Discrimination Of Cigarettes Brands Based On Their Routine Testing Chemical Indicators And Near Infrared Spectra

In this paper near infrared (NIR) spectra and routine testing chemical indicators of different brands cigarettes were selected as original information to discriminate brand of cigarettes. Reversed sequence method was applied to test stability of NIR data. The method of extracting classification features of discriminating brands of cigarettes was advanced based on information gain. Nine routine testing chemical indicators, six common chemical indicators selected by experience, principal component scores (PCS), five extracted PCS and twelve extracted synthetical indictors of cigarettes were taken as classification features, respectively. Improved and simplified KNN (IS-KNN) and simplified KNN (SKNN) pattern recognition methods were applied to build a series of classification models for discriminating brands of cigarette.Two hundred and seventy-two cigarettes belonging to six brands were collected as samples to validate the pattern recognition models established based on the above five features. For every pattern recognition model, four validation sample sets were selected randomly and the average prediction accuracy of the four validation sets was adopted to evaluate prediction ability of the model. Finally the important features affecting style of brand cigarettes were determined according to discrimination results of the models. The results would provide theoretical basis and introduction for maintaining brand and formula of cigarettes, and provide help for design of cigarette formula.The results indicated that the models built on the first sixteen PCS and 12 extracted synthetical features gave very high average prediction accuracy of 92.65%～98.86% whether for two-class classification of two brands of cigarettes, three-class classification of three brands of cigarettes or six-class classification of six brands of cigarettes; When nine chemical indicators and six common chemical indicators selected by experience were used as classification features, the discrimination models gave average prediction accuracy higher than 95% for two-class classification of two brands of cigarettes and three-class classification of three brands of cigarettes, however the prediction accuracy for six-class classification of six brands of cigarettes was 83.09%～84.56%. Therefore, the prediction ability of the discrimination models based on NIR spectra is better than that of the models based on chemical indicators. The prediction accuracy decreases with the increasing of brand number and complexity of cigarette samples.Summing up the above, NIR spectra can well reflect main style of different brands of cigarettes. Brands of cigarettes can be accurately identified by NIR spectra. It is suggested that we can supervise cigarette brand, design and maintain formula of cigarettes by means of NIR information.