Rank extraction in tin-oxide sensor arrays

It is shown that data pre-processing by rank-order filtering can significan tly improve the odor discrimination capability of an array of chemical sens ors, while simultaneously reducing the amount of data to be processed. This work is a first example in feature extraction from tin-oxide sensors that both reduces the size of the data set and simultaneously improves the discr imination performance of the array. This work is aimed toward the design of remote sensor modules where bandwidth reduction and improved accuracy are both essential to system performance. The effectiveness of extracting rank from a 30-element array of tin-oxide sensors is presented. Results are extr apolated to other arrays of chemical sensors whose specificities and respon se characteristics overlap. Methods for processing data and extracting rank -related features from arrays of tin-oxide sensors are comparatively analyz ed. Processing parameters studied include those related to temporal filteri ng and window-averaging, pre-scaling (to remove baseline), sample acquisiti on time, and the number of ranks used in rank-order filtering of the data d uring the transient and steady state response. Cluster analysis, including principal component analysis (PCA) and a novel method described herein, is used to determine which of these processing techniques are most effective. Artificial neural networks, specifically multi-layer perceptrons and radial basis function networks, are used to further investigate the ability to di scriminate odors on the basis of the extracted features. The analysis is performed for an array of 30 tin-oxide sensors applied to d etecting a sampling of breath alcohol mixtures (beer, wine, vodka) and comm on interferents (acetone, formaldehyde, isopropyl). Ammonia is included as a contrast substance. For the set of seven odorants studied, it is found th at using rank-order filtering with 10 or more ranks improves odor recogniti on rate by a multi-layer perceptron neural network from 92% to 95%. If one odor (vodka) is removed from the study set, the recognition rate for the re maining odors improves from 95% (with no rank-order filtering) to 99%. Simu ltaneously, the dimensions of the data set for each odor are reduced from 3 0 sensors x 18,000 time steps (12 bit samples) to N integer values, where N is the number of ranks used in the rank-order filtering. (C) 2000 Elsevier Science S.A. All rights reserved.