CN104898414A - Key parameter identification method for high-speed train dynamics performance design - Google Patents
Key parameter identification method for high-speed train dynamics performance design Download PDFInfo
- Publication number
- CN104898414A CN104898414A CN201510145441.5A CN201510145441A CN104898414A CN 104898414 A CN104898414 A CN 104898414A CN 201510145441 A CN201510145441 A CN 201510145441A CN 104898414 A CN104898414 A CN 104898414A
- Authority
- CN
- China
- Prior art keywords
- design
- high speed
- dynamics
- parameter
- performance
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Landscapes
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention provides a key parameter identification method for high-speed train dynamics performance design, which relates to the field of dynamics simulation design and analysis for the high-speed trains. A single-output LM neural network agent model is effectively used, and overall design of the high-speed train is merged in multi-subject global simulation design. The method comprises the following steps: a high-speed train multi-rigid-body dynamics physical model and a simulation model are firstly used, high-speed train dynamics performance input and output design space is determined, expert domain knowledge is used, and high-speed train dynamics performance design parameters and response indexes are extracted to shorten the design space; then, LM algorithm is adopted to adjust weights and thresholds of the neural network so as to improve the convergence speed and the convergence accuracy of the single-output neural network, and according to a sensitivity formula of input parameters in relative to an output value in the neural network, sensitivity is calculated; and finally, sensitivity analysis and key parameter identification are carried out on high-speed train design parameters. The method of the invention is mainly used for high-speed train dynamics analysis and design.
Description
Technical field
The invention belongs to rolling stock design field, particularly the Parametric Design Analysis of bullet train on dynamics simulation basis.
Background technology
Bullet train from one-piece construction be one by machinery, control, electrically, communication etc. form, is by the complex electromechanical systems of the multiple mechanical behavior decision such as dynamics, fatigue strength, acoustics, aerodynamics from physical property.Wherein the index such as the stationarity of influential system, security and comfortableness is main or determined by the mechanical property of mechanical system, emphasis shows on the overall performance-high speed train dynamics of bullet train, and its design variable collection be made up of parameters such as the structure of bullet train parts and performances determines.Realize the design process of bullet train complexity, need various CAE digital modeling and the simulation analysis such as experience high speed train dynamics, structural strength, mode of oscillation, noise, fatigue, collision, acoustics, aerodynamics, heating ventilation, and the multidomain modeling and simulation that collaborative formation is huge mutually, by solving huge mechanic-mathematical model, and seek suitable optimization method to obtain rational design parameter.Especially when bullet train operation reaches the speed of more than 300 kilometers/hour, the effect of itself and running environment is also strengthened, the performance of bullet train is by itself extend to running environment system, thus bullet train is needed to be placed in running environment, comprise Pantograph-OCS system from above, comprise wheel track and line condition downwards, comprise fluid-structure coupling system etc. forward, set up a more complicated Coupled Dynamics system-bullet train system dynamics, make the design parameter optimized more reasonable.
The various fields of bullet train system, the modeling and simulation method comprising bullet train coupled system all has breakthrough, but the complexity that constructed system is suitable, have that nonlinearity, difficulty in computation are large, computing time and resources occupation amount large, be unfavorable for the difficult problem that finds rational optimization solution etc. many at problem space.Only with regard to bullet train Dynamic performance Optimization itself, needing, by considering Wheel-rail contact model, to set up the subordination principle based on many-body dynamics, finding the sensitivity relation between design parameter and output on this basis, then adopting Optimization Solution.I.e. advanced line sensitivity analysis, study the influence degree of each design parameter to target response index, remove the very little design variable parameter of those influence degrees according to certain rule again, the Optimized model of the little dimension variable be simplified, can reduce and solve difficulty.
In the research of bullet train, what river in Guangxi proposes " the iron nitride films modeling of Design-Oriented and sensitivity analysis ", [Chengdu: Southwest Jiaotong University's PhD dissertation, 2012] in literary composition, the typical spring damping power of wheel rail force is described, by carrying out based on primitive sensitivity analysis feature modeling to spring-damping force, obtain the Sensitivity Analysis of wheel rail force, achieve the numerical evaluation work of obtained model.In fact, in practical implementation, sensitivity analysis widely for solving Optimal Structure Designing problem, but due to the required system of equations amount solved large and complicated, be rarely used in many-body dynamics, multidisciplinary field.In this process, need to calculate a large amount of non-linear differential movement difference equations, and this difficulty is mathematically well-known.
The identification of the research of German on neural network agent model to Dynamics Performance of High Speed Trains key parameter opens another method [Computer-aided design sensitivity analysis for dynamic multibody systems [D] .Iowa:PhD Thesis, The University of Iowa, 2006], Eberhard is to the research of LM algorithm, solve again generalization ability and Generalization accuracy [Analysis and Optimization of Complex Multibody Systems Using Aadvanced Sensitivity Analysis Methods [D] the .Stuttgart:PhD Thesis of University of Stuttgart of neural network agent model to a certain extent, 1997], make key parameter identification more reliable, avoid the process solving large amount of complex system of equations simultaneously.Obviously, complicated multi-body system equation is set up in contrast, and this key parameter recognition methods has the advantages that efficiency is high, real-time, cost is low.The key content of research includes: the determination of bullet train rigid multibody dynamics design space and reduction technology; Based on single goal neural network and the sensitivity of sensitivity gradient formulae discovery; Key parameter criterion of identification.
