Dual-weight attention-based multi-source multi-stage alignment domain adaptation for industrial fault diagnosis

Abstract

Abstract Cross-domain fault diagnosis is crucial for industrial applications with various and unknown operating conditions. However, due to the significant differences in the distribution of features in multiple source domains, it may lead to mutual interference of features between different domains and reduce the accuracy of diagnosis, which is a problem not considered by most current researches. In addition, most of the existing methods focus only on the extraction of low-frequency global information and cannot adequately deal with high-frequency local information. Consequently, this paper provides a multi-stage processing integrated dual-weight attention-based multi-source multi-stage aligned domain adaptation (DAMMADA) method. Global fault features that are shared by various subdomains are extracted by three domain-specific feature extractors from various domains. In a local feature extractor, the dual-weight attention module not only uses shared weights to aggregate local information, but it also uses contextual weights to improve local features. In terms of loss handling, multiple pseudo-labels are used to reduce the loss of the local maximum mean discrepancy (LMMD) in order to learn the domain-invariant characteristics after improving the high-frequency and low-frequency information extraction. To modify the classification boundaries, the pseudo-labels' mean square errors (MSE) are combined. Comprehensive experiments were carried out on two platforms for fault diagnosis of SCARA robots and bearings respectively, and the results demonstrated that DAMMADA is superior to other methods in terms of accuracy and its ability to suppress negative migration for cross-domain tasks.