Wavelet transform of electronic tongue data

Abstract A measurement in a multi-sensor system is characterized by a large array of numbers (a vector or a matrix), sometimes several thousands. In order to increase the interpretability of the measurements, decrease the calculation demand on the computer, and/or to reduce noise, an alternative, more compact, representation of the measurement can be made which describes the important features of the measurement well but with a much smaller vector. The purpose of this paper is to show that for a particular wet-chemical sensor system (pulsed voltammetry, also called an electronic tongue) the data compression can be made using a wavelet transform together with different wavelet selection algorithms for different purposes. The resulting compressed data can also be used for easy interpretation of the measurements and to give hints for improvements or simplifications of the measurement procedure. Two different criteria for selection of wavelet coefficients have been used, variance and discriminance, in two different cases. The variance criterion was used when variations of any kind in the raw data was studied during monitoring of water in drinking water production plant. In this case, the number of variables was reduced with a factor of 18, without loosing relevant information. In the other case, the focus was to separate different microorganisms, therefore, the discriminance selection criterion was successfully used. The number of variables was reduced by a factor of 144, this smaller data set captured the important information for separating the microorganisms, which led to better classification of the test set.