Tool-grounded large language models for decision-level fusion in fault diagnosis

Abstract

Modern condition-monitoring systems generate heterogeneous information, such as high-rate sensor streams, derived health indicators, and discrete event histories, that must be combined into actionable diagnostic statements. Large language models (LLMs) are attractive as a decision-level fusion mechanism because they natively handle the textual modality, but their use in safety-critical fault diagnosis is impeded by hallucinated numerical content and stochastic outputs. This paper proposes a two-level information-fusion framework that uses an LLM as a constrained fusion operator rather than as a generator of facts. At the feature level, a recursive joint state-and-fault estimator built on a linear state-space model produces a vector of derived indicators, including an estimate of the actuator-fault parameter and its uncertainty. At the decision level, the LLM is restricted to interact with the estimator state and the event history exclusively through a finite library of typed deterministic tools, and must copy every numerical token in its response from a tool output. Grounding of the response is then satisfied by construction, and consistency follows once the sequence of tool calls is fixed. The framework is illustrated on a realistic actuator-fault scenario in which an autopilot-driven autonomous surface vessel encounters three fault events of different characters and the LLM produces reproducible operator-facing reports. A structured comparison against the two alternative reporting architectures available for this task, a rule-based template reporter driven by the same estimator and a flat-prompt LLM given the same numerical context, identifies open-ended query coverage at fixed numerical reliability as the property that neither alternative provides, and states the metrics under which that claim is to be tested.