MCCT-Net: A Hybrid Neural Network for Improving the Detection Performance of Electronic Nose System

Abstract

The electronic nose (e-nose) system utilizes a cross-sensitive gas sensor array to detect gas information. An effective gas information classification method is one of the key problems to improve the detection ability of e-nose. In this paper, a new deep learning algorithm is constructed to classify e-nose gas information. The Transformer structure is incorporated into the convolutional neural network (CNN) architecture to compensate for the limitation of local convolutions in capturing long-range dependencies and global feature correlations. Firstly, we construct a multi-scale convolutional encoder (MCE) to extract local gas features. Then, a channel-refiltering transformer block (CTB) is designed to learn global gas features. Finally, we combine MCE and CTB to establish a hybrid neural network for gas information classification, called multi-scale convolution and channel re-filtering transformer hybrid network (MCCT-Net). To verify the effectiveness of MCCT-Net, we conduct performance tests on two datasets. For Dataset 1, MCCT-Net obtains the 98.67% accuracy, 98.65% recall, and 98.69% precision. For Dataset 2, MCCT-Net obtains the 98.52% accuracy, 98.17% recall, and 98.69% precision. The results show that MCCT-Net can accurately classify gas information, which provides a method for improving the detection performance of the e-nose system.