SIGTN: A novel structural Infomax Graph Transfer Networks for rotating machinery fault diagnosis in cross-condition and cross-equipment scenarios

Abstract

Graph-based networks have proven effective in node classification for diagnosing faults in rotating machinery. However, current graph neural networks often prioritize local information over global influences, hindering cross-graph transfer diagnosis in unlabeled graphs. To address these challenges, we propose SIGTN (Structure Infomax Graph Transfer Network), a novel algorithm for cross-graph diagnosis. Initially, raw and corrupted graph data is individually fed into the feature extractor, enhancing learned node representations to capture global structural properties by maximizing local-global mutual information. The node classifier then predicts labels based on these representations. During training process, both the feature extractor and node classifiers are trained concurrently to minimize cross-entropy loss for labeled nodes. Additionally, a conditional domain adversarial network alleviates distributional disparities between source and target domain graphs. Finally, experimental validation across various datasets demonstrates SIGTN's effectiveness in handling cross-graph transfer across different rotation speeds, loads, and equipment.