Identification andquantification of individual volatile organic compounds in abinary mixture by SAW multisensor array andpattern recognition analysis

Abstract

We have developed a surface acoustic wave (SAW)multisensor array with five acoustic sensing elements configuredas two-port resonator 433.92 MHz oscillators and a referenceSAW element to recognize different individual components anddetermine their concentrations in a binary mixture of volatileorganic compounds (VOCs) such as methanol and acetone, in theranges 15-130 and 50-250 ppm, respectively. The SAW sensorshave been specifically coated by various sensing thin filmssuch as arachidic acid, carbowax, behenic acid,triethanolamine or acrylated polysiloxane, operating at roomtemperature. By using the relative frequency change as theoutput signal of the SAW multisensor array with an artificialneural network (ANN), a recognition system has been realizedfor the identification and quantification of tested VOCs. Thefeatures of the SAW multisensor array exposed to abinary component organic mixture of methanol and acetone have been extracted from the output signals of five SAW sensors bypattern recognition (PARC) techniques, such as principalcomponent analysis (PCA). An organic vapour pattern classifierhas been implemented by using a multilayer neural network witha backpropagation learning algorithm. The normalized responses of a reduced set of SAW sensors or selected principalcomponents scores have been used as inputs for a feed-forwardmultilayer perceptron (MLP), resulting in a 70% correctrecognition rate with the normalized responses of the four SAW sensors and in an enhanced 80% correct recognition rate with the firsttwo principal components of the original data consisting of thenormalized responses of the four SAW sensors. The prediction of theindividual vapour concentrations has been tackled with PCA forfeatures extraction and by using the first two principalcomponents scores as inputs to a feed-forward MLP consisting ofa gating network, which decides which of three specific subnets should be used to determine the output concentration: the firstsubnet for methanol only, the second subnet for acetone only andthe third subnet for methanol and acetone in the binarymixture. Good 0.941 and 0.932 correlation coefficients for thepredicted versus real concentrations of methanol and acetone,respectively, as individual components in a binary mixture have been obtained. The experimental results demonstrated that theproposed binary organic vapour mixture classifier is effectivein the identification of the tested VOCs of methanol and acetone.Also, the combination of PCA and ANN techniques provides arapid and accurate quantification method for the individualcomponents' concentration in the tested binary mixture ofmethanol and acetone.