The team led by Li Mingjie at Guangdong University of Technology published in ACS Sensors the results on gas identification using a MEMS dual-sensor array and multi-task deep learning
The team led by Li Mingjie at Guangdong University of Technology published a study in ACS Sensors, integrating PdO-modified SnO2/ZnO and MoS2/ZIF-67 onto MEMS microheaters to construct a dual-sensor array, combined with a multi-task model of a one-dimensional convolutional neural network and a bidirectional long short-term memory network, simultaneously achieving identification and concentration prediction of acetone, isopropanol, and mixed gases with simulated clinical ratios. On data collected at 5–25 ppm gases and separately at 7 and 10 ppm, the model used a 15-second time window, achieving an overall classification accuracy of 95.4% and a mean absolute error of 0.625 ppm for concentration prediction, with results updated every 15 seconds.
Body · AI translation
Recently, teacher Li Mingjie of our institute guided undergraduates to publish a research paper titled "Overcoming Cross-Sensitivity for the Accurate Identification of Acetone, Isopropanol, and Clinical Mixtures Using a MEMS Dual-Sensor Array and Multi-task Deep Learning" in 《ACS Sensors》 (CAS Zone 1, impact factor 10.9). Undergraduates of the 2023 cohort Yin Sihan, Li Jiasheng, and Chen Guanxi are the first, second, and third authors of the paper, respectively; Li Mingjie and Professor Zhang Ming of Hunan University serve jointly as corresponding authors, with Li Mingjie as the last corresponding author, and Guangdong University of Technology as the first affiliation of the paper.
Addressing the cross-sensitivity problem of semiconductor gas sensors, the research team integrated PdO-modified SnO2/ZnO and MoS2/ZIF-67 onto a micro-electromechanical systems (MEMS) micro-heater to construct a dual-sensor array, and combined a one-dimensional convolutional neural network with a bidirectional long short-term memory network to build a multi-task model, utilizing complementary dynamic response information to simultaneously achieve identification and concentration prediction of acetone, isopropanol, and their simulated clinical-ratio gas mixtures.
Figure | Schematic diagram of the electronic nose system combining the MEMS dual-sensor array with multi-task deep learning
The experiments covered 5–25 ppm of acetone, isopropanol, and three preset-ratio gas mixtures. On data of 7 and 10 ppm separately collected by the same array, the model used a 15-second time window to achieve an overall classification accuracy of 95.4% (including the air category) and a mean absolute error of 0.625 ppm for concentration prediction, with prediction results updated every 15 seconds, reducing dependence on the complete steady-state response.
Figure | Multi-task deep learning model and performance evaluation
Through the collaborative design of sensitive materials, MEMS devices, and multi-task deep learning, this study can simultaneously accomplish gas identification and concentration prediction using only two sensors, effectively improving the discrimination capability for single and mixed gases, providing a new technical path for overcoming cross-sensitivity and developing miniaturized intelligent electronic noses, and demonstrating application potential in the field of portable breath analysis. The research was supported by projects including the National Natural Science Foundation of China (U25A20491) and the National Key R&D Program (2024YFB3614200).
Figure | Group photo of the project team