文献类型：期刊文献

英文题名：Research on Binary Mixed VOCs Gas Identification Method Based on Multi-Task Learning

作者：Mei, Haixia[1,2];Yang, Ruiming[1];Peng, Jingyi[1];Meng, Keyu[1];Wang, Tao[3];Wang, Lijie[2]

机构：[1]Changchun Univ, Key Lab Intelligent Rehabil & Barrier Free Disable, Minist Educ, Changchun 130022, Peoples R China;[2]Jilin Univ, Coll Elect Sci & Engn, State Key Lab Integrated Optoelect, Changchun 130012, Peoples R China;[3]East China Univ Sci & Technol, Sch Mech & Power Engn, Shanghai Key Lab Intelligent Sensing & Detect Tech, Shanghai 200237, Peoples R China

年份：2025

卷号：25

期号：8

外文期刊名：SENSORS

收录：;EI(收录号：20251818328951);WOS:【SCI-EXPANDED(收录号:WOS:001475631400001)】；

基金：This work was supported by the Jilin Provincial Science and Technology Development Program Project (YDZJ202501ZYTS591) and the National Natural Science Foundation of China (52205586 and 62301314).

语种：英文

外文关键词：gas sensor; multi-task learning; mixed gases; feature fusion

摘要：Highlights What are the main findings? A multi-task residual network (MRCA) which generates dynamic feature depending on the cross-fusion module was invented to perform VOCs gas component identification and concentration prediction. The dynamic weighted loss function, which can dynamically adjust the weight according to the training progress of each task. What is the implication of the main finding? The MRCA model showed a high classification accuracy of 94.86%, as well as achieving an R2 score up to 0.95. Using only 35% of the total data length as input data leads to excellent identification performance.Highlights What are the main findings? A multi-task residual network (MRCA) which generates dynamic feature depending on the cross-fusion module was invented to perform VOCs gas component identification and concentration prediction. The dynamic weighted loss function, which can dynamically adjust the weight according to the training progress of each task. What is the implication of the main finding? The MRCA model showed a high classification accuracy of 94.86%, as well as achieving an R2 score up to 0.95. Using only 35% of the total data length as input data leads to excellent identification performance.Abstract Traditional volatile organic compounds (VOCs) detection models separate component identification and concentration prediction, leading to low feature utilization and limited learning in small-sample scenarios. Here, we realize a Residual Fusion Network based on multi-task learning (MTL-RCANet) to implement component identification and concentration prediction of VOCs. The model integrates channel attention mechanisms and cross-fusion modules to enhance feature extraction capabilities and task synergy. To further balance the tasks, a dynamic weighted loss function is incorporated to adjust weights dynamically according to the training progress of each task, thereby enhancing the overall performance of the model. The proposed network achieves an accuracy of 94.86% and an R2 score of 0.95. Comparative experiments reveal that using only 35% of the total data length as input data yields excellent identification performance. Moreover, multi-task learning effectively integrates feature information across tasks, significantly improving model efficiency compared to single-task learning.