Contrastive Learning for Fault Diagnosis Enhanced by System-Level Graph Representation

System-level fault diagnosis are key to improving the reliability of interconnected networks. However, traditional system-level diagnosis methods are often limited by the number of the regularity and the smallest node degree of an interconnected network. To address these challenges, we propose a graph-enhanced multi-view contrastive learning fault diagnosis algorithm (SGEMC) based on PMC and MM* model. Our method introduces a dual-view enhancement technique that integrates the diagnosis results of neighboring nodes and their relationship with target node, thereby enriching the output. Additionally, we incorporate an incremental contrast mechanism that allows the target node to be contrasted across multiple dimensions, including node-subgraph, node-node, and subgraph-subgraph contrasts, which significantly improves the model’s diagnosis accuracy. Finally, we propose a fault scoring method that leverages multiple rounds of positive and negative instance samplings to effectively differentiate normal nodes from faulty ones. Experimental results demonstrate that SGEMC outperforms the existing models such as CoLA, Sub-CR, and SL-GAD in terms of accuracy, precision, recall, F1 score, and AUC.