Robust Real-Time Embedded EMG Recognition Framework Using Temporal Convolutional Networks on a Multicore IoT Processor | JoVE Visualize

Area of Science:

Biomedical Engineering
Machine Learning
Wearable Technology

Background:

Surface electromyographic (sEMG) signals are crucial for advanced Human-Computer Interaction (HCI) but face challenges in long-term reliability due to signal variability.
Existing embedded sEMG solutions suffer from accuracy degradation over time, limiting their use in reliable gesture controllers.
Developing robust and accurate embedded systems for sEMG-based gesture recognition is essential for practical HCI applications.

Purpose of the Study:

To develop a wearable-class embedded system for robust sEMG-based gesture recognition using Temporal Convolutional Networks (TCNs).
To introduce a novel TCN topology, TEMPONet, and evaluate its performance against state-of-the-art methods.
To design an energy-efficient embedded platform for real-world deployment and assess its performance and memory constraints.

Main Methods:

A novel Temporal Convolutional Network (TCN) topology, TEMPONet, was developed for sEMG hand movement classification.
The TEMPONet was initially tested on the benchmark Ninapro dataset, achieving state-of-the-art accuracy.
An energy-efficient embedded platform based on the GAP8 processor was designed, and a new dataset was collected for validation in a deployment-representative setup. 8-bit quantization was employed to optimize the network for the embedded platform's memory constraints.

Main Results:

TEMPONet achieved 49.6% average accuracy on the Ninapro dataset, outperforming current state-of-the-art by 7.8%.
On the custom 20-session dataset using the GAP8 platform, the TCN achieved 93.7% average accuracy, comparable to a state-of-the-art Support Vector Machine (SVM) approach (91.1%).
An 8-bit quantization strategy reduced the memory footprint by 4× (to 460 kB) with only a 3% accuracy degradation. The quantized network classifies gestures in 12.84 ms with 0.9 mJ power consumption.

Conclusions:

The proposed TEMPONet and GAP8-based embedded system offer a robust and energy-efficient solution for real-time sEMG gesture recognition.
The system demonstrates high accuracy and low power consumption, making it suitable for long-lifetime wearable applications.
The developed system addresses the critical challenge of long-term reliability in sEMG-based control, paving the way for advanced HCI devices.