关键词: 自注意力编码器, 密度聚类算法, 数据清洗, 监控与数据采集系统, 风电机组

Abstract: The supervisory control and data acquisition( SCADA) system for wind turbines is often affected by various
factors, leading to data anomalies. This article proposes a A method for improving the density-based spatial clustering of
applications with noise( DBSCAN) clustering model based on Transformer autoencoder was proposed. By combining the
feature extraction ability of Transformer and the density clustering characteristics of DBSCAN, relative position encoding
and optimized multi head attention mechanism were introduced to enhance the recognition ability of SCADA abnormal data.
The experiment shows that compared with traditional methods, the data cleaning effect and model accuracy were better. The
abnormal data removal rate of this method reached 26. 58% and when fitting the wind speed power curve, MAE and RMSE
were the lowest and R² was the highest. The cleaned SCADA data used for unit fault diagnosis can significantly improve the
accuracy of wind turbine fault identification over 92%, the timeliness of fault warning 20% in advance, and the accuracy of
fault type classification 30% improvement. This method not only improved the operational efficiency and reliability of wind
turbines, but also provided more reliable data support for the operation management and decision-making of wind farms.

Key words: transformer autoencoder, density-based spatial clustering of applications with noise ( DBSCAN ), data
cleaning, supervisory control and data acquisition( SCADA) system, wind turbine unit