关键词: 风电机组, 齿轮箱, 生成对抗网络, 遗传算法, 峭度, 半监督学习, 故障诊断

Abstract: To solve the problem of low-fault diagnostic accuracy and unclear fault characteristics of wind turbine gearboxes under unbalanced data sets, a fault diagnosis method of Wasserstein generative adversarial networks optimized by kurtosis label and genetic algorithm is proposed in this paper. Firstly, the kurtosis label is mapped to the convolution layer as a semantic label to normalize the fault features. Secondly, the metagenomics is binary coded and the weights are initialized in deconvolution networks. Then, multi-point crossover and gaussian approximate variation are performed on the unbalanced sample sets to search for local faults. Finally, kurtosis is inputted to the discriminator network as labeled negative cases, and deconvolution and VGG neural networks are reconstructed to improve weight cutting. The WGAN network becomes a semi-supervised learning model, which can update weight forward judgment and output diagnostic results. Experimental results showed that the diagnostic accuracy of the proposed method could reach 98.69% under unbalanced data sets, indicating that it has a higher generalization ability and feature extraction ability, which can enhance fault features.

Key words: wind turbines, gearboxs, generative adversarial network, genetic algorithm, kurtosis, semi-supervised learning, fault diagnosis