Prediction of Knee Angle by EMG Regression Based on BP Neural Network

Accurate prediction of knee joint angles using surface electromyography (sEMG) signals is critical for applications such as exoskeleton control and rehabilitation, yet remains challenging due to the time-varying, nonlinear nature of sEMG and individual variability. This study proposes a backpropagation (BP) neural network-based regression model to address these challenges. The model employs a dual-hiddenlayer architecture to capture nonlinear relationships between multi-channel sEMG features (including time-domain, frequency-domain, and wavelet-based coefficients) and knee joint angles. Experimental data from quadriceps muscles were preprocessed with sliding window segmentation, time-domain shifting (200 ms ahead prediction), and feature normalization. Results demonstrated high prediction accuracy, with R -values exceeding 0.98 for arm joint angles and 0.94 for knee angles, alongside robust generalization on test datasets. The model's computational efficiency and ability to handle dynamic movements highlight its potential for real-time biomechanical applications.