OPTIMAL CONFIGURATION OF A THERMALLY CYCLED GAS SENSOR ARRAY WITH NEURAL-NETWORK PATTERN-RECOGNITION

Through the application of novel thermal modulation, pre-processing an d feature extraction techniques the performance of all 8-element tin o xide gas sensor array has been significantly enhanced when applied to the task of classifying the aromas of three loose leaf teas. Array sig natures were generated by thermally cycling the sensor array over the temperature range 250-500 degrees C, whilst exposing the array to the odorous headspace of the three teas. The application of thermal modula tion and an enhanced feature extraction algorithm, generating 208 para meters, proved to be highly successful giving a cross-validated classi fication rate of 90% for unseen samples of the three classes of tea. I n comparison, a fixed temperature steady state metric, based upon the same array of sensors, yielded a cross-validated classification rate o f only 69%. Furthermore, using a novel genetic algorithm optimisation technique to identify a near-optimal sensor parameter configuration fo r the task of tea classification, it was shown that a correct classifi cation rate of 93% could be achieved with only 21 dynamic parameters. (C) 1998 Elsevier Science S.A. All rights reserved.