The Research Of Evaluation Methods Of Feature Vector Identification Ability In Food Test By Electronic Nose

Feature extraction and selection are key issues in food testing by the electronic nose(E-nose). Accurate evaluation of feature vector identification ability is the premise ofselecting the robust one. But by now, few studies have been done on this subject.Vinegar, milk and wine were taken as the study objects and tested by E-nose. Severalevaluation methods of feature vector identification ability based on distance,correlation coefficient and wilks Λ statistics were proposed. Then these methods wereused to quantitatively analyze six kinds of feature vectors including Variance(Var),Average differential value(Adv), Integral value (Inv), Wavelet energy value(Wev),Average value in relative steady-state(Avrs) and Value of area divided by theslope(Vads). The main research results are as follows:1. The proposed methods were used to quantitatively evaluate the identificationability of the aforementioned feature vectors. The results showed that the evaluationeffect based on distance measure was not good. The methods of correlation coefficientand wilks Λ statistics which could provide feature vectors with quantitative evaluationresults were better. According to the evaluation results, Inv and Wev both had goodeffects on the identification of all types of food, which was consistent with FDA.2. Made a preliminary study on “combination feature” which was also a kind ofinformation representation method. Results showed that: The application of“combination feature” improved the correct rate of discrimination.3. Sensor array was optimized in order to improve the identification result. Analysisresult was that: The realization process of wilks Λ statistics was easier. Redundantinformation was eliminate effectively, which facilitated the subsequent patternrecognition.The identification ability of feature vectors was improved after optimizing.4. FDA was used to verify the evaluation result of feature vector identificationability. The result was consistent with the evaluation result of correlation coefficientand wilks Λ statistics.