Gas Discrimination Analysis of Neural Network Algorithms for a Graphene-Based Electronic Nose

Electronic noses, such as chemiresistive gas sensors, are important tools for environmental monitoring, public health and food quality. Throughout the measurement process, pattern recognition algorithms are used to map the electrical sensor measurements to a gas concentration. When analyzing gaseous mixtures with such devices, cross-sensitivities are likely to occur, which are related to the selectivity of the sensor materials. In the prediction process, cross-sensitive sensor signals can lead to false positive predictions and a loss in measurement accuracy. In this work, we thoroughly study the impact of the degree of crosssensitivity on the gas concentration prediction performance of neural networks in graphene-based gas sensor arrays. The study was conducted by using a simulation model of a chemiresistive gas sensor to simulate an array of two sensors with different levels of cross-sensitivity to two different target gases. Subsequently, two neural network algorithms were trained and evaluated on the datasets with the varying cross-sensitivity levels. Our analysis shows that a certain threshold regarding the independence of the array signals is necessary in order to ensure a sufficient gas discrimination and concentration estimation performance of the applied machine learning algorithms. Furthermore, it was also observed that a combination of a highly and poorly selective sensor, as implemented by filtering techniques, can also provide adequate results, when using suitable algorithms.