An All-in-One Electronic Nose With a Multi-Scale Temporal Shift and Depth Dynamic Aggregation Network for Low-Concentration Malodorous Gas Classification

Detecting low-concentration malodorous gases of down to single-digit ppm levels remains challenging due to the weak and overlapping transient responses of metal oxide semiconductor (MOS) sensors and limited, as well as imbalanced datasets. In this work, we propose an all-in-one electronic nose (E-nose) prototype and a multi-scale, one-dimensional convolutional neural network (CNN) incorporating a temporal-shift and depth dynamic aggregation (TS-DDA) module for robust odor classification. The E-nose instrument adopts a modular design comprising: 1) a sensing module with a 16-channel MOS sensor array enclosed in an annular gas chamber validated through k-ω computational fluid dynamics (CFD) simulation for uniform flow and rapid desorption (<40 s residual washout); and (2) a data-acquisition and control module, implemented on a single custom printed circuit board (PCB), providing precise sampling, pump and valve control. Additionally, the proposed MS-TS-DDA network enhances temporal feature density and multi-depth information fusion while maintaining low computational cost. A controlled laboratory dataset consisting of 497 samples covering eight types of low-concentration malodorous gases (0.5-60 ppm) was collected and balanced via a temporal oversampling strategy inspired by T-SMOTE. The proposed method achieves a mean classification accuracy of 95.15% under 5-fold cross-validation, outperforming classical CNN baselines. These results indicate that the proposed framework provides a compact, cost-effective and robust solution for low-concentration odor detection under resource-constrained conditions.