A Convolutional Neural Network with Hybrid Loss Function for Bearing Fault Diagnosis

This paper proposes a method based on a convolutional neural network (CNN) with hybrid loss function to solve a problem of bearing fault diagnosis in different working conditions. Firstly, Swish activation function is introduced into traditional convolutional neural network which can reduce the effect of vanishing gradient and optimize the training process. Secondly, a hybrid loss function combines margin loss function and cross entropy loss function (CEM) is used to improve the classification effect. Finally, the parameters of shallow layers pre-trained are frozen and the weights of deep layers are fine-tuned to accommodate the target domain sample distribution. In this paper, gray scale image is taken as the network input. The proposed method is experimentally verified on the bearing data of Case Western Reserve University and the fan data set. The experiment results show that the proposed method outperforms the comparing methods.