【摘要】 应用支持向量分类方法,将云南省9个地区缴获的1148个海洛因样品,用近红外漫反射光谱在4 000 cm^-1～10 000 cm^-1范围内吸收系数数据集合,构建判别毒品来路的分类器。光谱数据选取了指纹波数区段5 990 cm^-1～7 500 cm^-1,以及最大和较大吸收系数的41个波数的光谱数据。针对一对一算法的五分类问题,采用两种分类法C-SVC和v-SVC, 4种核函数,分别以默认参数和优化参数,得训练集模型有效率和检验集的预报总精度。比较各种模型后,确定了152个指纹区波数,线性核函数的L-152 C-SVC作为分类器模型。该模型对已知分类的5个地区随机选取的训练集样本,在10-交叉检验下的有效率是90.74%,对不包含训练集的其余全部已知样品,其预报总精度是88.71%。5地区分类统计计算的敏感性、特异性、相关系数的评价都较好。最后,又试用该分类器于未知地毒品的来路辨认。与报道的模式识别比较,工作没有止于训练集给出模型,检验集判断预报效果的已知样品,又走出了重要一步,即识别训练集和检验集之外的未知样品。更多还原

【Abstract】 Applying a support vector classification method,the classifiers of differentiating sources were established to a data set made using the near infrared diffuse reflectance spectra in the wave number range of 4 000 cm^-1～10 000 cm^-1 of the 1148 heroin drugs from the nine regions of Yunnan.The selected spectra involved the characteristic wave numbers,5 990 cm^-1～7500 cm^-1,and the other 41 wave numbers that possess the most and the more absorbance.For a five classification with one-against-one strategy,first,two machines C-SVC and v-SVC,as well as four kernels were applied to achieve the effectiveness of training set and the accuracy of test set in the different parameters,default and optimum.After comparing the various classification models,the L-152 C-SVC,which had 152 characteristic wave numbers of input vector and line kernel was determined.This classifier shows the following results:（1） 90.74%of the effectiveness under 10-fold cross validation for the training set selected at random from the known samples;（2） 88.71%of the overall accuracy of prediction for test set composed of the rest known samples except the training set;（3） the better sensitivity,distinctness, correlation coefficient for the five region classes.Finally,this classifier was tried to differentiate those source-unknown heroin drugs.Compared with the published pattern recognitions,the present study not only have brought to success in the modeling of training and test sets for known samples,but continued an important differentiation for the unknown samples.更多还原