Machine learning on fault diagnosis in wind turbines

• Main Authors: Ng, Eddie Yin Kwee, Lim, Jian Tiong
• Other Authors: School of Mechanical and Aerospace Engineering
• Format: Article
• Language: English
• Published: 2022
• Subjects: Engineering::Materials; Machine Learning; Artificial Neural Network; Python; Extreme Gradient Boosting; K-Nearest Neighbours
• Online Access: https://hdl.handle.net/10356/163954
• Institution: Nanyang Technological University

With the improvement of wind turbine (WT) operation and maintenance (O&M) technologies and the rise of O&M cost, the fault diagnostic in WT based on supervisory control and data acquisition (SCADA) system has become one of the cheapest and easiest method to detect early alarm of fault in WT. The O&M cost was 21% of the total cost of a wind turbine project. The replacement cost of the critical parts like gearbox is twice the replacement cost for generator or blade. Hence, it is necessary to monitor the change pattern of real time parameters from the WT and maintenance action could be taken in advanced before any major failures. Therefore, a SCADA-driven fault diagnosis in WT based on machine learning algorithm has been proposed in this study by comparing the performance of three different machine learning algorithms, namely k-nearest neighbours (kNN) with bagging regressor, extreme gradient boosting (XGBoost) and artificial neural network (ANN) on condition monitoring of gearbox oil sump temperature. Beside this, this study also compared the performance of two different features selections method, namely Pearson Correlation Coefficient (PCC) and Principle Component Analysis (PCA) and three hyperparameters optimization method on optimizing the performance of the models, namely grid search, random search and Bayesian Optimization. A set of 3 years SCADA data of WT located at France have been used to verify the selected method. The result showed the kNN with bagging regressor with PCA by applying grid search provide the best R2 score and the lowest root mean square error (RMSE). The trained model can detect the potential of WT faults at least 4 weeks in advanced. However, the proposed kNN model in this study was suggested to trained with Support Vector Machine hybrid algorithm to improve its performance and reduce fault alarm. ANN could also be enhanced by applying Bayesian Physics-Informed Neural Networks as this algorithm network is more compatible to the real world non-linear dynamic system like WT.