A comprehensive survey on domain adaptation for intelligent fault diagnosis

Deep learning-based intelligent fault diagnosis methods have typically been developed under the assumption that an abundant and diverse set of training samples and labels is available. Thus, it is crucial to develop models capable of generalizing effectively to distributions characterized by limited samples and insufficient labels. The transfer of knowledge from semantically related but distributionally different source domains has been recognized as an effective approach; however, discrepancies between distributions may result in negative transfer issues. Domain adaptation (DA), as a prominent research area within transfer learning, has been extensively studied to enhance generalization performance on target tasks. In this survey, the various concepts, formulations, algorithms, and applications of DA in industrial fault diagnosis are thoroughly reviewed. Broader DA solutions are covered, including (a) metric learning, adversarial adaptation, reconstruction, and generation within a homogeneous setting, and (b) source-free domain adaptation, domain generalization, partial domain adaptation, open-set domain adaptation, and universal domain adaptation within a heterogeneous setting, all of which extend beyond the traditional divisions of semi-supervised and unsupervised learning. This survey allows researchers to quickly and comprehensively grasp the research foundation, current status, theoretical limitations, and under-explored directions in the field, thereby facilitating the achievement of universally applicable methods in diverse industrial scenarios.