Recognition of the Bouquet of Chinese Spirits by Artificial Neural Network Analysis

Abstract

Forty-six samples of Chinese spirits, whose bouquets were determined by sensory evaluations, and 17 compounds characteristic of the flavors determined by gas chromatography/gas chromatography-mass spectrometry (GC/GC-MS), were subjected to neural network analysis and their corresponding factor scores developed. To make the bouquet recognition more efficient, an improved artificial back-propagation neural network (BPNN) was applied. In each kind of data, the BPNN was trained repeatedly until the error rate was less than the predetermined threshold error; then the trained network was applied to the test set that was not involved in the training process to establish the validity of the network, and a correct prediction rate of 100% was obtained. The BPNN provided a correlation between the data offered from sensory evaluations and the data of chemical compositions determined by instrumental analysis. The BPNN approach is feasible regardless of whether the crude data or the factor scores are used; however, recognition results were better with the latter than with the former. In a comparison of all the results obtained by BPNN, cluster analysis, and discriminant analysis, the method of artificial neural network analysis appeared to be the optimal technique for recognizing the bouquet of Chinese spirits.