文献类型：期刊文献

英文题名：A lightweight temporal convolutional network for real-time gas molecule recognition

作者：Zhang, Chao[1,2,3];Zhang, Guozhu[1,2,3];Ling, Xiaofeng[4];Wang, Zeyu[1,2,3];Zhang, Shunping[5];Nagashima, Kazuki[6,7];Wang, Tao[1,2,3];Xuan, Fu-Zhen[1,2,3]

机构：[1]Shanghai Key Lab Intelligent Sensing & Detect Tech, Shanghai, Peoples R China;[2]East China Univ Sci & Technol, Sch Mech & Power Engn, Shanghai, Peoples R China;[3]East China Univ Sci & Technol, Key Lab Pressure Syst & Safety, Minist Educ, Shanghai, Peoples R China;[4]East China Univ Sci & Technol, Sch Informat Sci & Engn, Shanghai, Peoples R China;[5]Huazhong Univ Sci & Technol, Sch Mat Sci & Engn, Wuhan 530074, Hubei, Peoples R China;[6]Hokkaido Univ, Grad Sch Chem Sci & Engn, N13W8 Kita, Sapporo, Hokkaido 0608628, Japan;[7]Hokkaido Univ, Res Inst Elect Sci, Green Nanotechnol Res Ctr, N20W10 Kita, Sapporo, Hokkaido 0010020, Japan

年份：2026

卷号：468

外文期刊名：SENSORS AND ACTUATORS B-CHEMICAL

收录：;EI(收录号：20263221255827);Scopus(收录号：2-s2.0-105046472455);WOS:【SCI-EXPANDED(收录号:WOS:001844744300001)】；

基金：This work was supported by the National Natural Science Foundation of China (Grant Number: 52375148 and 52321002), the Natural Science Foundation of Shanghai (Grant No. 23ZR1417000), and the Sakura Science Exchange Program in Japan Science and Technology Agency (JST) (Grant Number: S2024F0200227). G.Z. and K.N. were partly supported by the Cooperative Research Program of "Network Joint Research Center for Materials and Devices" from the Ministry of Education, Culture, Sports, Science and Technology (MEXT), Japan (20263004).

语种：英文

外文关键词：Machine olfaction; Electronic nose; Lightweight temporal convolutional network; Edge, computing; Gas molecule recognition

摘要：Deep learning has become a central component of intelligent machine olfaction, enabling effective analysis of complex gas response dynamics. However, existing time-series models, such as recurrent neural networks (RNNs), suffer from limited training efficiency and degraded long-sequence representation, which restrict their deployment in resource-constrained edge devices. Here, we propose a lightweight separable temporal convolutional network (STCN) that enables real-time gas molecule recognition and quantitative analysis on a custom-built electronic nose. Implemented on a low-power ESP32 microcontroller, the proposed STCN enables fully standalone operation without reliance on external computing resources. By integrating the STCN with a sliding-window strategy, transient and dynamic gas response features are effectively captured, enabling simultaneous molecule identification and concentration estimation. Experiments conducted on twelve mixed-gas combinations demonstrate an average multi-label classification accuracy of 94.4% and an average regression R2 of 0.985. These results indicate that the model successfully learns intrinsic coupling patterns among gas components and achieves component-level recognition in multi-component environments, confirming accurate and robust molecule-level identification and quantification under complex overlapping conditions.