摘要：

风力发电机故障分类的复杂性和多样性严重影响风能发电效率, 传统的人工方法效率低下, 准确率较低, 已有的深度学习模型在真实环境中易受数据噪声干扰而表现不佳。为提升风力发电机故障分类模型在真实环境下的分类性能与鲁棒性, 提出一种基于对抗训练与Transformer的故障分类方法。首先通过引入一维卷积与门控线性单元(GLU)增强注意力机制对局部特征的学习, 保留易被忽略的局部信息, 提升模型对于局部特征的敏感度。其次结合限制因子约束对抗样本, 提高对抗样本产生的准确性。最后在消除错误样本的同时反馈生成过程, 使其具备更好的抗干扰能力。实验结果表明, 与5种常用的分类模型相比, 所提模型分类性能平均提升7.76%, 与真实结果之间的误差最小。局部增强的注意力机制和所提的对抗训练方法分别使模型的分类性能平均提升4.51%、4.95%。所提模型在10%~20%噪声环境中仍保持较好性能, 增强了其在真实环境中的稳定性。该方法在提高分类准确率的同时使模型具备更强的泛化能力, 对于提升风力发电机故障分类性能与鲁棒性具有重要意义。

关键词: 风力发电机, 门控线性单元, Transformer模型, 对抗训练, 故障分类

Abstract:

The complexity and diversity of wind turbine fault classification severely affect the efficiency of wind power generation. Conventional manual methods display low efficiency and accuracy. Existing deep learning models perform ineffectively in real environments owing to the data noise interference. To improve the classification performance and robustness of wind turbine fault classification models in real environments, this paper proposes a fault classification method based on adversarial training and Transformer. First, by introducing a one-dimensional convolution and Gated Linear Unit (GLU) enhanced attention mechanism for learning local features, the paper improves the sensitivity of the model to local features by retaining local information that is overlooked straightforwardly. Second, combining with constraint factor-constrained adversarial samples improves the accuracy of adversarial sample generation. Finally, while eliminating incorrect samples, the feedback generation process enhances its anti-interference capability. The experimental results reveal that compared with five commonly used classification models, the proposed model achieves an average improvement in classification performance of 7.76% and minimal error compared with actual results. The locally enhanced attention mechanism and proposed adversarial training method improve the average classification performance of the model by 4.51% and 4.95%, respectively. The proposed model still maintains good performance in a noise environment ranging from 10% to 20%. This enhances its stability in real environments. The method improves the accuracy and enhances the generalization capability of the model. This is significant for improving wind turbine fault classification performance and robustness.

Key words: wind turbine, Gated Linear Unit(GLU), Transformer model, adversarial training, fault classification