm,
30 mm, 40 mm, 50 mm, 55 mm, 60 mm, 65 mm, 70 mm
2.1. NURBS theory and the full top width were selected as the control
NURBS is a parametric spline curve, which is used by the sections. Due to the obvious discontinuity at the
International Organization for Standardization as the only connection between the working edge curve and the non-
mathematical method to define the geometry of industrial working edge of the tip rail, it is easy to generate a large
products [6]. A NURBS curve of degree p is defined as fitting error. Considering that the non-working edge of
follows [7]: the tip rail usually does not participate in the wheel-rail
∑𝑛𝑖=0 𝑁𝑖,𝑝 (𝑢)𝜔𝑖 𝑃𝑖 contact calculation, only the control of the straight tip rail
𝐶(𝑢) = 𝑛 ,0 ≤ 𝑢 ≤ 1 (1) is performed. The profile of the working edge of the
∑𝑖=0 𝑁𝑖,𝑝 (𝑢)𝜔𝑖
In the formula, {Pi} is the control vertex position vector, section is discretized to obtain the coordinates of the
which is used to form the control polygon; {ωi } is the corresponding type value points. The corresponding
weight factor, which is used to modify the shape of the NURBS control vertex positions are obtained by using
NURBS curve in the interval u∈[ui,ui+1]; {Ni,p(u)} is the the De Boor algorithm, and the control vertex grid of the
p-order B-spline basis function defined on the aperiodic straight-point rail of the No. 18 turnout is obtained by
and non-uniform nodal vector U(u0, u1,…,um), which can combining the longitudinal mileage corresponding to
be calculated by the De Boor formula [8], and can be each control section, as shown in Fig. 4. The cross-
expressed as: sectional profile at any position in the switch area can be
1, 𝑖𝑓 𝑢𝑖 ≤ 𝑢 ≤ 𝑢𝑖+1 automatically generated by the NURBS surface, and at
𝑁𝑖,0 = { (2) the same time, the profile of the non-working side of the
0, 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒
point rail can be supplemented, and the three-dimensional
space profile of the point rail of the switch can be
obtained as shown in Fig. 5. The content of this section
is realized by Matlab programming. According to the generated by interpolation in the parametric model, so
literature [9], it can be known that the turnout profile that the transition between the optimized profile and the
accuracy obtained by this method can meet the needs of full top width profile is smooth. Based on the above
subsequent simulation calculations. In the same way, principles, the key sections of the switch area that need to
parameterize the rail profile of the No. 18 high-speed be optimized include straight basic rails, curved basic
turnout measured on site at Linfen West Station to obtain rails, and straight tip rail top widths of 20 mm, 30 mm,
the rail profile of the switch area of the corresponding 40 mm, and 50 mm. The grinding amount of each control
section, and compare the key section profile with the point in the wheel-rail contact area on each key section is
standard profile. The wear is more serious, so when selected as the decision variable, as shown in Fig. 7. The
optimizing it, in addition to improving the stability of the number of movable points in the curved basic rail section
vehicle when passing the fork, the development of wheel is 9, and the number of movable points in each control
and rail wear should also be delayed as much as possible. section of the straight-tip rail is 14, so the total dimension
of decision variables is 74. According to the requirements
for rail grinding in the high-speed railway section of
Linfen West Station, the maximum grinding amount of
each rail piece cannot exceed 1.0 mm, so the value range
of each variable is [0, 1] mm. By adjusting the values of
each variable through the NURBS parametric model, any
number of rail space profiles in the switch area can be
generated, thereby providing sufficient training samples
for subsequent calculations.

