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

A new framework submitted to arXiv on 26 Jun 2026 proposes using Large Language Model agents to perform active Fault-Tolerant Control, converting fault detection outputs into validated recovery actions for industrial processes [1]. The agentic framework, detailed in a preprint, couples a multi-agent workflow with a Digital Process Plant Twin and a Graph Retrieval-Augmented Generation layer built on the CPSMod ontology [1][2]. The multi-agent system decomposes operator responsibilities into monitoring, planning, action synthesis, simulation, validation, and reprompting [2]. Corrective actions are generated as minimal-risk state-machine recovery paths and corresponding discrete commands or continuous setpoint adaptations [2]. Before any actuation, these actions are validated deterministically against interlocks, envelopes, and dynamic feasibility [2]. If no acceptable plan is found within a bounded time window, control is handed to a safety fallback [2]. The framework was evaluated in simulation on two benchmarks: a discrete batch Mixing Module and a Continuous Stirred-Tank Reactor under closed-loop PID regulation [1][2]. Results with lightweight models GPT-4o-mini and GPT-4.1-mini showed that semantically grounded agents can derive valid recovery decisions within latency budgets compatible with the respective process dynamics [2]. The preprint was posted on arXiv, an open-access repository of electronic preprints that is not peer-reviewed, which was founded in 1991 and now receives about 24,000 articles per month [6]. The work appears under the Electrical Engineering and Systems Science category and is accessible through the arXiv abstract page, which also features community-developed tools via the arXivLabs framework [1][5]. arXivLabs, launched as a formalized collaboration space in 2020, allows third-party developers to create experimental features that appear as tabs on article record pages, with partners required to adhere to arXiv's values of openness, community, excellence, and user data privacy [5].