Reconstruction-based Deep Unsupervised Adaptive Threshold Support Vector Data Description for wind turbine anomaly detection

Abstract

The global deployment of wind turbines as a sustainable and clean energy source underscores the criticality of early anomaly detection to ensure their safe operation, improve power generation efficiency, and reduce downtime costs. Yet, acquiring sufficient labeled and faulty data is time-consuming and expensive in practical applications, limiting the use of supervised learning methods. To this end, this paper introduces a new approach, namely the Reconstruction-based Deep Unsupervised Adaptive Threshold Support Vector Data Description (DUA-SVDD) model, for wind turbine anomaly detection. DUA-SVDD integrates reconstruction-based and boundary-based anomaly detection paradigms, synthesizing comprehensive and detailed representation information from dynamic monitoring data, encoding the distribution and patterns of normal samples across multiple levels. This model employs a joint optimization mechanism to minimize reconstruction errors and hypersphere volume simultaneously in the latent space, resolving the hypersphere collapse issue observed in Deep Support Vector Data Description (DeepSVDD). It constructs a well-structured latent space proficient in handling data noise and variations, allowing SVDD to establish more robust spherical boundaries. Additionally, it proposes an adaptive threshold algorithm based on pseudo-data to accurately differentiate abnormal from normal patterns. The method is tested and evaluated on real wind farm SCADA datasets. A comparative analysis against state-of-the-art methods highlights the superior performance of the proposed model in detecting blade icing on wind turbines, achieving average AUC values of 97.54% and 99.45% across two specific cases.