中红外光谱和支持向量机的梅花鹿角帽粉假冒与掺假识别模型

A Model for the Identification of Counterfeited and Adulterated Sika Deer Antler Cap Powder Based on Mid-Infrared Spectroscopy and Support Vector Machines

梅花鹿角帽具有较高的药用和经济价值,因其质地坚硬,故一般选择打成粉末使用。消费者很难从外观上去判别梅花鹿角帽粉是否为正品,导致其假冒与掺假事件层出不穷。因此,提出利用中红外光谱(FTIR)结合机器学习探索一种识别梅花鹿角帽粉假冒与掺假的方法,以解决用马鹿角帽粉、梅花鹿骨粉假冒梅花鹿角帽粉和牛骨粉掺假梅花鹿角帽粉的问题。从黑龙江、吉林、辽宁3省共5个地区采集梅花鹿角帽、马鹿角帽、梅花鹿骨各120份,共360份样品;牛骨采购于长春市南关区农贸市场,使用牛骨粉掺假梅花鹿角帽粉,掺假比例分别为5%,10%,20%,30%,40%,50%,每种比例各20份,共120份。采集样品中红外光谱数据,多元散射校正(MSC)对光谱进行预处理,经K-S法抽样,按3∶1的比例划分训练集和测试集后,对光谱数据进行归一化(normalization)和主成分(PCA)分析降维处理。根据主成分个数累积贡献率≥85%,主成分特征值≥1原则,选择前7个主成分构成降维后的光谱数据;分别将全光谱(FS)数据与PCA降维后的光谱数据作为模型输入,建立支持向量机(SVM)、随机森林(RF)、极限学习机(ELM)识别模型。结果表明,梅花鹿角帽粉正品与假冒伪品、掺假次品的波谱在波段1300~1800和2800~3600 cm处存在差异,尤其是掺假比例≥10%以上的梅花鹿角帽粉与纯梅花鹿角帽粉差异明显。在识别梅花鹿角帽粉假冒与掺假模型中,FS-SVM,PCA-SVM,FS-RF模型均有很好的识别效果,其训练集与测试集识别率均为100%,其他模型识别率均低于98%。从简化模型的角度上比较,FS-SVM,FS-RF建模时间分别为4859.36和1818.96 s,而PCA-SVM建模时间仅为19.91 s。因此,PCA-SVM在6种识别模型中整体效果最佳。研究表明,中红外光谱结合支持向量机建模可以作为一种快速、准确、无损鉴别梅花鹿角帽粉假冒与掺假的有效识别方法。

Sika deer antler caps are of great medicinal and economic value.Because of its hard texture,the finished product is generally presented as powder.It is difficult for consumers to determine the authenticity of sika deer antler cap powder from its appearance,which leads to endless series of counterfeit and adulterated products.Therefore,this paper proposes a FTIR technology and machine learning method to identify counterfeited and adulterated sika deer antler cap powder.This method can identify counterfeited sika deer antler cap powder by horse stag deer antler cap powder,sika deer bone powder,and adulterated sika deer antler cap powder by beef bone powder.This research’s sika deer antler caps,stag deer antler caps and sika deer bones are from five regions of the three provinces of Heilongjiang,Jilin and Liaoning.The samples are divided into 360 portions,including 120 portions of sika deer antler caps,120 portions of red deer antlers caps and 120 portions of sika deer bones.The beef bone powder is purchased in Changchun Nanguan District Farmers’Market.Adulterate the beef bone powder into 120 portions of sika deer antlers powder with 5%,10%,20%,30%,40%,and 50%for every 20 portions.Sample spectral data were collected by mid-infrared spectroscopy,preprocessed by multiple scattering correction(MSC),and sampled by the K-S method.After the training and test sets were divided by 2∶1,Normalization and principal component analysis(PCA)dimension reduction was conducted on spectral data.According to the principle of cumulative contribution rate of the number of principal components≥85%and principal component characteristic value≥1,the first 7 principal components were selected to form the spectral data after dimensionality reduction.The recognition models of support vector machine(SVM),random forest(RF)and Extreme learning machine(ELM)were established by using full-spectrum(FS)data and PCA dimensional-reduction spectral data as model inputs.The results showed a difference between the authentic and counterfeit and adulterated products in the waveband of 1300~1800 and 2800~3600 cm.The difference between the pure sika deer antler cap powder and sika deer antler cap powder of the adulteration rate≥10%was the most obvious.FS-SVM,PCA-SVM and FS-RF models all have excellent recognition effects in identifying fake and adulterated sika deer antler hat powder.The recognition rate of the training and test set is 100%,and the recognition rate of other models is less than 98%.From the perspective of simplified models,the modeling time of FS-SVM and FS-RF is 4859.36 and 1818.96 s respectively,while the modeling time of PCA-SVM is only 19.91 s.Therefore,PCA-SVM has the best overall effect among the six recognition models.The research shows that the method based on mid-infrared spectroscopy combined with support vector machine modeling can be used as a fast,accurate and non-destructive identification method for counterfeiting and adulteration of sika deer antler cap powder.