Cross-device calibration of wearable sEMG gesture recognition using spatio-spectral encoding and frozen-statistic adaptation.

PACE-Net is a spatio-spectral few-shot calibration framework for wearable sEMG recognition under heterogeneous cross-device transfer. It converts multichannel sEMG into continuous-wavelet time-frequency representations and combines coordinate attention, relaxed-margin metric learning, frozen-statistic adaptation, and test-time soft voting to improve discrimination and stabilize calibration from few labeled samples. On a shared six-class NinaPro DB2/DB5 gesture space under mixed-domain subject-disjoint evaluation, PACE-FixBN reached 93.10%/97.21% (K=5/10) on DB2 target-like subjects and improved DB5 from 44.45% to 80.27%/86.49%; 50-subject accuracy was 90.23%/95.45%. Strict bidirectional validation gave 71.95% (DB2-to-DB5) and 61.54% (DB5-to-DB2) at K=10, confirming effective, practical rapid calibration.      .