摘要
提出在自变量空间寻找与未知样本最接近的K个已知样本,然后根据K个最近邻样本(KNN)与未知样本在自变量空间的关系去预测未知样本因变量的最优化保形映射方法(简称KNN-KSR)。以卷烟的稀释率、闭式吸阻、单支重、圆周、开式吸阻、硬度等6个物理指标及总糖、还原糖、总氮、总植物碱、氯等5个化学指标为自变量,以企业生产的595批卷烟测试数据为基础,采用KNN-KSR方法预测卷烟主流烟气中的焦油、CO、烟气烟碱,并将有关结果与传统多元线性回归(MLR)、主成分回归(PCR)及偏最小二乘(PLS)的结果进行了比较。留1/4样本检验结果表明:KNN-KSR方法各指标预测平均残差、平均相对误差(绝对值)、相关系数和准确率均优于传统的MLR、PCR及PLS的方法。以GB5606.5-2005所规定的误差范围为标准,用KNN-KSR方法对3个卷烟主流烟气指标的同时预测准确率可以达到94%。
A novel mathematical modeling method, by which the dependent variables of an unknown sample were determined according to the relationship between the sample and K samples that are mostly closed to it in the space of independent variables, was provided in this paper. The method was named as keeping the same relationship in dependent and independent variable spaces based on K nearest neighbor samples, and KNN-KSR for short. Furthermore, using 6 physical properties of cigarettes： ventilation, closing and opening resistance, rigidity, weight and circumference, and 5 chemical qualities, content of total sugar, reducing sugar, total plant alkali, total nitrogen and total chlorine, as independent variables, tar, CO and nicotine of main stream smoke of cigarettes were predicted by the KNN-KSR method based on inspection data of 595 batch produced cigarettes of tobacco manufactures. The predicted results were compared to those given by traditional mathematical modeling methods, such as Multi-component Linear Regression （MLR）, Principal Component Regression （PCR） and PLS. It was indicated that average residual errors, average absolute value of relative errors, correlative coefficients and predicting accuracy of the three smoke indices, given by KNN-KSR, were better than those given by the three traditional methods. The ratio of number of samples, whose error between predicted value and actual value of tar, nicotine and CO were in the allowable region of GB5606.5 2005, to the number of all validation samples, could be higher than 94 %.
出处
《华东理工大学学报（自然科学版）》
EI
CAS
CSCD
北大核心
2008年第4期547-552,共6页
Journal of East China University of Science and Technology
