A Fault Detection Method for Wind Turbine Using the Dynamic K-Nearest Neighbor and Autoencoder

With the increasing importance of wind energy utilization and the high costs of maintenance, effective fault warning is crucial for ensuring the long-term stability of wind turbines. This paper addresses the issue of fault detection in wind turbines by proposing a novel detection method based on the dynamic k-Nearest neighbor (DKNN) and Autoencoder models. This study first employs Zero-Component Analysis Whitening (ZCA) for the preprocessing of SCADA data, reducing feature correlations and enhancing the performance of subsequent machine learning models. By utilizing the DKNN algorithm, we can identify pattern features under both normal and abnormal conditions, while the Autoencoder model detects anomalies through reconstruction errors. Experimental results demonstrate that our method effectively identifies potential faults in wind power systems and successfully achieves a two-day advance warning, significantly improving the timeliness of fault detection. This approach provides a new pathway for the reliability and intelligent management of wind energy systems, showcasing the potential for further optimization of detection performance across broader datasets.