Advanced Multivariate Anomaly Detection for Wind Turbines Using SCADA Data and Machine Learning

Ensuring the reliability of wind turbines is essential for reducing maintenance costs and improving energy availability. This study introduces a multivariate anomaly detection framework that combines physical modelling and machine learning to identify early signs of mechanical faults under variable operating conditions. The approach uses SCADA data and nacelle vibration measurements, integrating support vector regression. Unlike traditional residual-based methods, the proposed solution extends detection to a multivariate domain through Mahalanobis distance, improving robustness against confounding factors. Validation on an open-access offshore dataset demonstrates that the multivariate approach achieves high accuracy (95% on average) in detecting pitch-related failures and other anomalies, outperforming univariate residual analysis. These findings highlight the potential of diagnostic strategies for predictive maintenance in wind energy systems.