Rolling Bearing Fault Diagnosis Method Based on Non-dimensionlity Reduction Attention Mechanism with Aggregate Residual Network

Abstract

Aimed at the difficulties of extracting fault features, poor model generalization, and low diagnostic accuracy under noisy environments in traditional bearing fault diagnosis algorithms, a fault diagnosis method, which combines a portable non-dimensionlity reduction attention mechanism with deep residual neural network, was proposed. This method uses a non-dimensionality reduction attention mechanism to redistribute the weights of the feature maps generated by residual block. Simultaneously, local cross-channel communication methods rarher than global cross-channel communication methods are adopted in achieving the effect of non-dimensionality reduction and adaptively learning the attention scores of neighboring channels. Case Western Reserve University's bearing fault datasets were used to verify the method. Experiments results show that the residual network fused with non-dimensionality reduction attention mechanism can accurately identify faulty bearing samples disturbed by noise under different loads. Specifically, the diagnosis accuracy under 12 dB signal-to-noise ratio is 99.5%, with strong anti-noise performance and certain generalization performance.

Keywords

• fault diagnosis
• residual network
• non-dimensionality reduction attention
• deep learning