摘要
磁探测技术用于金属目标物的探测等领域,具有分辨率高,成本低,操作方便等优点,但磁探测难以实现对探测目标的有效分类,且虚警率较高,使得后续排爆工作仍存在很大难度。本文通过构建典型场景,获取磁力仪探测金属目标物的数据,利用经典机器学习分类方法以及神经网络模型开展对金属目标物的磁探测目标识别分类研究。结果表明,经典机器学习方法的识别分类效果优于深度学习模型,其中随机森林方法效果最好,宏平均正确率达到96%。本文还进一步研究了各项特征对最终分类效果的贡献程度,通过暴力遍历、自编码器、PCA和LDA方法实现特征的提取,从中挖掘出影响识别分类效果的主要特征,通过对比分析发现由第一磁探测器信号数据组成的第一空间特征就可以较为准确的对各类弹药种类进行分类,而第三磁探测器的Z方向数据是对分类结果影响最大的单一特征。研究成果有利于提高磁探测爆炸装置的识别准确率,并可以指导相关探测设备的设计和应用。
Magnetic detection technology is widely used in the detection of metallic objects, because of its high resolution, low cost and convenient operation. However, magnetic detection also has high false alarm rate and can not classify targets, which increases the difficulty of following disposal work. In this paper, we use classical machine learning classification methods and convolutional neural networks to detect metallic objects. The results show that the classical machine learning method is better than the deep learning model, in which the random forest method is the best, and the macro average accuracy rate reaches 96%. Besides, we also use brute-force-traversal, auto encoder(AE), principal component analysis(PCA) and linear discriminant analysis(LDA) to extract features which have contributions to metal detection, and found the first magnetic detector signal data can be more comprehensive to classify various types of targets. In addition, the third Z direction of the magnetic detector data has the greatest impact on the classification results.
作者
张远鹏
单新伟
汪海涛
郝尚鹏
沙灜
李静海
Zhang Yuanpeng;Shan Xinwei;Wang Haitao;Hao Shangpeng;Sha Ying;Li Jinghai(State Key Laboratory of NBC Protection for Civilian,Beijing 102205,China;Hubei Engineering Technology Research Center of Agricultural Big Data,Huazhong Agricultural University,College of Informatics,Wuhan 430070,China;Nanjing Buyiren Technology Co.Ltd,Nanjing 320571,China)
出处
《科技通报》
2022年第6期60-65,共6页
Bulletin of Science and Technology
关键词
磁探测
金属目标物
机器学习
目标分类
magnetic detection
metallic object
machine learning
target classification
