Fault prognosis of wind turbines using multimodal machine learning

Wind turbines (WT) convert wind’s kinetic energy into electrical energy, providing sustainable electricity. However, all machines eventually experience performance decline due to faults. Predicting faults before they cause significant damage is crucial. Existing wind turbine fault detection studies use either regression or classification models alone, missing the opportunity to both accurately predict and classify different fault types using multimodal machine learning. We propose an ensemble multimodal ML system that uses various sensor data to predict and classify faults. A genetic algorithm selects the most relevant features from the dataset. The ensemble system integrates Recurrent Neural Network, Extra Tree Regressor, and Random Forest for regression, alongside Adaptive Boosting, Naive Bayes, and K-Nearest Neighbor classifiers for fault classification. A WT dataset with multiple sensors’ information is used to predict future values and classify faults. The model achieved high accuracy in classifying specific faults, while regression analysis showed stator temperature predictions had the lowest root mean squared error. The proposed model can be used for 48-h ahead fault prognosis. • Development of a comprehensive MMML framework for improved fault prognosis. • Implementation of advanced feature selection and model architecture. • Comprehensive evaluation using multiple performance metrics.