Abstract:

Given that it is difficult to accurately diagnose rolling bearing faults in a strong noise environment， a deep residual shrinkage network is proposed with pooling fusion for rolling bearing fault diagnosis. Firstly， the proposed method employs residual connections to avoid the risk of gradient vanishing or explosion due to excessive network depth. Then， the approach uses multi?scale pooling feature fusion to extract more comprehensive local features from vibration signals， and an attention mechanism to automatically derive the optimal threshold of a soft threshold function. Finally， the network is trained through labeled data to achieve the accurate fault diagnosis of rolling bearings in a strong noise environment. Experiment results show that the deep residual shrinkage network based on pooling fusion can outperform the traditional models in diagnosing rolling bearing faults under noise conditions with different SNRs.

Key words: deep residual shrinkage network, pooling fusion, attention mechanism, rolling bearing, fault diagnosis