DeepmmWave: Attention-Guided Framework for Contactless Liquid Identification and Adulteration Using mmWave Radar

Traditional millimeter-wave radar-based liquid sensing systems face practical deployment challenges due to weak interliquid contrast, frequency-dependent dispersion, and sensitivity to container geometry and measurement variability. This work presents DeepmmWave, an attention-guided framework for contactless mmWave radar-based multiclass liquid identification and adulteration that integrates frequency diversity with adaptive feature modeling to improve robustness under realistic sensing conditions. In particular, a compact 12-D physics-guided descriptor is extracted from peak-centered FMCW range responses and processed by a feature-gated residual neural architecture that performs sample-adaptive recalibration of heterogeneous features. The end-to-end pipeline includes background-suppressed range processing, peak-aligned region-of-interest extraction, descriptor standardization, attention-based feature gating, and softmax-based inference. Operating in the 60-64-GHz band on a millimeter-wave FMCW radar platform, the proposed framework enables liquid identification and adulteration assessment without using container labels at inference time within the three evaluated container materials. The descriptor combines conventional radar statistics with electromagnetic proxy features reflecting attenuation, phase dispersion, spatial energy morphology, and temporal stability. Performance is assessed using grouped stratified fivefold cross-validation for ten-class liquid identification and ten-class adulteration classification. Liquid identification achieves a mean test accuracy of 0.978 ± 0.018 , with macroaveraged precision, recall, and F1-score of 0.96, 0.95, and 0.955, respectively. Adulteration classification attains a pooled accuracy of 97.60%, with macroprecision of 0.9766, macrorecall of 0.9720, and macro F1-score of 0.9743. Overall, the adaptive frequency-aware model consistently outperforms fixed-weight and nonattention baselines, providing a robust and interpretable solution under practical measurement variability.