电子测量与仪器学报

曾勇杰,范必双,杨涯文,蒋冲.改进YOLOV8算法在风机叶片缺陷检测上的应用[J].电子测量与仪器学报,2024,38(8):26-35

改进YOLOV8算法在风机叶片缺陷检测上的应用

YOLOv8 algorithm is improved in the defect detection ofwind turbine blades applications

DOI：

中文关键词: 风机叶片缺陷检测 YOLOv8n 高效多尺度卷积 Dyhead 损失函数

英文关键词:wind turbine blades defect detection YOLOv8n efficient multi-scale convolution Dyhead loss function

基金项目:国家自然科学基金（52277077）项目资助

作者	单位
曾勇杰	长沙理工大学电气与信息工程学院长沙410000
范必双	长沙理工大学电气与信息工程学院长沙410000
杨涯文	长沙理工大学电气与信息工程学院长沙410000
蒋冲	长沙理工大学电气与信息工程学院长沙410000

Author	Institution
Zeng Yongjie	School of Electrical and Information Engineering, Changsha University of Science and Technology, Changsha 410000，China
Fan Bishuang	School of Electrical and Information Engineering, Changsha University of Science and Technology, Changsha 410000，China
Yang Yawen	School of Electrical and Information Engineering, Changsha University of Science and Technology, Changsha 410000，China
Jiang Chong	School of Electrical and Information Engineering, Changsha University of Science and Technology, Changsha 410000，China

摘要点击次数: 92

全文下载次数: 132

中文摘要:

风力发电机叶片，作为风力发电系统的核心组件，其健康状况直接关乎整个发电效率与运行安全。针对叶片缺陷检测的挑战，深入研究了YOLOv8n网络，并创新性地提出了高效多尺度卷积模块（EMSConv），该模块有效替代了传统残差块中的卷积层，通过分组卷积技术显著降低了冗余特征对检测结果的干扰，从而提升了检测的精确性。此外，在检测头部分，融入了Dynamic Head的多元化注意力机制，这些自注意力机制协同工作，跨越不同特征层，实现了对目标尺度、空间位置及检测任务的精准感知，极大地增强了目标检测模块的综合能力。还创新性地整合了Inner-IoU、Wise-IoU与MPDIoU，创造性地提出了Inner-Wise-MPDIoU，以替代传统的CIoU损失函数，这一举措不仅提高了网络的检测精度，还加速了收敛过程。在针对自制风机叶片缺陷数据集的测试中，YOLOv8-EDI展现出了卓越的性能，其mAP50值高达81.0%，相比原始YOLOv8n提升了2.3%；召回率也达到了76.8%，提升了3.7%。该模型在提升检测效果的同时，还实现了计算量的降低，降幅达5.5%，充分满足了工业环境下对风机叶片进行高效、准确、大批量检测的需求。

英文摘要:

Wind turbine blades, being the core component of wind power generation systems, have their health status directly impacting the overall power generation efficiency and operational safety. Addressing the challenges of blade defect detection, researchers delved into the YOLOv8n network and innovatively proposed the Efficient Multi-Scale Convolutional module (EMSConv). This module effectively replaces the convolutional layers in traditional residual blocks, significantly reducing the interference of redundant features on detection results through grouped convolution techniques, thereby enhancing detection accuracy. Furthermore, in the detection head, a diverse set of attention mechanisms from Dynamic Head are incorporated. These self-attention mechanisms work in concert, spanning across different feature layers, to achieve precise perception of target scales, spatial locations, and detection tasks, vastly strengthening the comprehensive capabilities of the target detection module. Moreover, researchers innovatively integrated Inner-IoU, Wise-IoU, and MPDIoU, creatively proposing Inner-Wise-MPDIoU to replace the traditional CIoU loss function. This move not only improved the network’s detection precision but also accelerated the convergence process. During testing on a self-made dataset of wind turbine blade defects, YOLOv8-EDI exhibited remarkable performance, achieving an mAP50 value of 81.0%, a 2.3% increase compared to the original YOLOv8n. The recall rate also reached 76.8%, marking a 3.7% improvement. While enhancing detection performance, this model managed to reduce computational requirements by 5.5%, fully meeting the need for efficient, accurate, and large-scale blade defect detection in industrial settings.

查看全文查看/发表评论下载PDF阅读器