A wind turbine frequent principal fault detection and localization approach with imbalanced data using an improved synthetic oversampling technique

Frequent principal fault detection and localization (FPFDL), as a new problem of fault diagnosis of the wind turbine system in practice, has gained a growing concern in wind power industries. The knowledge-based fault diagnosis method with historical wind buffer data is a feasible way to solve this problem. However, due to the uncertainty of the principal fault and the incompletion of the practical data, the wind buffer data used is imbalanced, inadequate, and unfixed-length, which leads to higher misclassification rates for the minority classes by traditional machine learning methods. To overcome these challenges, a novel FPFDL approach is proposed in this paper. Firstly, we design an improved oversampling algorithm to generate and develop the balanced dataset based on the imbalanced dataset of unfixed-length. This algorithm combines the dependent wild bootstrap oversampling and the modified synthetic minority oversampling technique. So, it can consider the temporal dependence and the relationship between the samples during data oversampling. Secondly, we introduce the one-dimensional convolutional neural networks to achieve automatic high-level local feature extraction and fault identification. Finally, the experimental results of seven cases using the datasets collected from two real wind farms in China validate our proposed approach's effectiveness and robustness with imbalanced wind buffer data of unfixed-length.