sEMG gait phase classification based on CNN-transformer and transfer learning

Accurate motion intention recognition is essential for active control of lower-limb rehabilitation exoskeletons. To address inter-subject variability in sEMG signals and real-time requirements, a lightweight CNN-Transformer hybrid with two-stage transfer learning is proposed. Inverted residual CNNs extract multi-dimensional local features while Transformer modules capture long-range dependencies, achieving merely 0.35M parameters. Personalized adaptation is realized via source-domain pre-training and target-domain fine-tuning. The model achieves 5.6 ms inference latency, with single-subject and cross-subject accuracies of 93.46% and 92.97%, respectively, providing a robust patient-adaptive solution for real-time exoskeleton control.