A novel two-stage variables contribution analysis method toward explainable graph convolutional network-based industrial fault diagnosis

• Published in: Engineering applications of artificial intelligence Vol. 143; p. 110071
• Main Authors: Xu, Jiamin, Mo, Siwen, Chen, Zhiwen, Ke, Haobin, Jiang, Zhaohui
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.03.2025
• Subjects: Contribution analysis; Deep neural network; Explainability; Fault diagnosis; Graph convolutional network; Fault diagnosis; Deep neural network; Graph convolutional network; Explainability; Contribution analysis
• ISSN: 0952-1976
• DOI: 10.1016/j.engappai.2025.110071

In industrial systems, fault diagnosis methods play a crucial role for ensuring their safe and stable operations. Among these methods, graph convolutional network (GCN)-based fault diagnosis stands out for its ability to capture complex relationships within data. However, conventional GCN-based methods regard fault diagnosis solely as a classification problem, terminating the diagnostic process once classification decisions are made. This approach overlooks the analysis of the contributions of various relevant variables to the diagnostic results, leading to a lack of explainability and reliability, which hinders their industrial adoption. To address this challenge, the paper proposes a two-stage variables contribution analysis method called graph self-reproductive non-dominated sorting genetic algorithm (Graph-SRGA). The method sequentially conducts neighbor-stage and sensor-stage contribution analysis for the explainability of GCN-based methods. Specifically, two objective functions are formulated as evaluation indicators, and the selection process is encoded into a coding string. The proposed SRGA algorithm, integrating a self-reproductive strategy into the conventional non-dominated sorting genetic algorithm, identifies the required influential set. Experimental validation from various perspectives, using application cases in the high-speed train traction system and the Tennessee Eastman process, demonstrates that the proposed method offers a distinctive solution for the contribution analysis of GCN-based fault diagnosis methods.