AANet: adaptive attention network for rolling bearing fault diagnosis under varying loads

Recently, modern intelligent fault diagnosis algorithms based on deep learning have been widely used to recognize the health state of rolling bearings. However, the constantly varying load in real industry leads to unsatisfactory diagnosis results. How to make the models effectively diagnose the health state of rolling bearings under varying loads is a key issue. In this paper, an Adaptive Attention Network (AANet) is proposed to resolve the issue. That the interference is introduced by the Multi-scale Convolution Module with wide kernels (MCM) at the head of the AANet is the premise for extending the model to other loads. And the Adaptive Attention Modules (AAMs) embedded in the AANet distinguishe state-related features and unrelated features, which enhances the diagnostic ability of the model across loads. In order to verify the effectiveness of the algorithm, experiments have been performed on a public data set. Experimental results show that the average accuracy of this algorithm achieves 0.976, which can effectively recognize the health state of rolling bearings under varying loads, compared to other algorithms.