MBWCT-Net: Multiscale Transient Learning for Early Discrimination of Wood-Extracted VOC Mixtures via an Electronic Nose

Abstract

Metal oxide semiconductor electronic noses offer a compact and nondestructive method for volatile organic compound (VOC) analysis. However, reliable recognition of multicomponent mixtures remains challenging because of sensor cross-sensitivity, nonlinear adsorption–desorption dynamics, sensor drift, and long acquisition cycles. Eucalyptus, Rosewood, and Sandalwood essential oils were used as controlled plant-derived VOC benchmarks. Static-headspace gas chromatography–mass spectrometry showed partially shared constituents but distinct dominant headspace fingerprints: Eucalyptus was rich in eucalyptol and monoterpenes, Rosewood was linalool-dominant, and Sandalwood showed a substantially different and partly unresolved profile. The three samples therefore did not exhibit uniformly high compositional similarity. We developed MBWCT-Net, a physics-motivated Multi-Branch Wavelet CNN–Transformer Network that combines a Multi-Scale Wavelet Parallel Network, Lightweight Cross-Attention, and a Multi-Scale Adaptive Transformer to model multiscale response components, cross-band coupling, and local–global temporal dependencies. Under an identical five-fold protocol, MBWCT-Net achieved 95.37 ± 0.61% accuracy and 95.37 ± 0.66% weighted F1, outperforming both raw-input and wavelet-enhanced baselines. Using the first 45 s of the transient response, accuracy reached 94.61 ± 0.70%, within 0.76 percentage points of the complete 60 s response. Deployment on an RK3588 platform yielded a single-sample inference latency of 12.4 ± 0.8 ms and a memory footprint of 48.5 MB. These measurements characterize computational feasibility and do not represent the duration of the complete sensing cycle. Response plots from three independent measurements and normalization-shape checks assessed short-term repeatability and preprocessing effects. Quantitative high-dimensional feature metrics and measured-noise physical-validity analyses evaluated class separability and the preservation of temporal morphology, respectively.