A multi-scale graph-guided dynamic enhanced alignment network for mechanical fault diagnosis considering domain shift and data imbalance

Transfer learning is widely applied in the cross-domain diagnosis of machines, often under the assumption of balanced data. However, machines mostly operate in the normal state in practical industry, and only limited fault samples can be collected before maintenance shutdowns, leading to class imbalance. This imbalance, combined with the inherent domain shift between the source and target domains, makes diagnosis challenging, including the neglect of minority fault classes and misalignment between domains. To address this, a Multi-scale graph- guided Dynamic enhanced alignment network (Mul-Dean) is proposed. Mul-Dean consists of two main components. One component, the Multi-scale Graph Construction Method (MGCM), tackles intra-domain imbalance by learning diverse and distinguishable relationships between classes. The other component, the Dynamic Enhancement Alignment Loss (DEAL) algorithm, reduces the sensitivity to proportion disparities of imbalance by introducing class-contrastive labels. Additionally, the inter-domain distance is proposed to quantify the alignment degree between domains. Experimental results on the bearing and reciprocating compressor datasets showed that Mul-Dean achieved better classification and alignment capability than the other 11 methods.