Feature selection for unsupervised defect detection of a wind turbine blade considering operational and environmental conditions

Structural Health Monitoring (SHM) of wind turbine blades (WTBs) is beneficial for timely detection of defects and indicating lack of performance. Methods for these rely on Damage Sensitive Features (DSFs), often derived from vibration responses of the blades, requiring effective feature selection to extract relevant information. Despite the success of various feature selection and damage detection methods, challenges remain in capturing sufficient DSFs, especially in the presence of environmental and operational variability (EOV). This challenge is particularly important for unsupervised machine learning approaches, which are rapidly evolving and being adapted to the wind sector. This paper proposes a novel feature selection approach for data-driven anomaly detection in an operating V27 blade. The proposed approach integrates EOV-related insights from a healthy baseline, with data-driven anomaly detection techniques identifying an effective and minimal feature set for damage detection using acceleration responses from twelve sensors installed on the blade. Initially, twelve time-domain, fifteen frequency-domain, and thirty-five time–frequency domain features are employed to capture information from the vibration responses. Subsequently, the best set of features is selected using their variation in the healthy state, and their correlation with temperature. Finally, two unsupervised anomaly detection methods, One-class Support Vector Machine (OC-SVM) and Autoencoder (AE) are employed to show the performance of the selected features. The results show how selecting optimal features can significantly improve the performance of anomaly detection for structures under operating and complex conditions such as WTBs. The proposed method can also be adapted for a wide range of sensors and unsupervised detection.