A Novel Transformer Network for Anomaly Detection of Wind Turbines

The supervisory control and data acquisition (SCADA) system of wind turbines include various state parameters, such as oil temperature, bearing temperature, and generator speed. By analyzing SCADA data, the operating status of wind turbines can be evaluated, and then anomalies can be detected. Previous studies have used stationarization (normalization and denormalization) for better predictability. However, over-stationarization may remove useful non-stationary characteristics, thereby reducing the informativeness of the data. Non-stationary Transformer model, which includes series stationarization and de-stationary attention modules, is employed to handle SCADA data's non-stationarity. De-stationary attention can estimate attention without normalization and specific temporal dependencies in the original SCADA data. It will replace the Self-Attention of Transformer network, in this way can processing non-stationary information of original data effectively. As a result, the proposed model outperforms the Transformer model for anomaly detection on real SCADA data.