Selective Kernel Network based on Joint Attention Mechanism for Rolling Bearing Fault Diagnosis

Abstract

Deep learning has demonstrated great vitality in various fields in recent years. However, most deep learning models lack kernel size selection capability and feature importance distinction mechanism. Although there have been studies that have considered using the channel attention mechanism to help the automatic selection of kernel size, few researchers have used the spatial attention mechanism to recalibrate the choice of kernel size. Therefore, in this paper, a selective kernel network, based on joint attention mechanism (SKNJA), is proposed for the first time to further improve the ability of the kernel size selection for the convolutional neural network. The network incorporates the newly designed selective-kernel joint attention module (SKJAM), which consists of a series connection of selective-kernel channel attention module (SKCAM) and selective-kernel spatial attention module (SKSAM). Both SKCAM and SKSAM use two parallel branches with different kernel sizes. The SKNJA is applied in the field of mechanical fault diagnosis. The Case Western Reserve University bearing dataset is used to validate the proposed model. Experiment results indicate that the SKSAM can enhance fault-related kernel size automatically. And it can be further improved by integrating SKCAM. Compared with four advanced deep learning methods, the SKNJA performs best in the bearing fault diagnosis task.