Dynamic Graph Tranformer-Convolutional Network for Multi-Channel EMG Gesture Recognition.

Area of Science:

• Biomedical Engineering
• Signal Processing
• Machine Learning

Background:

• Electromyography (EMG) signal classification is vital for applications like hand gesture recognition.
• Accurate modeling of temporal and spatial dependencies in multi-channel EMG data presents significant challenges.
• Existing methods often struggle to dynamically capture evolving muscle coordination during gestures.

Purpose of the Study:

• To propose a novel Dynamic Graph Transformer-Convolutional Network (GTCN) for improved EMG signal classification.
• To effectively model both long-range temporal dependencies and dynamic inter-channel relationships in EMG data.
• To enhance the accuracy of hand gesture recognition using EMG signals.

Main Methods:

• Developed a GTCN integrating a dynamic graph convolutional network (GCN) and a Transformer architecture.
• The dynamic GCN captures evolving inter-channel relationships, representing muscle coordination.
• The Transformer models long-range temporal dependencies in the EMG signal sequences.

Main Results:

• The GTCN achieved a classification accuracy of 97.24% on a public EMG dataset for hand gesture recognition.
• The model outperformed several existing state-of-the-art methods.
• Demonstrated effective exploitation of temporal and spatial characteristics, including muscle synergy.

Conclusions:

• The proposed GTCN is a highly effective method for EMG signal classification and hand gesture recognition.
• The model's ability to adaptively represent evolving spatial dependencies and temporal patterns is key to its success.
• GTCN offers a promising advancement for sophisticated EMG-based human-computer interaction.