Evaluation method of feature vector in vinegar identification by electronic nose

Abstract: In a test of the electronic nose (E-nose), two evaluation methods of the feature vector identification ability are commonly used: 1) visual analysis for the discrimination result, 2) statistical computation for the correct rate of the discrimination result. However, when the correct rates of the discrimination result are same, for the different feature vectors, the identification ability can not be evaluated accurately and quantitatively by the above two methods. In order to achieve the precise evaluation of the feature vector identification ability, different kinds of vinegar were taken as the study object and tested by the E-nose. Six kinds of feature vectors including Variance (Var), Integral value (Inv), Average value in relative steady-state (Avrs), Value of area divided by the slope (Vads), Average differential value (Adv), and Wavelet energy value (Wev) were extracted from the acquisition data after removing the background signal. The correlation coefficient between feature vectors and categories was used as an evaluation index of the feature vector's identification ability, the evaluation and comparison for the feature vectors is achieved by this index. Absolute values of the correlation coefficient between feature vectors and categories are respectively 0.2936, 0.6455, 0.6182, 0.1027, 0.6176 and 0.6189. Among them the absolute value of correlation coefficient was least between the 'Vads' feature vector and the categories, and the absolute value of correlation coefficient was greatest between the 'Inv' feature vector and the categories. These results show that the identification ability of the 'Vads' feature vector is the lowest, and the identification ability of the 'Inv' feature vector is the highest.The correct rate of the Fisher discrimination result was calculated, and the classification effect graph of the Fisher Discrimination was analyzed for every feature vector. The correct rate of the 'Vads' feature vector was the lowest (39.2%). Its corresponding classification effect was the worst; all kinds of vinegar samples mixed, and the group centroids of three categories almost overlapped. The correct rates of the discrimination result were all 100% for the other feature vectors, so the comparison of feature vector identification ability can not be carried out only by the correct rates of discrimination. But the classification effect graphs of Fisher Discrimination for these feature vectors show that the clustering degree within and between the groups for the three categories of vinegar samples were very different, which indicates that there may still be differences among the feature vectors' identification ability, even though their discrimination correct rates were all 100%. Specifically, the classification effect of feature vector 'Inv' was the best, the clustering degree of samples within the groups was the highest, and the boundaries between groups were the most distinct.The discriminating results of Fisher Discrimination prove that the higher the feature vector identification ability is, the better the classification result is. This result is in accord with that of the absolute values of the correlation coefficient. Therefore, it is right and effective that the correlation coefficient between feature vectors and categories be used as an evaluation index of the feature vector identification ability. The proposed method will provide a new train of thought for studies of quantitative evaluation in the E-nose system.