A cross-domain intelligent fault diagnosis method based on feature transfer with improved Inception ResNet for rolling bearings under varying working condition

Abstract

With the popularity of smart manufacturing, data-driven fault diagnosis methods for rolling bearings have been extensively studied in recent years. Existing rolling bearing fault diagnosis method has problems such as low precision and poor generalization ability when diagnosing multi-working condition bearings. In actual industrial scenarios, bearings usually operate under different operating conditions, causing differences in the probability distribution of the vibration data. Considering existing problem, this article proposes a diagnostic method of Inception ResNet Network (TL-IResnet) based on feature transfer learning. First, we utilize the Inception network to derive multiple scales of features from the original vibration signal. This enhances the capacity for feature expression in the model, and addresses the over-fitting issue in the deep model. Then the residual network is used to carry out deep learning on the fused multi-scale features to improve the residual network's ability to pay attention to important information, the self-attention mechanism is integrated into the residual network, and a new residual network structure is proposed. Finally, the maximum mean difference (MMD) is employed in output layer to measure the degree to which the probability distribution differs between the source and target domains to enhance the ability of model to transfer knowledge and complete the task of diagnosing the bearing of a machine. TL-IResnet is evaluated using the bearing dataset from Case Western Reserve University (CWRU) and the gearbox dataset from Southeast University. Experimental results demonstrate that TL-IResnet has a strong capacity to generalize information in addition to a high degree of accuracy under different conditions of operation, and has certain advantages over existing fault diagnosis methods.