Cross-Domain Fault Diagnosis for Rotating Machines with Multi-Scale Domain Adaptation

Abstract

Transfer learning (TL), especially domain adaptation (DA), has greatly enhanced the cross-domain fault diagnosis of rotating machines. However, the existing methods based on feature alignment at a single scale are still inadequate for complex cross-domain generalization, and thus have much room for improvement. Therefore, this work proposed a multi-scale domain adaptation network (MSDAN) to achieve representation alignment with multiple scales. By minimizing the uniquely designed combined mean maximum discrepancy (CoMMD) metrics, MSDAN is able to learn more domain-invariant representations on multi-scale branches. The case study that learns multi-scale domain adaptation (MSDN) with vibration signals of cross-domain bearings fully validates the feasibility of this method. Comparison with state-of-the-art methods also shows the necessity and advantages of simultaneous domain adaptation on multi-scale representations.