融合大卷积核的风电锚栓裂纹检测  

Fusion of Large Convolution Kernels for Wind Power Anchor Bolt Crack Detection

作　　者：孙前来[1,2] 荆佳鹏 张帅 胡啸 刘瑞珍[1] SUN Qian-lai;JING Jia-peng;ZHANG Shuai;HU Xiao;LIU Rui-zhen(School of Electronic Information Engineering,Taiyuan University of Science and Technology,Taiyuan030024;Shanxi Normal University Modern College of Arts and ScienceTransfer Preparatory Office,Linfen 041000)

机构地区：[1]太原科技大学电子信息工程学院,山西太原030024 [2]山西师范大学现代文理学院转设筹备处,山西临汾041000

出　　处：《制造业自动化》2025年第3期142-148,共7页Manufacturing Automation

基　　金：山西省重点研发计划(202102020101005);太原科技大学博士科研启动基金(20202038);来晋工作优秀博士奖励资金(20222088);山西省高等学校科技创新项目(2023L185)。

摘　　要：风电锚栓在加工过程中通常会产生表面裂纹等缺陷,针对锚栓表面细长裂纹检测效率低、精度差的问题,提出了一种融合大卷积核的YOLOv5s网络。首先,在特征提取网络中融合大卷积核,来获得更大的有效感受野、提取更多的空间信息。其次,引入单卷积核的全维动态卷积,采用并行策略,同时学习四个不同维度的特征,不仅减少了计算量,而且提高了特征提取能力。最后添加协调注意力机制,增强对位置信息的提取能力。实验结果表明,该算法较原YOLOv5s模型在风电锚栓裂纹数据集上mAP提高了3%,FLOPs减少了21.5%,FPS达到了85帧/秒。可以满足工业生产的实时性、准确性要求。A YOLOv5s network with fused large convolution kernels is proposed to address the problem of low efficiency and poor accuracy in the detection of surface cracks on anchor bolts.Firstly,the large convolutional kernels are fused in the feature extraction network to obtain a larger effective receptive field and extract more spatial information.Secondly,the Omni-dimensional dynamic convolution with a single convolutional kernel is introduced,and a parallel strategy is used to study four different dimensions of features simultaneously,which not only reduces the computational effort but also improves the feature extraction capability.Finally,a coordinated attention mechanism is added to enhance the extraction of location information.The experimental results show that the algorithm improves mAP by 3%,reduces FLOPs by 21.5%and achieves FPS of 85.0 frames per second compared to that of the original YOLOv5s model on the wind power anchor bolt crack dataset,which can meet the real-time and accuracy requirements of industrial production.

关 键 词：Yolov5s 锚栓裂纹检测 全维动态卷积 大卷积核 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]