Deep learning-based fatigue damage segmentation of wind turbine blades under complex dynamic thermal backgrounds

Abstract

Passive thermography is an efficient method to inspect fatigue damage of large-scale structures such as wind turbine blades under cyclic loads. Quantitative damage evaluation often requires the damage region to be segmented from the thermal image, which challenges conventional image process techniques, especially when the structure is moving, and the thermal background is changing. This study proposes a model based on deep learning and thermography to automatically segment complex dynamic background of images taken from a wind turbine blade during cyclic loading and subsequently segment the fatigue blade damages. An automated background segmentation algorithm is developed to isolate the blade from the background using six state-of-the-art deep learning models. The most accurate model is then chosen and improved for the second step of damage segmentation, achieving a level of accuracy comparable to that of human observation, even with fewer images in the training process. The proposed background and damage segmentation methods have recall of 99% and 82%, respectively, indicating that the proposed approach is accurate, efficient, and robust.