A Batch-Normalized Deep Neural Networks and its Application in Bearing Fault Diagnosis

At present, in the field of fault diagnosis, deep learning has shown state-of-the-art performance in processing mechanical big data. This paper studies the deep neural networks(DNN) model based on auto-encoder, which has high performance in bearing fault diagnosis. However, the traditional structure of stacked auto encoders has the problem of internal covariant transfer, that inhibits the training efficiency and generalization ability of the network. To overcome the aforementioned deficiency and further explore the performance of DNN, a batch normalization layer is employed in the fully connected layer of the DNN during training, so the network can obtain the stable distribution of activation values. Therefore, this paper proposes a new intelligent diagnosis method named batch normalization deep neural networks(BN-DNN). Finally, the experimental results show that: (1) The performance of BN-DNN is better than DNN. (2) BN-DNN can directly process the raw vibration signals, and the diagnostic accuracy can be maintained above 99% under different working conditions.