Speaker
Description
Operation and maintenance costs account for 25–35% of the total lifetime cost of an offshore wind farm. This study addresses that gap by developing and testing a prototype agentic AI framework for offshore wind turbine maintenance decision support.
The proposed framework uses four specialized agents in sequence, with each agent producing a validated output for the next stage. The framework was validated on the publicly available EDP SCADA dataset across 2016–2017. The pipeline integrates ML and large language model reasoning at each stage: anomaly detection outputs from the Isolation Forest are interpreted by a language model root cause agent grounded in ISO 15243 bearing failure taxonomy, whose diagnosis informs the degradation model's health indicator construction, whose prognostic output in turn is interpreted by a second language model layer before passing urgency and remaining useful life estimates to the scheduling agent.
Agent 1 detected 16 of 17 failure events (94% recall) across four component groups. Agent 2 classified generator bearing failures as rolling contact fatigue (T07) and adhesive wear (T06), both at 0.75 confidence. Agent 3 achieved a prognostic horizon prediction of 8.76 days Remaining Useful Life at 14.5% relative error within the ±20% accuracy criteria. Agent 4 identified a vessel-valid three-day inspection window closing one day before the recorded failure. The main limitation was the small number of failure events; this limited the calibration of the prognostic model for hydraulic components.
| Topic Area | Smart Operations and Maintenance |
|---|