Fault Diagnosis Method Of The On-board Equipment Of CTCS Based On LSTM-BP Neural Network

Chinese Train Control System(CTCS)is the core of guaranteeing the high-speed and safe operation of the train,and it is very important for the fast and effective fault diagnosis of the on-board equipment as its key component.At present,the fault diagnosis of CTCS on-board equipment mainly relies on the human’s experience,and the artificial intelligence fault diagnosis method has less research results with focuses on system-level fault analysis,which has little research on the equipment running data.In this thesis,according to the log data of on-board equipment,the data features are analyzed deeply,the LSTM-BP optimization network model is established,an intelligent diagnosis system is built which can obtain and use the key fault information flexibly,and its performance is simulated and validated.The research contents of this thesis main include:(1)This thesis takes the running information of 300T on-board equipment as the research object.First,build the information corpus of on-board equipment through the text data mining method.Aiming at the problem that one-hot representation is easy to cause dimension disaster and lack of semantic expression,a language model based on deep learning is used to realize the vector expression of key information of on-board equipment operation,to express semantic information in the form of word vectors,and to provide accurate and effective data support for subsequent models.(2)A fault diagnosis method of on-board equipment based on BP Neural network model is proposed to solve the problem that the same fault phenomenon may be caused by a variety of fault causes and the same fault causes corresponding to different symptoms.The neural network model was optimized by the conjugate gradient method,LM algorithm and Bayesian regularization algorithm,which improved the disadvantage that the training process was easy to fall into the local optimum and the poor classification ability.Then the generalization ability of the three optimization model is compared,and the best improved algorithm is found for the object in this thesis.(3)A cascade model of LSTM and BP Neural network was present in this thesis because of BP network is inaccurate for fault classification with time feature,which makes full use of the memory characteristics of LSTM network model and combines the previous operation information of on-board equipment to judge the current sample’s type.Cascade Neural network structure can further improve the fault classification ability of on-board equipment.The experimental results show that the improved Bayesian regularization model has the advantages of stable performance and high generalization ability,compared with other two algorithms.Compared with the original BP network model,the accuracy of the optimized model to the unknown sample classification is increased from 48.76%to 85.06%.The LSTM-BP network cascade Model presented in this thesis has excellent resolving power for the faults caused by other faults,and the classification accuracy of unknown samples is 95.13%,which satisfies the requirement of the field and verifies the feasibility of this scheme.