基于PSO聚类和特征贡献度的油液监测信息特征选择方法  

Feature Selection of Oil Monitoring Information Based on PSO Clustering and Feature Contribution

作　　者：石新发[1] 刘东风[1] 周志才[1] 杨琨[2] 

机构地区：[1]海军工程大学青岛油液检测分析中心,山东青岛266012 [2]武汉理工大学能源与动力工程学院,湖北武汉430063

出　　处：《润滑与密封》2016年第1期86-89,114,共5页Lubrication Engineering

基　　金：国家自然科学基金项目(51309185)

摘　　要：特征选择是实现油液监测多技术手段综合应用的关键问题之一。针对油液监测信息特点,提出一种油液监测信息特征选择方法。该方法首先采用K均值PSO聚类算法对样本实施无监督聚类,实现样本的预先分类;然后采用定义的特征贡献度,计算各特征对聚类结果的贡献度,并以此作为特征选择的依据,实现无监督的过滤式特征选择。通过在某型柴油机润滑油原子发射光谱和红外光谱信息中的应用表明,该算法能够很好的实现油液监测信息的特征选择,减少特征指标数量,而且能够避免由于油液监测信息依存度和相关度高的特点而造成特征选择时可能会将重要信息删除的问题。Feature selection is one of key-issues in synthesis application of the oil monitoring technologies.According to the characteristics of oil monitoring information, K mean particle swarm clustering was used to realize samples unsupervised clustering.The defined feature contribution degree was calculated and used as the basis of feature selection,and the unsupervised feature selection based filter for the oil monitoring information was realized.The algorithm was applied in the information of atomic emission spectrum and FT-IR spectrum gained from one marine diesel＇ s lubricant. The result indicates that the proposed algorithm can well realizes the feature selection of oil monitoring information, and reduce the number of the parameters, and avoid the problem that the important information is possibly pruned because of the existence of the characteristics of high relevance objective in oil monitoring information.

关 键 词：机械装备 油液监测 特征选择 粒子群聚类 特征贡献度 

分 类 号：TH117.1[机械工程—机械设计及理论]