A Knowledge-Guided Hierarchical Information Fusion-Based Quality-Related Fault Diagnosis Method for Complex Industrial Processes

Complex industrial processes exhibit characteristics of multi-unit coupling and multi-level integration, whose information interaction exists between the device level and cyber level, and the data present the characteristics of spatiotemporal multi-scale distribution at different levels, which bring significant challenges for traditional quality-related fault diagnosis. Furthermore, existing fault diagnosis methods rarely consider to integrate mechanism knowledge with process data, potentially leading to interpretability deficiencies in process modeling. To address these challenges, we propose a knowledge-guided hierarchical information fusion-based quality-related fault diagnosis method. First, a graph structure learning module is developed to discover complete posterior knowledge, which converts the dependencies among variables into directed graphs. This module incorporates mechanism knowledge as learning bias while considering the associations between process variables and quality indicator, thereby constructing knowledge-guided quality-related graphs. Second, temporal information extractors (TIEs) are embedded into the graph model to capture spatiotemporal multi-scale features at different hierarchical levels while preserving original spatial relationships. Then, a hierarchical information fusion strategy that perceives the interactions between device level and cyber level is developed. The high-level information is transmitted to the device level, and the quality-related features are extracted to complete the fault detection. Finally, after fault detection, variable oblivion contribution plot based on TIEs is adopted to identify fault variables. The effectiveness of the proposed method is validated through comprehensive experiments on an actual hot strip mill process.