3. Service performance prediction model of

Figure 4 The control grid of the NURBS surface turnout
In this paper, the service performance of the turnout is
comprehensively evaluated by the stability of the vehicle
passing the switch and the wheel-rail contact damage
index, and the profile of the rail combination in the switch
area is optimized. In this process, it is necessary to drive
the corresponding physical simulation analysis model to
perform a large number of iterative calculations, and the
time cost is high. Deep learning has been widely used in
various fields in recent years, and it is very effective in
solving supervised learning problems, which can greatly
improve the solution efficiency while ensuring
Figure 5 Three-dimensional profile of switch rail computational accuracy. The service performance
prediction of the turnout after rail grinding can be
regarded as a kind of supervised learning problem, in
which the sample feature is the grinding amount of the
turnout rail, and the label is the stability index and the rail
wear of the turnout when the gross weight is 10Mt. It
should be noted that the long-term service performance
of the turnout can be reflected to a certain extent when
the gross weight reaches 10Mt, and the calculation cost is
also within the acceptable range. Therefore, the
performance of the turnout under the total passing weight
is selected as the sample label. In this section, the deep
Figure 6 Comparison of standard profile and measured profile
neural network is combined with the dynamic
performance of the vehicle and the contact damage of the
wheel and rail, and the mapping relationship between the
grinding amount of the rail in the switch area and the
service performance of the switch is established, and the
long-term service of the switch under different grinding
schemes is realized. Performance predictions.

3.1. Vehicle turnout coupling dynamics model

Figure 7 Decision variables for critical sections In this section, SIMPACK software is used to establish a
The small top width section of the turnout is usually the vehicle-turnout multi-body dynamics model to analyze
limited area for rail grinding, and the 20mm top width the dynamic performance of the vehicle, and the lateral
section of the straight tip rail is selected as the starting vibration acceleration of the axle box is used as the
point for profile optimization. At the same time, the evaluation index for the stability of the vehicle at the
profile of the straight tip rail with a top width of 50mm is turnout. Input parameters. The vehicle-turnout coupling
no longer used as an active optimization object, but is dynamics model includes three parts: the vehicle model,
the turnout model and the wheel-rail contact model. The 𝑝𝑧 (𝑥, 𝑦)Δ𝑑(𝑥, 𝑦)
Δ𝑧(𝑥, 𝑦) = 𝑘𝑤 (𝑝𝑧 , 𝑠) (7)
vehicle model is a Chinese CRH2 car, which consists of 𝐻
1 car body, 2 bogies, 4 wheel-rails and suspension units. where pz(x,y) is the normal contact stress of the wheel and
composition, as shown in Fig. 8. Since the establishment rail in any unit area in the contact spot, and Δd(x,y) is the
of a deep neural network model requires a large number relative sliding distance of the wheel and rail when the
of samples, in order to shorten the calculation time, this wheel rolls through any unit of the contact spot.
paper regards the car body, bogie and wheelset as rigid In order to reduce the amount of calculation, it is
bodies, and each suspension unit is simulated by springs approximately considered that the state of the contact
or damping elements. In the model, the car body and spot does not change during the process of the wheel
bogie have 5 degrees of freedom of ups and downs, lateral rolling through the rail section. According to the method
movement, nodding, shaking his head, and rolling, and in the literature [11], the wear is superimposed, and the
the wheels have 4 degrees of freedom of ups and downs, worn rail profile is updated at the same time as the input
lateral movement, shaking his head, and rolling. of the next iterative calculation. Profile. According to this
The switch model is established according to the China process, the wear distribution of each section of the
Passenger Dedicated Line 07 (009) switch design turnout after a period of operation can be calculated.
standard (the radius of the guide curve is 1100 m, the
switch angle is 1:18), and the mobile elastic track module 3.3 Deep Neural Network Prediction Model
in SIMPACK is used, as shown in Fig. 9. By importing The deep feedforward network can be regarded as an
the profile sample of the turnout rail generated in Section efficient nonlinear function approximator. It is based on
2.2 into the model along a specific position in the the use of gradient descent to minimize the loss function.
longitudinal direction, the variable section features of the It can model the complex nonlinearity between input and
rail in the switch area can be considered. The wheel-rail output. Its basic structure is in the form of As shown in
contact model is the link between the coupling vehicle Figure 10, it consists of input layer, hidden layer, and
model and the switch model. The Hertzian contact theory output layer [12]. In this paper, the grinding amount of
and the FASTSIM algorithm are used to solve the normal the rail control point of the key section of the turnout is
and tangential problems respectively. selected as the sample feature, the lateral acceleration of
the axle box and the rail wear of the high-speed turnout
after a period of operation are used as the label, and the
deep neural network model of the above indicators is
constructed based on a large number of training samples.
By adjusting the appropriate hyperparameters, it can be
seen that the training error and verification error of the
neural network model continue to decrease as the number
of training increases until they converge. At this time,
Figure 8 The vehicle model in SIMPACK according to the input rail grinding amount, the predicted
value of each performance index of the turnout when the
total weight is 10Mt can be obtained. It can be seen from
Figure 11 that the neural network model constructed in
this paper has high accuracy, and its prediction results can
meet the optimization calculation requirements.

