基于聚类的航天器多余物粒径特征识别方法  

A cluster-based method for identifying the feature of particle size for spacecraft remnant

作　　者：刘海江[1] 张恒 汪乾 刘劲松 LIU Haijiang;ZHANG Heng;WANG Qian;LIU Jinsong(School of Mechanical Engineering,Tongji University,Shanghai 201804,China)

机构地区：[1]同济大学机械与能源工程学院,上海201804

出　　处：《航天器环境工程》2022年第1期111-118,共8页Spacecraft Environment Engineering

基　　金：上海市科学技术委员会基金项目(编号:15111103400)。

摘　　要：针对焊锡粒多余物粒径特征识别过程中,粒径区分度不足和粒径特征参数类间交叉对分类准确率的不利影响,提出基于聚类的高精密航天器多余物粒径特征识别方法。从信号时域与频域分析技术出发,选取多个特征参数构建多余物粒径初始特征参数向量;采用Fisher比量化各个特征参数对粒径的区分能力并削除贡献率较低的特征参数,从而构建最终多余物粒径特征识别模型;用K均值聚类算法对无标记的不同粒径等级训练样本进行学习后揭示不同粒径等级下输入特征参数的分布规律,实现混合粒径的识别。验证试验表明,在含单个和2个多余物的情况下,多余物粒径的总体识别准确率达81.8%,满足实际要求。In view of the undesirable effect of the classification accuracy in the process of the particle size identification of the excess solder particles caused by poor particle size discrimination and the intersection of the particle size characteristic parameters,a cluster-based method for the particle size identification is proposed.Based on the signal analysis techniques in the time-domain and the frequency-domain,we select a group of characteristic parameters to construct the eigenvectors for representing the particle size.The Fisher ratio is used to quantify the distinguishing ability of each feature parameter for the particle size identification,and those feature parameters with a low contribution rate are ignored in constructing the final particle size feature identification model.The K-means clustering algorithm is used to learn from the unlabeled training samples of different particle size levels and to obtain the distribution of the input characteristic parameters for different levels of particles,to realize the size identification for the mixed particle size.Testing results show that the overall recognition accuracy rate reaches 81.8%in the case of a single or two redundant objects,which can meet the application requirement.

关 键 词：高精密航天器 多余物 粒径识别 特征参数 Fisher比 K均值聚类算法 

分 类 号：TB52.9[理学—物理] V416.6[理学—声学]