【摘要】 针对新建风电机组历史数据不足及不同机组间数据分布差异大的问题，提出一种结合自注意力机制与域自适应网络的风电机组异常状态检测方法。首先，采用编码器-解码器结构对源域和目标域风电机组运行数据进行特征重构，以捕捉潜在的风电模式和领域信息。然后，设计自注意力模块，通过与域判别器的对抗学习提取跨域共享特征，根据跨域信息的匹配度自动加权不同机组的领域信息，实现动态特征重构，从而提升模型对不同机组数据分布变化的适应性。最后，计算重构误差作为异常分数用于异常检测。实际风电机组运行数据的结果表明，该方法在历史数据有限的条件下能够高效地识别风机异常状态，相较于其他深度学习和深度迁移学习方法，显著提升了检测精度。更多还原

【Abstract】 To address the issue of insufficient historical data for newly installed wind turbines and the significant data distribution differences among various turbines, an abnormal status detection method was proposed for wind turbines that integrated the self-attention mechanism with domain adaptation networks. Firstly, an encoder-decoder structure was employed to perform feature reconstruction from the operation data of both source and target domain turbines in order to capture latent wind power patterns and domain-specific information. Then, a self-attention module was designed to extract cross-domain shared features through adversarial learning with a domain discriminator, and domain-specific information was automatically weighted based on the matching degree of cross-domain shared features, enabling dynamic feature reconstruction and thereby improving the model’s adaptability to changes in the data distribution across different units. Finally, the reconstruction error was calculated as the abnormal score for anomaly detection. The results from actual wind turbine operation data demonstrate that the method can efficiently identify abnormal data with limited historical data and significantly improve detection accuracy compared to other deep learning and deep transfer learning methods.更多还原