Abstract: The key to determining whether the wind turbine can operate properly is to detect surface defects on the turbine blades.After field inspections and the dataset making,a defect detection algorithm based on improved YOLOv7 named YOLOv7-BAHS was proposed to address the unique challenges in surface defect detection of generators and the shortcomings of existing deep learning methods.This algorithm incorporated the BoTNet network into the original model to reduce parameter count,and thus the speed of computation was improved.Additionally,the SiLU activation function was replaced with Hardswish to optimize sample imbalance,increase convergence speed and improve regression accuracy.Furthermore,the ACmix module was introduced in the Head layer to fuse convolution and self-attention mechanisms,thereby enhancing the model's predictive capability and improving detection performance.Finally,the SPPCSPC structure was optimized by designing a new SPPCSPC_N module,which reduced computational complexity while maintaining the same receptive field.Experimental results demonstrated that compared to other detectors,the proposed approach achieved 6.5% increase in mAP value and 7% increase in F1-score,and the speed was improved from 52.36 FPS to 74.66 FPS.

Key words: YOLOv7, wind power generation, defect detection, deep learning, attentional mechanisms

摘要： 风力发电机表面缺陷检测是判断风机能否正常运行的关键。在实地考察并制作数据集之后,针对发电机表面缺陷检测特有难点,及现有深度学习方法的不足,提出基于改进YOLOv7的缺陷检测算法——YOLOv7-BAHS算法。该算法首先在原网络中引用BoTNet网络,使模型参数量降低,进而提高计算速度,并用Hardswish激活函数替代SiLU激活函数,优化样本不平衡的问题,提高收敛速度和回归精度。其次在Head层引入ACmix模块,将卷积和自注意力进行融合,增强模型的预测能力,进而提升检测效果。最后将改进的Spatial Pyramid Pooling结构（SPPCSPC）进行优化,设计出新的SPPCSPC_N模块,可以在保持感受野不变的情况下减少计算量。结果表明,该方法相比其他检测器,mAP值提升6.5%,F1分值提升7%,速度从52.36 FPS提升至74.66 FPS。

关键词: YOLOv7, 风力发电, 缺陷检测, 深度学习, 注意力机制