Bio-inspired sensor data fusion for herbal tea flavour assessment

Abstract

Herbal-based products are becoming a widespread production trend among manufacturers for the domestic and international markets. As the production increases to meet the market demand, it is very crucial for the manufacturer to ensure that their products have met specific criteria and fulfil the intended quality determined by the quality controller. One of famous herbal-based product is herbal tea. This thesis investigates bio-inspired flavour assessments in a data fusion framework involving an E-nose and E-tongue. The objectives are to attain good classification of different types and brands of herbal tea, classification of different flavour masking effects and finally classification of different concentrations of herbal tea. Two data fusion levels were employed in this research, low level data fusion (LLDF) and intermediate level data fusion (ILDF). Four classification approaches; Fisher Linear Data Analysis (FDA), Support Vector Machine (SVM), k-Nearest Neighbour (KNN) and Probability Neural Network (PNN) were examined in search of the best classifier in order to achieve the research objectives. In order to evaluate the classifiers‘ performance, an error estimator based on k-fold cross validation and leave-one-out were applied. Classification based on GC/MS TIC data was also included as a comparison to the classification performance using fusion approaches. Generally, KNN outperformed the other classification techniques for the three flavour assessments in the low level data fusion and intermediate level data fusion. However, the classification results based on GC/MS TIC data varies in different application. Since KNN provide the highest classification performance, automatic grading system was developed based on this technique.