Figure 9 A moving elastic turnout model

3.2 Rail Wear Analysis Model

In this paper, the rail wear is calculated using the Archard
wear model, and the formula is as follows [10]:
𝐹𝑁 𝑑
𝑉𝑤 = 𝑘𝑤 (𝑝𝑧 , 𝑠) (6) Figure 10 Structure of deep neural network
𝐻
where Vw is the volume of material worn on the wheel-
rail contact spot, FN is the normal force of the wheel-rail
contact, d is the relative sliding distance of the wheel-rail
contact spot interface, and H is the material hardness,
assuming that the hardness of the wheel and rail materials
is the same. kw is the Archard wear coefficient related to
the normal contact stress pz and the wheel-rail relative
sliding velocity s, which can be obtained from the wear
coefficient distribution diagram, as shown in Fig. 10.
In order to obtain the distribution of the material wear Figure 11 MSE loss function
depth in the wheel-rail contact spot, the contact spot is
divided into multiple grids, and the calculation formula 4. Multi-objective optimization of rail profile in
of the (x, y) wear amount in any grid can be obtained as: switch area
The multi-objective optimization mathematical model in
this paper consists of decision variables, objective
functions and constraints, and the expressions are as
follows.
Decision variables. According to the requirements of the
on-site grinding operation, the grinding amount of the rail
should be less than 1.0 mm, so the value range of the
variable is:
0≤𝑿≤1 (8)
In the formula, X represents the grinding amount of the
key control points of each section of the turnout.
Objective function. Taking the prediction value of the
deep neural network model constructed in Section 3.3 as Figure 12 MSE loss function
the objective function, it mainly includes the lateral
acceleration amplitude of the axle box and the maximum
wear of the turnout rail when the total passing weight is
10Mt, which can be expressed as:
𝑚𝑖𝑛 𝑦1 (𝑿) = max{|𝑎𝑥 |}
(9)
𝑚𝑖𝑛 𝑦2 (𝑿) = max{𝑊𝑠 }𝐺𝑊=10 𝑀𝑡
In the formula, ax is the lateral acceleration of the vehicle
axle box, Ws is the rail wear in the switch area, and GW
is the total passing weight of the vehicle.
Restrictions. Taking the monotonicity and concave-
convexity of the rail profile as constraints, assuming that
the fitted profile curve is f(x), the first-order and second-
order derivatives of the straight tip rail, straight basic rail (a)
and curved basic rail profile can be expressed as follows:
𝑓 ′ (𝑥𝑠𝑤𝑖 ) ≥ 0, 𝑖 ∈ [2,14] 𝑠𝑤𝑖𝑡𝑐ℎ 𝑟𝑎𝑖𝑙
𝑓 ′ (𝑥 ) ≤ 0, 𝑖 ∈ [1,5]
{ ′ 𝑠𝑡𝑖 𝑠𝑡𝑜𝑐𝑘 𝑟𝑎𝑖𝑙
𝑓 (𝑥𝑠𝑡𝑖 ) ≥ 0, 𝑖 ∈ (5,9] (10)
″
𝑓 (𝑥𝑠𝑤𝑖 ) ≤ 0, 𝑖 ∈ [1,14] 𝑠𝑤𝑖𝑡𝑐ℎ 𝑟𝑎𝑖𝑙
𝑓 ″ (𝑥𝑠𝑡𝑖 ) ≤ 0, 𝑖 ∈ [1,9] 𝑠𝑡𝑜𝑐𝑘 𝑟𝑎𝑖𝑙
As an efficient evolutionary algorithm, NSGA-II obtains
offspring individuals through parameter evolution, and
evaluates its quality according to the fitness of each
solution. After multiple rounds of sorting and searching,
the optimal solution of the multi-objective optimization
problem is obtained. The algorithm is detailed The (b)
principle and process can be found in the literature [13],
which will not be repeated here. The computer
configuration and algorithm parameter settings are shown
in Table 1. Figure 12 shows the Pareto front surface
formed by convergence after multiple iterations. It can be
considered that a stable non-inferior solution has been
obtained at this time, and one set of results is extracted
and substituted into the switch space. In the profile
parameterization model, the optimized key section
profile of the turnout is obtained as shown in Figure 13.
Table 1 NSGA-II parameter settings.
Parameter Value
(c)
Population size 60

