Deep Scattering Transform with Attention Mechanisms Improves EMG-based Hand Gesture Recognition.

Area of Science:

• Biomedical Engineering
• Signal Processing
• Machine Learning

Background:

• Electromyogram (EMG) signals are crucial for understanding hand muscle activity but are complex due to noise and non-stationarity.
• Accurate analysis of EMG signals is vital for advanced prosthetic control and human-computer interfaces.
• Existing methods face challenges in robustly classifying EMG patterns amidst signal variability.

Purpose of the Study:

• To develop and validate a novel approach for classifying EMG patterns using wavelet scattering transform (WST) and deep learning attention mechanisms.
• To investigate the effectiveness of WST and attention mechanisms in handling the stochastic and non-stationary nature of EMG signals.
• To improve the accuracy of myoelectric pattern recognition (PR) for hand movements.

Main Methods:

• Utilized wavelet scattering transform (WST) for signal decomposition and robust feature extraction from EMG data.
• Integrated attention mechanisms from deep neural networks to focus on salient features and muscle activation correlations.
• Applied the combined WST and attention approach to three standard EMG datasets from laboratory and wearable devices.

Main Results:

• Achieved significant improvements in myoelectric pattern recognition (PR) accuracy compared to existing methods.
• Demonstrated average classification accuracies of up to 98% across different EMG datasets.
• Validated the hypothesis that focusing on inter-muscle activation correlations enhances EMG classification.

Conclusions:

• The combination of WST and attention mechanisms offers a powerful and accurate method for EMG signal classification.
• This novel approach enhances the robustness and reliability of myoelectric pattern recognition.
• The findings have significant implications for developing more sophisticated EMG-based control systems.