Domain Generalization for Cross-Domain Fault Diagnosis: an Application-oriented Perspective and a Benchmark Study

Most data-driven methods for fault diagnostics rely on the assumption of independently and identically distributed data of training and testing. However, domain shift between the phases of training and testing is common in practice. Recently, domain generalization-based fault diagnosis (DGFD) has gained widespread attention for learning fault diagnosis knowledge from multiple source domains and applying it to unseen target domains. This paper summarizes the developments in DGFD from an application-oriented perspective. Firstly, basic definitions of DGFD and its variant applications are formulated. Then, motivations, goals, challenges and state-of-the-art solutions for different applications are discussed. The limitations of existing technologies are highlighted. A comprehensive benchmark study is carried out on eight open-source and two self-collected datasets to provide an understanding of the existing methods and a unified framework for researchers. Finally, several future directions are given. Our code is available at https://github.com/CHAOZHAO-1/DG-PHM.