A novel data-driven technique to produce multi- sensor virtual responses for gas sensor array-based electronic noses

Abstract

Accurate detection of gas/odor requires highly selective gas sensor. However, the high-performance classification of gases/odors can be achieved using partial-selective gas sensors. Since 1980s, an array of broadly tuned (partial-selective) gas sensors have been used in several fields of science and engineering, and the resulting gas sensing systems (GSS) are popularly known as electronic noses (e-Noses). The combination of similar or different sensors in the array indirectly compensates for the requirement of high selectivity in GSS. Further, e-Nose's performance inevitably depends on the salient features drawn from the initial responses of the gas sensor array (GSA). So obtained features are referred to as the responses of virtual sensors (VS). In this paper, we have proposed the three-input and three-output (TITO) technique to derive efficient virtual sensor responses (VSRs) which outperform its well-published peer technique. A GSA consisting of four elements is used to demonstrate the proposed technique. Our proposed technique augments the VSRs by four times compared to its peer. The efficacy of our proposed technique has been tested using nine fundamental classifiers, viz., linear support vector machine (100%), decision tree (97.5%), multi-layer perceptron neural network (100%), K-nearest neighbor (85%), logistic regression (100%), Gaussian process with radial basis function (95%), linear discriminant analysis (97.5%), random forest (100%), and AdaBoost (95%). Ten-fold cross-validation has been used to minimize the biasing impact of the intra- and inter-class variance. With the result, four classifiers successfully provide an accuracy of 100 percent. Hence, we have proposed and vindicated an efficient technique.