Broadly responsive vapor sensor arrays, or so-called “electronic noses,” have been explored and/or used for many years as a means to detect the vapors present in the headspace of a variety of targets, such as coffees, wines, vinegars, oils, explosives, and nerve agents, and for disease diagnosis. Electronic nose sensing modalities often exhibit a response that is linear with concentration, and additive with respect to multiple vapors. Ideally, one could simply train an array towards the pure vapors of interest, and use that pure vapor training to identify either pure vapors or vapor mixtures during field-testing. This, however, has proven difficult, and has limited the utility of this vapor detection approach for a number of applications. This thesis utilized a low-cost, low-power sensing modality, insulator – carbon black composite chemiresistors, and exploited their linear response properties to enhance the classification rates of both pure vapors and vapor mixtures, based on pure vapor training. Sensors utilizing non-polymeric, small organic molecules as the insulating component were demonstrated to offer enhanced separation between pure vapor response clusters, and lower detection limits, relative to the traditional use of polymers as the insulating phase. These sensors were then used in a sensing geometry that induced a space- and time, or spatiotemporal (ST) dependence, to the sensor response, which increased the amount of chemical information extracted from the sensor response. This ST response information allowed for the correct classification of vapor mixtures consisting of up to 5 components, with training on only pure vapors. A mass uptake model for the ST response of the sensors was developed, and vapor detection and mixture analysis was simulated for chamber geometries and vapor delivery flow rates spanning several orders of magnitude. The data were first used to define an optimized ST sensing regime for mixture analysis, based on two dimensionless Peclet number analogs. The data were then used to identify the inherent properties of the pure vapor training data that allowed for mixture analysis to be performed at high levels, specifically that the minimum resolution factor between all binary vapor combinations in the training library was sufficiently high. Finally, the utility of the ST response was demonstrated to offer enhanced pure-vapor classification rates, relative to the traditional steady state approach typically employed with broadly responsive array-based sensing. These enhanced classification rates were demonstrated using a number of classification algorithms, including a bioinspired algorithm based on Fisher’s Linear Discriminant. In summary, the results demonstrated herein quantify, in different ways, what is required for classification optimization, and in doing so increase the utility of this approach to vapor detection for a number of applications.
Marc D. Woodka
Journal name not available for this finding