Research On Fault Diagnosis Of Electro-hydraulic Turnout Switch Machine Based On Deep Learning

As one of the basic equipments for railway signals,the switch machine is safe,stable and reliable,which plays an important role in ensuring the safe transportation of railways.At present,most of the maintenance of switch machines are periodically checked and repaired based on regulations,which reduces the possibility of faults to a certain extent.However,artificial long-term inspection of the same type of equipment will cause visual fatigue,regardless of the periodicity of the actual operating state of the equipment.Inspection will greatly increase the amount of manual labor.Frequent switch box operations are most likely to cause normal operation of the equipment due to improper operation of the electrician.The action current curve of the switch can reflect the running state of the switch machine,obtain the raw data of the switch machine in the microcomputer monitoring system,analyze and count the common fault type of the switch machine,based on the original data type of the switch machine and the applicable range of the deep learning model.The fault diagnosis model of switch machine is established.Through the continuous optimization and adjustment of the model structure,the fault classification of switch machine has higher accuracy rate.The fault diagnosis system is built based on the model to verify the feasibility of its practical application.In this paper,the ZYJ7 electro-hydraulic ballast switcher with more railways in China is taken as the research object.Based on the action sequence of the switch machine and the advantages of LSTMs for time series data processing,four fault diagnosis models based on LSTMs are established.Four kinds of model fault classification effects are obtained and optimized.The model based on uniform distribution sampling is used to optimize the number of hidden layers and the number of nodes.The influence of different optimization algorithms on the accuracy of model fault classification is compared and analyzed.The search method optimizes the learning rate,and based on the performance of the four models after ROC comparison optimization,the optimal model is selected.Based on the PCA-LDA data denoising and dimension reduction,the variance percentage of each component is different.The Attention mechanism is integrated into the optimal model,and the different coefficients of the model are assigned to different lengths for training.The fault diagnosis system is established by calling the trained model.In the system,the increase of data collection of the switch machine realizes incremental training of the model.The initial four LSTMs models are optimized to achieve optimal multi-layer bidirectional LSTMs,and the fault classification accuracy rate can reach 97.2%.After theconvergence of the Attention mechanism,the fault classification accuracy can reach 97.6%,and zero miss diagnosis can be achieved.It is indicated that the multi-layer bidirectional LSTMs with integrated Attention mechanism have better fault diagnosis for switch machine.The established fault diagnosis system has good applicability and can realize fast and accurate diagnosis of fault type.After incremental training,the model can maintain about98%.The accuracy rate proves that the system has high reliability for the fault diagnosis of the ZYJ7 electro-hydraulic ballast switch machine.