Bidirectional Shrinkage Gated Recurrent Unit Network With Multiscale Attention Mechanism for Multisensor Fault Diagnosis

Abstract

Fault diagnosis is of critical significance to intelligent manufacturing, and data-driven methods have been successfully explored in fault diagnosis. However, in actual industry scenarios, the collected signals are not only contaminated by strong background noise caused by equipment aging, human interference, and environmental disturbances but also exhibit complicated nonstationary characteristics. Therefore, a bidirectional shrinkage gated recurrent unit network with a multiscale attention mechanism (BiSGRU-MAM) is proposed for multisensor fault diagnosis in this article. In particular, the bidirectional shrinkage gated recurrent unit (GRU) that combines GRU and soft thresholding denoising strategy is designed to adaptively filter out the noise-related feature information. Besides, a multiscale feature learning strategy that consists of multiscale dilated convolution and multiscale attention mechanism is established to learn discriminative multiscale features from nonstationary mechanical signals. The proposed BiSGRU-MAM is evaluated through extensive experiments on multisensor datasets. Compared with some data-driven fault classification methods, the BiSGRU-MAM achieves significantly better diagnostic accuracies with 99.85%, 99.79%, 99.84%, and 99.78% in the four subdatasets, respectively. In addition, under noisy and complex working conditions, the experimental results validated that the BiSGRU-MAM has excellent antinoise performance and multiscale feature learning ability.