EMG-Based Dual-Branch Deep Learning Framework With Transfer Learning for Lower Limb Motion Classification and Joint Angle Estimation

Abstract

Wearable surface electromyography (sEMG) sensors capture neuromuscular signals for analyzing lower limb movements, exoskeleton robotics control, and rehabilitation application. However, simultaneous motion classification and continuous joint angle prediction remain challenging, particularly with limited patient data. This study introduces DBWCT-EMGNet, a novel deep learning framework with a dual-branch architecture augmented with transfer learning. The main structure integrates a Improve WaveNet fusion layer for multi-scale feature extraction, convolutional block attention module (CBAM) attention for enhanced feature focus. The classification branch integrates a Transformer encoder for robust motion recognition. The regression branch employs a Temporal Convolutional Attention network for precise joint angle prediction. Transfer learning adapts models trained on healthy subjects to patient data to mitigate data scarcity issues. Compared to models such as CNN-BiLSTM and CNN-TCN, DBWCT-EMGNet achieved superior intra-subject performance (classification accuracy: 99.86% ± 0.11%; joint angle $R^2$: 0.98 ± 0.04, RMSE: 1.40° ± 1.64°). Transfer learning improved inter-subject results by 21.7% in accuracy, 24.7% in $R^2$, and 67.6% in RMSE. By enabling accurate motion analysis and generalization across subjects, DBWCT-EMGNet shows strong potential for developing advanced sensor-based assistive and rehabilitative technologies.