A Lightweight Multi-Scale Convolutional Network with Gramian Angular Field Encoding for VOC Classification

Accurate classification of volatile organic compounds (VOCs) is important for environmental monitoring and industrial safety via electronic nose (E-nose) systems. However, extracting discriminative features from dynamic one-dimensional sensor responses remains challenging, especially when the recognition model is expected to maintain low computational complexity. This study introduces MSD-GasNet, a lightweight multi-scale depthwise convolutional network combined with Gramian Angular Summation Field (GASF) encoding, for VOC classification using E-nose response signals. The gas-sensing response curves are first transformed into two-dimensional GASF images to preserve temporal correlation information and provide structured inputs for convolutional feature learning. MSD-GasNet further adopts parallel 3 × 3 and 5 × 5 depthwise convolutional branches with feature fusion to capture local response details and broader morphology-related patterns while reducing parameter redundancy. Evaluated on Dataset 1, which contains five representative VOC categories including 1-butanol, acetone, benzaldehyde, butyl acetate, and dimethylbenzene, MSD-GasNet achieves an accuracy of 96.80 ± 0.78%, with 796.6 K parameters and 2.54 ms inference time per sample. Compared with traditional machine learning classifiers, conventional CNN baselines, recent lightweight networks, and a single-scale ablation model, MSD-GasNet shows better classification performance under the current five-class setting. An additional independent validation on Dataset 2 achieves an accuracy of 95.12 ± 1.11% under a chronological train/test split, further supporting the generalization potential of the proposed method. This work provides a GASF-based lightweight multi-scale framework with potential for efficient VOC recognition in portable or resource-limited E-nose applications.