From Detection to Action: Using LLM Agents for Fault-Tolerant Control

🕒 Published (v1): 2026-06-26 12:13 UTC · Source: Arxiv · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

A multi-agent LLM framework for active Fault-Tolerant Control (FTC) that couples a six-agent workflow with a Digital Process Plant Twin (DPPT) for pre-execution simulation and a Graph RAG layer built on the CPSMod ontology for relation-aware, multi-hop retrieval of plant knowledge. Evaluated on a discrete batch mixing module and a continuous stirred-tank reactor (CSTR) using GPT-4o-mini and GPT-4.1-mini, showing semantically grounded agents can derive valid recovery decisions within latency budgets compatible with process dynamics.

Problem

• Process plant operators must interpret noisy/faulty evidence under alarm floods and uneven expertise across shifts; existing automation struggles with unforeseen conditions
• RL-based active FTC suffers from low sample efficiency, limited generalizability across systems, and requires carefully designed reward functions for stable policy learning
• Document-level RAG cannot support fault recovery because it requires explicit traversal of distributed relations among equipment, faults, interlocks, operating modes, and control actions — isolated document chunks are insufficient
• Three specific gaps in prior LLM-for-control work: (i) Grounding — plant knowledge is fragmented across heterogeneous artifacts with no relation-aware retrieval; (ii) Assurance — proposed actions lack auditable feasibility checks against interlocks, envelopes, and physics before execution; (iii) Timeliness/Robustness — recovery must handle hybrid (discrete+continuous) dynamics under bounded latency

Method

• Agent Layer (six roles): Monitoring Agent (deterministic/data-driven fault detection from sensor/actuator streams); Planning Agent (LLM-based, generates minimal-risk state-machine recovery paths); Action Agent (LLM-based, synthesizes discrete commands or continuous setpoint adaptations); Simulation Agent (invokes DPPT simulation to test proposed actions); Validation Agent (checks interlocks, operating envelopes, dynamic feasibility); Reprompting Agent (iterative refinement feedback loop with bounded iterations)
• Tool Layer — DPPT: Digital Process Plant Twin exposing time-synchronized sensor/actuator data, physics-based and reduced-order simulation models, and validation services. Supports condition monitoring and simulation-based pre-execution testing.
• Tool Layer — Graph RAG: Built on the CPSMod alignment ontology integrating five standards-based ontology design patterns (ODPs): VDI 2206 (structural decomposition), VDI 3682 (function/product/energy/information flows), UML State Machine (discrete behavior), OpenMath (continuous behavior via equations), DIN 17359 (condition monitoring and fault diagnosis). Agents invoke predefined SPARQL templates to retrieve compact, relevant subgraphs as structured context appended to LLM prompts.
• Information flow: Monitoring → Planning → Action → Simulation → Validation (with reprompting feedback loops); if no admissible plan within bounded time window, control hands to a safety fallback in the Physical Space
• LLM role: Used only in Planning and Action agents where interpretation of semantically enriched knowledge and synthesis of constraint-compliant recovery decisions is needed; Monitoring, Simulation, Validation remain deterministic/rule-based

Key Contributions

• Relation-aware grounding for LLM agents via CPSMod ontology + Graph RAG, enabling multi-hop traversal of equipment-function-behavior-fault semantics (addresses RQ1)
• Safety-assured action generation via minimal-risk state-machine recovery paths, deterministic pre-execution validation (interlocks, envelopes, dynamic feasibility), and an auditable decision trace — all wrapped in a multi-agent workflow (addresses RQ2)
• Bounded-latency operation through tool-assisted monitoring, constrained action generation, runtime verification, and a hard safety fallback if no admissible plan is found (addresses RQ3)
• Framework handles both discrete and continuous recovery tasks (supervisory sequences and setpoint adaptations)
• Separation of T-Box (reusable schema) and A-Box (plant-specific instances) enables ontology reuse across different process plants

Results

• Evaluated on two simulation benchmarks: a discrete batch Mixing Module and a Continuous Stirred-Tank Reactor (CSTR) under closed-loop PID regulation
• Lightweight LLMs tested: GPT-4o-mini and GPT-4.1-mini
• Semantically grounded agents derived valid recovery decisions within latency budgets compatible with both discrete and continuous process dynamics
• (Full quantitative results from Section 5 are in the truncated portion of the paper)

Limitations

• Evaluated only in simulation, not on a physical process plant — real-world sensor noise, actuator lag, and model–plant mismatch are not tested
• Two case studies only (mixing module and CSTR); generalizability to larger, more complex plants with hundreds of interacting units is unverified
• Requires upfront engineering effort to construct and maintain the CPSMod ontology A-Box and the SPARQL template registry
• Safety guarantee depends on bounded-iteration reprompting; if the LLM repeatedly generates infeasible plans, the fallback (plant shutdown) is the only option
• LLM inference latency, while bounded, may not scale to sub-second control loops in fast process dynamics

Relevance to Agentic AI / LLM Agents

This paper demonstrates how to embed LLM agents in a safety-critical closed-loop control setting by constraining their outputs with deterministic validation gates and grounding them in structured domain knowledge via Graph RAG. It addresses the practical deployment challenge of making LLM agents reliable enough for systems where incorrect actions have physical consequences — a template for agentic AI in cyber-physical systems. The multi-agent decomposition (monitor → plan → act → simulate → validate → reprompt) is a reusable architecture pattern for any domain requiring auditable, intervention-safe LLM-based decision making.