Learning Deep Representation for Blades Icing Fault Detection of Wind Turbines

The blades icing accretion detection of wind turbines is a quite challenge problem in wind energy industry. One of the fair important aspects is that wind turbines work on the varying condition, in which the wind speed changes a lot. In this paper, the feasibility of it based on the power-related data, wind data, and temperature data in the supervisory control and data acquisition (SCADA) system is investigated. A deep neural network (DNN)-based framework is proposed to monitor the health condition of wind turbines. The core of it is that the discriminative deep feature representation is learned by taking multiple operation stages of wind turbines into consideration in the network structure, such as the stage of maximal power point tracking, the stage of constant rotating speed, the stage of rated power output, and so on, which handles two problems well, i.e., the data imbalance problem and high degree of feature variability in original SCADA data. The optimization goal of the proposed DNN is the triplet-header hinge loss. It preserves locality across operation stages and discrimination between different health conditions. The effectiveness and generalization ability of the proposed method is demonstrated by the SCADA data of three wind turbines from one wind farm in north China. The results show that the proposed DNN outperforms the traditional method named normal behavior modeling based on artificial neural network.