基于深度学习的油封缺陷检测方法研究  

Oil seal defect detection method based on deep learning

作　　者：夏桂方 于正林[1] XIA Gui-fang;YU Zheng-lin(School of Mechanical and Electrical Engineering,Changchun University of Science and Technology,Changchun 130022,China)

机构地区：[1]长春理工大学机电工程学院,吉林长春130022

出　　处：《机电工程》2023年第1期69-75,共7页Journal of Mechanical & Electrical Engineering

基　　金：吉林省科技厅重点资助项目(20190302069GX)。

摘　　要：在传统的油封缺陷检测中,存在油封表面质量缺陷尺寸小、人工检测效率低、漏检错检率高、成本高等问题,为此,以油封缺陷为研究对象,提出了一种基于深度学习Faster R-CNN框架的多尺度特征融合的改进算法。首先,构建了油封缺陷检测系统,采集了油封缺陷图像,经扩增及标注等预处理后制作了数据集;然后,研究了油封缺陷尺寸较小导致的识别精度低问题,设计了Faster R-CNN网络基于FPN+ResNet50框架进行特征多尺度融合改进的方法;最后,采用了预训练参数送入改进的Faster R-CNN网络模型,并对油封缺陷数据集进行深度训练的方法,进行了油封缺陷的检测实验。研究结果表明:该模型的检测精确度和速度综合性能优于固有的Faster R-CNN网络模型,划痕、毛刺和凹缺的检测精确度分别达到0.96、0.95和0.97,召回率分别达到0.89、0.88和0.91,mAP可达85.5%,高于改进前模型1.4%,识别速度可达16 fps,高于油封生产速度;该检测方法可以满足油封缺陷的检测要求。Aiming at the problems of small size of defects in the surface quality of automobile oil seals,low efficiency of manual detection,high rate of missed detection and false detection,and high cost,an optical detection method for automobile oil seal defects based on the improved deep learning Faster region with convolutional neural network(R-CNN)algorithm was proposed.Firstly,the oil seal defect detection system was constructed,the oil seal defect images were collected,and the data set was made after preprocessing such as amplification and labeling.Then,the problem of low recognition accuracy caused by the small size of the oil seal defect was studied,and the Faster R-CNN network was designed based on the feature pyramid network(FPN)+ResNet50 framework for feature multi-scale fusion and improvement.Finally,the oil seal defect detection experiment was carried out by using the pre-training parameters to send into the improved Faster R-CNN network model and deeply training the oil seal defect data set.The research result indicates,it can be seen that the comprehensive performance of the proposed model is better than the inherent Faster R-CNN network model,and the detection accuracy of scratches,burrs and dents reaches 0.96,0.95 and 0.97,respectively.The recall rate reaches 0.89,0.88 and 0.91,and the mean average precision(mAP)can reach 85.5%,which is 1.4%higher than the model before improvement,and the recognition speed can reach 16 fps,which is higher than the oil seal production speed,so it can meet the detection requirements of oil seal defects.

关 键 词：机械密封 主唇口问题 密封失效 Faster R-CNN网络模型 深度训练 多尺度特征融合 

分 类 号：TH136[机械工程—机械制造及自动化] TB114.3[理学—概率论与数理统计]