An Intelligent Fault Diagnosis Method of Small Sample Bearing Based on Improved Auxiliary Classification Generative Adversarial Network

Abstract

Intelligent diagnosis is one of the key points of research in the field of bearing fault diagnosis. As a representative unsupervised data expansion method, generative adversarial networks (GANs) may solve the problem of insufficient samples in rotating machinery fault diagnosis. In this article, a new method based on an improved auxiliary classification generative adversarial network (ACGAN) is proposed. First, using the ACGAN framework to add fault label information to the input sample improves the quality of the generated sample. The Wasserstein distance is introduced into the loss function in ACGAN, which effectively solves the problems of gradient collapse and gradient disappearance. Second, the weight penalty is added to the loss function, which increases the expression ability of the network compared with the weight clipping and effectively solves the problem of unstable network training. Finally, an independent fault diagnosis classifier is introduced to balance the compatibility of discrimination and classification. The convolutional block attention module (CBAM) based on the attention mechanism is introduced to fuse its output with the features extracted by the convolution module, which improves the representation ability of the fault diagnosis network and realizes high-precision fault diagnosis. The method is applied to the public bearing dataset of Western Reserve University and the bearing simulation dataset of the Yanshan University Laboratory, which can generate high-quality multimode fault samples more efficiently and can be used to help the training of fault diagnosis models based on deep learning. It has high accuracy, generalization, and stability.