Deep Learning-Based Predictive Maintenance of Wind Turbine Components

Abstract

This research examined predictive maintenance strategies for wind turbines by tracking operational variables, including wind speed, rotor speed, gearbox temperature, generator temperature, vibration, power output, and operating hours. To forecast component failures or abnormal operating states, we built and tested several machine-learning and deep-learning architectures, specifically XGBoost, Random Forest, and Long Short-Term Memory (LSTM) networks. We gauged classification quality through accuracy, precision, recall, F1-score, and the area under the receiver operating characteristic curve (AUC), while regression outcomes were judged on mean absolute error and root mean squared error. We pinpointed which operating variables tied most directly to turbine failures by combining feature-importance rankings with sensitivity analysis. Overall, the LSTM network led across both classification and regression measures; XGBoost trailed just behind it, while Random Forest scored lowest across every metric evaluated. Across all three models, gearbox temperature and vibration proved to be the two strongest indicators of turbine faults, with generator temperature and operating hours ranking next; our sensitivity checks confirmed that estimated fault risk climbed most steeply under spikes in gearbox heat and vibration. These outcomes indicate that deep-learning and gradient boosting frameworks offer a viable foundation for catching faults early and organizing condition-based maintenance schedules during wind turbine operation.