Knowledge-Enhanced Industrial Fault Detection via FMEA Graph Learning and Cross-Modal Feature Alignment

The increasing complexity and integration of industrial equipment has elevated the importance of reliable operation, making accurate fault detection more critical than ever. While modern detection models predominantly rely on neural networks to learn patterns from operational data, they often overlook valuable, structured expertise from design and maintenance. To effectively integrate data-driven modeling with domain knowledge, this article introduces a novel fault detection framework that incorporates failure mode and effects analysis (FMEA) knowledge with operational time-series data. A task-oriented FMEA knowledge graph is first constructed to organize expertise from FMEA documents, which is then encoded using a newly designed FMEA knowledge graph convolutional network (FKGCN). The FKGCN employs type-aware neighbor sampling and adaptive category weighting to extract fault-related features. Following this, a generative adversarial network-based mapping with contrastive supervision aligns operational data features with FMEA knowledge embeddings. A fault-mode cross-attention module is then applied for feature fusion, enabling effective fault detection. Experimental results on the JNU Bearing, SWaT, and Metro Door datasets demonstrate consistent improvements in precision, recall, and F1score. The framework exhibits notable robustness and practical effectiveness, particularly under data-scarce or noisy conditions.