Optimizing wind turbine early fault identification: a multi-sensor approach with an enhanced DBSCAN algorithm

Abstract To address the limitations of single-sensor data in accurately detecting early abnormal states of wind turbine units, particularly considering the difficulty of obtaining labeled data and the distribution differences in datasets caused by changes in operating conditions, this study employs multi-sensor signal fusion and unsupervised learning methods to enhance fault identification from two aspects. First, the WA-VMD-CC framework (Signal Decomposition and Fusion module) synchronizes multi-sensor signal frequencies using the window average method, applies the Variational Mode Decomposition (VMD) algorithm to decompose signals, and utilizes the correlation coefficient (CC) method to select modal signals for fusion. Second, a local cluster optimization algorithm based on cluster center density differences is proposed to improve the clustering performance of the DBSCAN algorithm for different turbine samples. Experimental results demonstrate that the proposed method significantly outperforms HDBSCAN and other clustering algorithms in evaluation metrics such as the Silhouette Coefficient, achieving a maximum performance improvement of 41.70% over HDBSCAN. This indicates superior clustering compactness and separability, ensuring the reliability and robustness of the wind turbine status early warning system. This enhancement aligns with the practical engineering requirements of unsupervised learning in fault diagnosis applications.