Continuous chemical classification in uncontrolled environments with sliding windows

Electronic noses are sensing devices that are able to classify chemical volatiles according to the readings of an array of non-selective gas sensors and some pattern recognition algorithm. Given their high versatility to host multiple sensors while still being compact and lightweight, e-noses have demonstrated to be a promising technology to real-world chemical recognition, which is our main concern in this work. Under these scenarios, classification is usually carried out on sub-sequences of the main e-nose data stream after a segmentation phase which objective is to exploit the temporal correlation of the e-nose's data. In this work we analyze to which extent considering segments of delayed samples by means of fixed-length sliding windows improves the classification accuracy. Extensive experimentation over a variety of experimental scenarios and gas sensor types, together with the analysis of the classification accuracy of three state-of-the-art classifiers, support our conclusions and findings. In particular, it has been found that fixed-length sliding windows attain better results than instantaneous sensor values for several classifier models, with a high statistical significance.