Abstract:

To address the challenges of data imbalance and difficult feature extraction in fault diagnosis of wind turbine pitch systems， a fault diagnosis method was proposed based on an improved conditional generative adversarial network（ICGAN） and a multimodal feature fusion network（MMFNet）. Based on the traditional conditional generative adversarial network（CGAN） architecture， an attention layer was introduced to strengthen the model's ability to capture key information from input features. An autoregressive model was adopted to generate random noise， which improved the diversity and authenticity of generated samples. A reward mechanism was embedded to optimize the model training process and enhance the quality of generated samples. Based on this， the MMFNet model was constructed to achieve accurate diagnosis of multiple faults in wind turbine pitch systems. Experimental results demonstrate that the proposed ICGAN-MMFNet model achieves precision and F1-score above 0.96， as well as recall above 0.97 under various sample quantity ratios. The effectiveness and generalization of the proposed fault diagnosis model were verified when dealing with imbalanced datasets.

Key words: wind turbine pitch system, fault diagnosis, imbalanced sample, generative adversarial network

摘要：

针对风电机组变桨系统的故障诊断面临数据不平衡、特征提取困难等问题，提出了一种基于改进条件生成对抗网络（ICGAN）与多模态特征融合网络（MMFNet）的故障诊断方法。在传统条件生成对抗网络架构基础上引入注意力层，增强模型对输入特征关键信息的捕捉能力；采用自回归模型生成随机噪声，使生成样本更具多样性与真实性；加入奖励机制，优化模型训练过程，提高生成样本质量。在此基础上构建MMFNet模型，用于实现对风电机组变桨系统多故障的精准诊断。实验结果表明，所提出的ICGAN-MMFNet模型在不同样本数量比例下，精确率和F1分数都在0.96以上，召回率都在0.97以上，说明了所提故障诊断模型在不平衡样本下的有效性和泛化性。

关键词: 风电机组变桨系统, 故障诊断, 不平衡样本, 生成对抗网络