Simple trigger logic is commonly used in actual wind farms to monitor unit conditions, which face problems such as a high false-alarm rate and overlapping alarms. In addition, the characteristics of SCADA data, such as large quantity, complexity, and variable correlation, lead to insufficient accuracy of fault diagnosis. To address this problem, an improved fault diagnosis method based on a Multi-Channel Attention Mechanism Convolutional Neural Network (MCAMCNN) is proposed. Firstly, feature analysis is performed after preprocessing SCADA data to fully explore the coupling characteristics between data, and a dataset is established. Then, the proposed fault diagnosis model is used for feature screening. Innovatively, a structure combining double-layer multi-scale convolution and multi-channel attention is adopted to extract multi-domain features and dynamically calibrate the weights of feature channels. Fault classification is realized after adaptive fusion of features by Efficient Channel Attention (ECA). Finally, experiments are designed based on real data from an onshore wind farm in China, which verify that the method is timely and accurate in fault diagnosis, with significantly improved accuracy and F1-score, and has obvious advantages over comparative methods.
Zheng et al. (Mon,) studied this question.