Knowledge graph augmented meta-learning with condition-sensitive pseudo-labeling for semi-supervised fault diagnosis under multiple working conditions

Owing to the low occurrence rate of faults and the high cost of manual labeling, industrial scenarios under multiple working conditions suffer from both sample scarcity and pronounced distribution shifts. To address this challenge, a semi-supervised fault diagnosis method termed Knowledge Graph-Augmented Meta-Learning with Condition-Sensitive Pseudo-Labeling (KG-CAML-SS) is proposed. Within a meta-learning framework, KG-CAML-SS constructs a domain knowledge graph by integrating equipment structural relationships with process variable correlations. The knowledge graph is embedded via a relational graph convolutional network combined with global mean pooling to obtain task-level structural representations, thereby enhancing the model’s ability to capture variations across multiple working conditions. Subsequently, a metric-based condition-sensitive network is designed to evaluate similarity differences between unlabeled samples and support-set prototypes relative to their nearest and second-nearest class centers, based on which a confidence-ratio criterion is formulated for adaptive high-reliability pseudo-label selection. Furthermore, an adaptive prototype fusion mechanism is introduced to dynamically fuse each class prototype with its corresponding high-confidence pseudo-labeled features. Finally, the refined prototypes are used to classify query samples, enabling accurate fault diagnosis across unseen working conditions. Experimental results on two real industrial process datasets demonstrate that the proposed KG-CAML-SS method achieves significant diagnostic performance and generalization capability under multiple working condition scenarios.