Simultaneous estimation of classes and concentrations of odors by an electronic nose using combinative and modular multilayer perceptrons

It is relatively suitable to take a simultaneous estimation problem of multiple odor classes and concentrations for multiple function approximation ones. To solve such problems, this paper decomposes a many-to-many approximation task into multiple single-output approximation ones, and correspondingly presents a kind of combinative and modular multilayer perceptrons (MLPs). Every module is made up of multiple one-to-one MLPs and one many-to-one MLP, and one MLP is regarded as an expert. One module consists of several such experts and implements a many-for-one approximation task. By means of enlarging the input components to the range of [0,6.0] and making the sigmoid activation functions s(x) = 3(1 + exp(−x/3))−1, the learning processes of MLPs are sped up. In an electronic nose, one MLP module is on behalf of a kind of odor, and determines how similar an undecided odor is to a known odor, namely its strength. The most similar module gives the class label and strength value of the odor. The experiment for simultaneously estimating the classes and concentrations of four kinds of fragrant materials, namely ethanol, ethyl acetate, ethyl caproate and ethyl lactate, 21 different concentrations in all, shows that the proposed combinative and modular approximation method is quite effective.