Reconstruction-Based Anomaly Detection in Wind Turbine Operation Time Series Using Generative Models

Unsupervised time series anomaly detection is a common tasks in many real world problems, in which the normal/anomaly labels are extremely unbalanced. In this work, we propose to use three generative models (namely, a basic autoencoder, a transformer autoencoder and a diffusion model) for a reconstruction-based anomaly detection pipeline applied to failure detection in wind turbine operation time series. Our experiments show that the transformer autoencoder yields the most accurate reconstructions of the original time series, whereas the diffusion model is not able to obtain good reconstructions. The reconstruction error, which is used as an anomaly score, seems to follow different distributions for the anomalies and for the normal data in 2 of the 3 models, which is confirmed by our quantitative evaluation. The transformer autoencoder is the best performing generative model, achieving a AUC score of 0.98 in the detection of the anomalies. However, the same result is obtained by standard (i.e. non-generative) outlier detection algorithms, exposing that although the anomalies in this problem are sequence anomalies - with a temporal nature -, they can be effectively modeled and detected as point outliers.