Identifying Bearing Faults Using Multiscale Residual Attention and Multichannel Neural Network

Abstract

To solve the problem of the low signal-to-noise ratio and fault features can only be extracted from a single scale of traditional convolutional neural network (CNN) in vibration-based bearing fault diagnosis, this paper proposes a new multi-scale residual attention and multi-channel network (MSCNet), which can effectively reduce noise and fully extract meaningful features from different scales of the signal. The proposed method combines filtering methods to remove redundant parts and noise in the signal, and multiple filtered signals are input into the proposed CNN. The proposed CNN can perform multi-scale feature extraction on the signal and make the network focus on valuable information in the feature through the residual attention mechanism. Therefore, MSCNet achieves better performance. Experimental results on the published bearing datasets at the Paderborn University and the University of Ottawa show that MSCNet achieves 94.28% and 96.6% accuracy in strong noise environments, while outperforming five state-of-the-art (SOTA) networks in terms of accuracy.