Universal Source-Free Knowledge Transfer Network for Fault Diagnosis of Electromechanical System With Multimodal Signals

Recently, deep learning (DL) has found widespread applications in electromechanical systems due to its powerful feature extraction capabilities. However, DL-based intelligent fault diagnosis (IFD) models encounter challenges including inadequate data privacy protection, limited generalization performance, and insufficient use of multimodal data in real-world industrial settings. In this study, a novel universal source-free knowledge transfer network (USKTN) is proposed for the IFD of electromechanical systems. First, representative features of fault classes are effectively learned from multiple modalities to ensure that comprehensive fault information is obtained. Second, unlabeled target samples are labeled using a high-confident pseudo-label strategy. Specifically, a self-supervised deep clustering mechanism and an uncertainty-estimated pseudo-labeling filtering mechanism are developed to construct a high-confidence pseudo-label dataset. Finally, a novel mixed loss function is designed to train the diagnostic model, which not only addresses the noise interference issue caused by low-quality pseudo-labeled samples, but also effectively resolves the data distribution discrepancy, achieving knowledge transfer from the source domain to the target domain. The effectiveness of the proposed USKTN is verified through experiments on two different datasets of electromechanical systems. The experiments indicate that the proposed USKTN is capable of accurately diagnosing unlabeled target domain data without requiring any source samples.