基于参数调优Xgboost算法的多余物信号检测技术  

Excess signal detection technology based on parameter tuning Xgboost algorithm

作　　者：李超然 赵娜靖 李硕 王国涛[1,2] LI Chao-Ran;ZHAO Na-Jing;LI Shuo;WANG Guo-Tao(School of Electronic Engineering, Heilongjiang University, Harbin 150080,China;Military Apparatus Research Institute of Harbin Institute of Technology, Harbin 150001,China)

机构地区：[1]黑龙江大学电子工程学院,哈尔滨150080 [2]哈尔滨工业大学军用电器研究所,哈尔滨150001

出　　处：《黑龙江大学工程学报》2020年第3期71-77,共7页Journal of Engineering of Heilongjiang University

基　　金：国家自然科学基金项目(51607059,51077022);黑龙江省自然科学基金项目(QC2017059);黑龙江省博士后基金项目(LBH-Z16169);黑龙江省高校基本科研业务费项目(HDRCCX-201604);黑龙江省教育厅科技成果培育项目(TSTAU-C2018016)。

摘　　要：密封继电器在航空航天领域有着非常关键的作用,继电器内部的多余物严重影响着它的可靠性和稳定性。现在的密封继电器多余物检测技术大多采用微粒碰撞噪声检测法(PIND),这种传统的检测技术保障了航天系统的可靠性和稳定性,但对于密封继电器多余物信号判断的准确率只能达到75%。实验使用基于参数调优的Xgboost算法对检测信号进行分类,首先,通过对比多余物信号和组件信号的时域和频域波形提取出13个特征构建训练样本,然后通过网格搜索和k折交叉检验结合的方法,搜索出Xgboost树的最大深度和每棵树随机采样的最好比例,进行模型训练和模型调整。结果表明,基于参数调优的Xgboost算法在密封继电器多余物的判断上,准确率最高可以提升到90%,精确率和召回率等各项分类指标也有大幅度提升。The sealed relay has a very important role in the aerospace field,and the excess inside the relay seriously affects its reliability and stability.Most of the current sealed relay redundant object detection technology uses the particle collision noise detection method(PIND).This traditional detection technology guarantees the reliability and stability of the aerospace system,but the accuracy of the sealed relay redundant signal judgment can only reach 75%.Xgboost algorithm based on parameter tuning is used to classify the detection signals.First,13 features are extracted by comparing the time domain and frequency domain waveforms of redundant signals and component signals to build training samples,and then grid search and k-fold cross-check are performed.The combined method searches for the maximum depth of the Xgboost tree and the best ratio of random sampling of each tree for model training and model adjustment.The final results show that the Xgboost algorithm based on parameter tuning can improve the accuracy of the sealed relays to 90%,and the classification indicators such as accuracy and recall have also been greatly improved.

关 键 词：多余物 微粒碰撞鼓噪检测法(PIND) Xgboost 网格搜索 k折交叉检验 

分 类 号：TM58[电气工程—电器]