A bearing fault diagnosis method based on improved convolution neural network and transfer learning

Abstract

Aiming at these issues of complex working conditions and small sample dataset in practical bearing fault diagnosis, a novel method based on improved convolution neural network and transfer learning is proposed. In this method, the global average pooling layer is firstly used to replace the traditional full connection layer as output layer, which can reduce the training parameters of whole network. Meanwhile, the original ReLu activation function is removed and then LeakyReLu activation function is introduced to make the network model be hard to produce Dead ReLu. The improved network is pretrained with the source domain data to learn the fault diagnosis feature. Here, the bottom network parameters are frozen, and the top network is also fine-tuned with a small amount of target domain data to obtain the transfer model. Finally, the target domain data is input to the transfer model for bearing fault diagnosis. The proposed method is verified by considering the scenario of small sample and cross working conditions with the bearing dataset of Case Western Reserve University. The results show that the average classification accuracy of the proposed method can reach 95.49%, and is higher than that of normal convolution neural network.