Deep Residual U-Net Autoencoder with Weighted Overlapping Reconstruction for EMG Signal Denoising

Electromyography (EMG) signals, crucial for neuromuscular assessment, are frequently corrupted by noise, impairing signal fidelity and subsequent analysis across diverse applications. Conventional filters often inadequately address non-stationary noise or introduce signal distortion. This paper introduces an advanced deep learning framework for EMG denoising, centred on a U-Net-inspired convolutional autoencoder with integrated residual blocks and skip connections. Training utilised synthetic EMG data, closely emulating physiological frequency bands and burst dynamics, subsequently corrupted by a comprehensive noise model encompassing electrode, crosstalk, electronic, drift, and contact artefacts. Training was guided by a custom loss function that combined weighted mean squared error (MSE) with signal-to-noise ratio (SNR). The proposed autoencoder achieved substantial improvements, SNR increased from -0.95 dB (noisy) to 14.64 dB (denoised), and MSE was drastically reduced from 0.001493 V² to 0.000041 V² on the test dataset. Qualitative analysis confirmed effective noise suppression while retaining crucial EMG burst characteristics. This advanced framework offers a promising solution for robust restoration of EMG signals in practical settings.