首页|基于改进YOLOv8的风电叶片表面损伤检测与识别方法

基于改进YOLOv8的风电叶片表面损伤检测与识别方法

扫码查看
原文链接
  • NETL
  • NSTL
  • 万方数据
针对风电叶片极易出现损伤和故障,且制造和维护成本高昂等问题,提出了一种基于改进YOLOv8 模型的风电叶片表面损伤检测与识别方法.首先,将现场拍摄到的高清叶片图像作为实验数据集,并将其按比例随机划分为训练集、验证集和测试集;然后,在YOLOv8 模型中引入了动态数据增强算法Mosaic、Mixup及离线数据增强算法Albumentations,对训练数据集进行了扩充,解决了模型在有限数据集下的泛化性问题;最后,使用卷积注意力模块(CBAM)和梯度协调机制(GHM)/Focal loss算法等手段加强了模型的损伤检测能力,改进了样本分布不均衡问题,建立了一种先进的风电叶片表面损伤检测与识别方法,提升了YOLOv8 模型对叶片损伤的检测精度.研究结果表明:改进后的YOLOv8 模型在计算量和参数量都较低的情况下,其平均精度(AP)、平均召回率(AR)都超越了同等配置下的快速区域卷积神经网络(Faster R-CNN)模型.改进后的YOLOv8 模型在交并比(IoU)阈值为0.5 时的AP和AR分别达到了73.2%和58.8%,验证了该方法在风电叶片损伤检测方面具有一定的可靠性和有效性.
Surface damage detection and identification method of wind turbine blades based on improved YOLOv8
Aiming at the problem of the high manufacturing and maintenance costs of wind turbine blades,which were prone to damage and failure,a method for detecting and recognizing surface damages on wind turbine blades based on the improved YOLOv8 model was proposed.Firstly,the high-definition blade images captured in the field were used as the dataset,which was randomly divided into a training set,validation set,and test set according to the proportion.Then,the dynamic data enhancement algorithms Mosaic and Mix up were introduced,as well as the offline data enhancement algorithm Albumentations,to expand the experimental training dataset and address the generalization problem of the model under a limited dataset.Finally,the convolutional block attention module(CBAM)and gradient harmonizing mechanism(GHM)/Focal loss algorithms were employed to strengthen the model's ability to detect damage and improve the problem of sample distribution imbalance.These approaches aimed to establish an advanced wind turbine blade surface damage detection algorithm and enhance the accuracy of damage detection with theYOLOv8 model.The research results show that the improved YOLOv8 model outperforms the faster region convolutional neural network(Faster R-CNN)model with the same configuration in terms of average precision(AP)and average recall(AR)with lower computational and parameter counts.The improved model achieves an AP and AR of 73.2% and 58.8% respectively at an Intersection over Union(IoU)threshold of 0.5,verifying the reliability and effectiveness of this method in wind turbine blade damage detection direction.

wind turbine blade damage identificationYOLOv8target detectiondata augmentation algorithmconvolutional block attention module (CBAM)gradient harmonizing mechanism (GHM)average precision (AP)average recall (AR)faster region convolutional neural netw

吴博阳、毛胜轲、林特宇、任浩杰、蔡海洋、李扬

展开 >

运达能源科技集团股份有限公司,浙江 杭州 310012

风电叶片损伤识别 YOLOv8 目标检测 数据增强算法 卷积注意力模块 梯度协调机制 平均精度 平均召回率 快速区域卷积神经网络 交并比

国家重点研发计划项目

2022YFB4201205

2024

10.3969/j.issn.1001-4551.2024.07.014

机电工程
浙江大学　浙江省机电集团有限公司

机电工程

CSTPCD北大核心
影响因子：0.785
ISSN：1001-4551
年,卷(期)：2024.41(7)
  • 1
  • 11