Real-Time Hand Rehabilitation: FPGA-Accelerated Neural Network for Muscle Activity Classification.

Area of Science:

• Biomedical Engineering
• Computer Engineering
• Artificial Intelligence

Background:

• Detecting hand movement intention from muscle activity is vital for assistive and rehabilitation devices.
• Current microcontroller-based systems may lack the real-time responsiveness required for seamless human-machine interaction.

Purpose of the Study:

• To develop and evaluate an FPGA-based neural network for classifying muscle activity indicative of hand movement intention.
• To demonstrate the feasibility of hardware acceleration for real-time biomedical signal processing.

Main Methods:

• A three-layer neural network was designed and implemented on an FPGA using High-Level Synthesis (HLS).
• The system processed muscle sensor data in real-time to classify muscle states as stressed (movement intention) or relaxed (no movement intention).
• The FPGA system was integrated into a wearable hand movement assistance prototype.

Main Results:

• The FPGA implementation achieved low latency and power-efficient operation, outperforming traditional microcontroller approaches.
• The hardware-accelerated neural network demonstrated efficient classification of muscle activity with a compact design.
• The system successfully triggered controlled hand motions in the prototype based on classified muscle states.

Conclusions:

• FPGA-based neural networks offer a promising, fast, and efficient alternative for real-time biomedical signal processing.
• The developed system shows significant potential for enhancing assistive and rehabilitation applications.
• Future work will focus on optimizing hardware resources and expanding applications in prosthetics and broader rehabilitation scenarios.