Summary of the invention
The object of this invention is to provide the key parameter recognition methods of a kind of Dynamics Performance of High Speed Trains design, it can effectively utilize single goal agent model, bullet train overall design is dissolved in the design of multidisciplinary global simulation, thus under multidisciplinary field conditions, set up the Sensitivity Analysis Method of High-Speed Train Design parameter, identify key parameter required in design, thus improve Dynamics Performance of High Speed Trains.
The object of the invention is to be achieved through the following technical solutions:
A kind of Dynamics Performance of High Speed Trains design key parameter identification method, utilize and improve single goal LM neural network agent model, set up the Sensitivity Analysis Method of Dynamics Performance of High Speed Trains design parameter, detailed process is as follows:
First, utilize bullet train rigid multibody dynamics physical model and realistic model, determine Dynamics Performance of High Speed Trains input and output design space, and utilize expert's domain knowledge, extraction Dynamics Performance of High Speed Trains design parameter and response index reduce design space, to reduce scale and the capacity of neural network, comprising:
1) based on bullet train structural topologies, extract the physical model parts involved by high speed train dynamics analytical model, and according to the expression way of each parts in analogue system, obtain the abstract form involved by simulation analysis model;
2) according to physical model parts and emulation abstract form, the whole design spaces involved by the input and output of Dynamics Performance of High Speed Trains design parameter are determined;
3) seven the output performance indexs affecting stationarity, security and curve negotiation ability are extracted according to High-Speed Train Design standard;
4) for Dynamics Performance of High Speed Trains design parameter input variable, draft and provide expert investigation table, utilizing expert's domain knowledge to carry out the reduction of design variable and the sampler space to Dynamics Performance of High Speed Trains design space;
Secondly, adopt the weights and threshold of LM algorithm to single output nerve network to adjust, to improve speed of convergence and convergence precision, according to the sensitivity formula of input parameter in neural network relative to output valve, calculate sensitivity, comprising:
1) adopt Latin hypercube sampling method, obtain bullet train vehicle dynamics simulation standardization sample;
2) according to the span of Dynamics Performance of High Speed Trains design parameter variable, standardization sample is converted into actual design parameter sample;
3) determining design condition, in simulation software, set up realistic model, after simulation calculation, obtain performance parameter sample, and be converted into standardization numerical value, obtaining training and precision test sample point, for building three layers of Feedback Neural Network agent model;
4) adopt three layers of Feedback Neural Network training sample based on LM algorithm, the list obtaining bullet train exports LM neural network agent model;
5) whether the precision of inspection sheet output LM neural network agent model reaches requirement, and in this way, then determine structure and the scale of the neural network improved, modeling terminates; As no, then adjust the neural network structure of improvement and scale or reconstructed sample point, re-training, iterative cycles, till reaching requirement;
6) utilize the list of this improvement to export LM neural network agent model, in conjunction with Calculation of Sensitivity formula, carry out the Calculation of Sensitivity of Dynamics Performance of High Speed Trains design parameter;
Finally, carry out sensitivity analysis and the key parameter identification of Dynamics Performance of High Speed Trains design parameter, comprising:
1) relative sensitivity is calculated, so that sensitivity for analysis result of calculation;
2) draft key parameter recognition rule, set up key parameter recommendation tables, enter to identify to key parameter during High-Speed Train Design.
Describedly utilize expert's domain knowledge, adopt Delphi (Delphi) method, draft importance degree and the span questionnaire of Dynamics Performance of High Speed Trains design parameter, send to expert of the art, require that brainstrust is based on experience and domain knowledge, provides the span of each design variable in table and the importance degree to performance.
The importance degree span of described Dynamics Performance of High Speed Trains design parameter is between 0 ~ 1, value is more close to 1, then larger on the impact of the Dynamics properties of train index that system exports, extract design parameter that importance degree is greater than 0.5 as important input variable corresponding to Dynamics Performance of High Speed Trains, reduction dimension.
Described three layers of Feedback Neural Network, refer to that input layer adopts non-linear transfer to the neuron of hidden layer, hidden layer to output layer then apply Multiple Linear Regression model.
Described employing LM algorithm is used to the weights and threshold controlling single output nerve network, to improve the generalization ability of neural network.
The neuronic number of described input layer and output layer, be arranged at the variable number 1 of Dynamics Performance of High Speed Trains design parameter variables number n and one of them output-index parameter respectively, the interstitial content of hidden layer is defined as m, according to above principle, set up n-m-1 mono-output LM neural network agent model that seven output-indexes are corresponding respectively.
Described Calculation of Sensitivity formula is according to the computing formula of input parameter relative to the disturbance degree of output performance index, obtain the sensitivity number of input parameter relative to output performance index, and convert it into relative sensitivity, with to its further analysis of key parameter.
Described key parameter recognition rule is greater than the first five items of 65% for standard with relative sensitivity absolute value.
Adopt this method, network input parameter in bullet train agent model technology can be effectively utilized, export the Sensitivity Analysis Method between response and weights, carry out the key parameter identification based on high speed train dynamics emulation, solve the bullet train analogue system related in multidisciplinary field when carrying out train dynamics analysis, design analysis is caused to be difficult to the problem of carrying out smoothly because huge design parameter need be processed, this provides new method for designing and thinking for bullet train with the design and analysis of similar Complex Mechatronic Products, there is great guidance value.
The present invention's advantage compared with prior art and effect:
One, solving of complicated differential equation group is avoided
Existing research mainly adopts the mode setting up differential equation group to carry out sensitivity analysis, but due to the complicacy of design space, to derive and in the process that solves in differential equation group, difficulty still very greatly.The identification of research to bullet train key parameter of agent model opens another method, and this method avoids the process solving large amount of complex system of equations, has the advantages that efficiency is high, real-time, cost is low.
