Abstract:

To solve the issues of wind turbine blades in terms of classification difficulty and blurry segmentation of small defects in surface defect detection, an improved U-Net semantic segmentation network was proposed based on dilated convolution and convolutional attention modules. Based on the encoding-decoding structure of the network model, a transferable VGG16 feature extraction layer was used to replace the encoding part of the U-Net network. Then, a convolutional attention module was added to the skip module between encoding and decoding. The global weight was enhanced by selecting small defect information. Dilated convolution was used in the decoding section to enhance the network's feature extraction ability, and the pre-trained VGG16 model was used to realize transfer learning. The hybrid loss function of Focal and Dice was validated against the models of DeeplabV3+, Pyramid Scene Parsing Network(PSPnet), High-Resolution Network(HRNet), and U-Net. The results showed that the improved U-Net network had higher prediction accuracy in blade defect classification and segmentation tasks, mean intersection over union, mean pixel accuray, and recall values were 83.60%, 92.84%, and 88.50%, respectively. The mean intersection over union simulated by the improved U-Net model was 13.98% and 9.38% higher than that by the DeeplabV3+ and standard U-Net model, respectively. Therefore, the proposed model can improve the sensitivity of blade defect detection, effectively reduce false positives of detection results, and provide guidance to wind turbine blade defect detection.

Key words: attention mechanism, U-Net network, wind turbine blades defect, image segmentation, transfer learning, convolutional block attention module (CBAM)