Lower Limb Movement Recognition Based on a Hybrid Deep Learning Model Using Surface Electromyography

In the fields of intelligent rehabilitation and human-computer interaction, the accurate recognition of lower limb motion patterns is crucial for achieving efficient assistance and precise control. This study proposes a TCN-Transformer-LSTM (TTLNet) model based on multi-module fusion for lower limb motion recognition. By integrating the advantages of three advanced architectures—Temporal Convolutional Network (TCN), Long Short-Term Memory (LSTM), and Transformer—the model constructs an efficient framework for lower limb motion recognition. First, surface electromyography (sEMG) signals from five types of lower limb movements (i.e., walking, standing, sitting, ascending stairs, and descending stairs) are collected and preprocessed. Then, six time-domain features are extracted from the preprocessed sEMG signals, including integrated electromyography (iEMG), variance of electromyography (VAR), waveform length (WL), zero crossing (ZC), slope sign change (SSC), and Willison amplitude (WAMP), which serve as inputs for the TTLNet model. Comparisons are made with TCN, LSTM, convolutional neural networks (CNN), and Support Vector Machine (SVM) models. Experimental results demonstrate that TTLNet achieves a mean accuracy of 98.82%, significantly outperforming TCN by 2.50%, LSTM by 4.46%, CNN by 3.24%, and SVM by 5.28%. Through systematic experiments, this study not only validates the superior performance of TTLNet in handling complex motion patterns but also provides an efficient and precise solution for lower limb motion recognition in the fields of intelligent rehabilitation and human-computer interaction.