Two, the present invention is based on the analysis to high speed train dynamics combination property
Be applied to the agent model technology that Dynamics Performance of High Speed Trains is analyzed at present, often be limited to certain aspect of analytical performance or the problem of certain several aspect, and the present invention is based on the research of bullet train overall performance, relate to bullet train safety in operation, stationarity and comfortableness almost over-all properties index.Namely in order to meet the requirement of these dynamic performances, except abrasion index, other lateral stability index, vertical riding index, derailment coefficients, rate of wheel load reduction, wheel shaft transverse force, Overturning Coefficient and critical velocity seven performances, all in studied analyst coverage, more can embody its engineering practical value.
Accompanying drawing explanation
Fig. 1 process flow diagram of the present invention
The bullet train vehicle system dynamics performance index corresponding relation that Fig. 2 is involved in the present invention
The mono-output nerve network structure of Fig. 3 n-m-1 of the present invention
Embodiment
Below in conjunction with accompanying drawing, the present invention is described further:
1. the determination of Dynamics Performance of High Speed Trains design space and reduction
According to the structural topologies between all parts, realistic model is by car body, bogie (framework, axle box and wheel to) and one be, secondary suspension power unit forms high speed train dynamics analytical model.In simulation process, physical model parts are stated according to following rule: car body, framework, axle box and wheel are to the body be abstracted in multi-body system, each spring force adopts three-dimensional component force unit to express, the power unit of vibration damper adopts serial spring-damping force unit to describe, axle box and wheel between rotation adopt rotary gemel to retrain, thus obtain high speed train dynamics and analyze the abstract form of realistic model and often save the number that car relates to, in table 1.
Table 1 physical model parts and analogue system abstract form
| Physical system describes (often saving car number) | Analogue system describes (number is corresponding) | Explanation |
| Car body (1) | Car body (body) (1) | Body in multi-body system |
| Framework (2) | Framework (body) (2) | Body in multi-body system |
| Axle box (8) | Axle box (body) (8) | Body in multi-body system |
| Wheel is to (4) | Wheel is to (body) (4) | Body in multi-body system |
| One is journal box spring (8) | Three-dimensional component force unit (No. 05 power unit) (8) | Power unit in multi-body system |
| One is vertical damper (8) | Serial spring-damping force unit (No. 06 power unit) (8) | Power unit in multi-body system |
| One is pivoted arm node (8) | One is pivoted arm power unit (No. 05 power unit) (8) | Power unit in multi-body system |
| Two is air spring (4) | Three-dimensional component force unit (No. 05 power unit) (4) | Power unit in multi-body system |
| Two is vertical damper (4) | Serial spring-damping force unit (No. 06 power unit) (4) | Power unit in multi-body system |
| Two is lateral damper (4) | Serial spring-damping force unit (No. 06 power unit) (4) | Power unit in multi-body system |
| Anti-hunting damper holder (4) | Serial spring-damping force unit (No. 06 power unit) (4) | Power unit in multi-body system |
| Drawing pull bar spring (2) | Three-dimensional component force unit (No. 05 power unit) (2) | Power unit in multi-body system |
| Horizontal backstop (4) | Serial spring-damping force unit (No. 06 power unit) (4) | Power unit in multi-body system |
| Axle box and wheel between rotation (8) | Rotary hinge (8) | Hinge in multi-body system |
| Anti-side rolling torsion rod (2) | Power unit and moment (2) | Power unit in multi-body system and moment |
The physics analyzing out according to the Multi-body dynamic model of bullet train and emulation abstract model (see table 1), can extract the whole input variable of bullet train when simulation modeling and output variable.In SIMPACK analogue system, the input variable extracted relates to: the variablees such as travelling speed, simulation time, integration step, integral way, bullet train structure and performance parameters, and wherein bullet train parameter contains vehicle basic parameter, quality/moment of inertia, center of gravity, single stage suspension/secondary suspension/horizontal backstop/anti-side rolling torsion rod/structural parameters, single stage suspension/secondary suspension/drawing pull bar/parameter such as anti-side rolling torsion rod suspension parameter and nonlinear characteristic; The output variable extracted relates to: the parameters such as translation displacements, rotation displacement, speed of moving body, speed of moving body, body acceleration of motion, wheel-rail contact force, Bearing pattern, wheel shaft transverse force, constraining force, power unit and tie point displacement.The design space that these inputs, output variable form bullet train often reaches up to a hundred variablees respectively, and obviously involved variable number is huge, needs the pre-service carrying out design space, namely reduces design space.
In analogue system, output variable is many especially, but but can not directly be used for evaluating the runnability of bullet train.When not considering wear hardness, according to dynamic performance: safety in operation, the requirement of curve negotiation ability and steady comfortableness three aspects, in up to a hundred output parameters of analogue system, find out each self-corresponding simulation data item of performance index, calculate further according to " iron nitride films assessment of performance and test for identification specification " (GB5599-85) and integrate and can obtain seven dynamics output performance indexs and value thereof: lateral stability index, vertical riding index, derailment coefficients, rate of wheel load reduction, wheel shaft transverse force, Overturning Coefficient and critical velocity.
