Locomo-Net: A Low -Complex Deep Learning Framework for sEMG-Based Hand Movement Recognition for Prosthetic Control.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Rehabilitation Technology

Background:

• Deep learning for myoelectric pattern recognition is computationally intensive and memory-demanding.
• Surface electromyography (sEMG) based movement recognition is crucial for advanced prosthetics.

Purpose of the Study:

• To introduce LoCoMo-Net, a novel deep learning framework for low-complexity sEMG pattern recognition.
• To enable accurate recognition of various movements and force patterns using a single sEMG channel.

Main Methods:

• Developed a two-stage deep learning pipeline: input data compression and data-driven weight sharing.
• Validated the LoCoMo-Net framework on two datasets (DS1 and NinaPro) for diverse movement recognition.
• Prototyped LoCoMo-Net on a Xilinx FPGA for real-time execution feasibility.

Main Results:

• LoCoMo-Net achieved higher classification accuracy than Twin-SVM (8.5% increase) and existing CNN models (16.0% increase).
• Demonstrated significant hardware savings: 27% in LUTs, 49% in registers, 50% in memory, 23% in power, and 43% in computation time.
• Successfully validated real-time execution for 15 movements on an FPGA platform.

Conclusions:

• LoCoMo-Net provides an accurate and computationally efficient solution for sEMG-based movement recognition.
• The proposed system has significant clinical potential to improve the quality of life for amputees.
• This low-complexity framework facilitates the development of advanced, real-time myoelectric control systems.