文献类型：期刊文献

英文题名：A transferable multi-task model for enhanced volatile organic compounds detection across multiple domains

作者：Cheng, Weiwei[1];Wang, Tao[2];Zhu, Jiaqing[2];Chen, Lechen[3];Ni, Wangze[3];Yang, Zhi[3];Zhang, Bowei[2];Xu, Shusheng[1];Xuan, Fuzhen[2]

机构：[1]Shanghai Univ Engn Sci, Sch Mat Sci & Engn, Shanghai 201620, Peoples R China;[2]East China Univ Sci & Technol, Sch Mech & Power Engn, Shanghai Key Lab Intelligent Sensing & Detect Tech, Shanghai 200237, Peoples R China;[3]Shanghai Jiao Tong Univ, Sch Elect Informat & Elect Engn, Natl Key Lab Adv Micro & Nano Manufacture Technol, Shanghai 200240, Peoples R China

年份：2026

卷号：447

外文期刊名：SENSORS AND ACTUATORS B-CHEMICAL

收录：;EI(收录号：20253919215184);WOS:【SCI-EXPANDED(收录号:WOS:001582846900005)】；

基金：This work was supported by the National Natural Science Foundation of China (62301314 and 52321002) . This work also received support from the In Situ Devices Center of East China Normal University.

语种：英文

外文关键词：Transfer learning; Multi-task deep learning; Convolutional neural networks; Gas sensors; VOCs detection

摘要：As a data-driven approach, top-performing deep architectures often require extensive and high-quality data. However, data scarcity is constantly one of the critical bottlenecks to utilizing these data-driven methods in volatile organic compounds detection. In the absence of sufficient data for comprehensive training, transfer learning always provides an attractive option. Here we propose a transferable multi-task model that combines transfer learning and multi-task learning, enabling the simultaneous completion of four tasks with high precision while minimizing training time and the trainable parameters. We design four basic tasks inspired by the knowledge of various modalities, including gas classification, concentration prediction, state assessment, and anomaly detection. The proposed model employed a two-stage pipeline consisting of a feature extractor and a multi-head predictor. A large-scale source dataset is first used to train the model, followed by transfer learning to adapt and fine-tune two smaller-scale target datasets. By migrating the parameters of the feature extractor and fine-tuning label predictor across the target dataset, the approach achieves comparable accuracy to the pretrained model, using nearly 1/5 of the training parameters and 1/4 of the training time. Specifically, the model demonstrates exceptional accuracy across the four tasks: gas classification accuracy exceeds 96 %, state assessment accuracy surpasses 98 %, concentration R2 score exceeds 0.95, and anomaly detection accuracy exceeds 99 %. This innovative approach establishes the potential of leveraging transfer learning and multi-task learning to overcome data limitations, setting a new benchmark for versatile volatile organic compounds detection systems.