Lightweight Neural Network for Gas Identification Based on Semiconductor Sensor
Jianbin Pan, Aijun Yang, Dawei Wang, Jifeng Chu, Fangfei Lei, Xiaohua Wang, Mingzhe Rong
Abstract
This article proposes a lightweight network called multiscale convolutional neural network with attention (MCNA), which combines a multiscale deep convolutional network with a self-attention mechanism. MCNA identifies ambient gases through signals of semiconductor gas sensor arrays, despite poor selectivity and drift problems. Notably, MCNA extracts temporal features of each signal and relevance among different signals more effectively than deep convolutional networks. MCNA requires much fewer parameters and computation costs than previous deep learning networks, but it still achieves the same high gas identification accuracy; this is crucial for gas sensing embedded systems. When the operating conditions of the gas sensor array change, it also exhibits better generalization ability and identification accuracy. We also discuss the effects of different MCNA architecture parameters and compare MCNA and other baseline approaches.