Crossover probability 0.9

Mutation probability 0.02

Maximum evolutionary algebra 60

 (b)
Figure 14 Wheel-rail dynamic interaction: (a)Wheel-rail vertical
force; (b)Wheel-rail lateral force
Stability. The acceleration of the axle box is selected as
the stability index when the vehicle crosses the fork, and
the results of the vertical acceleration of the axle box and
the lateral acceleration of the axle box are shown in Fig.
15(a) and (b), respectively. After the rail profile is
optimized, the maximum vertical acceleration of the axle
box is reduced from 13.47 m·s-2 to 8.74 m·s-2, a decrease
of 35.11%. The maximum lateral acceleration of the axle
box is reduced from 2.61 m·s-2 to 2.19 m·s-2, a decrease
(d)
of 16.09%. The optimized profile of the switch can
Figure 13 Comparison of measured profile and optimized profile: significantly improve the smoothness of the vehicle when
(a) Top width 20mm of straight switch rail;(b) Top width 50mm of passing the switch.
straight switch rail;(c) Straight stock rail;(d) Curved stock rail

5. Results and discussion

This section takes the case of the vehicle passing through
the switch area of the No. 18 high-speed switch in a
straight and reverse direction at a speed of 300km/h. The
dynamic performance of the turnout before and after rail
grinding and the rail wear under long-term service
conditions are calculated and analyzed to verify the
effectiveness of the optimized profile. (a)

5.1 Turnout Dynamics Performance

This section mainly evaluates the dynamic performance
of the vehicle laterally crossing the switch from the two
aspects of wheel-rail dynamic interaction and driving
stability. Since the first wheelset of the first bogie is most
severely impacted when the vehicle crosses the switch,
the dynamic responses in this section are the calculation
results of the guide wheelset unless otherwise specified.
Wheel-rail dynamic interaction. Fig. 14 shows the results (b)
Figure 15 Stationarity Indicator: (a)Vertical acceleration of
of the wheel-rail vertical force and the wheel-rail lateral
axle box; (b)Lateral acceleration of axle box
force when the vehicle passes through the switch area
before and after the profile optimization. It can be seen 5.2 Rail Wear Analysis
that the amplitudes of the vertical force and lateral force The Archard wear model is used to calculate the rail wear
of the wheel-rail are significantly reduced after the rail distribution in the switch area when the vehicle passes
profile is optimized. The maximum vertical force is through the switch. Fig. 16 shows the rail wear
reduced from 71.22 kN to 68.23 kN, and the maximum distribution when the four wheelsets of the vehicle roll
lateral force is reduced from 4.25 kN to 3.44 kN, a through the switch area one time in sequence. It can be
decrease of 19.06%. The optimized profile slows down seen from the calculation results that when the vehicle
the dynamic interaction between wheel and rail. passes through a turnout, the maximum wear of the
measured profile is 1.04 × 10-5 mm, and the maximum
wear of the optimized profile is 1.42 × 10 -4 mm, and the
section where the maximum wear is located slightly
moves forward. Although the wear amplitude increases
after the profile optimization, the wear of the remaining
sections of the turnout is obviously smaller than the
measured profile. And this result can only reflect the wear
behaviour of turnouts in the short term. In order to
explore the evolution law of rail wear under long-term
(a)
service conditions of turnouts, the method described in
Section 3.2 is used for iterative calculation. Considering
the time cost, calculation stops when the total passing
weight reaches 20 Mt.
 box in different operation stages is smaller than the
measured profile, which shows that the optimized profile
can ensure a smaller wheel. While the rail contact damage
is improved, the stability of the vehicle crossing the fork
under long-term service conditions is improved.

