The Application And Research Of Kernel Transformation Used In Food Test By Electronic Nose

Pattern recognition is one of the key issues in food test using the electronic nose (E-nose). Different pattern recognition algorithms have different classification results. Some pattern recognition methods used in many fields such as principal component analysis, fisher liner discriminant analysis belong to liner discriminant analysis and can not solve nonlinear problems. But kernel method can provide a good solution to the problems. This article introduced kernel method to E-nose analysis; at the same time vinegar and white spirit were taken as the study objects. On the basis of six kinds of feature vectors including variance (Var), integral value (Inv), average value in relative steady-state (Avrs), average differential value (Adv), wavelet energy value (Wev), and value of area divided by the slope (Vads) were extracted from E-nose signals, the kernel discriminant method was used to identified the two kinds of samples, respectively. The main research results are as follows.1. A measure method of matrix similarity based on distance index was presented to define the kernel parameter of radial basis kernel function (RBF). For vinegar samples, the kernel parameters corresponding to Var, Inv, Avrs, Adv, Wev, and Vads were4.4797,5.7770,5.3878,5.6927,4.357and0.0015625, respectively. For white spirit samples, they were5.8922,3.61,2.4697,3.1027,0.58125and0.0022095, respectively.2. When some principal components were selected by their contribution rate, the discrimination results of kernel fisher discriminant analysis (KFDA) showed that:(1) when the accumulative contribution rate was0.9, the identification correct rates based on the six kinds of feature vectors corresponding to vinegar samples were95.0,98.3,98.3,96.1,97.2and88.3%, respectively; for white spirit samples, their identification correct rates were86.7,84.4,84.4,82.8,93.9and83.9%,respectively;(2) when all principal component were selected, i.e. the accumulative contribution rate was1.0, the identification correct rates of vinegar samples based on the six kinds of feature vectors were all100%; for white spirit samples, the identification correct rate based Wev was99.4%, but that of other five kinds of feature vectors were all100%.3. When some principal components were selected by wilks Λ statistics, the results of KFDA showed that even if we did not select all principal components, the identification correct rates could reach100%for vinegar and white spirit samples. This indicates clearly that wilks Λ statistics is used to select the principal component in KFDA is not only appropriate but also necessary, and its calculation workload can also be reduced.4. In conclusion, the practical application illustrates that the E-nose based on KFDA was employed to discriminate vinegar samples and white spirit samples is very effective.