Transportability of data between electronic noses: mathematical methods

Accumulation of databases containing volatile attributes of foods measured by electronic noses requires transportability of data between machines. Due to differences in each machine, or sensor drift, the same sample may result in slightly different readings. Three functions (coefficient, F1; coefficient with intercept, F2; matrix, F3) and two neural network configurations (NN1, NN2) were tried to assess their suitability for acceptable transformation. Two EEV electronic noses with identical sensors were used to read milk samples stored for up to 12 days. Six replicate readings were separated to ‘calibration’ (four replicates) and ‘unknown’ (two replicates) sets. Discriminant function analysis was applied to evaluate ‘unknowns’ in ‘calibration’ sets. When ‘old unknowns’ were converted and classified in ‘new calibrations’, correct classification rates (CCR) were 2/12, 3/12 and 7/12 for F1, F2 and F3, respectively. Also, neural networks were used to either transform one ‘calibration’ set to another (CCR=8/12 and 9/12), or to evaluate the ‘unknown’ sets in the ‘calibration’ sets (CCR=7/12). More research is needed in this area.