| || Feature Extraction by Wavelet Decomposition of Surface Acoustic Wave Sensor Array Transients
Author : Singh, Prashant;Yadava, R.D.S.
Source : Defence Science Journal ; Vol:60(4) ; 2010 ; pp 377-386
Subject : 53 Applied Physics;681.586 Sensors;Defence Science Journal
Keywords : SAW sensor transients;wavelet decomposition;feature extraction;VOC identification;machine olfaction;surface acoustic wave;discrete wavelet transformer
Abstract : The paper presents a new approach to surface acoustic wave (SAW) chemical sensor array design and data processing for recognition of volatile organic compounds (VOCs) based on transient responses. The array is constructed of variable thickness single polymer-coated SAW oscillator sensors. The thickness of polymer coatings are selected such that during the sensing period, different sensors are loaded with varied levels of diffusive inflow of vapour species due to different stages of termination of equilibration process. Using a single polymer for coating the individual sensors with different thickness introduces vapour-specific kinetics variability in transient responses. The transient shapes are analysed by wavelet decomposition based on Daubechies mother wavelets. The set of discrete wavelet transform (DWT) approximation coefficients across the array transients is taken to represent the vapour sample in two alternate ways. In one, the sets generated by all the transients are combined into a single set to give a single representation to the vapour. In the other, the set of approximation coefficients at each data point generated by all transients is taken to represent the vapour. The latter results in as many alternate representations as there are approximation coefficients. The alternate representations of a vapour sample are treated as different instances or realisations for further processing. The wavelet analysis is then followed by the principal component analysis (PCA) to create new feature space. A comparative analysis of the feature spaces created by both the methods leads to the conclusion that both methods yield complimentary information: the one reveals intrinsic data variables, and the other enhances class separability. The present approach is validated by generating synthetic transient response data based on a prototype polyisobutylene (PIB) coated 3-element SAW sensor array exposed to 7 VOC vapours: chloroform, chlorobenzene o-dichlorobenzene, n-heptane, toluene, n-hexane and n-octane.