CN102521651A - Bow net contact force prediction method based on NARX neural networks - Google Patents

Bow net contact force prediction method based on NARX neural networks Download PDF

Info

Publication number
CN102521651A
CN102521651A CN2011104362224A CN201110436222A CN102521651A CN 102521651 A CN102521651 A CN 102521651A CN 2011104362224 A CN2011104362224 A CN 2011104362224A CN 201110436222 A CN201110436222 A CN 201110436222A CN 102521651 A CN102521651 A CN 102521651A
Authority
CN
China
Prior art keywords
data
normalization
test
bow net
contact force
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
Application number
CN2011104362224A
Other languages
Chinese (zh)
Inventor
秦勇
贾利民
张媛
陈皓
张道于
朱跃
邢宗义
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Beijing Jiaotong University
Original Assignee
Beijing Jiaotong University
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Beijing Jiaotong University filed Critical Beijing Jiaotong University
Priority to CN2011104362224A priority Critical patent/CN102521651A/en
Publication of CN102521651A publication Critical patent/CN102521651A/en
Pending legal-status Critical Current

Links

Images

Landscapes

  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)
  • Feedback Control In General (AREA)

Abstract

The invention, which belongs to the railway safe operation control technology field, discloses a bow net contact force prediction method based on nonlinear auto-regressive with eXogenous input (NARX) neural networks. The method comprises: test data are obtained by a simulation test; Normalization processing is carried out on the test data; an NARX neural networks prediction model is established; the first set number of data pairs are extracted from the test data that have been processed by the normalization and are used as training samples, and a bayesian regularization algorithm is employed to train the NARX neural networks prediction model; the second set number of data pairs are extracted from the test data that have been processed by the normalization and are used as testing samples; and data in the test samples are used as input data and are input into the trained NARX neural networks prediction model, reverse normalization processing is carried out on an output result and then the processed output result is used as a bow net contract force predicted value. According to the invention, the NARX neural networks model is employed to predict a bow net contact force, so that prediction precision of the bow net contact force is improved.

