Deep Learning for Electromyographic Hand Gesture Signal Classification Using Transfer Learning.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Human-Computer Interaction

Background:

• Deep learning excels at feature learning but requires extensive data for electromyography (EMG)-based gesture recognition.
• Manual data generation for deep learning in EMG gesture recognition is labor-intensive.
• Aggregating data across users offers a potential solution to reduce individual recording burden.

Purpose of the Study:

• To investigate the efficacy of transfer learning on aggregated multi-user EMG data for improved gesture recognition.
• To reduce the data recording effort required for training deep learning models in EMG gesture recognition.
• To enhance the generalization capabilities of deep learning models by leveraging diverse user signals.

Main Methods:

• Collected EMG data from multiple participants using Myo armbands.
• Applied transfer learning techniques on aggregated datasets for pre-training deep learning models.
• Evaluated three deep learning network architectures using raw EMG, spectrograms, and continuous wavelet transform (CWT) as input modalities.

Main Results:

• The proposed transfer learning scheme significantly improved performance across all tested deep learning networks and modalities.
• Achieved high offline accuracies, including 98.31% for 7 gestures (CWT-based ConvNet) and 68.98% for 18 gestures (raw EMG-based ConvNet).
• A use-case study demonstrated that real-time feedback enables users to adapt strategies, mitigating accuracy degradation over time.

Conclusions:

• Transfer learning on aggregated multi-user EMG data is a viable strategy to enhance deep learning-based gesture recognition.
• This approach effectively reduces the data collection burden while improving model performance and generalization.
• Real-time feedback can further optimize user performance and model robustness in practical applications.