Integration of infrared thermography and convolutional neural networks for early detection of roller damage in wind turbine main bearings

The increased concern in wind turbine main bearing reliability due to ever growing turbine rotor sizes is driving increased research into more effective main bearing fault detection strategies. This study presents a data-driven framework for detecting roller damage in wind turbine main bearings using convolutional neural networks (CNN) and thermal imaging. Thermal images were collected from an experimental test rig, with both healthy and damaged bearings. A CNN model trained on these images achieved an initial classification accuracy of 83%. After hyperparameter optimization using grid search, the model's accuracy improved to 93.4%, with similar improvements in recall, precision, and F1-score. The confusion matrix showed a reduction in misclassifications from 10 to 5, and the area under curve (AUC) increased from 0.90 to 0.95, highlighting a significant improvement in performance. By leveraging thermal imaging, the proposed method offers a non-invasive, efficient approach for early fault detection in wind turbine bearings, supporting predictive maintenance and reducing the risk of critical failures.