Fault detection in wind turbines using health index monitoring with variational autoencoders

Abstract. As wind energy capacity expands globally, ensuring the operational reliability and economic viability of wind turbines has become a critical industrial challenge. Effective fault detection systems are essential for minimizing high maintenance costs and preventing catastrophic failures. To address this need, this paper presents a semi-supervised framework designed to identify anomalies in wind turbines using only healthy operational data. The methodology begins by extracting a comprehensive set of features from the time and frequency domains of raw vibration signals to capture a rich representation of the dynamics of wind turbines. A variational autoencoder, a deep generative model, is then trained exclusively on these features from healthy operational periods to learn a robust model of normal behavior and generate reconstruction errors as health indicators with exponentially weighted moving average smoothing to enhance robustness and reduce false alarms. The framework is evaluated using public data from the Aventa AV-7 ETH Zurich research wind turbine, which includes multiple failure events. Experimental results demonstrate effective and early detection of pitch faults, as well as accurate detection of icing events and aerodynamic imbalances. The proposed approach therefore offers a robust and practical solution to improving operational safety and enabling proactive maintenance of wind turbines.