Physics-Guided Domain-Quality Selective Learning for Cross-Domain Wi-Fi CSI Fall Recognition

Source abstract

The Physics-Guided Domain-Quality Selective Network (PG-DQSNet) combines measurement-derived motion channels with selective Wi-Fi fall recognition. Amplitude dynamics and sanitized relative-phase differences feed a multi-scale temporal encoder. Participant and room adversaries supervise the training representation. A separate class-conditional selector estimates prediction error from capture quality, transformed probability, dropout disagreement, and prototype distance. We evaluate 2,011 audited recordings with 429 demonstrated falls across three participants, three rooms, six dates, and 13 acquisition runs. Fitting, calibration, and testing use disjoint acquisition groups. Eight model configurations share the outer partitions and three random seeds. PG-DQSNet achieves person-disjoint macro F2 of 0.609 and room-disjoint macro F2 of 0.522. The gains over the amplitude-only configuration are 0.082 and 0.148. Recall is 94.1% and 76.2% under calibration-frozen thresholds. The corresponding specificity values are 6.4% and 31.9%. Learned selection reduces room-disjoint area under the risk–coverage curve from 0.653 to 0.507 at matched coverage. Complete domain results, threshold-independent ranking scores, and frozen-threshold review fractions characterize the measured decisions. The equations, editable architecture, and record-level outputs support reproduction of the grouped benchmark.