Intelligent method for sensor subset selection for machine olfaction

Abstract

A fundamental design concept for an array of sensors used in machine olfaction devices, electronic noses (e-noses), is that each sensor should maximize the overall sensitivity and provides different selectivity profiles over the range of target odor application. Ideally, each sensor should produce a different response to a given odor so that a unique odor pattern is created. Since this is rarely the case, sensor selection or reduction is needed when classification performance, cost, and technology limitations are issues of concern. The first step in the selection/reduction process is to generate features from each sensor's output waveform. In practice, some of the features obtained from an array of sensors are redundant and irrelevant due to cross-sensitivity and odor characteristics. As a result, inappropriate features or a poor configuration of features can lead to a deterioration of classification performance, or a more complex classification algorithm may be required. Hence, sensor selection for e-nose systems is of great important. In this study, a novel computationally efficient method is introduced by selecting the first few critical sensors based on a maximum margin criterion among different odor classes. Then, a stochastic search algorithm, a genetic algorithm (GA), uses those features as an initial step to optimize our sensor selection problem. The advantages of the proposed method are not only to avoid any initial misstep starting the search, but also to reduce the searching space for the optimal sensor array. From the experimental results on coffee and soda data sets, the number of selected sensors is significantly reduced (up to 90%) and classification performance is near 100%.