Description

Bow net contact force Forecasting Methodology based on the NARX neural network
Technical field
The invention belongs to safe railway operation control technology field, relate in particular to a kind of bow net contact force Forecasting Methodology based on the NARX neural network.
Background technology
Develop High-Speed Railway is the inexorable trend of China railways modernization construction, and electric locomotive is big with its freight volume, speed is fast, energy consumption is low, pollution is little, freight rate is honest and clean and advantage such as safe and reliable, becomes the main force in the express locomotive.In the high-speed electric railway system, directly related with train speed is the pantograph-catenary current collection system, promptly when the train high-speed cruising, must keep the stable stream mode that receives, and that is to say between pantograph and osculatory to have certain contact pressure.When contact pressure is too small, be prone to cause off-line, promptly pantograph disengages line and produces electric arc; When contact pressure was excessive, osculatory lifting amount was excessive, made the osculatory local bending, caused the osculatory fatigue damage, and the osculatory abrasion are increased, and caused accident between pantograph when serious.Therefore, bow net contact force measurement is all significant for the development that guarantees train traffic safety and China Express Railway.
The method that former China, people such as Bi Jihong were employed between bow net and set up " contact to " realizes the bow net coupling, has set up the rigid suspension catenary coupling model, and model is carried out non-linear TRANSIENT DYNAMIC ANALYSIS, obtains the change curve of contact force and slide plate vertical displacement.Xu Haidong; People such as Xu Min have set up the bow net kinetic model based on the big System Dynamics Theory of railway; And the vibratory response at locomotive top pantograph pedestal place is rich as swashing of bow net system, be applied to kinetic model, inquired into of the influence of track coupled vibrations to the bow net contact pressure.
Chinese scholars has site test, half in kind half virtual (pantograph is in kind, and contact net obtains through Computer Simulation) test and computer simulation etc. to the research method of bow net.Site test requires highly to measuring method, data processing etc., has only few countries such as Germany can directly measure the contact force between pantograph and contact net.Along with the appearance of high-speed electronic computer, make and utilize numerical method to come authentic this bulky structure of simulation contact net to become possibility, computer simulation method has become the most general research method.
The implementation method of contact net and pantograph coupling is divided into two kinds basically: proposition such as a kind of Wu Tianhang of being pass through bow, net equates to realize coupling in the displacement of contact point place, a kind of in addition spring realization of pass through that is Zhang Weihua etc. proposes is coupled.The former needn't select contact stiffness, but can't consider off-line; And for the latter, because the dynamic contact force between bow net and the relation of sledge displacement are in the reality: when contacting, both keep in touch the relation of rigidity; In case off-line, contact force will be always zero, no longer have any contact between the two.Therefore can not study the vibration of off-line with the coupling of spring simulation bow net with cantle, net.
Can also use the life and death monotechnics among the engineering design software ANSYS to simulate pantograph moving along contact net; Specific practice is all to set up a pantograph at each node place of osculatory; In solution procedure,, activate or kill corresponding pantograph through life and death monotechnics.The maximum defective of this method is to consider the continuity and the transitivity of the node speed of pantograph, displacement, so result of calculation and the fact differ greatly, and is very inaccurate.
Can find out that through above-mentioned introduction therefore the defective of the existing equal various degrees of bow net contact force Forecasting Methodology is necessary to propose a kind of new bow net contact force Forecasting Methodology, to improve bow net contact force prediction accuracy.
Summary of the invention
The objective of the invention is, a kind of bow net contact force Forecasting Methodology based on the NARX neural network is provided, in order to solve the not high problem of bow net contact force precision that bow net contact force Forecasting Methodology commonly used is calculated.
To achieve these goals, the technical scheme of the present invention's employing is that a kind of bow net contact force Forecasting Methodology based on the NARX neural network is characterized in that said method comprises:
Step 1: obtain test figure through l-G simulation test; Wherein, test figure comprises osculatory irregularity data and the bow net contact force data corresponding with it;
Step 2: test figure is carried out normalization handle;
Step 3: set up the NARX neural network prediction model;
Step 4: the test figure after normalization is handled, extract first and set quantity data to as training sample; Wherein, the bow net contact force data after data are handled the osculatory irregularity data after being meant normalization and handling and the normalization corresponding with it;
Step 5: the bow net contact force data after osculatory irregularity data after the normalization in the training sample handled and the normalization corresponding with it are handled are respectively as input data and output data, employing Bayesian regularization algorithm training NARX neural network prediction model;
Step 6: again the test figure after normalization is handled, extract second and set quantity data to as test sample book; Wherein, the bow net contact force data after data are handled the osculatory irregularity data after being meant normalization and handling and the normalization corresponding with it;
Step 7: be input in the NARX neural network prediction model of step 5 training as the input data with the osculatory irregularity data after the processing of the normalization in the test sample book; The output result is carried out anti-normalization handle, the output result after anti-normalization is handled is as bow net contact force predicted value.
Saidly obtain test figure through l-G simulation test and specifically be, set up bow net Coupled Dynamics model earlier, utilize MATLAB/Simulink software to carry out dynamic simulation again and obtain osculatory irregularity data and the bow net contact force data corresponding with it.
It is said that test figure is carried out that normalization handles specifically is to utilize formula
x i scal = x i - x min x max - x min
To test figure x iCarrying out normalization handles; Wherein,
Figure BDA0000123796280000032
Figure BDA0000123796280000033
