Fault detection in wind turbines is essential for ensuring their safety, reliability, and optimal performance. However, some of the existing approaches for sensor fault detection in wind turbines face several challenges that hinder their effectiveness. These challenges include handling multivariate and non-Gaussian data, low sensitivity to small changes, and setting an appropriate detection threshold to avoid false alarms. Additionally, constructing an analytical model for monitoring wind turbines becomes particularly challenging and time-consuming, especially for large-scale wind turbines. This paper proposes a semi-supervised data-driven approach for sensor fault detection in wind turbines using supervisory control and data acquisition (SCADA) data. The proposed approach combines the advantages of independent component analysis (ICA) and the Kantorovich Distance (KD)-based fault detection scheme. ICA enables efficient handling of multivariate non-Gaussian data, while the KD scheme provides a sensitive indicator for assessing the residuals obtained from ICA. The ICA-based KD scheme needs only fault-free data in training, making it more attractive for fault detection in practice. Kernel density estimation is employed to compute the detection threshold of the KD scheme, making it more flexible. Experimental evaluations using simulated sensor faults based on real wind turbine data demonstrate the superior detection performance of the proposed approach, achieving an average F1-score of approximately 0.96 and outperforming conventional approaches.
All Science Journal Classification (ASJC) codes
- Control and Systems Engineering
- Renewable Energy, Sustainability and the Environment
- Energy Engineering and Power Technology
- Electrical and Electronic Engineering