基于注意力机制和隐马尔科夫的高精密螺纹全自动精确测量  

Automatic and Accurate Measurement of High Precision Thread Based on AA R2Unet and HMM

作　　者：张堃[1,5] 李子杰 瞿宏俊 吴建国 华亮 ZHANG Kun;LI Zijie;QU Hongjun;WU Jianguo;HUA Liang(School of Electrical Engineering,Nantong University,Nantong 226019,China;Hito Technology Co.,Ltd.,Wuxi 214122,China;Department of Electronic Engineering,Jiangnan University,Wuxi 214122,China;School of Electrical and Energy Engineering,Nantong Institute of Technology,Nantong 226002,China;Nantong Key Laboratory of Intelligent Control and Intelligent Computing,Nantong 226002,China)

机构地区：[1]南通大学电气工程学院,江苏南通226019 [2]中科海拓(无锡)科技有限公司,江苏无锡214122 [3]江南大学物联网工程学院,江苏无锡214122 [4]南通理工学院电气与能源工程学院,江苏南通226002 [5]南通市智能控制与智能计算重点实验室,江苏南通226002

出　　处：《南通大学学报（自然科学版）》2021年第3期57-66,共10页Journal of Nantong University(Natural Science Edition)　

基　　金：江苏省高校自然科学基金项目(18KJB510038);江苏省“333工程”项目(BRA2018218);国家级大学生创新创业训练计划项目(202010304065Z)。

摘　　要：基于机器视觉的螺纹测量易受到工业环境(例如灰尘、铁屑、油渍等)的干扰,且需要人工半自动干预,导致测量结果不稳定。通过加入Attention机制对R2Unet模型进行改进,提出一种基于AA R2Unet深度学习模型和隐马尔科夫模型的高精密螺纹全自动精确测量方法。首先,为了克服工业环境中灰尘、铁屑等因素的干扰,设计了AA R2Unet模型对外螺纹进行有效边缘识别与提取;然后,通过计算螺纹边缘点梯度方向特征信息,使用隐马尔可夫模型对螺纹边缘点进行分类,达到螺纹零件在测量过程中可以任意角度放置的目的。通过实际采集工件图像制作数据集进行实验验证,结果表明,基于AA R2Unet的螺纹边缘提取方法分割精度达到95.92%,基于隐马尔可夫模型的螺纹边缘点分类准确率达到86%以上,外径测量误差在0.01 mm以内。The thread measurement methods based on machine vision are easily disturbed by the environment(e.g.dust,iron filings,oil stains,etc.),resulting in inaccurate measurement results.This paper improves the R2Unet model by adding the Attention mechanism,and proposes an external thread measurement method based on AA R2Unet and hidden Markov model(HMM).Firstly,to overcome the interference of dust,iron filings et al,the AA R2Unet model was designed to identify and extract the external threads.Secondly,the feature information on gradient direction of thread edge points is calculated,HMM was used to classify the thread edge points so that the threaded parts can be placed at any angle during the measurement.Finally,the method with the gathered dataset was evaluated.The results show that the segmentation accuracy of the thread edge extraction method based on AA R2Unet is 95.92%,the classification accuracy of thread edge points based on HMM is above 86%and the comprehensive measurement error is within 0.01 mm.

关 键 词：Unet模型 注意力机制 隐马尔可夫 螺纹测量 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术] TP24[自动化与计算机技术—计算机科学与技术]