DCIR-Net: A Temporal-Scale Deformable Architecture for Cross-Speed sEMG-Based Upper Limb Kinematics Decoding

Abstract

Continuous joint angle prediction from surface electromyography (sEMG) is a core task in upper-limb motion intent decoding for prosthetics, human-robot interaction, and rehabilitation. Speed variation in natural movement causes sEMG amplitude, frequency content, and temporal activation duration to shift systematically, yet existing deep learning decoders are predominantly trained and evaluated under speed-matched conditions, leaving cross-speed generalization largely uncharacterized. DCIR-Net, a temporal-scale deformable pyramid with learnable residual gate fusion cascaded with ConvLSTM and a sparseattention encoder, is proposed for sEMG-based decoding of 4-DOF upper limb joint angles under unseen motion speeds. A systematic cross-speed benchmark is further established under a strict leave-one-subject-out protocol that simultaneously withholds unseen subjects and unseen speed conditions from training. Evaluated on the 13-subject, four-motion ULTRA-MoCap dataset using only sEMG, DCIR-Net achieves a mean NRMSE of 0.2422 across 12 action-speed test conditions, reducing mean NRMSE by 5.9\% over Informer (0.2574), and attains a positive macroaverage $R^{2}$ of 0.042. Baseline models yield negative macroaverage $R^{2}$; DCIR-Net achieves positive $R^{2}$ in seven of twelve conditions, including $\mathrm{R}^{2}=0.46$ on Crossbody Reach Normal, and reduces NRMSE by up to 21\% over Transformer on Shoulder Rotation.