Cognitive Supervisory Control with LLM Reasoning Agent for Fault-Tolerant Process Systems: A Digital Twin Perspective

Fault-tolerant control systems deployed in manufacturing and process industries must maintain output regulation under actuator degradation and parametric drift while producing operator-readable intervention records. This paper proposes a three-layer cognitive supervisory control framework: a continuously synchronized nominal plant model serves as a lightweight digital twin, driving residual-based anomaly detection that responds to model-mismatch faults before tracking error accumulates; an event-triggered supervisory layer applies composite anomaly scoring, risk estimation, and dwell-time-constrained mode switching over a bank of pre-verified stable controllers; and an optional bounded LLM reasoning agent forming Layer 3 of the cognitive architecture produces real-time natural-language audit records at each supervisory event. The framework is functionally non-intrusive—the closed-loop response is numerically identical to standalone LQI in three of four benchmark scenarios and on an unstable plant variant, with zero supervisory interventions. Under abrupt 75% actuator loss, a single targeted intervention achieves lower overshoot and better tracking than all baselines, including MPC whose feasibility-degradation under reduced actuator authority is confirmed empirically. On near-undamped plants representative of compressor and pipeline dynamics, integral absolute error improves significantly relative to standalone LQI; on a distributed thermal process representative of laser welding and induction heating, IAE improves substantially under abrupt heater power loss. Parametric sensitivity analysis across fault severity, detection threshold, onset timing, measurement noise, score weights, and dwell-time confirms robustness of the detection architecture. The LLM layer’s contribution is regulatory rather than performative: its value is the causally connected audit record generated at the moment of each decision, not IAE improvement—a form of real-time explainability that after-the-fact attribution methods such as SHAP and LIME cannot provide by construction.