Wind turbine fault detection based on spatial-temporal feature and neighbor operation state

Wind turbines (WTs) are complex pieces of engineering. They are generally exposed to harsh environmental conditions. Thus, major changeable working conditions and multidimensional dynamic correlation arise, resulting in unsatisfactory fault detection accuracy. A WT fault detection method based on spatio-temporal features and neighbor operation state is proposed in this study to tackle the aforementioned challenge. A spatio-temporal feature extraction method based on accumulated mutual information is proposed to describe the dynamic correlation of multiple features during the offline phase. Then, the spatio-temporal features are applied to build the weighted k-nearest neighbor fault detection model. In the online phase, the weighted distance between the online sample and the neighbor operation state is adopted to identify the anomaly state. Moreover, a dynamic threshold based on the neighbor sample is designed to cope with the changeable operating conditions. The proposed methods are demonstrated using FAST data and supervisory control and data acquisition data, which consider different fault types. The experiment results show that compared with similar methods and traditional fault detection methods, the operational characteristics of WTs’ components can be better described by the proposed spatio-temporal feature extraction method, and the false alarm rates can be considerably reduced by the dynamic threshold.