Journal of Pattern Recognition Research

Steady state response of chemiresistive sensors like metal oxide and conducting polymer composite (CPC) exhibit a power law dependency on exposed analyte concentration/vapor pressure. In present research, we suggest linearization of the measured (raw) sensor signal, for both metal-oxide and CPC sensor, by an inverse power scaling, before preprocessing by mean-centering and variance-normalization, and analyze its influence on feature extraction by principal component analysis and chemical vapor classification by backpropagation neural network method (BPNN). For analysis we have employed two tin-oxide (SnO2) sensor array response based data sets and one CPC sensor array response based data set available in published literature. We find that preprocessing the data sets by suggested power scaling method improves the visual discrimination of chemical analyte in Principal component (PC) space and neural network identification rate by 14% and 16% with SnO2 sensor array and by 30% with CPC sensor array. Empirical values of scaling coefficient, for proposed preprocessing method have been found in accordance with the existing theoretical models and working physics of SnO2 and CPC sensors.