(a)

Figure 18 Maximum accumulated wear of turnout

(b)
Figure 16 Distribution of wear in the switch panel: (a)Wear
of measured profile; (b) Wear of optimized profile
Through the calculation of the wear distribution of the
turnout when the total weight reaches 20 Mt, it can be
seen that the maximum wear section of the measured
turnout occurs at the position of the top width of the tip
Figure 19 Axle box lateral acceleration amplitude
rail with a width of 37.9 mm, and the maximum wear of
the optimized switch occurs at the position of the top
width of the tip rail with a width of 33.8 mm. Fig. 17 6. Conclusions
shows the wear evolution law of the measured profile and This paper proposes a multi-objective optimization
the optimized profile, respectively. It can be seen from method for the profile of the rail combination considering
the calculation results that the maximum cumulative wear the long-term service performance of the turnout. Taking
of the optimized profile is 0.273 mm, which is smaller the measured profile of the rail in the switch area of a
than the measured cumulative wear of 0.304 mm, group of No. 18 high-speed turnouts at Linfen West
indicating that the optimized profile is beneficial to Station as the research object, the grinding amount of the
prolong the service life of the weak rail. control point of the key section is selected as the design
variable. Taking the lateral acceleration of the axle box
and the accumulative wear of the rail as the objective
functions, a deep neural network model of the service
performance of the turnout is constructed. Combined
with the NSGA-II algorithm, the multi-objective
optimization problem is solved, and the target profile of
the turnout can be obtained, which can take into account
(a) the stability of driving and the reduction of wheel and rail
wear. The relevant conclusions of this paper are as
follows:
1. After the rail profile is optimized, the maximum
vertical acceleration of the axle box is reduced from
13.47 m·s-2 to 8.74 m·s-2, a decrease of 35.11%, and
the maximum lateral acceleration of the axle box is
(b) reduced from 2.61 m·s-2 to 2.19 m·s-2, a reduction of
Figure 17 The change law of wear with the total passing 16.09%. The optimized profile significantly
weight: (a)37.9 mm top width of measured; (b)33.8 mm top width improves the stability of the vehicle when passing
of optimized
the switch.
The variation law of the maximum wear of the switch
2. When the vehicle passes through a turnout, the
with the total passing weight is shown in Fig. 18. It can
maximum wear of the measured profile is 1.04×10 -5
be seen that although the initial wear of the optimized
mm, and the maximum wear of the optimized profile
profile is greater than the measured profile, with the
is 1.42×10-4 mm. However, with the increase of the
increase of the total passing weight, the cumulative wear
total passing weight, the cumulative maximum wear
of the optimized profile is eventually smaller than the
of the optimized profile is 0.273mm, which is less
measured profile. The cumulative wear of the shape
than the cumulative wear of the measured profile,
shows that the optimized profile can effectively delay the
which is 0.304mm. It shows that the optimized
wear rate of the switch. Fig. 19 shows the variation law
profile can effectively delay the wear rate of the
of the lateral acceleration amplitude of the axle box with
switch, which is beneficial to prolong the service life
the total passing weight. After the turnout profile is
of the weak rail.
optimized, the lateral acceleration amplitude of the axle
3. The amplitude of the lateral acceleration of the axle
box with the optimized profile is smaller than that of
the measured profile in different operation stages,
which indicates that the optimized profile can
improve the stability of vehicle crossing fork under
long-term service conditions while ensuring less
wheel-rail contact damage.
Combining the deep neural network model and the
intelligent evolutionary algorithm, this paper proposes a
design method for the profile of the turnout grinding
target, and obtains ideal results. Compared with the
existing research, this method incorporates the long-term
service performance of the turnout into the optimization
objective, which can scientifically guide the maintenance
and repair of the in-service high-speed turnout, and
prolong the service life of the turnout while maintaining
the stability of the train crossing the turnout.