TY - GEN
T1 - Prediction of Exoskeleton Knee Angle Based on Unilateral Lower Limb Information
AU - Song, Zihe
AU - Liu, Yali
AU - Ma, Xunju
AU - Zhang, Keshi
AU - Li, Xiao
AU - Song, Qiuzhi
N1 - Publisher Copyright:
© 2026 Copyright held by the owner/author(s).
PY - 2026/7/23
Y1 - 2026/7/23
N2 - Unilateral lower-limb exoskeletons improve system convenience and cost-effectiveness through reduced sensor and mechanical components. Unilateral lower-limb biomechanical signals alone pose challenges for precise, continuous gait perception in exoskeleton research. An LSTM-TPA (Long Short-Term Memory with Temporal Pattern Attention) model is proposed for continuous joint angle prediction based on unilateral lower limb information. Built on the LSTM framework, the model integrates a temporal pattern attention mechanism that adaptively focuses on critical time intervals within historical gait sequences crucial for current predictions, efficiently capturing the dynamic evolution of movement patterns. Offline prediction results based on patient experiments show a Pearson Correlation Coefficient of 0.9713 for the 200 ms prediction task, with MAE and RMSE reduced to 1.9149°and 3.4737°, respectively. This corresponds to 29.97% and 30.02% reductions in MAE and RMSE compared with the CNN-LSTM model. The proposed method exhibits effectiveness for continuous joint angle prediction under unilateral signal conditions, offering a reliable technical pathway for intelligent control of lightweight, low-cost lower-limb exoskeletons.
AB - Unilateral lower-limb exoskeletons improve system convenience and cost-effectiveness through reduced sensor and mechanical components. Unilateral lower-limb biomechanical signals alone pose challenges for precise, continuous gait perception in exoskeleton research. An LSTM-TPA (Long Short-Term Memory with Temporal Pattern Attention) model is proposed for continuous joint angle prediction based on unilateral lower limb information. Built on the LSTM framework, the model integrates a temporal pattern attention mechanism that adaptively focuses on critical time intervals within historical gait sequences crucial for current predictions, efficiently capturing the dynamic evolution of movement patterns. Offline prediction results based on patient experiments show a Pearson Correlation Coefficient of 0.9713 for the 200 ms prediction task, with MAE and RMSE reduced to 1.9149°and 3.4737°, respectively. This corresponds to 29.97% and 30.02% reductions in MAE and RMSE compared with the CNN-LSTM model. The proposed method exhibits effectiveness for continuous joint angle prediction under unilateral signal conditions, offering a reliable technical pathway for intelligent control of lightweight, low-cost lower-limb exoskeletons.
KW - Joint angle prediction
KW - Lower-limb exoskeleton
KW - Temporal pattern attention
KW - Unilateral lower limb information
UR - https://www.scopus.com/pages/publications/105046436143
U2 - 10.1145/3820709.3820718
DO - 10.1145/3820709.3820718
M3 - Conference contribution
AN - SCOPUS:105046436143
T3 - Proceedings of 2026 6th International Conference on Robotics and Control Engineering, RobCE 2026
SP - 172
EP - 178
BT - Proceedings of 2026 6th International Conference on Robotics and Control Engineering, RobCE 2026
A2 - Dan, Zhang
A2 - Song, Aiguo
A2 - Habib, Maki
PB - Association for Computing Machinery, Inc
T2 - 2026 6th International Conference on Robotics and Control Engineering, RobCE 2026
Y2 - 21 May 2026 through 23 May 2026
ER -