Cross-Modal Commonality Graph Matching Frame: A Fault Diagnosis Method for Multimode Process

Abstract

Fault diagnosis in complex industrial systems, such as large-scale chemical plants and advanced manufacturing lines, is critically challenged by data acquired in multiple and varying operating modes. These modal shifts, driven by different production demands and process parameters, often obscure fault signatures and undermine the reliability of conventional diagnostic models. Despite these operational variations, it is observed that variables within such multimode processes often exhibit consistent correlations. Consequently, the same fault can manifest itself as a shared propagation characteristic across different modes. To leverage these invariant features for robust fault identification, this paper proposes a novel Cross-Modal Commonality Graph Matching (CMCGM) framework. Our approach first extracts the shared causal structures of a specific fault from data across multiple modalities to construct a modality-independent “commonality fault dictionary,” which captures the essential fault signature. By transforming the diagnosis task into a local subgraph matching problem against this dictionary, the CMCGM method achieves robust and accurate fault identification while circumventing the computational costs of online feature propagation. The effectiveness and superior performance of the proposed framework are validated through extensive experiments on the Tennessee-Eastman (TE) process, a widely-recognized benchmark for complex chemical systems. Note to Practitioners—This work addresses the challenge of diagnosing faults in industrial processes where data comes from multiple working modes. In such systems, faults often propagate in similar ways across different modalities, but traditional methods struggle to leverage these shared patterns effectively. Our solution, the Cross-Modal Commonality Graph Matching (CMCGM) framework, identifies and matches these common fault signatures across modalities, enabling faster and more reliable fault diagnosis without requiring extensive computational resources. The method works by analyzing how faults propagate across different data sources, extracting their shared structural features, and building a reusable fault reference library. Instead of treating each data stream separately, it compares faults in a way that reduces redundant computations. This makes the diagnosis process more efficient while maintaining reliability. One current limitation is that it requires sufficient historical fault records to build an effective reference model. Future improvements could focus on making the system more responsive to newly emerging faults and integrating it with real-time monitoring systems. Potential applications include predictive maintenance in factories, power plants, and other complex industrial systems where early and accurate fault detection is critical.