241. Hybrid Stacked-Ensemble and LLM-Augmented Framework for Predictive Maintenance of Wind Turbines
Invited abstract in session WB-3: Uncertainty and energy, stream Energy and Sustainability.
Wednesday, 10:45-12:15Room: H5
Authors (first author is the speaker)
| 1. | Mohamed Saâd El Harrab
|
| Industrial Engineering Department, Co-Innovation Lab, École nationale des ponts et chaussées | |
| 2. | Quoc-Tuan Tran
|
| Liten, CEA |
Abstract
Modern wind turbines generate dense SCADA and vibration logs in which early fault signals are easily masked by noise. We introduce a hybrid framework that pairs a stacked ensemble of anomaly-detection models with a large language model (LLM). Cleaned, windowed sensor data are processed by complementary detectors (Isolation Forest, ν-SVM, Prophet residual analysis and a convolutional auto-encoder...). Their calibrated scores feed a sparsity-regularised logistic combiner that minimises a cost-weighted sum of false and missed alarms. A Kolmogorov–Smirnov guard monitors distribution drift and triggers lightweight retraining as conditions evolve.
The LLM enlarges the feature space with automatically generated descriptors and periodically adjusts detector weights using compact context vectors such as season and wind speed. Operator feedback is distilled into concise rules that mute recurring false alerts, closing the loop. Tests on multi-year fleet records improve precision–recall while halving unnecessary alarms, demonstrating a scalable and interpretable tool for turbine-health management.
Keywords
- Artificial Intelligence
- Machine Learning
- Time Series Analysis
Status: accepted
Back to the list of papers