A novel transformer network for anomaly detection of wind turbine

The supervisory control and data acquisition (SCADA) system of wind turbine includes 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, allowing for detection of anomalies. It is demonstrated that the Transformer model and its enhanced variants exhibit strong feature extraction capabilities for SCADA data. However, they face limitations in handling non-stationary features, thereby reducing the informativeness of data. In this paper, a novel Transformer network called non-stationary Transformer is proposed for anomaly detection of wind turbines. The model is employed to improve performance on extracting non-stationary features. In model, a multilayer perceptron (MLP) is proposed to adaptively learn de-stationary factors from raw SCADA time series data. These learned factors modify the Transformer’s self-attention mechanism, replacing it with de-stationary attention. Kullback-Leibler (KL) divergence is performed to quantify the differences between predicted and actual values. Using KL divergence, the probability density function (PDF) of normal and abnormal data are formulated, illustrating their distribution differences. By comparing with other models using real wind farm dataset, the proposed model is demonstrated to achieve superior performance on anomaly detection of wind turbine.