A comparative analysis of basic pattern recognition techniques for the development of small size electronic nose

In this paper, we propose the investigation and the realization of an artificial sensor system and an analysis system able to extract analytical information from odors, under the constraint of being suitable for miniaturization and portability. A sensor array was realized with a series of carbon black–polymer detectors. The lack of reproducibility of those sensors was compensated using a very flexible calibration and recognition tool based on neural networks. The training strategy used in this work, that performs better than derivative-based optimization techniques like standard back-propagation, permits a very low cost VLSI realization, necessary condition for deep miniaturization of the system. A comparative analysis of different pattern recognition approaches was performed in order to evaluate the suitability of this kind of neural networks, which allow deep computing circuit miniaturization. Moreover, we used dimensionality reduction techniques to decrease the computational complexity of the classification technique. The analyses carried out in this study could allow the development of a compact and self-contained electronic nose, in which the analysis system is directly embedded in the sensor device. This should permit to minimize the costs and to obtain better portability and performances.