Optimized approach for wind turbine anomaly detection: integrating uncertainty quantification and dimensionality reduction

Abstract

Accurate anomaly detection in wind turbines is critical for operational stability, yet existing methods are hindered by a persistent trade-off among uncertainty quantification, high-dimensional complexity, and computational efficiency, a challenge termed the "Anomaly Detection Trilemma". To resolve this, we propose a novel Bayesian-refined Random Forest (BPRF) model featuring a synergistic three-stage architecture. The framework first employs a Bayesian module to generate a probabilistic feature, then uses a Random Forest as a non-linear amplifier to map data into a high-dimensional space, and finally applies Principal Component Analysis (PCA) to distill a robust low-dimensional "normalcy" subspace for anomaly scoring. Validated on a real-world public dataset, the BPRF model demonstrated superior performance over a comprehensive set of mainstream baselines, achieving a leading F1-score of 0.59, an accuracy of 0.87, and an AUC of 0.86. The results confirm that the BPRF model effectively escapes the "Anomaly Detection Trilemma", providing a solution that is not only accurate but also reliable and computationally tractable, thereby establishing a new performance benchmark for practical industrial deployment.