Research On Fault Diagnosis Methods Of High-speed Train Bogie Based On Convolutional Neural Network

As the only connecting part of the train body and the track,the normal working condition of the bogie is one of the key factors to ensure the steady running of the high-speed train,the ability to pass curves smoothly and the comfort of passengers.In order to monitor the running condition of the train in real time and avoid emergencies during operation,installing sensors on the bogie and analyzing the vibration signal is an effective way to conduct fault diagnosis However,the vibration signal of the running train is very complicated,and there are often a lot of noise and redundancy in the signal,as well as coupling among different channels.To solve these problems,this thesis mainly studies the automatic feature extraction and fault diagnosis based on convolutional neural network.The main contents are as followFirstly,the related concepts of convolutional neural networks and the evolution of state-of-the-art models are briefly introduced.In order to extract more complex and abstract features and achieve a more accurate classification,the number of convolutional neural network layers is deepening.So as to achieve an effective deep network,various techniques have emerged to solve the problems of gradient disappearance and degradation when the model is deepened,and modular methods and optimization methods are proposed to make full use of computing resources.This part is the theoretical basis and basis of the following textSecondly,the research object was elaborated.Starting from the mechanical construction of high-speed trains,the key components associated with bogie failures and the possible consequences of failures are described.The model and schemes of the simulation based on SIMPACK is described,and the data composition and effectiveness of the simulation are explained.The collected signals are briefly analyzed from time domain and through spectrograms.Based on the above analysis of spectrograms,a fault type recognition algorithm based on random forest and voting method is designed.By training a random forest classifier for each channel and then using the voting method to fuse the results of each classifier,a high recognition accuracy is achieved,which is much higher than the classification accuracy of each channel.In order to make full use of the information contained in the original signal and simplify the process of the method above,a deep convolutional neural network model named RSNet(Residual squeeze network)is designed.This model takes the original time domain sampling points as input and automatically fuses the channel fusion in the feature extraction process.The basic architecture of this model is one-dimensional convolutional neural network,which is very suitable for time series with multiple channels,and has less time complexity comparing to traditional two-dimensional convolutional neural networks.The experimental results show that the algorithm has the accuracy around 99%in both fault type recognition and fault component’s location recognition tasks,and has strong robustness regardless of running speeds.Finally,for the more complex problem,the evaluation of performance degradation of key components of bogies,the above algorithm has been studied and adapted from the perspectives of classification and regression.The t-SNE(t-distributed stochastic neighbor embedding),a technique of data dimensionality reduction is applied to visually analyze the proposed convolutional neural network model.