Semi-supervised multivariate time series anomaly detection for wind turbines using generator SCADA data

The maintenance cost and unplanned downtime caused by faults are an important part of the operation cost of wind turbines. Supervisory control and data acquisition (SCADA) data is a multivariate time series (MTS) for monitoring the status of wind turbines, in which anomaly patterns may indicate potential faults. The existing anomaly detection methods can neither extract and process pattern information in MTS stably, nor make reasonable use of a small amount of valuable labeled data. In this paper, we propose an end-to-end semi-supervised anomaly detection model including reconstruction model, prediction model and auxiliary discriminator, with a joint objective function. Combining reconstruction model and prediction model, the unsupervised model can effectively extract the inter-variable correlation and temporal dependence of MTS data. Further, using the semi-supervised auxiliary discriminator based on adversarial training, the proposed model can integrate expert knowledge to incrementally upgrade performance from unsupervised to supervised level. Our evaluation experiments are conducted on a public server dataset and a real-world wind turbine SCADA dataset. The results show that the F1-score of unsupervised model can exceed the several state-of-the-art baseline methods by 3.86% and 2.89%, and the F1-score can be increased to 98.60% and 98.30% after using the auxiliary discriminator.