.png){kind=icon}
.png){kind=icon}
Return
Generalization-Enhanced Cross-Set Upper-Limb Multi-Joint Torque Prediction via a Dual-Stream Time–Frequency Attention Network
.png){kind=icon}
OA
AI
K
K
X
Y
B
C
Z
DOI：10.1109/tnsre.2026.3715302.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Mapping surface electromyography (sEMG) to multi-joint torque is critical for active stroke rehabilitation. While compounded non-ideal factors between discontinuous exercise sets impose stringent demands on amplitude–phase robustness, synergy preservation, and noise resistance. This study designed and validated a dual-stream time-frequency attention network incorporating Luenberger observation feedback (DSA-Luen) to address the issue of sEMG pattern inconsistency arising from inter-set variations in repetitive upper-limb rehabilitation. The framework integrates DSA based on Real-Time Fast Fourier Transform (RTFFT) with a Luen-Mamba module inspired by the Luenberger observer. The former captures more robust dynamic sEMG information, while the latter utilizes output feedback for dynamic correction, effectively suppressing noise and enhancing prediction accuracy. On datasets from healthy subjects and stroke patients performing multiple upper-limb movements, DSA-Luen consistently outperforms time-domain sequence models and attention baselines. Averaged across four Degrees of Freedom (DOFs), cross-set evaluation attains NRMSE 0.130 and <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\text {R}^{{2}}~0.703$ </tex-math></inline-formula>, representing a 23% reduction in NRMSE and a 36% increase in <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\text {R}^{{2}}$ </tex-math></inline-formula> relative to the strongest attention baseline; cross-movement evaluation further achieves NRMSE 0.122 and <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\text {R}^{{2}}~0.704$ </tex-math></inline-formula>. In stroke patient trials, DSA-Luen maintained a positive mean <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\text {R}^{{2}}$ </tex-math></inline-formula> of 0.527, reduced NRMSE by approximately 30%, and increased <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\text {R}^{{2}}$ </tex-math></inline-formula> by more than five times compared to the strongest baseline (Informer), while several baselines deteriorated to near-zero or negative values. These results demonstrate that the method maintains strong adaptability and robustness under complex non-ideal conditions. In clinical settings, it provides a transferable foundation for intent decoding in personalised active rehabilitation and reduces the recalibration burden.
Keywords:
Rehabilitation engineering
electromyography (EMG)
time–frequency analysis
torque estimation
robustness
Journal
IF:
5.2
Papers:
448
Citations:
1.6W
Organization
Cited Papers