Find out that the variable number related to is huge according to the quantitative analysis of In-put design change above, when adopting any one model to act on behalf of, all there is the feature of higher-dimension high non-linearity, when modeling, bring unnecessary degree of difficulty.Assessing the cost to reduce, shortening computing time, adopt Delphi (Delphi) method for this reason, draft a series of parameter importance degree and span questionnaire, send to domain expert, select variable according to the priori of expert, effectively reduce the dimension of input variable.In whole questionnaires, for the design variable of bullet train in SIMPACK needed for modeling and simulating, correspond to each vehicle original reference data, require that brainstrust is based on experience and domain knowledge, provide the span of each design variable and the importance degree to performance.Importance degree span is between 0 ~ 1, and value is more close to 1, then larger on the impact of the Dynamics properties of train index that system exports, thus also larger on the impact of security during train operation, stationarity, comfortableness and curve negotiation ability.Like this based on the knowledge of specialist field, and consider unstability situation, extract design parameter that importance degree is greater than 0.5 as important input variable corresponding to bullet train, determine further according to experiment and reduce span.By this principle, finally obtain the input variable collection of reduction.
2. build and improve agent model and meter sensitivity
Choose Latin hypercube function, build bullet train vehicle dynamics simulation sample design.In order to construct neural network agent model, necessary first contrived experiment method, obtaining sampling collection could set up and assess neural network.Therefore, the foundation of simulation calculation sample design is based on the important design parameter of initial selected and seven performance index, consider High-Speed Train Design spatial complex, spatially be evenly distributed rationally for making sample point and easily adjust, adopt the Latin hypercube Sampling Strategy that sample point quantity is the amount doesn't matter, dirigibility is high, obtain bullet train vehicle dynamics simulation standardization sample.
According to normalized design parameter sample, according to the span of design variable, standardization sample is converted into actual parameter value, obtains actual design parameter sample.
Determine design condition, consider the problem such as space geometry, nonlinear element, Wheel Rail Contact geometric relationship spatially of track irregularity, circuit, realistic model is set up in SIMPACK v8.904 software, the actual parameter sample of In-put design variable, after emulation, calculate seven output-index values, obtain output parameter sample, be converted into standardization numerical value again, finally like this obtain training and the normalization design parameter of precision test and the sample point of output parameter.
Adopt three layers of Feedback Neural Network, each network is made up of an input layer, a hidden layer and an output layer.The neuron arranged from input layer to hidden layer adopts non-linear transfer, and applies Multiple Linear Regression model on output layer.The neuronic number of input layer and output layer, be arranged at design parameter variables number n and index parameter variables number 1 respectively, the nodes of hidden layer is defined as m.The stationarity of this model train, comfortableness, security, seven performance index groups that difference is corresponding: lateral stability, vertical stationarity, derailment coefficients, rate of wheel load reduction, wheel shaft transverse force, Overturning Coefficient and critical velocity.
Adopt Levenberg-Marguart algorithm, namely the weights and threshold of LM algorithm to network adjusts, to improve the generalization ability of neural network.The important part of this algorithm adds adjustment item in error sum of squares formula, a kind of between gauss-newton method (Guass-Newton, the optimization method of the lsqnonlin GN) and between gradient method, not only avoid in gradient descent method and start to search for the shortcoming that but fast last convergence is slow and precision is low, the adjustment item added has processed the phenomenon of the unusual and false convergence of Jacobian matrix in GN, ensure that each adjustment of weights and threshold all makes error reduce simultaneously, ensure that the stability of network.The neural network agent model exported for only having one, because its structure is simple, scale capacity is relatively little, adopt the improvement neural net method based on LM algorithm to build, its speed and precision can reach requirement.In MATLAB software, set up that the n-m-1 corresponding with seven output-index groups is mono-exports LM neural network agent model respectively.
Whether the precision that inspection sheet exports LM neural network agent model reaches requirement, and in this way, then determine structure and the scale of the neural network of this improvement, modeling terminates; As no, then adjust the neural network structure of this improvement and scale or reconstructed sample point, re-training, iterative cycles, until reach requirement to carry out subsequent calculations again.
The value of sensitivity can be calculated according to computing formula:
Wherein S
ikinput (variable) X
i(1≤i≤n) is at single output neuron (performance response index) Y
kthe disturbance degree of (1≤k≤7), i.e. Sensitirity va1ue, W
ijthe weight matrix of i-th input layer to jth (1≤j≤m, m is hidden layer neuron number) hidden layer neuron, w
jkthat a jth hidden layer neuron is to the neuronic weight matrix of the k output layer.The sensitivity absolute value calculated | S
ik| larger correlativity is stronger, | S
ik| less correlativity is less.Negative value represents negative correlation.
3. sensitivity analysis and key parameter identification
Obtain weight matrix according to neural network and bring formula 1 into, after calculating, obtain the Sensitirity va1ue of n input variable relative to seven performance output valves.Again according to the following formula, relative sensitivity S ' is calculated
ik:
S'
ik＝S
ik/S
ikmax2
S in formula
ikmax---sensitivity maximum in table.
Key parameter recognition rule is as follows: be greater than the first five items of 65% with relative sensitivity absolute value for standard, sets up each single performance key parameter recommendation tables effective, for identifying bullet train key parameter.Design recommendation is provided to instruct bullet train subsequent design manufacture work: according to recommendation tables, global design key parameter can be obtained, greatly reduce major design variable, carry out dimensionality reduction to design space according to key parameter recognition rule; According to this recommendation tables, the key parameter involved by different response indexs can be found out, can regulate according to sensitivity when actual design, to reach optimal effectiveness, as strengthened weighted value according to sensitivity in design optimization, the then weighted value that namely sensitivity is larger is larger, and some parameter has significant impact to many index for another example, even influence degree is contrary, needs emphasis to consider.
