A two-stage graph spatiotemporal model with domain-class alignment for fault diagnosis under multi-source long-tailed distributions

In practical engineering, monitoring data often follow multi-domain long-tailed distributions (MDLT), where label imbalance, domain shift, and cross-domain label divergence are deeply intertwined, posing significant challenges for intelligent fault diagnosis. To address these, we propose a novel two-stage decoupled graph spatiotemporal network guided by a balanced domain-class alignment loss. This framework introduces domain-class pairs and constructs a domain-class transferability graph using distance metrics. Building upon this, we propose an intensified Balanced Domain-Class Distribution Alignment (iBoDA) loss, which strengthens the similarity of intra-domain and cross-domain features within the same class while attenuating the similarity across different classes. This loss function calibrates and aligns domain-class distributions in imbalanced datasets, enhancing generalization for out-of-distribution samples. Furthermore, we design a multi-source fusion two-stage decoupled graph spatiotemporal network to extract domain-invariant, noise-resistant representations by capturing multi-dimensional spatiotemporal dependencies. Extensive experiments on three MDLT datasets, benchmarked against 15 state-of-the-art algorithms, validate the method's effectiveness, robustness, and computational efficiency in addressing MDLT challenges in industrial fault diagnosis.