Study on the odor classification in dynamical concentration robust against humidity and temperature changes

In this paper, we propose a method for enhancing the robustness of odor classification against the changes of humidity and temperature when the odor concentration is changing dynamically. We used amplitudes of frequency components of sensor responses at particular frequencies, instead of response magnitudes, to compose a pattern vector for the odor classification. The frequency analysis was done by using a short-time Fourier transform (STFT) and the selection of the frequency components by using a stepwise discriminant analysis. Besides the use of the STFT, we also improved the classification performance by including the humidity and temperature values to the pattern vector. Using a learning vector quantization (LVQ) neural network and training the network with wide-range data, we successfully achieved high robustness against various environment conditions even if the odor concentration was changing dynamically and irregularly under various humidity and temperature.