According to the abstractdesription method of the key parameter identification of above-mentioned Dynamics Performance of High Speed Trains design, the key parameter recognition methods that certain Dynamics Performance of High Speed Trains designs specifically is implemented as follows:
1. design space is determined and is reduced
According to the topological structure of certain type car, extract the physical model parts involved by its dynamic analysis model, and according to the expression way of power unit in analogue system, obtain the abstract form involved by simulation analysis model, again according to physical model parts and emulation abstract form, determine the whole design spaces involved by input and output of this model bullet train, according to standard and regulation extract affect stationarity, security and curve negotiation ability seven output-indexes in table 2.And for design input variable, draft and provide expert investigation table, and based on expert's domain knowledge, High-Speed Train Design space is reduced, the design variable after reduction is in table 3.
Table 2 dynamic performance index
Table 3 CRH type car input variable and span
2. build agent model and meter sensitivity
The minimum sample number of Latin hypercube should be generally more than three times of variable number, considers the phenomenon of unstability, with reference in the former design load of CRH type car and the basis of domain experts' suggestion, reduces design variable span (see table 3).Last in span, determine corresponding Sampling Strategy: each parametric variable is divided into 100 levels, Latin hypercube sample design method is adopted to generate the design sample of Dynamics Performance of High Speed Trains design parameter, generate 29*100 standardization sample parameter value, be converted to corresponding actual design variable parameter value.Have 3 groups of data unstabilitys after carrying out Computer Simulation according to actual design variable parameter value, the level of remaining 97 successful operations, design parameter sample becomes 29*97 group.Again in normalized space, each parametric variable is divided into 6 levels, generates 29*6 standardization sample parameter value.For 97+6 group normal value, whole design parameter samples is formed after being converted to corresponding actual design variable parameter value, wherein 97 design parameter samples are used for setting up neural network agent model, other 6 design parameter samples as checking agent model precision.
According to physics and the emulation abstract form of bullet train, in SIMPACK v8.904, set up realistic model.The line condition that design condition is chosen is: choose 300km/h as computing velocity.Line condition defines: the total length of circuit is 1.5 kilometers, has the segmentation that straight-line segment, mild wet air oxidation and circular curve three kinds is different.Wherein straight-line segment is divided into two sections, and length is respectively 500m and 270m; Adjustment curve is also divided into two sections, and length is all 290m; Circular curve length is 150m, and track superelevation is taken as 102.6mm.Incentive condition: the Beijing-Tianjin track spectrum of actual measurement.Tread profile: LAM.Sample data embeds realistic model, can obtain the original output response of constructing needed for agent model, after these values are carried out analytical calculation process, changes into and exports response index sample, finally obtain 97+6 group input and output sample point.
A neural network network structure, is optimized neural network structure based on LM algorithm, namely adjusts the weights and threshold of network, to improve speed of convergence and the Generalization accuracy of network, realize the training of the neural network of 97 groups of input and output values.The comfortable stationarity of this model train, security and curve negotiating rate, corresponding respectively seven performance index groups: lateral stability, vertical stationarity, derailment coefficients, rate of wheel load reduction, wheel shaft transverse force, Overturning Coefficient and critical velocity.In MATLAB software, establish single output agent model of the corresponding improvement neural network based on design space pre-service-LM optimized algorithm respectively for seven performance index.
For lateral stability, by constantly debugging, hidden layer neuron number is taken as 12, sets up the BP neural network of 29-12-1.By debugging, other index: vertical stationarity, derailment coefficients, rate of wheel load reduction, wheel shaft transverse force, Overturning Coefficient and critical velocity all adopt the BP neural network of 29-10-1, write code, obtain correlation coefficient r 1, r2 through training ... r7 result all relatively 1.
By calculating, the error amount of the response that 97 groups of experiment values and agent model draw is all smaller.Maximum relative error as lateral stability is 0.18%, the maximum relative error of vertical stationarity is 0.22%, the maximum relative error of derailment coefficients is 1.50%, the maximum relative error of rate of wheel load reduction is 0.30%, the maximum relative error of wheel shaft transverse force is 0.11%, the maximum relative error of Overturning Coefficient is 4.72%, and the maximum relative error of critical velocity is 0.32%.
Six groups are not participated in and sets up agent model and the comparing of experiment value that each response of directly being obtained by agent model obtains with SIMPACK modeling, lateral stability maximum relative error is 6.09%, vertical stationarity maximum relative error is 5.88%, derailment coefficients maximum relative error is 7.88%, rate of wheel load reduction is 8.56%, wheel shaft transverse force is 9.38%, Overturning Coefficient is 9.61%, critical velocity is 4.27%, the error of visible seven agent models is all be less than 10%, simultaneously all average relative errors are less than 6%, so precision is enough.
According to formula 1, calculate 29 input variables to the sensitivity of seven performance index.
3. key parameter identification
29 input variables are calculated to the relative sensitivity of seven performance index according to formula 2.
The first five items of 65% is greater than for standard with relative sensitivity absolute value, set up the key parameter recommendation tables of effective influence dynamic performance index in table 4, before what wherein correlativity was large come, the design parameter meeting the vertical stationarity of this condition only has one, add one again and be convenient to Design coordination, band * is negative correlation, and other is positive correlation.
