Attention gate guided multiscale recursive fusion strategy for deep neural network-based fault diagnosis

Rolling bearings are crucial for ensuring the safe and stable operation of electromechanical systems. Although deep learning has been widely used in fault diagnosis of rolling bearings, it is unable to accurately diagnose faults when the system operates under multiple working conditions. Therefore, it is essential to conduct research on fault diagnosis of rolling bearings under multiple working conditions to ensure the reliable operation of electromechanical systems. The potential features related to working conditions may be reflected in the different layers of the deep neural network (DNN). However, information loss during the process of layer-by-layer feature extraction may result in the loss of potential features related to changes in working conditions, which in turn affects the fault diagnosis results. This study focused on developing a multiscale recursive fusion strategy for a DNN by designing a new attention model with a lower computational burden. The proposed multiscale recursive fusion strategy guided by the attention mechanism can help correctly characterize the potential features related to variations in working conditions by allocating more attention to useful information and less attention to useless information on the adjacent layers of the DNN. Experimental tests for fault diagnosis of rolling bearings verified that the proposed method is superior to existing methods for fault diagnosis when the system is operated under multiple working conditions.