第61卷第8期Vol.61No.82023年8月August2023农业装备与车辆工程AGRICULTURALEQUIPMENT&VEHICLEENGINEERINGdoi:10.3969/j.issn.1673-3142.2023.08.019基于改进YOLOv4风机叶片缺陷检测方法高文俊，张海峰（201620上海市上海工程技术大学机械与汽车工程学院)[摘要]随着风电行业发展，风电设备越来越普及。风力发电设备长期处于恶劣环境下，设备叶片关键部件会发生损伤，降低整机发电效率。为了解决风力发电机叶片传统检测耗时长、效率低、精度低等问题，提出一种基于改进YOLOv4的风机叶片缺陷检测方法。首先采用GhostNet特征提取网络更换原有YOLOv4的特征提取网络，使得模型轻量化的同时保持良好的检测精度；其次，采用基于COCO数据集权重的迁移学习，减少模型训练时间并加快模型收敛；最后，采用Focalloss分类损失函数解决数据集缺陷类别不平衡问题，且使得目标检测模型收敛。实验结果表明，相比原有的YOLOv4，map值提高了3.66%且能满足实时性需求。[关键词]改进YOLOv4；风机叶片；缺陷检测；特征提取[中图分类号]TM315[文献标志码]A[文章编号]1673-3142(2023)08-0094-05引用格式：高文俊,张海峰.基于改进YOLOV4风机叶片缺陷检测方法[J].农业装备与车辆工程,2023,61(8):94-98.WindturbinebladedefectdetectionbasedonimprovedYOLOv4GAOWenjun,ZHANGHaifeng(CollegeofMechanicalandAutomotiveEngineering,ShanghaiUniversityofEngineeringScience,Shanghai201620,China)[Abstract]Withthedevelopmentofwindpowerindustry,windpowerequipmentisbecomingmoreandmorepopular.However,thelong-termadverseenvironmentofwindpowerequipmentwillcausedamagetothekeycomponentsofthebladeandreducethepowergenerationefficiencyofthewholemachine.Inordertosolvetheproblemsoflongtimeconsuming,lowefficiencyandlowprecisionintraditionalwindturbinebladedetection,adefectdetectionmethodbasedonimprovedYOLOv4fanbladewasproposed.Firstly,GhostNetfeatureextractionnetworkwasusedtoreplacetheoriginalYOLOv4featureextractionnetwork,whichmadethemodellighterandkeptgooddetectionaccuracy.Secondly,transferlearningbasedonCOCOdatacentralizationwasusedtoreducemodeltrainingtimeandacceleratemodelconvergence.Finally,theFocalLossclassificationfunctionwasusedtosolvetheproblemofunbalanceddefectcategoriesinthedatasetandmakethetargetdetect...
