Fault Diagnosis of Wind Turbines with A Dilated Convolution and Hybrid Attention Neural Network

Abstract

the class-imbalanced problems of data and the low generalization of the model are two major nuts to crack in fault diagnosis of wind turbines. Effective fault diagnosis guarantees the overall stability of power systems and huge economic benefits. Hereby, this paper proposes a fault diagnosis framework based on neural networks. First, the dilated convolution is used to capture complex data patterns with a receptive field which increases exponentially with dilation size, and the common characteristics among wind turbines are fully extracted locally to enhance the generalization ability. Then, the hybrid attention is successively introduced into the convolutional network to obtain the representation ability of different classes globally to reduce the influence of class-imbalance. The experiments are carried out on a real dataset from a wind farm in China. The proposed framework diagnoses the blade icing faults of different wind turbines with a variety of imbalanced degrees to verify the validity and feasibility. The proposed fault diagnosis framework obtains better performance than the existing methods.