A Tutorial on Autonomous Fault-Tolerant Control Using Knowledge-Grounded LLM Agents

Fault recovery in industrial process plants is traditionally handled by human operators who interpret complex data, such as P&IDs and process trends, to maintain safety during faults. This paper proposes a framework to automate this process using Large Language Model (LLM) agents. By acting as a constrained supervisory planner, the LLM suggests recovery actions that are grounded in plant-specific knowledge. To ensure safety, these suggestions are not executed directly; instead, they are passed through an external validator—either symbolic or simulation-based—that confirms the action is safe and feasible before it is applied to the plant.

The authors decompose the recovery process into three key design dimensions:

1. Recovery Patterns: The framework categorizes problems into discrete supervisory routing (Pattern A), continuous setpoint adaptation (Pattern B), or hybrid approaches (Pattern C). This classification helps determine the appropriate action space and the type of validator required.
2. Validation Strategies: The validator acts as the critical safety boundary. Symbolic validators check for structural consistency (e.g., reachability), while simulation-based validators use a digital twin to test if a proposed action is dynamically feasible under current plant conditions.
3. Deployment Considerations: Practical implementation requires addressing latency, the significant effort of building a plant-specific knowledge graph, integration with existing safety systems, and managing the lifecycle of the LLM model to ensure reproducible behavior.

This research provides a structured approach to integrating stochastic AI models into deterministic industrial environments. By keeping the LLM in a supervisory role and enforcing strict validation, the framework allows for the flexibility of LLM reasoning while maintaining the rigorous safety standards required in process engineering. The authors provide two open-source Python environments—a mixing module and a continuous stirred-tank reactor—to help researchers test and refine their own recovery and validation logic.