Table 4 bullet train key parameter recommendation tables
Data in analytical table, can identify key parameter, and what and guide later bullet train manufactures and designs work further:
Global design key parameter is for having: X9, X10, X11, X12, X14, X15, X18, X20, X21, X23, X24, X25, X26, X27, X28, X29.These parameters correspond to respectively: wheel diameter (rolling circle nominal diameter), distance inside wheel, wheel is to quality, sidewinder moment of inertia-X, shake the head moment of inertia-Z, one is coil spring longitudinal rigidity (every axle box), one is vertical damping (every axle box), axle box swivel arm node longitudinal rigidity (every axle box), axial pivoted arm node lateral stiffness (every axle box), air spring longitudinal rigidity (every spring), air spring lateral stiffness (every spring), air spring vertical stiffness (every spring), two is vertical damping, two is horizontal damping and anti-hunting damper holder connection stiffness (each).So far, Dynamics Performance of High Speed Trains design variable identifies 16 key parameters from 29, greatly reduces the dimension of design space.
In 16 key parameters identified, except X9, X10 are parameter of structure design, other is all performance design parameter.This illustrates that the performance design parameter of input and parameter of structure design have impact to the stationarity of bullet train, comfortableness and security, but performance parameter is wider on the impact of bullet train, and this and practical experience are substantially consistent.
According to this recommendation tables, identify for the key parameter involved by different response indexs, when actual design, key parameter values can have been regulated according to sensitivity, to reach optimal effectiveness.As strengthened weighted value according to sensitivity in design optimization, namely sensitivity is larger, and weighted value is larger.And for X18, X20, X25 and X28, corresponding one be vertical damping (every axle box) respectively, axle box swivel arm node longitudinal rigidity (every axle box), air spring lateral stiffness (every spring) and two be horizontal damping, these design parameters have considerable influence to two in response index, more want emphasis to consider during optimal design.
Claims (8)
1. a Dynamics Performance of High Speed Trains design key parameter identification method, utilizes and improves single goal LM neural network agent model, set up the Sensitivity Analysis Method of Dynamics Performance of High Speed Trains design parameter, it is characterized in that:
First, utilize bullet train rigid multibody dynamics physical model and realistic model, determine Dynamics Performance of High Speed Trains input and output design space, and utilize expert's domain knowledge, extraction Dynamics Performance of High Speed Trains design parameter and response index reduce design space, to reduce scale and the capacity of neural network, comprising:
1) based on bullet train structural topologies, extract the physical model parts involved by high speed train dynamics analytical model, and according to the expression way of each parts in analogue system, obtain the abstract form involved by simulation analysis model;
2) according to physical model parts and emulation abstract form, the whole design spaces involved by the input and output of Dynamics Performance of High Speed Trains design parameter are determined;
3) seven the output performance indexs affecting stationarity, security and curve negotiation ability are extracted according to High-Speed Train Design standard;
4) for Dynamics Performance of High Speed Trains design parameter input variable, draft and provide expert investigation table, utilizing expert's domain knowledge to carry out the reduction of design variable and the sampler space to Dynamics Performance of High Speed Trains design space;
Secondly, adopt the weights and threshold of LM algorithm to single output nerve network to adjust, to improve speed of convergence and convergence precision, according to the sensitivity formula of input parameter in neural network relative to output valve, calculate sensitivity, comprising:
1) adopt Latin hypercube sampling method, obtain bullet train vehicle dynamics simulation standardization sample;
2) according to the span of Dynamics Performance of High Speed Trains design parameter variable, standardization sample is converted into actual design parameter sample;
3) determining design condition, in simulation software, set up realistic model, after simulation calculation, obtain performance parameter sample, and be converted into standardization numerical value, obtaining training and precision test sample point, for building three layers of Feedback Neural Network agent model;
4) adopt three layers of Feedback Neural Network training sample based on LM algorithm, the list obtaining bullet train exports LM neural network agent model;
5) whether the precision of inspection sheet output LM neural network agent model reaches requirement, and in this way, then determine structure and the scale of the neural network improved, modeling terminates; As no, then adjust the neural network structure of improvement and scale or reconstructed sample point, re-training, iterative cycles, till reaching requirement;
6) utilize the list of this improvement to export LM neural network agent model, in conjunction with Calculation of Sensitivity formula, carry out the Calculation of Sensitivity of Dynamics Performance of High Speed Trains design parameter;
Finally, carry out sensitivity analysis and the key parameter identification of Dynamics Performance of High Speed Trains design parameter, comprising:
1) relative sensitivity is calculated, so that sensitivity for analysis result of calculation;
2) draft key parameter recognition rule, set up key parameter recommendation tables, enter to identify to key parameter during High-Speed Train Design.
2. a kind of Dynamics Performance of High Speed Trains design key parameter identification method according to claim 1, it is characterized in that: describedly utilize expert's domain knowledge, adopt Delphi (Delphi) method, draft importance degree and the span questionnaire of Dynamics Performance of High Speed Trains design parameter, send to expert of the art, require that brainstrust is based on experience and domain knowledge, provides the span of each design variable in table and the importance degree to performance.
3. a kind of Dynamics Performance of High Speed Trains design key parameter identification method according to claim 2, it is characterized in that: the importance degree span of described Dynamics Performance of High Speed Trains design parameter is between 0 ~ 1, value is more close to 1, then larger on the impact of the Dynamics properties of train index that system exports, extract design parameter that importance degree is greater than 0.5 as important input variable corresponding to Dynamics Performance of High Speed Trains, reduction dimension.
4. a kind of Dynamics Performance of High Speed Trains design key parameter identification method according to claim 1, it is characterized in that: described three layers of Feedback Neural Network, refer to that input layer adopts non-linear transfer to the neuron of hidden layer, hidden layer to output layer then applies Multiple Linear Regression model.
5. a kind of Dynamics Performance of High Speed Trains design key parameter identification method according to claim 1, is characterized in that: described employing LM algorithm is used to control the weights and threshold of single output nerve network, to improve the generalization ability of neural network.
