基于改进YOLO算法的多目标铁谱磨粒智能识别  

Intelligent Recognition of Multi-Objective Ferrographic Wear Particles Based on Improved YOLO Algorithm

作　　者：张子杨 魏海军[1] 刘竑[1] 贾风光 ZHANG Ziyang;WEI Haijun;LIU Hong;JIA Fengguang(Merchant Marine College,Shanghai Maritime University,Shanghai 201306,China;School of Marine and Marine Engineering,Shandong Jiao Tong University,Jinan Shandong 264209,China)

机构地区：[1]上海海事大学商船学院,上海201306 [2]山东交通学院船舶与轮机工程学院,山东济南264209

出　　处：《润滑与密封》2021年第5期27-33,共7页Lubrication Engineering

摘　　要：为提高铁谱磨粒中相似磨粒的识别率,降低小颗粒磨粒的漏检率,并确保检测速度的实时性,基于YOLO算法,提出了大尺度yolo层检测(yolov3mod)和全尺度yolo层检测(yolov35l)两种改进模型。该改进模型通过添加空间金字塔池化模块、拓展yolo层尺度,来改善网络结构,提高了相似磨粒的识别率,降低了小颗粒磨粒的漏检率;通过融合卷积层与批量归一化(BN)层,减少了模型计算量,提高了模型检测速度。实验结果表明:与原始模型相比,yolov3mod模型对相似磨粒的识别率提高了8%,总平均准确率提高了5%,yolov35l模型对相似磨粒的识别率提高了14%,总平均准确率提高了10%;2种改进模型的推理速度相比原始模型提高了8%,且磨粒的定位更加精确,基本实现了复杂背景下多目标磨粒的识别;yolov3mod拥有较快的检测速度,yolov35l则有着更高的检测精度,可根据实际工况需求进行取舍。To improve the recognition rate of similar abrasive particles in ferrous abrasive particles,reduce the missed detection rate of small particles,and ensure the real-time detection speed,two improved models,large detection region of yolo layers(yolov3mod)and fully detection region of yolo layers(yolov35 l),were proposed based on the YOLO algorithm.The two improved models increase the recognition rate of similar abrasive particles and reduce the missed detection rate of small particles by adding the Spatial Pyramid Pooling Module(SPP Module)and expanding the scale of the yolo layer to improve the backbone network structure,and reduce the amount of model calculation and increase the speed of model detection by merging the convolutional layer with batch normalization(BN)layer.The experimental results show that compared with the original model,the recognition rate of yolov3mod for similar abrasive particles is increased by 8%and the mean average precision is increased by 5%,and the recognition rate and mean average precision of yolov35 l are increased by 14%and 10%,respectively.Compared with the original model,the inference speed of the two improved models is increased by 8%,and the models have better positioning accuracy,which basically realize the recognition of multi-target wear particles in complex background.The yolov3mod model has a faster detection speed,and the yolov35 l model has a higher detection accuracy,which can be selected based on actual working conditions.

关 键 词：铁谱磨粒 空间金字塔池化模块 卷积层 批量归一化层 

分 类 号：TH117.1[机械工程—机械设计及理论]