Abstract

Despite their demonstrated effectiveness in modeling sensor interaction networks for industrial fault diagnosis, graph neural networks (GNNs) still encounter two key limitations: black-box operation that lacks transparency in fault identification and propagation analysis, and unreliable attention mechanisms whose weights fail to faithfully reflect the genuine relevance of sensors or their interactions. To tackle these challenges, we put forward the Stochastic Mask Causal Graph Network, a novel framework that integrates a learnable stochastic masking mechanism guided by the information bottleneck principle. Unlike conventional attention-based or post-hoc approaches, our method automatically suppresses label-irrelevant graph components while preserving causally relevant structures, thereby providing faithful inherent interpretability without biased assumptions and effectively removing spurious correlations to enhance generalization. Comprehensive experiments on realistic complex industrial system datasets demonstrate that the proposed method achieves superior diagnostic accuracy and enhanced interpretability compared with existing advanced approaches.

IPC Classification

Keywords