A Method for Optimizing the Design of Heterogeneous Nano Gas Chemiresistor Arrays

A method for optimizing the design of heterogeneous gas chemiresistor arrays to maximize classification accuracy is presented. The features used for classification include coefficients from a Fast Fourier Transform, properties of a piecewise representation of the response shape, and features characterizing the response time. The novel response time features are designed to be insensitive to noise and sensor drift. This work introduces a novel approach for leveraging experimental data to create large synthetic datasets for training classifiers. Pairs of time-series sensor response measurements are randomly combined with a noise model to create synthetic sensor responses, which are then combined to create synthetic array responses. J48 decision trees are used to classify species and support vector machine regression models are used to determine the concentration once the species is known. The results demonstrate the value of array optimization, as the highest classification accuracy is achieved using a subset of the available sensor designs. J48 decision trees proved to be efficient for use in optimization and achieved high accuracy. Separating the tasks of classifying species and identifying concentration also proved to be effective. The techniques were applied to the design of an array for classifying H-2, H2S, NH3, and NO2. The optimal array achieved 95.7 % accuracy at classifying species and an average correlation coefficient of 0.92 for determining concentration.