Research On Fault Diagnosis Of Rolling Element Bearings In High-speed Train Based On Convolutional Neural Network

As key parts of bogie in railway vehicle,the healthy status of rolling element bearings is directly related to the safety operation of high-speed train,so it is very important to diagnose the faults as early as possible.Aiming at the diagnosis tasks under different working condition or with multi-source information in high-speed train,a serious of bearing fault diagnosis methods based on convolutional neural network were proposed in this paper.The main contents are listed as follows.Firstly,a deep transfer diagnosis model(Transfer convolutional neural networks,TCNN)based on Res Ne Xt-50 was proposed to solve the problem that the shallow neural network could not guarantee the classification accuracy with insufficient samples of the target task.Combined with the strategy of parameter transfer learning,a pre-trained model Res Ne Xt-50 based on Image Net dataset was used as feature extraction module to diagnose the cross-operation and cross-physics data of rolling element bearings.Experimental results show that the proposed model not only has strong classification performance,but also greatly reduces the training difficulty of deep network model in diagnosis task based on small volume labeled samples.Secondly,an improved domain adaptive transfer learning model(Domain Adaptation Neural Networks,DANN)is proposed to solve the problem that the portability of deep features is significantly reduced with a maximum mean difference(MMD)regularization constraint only at the end of deep network.Different from TCNN,DANN does not need a pre-trained deep network for parameter migration.Instead,the model parameters are trained by embedding MMD regularization terms in multiple specific task layers to reduce the differences of data distribution between the source domain and target domain.The experimental results also show that the proposed model has strong ability of learning domain invariance and can enhance the transferability of deep features.Finally,an improved dual-flow CNN fusion model was proposed to solve the problem that single sensor data cannot reflect all of the fault information in characterizing the condition of rolling element bearings.In this model,the data from accelerometer and acoustic sensor were used as the inputs of two single-channel CNN models respectively to fuse the multi-source information at feature level.The experimental results show that the proposed model can integrate the sensor information of various types and various directions and effectively expand the feature space.