Dynamic Transfer Multi-Source Domain Adaptation Network for Fault Diagnosis

Presently, the majority of research in the field of multi-source domain adaptation is centered on the development of a network that is universally applicable across various domains, with its parameters kept static. However, such a network struggles to resolve the inconsistencies that arise when dealing with multiple domains, as there are not only differences between the source and target domains but also among the various source domains themselves. Hence, this study introduces a novel approach called the dynamic transfer multi-source domain adaptation network (DTMDA), which features model parameters that adjust in response to the input data. The DTMDA is composed of two main parts: a feature extractor equipped with a dynamic transfer module (DTM) and a pair of classifiers incorporating an attention mechanism. The feature extractor, enhanced by the DTM, is capable of transforming multiple source domains into a single-source domain, facilitating the alignment of data distributions between source and target domains. This is achieved by aligning the target domain with any subset of the multi-source domains. Additionally, the classifiers are augmented with an attention mechanism to optimize the utilization of knowledge from multiple source domains for aligning data distributions. The experimental findings confirm that DTMDA demonstrates exceptional performance in diagnosing bearing faults across various domains.