<p>Accurate detection of wind turbine blade damage is critical for enhancing power generation efficiency and extending equipment lifespan. To address the limitations of multi-scale target recognition in wind turbine blade defect detection, a novel model named MSA-YOLO is introduced, which integrates multi-scale feature fusion and attention mechanisms for enhanced detection accuracy. A lightweight backbone network, ACINet, is developed by combining channel attention and dynamic channel shuffling to enable efficient extraction of image features while improving computational performance. Additionally, the MSBlock and ADown modules are integrated into the neck of the network to facilitate effective multi-scale feature fusion and optimization, thereby enhancing detection accuracy and robustness under complex visual conditions. The model further incorporates the Adaptive Spatial Feature Fusion Head(ASFFHead) detection head, which adaptively fuses features across different scales to optimize detection outcomes. Experimental results demonstrate that MSA-YOLO outperforms the original YOLO11, achieving a 6.4% improvement in mAP@0.5 and a 9.5% improvement in mAP@0.5–0.95, with only 2.9 M parameters and a computational cost of 6.1 Giga Floating-point Operations Per Second(GFLOPs). These characteristics make the proposed method highly suitable for accurate multi-scale defect detection in memory-constrained deployment scenarios.</p>