N is the number of test figure.
The middle layer node of said NARX neural network prediction model adopts the tan-sigmoid function, and the output layer node adopts linear function, and the input layer number is 1, and the middle layer node number is 15, and the output layer interstitial content is 1, and it all is 45 that input and output postpone; Wherein, Said tan-sigmoid function is the input data of x for latent layer; T is a zoom factor, and θ is a displacement coefficient.
The said first setting quantity data is right to being 1300 data.
The said second setting quantity data is right to being 700 data.
Also comprising the step that adopts the root-mean-square error method to estimate NARX neural network prediction model performance after the said step 7, specifically is to utilize formula
RMSE ( y , y m ) = 1 N Σ i = 1 N ( y ( i ) - y m ( i ) ) 2
Estimate the performance of the NARX neural network prediction model after training; Wherein, y (i) is the desired value in the test sample book, y m(i) be predicted value after anti-normalization is handled, N is the data number in the test sample book.
The present invention adopts NARX Neural Network model predictive bow net contact force, has improved the precision of prediction of bow net contact force.
Description of drawings
Fig. 1 is based on the bow net contact force Forecasting Methodology process flow diagram of NARX neural network;
Fig. 2 is NARX neural network structure figure;
Fig. 3 is output of bow net contact force test data and NARX neural network output comparison diagram;
Fig. 4 is the correlation analysis of output of bow net contact force test data and the output of NARX neural network.
Embodiment
Below in conjunction with accompanying drawing, preferred embodiment is elaborated.Should be emphasized that following explanation only is exemplary, rather than in order to limit scope of the present invention and application thereof.
Fig. 1 is based on the bow net contact force Forecasting Methodology process flow diagram of NARX neural network.Among Fig. 1, comprise based on the bow net contact force Forecasting Methodology of NARX neural network:
Step 1: obtain test figure through l-G simulation test; Wherein, test figure comprises osculatory irregularity data and the bow net contact force data corresponding with it.
Obtaining test figure through l-G simulation test specifically is, sets up bow net Coupled Dynamics model earlier, utilizes MATLAB/Simulink software to carry out dynamic simulation again and obtains osculatory irregularity data and the bow net contact force data corresponding with it.
Step 2: test figure is carried out normalization handle.
Test figure is carried out normalization to be handled and to comprise that butt joint touches the normalization of line irregularity data and bow net contact force data corresponding with it and handle.It specifically is to utilize formula that normalization is handled
x i scal = x i - x min x max - x min
Osculatory irregularity data and bow net contact force data are carried out the normalization processing; Wherein,
Figure BDA0000123796280000053
N is the number of test figure, x iBe osculatory irregularity data/bow net contact force data.
Step 3: set up the NARX neural network prediction model.
The structure of NARX neural network (Nonlinear Auto-Regressive with eXogenous input Neural Networks) is as shown in Figure 2.The middle layer node of NARX neural network prediction model adopts the tan-sigmoid function, and the output layer node adopts linear function, and the input layer number is 1, and the middle layer node number is 15, and the output layer interstitial content is 1, and it all is 45 that input and output postpone.Wherein, The tan-sigmoid function is the input data of
Figure BDA0000123796280000054
x for latent layer; T is a zoom factor, and θ is a displacement coefficient.
Step 4: the test figure after normalization is handled, extract 1300 data to as training sample; Wherein, the bow net contact force data after data are handled the osculatory irregularity data after being meant normalization and handling and the normalization corresponding with it.
Step 5: the bow net contact force data after osculatory irregularity data after the normalization in the training sample handled and the normalization corresponding with it are handled are respectively as input data and output data, employing Bayesian regularization algorithm training NARX neural network prediction model.
Bayesian regularization (BR; Bayesian Regularization) algorithm; Be meant in order to improve network promotion ability; To set up a property parameter that constitutes by each layer output error, weights and threshold values in the training process,, this parameter minimized through the weights and the threshold values of network being adjusted according to the Levenberg-Martquartdt optimum theory.
Step 6: again the test figure after normalization is handled, extract 700 data to as test sample book; Wherein, the bow net contact force data after data are handled the osculatory irregularity data after being meant normalization and handling and the normalization corresponding with it.
Step 7: be input in the NARX neural network prediction model of step 5 training as the input data with the osculatory irregularity data after the processing of the normalization in the test sample book; The output result is carried out anti-normalization handle, the output result after anti-normalization is handled is as bow net contact force predicted value.
Can also adopt the root-mean-square error method to estimate the performance of above-mentioned NARX neural network prediction model, specifically be to utilize formula
RMSE ( y , y m ) = 1 N Σ i = 1 N ( y ( i ) - y m ( i ) ) 2
Estimate the performance of the NARX neural network prediction model after training; Wherein, y (i) is the desired value in the test sample book, y m(i) be predicted value after anti-normalization is handled, N is the data number in the test sample book.
The RMSE value is more little, and the precision of prediction of representation model is high more, and predicted value is more near desired value.Secondly, model output and target output are carried out curve fitting, can reflect the degree of approximation between target output value and the model output valve more intuitively, as shown in Figure 3.At last, linear regression analysis is carried out in model output and target output, can accurately be calculated the related coefficient between target output value and the model output valve, as shown in Figure 4, wherein, A neural network output data, T representative test output data.
The above; Be merely the preferable embodiment of the present invention, but protection scope of the present invention is not limited thereto, any technician who is familiar with the present technique field is in the technical scope that the present invention discloses; The variation that can expect easily or replacement all should be encompassed within protection scope of the present invention.Therefore, protection scope of the present invention should be as the criterion with the protection domain of claim.

