Wind Turbine Surface Defect Detection Analysis from UAVs Using U-Net Architecture

Effective and timely identification of defects for wind turbine surface are crucial to propitiously repair and limit any further degradation. To date, the damage is detected via drones, on-site high-resolution photography, or physical examination, which is usually done after an annual rotor blade assessment. Obviously, any damage that goes unnoticed can worsen over time, raising the cost of repair and maintenance while also increasing the potential of catastrophic collapse. Maintenance after the discovery of a blade problem, or in the worst-case scenario, blade failure, can result in substantial repair and replacement expenses and significant revenue losses, especially in the case of offshore wind farms. If the damage had been discovered sooner, it could have been fixed for a far lower price. In this work, we propose an automated wind turbine surface distress analysis system based on the UNet deep learning framework. The system is trained using images acquired from UAVs and tested on several different datasets. The detection and classification accuracy of the proposed distress analyser is measured using the precision and recall values. Successful application of this study can help identify structure anomalies in need of urgent repair, thereby facilitating a much better civil infrastructure monitoring system. Experimental validation demonstrated the ability of the method to detect micro-crack locations with an accuracy of 96%. Experimental results are presented and discussed.