Enhanced bearing fault diagnosis under variable working conditions using ShuffleNet V2 with attention and adaptive normalization

Abstract

With the increasing complexity of industrial equipment, bearing fault diagnosis faces challenges such as insufficient feature extraction capability and poor adaptability, highlighting the urgent need to improve diagnostic accuracy and efficiency. This study proposes a bearing fault diagnosis method based on the Shu-CBAM-Ada model, which integrates the Convolutional Block Attention Module (CBAM) and the Adaptive Batch Normalization (AdaBN) optimization strategy to enhance the performance of the ShuffleNet V2 convolutional neural network. The proposed method is validated using bearing data sets from Case Western Reserve University and a power plant in Northwest China. The results demonstrate that embedding the CBAM module into the third layer of ShuffleNet V2 significantly improves feature extraction and information fusion capabilities. In addition, the AdaBN optimization algorithm enhances the parameter stability, generalization ability, and computational efficiency of the ShuffleNet V2 model. The combination of CBAM and AdaBN enables the Shu-CBAM-Ada model to achieve a favorable balance between classification accuracy and generalization performance, demonstrating strong robustness and adaptability. Furthermore, the Shu-CBAM-Ada-based fault diagnosis method exhibits high accuracy and fast inference speed in practical applications, outperforming other deep learning approaches.