6. a kind of Dynamics Performance of High Speed Trains design key parameter identification method according to claim 1, it is characterized in that: the neuronic number of described input layer and output layer, be arranged at the variable number 1 of Dynamics Performance of High Speed Trains design parameter variables number n and one of them output-index parameter respectively, the interstitial content of hidden layer is defined as m, according to above principle, set up n-m-1 mono-output LM neural network agent model that seven output-indexes are corresponding respectively.
7. a kind of Dynamics Performance of High Speed Trains design key parameter identification method according to claim 1, it is characterized in that: described Calculation of Sensitivity formula is according to the computing formula of input parameter relative to the disturbance degree of output performance index, obtain the sensitivity number of input parameter relative to output performance index, and convert it into relative sensitivity, with to its further analysis of key parameter.
8. a kind of Dynamics Performance of High Speed Trains design key parameter identification method according to claim 1, it is characterized in that: described key parameter recognition rule, be greater than the first five items of 65% for standard with relative sensitivity absolute value, set up effective dynamic performance key parameter recommendation tables, obtain bullet train key parameter recognition result.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510145441.5A CN104898414A (en) | 2015-03-31 | 2015-03-31 | Key parameter identification method for high-speed train dynamics performance design |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510145441.5A CN104898414A (en) | 2015-03-31 | 2015-03-31 | Key parameter identification method for high-speed train dynamics performance design |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN104898414A true CN104898414A (en) | 2015-09-09 |
Family
ID=54031133
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201510145441.5A Pending CN104898414A (en) | 2015-03-31 | 2015-03-31 | Key parameter identification method for high-speed train dynamics performance design |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN104898414A (en) |
Cited By (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106096108A (en) * | 2016-06-06 | 2016-11-09 | 武汉理工大学 | The optimization method of semi-mounted truck side wind stability based on agent model |
| CN106951657A (en) * | 2017-03-31 | 2017-07-14 | 西安石油大学 | One kind abrasion steel rail grinding target profile Fast design method |
| CN109145435A (en) * | 2018-08-16 | 2019-01-04 | 东北大学 | A method of reducing vehicle safety belt concentration phenomenon probability of occurrence |
| CN109918836A (en) * | 2019-03-25 | 2019-06-21 | 西南交通大学 | A kind of bogie suspension parameter matching and optimization method |
| CN110245424A (en) * | 2019-06-17 | 2019-09-17 | 中国科学院力学研究所 | A kind of pneumatic and many-body dynamics close coupling method suitable for bullet train |
| CN111913405A (en) * | 2020-06-23 | 2020-11-10 | 中车工业研究院有限公司 | Multi-software joint simulation system |
| CN112800613A (en) * | 2021-02-04 | 2021-05-14 | 西南交通大学 | Railway vehicle on-line monitoring algorithm development system |
| CN114564878A (en) * | 2021-12-31 | 2022-05-31 | 哈尔滨工业大学（深圳） | Structural dynamics model generation method based on graph neural network |
| CN115056825A (en) * | 2022-05-17 | 2022-09-16 | 北京理工大学 | Adaptive train acceleration method |
| CN116577997A (en) * | 2023-07-06 | 2023-08-11 | 西北工业大学 | Omnidirectional trolley parameter identification method based on concurrent learning |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US5673368A (en) * | 1993-11-11 | 1997-09-30 | Siemens Aktiengesellschaft | Method and device for conducting a process in a controlled system with at least one precomputed process parameter determined using a mathematical model having variable model parameters adjusted based on a network response of a neural network |
| CN1586947A (en) * | 2004-09-09 | 2005-03-02 | 上海交通大学 | Complex test visible system for metro train main loop |
| CN102789531A (en) * | 2012-07-27 | 2012-11-21 | 北京交通大学 | Method for designing jointless track of long and large bridge girder longitudinal butt plate type ballastless track for high-speed railway |
| CN102880759A (en) * | 2012-09-27 | 2013-01-16 | 西南交通大学 | High-speed train large system coupling dynamic simulation platform |
| CN103630378A (en) * | 2012-08-20 | 2014-03-12 | 南车青岛四方机车车辆股份有限公司 | Method for detecting high-speed train passenger comfort degree |
-
2015
- 2015-03-31 CN CN201510145441.5A patent/CN104898414A/en active Pending
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US5673368A (en) * | 1993-11-11 | 1997-09-30 | Siemens Aktiengesellschaft | Method and device for conducting a process in a controlled system with at least one precomputed process parameter determined using a mathematical model having variable model parameters adjusted based on a network response of a neural network |
| CN1586947A (en) * | 2004-09-09 | 2005-03-02 | 上海交通大学 | Complex test visible system for metro train main loop |
| CN102789531A (en) * | 2012-07-27 | 2012-11-21 | 北京交通大学 | Method for designing jointless track of long and large bridge girder longitudinal butt plate type ballastless track for high-speed railway |
| CN103630378A (en) * | 2012-08-20 | 2014-03-12 | 南车青岛四方机车车辆股份有限公司 | Method for detecting high-speed train passenger comfort degree |
| CN102880759A (en) * | 2012-09-27 | 2013-01-16 | 西南交通大学 | High-speed train large system coupling dynamic simulation platform |
Non-Patent Citations (1)
| Title |
|---|
| J.ZHANG 等: "Indentification of key designs parameters of high-speed train for optimal design", 《INT J ADV MANUF TECHNOL》 * |
