A novel deep multi-source domain adaptation framework for bearing fault diagnosis based on feature-level and task-specific distribution alignment. (June 2021)

• Record Type: Journal Article
• Authors: Rezaeianjouybari, Behnoush; Shang, Yi
• Is Part Of: Measurement. Volume 178(2021)
• Publication Date: 2021-06
• Subjects: Multi-source domain adaptation -- Fault diagnosis -- Sliced Wasserstein Distance -- Deep learning; CMD Central Moment Discrepancy -- CNN Convolutional Neural Network -- CORAL Correlation alignment -- CWRU Case Western Reserve University -- DA Domain Adaptation -- FTD-MSDA Feature-level and task-specific distribution alignment multi-source domain adaptation -- HoMM Higher-order Moment Matching -- KAT Konstruktions- und Antriebstechnik -- KL Kullback-Leibler -- MMD Maximum Mean Discrepancy -- MSDA Multi-source domain adaptation -- OT Optimal Transport -- PHM Prognostics and Health Management -- RKHS Reproducing Kernel Hilbert Space -- SWD Sliced Wasserstein distance -- TL Transfer Learning -- UDA Unsupervised Domain Adaptation -- WD Wasserstein Distance; Weights and measures -- Periodicals; Measurement -- Periodicals; Measurement; Weights and measures; Periodicals; 530.8
• DOI: 10.1016/j.measurement.2021.109359
• Languages: English
• ISSNs: 0263-2241

Abstract

Highlights: A novel multi-source domain adaptation framework for fault diagnosis is proposed. Local feature extractors align source and target feature distributions by MMD loss. Task-sepecific decision boundries are made based on SWD metric. Abstract: In recent years, deep learning has been extensively applied for intelligent fault diagnosis systems. Most of the developed algorithms ignore the domain shift problem and assume distribution of the training data, known as the source domain, is similar to that of testing data, denoted as the target domain. However, in real-world applications, this assumption is not necessarily true. In current work, a novel multi-source domain adaptation deep learning framework for fault diagnosis of rotary machinery is proposed, which aligns the domains in both feature-level and task-level. The proposed Feature-level and Task-specific Distribution alignment multi-source domain adaptation (FTD-MSDA) framework transfers the knowledge from multiple labeled source domains into a single unlabeled target domain by reducing the feature distribution discrepancy between the target domain and each source domain. Sliced Wasserstein discrepancy is utilized to shape task-specific decision boundaries. Besides, the model can be easily reduced to a single-source domain adaptation problem. Also, the model can be readily updated to unsupervised domain adaptation problems in other fields such as image classification and image segmentation. The experimental verification results show the superiority of the proposed framework over state-of-the-art multi-source domain-adaptation models.