基于特征加权K-means聚类的多模型软测量建模  

Multi-model soft-sensor modeling based on feature-weighted K-means clustering

作　　者：董陶[1] 杨慧中[1] 

机构地区：[1]江南大学物联网工程学院,教育部轻工过程先进控制重点实验室,江苏无锡214122

出　　处：《计算机与应用化学》2013年第4期361-364,共4页Computers and Applied Chemistry

基　　金：国家自然科学基金资助项目(No.61273070);江苏高校优势学科建设工程资助项目;高等学校学科创新引智计划资助(B12018);江南大学博士研究生科学研究基金(JUDCF12030)

摘　　要：针对传统K-means聚类算法的聚类结果易随不同的初始聚类中心波动的问题,采用最大距离积法优化K-means聚类算法的初始聚类中心。传统的K-means聚类算法都假定样本的各维特征对聚类的贡献相同,影响了聚类效果和模型估计精度。为了考虑样本各维特征对聚类的不同影响,利用一种新型的特征加权K-means聚类算法逐步调整特征权值,最终有效改善了聚类效果。利用本文方法建立组合支持向量机模型,将其用于双酚A生产过程质量指标的软测量建模中,仿真结果表明该算法能够有效改进数据的分类效果并提高软测量模型的估计精度。Clustering results of traditional K-means clustering algorithm easily fluctuates with random initializing cluster centers. In order to solve this problem, maximum distances product algorithm is adopted. Secondly, clustering assumes that each feature of the samples is the same to the contribution of clustering. Therefore, a new feature-weighted K-means clustering algorithm is proposed which adjusts feature weights gradually on the basis of traditional K-means flustering, and improves the clustering result finally. A combination model based on support vector machine（SVM） is established and is applied to a soft sensor modeling for the quality index in a Bisphenol A production process. The simulation result shows that the data classification result is effectively improved by the algorithm, and the estimation precision is improved.

关 键 词：K-MEANS聚类算法 初始聚类中心 特征权值 组合支持向量机 

分 类 号：TP274[自动化与计算机技术—检测技术与自动化装置]