Cross-Domain Augmentation Diagnosis : An Adversarial Domain-Augmented Generalization Method for Fault Diagnosis under Unseen Working Conditions

Li, Qi and Chen, Liang and Kong, Lin and Wang, Dong and Xia, Min and Shen, Changqing (2023) Cross-Domain Augmentation Diagnosis : An Adversarial Domain-Augmented Generalization Method for Fault Diagnosis under Unseen Working Conditions. Reliability Engineering and System Safety, 234: 109171. ISSN 0951-8320

Text (JRESS-D-22-02067_R1 (1))
JRESS_D_22_02067_R1_1_.pdf - Accepted Version
Available under License Creative Commons Attribution-NonCommercial-NoDerivs.
Download (6MB)

Abstract

Intelligent fault diagnosis based on domain adaptation has recently been extensively researched to promote reliability of safety-critical assets under different working conditions. However, target data may be inaccessible in the model training phase, resulting in the degradation or failure of the diagnosis model. Therefore, this paper introduces a new idea called cross-domain augmentation (CDA) to achieve diagnosis under unseen working conditions, which are frequently occurred in industrial scenarios. To realize this idea, an adversarial domain-augmented generalization (ADAG) method is proposed with domain augmentation via convex combination of data and feature-label pairs. Through adversarial training on multi-source domains and the augmented domain, ADAG enables learning generalized and augmented features, which are proximal representation in the unseen domain, facilitating the generalization ability of the model. Moreover, feature extractor and domain classifier are optimized as adversaries in model training to obtain domain-invariant features, while the fault classifier is trained to identify the features. Extensive experiment studies indicate that ADAG can successfully solve the cross-domain diagnosis problem under unseen working conditions. For SDUST case study, ADAG promotes the model accuracy by 1.44%; while for a more challenging Ottawa case study, it promotes the model accuracy by 5.34%. Moreover, the domain discrepancy is reduced by 4.6%.