FPGA-based Lightweight QDS-CNN System for sEMG Gesture and Force Level Recognition.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Embedded Systems

Background:

• Deep learning (DL) excels at electromyographic (EMG) signal recognition but faces implementation challenges in resource-limited prosthetics and human-computer interaction (HCI) devices.
• Efficient processing of complex EMG data is crucial for real-time control in wearable applications.

Purpose of the Study:

• To implement a low-power system for EMG gesture and force level recognition on a Zynq architecture.
• To develop a lightweight deep learning model and a dedicated hardware accelerator for efficient EMG signal processing.

Main Methods:

• Proposed a lightweight network using Ultra-lightweight depth separable convolution (UL-DSC) and channel attention-global average pooling (CA-GAP) to reduce computational load.
• Developed a compact wearable EMG acquisition device for real-time data collection.
• Designed a highly parallelized hardware accelerator for efficient inference computation with 8-bit precision.

Main Results:

• Achieved an average accuracy of 94.92% for recognizing 18 gestures and force levels from 22 healthy subjects.
• The lightweight model has only 5.0k parameters and a size of 0.026MB.
• The hardware accelerator demonstrated a single-frame inference time of 41.9μs, power consumption of 0.317W, and data throughput of 78.6 GOP/s.

Conclusions:

• The proposed low-power system effectively enables accurate EMG gesture and force level recognition for resource-constrained applications.
• The combination of a lightweight DL model and a dedicated hardware accelerator offers a viable solution for advanced prosthetic and HCI devices.
• The system's high accuracy and efficiency pave the way for more sophisticated and accessible wearable technology.