Claims (7)

1. bow net contact force Forecasting Methodology based on the NARX neural network is characterized in that said method comprises:
Step 1: obtain test figure through l-G simulation test; Wherein, test figure comprises osculatory irregularity data and the bow net contact force data corresponding with it;
Step 2: test figure is carried out normalization handle;
Step 3: set up the NARX neural network prediction model;
Step 4: the test figure after normalization is handled, extract first and set quantity data to as training sample; Wherein, the bow net contact force data after data are handled the osculatory irregularity data after being meant normalization and handling and the normalization corresponding with it;
Step 5: the bow net contact force data after osculatory irregularity data after the normalization in the training sample handled and the normalization corresponding with it are handled are respectively as input data and output data, employing Bayesian regularization algorithm training NARX neural network prediction model;
Step 6: again the test figure after normalization is handled, extract second and set quantity data to as test sample book; Wherein, the bow net contact force data after data are handled the osculatory irregularity data after being meant normalization and handling and the normalization corresponding with it;
Step 7: be input in the NARX neural network prediction model of step 5 training as the input data with the osculatory irregularity data after the processing of the normalization in the test sample book; The output result is carried out anti-normalization handle, the output result after anti-normalization is handled is as bow net contact force predicted value.
2. method according to claim 1; It is characterized in that saidly obtaining test figure through l-G simulation test and specifically being; Set up bow net Coupled Dynamics model earlier, utilize MATLAB/Simulink software to carry out dynamic simulation again and obtain osculatory irregularity data and the bow net contact force data corresponding with it.
3. method according to claim 1 is characterized in that said test figure is carried out that normalization handles specifically is to utilize formula
x i scal = x i - x min x max - x min
To test figure x iCarrying out normalization handles; Wherein,
Figure FDA0000123796270000022
Figure FDA0000123796270000023
N is the number of test figure.
4. method according to claim 1; The middle layer node that it is characterized in that said NARX neural network prediction model adopts the tan-sigmoid function; The output layer node adopts linear function, and the input layer number is 1, and the middle layer node number is 15; The output layer interstitial content is 1, and it all is 45 that input and output postpone; Wherein, Said tan-sigmoid function is the input data of x for latent layer; T is a zoom factor, and θ is a displacement coefficient.
5. method according to claim 1 is characterized in that the said first setting quantity data is right to being 1300 data.
6. method according to claim 1 is characterized in that the said second setting quantity data is right to being 700 data.
7. method according to claim 1 is characterized in that also comprising the step that adopts the root-mean-square error method to estimate NARX neural network prediction model performance after the said step 7, specifically is to utilize formula
RMSE ( y , y m ) = 1 N Σ i = 1 N ( y ( i ) - y m ( i ) ) 2
Estimate the performance of the NARX neural network prediction model after training; Wherein, y (i) is the desired value in the test sample book, y m(i) be predicted value after anti-normalization is handled, N is the data number in the test sample book.
CN2011104362224A 2011-12-22 2011-12-22 Bow net contact force prediction method based on NARX neural networks Pending CN102521651A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN2011104362224A CN102521651A (en) 2011-12-22 2011-12-22 Bow net contact force prediction method based on NARX neural networks

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN2011104362224A CN102521651A (en) 2011-12-22 2011-12-22 Bow net contact force prediction method based on NARX neural networks

Publications (1)

Publication Number Publication Date
CN102521651A true CN102521651A (en) 2012-06-27

Family

ID=46292562

Family Applications (1)

Application Number Title Priority Date Filing Date
CN2011104362224A Pending CN102521651A (en) 2011-12-22 2011-12-22 Bow net contact force prediction method based on NARX neural networks

Country Status (1)

Country Link
CN (1) CN102521651A (en)

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110457822A (en) * 2019-08-13 2019-11-15 西南交通大学 A kind of contact force threshold model construction method for determining bow net contact electric arc and occurring
CN111209999A (en) * 2018-11-21 2020-05-29 成都唐源电气股份有限公司 Contact network performance degradation prediction method based on recurrent neural network
CN111367173A (en) * 2020-03-06 2020-07-03 西南交通大学 High-speed railway pantograph robust prediction control method based on state estimation
CN113028999A (en) * 2021-02-24 2021-06-25 河南辉煌科技股份有限公司 Contact line lift measurement method and system based on convolutional neural network
CN113076949A (en) * 2021-03-31 2021-07-06 成都唐源电气股份有限公司 Method and system for quickly positioning parts of contact net
CN113267286A (en) * 2021-07-02 2021-08-17 中国国家铁路集团有限公司 Railway bow net contact force identification method and device

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101819407A (en) * 2010-04-02 2010-09-01 杭州电子科技大学 Sewage pump station water level prediction method base on neural network
CN101963638A (en) * 2010-09-07 2011-02-02 西南交通大学 Bow net electric contact characteristic testing device

