Prediction of odor pleasantness using electronic nose technology and artificial neural networks

An electronic nose was used in conjunction with a human panel and artificial neural networks to predict human assessments on a hedonic (pleasantness) scale when confronted with spent and phase I mushroom substrate odors. Odor samples were collected and presented to both the electronic nose and the human panelists. The human panelists were required to rate the odor samples on a hedonic scale. After calculating the average rating from the human panel for each odor sample, the hedonic rating was matched with the 32 electronic nose sensor readings from the corresponding sample. Once the hedonic ratings had been matched to the corresponding sensor readings, the data were then used to develop and train an artificial neural network.; Two artificial neural networks were developed from these training data. The first artificial neural network was developed using Ward Systems Neuroshell 2. This neural network used 10 of the electronic nose sensor readings to predict the hedonic rating of the sample odor. This artificial neural network produced predictions with an r2 of 0.95.; The second artificial neural network was developed using Ward Systems Neuroshell Predictor. This neural network used all 32 of the electronic sensor readings to predict the hedonic rating of sample odors and was examined in two modes. The first mode of operation was termed Non-Enhanced Generalization; in this mode the artificial neural network attempted to fit the data tightly. In the Enhanced Generalization mode, the neural network predictions were not as tightly fit; this method is generally used for new inputs for which the neural network had not been trained. In Non-Enhanced Generalization mode, an r2 of 0.95 was determined. In the Enhanced Generalization mode, the predictions were a bit more relaxed; an r2 of 0.89 was determined.; The results obtained from the electronic nose data and the developed neural networks prove promising in the area of hedonic rating. It should be emphasized that this type of modeling or simulation is very specific and that any developed neural network is only as good as the data on which it is was trained. There are many considerations that need to be addressed when taking on a matter of this magnitude. First, a small change in chemical composition can radically affect the odor of a compound. Secondly, the intensity of an odor can affect the perception of an odor. For example, some odorants at small quantities are perceived as pleasant, but at higher concentrations they can become offensive. The first two considerations would require the collection of an infinite amount of data to represent all of the possible combinations. Third, in a subjective matter such as this, life experiences, cultural differences, etc., contribute to each panelist's perception of an odor. Carefully choosing and training the human panel can help minimize the third consideration. Fourth, it is well known that the human nose is much more sensitive than any electronic nose to subtle changes in an odorant. Until advances are made in the electronic nose area, this technology may not be sensitive enough to simulate human perception except for very specific and limited applications. This research may be small step toward much larger research and applications.