A dual-branch transformer-based approach for corrosion detection of lightning receptors in wind turbine blades

With the accelerating global energy transition, wind power has become a core clean energy technology attracting extensive attention. However, the increasing size of wind turbine blades has markedly elevated lightning strike risks. Surface corrosion of lightning receptors, key components of lightning protection systems, severely degrades electrical conductivity and current dissipation efficiency. To address the limited recognition accuracy of existing vision-based detection methods under complex conditions such as strong reflections, occlusions and non-uniform illumination, this study proposes a dual-branch fusion architecture, the Swin-BiFormer fusion network (SB-FNet). The model adopts Swin Transformer (Swin-T) as the backbone and integrates BiFormer’s dynamic sparse attention mechanism to capture global contextual information. An attention gate (AG) mechanism is introduced at intermediate layers to adaptively fuse local detail and global semantic features, enhancing robustness and generalisation under varying illumination, viewpoint and occlusion conditions. Experimental results demonstrate that SB-FNet achieves 94.32% classification accuracy on lightning receptor corrosion images, outperforming ResNet-18, ResNet-50 and Swin-T. The model provides precise, quantitative evaluation of lightning receptor service states, offering essential technical support for predictive maintenance and improved operational safety of wind farms.