Authors: Edwin Nazario, Mason Davis, Mohammad Shekaramiz, Mohammad Masoum, Abdennour Seibi
Mentors: Mohammad Shekaramiz
Insitution: Utah Valley University
Wind turbine blade maintenance is expensive, time exhaustive, and prone to human error and misdiagnosis. With our energy production rapidly increasing in the green sector, these issues are only exacerbated. As wind energy production is expanded in Utah, and the wider Mountain West, the capital cost of wind turbine damage and the subsequent downtime of the turbines will cause strain on our energy output capabilities. One such solution is the development of autonomous classification and identification of these anomalies through deep learning. In this research, a novel dataset is created using a small wind turbine and multiple deep learning architectures and techniques are deployed for comparative analysis. Here, ResNet-50, VGG-19, Xception, and a custom CNN are deployed for the purpose of anomaly detection. Transfer learning is also investigated for further performance gains with each of the existing architectures as the backbone network.
For this research, a new dataset was created that combines both indoor and outdoor images of a small wind turbine. A total of 6 blades were used, 3 representing healthy and 3 representing faulty, resulting in 6000 images. The faults on the blades had a combination of cracks, holes, and erosion to simulate the damage found on commercial grade turbine blades. Indoor images were taken using different cameras and backgrounds to simulate human inspection. To introduce realistic environmental features to the dataset, such as sunlight and clouds, drones were utilized for outdoor imaging.
After extensive hyperparameter search and simulations, it was found that the Xception architecture provided the best classification accuracy of 99.33% followed by ResNet-50 and VGG-19 attaining 98.412% and 97.418%, respectively. This accuracy shows promising performance in the autonomous detection of wind turbine faults for the purpose of health monitoring and maintenance scheduling. To expand on our work, fault localization and size analysis techniques will be investigated to provide more detailed information to maintenance personnel.