Adaptive wind turbine blade surface defect detection based on deep learning and transfer learning

Abstract

Addressing the issues of low efficiency and insufficient accuracy in traditional methods, a novel adaptive automatic detection framework for wind turbine blade surface defects is proposed, which combines a deep learning model with a transfer learning strategy. First, the super-resolution generative adversarial network deep learning model, which has been pre-trained using transfer learning, is employed to perform super-resolution reconstruction on the acquired images of wind turbine blade surface defects. Second, the pretrained weights from transfer learning and frozen training are utilized to optimize the training speed and detection accuracy of the You Only Look At Coefficients deep learning model. This optimized model is then employed to detect multiple types of defects on the blade, resulting in instance segmentation outcomes for the surface defects on the blade. Finally, defect detection was conducted on four types of blade surface defects at a wind farm in western China and across multiple datasets. The results demonstrated that the integration of deep learning with transfer learning achieved high-precision instance segmentation detection for various multi-scale defect types. Furthermore, this approach enabled the direct output of masks, locations, and size information for blade surface defects, providing a foundation for the assessment of blade surface defects. This study offers a novel solution for the detection of surface defects on wind turbine blades and provides valuable references for the application of deep learning in related fields.