Blade Segmentation Based on Improved FCN Network

To address the issues of limited dataset size and inaccurate fault region localization, this paper proposes an improved fully convolutional neural network (CBAM-FCN8s) based on the attention mechanism convolutional block embedding. This method introduces the CBAM module into the classic FCN8s network, enhancing attention to key regions and features through the attention mechanism, thereby improving fault segmentation and localization accuracy. To tackle the issue of insufficient datasets, this paper employs data augmentation techniques, such as random cropping, flipping, and scaling, to enhance the model’s generalization ability. For wind turbine blade faults in complex scenarios, combining the pixel-level segmentation capability of FCN with the CBAM module allows for precise segmentation of the fault area and enhances attention to key regions. Experimental results show that the CBAM-FCN8s network improves mean pixel accuracy and segmentation accuracy by 2.52% to 11.63% and 2.06% to 5.60%, respectively, compared to the traditional FCN32s, FCN16s, and FCN8s, significantly improving the accuracy and localization precision of fault segmentation.