Enhanced fault diagnosis of rolling bearings using attention-augmented separable residual networks

Abstract

Recently, with the quick development of industrial equipment automation, convolutional neural networks (CNN) have been broadly applied to the intelligent fault diagnosis of rolling bearings. In order to solve the problems of gradient vanishing, gradient explosion, and too many training parameters in deep convolutional networks that lead to low diagnostic accuracy and training efficiency of network models, a bearing fault diagnosis method based on an attention-augmented separable convolutional residual network (ASResnet) is proposed. First, the bearing vibration signal data is converted into a 2D grayscale map as an input to the network. Then, residual blocks with separable convolutions were constructed, allowing automatic learning of high-level representations from input images by stacking multiple separable convolutional residual blocks. Separable convolution effectively reduces the number of network parameters and improves computational speed. Finally, a feature extractor based on the Convolutional Block Attention Module (CBAM) is constructed so that the network focuses on the key feature regions to further improve the diagnostic performance. Validation was conducted using a Case Western Reserve University bearing dataset and three actual engineering datasets of production equipment in a cement plant. The experimental results show that ASResnet is able to improve the diagnostic accuracy and reduce the network training time of the CWRU dataset, and it also obtains a high accuracy rate in fault diagnosis for engineering applications in the cement production equipment industry.