Multi-Relational Fusion Graph Convolution Network with Multi-Scale Residual Network for Fault Diagnosis of Complex Industrial Processes

• Published in: IEEE transactions on instrumentation and measurement p. 1
• Main Authors: Han, Yinghua, Tuo, Siwei, Li, Yuan, Zhao, Qiang
• Format: Journal Article
• Language: English
• Published: IEEE 04.01.2024
• Subjects: Complex industrial processes; Convolution; Correlation; Data mining; Fault diagnosis; Feature extraction; graph convolution network (GCN); heterogeneous graph; meta-path graph; multi-scale residual network( MSRN); Residual neural networks; Sensors
• ISSN: 0018-9456; 1557-9662
• DOI: 10.1109/TIM.2024.3350143

Typically, complex industrial processes possess heterogeneous spatial relationships and dynamic temporal correlations due to intervariable coupling and intravariable dynamics. Unfortunately, with the difficulty in modeling heterogeneous relationships and the high computational complexity, existing fault diagnosis methods mainly process heterogeneous relationships as homogeneous relationships, and the influence of temporal correlations is also overlooked, resulting in inappropriate representations and inaccurate fault diagnosis results. To address those issues, this article proposes a multi-relational fusion graph convolution network with multi-scale residual network (MR-MFGCN) for fault diagnosis of complex industrial processes. Initially, a multi-scale residual network (MSRN) is designed to capture dynamic temporal correlations for multisensor data. Then, the extracted features are transformed into a heterogeneous graph, modeling the heterogeneous spatial relationships among multiple sensors. Furthermore, an adaptive multi-relational fusion strategy is proposed to construct a lightweight meta-path graph, reducing the complexity of the graph structure and enhancing the overall quality of the graph. Finally, the meta-path graph is modeled by a graph convolution network, and a coarse-grained graph-level representation is obtained for fault classification. Comprehensive experiments on two industrial processes along with theoretical interpretation validate the effectiveness and reliability of our proposed MR-MFGCN for fault diagnosis of complex industrial processes.