2001 AACC Annual Meeting Charlotte, North Carolina October 14-18, 2001 Charlotte Convention Center

Chemical information for classifying grain odors was obtained by using dynamic headspace technology coupled with gas chromatography-mass spectroscopy to determine volatiles in a set of 745 samples consisting of corn, sorghum, soybeans, and wheat. Sensory data for each sample was obtained from at least two panels. Previous processing of the chemical and sensory data by multivariate analyses such as principal component analysis and partial least squares methods helped to determine what volatiles could be used to classify grain odors. In this study, we used artificial neural network methods to classify odors in the samples. Proper choice of samples and use of optimized variables (compounds indicating off-odors), as well as some preprocessing of raw data were necessary for training the networks. Properly trained, networks could classify samples into two odor categories (normal and off) or as many as five categories (normal, sour, musty, smoke, and insect) from analysis of chemical data concerning relative amounts of specific volatile compounds purged from each sample. Networks trained to classify into several specific categories could also identify mixed odors, i.e., having both musty and sour odors, in some samples. Panelists had given a consensus assignment of only a single odor with some samples, but neural network results and the corresponding presence of odor-indicating compounds suggested that more than one type of odor could have been assigned and may have been detected by some individual panelists.