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101819407A (en) * 2010-04-02 2010-09-01 杭州电子科技大学 Sewage pump station water level prediction method base on neural network
CN101963638A (en) * 2010-09-07 2011-02-02 西南交通大学 Bow net electric contact characteristic testing device

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
XUE-MIAO PANG ET AL: "The Prediction of Derailment Coefficient Using NARX Neural Networks", 《ICTIS 2011》, 2 July 2011 (2011-07-02) *
蔡磊,等: "利用NARX神经网络由IMF与太阳风预测暴时SYM_H指数", 《中国科学》, vol. 40, no. 1, 31 December 2010 (2010-12-31) *

Cited By (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111209999A (en) * 2018-11-21 2020-05-29 成都唐源电气股份有限公司 Contact network performance degradation prediction method based on recurrent neural network
CN111209999B (en) * 2018-11-21 2023-04-07 成都唐源电气股份有限公司 Contact network performance degradation prediction method based on recurrent neural network
CN110457822A (en) * 2019-08-13 2019-11-15 西南交通大学 A kind of contact force threshold model construction method for determining bow net contact electric arc and occurring
CN110457822B (en) * 2019-08-13 2022-04-29 西南交通大学 Contact force threshold value model construction method for judging generation of bow net contact arc
CN111367173A (en) * 2020-03-06 2020-07-03 西南交通大学 High-speed railway pantograph robust prediction control method based on state estimation
CN111367173B (en) * 2020-03-06 2021-06-25 西南交通大学 High-speed railway pantograph robust prediction control method based on state estimation
CN113028999A (en) * 2021-02-24 2021-06-25 河南辉煌科技股份有限公司 Contact line lift measurement method and system based on convolutional neural network
CN113076949A (en) * 2021-03-31 2021-07-06 成都唐源电气股份有限公司 Method and system for quickly positioning parts of contact net
CN113267286A (en) * 2021-07-02 2021-08-17 中国国家铁路集团有限公司 Railway bow net contact force identification method and device
CN113267286B (en) * 2021-07-02 2022-12-13 中国国家铁路集团有限公司 Railway bow net contact force identification method and device

Similar Documents

Publication Publication Date Title
CN102521651A (en) Bow net contact force prediction method based on NARX neural networks
Pombo et al. Multiple pantograph interaction with catenaries in high-speed trains
Polach Comparability of the non-linear and linearized stability assessment during railway vehicle design
CN101281116A (en) Wide span rail traffic bridge damnification detecting system
Spiryagin et al. Problems, assumptions and solutions in locomotive design, traction and operational studies
CN102360454A (en) Wheel-track force prediction method based on NARX (Nonlinear Auto-regressive with Extra Inputs) neural network
Li et al. Wheel wear prediction on a high-speed train in China
CN102103061A (en) Isoconcentration line method for atmosphere monitoring, prediction and assessment
CN112989660B (en) Method for predicting corrosion of pipeline under subway stray current based on partial least square method
CN102567576A (en) Method for predicting rate of wheel load reduction
Zhang et al. Sensitivity analysis and research on optimisation methods of design parameters of high‐speed railway catenary
CN105389656A (en) Performance evaluation method of secondary equipment of power system
CN105489010B (en) A kind of through street journey time reliability monitoring analysis system and method
Yin et al. Reliability analysis of subway vehicles based on the data of operational failures
Li et al. Active control strategy for the running attitude of high-speed train under strong crosswind condition
Hou et al. Delay recovery model for high-speed trains with compressed train dwell time and running time
CN102567786A (en) Method for predicting derailment coefficients
CN105303329B (en) A kind of equipment health state evaluation method based on monitoring information
CN103903181A (en) Model for evaluating and detecting subway accident device damage condition
CN103149278A (en) Correction method for on-line monitoring noisy data of oil chromatography
Yang et al. Braking process modeling and simulation of CRH2 electric multiple unit
Wang et al. Wear prediction of metro wheels based on the ARIMA model
CN110795877B (en) Fluid-solid coupling-based pantograph actuating motor torque compensation amount calculation method
Deng et al. A data-driven wheel wear prediction model for rail train based on LM-OMP-NARXNN
CN113484678A (en) Power grid fault impact strength quantitative index analysis method based on EWM

Legal Events

Date Code Title Description
C06 Publication
PB01 Publication
C10 Entry into substantive examination
SE01 Entry into force of request for substantive examination
C12 Rejection of a patent application after its publication
RJ01 Rejection of invention patent application after publication

Application publication date: 20120627