Imbalanced few-shot bearing fault diagnosis via hybrid enhanced VAE-WGAN and attention-guided WDCNN

Abstract

Abstract In order to address the issue that the traditional neural network models fail to train adequately and achieve unsatisfactory diagnostic results due to the scarcity and imbalance of rolling bearing fault data samples in practical industrial scenarios, this paper proposes a fault diagnosis method that combines an improved generative network (variational autoencoder with Wasserstein generative adversarial networks (VAE-WGAN)) with the wide convolutional neural network (WDCNN) based on the convolutional block attention module (CBAM). On the basis of the existing generative network model VAE-WGAN, this method improves the feature learning capability of the generative network by incorporating the self-attention mechanism and a spectral information loss function. Furthermore, considering the limited diagnostic performance of the features from single-domain of raw vibration signal in real-world industrial scenarios, this method innovatively designs a dual-channel network structure that integrates time-frequency domain features, building upon the original WDCNN network model. Additionally, a CBAM is introduced to enhance the network’s feature extraction ability. The experimental results demonstrate that the model integrating the enhanced VAE-WGAN with WDCNN-CBAM-DUAL exhibits superior performance on datasets characterized by few samples and data imbalance. In comparison to other generative network models, it showcases stronger learning and feature extraction capabilities. Notably, it can effectively accomplish the bearing fault diagnosis task even under conditions of insufficient data samples, unbalanced samples, and noisy data. This indicates its promising applicability in complex industrial environments.