Lower limb motion intention recognition using multi-source able-bodied gait signals.

Abstract

Accurate motion intention recognition for lower limb prostheses is challenging due to the limited use of distributed muscle signals. This study proposed a framework integrating musculoskeletal modelling, advanced feature processing, and feature selection to enhance motion intention recognition. A muscle screening method was developed to identify key muscles for accurately representing joint motion. Variational Mode Decomposition (VMD) was used to extract nonlinear features from sEMG signals, which were fused with linear features from joint angles and moments. Random Forest with Bagging (RF-Bagging) selected key features, reducing redundancy. An LSTM regression model was trained for continuous joint motion prediction. Experimental results from 20 subjects showed that using carefully screened muscles significantly improved prediction accuracy. Knee joint predictions (angle/moment) achieved MAE of 2.63±0.68 deg/0.020±0.006 Nm/kg, RMSE of 4.03±1.13 deg/0.029±0.010 Nm/kg, R² of 0.93±0.05/0.95±0.036, MSE of 17.49±10.78 deg/0.0009±0.0006 Nm/kg, and RPD of 1.011±0.049/1.019±0.034, while ankle joint predictions showed MAE of 1.32±0.31 deg/0.035±0.012 Nm/kg, RMSE of 1.72±0.43 deg/0.050±0.020 Nm/kg, R² of 0.93±0.03/0.98±0.021, MSE of 3.13±1.47 deg/0.002±0.002 Nm/kg, and RPD of 1.004±0.051/1.020±0.040. This study demonstrated that integrating optimal muscle selection, nonlinear feature enhancement, and intelligent feature selection significantly improves motion intention recognition.