Biosignal-based transferable attention Bi-ConvGRU deep network for hand-gesture recognition towards online upper-limb...

Area of Science:

• Biomedical Engineering
• Machine Learning
• Rehabilitation Technology

Background:

• Upper-limb amputation significantly impacts daily life and quality of life.
• Surface electromyogram (sEMG) signals offer a non-invasive method to monitor muscle activity for gesture recognition.
• Accurate recognition of hand gestures is crucial for restoring functionality after amputation.

Purpose of the Study:

• To develop a real-time deep learning model for automatic and reliable recognition of hand gestures using sEMG signals.
• To create a model capable of recognizing complex signals from both amputees and non-amputees.
• To improve the control of prosthetic devices through advanced bio-signal processing.

Main Methods:

• An attention bidirectional Convolutional Gated Recurrent Unit (Bi-ConvGRU) deep neural network was proposed for hand-gesture recognition.
• The model was trained on sEMG data from both amputees and non-amputees, utilizing 1D CNNs for feature extraction and Bi-GRU for sequential analysis.
• An attention mechanism was incorporated to enhance robustness to noise and irregularity in bio-data, alongside a novel transfer learning approach.

Main Results:

• The attention Bi-ConvGRU model achieved an average accuracy of 88.7% on the Ninapro benchmark database.
• This performance surpassed the state-of-the-art in 18-gesture recognition by 6.7%.
• The model demonstrated effective transfer learning, refining a baseline model with amputee data for personalized recognition.

Conclusions:

• The developed end-to-end deep learning model is the first to enable reliable predictive decision-making in short time windows (160ms).
• Reduced latency facilitates real-time, online bio-control of prosthetic devices.
• This technology holds significant potential for more intuitive and effective prosthetic limb control for amputees.