Anomaly detection and diagnosis for wind turbines using long short-term memory-based stacked denoising autoencoders and XGBoost

Abstract

An anomaly detection and diagnosis method for wind turbines using long short-term memory-based stacked denoising autoencoders (LSTM-SDAE) and extreme gradient boosting (XGBoost) is proposed in this paper. First, an abnormal data recognition algorithm based on the local outlier factor and adaptive K-means was developed to implement data preprocessing and noise extraction. The LSTM-SDAE model was then established to obtain the nonlinear temporal relationship among multivariate variables in normal behavior modes. The Mahalanobis distance was calculated based on reconstruction errors and the threshold for anomaly detection was set with a 99.7% confidence interval for the distribution curve fitted by kernel density estimation. An alarm mechanism based on the sliding window technique was set up to detect abnormalities in real time. Finally, contribution analysis was conducted to extract the parameter features under different abnormal modes, and the XGBoost was trained by extended data from wind turbines of the same type in the same wind farm to realize anomaly location and diagnosis. To verify the proposed method, real SCADA data from a wind farm located in northeastern China were applied. The results show the capability of the proposed method in anomaly detection and diagnosis for wind turbines.