Intelligent fault diagnosis of rotating machinery using a multi-source domain adaptation network with adversarial discrepancy matching. (March 2023)

• Record Type: Journal Article
• Authors: Liu, Shaowei; Jiang, Hongkai; Wu, Zhenghong; Yi, Zichun; Wang, Ruixin
• Is Part Of: Reliability engineering & system safety. Volume 231(2023)
• Publication Date: 2023-03
• Subjects: Fault diagnosis -- Rotating machinery -- Multi-source domain adaptation -- Self-attention mechanism -- Discrepancy matching technique; 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.2022.109036
• Languages: English
• ISSNs: 0951-8320

Abstract

Highlights: The self-attention mechanism adjusts the weight factors of multi-source domains. The discrepancy matching technique aligns different feature distributions. An adversarial classifier training method raises the adaptation transferability. Abstract: In the health management of modern rotating machinery, domain adaptation is an effective method to solve the diagnostic problems of insufficient labeled signals and poor generalization performance. In engineering scenarios, obtaining signals from various source domains can ensure abundant feature information and contribute to diagnostic ability improvement compared with learning from a single source domain. This paper presents a deep multi-source adversarial discrepancy matching adaptation network (MADMAN) for enhancing the accuracy of cross-domain intelligent diagnosis. Firstly, the proposed method makes use of the generalization knowledge learned from multiple domains to diagnose the unknown task, and adaptively adjusts the weight factors of multiple source domains utilizing the self-attention mechanism. Secondly, to better alleviate the domain shift phenomenon between different domains, the discrepancy matching technique is applied to dynamically align the feature distributions of different domains. Thirdly, an adversarial classifier training method is incorporated to raise the transferability by considering the decision boundary of specific tasks. The proposed method is verified by extensive experiments using two bearing datasets, and the superiority of the presented approach is demonstrated by comparison with advanced methods.