Bearing fault diagnosis based on parallel multi-channel deep CNN and optimization with local sparse structure

Abstract

To solve the problem that signal processing-based bearing fault diagnosis methods require rich professional knowledge and experience and cannot cope with the large amount of data, an end-to-end one-dimensional parallel multi-channel deep convolutional neural network (PMDCNN) has been proposed. The experimental results show that the network model can effectively mine the different signal characteristics of different channels, which can realize the bearing fault diagnosis with high accuracy. The performance of the model is improved by increasing the depth of the network, using batch normalization (BN), Dropout technique, and choosing the right optimizer and learning rate. To reduce the number of model parameters, a local sparse structure is employed. Firstly, the larger convolutional kernels are replaced by smaller sized ones, thus greatly reducing the redundant parameters of that convolutional layer. Secondly, the fully connected layer is substituted by a global average pooling layer to further reduce the number of parameters. The final result is that the number of parameters is only 1/66th of the previous model with similar accuracy, and the training time is reduced, which shows the optimization method of local sparse structure achieves good results. The average fault diagnosis accuracy of 99.90 $\pm$ 0.03% in Case Western Reserve University (CWRU), 99.80 $\pm$ 0.05% in comprehensive experimental platform, and 99.65 $\pm$ 0.10% in University of Ottawa Rolling-element Dataset-Vibration and Acoustic Faults under Constant Load and Speed conditions (UORED-VAFCLS) also indicates that the optimized network model is good at extracting fault features from the original fault data, reflecting the superiority of the proposed method.