Vision transformer-based electronic nose for enhanced mixed gases classification

Abstract

The classification of mixed gases is one of the major functions of the electronic nose. To address the challenges associated with complex feature construction and inadequate feature extraction in gas classification, we propose a classification model for gas mixtures based on the vision transformer (ViT). The whole-process signals of the sensor array are taken as input signals in the proposed classification model, and self-attention mechanism is employed for the fusion of global information and adaptive feature extraction to make full use of the dependence of responses at different stages of the whole-process signals to improve the model’s classification accuracy. Our model exhibited a remarkable accuracy (96.66%) using a dataset containing acetone, methanol, ammonia, and their binary mixtures. In comparison, experiments conducted by support vector machine and a one-dimensional deep convolutional neural network model demonstrated classification accuracy of 90.56% and 92.75%, respectively. Experimental results indicate that the ViT gas classification model can be effectively combined with multi-channel time series data from the sensor array using the self-attention mechanism, thereby improving the accuracy of mixed gases classification. This advancement can be expected to become a standard method for classifying mixed gases.