A Capsule Network with Deep Convolution Generative Adversarial Network Model for Rolling Bearing Fault Diagnosis

Abstract

In this study, a new fault diagnosis method named Capsule Network with Deep Convolution Generative Adversarial Network (DCGAN-CapsNet) is proposed for rolling bearings to cope with the insufficiency of CWRU(Case Western Reserve University) bearing data samples and the difficulty of network model training. First, a DCGAN model with four convolutional layers is designed to expand one-dimensional sample data. Second, a CapsNet with two convolutional layers is constructed, and the deep features of bearing fault data are extracted through the convolutional layer. Finally, the dynamic routing algorithm in the capsule neural network is applied to dynamically assign the features to the fault capsule category. By setting the threshold size to determine the probability of multiple faults, so as to complete the task of multi-fault identification. The proposed method alleviates the impact of insufficient data and the fault diagnosis accuracy rate exceeds 99%. The data set expanded by DCGAN is compared with the existing overlapping sampling expansion method, The prediction result of this model is compared with those of existing models such as the convolutional neural network. The effectiveness and potential of the proposed method are verified by experiments.