An open-weight Large Language Model agent for condition monitoring and predictive maintenance in wind energy systems

Condition monitoring and predictive maintenance of large wind farms require fusing heterogeneous information from SCADA measurements, maintenance logs, manufacturer manuals, weather forecasts, and historical fault records into rapid, safety-critical dispatch decisions. Cloud-hosted commercial Large Language Models (LLMs) are increasingly capable at this kind of cross-source synthesis, but they are unsuitable for regulated energy operators that cannot route turbine telemetry to third-party APIs. This paper proposes an open-weight LLM agent framework that bridges condition monitoring and predictive maintenance, running entirely within the operator’s security perimeter on commodity hardware. The LLM acts as a reasoning and orchestration layer, calling a curated library of five domain tools (SCADA query, anomaly detector, thermal/physics check, retrieval-augmented generation, maintenance scheduler) to produce grounded, traceable, and actionable recommendations. We formalise the task as a two-stage information-fusion problem and prove that the tool-grounded decomposition guarantees grounding (every numerical token traceable to a deterministic procedure) and consistency (numerical content invariant under repeated calls). These guarantees hold for any open-weight LLM with native function-calling support, making the framework model-agnostic. The reference implementation uses the Qwen3 family via Ollama. On a 30-day SCADA record from a 2 MW Siemens turbine with a documented gearbox-bearing failure, pre-fault signatures (a + 9.9 ° C rise in the 72-hour rolling mean of generator-temperature 1 above its normal-operation baseline, and a 0.19 drop in the gearbox–generator temperature correlation) become detectable in the SCADA stream up to 72 h ahead of failure. On 24 held-out operator queries, the agent achieves 91.7% accuracy on health-checks and 83.3% on predictive alerts, outperforming a fixed-threshold baseline, an isolation-forest detector, and a zero-shot tool-less LLM.