Deep learning-based anomaly detection for wind turbines

Abstract

Data-driven strategies have the potential to significantly improve diagnostic and prognostic analyses in condition monitoring (CM) of wind turbine (WT) operations, leading to improve maintenance planning and reduce operation and maintenance (O&M) costs. This study presents a deep learning (DL) approach for anomaly detection in WT gearboxes, a crucial component of WT systems. The anomaly detection models were developed based on vibration data collected from the healthy and damaged gearboxes.

The results demonstrate the effectiveness of the proposed long short-term memory (LSTM) autoencoder approach, with no signs of overfitting or underfitting observed in the training and validation loss curves. Additionally, the mean squared error (MSE) loss functions for all features employed during validation are deemed satisfactory. Evaluations of the model performance have shown practical values for precision, recall, F1-score, and accuracy ranging from 99.4–99.9%, 53.5–100%, 69.6–100%, and 76.6–100%, respectively.

The developed anomaly detection method provides critical insights into the health of WT gearboxes. Analysing vibration data collected from these gearboxes enables early detection of anomalies.