Intelligent information fusion for safety critical icing detection via machine learning and reliability analysis in wind turbine systems

Abstract Ensuring the safety and reliability of wind turbine systems in cold climates requires robust icing detection, as missed events can lead to structural compromise. This study proposes an intelligent information fusion framework that integrates vibration-based structural dynamics with environmental data through a pipeline of automated Operational Modal Analysis (OMA) and temperature compensation. Five classification architectures—Optimized Threshold (OT), Random Forest (RF), XGBoost (XGB), Support Vector Machine (SVM), and Gated Recurrent Unit (GRU)—were rigorously evaluated using rolling-window time-series cross-validation on high-fidelity experimental data from a large-scale climate chamber. To address the safety-critical nature of the application, the framework prioritizes Recall to minimize False Negative (FN) errors and incorporates non-parametric statistical validation (Kruskal-Wallis and Dunn’s post-hoc tests) to quantify model stability. Results reveal a “simplicity paradox” where the physics-informed OT model achieved the highest operational reliability (Recall = 0.9600), outperforming advanced ensemble and deep learning methods in both safety and statistical stability. This work demonstrates that meticulous feature-level fusion and domain-specific engineering enhance system resilience more effectively than increased algorithmic complexity. The proposed methodology provides a scalable foundation for multi-sensor fusion and adaptive predictive maintenance in renewable energy infrastructures.