TrustEMG-Net: Using Representation-Masking Transformer With U-Net for Surface Electromyography Enhancement

Area of Science:

• Biomedical Engineering
• Signal Processing
• Machine Learning

Background:

• Surface electromyography (sEMG) is crucial for muscle activity analysis but susceptible to contaminants.
• Existing denoising methods are often heuristic-based, lack robustness, and are contaminant-specific.

Purpose of the Study:

• To develop a potent, robust, and generalized sEMG denoising approach.
• To introduce TrustEMG-Net, a novel neural network-based method for sEMG signal cleaning.

Main Methods:

• Utilized a denoising autoencoder structure combining U-Net with a Transformer encoder.
• Employed a representation-masking approach for enhanced feature learning.
• Evaluated TrustEMG-Net on the Ninapro sEMG database with diverse contamination types and SNR levels.

Main Results:

• TrustEMG-Net demonstrated exceptional performance across five evaluation metrics, exceeding existing methods by at least 20%.
• The method showed consistent superiority across varying Signal-to-Noise Ratios (SNRs) from -14 to 2 dB and five contaminant types.
• An ablation study confirmed the effectiveness of TrustEMG-Net's architectural components.

Conclusions:

• TrustEMG-Net provides a highly effective, robust, and generalized solution for sEMG denoising.
• The proposed neural network approach significantly enhances sEMG signal quality for various applications.
• This advancement supports improved accuracy in healthcare and human-computer interaction systems utilizing sEMG data.