Novel Wearable HD-EMG Sensor With Shift-Robust Gesture Recognition Using Deep Learning.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Wearable Technology

Background:

• Myoelectric control for hand gesture recognition faces challenges with robustness to confounding factors like electrode movement and forearm orientation.
• Existing systems often lack adaptability to variations in limb size and inter-session data consistency.

Purpose of the Study:

• To develop a hardware-software solution enhancing the robustness of myoelectric hand gesture recognition.
• To address confounding factors in myoelectric control through novel sensor design and advanced machine learning algorithms.

Main Methods:

• Development of EMaGer, a novel, full-circumference, flexible, 64-channel high-density electromyography (HD-EMG) sensor.
• Implementation of an array barrel-shifting data augmentation (ABSDA) approach with a convolutional neural network (CNN).
• Utilized an anti-aliased CNN (AA-CNN) for improved shift invariance and robustness to electrode displacement.

Main Results:

• The ABSDA-CNN method demonstrated an average improvement of 25.67% in inter-session accuracy for 6 gesture classes compared to conventional CNNs.
• The AA-CNN achieved accuracy improvements of up to 63.05% over non-augmented methods with electrode displacements of ±45°.
• The EMaGer sensor's unique design was crucial in enabling these performance benefits.

Conclusions:

• Co-designing sensor systems, processing methods, and inference algorithms offers synergistic benefits for state-of-the-art challenges.
• The proposed hardware-software solution significantly enhances the robustness and accuracy of myoelectric hand gesture recognition.
• This approach holds promise for more reliable prosthetic control and human-computer interfaces.