| Smartongue system is a novel electronic tongue researched by our laboratory which consists of sensor array, signal excitation and reception circuit and intelligent algorithm. The early research was focused on the study of sensor array and the structure of novel system of signal excitation. Given the many advantages of electronic tongue, it is necessary to make the further optimization of the data although the related research was less than that of other parts for electronic tongue. Based on earlier research, this research has been systematically studied the pattern recognition of Smartongue via three pattern recognition algorithm on the data, such as Kernel PCA, Locally Linear Embedding(LLE), Sammon mapping and pictorial pattern recognition of the results with different methods, just like segmental surface fitting edge detection. The main research work and results are as follows.(1) Research on the data of Smartongue with Kernel PCA.Based on the limitation of the data processing, linear Principal Component Analysis (PCA) is not applicable to various samples of Smartongue. Therefore, a nonlinear transform technology used a kernel function——KPCA has replaced PCA. KPCA was selected as the method of pattern classification which used different kernel functions, such as gauss function,polynomial function and sigmoid function, was researched by three samples those are difficult to distinguish of Smartongue as usual. The result showed that the ability to distinguish samples of KPCA has significant difference while it can not distinguish in PCA method. KPCA-Gauss can classify all samples adequately; KPCA-Poly just can distinguish the bitter solution; KCPA-sigmoid just can not distinguish green tea. As a result, comparing the value of DI, KPCA-Gauss is more effective on the data processing and more applicable to Smartongue.(2) Research on the data of Smartongue with LLE.LLE has replaced PCA because of the characteristic of local linear of Smartongue's data as the same as the KPCA. LLE was choosed as the method of pattern classification was researched by the same three samples those are difficult to distinguish of Smartongue as usual. The result showed that the ability to distinguish samples of LLE has significant difference. LLE can classify all samples adequately while that can not distinguish in PCA method.(3) Research on the data of Smartongue with Sammon mapping.As before, Sammon mapping keeps Euclidean distance between the Smartongue's data in order to make visualization effect from high dimension into low dimension. Sammon mapping was choosed as the method of pattern classification was researched by the same three samples those are difficult to distinguish of Smartongue as usual. The result showed that the ability to distinguish samples of Sammon mapping has significant difference. Sammon mapping can classify all samples adequately while that can not distinguish in PCA method.(4) Research of the contour edge detection on Smartongue.The contour edge detection is taken as the core of image pattern recognition usually. To solve the problem of qualitative edge detection of Smartongue, this article has proposed a novel method——segmental surface fitting edge detection. The method was selected the white wine model of Smartongue to optimize and then was used to classify the multi-sample model of Smartongue. The result indicates that the optimal conditions as follow:the number of segments is8^the number of reserved edge points is22,the weighing of threshold value is2/3. In multi-sample detection, quantificational edge detection has a qualitative leap and optimize the structure of the edge.In summary, although three nonlinear methods are better than PCA for Smartongue, comparing the value of DI, this research indicates that the optimal pattern recognition of Smartongue combines Sammon mapping and segmental surface fitting contour edge detection. |
PDF Full Text Request