Physics-Informed hierarchical reasoning (PIHR): A structured framework for generalizable fault diagnosis in industrial thickeners

Abstract

In the minerals processing industry, ensuring the safe and stable operation of thickeners is a primary objective. However, their complex dynamic characteristics present challenges for conventional monitoring systems in providing early warnings and diagnosing faults, especially under novel operating conditions. Existing diagnostic methods generally lack the generalization capability for unseen faults. They can exhibit limitations in resolving diagnostic ambiguity, identifying concurrent faults, or distinguishing between process anomalies and sensor failures, which can lead to delayed warnings. To address this, this study proposes a Physics-Informed Hierarchical Reasoning (PIHR) framework for thickener diagnostics, centered on a ’Decomposition-Analysis-Synthesis’ workflow. The architecture first decomposes the sensor data stream into distinct features intended to reflect different physical phenomena, such as slowly evolving bed compaction trends versus rapidly changing flocculation chemistry instability. Subsequently, specialized analytical paths, including a Transformer, assess the morphology of these trends to distinguish between, for example, linear and exponential growth, while other paths quantify the system’s degree of instability. Finally, the model performs a dynamic weighted fusion of this evidence based on physics-inspired heuristics (e.g., ’a trend is of greater significance during stable conditions’) to assess the operational risk. PIHR’s performance was evaluated across multiple dimensions. Quantitatively, it achieved a Macro F1-score of 0.885 in zero-shot diagnosis on fault types unseen during training, outperforming the baseline models. It also increased the fault warning lead time by over 12 min on average, while maintaining a True Positive Rate (TPR) above 0.91 and a False Positive Rate (FPR) below 0.015. Qualitatively, case studies illustrated the model’s ability to resolve diagnostic ambiguity, handle concurrent faults, and distinguish between process and sensor failures. Moreover, a physical consistency analysis indicated that the architecture extracts multivariate dynamic signatures from observational data that correlate with different physical states. Finally, its performance was evaluated in a 60-day continuous industrial deployment, where the system delivered 14 valid early warnings with no missed detections and assisted in averting a rake-bogging incident.