A Novel Event-Driven Spiking Convolutional Neural Network for Electromyography Pattern Recognition.

Area of Science:

• Biomedical Engineering
• Neuroscience
• Artificial Intelligence

Background:

• Electromyography (EMG) pattern recognition is crucial for prosthetics and human-computer interaction.
• Practical applications are limited by electrode shift, poor accuracy, and extensive training data requirements.
• Spiking neural networks (SNNs) offer potential for improved robustness and reduced training burden.

Purpose of the Study:

• To propose and implement a spiking convolutional neural network (SCNN) for EMG pattern recognition.
• To evaluate the SCNN's performance against traditional methods in terms of accuracy, robustness to electrode shift, and training data requirements.
• To assess the SCNN's power consumption compared to conventional convolutional neural networks (CNNs).

Main Methods:

• Developed a SCNN integrating cyclic convolutional neural network (CNN) and fully connected modules.
• Utilized high-density surface electromyography (HD-sEMG) signals from 10 subjects performing 6 gestures.
• Conducted experiments focusing on small sample training and electrode shift scenarios.
• Compared SCNN performance against CNN, CNN-LSTM, LDA, and Spiking MLP.

Main Results:

• SCNN achieved higher accuracy than CNN, CNN-LSTM, LDA, and Spiking MLP in small sample training and electrode shift experiments.
• Demonstrated significant improvements in accuracy, particularly in mitigating electrode shift effects.
• Reported a substantial reduction in power consumption, approximately 1/93 of that of a standard CNN.

Conclusions:

• The proposed SCNN framework effectively addresses challenges in EMG pattern recognition, including electrode shift and training burden.
• SCNN offers a promising solution for developing user-friendly, real-time myoelectric control systems with reduced power demands.
• The findings highlight the potential of SNNs in advancing wearable technology and assistive devices.