Cited By (16)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106096108A (en) * | 2016-06-06 | 2016-11-09 | 武汉理工大学 | The optimization method of semi-mounted truck side wind stability based on agent model |
| CN106951657B (en) * | 2017-03-31 | 2020-07-14 | 西安石油大学 | Method for quickly designing grinding target profile of worn steel rail |
| CN106951657A (en) * | 2017-03-31 | 2017-07-14 | 西安石油大学 | One kind abrasion steel rail grinding target profile Fast design method |
| CN109145435A (en) * | 2018-08-16 | 2019-01-04 | 东北大学 | A method of reducing vehicle safety belt concentration phenomenon probability of occurrence |
| CN109145435B (en) * | 2018-08-16 | 2022-11-29 | 东北大学 | Method for reducing occurrence probability of vehicle safety belt concentration phenomenon |
| CN109918836B (en) * | 2019-03-25 | 2022-05-24 | 西南交通大学 | Bogie suspension parameter optimization matching method |
| CN109918836A (en) * | 2019-03-25 | 2019-06-21 | 西南交通大学 | A kind of bogie suspension parameter matching and optimization method |
| CN110245424A (en) * | 2019-06-17 | 2019-09-17 | 中国科学院力学研究所 | A kind of pneumatic and many-body dynamics close coupling method suitable for bullet train |
| CN111913405A (en) * | 2020-06-23 | 2020-11-10 | 中车工业研究院有限公司 | Multi-software joint simulation system |
| CN112800613A (en) * | 2021-02-04 | 2021-05-14 | 西南交通大学 | Railway vehicle on-line monitoring algorithm development system |
| CN112800613B (en) * | 2021-02-04 | 2022-09-20 | 西南交通大学 | Railway vehicle on-line monitoring algorithm development system |
| CN114564878A (en) * | 2021-12-31 | 2022-05-31 | 哈尔滨工业大学（深圳） | Structural dynamics model generation method based on graph neural network |
| CN115056825A (en) * | 2022-05-17 | 2022-09-16 | 北京理工大学 | Adaptive train acceleration method |
| CN115056825B (en) * | 2022-05-17 | 2024-04-05 | 北京理工大学 | Train acceleration method with self-adaptive performance |
| CN116577997A (en) * | 2023-07-06 | 2023-08-11 | 西北工业大学 | Omnidirectional trolley parameter identification method based on concurrent learning |
| CN116577997B (en) * | 2023-07-06 | 2023-10-03 | 西北工业大学 | Omnidirectional trolley parameter identification method based on concurrent learning |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN104898414A (en) | Key parameter identification method for high-speed train dynamics performance design | |
| CN104765916A (en) | Dynamics performance parameter optimizing method of high-speed train | |
| Ye et al. | Fault diagnosis of high-speed train suspension systems using multiscale permutation entropy and linear local tangent space alignment | |
| Auciello et al. | Development and validation of a wear model for the analysis of the wheel profile evolution in railway vehicles | |
| Wang et al. | Semi-active suspension systems for railway vehicles using magnetorheological dampers. Part I: system integration and modelling | |
| Ye et al. | Deep learning-based fault diagnostic network of high-speed train secondary suspension systems for immunity to track irregularities and wheel wear | |
| Mousavi Bideleh et al. | Global sensitivity analysis of bogie dynamics with respect to suspension components | |
| Kong et al. | Optimization of spring fatigue life prediction model for vehicle ride using hybrid multi-layer perceptron artificial neural networks | |
| Mousavi Bideleh et al. | Wear/comfort Pareto optimisation of bogie suspension | |
| Du et al. | A hierarchical framework for improving ride comfort of autonomous vehicles via deep reinforcement learning with external knowledge | |
| Shrestha et al. | Review of adhesion estimation approaches for rail vehicles | |
| Kraft et al. | Black-box modelling of nonlinear railway vehicle dynamics for track geometry assessment using neural networks | |
| Ignesti et al. | Development of a wear model for the wheel profile optimisation on railway vehicles | |
| Ignesti et al. | Wheel profile optimization on railway vehicles from the wear viewpoint | |
| Onat et al. | A new friction condition identification approach for wheel–rail interface | |
| Gadhave et al. | Rail-wheel contact forces and track irregularity estimation from on-board accelerometer data | |
| Magalhaes et al. | Rail vehicle design optimization for operation in a mountainous railway track | |
| Auciello et al. | Development of a model for the analysis of wheel wear in railway vehicles | |
| Bosso et al. | Application of machine learning techniques to build digital twins for long train dynamics simulations | |
| Zou et al. | Research on optimization design of suspension parameters of railway vehicle bogies based on surrogate model | |
| de Paula Pacheco et al. | Optimization of heavy haul railway wheel profile based on rolling contact fatigue and wear performance | |
| Teng et al. | A lightweight model of wheel‐rail force inversion for railway vehicles | |
| Meli et al. | An innovative degraded adhesion model for railway vehicles: development and experimental validation | |
| Zhang et al. | Research on the simulation of wheelset response characteristic identification of railway fastener loosening | |
| Li et al. | Improved generalized energy index method for comprehensive evaluation and prediction of track irregularity |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| C41 | Transfer of patent application or patent right or utility model | ||
| TA01 | Transfer of patent application right |
Effective date of registration: 20170124 Address after: 610031 Sichuan City, Chengdu Province, No. two North Ring Road, No. 111 Applicant after: Xi-nan Jiatoong Univ. Applicant after: Chengdu Tianyou hit soft Technology Co. Ltd. Address before: 610031 Sichuan City, Chengdu Province, No. two North Ring Road, No. 111 Applicant before: Xi-nan Jiatoong Univ. |
|
| WD01 | Invention patent application deemed withdrawn after publication | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20150909 |