Aggregation of sensory input for robust performance in chemical sensing microsystems

This paper demonstrates the usefulness of aggregating information generated from arrays of chemical sensors for improving the ability to discriminate among target chemicals and their potential interferents. Two types of aggregation methods are evaluated; the first set do not compress the data, but incorporate effects from neighboring sensors into the output of each sensor in an array. The second method does result in compression of data and aggregates multiple sensor outputs into a single, more robust signal. Methods for processing data and aggregating and smoothing outputs from arrays of tin-oxide sensors are comparatively analyzed. Processing parameters studied include those related to simple averaging, linear-weighted averaging, and exponential smoothing across operating temperature and across type of sensing film in the dimensionality of the array. Aggregation techniques are evaluated during various stages of both the transient and steady-state response of the array to quantify the early decision-making capability of the array over that of a single or small number of unprocessed sensors. Aggregation strategies are studied in combination, and results are extracted by quantitatively measuring the goodness of clustering for each case. Cluster analysis, including principal component analysis (PCA), is used to determine which of these processing techniques are most effective. shown that aggregation methods, whether they reduce transmission bandwidth or not, improve the performance of a 30-element, tin-oxide heterogeneous sensor array in discriminating among common breath alcohol components (ethanols), their interferents (acetone, formaldehyde, isopropyl), and a contrast substance (ammonia). Aggregation generates a best-case 42% improvement in separability of clusters and 6.25% improvement in the tightness of clusters. Results are shown that clearly demonstrate the usefulness of aggregation in heterogeneous arrays among sensors whose outputs possess an appreciably degree of correlation (overlapping specificity).