A novel sub-label learning mechanism for enhanced cross-domain fault diagnosis of rotating machinery. (September 2022)

• Record Type: Journal Article
• Authors: Deng, Minqiang; Deng, Aidong; Shi, Yaowei; Liu, Yang; Xu, Meng
• Is Part Of: Reliability engineering & system safety. Volume 225(2022)
• Publication Date: 2022-09
• Subjects: Domain adaptation -- Transfer learning -- Statistical moment matching -- Adversarial training -- Fault diagnosis -- Rotating machinery; Reliability (Engineering) -- Periodicals; System safety -- Periodicals; Industrial safety -- Periodicals; Fiabilité -- Périodiques; Sécurité des systèmes -- Périodiques; Sécurité du travail -- Périodiques; 620.00452
• DOI: 10.1016/j.ress.2022.108589
• Languages: English
• ISSNs: 0951-8320

Abstract

Highlights: A sub-label learning mechanism is proposed to enhance domain adaptability. Traditional domain adaptation ignores the correlation between samples. Exploring structural connectivity can effectively reduce mismatches. The proposed method is insensitive to trade-off parameters. Abstract: Deep Domain Adaptation (DDA), which transfers the knowledge learned in the source domain to the target domain, has made remarkable achievements in intelligent fault diagnosis. However, the existing DDA technology mainly focuses on eliminating cross-domain distribution discrepancies, while ignoring the exploration of intra-domain distribution characteristics, resulting in unsatisfactory performance in complex scenarios. To overcome this drawback, a novel sub-label learning mechanism (SLLM) is proposed in this paper, which exploits the structural connectivity of the original sample space to guide distribution alignment, thereby enhancing domain adaptability. Specifically, SLLM consists of two parts. First, the unsupervised target domain is annotated with sub-labels according to the probability distribution of the sample space, so that similar data can be recognized. Then, the intra-domain connectivity of the associated data is preserved during feature matching. In this way, samples belonging to the same category can be aggregated together in the feature space, and mismatches can be effectively alleviated. Extensive experiments on two datasets indicate that the proposed SLLM can significantly improve the domain adaptability of traditional DDA methods.