STCGNN: Spatio-Temporal Causal Graph Neural Network for Interpretable Fault Diagnosis
Year: 2026
Type: conference-paper
Abstract: In modern complex systems, increasing equipment complexity makes early and accurate fault diagnosis critical for safety and operational continuity. Traditional methods rely on expert experience and fail to capture hidden sensor correlations, while deep learning and graph neural networks (GNNs) improve diagnostic performance but focus solely on accuracy, lacking a causal analysis of fault progression (e.g., identifying critical time steps). To address these limitations, this paper proposes a Spatio-Temporal Causal Graph Neural Network (STCGNN). STCGNN introduces three key innovations: (1) a unified framework that jointly performs fault diagnosis and causal analysis on multi-variate time-series data; (2) a chain-based graph construction method that optimizes the computational overhead compared to fully connected graph approaches; and (3) a time-step causal analysis module that enhances interpretability by pinpointing the most critical windows during fault evolution. Experiments on multiple public benchmark datasets demonstrate that STCGNN outperforms baselines in diagnostic accuracy while providing causal interpretability. This work represents a step toward efficient and explainable fault diagnosis in complex engineering scenarios.
Authors: Mingxuan Shi, Leiming Chen
Institutions: Beihang University, China University of Petroleum, East China
Language: English
Subfield: Control and Systems Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7202390413