摘要
漏磁内检测是长输管道主要检测方式。目前漏磁内检测数据分析中的缺陷检测方法环境适应性差,需要大量样本为不同环境分别建立检测模型,难以满足实用需求。本文提出一种自监督的缺陷检测方法,可以在少量样本下建立精确缺陷检测模型,克服目前缺陷检测方法需要大量样本才能训练精确模型的问题,并且在不同环境下检测效果都有所提升,因此具有良好的适用性和迁移性。首先对漏磁内检测缺陷数据进行预处理,接着将缺陷数据自适应的可视化,然后利用视觉表示对比学习的简单框架(SIMCLR)对可视化后的缺陷进行训练获得预训练权重,最后采用深度学习完成对缺陷的识别与定位。试验研究表明,本文设计的自监督检测方法能够有效解决可标记数据少的问题,具有检测精度高,迁移性好,泛化能力强的特点。
Magnetic flux leakage internal detection is the main detection method for long-distance pipelines.Current defect detection method in magnetic flux leakage internal detection data analysis has poor environmental adaptability,require a large number of samples to establish detection models for different environments,which is difficult to meet practical requirement.This paper proposes a self-supervised defect detection method,which can establish an accurate defect detection model with a small number of samples,overcomes the problem that current defect detection method requires a large number of samples to train an accurate model,and the detection effects are all improved in different environments.So the proposed method has good applicability and migration.Firstly,the defect data in the magnetic flux leakage internal detection are pre-processed,the defect data are adaptively visualized,then simple framework for contrastive learning of visual representations(SIMCLR)is sued to train the visualized defects to obtain pre-training weights,and finally deep learning is used to complete the identification and localization of the defects.Experiment research shows that the self-supervised detection method designed in this paper can effectively solve the problem of less tagged data,and has the characteristics of high detection accuracy,good migration and strong generalization ability.
作者
刘金海
赵贺
神祥凯
鲁丹宇
唐建华
Liu Jinhai;Zhao He;Shen Xiangkai;Lu Danyu;Tang Jianhua(School of Information Science and Engineering,Northeastern University,Shenyang 110004,China;Shenyang Zhigu Technology Co.,Ltd.,Shenyang 110004,China;Energy Development Equipment Technology Co.,Ltd.,CNOOC,Tianjin 300452,China)
出处
《仪器仪表学报》
EI
CAS
CSCD
北大核心
2020年第9期180-187,共8页
Chinese Journal of Scientific Instrument
基金
国家重点研发计划(2017YFF0108800)
国家自然科学基金(61973071,61627809,61703087)
辽宁省自然科学基金(2019-KF-03-04)项目资助
关键词
缺陷检测
深度学习
自监督
视觉表示对比学习的简单框架
defect detection
deep learning
self-supervision
simple framework for contrastive learning of visual representations
