Research On Identification Of Degradation State Of Key Components Of EMU Bogies Based On Machine Learning

Bogie is the key component of high-speed EMU,which will inevitably deteriorate during service,such as performance degradation of dampers,aging of rubber joints,wheel tread wear,etc.,which may cause abnormal shaking of vehicles and endanger running safety in serious cases.In addition,if the degradation state can meet the usage requirements,premature maintenance or replacement will undoubtedly result in resource waste.Therefore,identifying the degradation status of key components of the bogie is very important.At present,PHM technology,which is widely used in the railway industry,mainly monitors the temperature and/or vibration parameters of key components such as train bearings and gears directly by deploying sensors,in order to determine their service status or predict their remaining life.However,for key components of bogies such as dampers,rubber joints and wheel treads,it is not easy to obtain their service status by directly deploying sensors.Therefore,it is of great practical significance to identify the degradation status of the key components of the bogie by using the vehicle dynamic performance data.In view of the above problems,this paper focuses on the research of yaw damper and wheel tread degradation identification methods that affect the lateral dynamic performance of high-speed EMU.First,based on the bench dynamic test data,a non parametric model of yaw damper was constructed,and on this basis,a joint simulation model of the whole vehicle with different degradation status was constructed.Second,through the simulation model,the response indices of vehicle dynamics in different deteriorated states were compared and analyzed,and the machine learning algorithm was used to identify the deteriorated states.Finally,the selected machine learning model was used for degradation identification of measured dynamic response data,and the accuracy of the model was further improved through optimization algorithms.The research results of this paper mainly include the following four parts:(1)Through bench dynamic tests,it was found that the yaw damper had strong nonlinear frequency and amplitude variation characteristics under high-frequency dynamic conditions.A non parametric model of yaw damper based on BP neural network had been constructed.Compared with traditional equivalent parametric models,the non parametric model could more accurately describe the actual dynamic behavior of yaw damper for high-speed EMU.A modeling method based on test data was proposed to establish the vehicle dynamics model,and a joint simulation model of the non parametric yaw damper model and the vehicle dynamics model is constructed.The simulation results were in good agreement with the measured data.(2)Based on the actual service condition of the degradation status of the yaw damper and the wheel tread,a joint simulation model of the vehicle dynamics under the degradation state was constructed,and the impact of different factors on the dynamic performance was analyzed.Based on the application of statistics,five yaw damper states with changes in damping force and joints stiffness were selected,as well as four typical wheel wear states for degradation state analysis.Research had found that it was difficult to describe the changes in dynamic performance of yaw dampers under degradation conditions based solely on their static damping and stiffness characteristics.The dynamic indicators under the condition of wheel wear were more sensitive to changes in the parameters of the yaw damper,but the trend of changes in vehicle dynamics indicators and the damping and stiffness of yaw dampers do not show regular changes.(3)A study was conducted on the recognition effect of different machine learning models for the degradation status of key components of bogies based on the dynamic performance of multiple unit trains.Based on the simulation data of vehicle dynamics,15 classification problems of five yaw damper states and three wheel tread states were designed.The feature data was extracted through the feature extraction methods in the time domain,frequency domain and time-frequency domain,and the feature dimension reduction was carried out.The recognition effect of the model using the support vector machine(SVM)algorithm was average,while the model using the Convolutional Neural Network(CNN)algorithm achieved high recognition accuracy.Considering the possible application demand in the future,the 15 classification problem was simplified into two yaw damper states and three wheel tread states to form a six classification problem,and CNN could still achieve high recognition accuracy.(4)Based on Gramian Angular Summation Fields(GASF)algorithm,a GASF-CNN model was developed,and the identification effect of degradation states was studied based on measured data.To solve the problem of poor classification accuracy of CNN models trained on simulation data on measured data,a decomposition data conversion layer based on the GASF algorithm was constructed to improve the CNN model’s ability to understand data during feature extraction and learning processes.By conducting an analysis of the factors affecting the accuracy of model recognition,the data construction method,data filtering,sliding window length,and number of channels were optimized.The final model achieved high classification accuracy and achieved good model generalization ability for different train operating directions,speed levels,and environments.87 figures,22 tables,and 158 references.