Graph Contrastive Learning With Adaptive Distribution Transformation for Fault Detection and Isolation Using Limited Labeled Faults

Fault detection and isolation (FDI) in process industries remain challenging due to misdiagnosed dynamic variations, limited fault labels, and complex variable interactions. To address these challenges, an adaptive distribution transformation-graph contrastive learning (ADT-GCL) is proposed to differentiate between global dynamics and local faults, even when labeled faults are limited. To capture variable interactions at both global and local levels, global and local graphs are established, and a local-global learning strategy is introduced. To prevent the few available fault labels from being misidentified as normal, contrastive tasks first cluster data embeddings before mapping them to labels. For fault isolation, an adaptive distribution transformation loss reconstructs fault embeddings as normal data, thereby increasing reconstruction errors. The T2 statistic is then used to identify fault variables based on these errors. Lastly, case studies on a sugar production process and a multiphase flow process (MFP) demonstrate ADT-GCL’s superiority in distinguishing global dynamics from local faults.