A neural network-based odor recognition system

Recognition and identification of odor or smell in the air have many potential applications, such as evaluation of foods, estimations for environmental conditions, and so on. Gas Chromatograph is one of the most precise analyzers for the odor recognitions, but they are not portable and tend to take much time to analyze the odors in the air. Thus, it is desirable to develop portable odor recognizer with high recognition abilities. This paper presents an odor recognition system, which primarily aims at discriminating diacetyl odor that is a cause of unpleasant smell from middle-aged men. The proposed system has four types of MOS-type gas sensors as sensory component in a sealed acyl box as a sensory components. For recognizing the sensory outputs, Learning Vector Quantization (LVQ), which is a neural network with supervised learning algorithm, is adopted. The LVQ in the proposed system accepts temporal changes of signals output from four sensors, and makes the feature map from these high dimensional sensory signals in a supervised manner. We conduct the recognition experiments by changing the duration of moving-average and the durations for sensory outputs. We use diacetyl, brewed-ethanol, isopropyl ether, and butyl acetate as target materials of odor recognition. Five-fold cross validation method is used for evaluating the recognition accuracy. From the experimental results, 96.5% of recognition accuracy on test data sets can be achieved by using 20 seconds of sensory signals and 5 seconds of moving-averaged.