Data Fusion Research Of Odor And Taste Of Pungent Chinese Herbal Medicines Based On Multi-Sensors

The quality control of Chinese Herbal Medicines(CHMs) were mainly based on character and microscopic identification to determine the authenticities, the physical and chemical identification to review the merits and quality, rarely be detected and controlled in accordance with the index components of smell and taste. The traditional control of index components were difficult to control the efficacy of traditional CHMs in reality, the detection of any kind of active ingredient cannot reflect the overall effect which they embody. In China, the differential treatment using the taste and smell of the CHMs, not the separate chemical composition. Therefore, the study of the variety and quality of CHMs should not only for certain kinds of chemical composition, but also control the whole substance group in CHMs which caused the reaction of taste and smell.In this work, Electronic nose based on gas sensors and electronic tongue based on taste sensors were chose to detected the overall odor and taste data information of the Pungent Chinese Herbal Medicines (PCHMs). The odor data were stored in files with ".nos " as a suffix, per data sample is a120X10matrix; the taste data were stored in files with ".tog" as a suffix, each data sample is a120X7matrix.While doing the data fusion of odor and taste got by the multi-sensors, the data fusion on the Data Level Fusion was adopted, which could maintain the integrity of the overall original data and get more accuracy and reliability.In consideration of the nonlinear characteristic of odor and taste data information, while doing the data fusion, the manifold algorithm-Locally Linearly Embedding (LLE) and Fisher Linear Discriminant Analysis (LDA) were employed to finish data dimension reduction, feature extraction and PCHMs samples’classification. This data fusion method was effective to reduce signal processing amount to improving the computing speed; maintain the integrity of the original sample space geometry structure to improving the accuracy of the sample classification.According to the principle of operation LLE methods, the data format should be changed. In the MATLAB environment, the odor data set was transposed into a column matrix, and then turned the corresponding taste data matrix set into a column matrix following the odor column matrix, thus one sample data fusion was finished on the data fusion level.While doing the data dimensionality reduction and feature extraction, select a appropriate value of K neighboring points according to the number of data samples of each PCHMs to get the best local reconstruction weight matrix; then according to this matrix to get the best regularization parameter r; based on the got value of K and r, set a different d dimension embedding manifolds to obtain the most suitable d, thus realized the dimensionality reduction and characteristics extraction of the odor and taste data fusion. At this moment, the low dimensional sub-manifold of all the samples was obtained (namely the embedded sub-manifold). Finally, use the LDA to finish the classification of the embedded sub-manifold.In this paper, there were two sets of PCHMs samples, one is three Guangdong Rhizaoma Atractylodis Albas in different production time, the other is six different kind of PCHMs. Studies of odor and taste data fusion were taken at different total number of data samples, based on the LLE and LLE+LDA method. The emphasis was focused on the use of data processing method and laws to determine parameters. The results showed that with the use of LLE+LDA data analysis methods, meanwhile using controlled variable method to acquire the optimal parameters, can accurately finished the odor and taste data fusion of PCHMs based on multi-sensors.