A physics-regularized reliability-aware learning framework for non-contact blink monitoring using millimeter-wave radar

Continuous monitoring of eye-blink activity provides a non-contact behavioral indicator relevant to fatigue-related applications and human-state monitoring. Reliable radar-based sensing remains difficult under the tested single-user office-like desktop conditions because blink-induced eyelid motion is weak, short-lived, and easily contaminated by respiration, head motion, and transient clutter. This study proposes a physics-regularized, reliability-aware artificial intelligence framework for 77-GHz millimeter-wave radar that combines adaptive variational mode decomposition with the Multi-modal Adaptive Deep Network (MAD-Net). Time-domain phase and frequency-domain representations are processed in parallel, aligned, fused, and temporally modeled for blink recognition. Doppler bandwidth is used only as a training-time supervisory signal for physics regularization and is not required during inference. Under subject-level macro-averaging across the 30 outer leave-one-subject-out folds on the retained labeled-window evaluation set, the proposed framework achieves 92.5% accuracy, 91.0% F1-score, and a 1.5% window-level false-positive rate, with an area under the receiver operating characteristic curve of 0.943. The results support non-contact blink monitoring under the tested single-user office-like desktop conditions.