摘要
压力容器对接纵焊缝是重要受力部位,对其测量是焊接质量评估的重要环节,从焊缝激光线中提取焊缝参数测量特征点是视觉焊缝测量重要内容。针对深度学习的焊缝图像参数特征点提取网络,其复杂骨干网络在模型部署、模型训练、模型正向传播时间成本而带来检测实时性限制问题,研究基于深度学习的压力容器焊缝图像特征点提取网络、焊缝参数特征点提取的相邻阶段数据融合KD模型的实时性提升方法,并以压力容器焊缝实验对象进行验证。结果表明,知识蒸馏KD训练后学生模型特征点提取精度有明显提升,且部署时间、单张图片提取时间分别减少70.8%、79.7%。
The pressure vessel butt longitudinal weld is an important force part,and its measurement is an important part of the weld quality assessment,and the extracting of weld parameter measurement feature points from the weld laser line is an important content of visual weld measurement.This paper addresses the deep learning weld image parameter feature point extraction network with complex backbone network in model deployment,model training,model forward propagation time cost and bring detection real-time limitation problem.The real-time enhancement methods of the pressure vessel weld image feature point extraction network based on deep learning and the adjacent stage data fusion knowledge distillation(KD)model for weld parameter feature point extraction are investigated and validated with pressure vessel weld experimental objects.The results show that the student model feature point extraction accuracy is significantly improved after knowledge distillation KD training,and the deployment time and single image extraction time are reduced by 70.8%and 79.7%,respectively.
作者
廖普
刘桂雄
杨宁祥
LIAO Pu;LIU Guixiong;YANG Ningxiang(School of Mechanical and Automotive Engineering,South China University of Technology,Guangzhou 510640,China;Guangdong institute of Special Equipment Inspection and Research,Zhuhai Guangdong 519000,China)
出处
《激光杂志》
CAS
北大核心
2023年第3期69-74,共6页
Laser Journal
基金
国家市场监督管理总局科技计划项目(No.2019MK143)。
关键词
压力容器
焊缝表面参数
深度学习
模型压缩
pressure vessel
weld surface parameters
deep learning
model compression
