TrustEMG-Net: Robust sEMG Denoising Framework

• The paper’s main contribution is integrating U-Net with a representation-masking Transformer to effectively suppress noise in sEMG signals.
• It employs a synergistic design where local convolutional feature extraction is combined with global self-attention, validated by improvements in SNR and RMSE.
• The framework outperforms traditional denoising methods, enabling robust applications in clinical monitoring, prosthetics, and human-computer interaction.

TrustEMG-Net is a neural network-based framework for single-channel surface electromyography (sEMG) denoising, employing a synergistic architecture that combines U-Net and a representation-masking Transformer encoder. Designed specifically to enhance robustness and generalizability across diverse noise sources and signal-to-noise ratio (SNR) conditions, TrustEMG-Net advances the state of sEMG enhancement by leveraging nonlinear representation learning in a denoising autoencoder structure (Wang et al., 2024).

1. Architectural Design: U-Net and Representation-Masking Transformer

TrustEMG-Net integrates the local feature extraction capacity of U-Net with global context modeling via a Transformer encoder, connected through a learnable feature-space mask. The network is configured for end-to-end single-channel waveform denoising.

• U-Net Encoder: Accepts a noisy waveform $\tilde{x} \in \mathbb{R}^d$ and processes it through five 1D convolutional layers (kernel size 8, ReLU activations). Stride-1 at the first layer yields $d \times 64$ features, with subsequent layers doubling feature channels and halving time resolution, ending at $\tfrac{d}{16} \times 1024$. Features from each encoding stage are preserved for skip connections.
• Transformer Bottleneck: The encoded tensor $X \in \mathbb{R}^{\tfrac{d}{16} \times 1024}$ receives sinusoidal positional encoding before passing through a single Transformer encoder layer (multi-head self-attention, 8 heads, $D=1024$). The output $f(X)$ is mapped to $[0,1]$ via a sigmoid, producing a learnable representation mask $M(X)$. The mask modulates $X$ via elementwise product: $\hat X = M(X) \odot X$.
• Intuition: Channels with $d \times 64$0 close to 1 preserve sEMG information, while those near 0 are suppressed as likely noise.
• U-Net Decoder: Four upsampling modules (transpose convolution and convolutional layer, with batch norm and ReLU) reconstruct temporal resolution; skip connections inject preserved features. A final transposed convolution restores the signal to $d \times 64$1.

The end-to-end process begins with a 2-second noisy sEMG segment, transforms its latent space through U-Net encoding, globally contextualizes features with the masked Transformer, and reconstructs a denoised waveform using the decoder with multi-scale skip connections.

2. Mathematical Formulation

The U-Net encoder's per-layer operation for the $d \times 64$2-th block is: $d \times 64$3 where $d \times 64$4 denotes 1D convolution, and $d \times 64$5, $d \times 64$6 are the layer-specific weights and biases.

Self-attention in the Transformer is computed as: $d \times 64$7 with multi-head formulation: $d \times 64$8

The representation-masking operation is: $d \times 64$9 where $\tfrac{d}{16} \times 1024$0 denotes the elementwise sigmoid and $\tfrac{d}{16} \times 1024$1 the Transformer mapping.

The network is supervised via an L1 loss between ground-truth $\tfrac{d}{16} \times 1024$2 and denoised output $\tfrac{d}{16} \times 1024$3: $\tfrac{d}{16} \times 1024$4 No adversarial or perceptual losses are incorporated.

3. Data Preparation and Training Protocol

• Signal Source: Clean sEMG is sourced from the Ninapro DB2 database (40 subjects, 12 channels, Exercises 1–2), undergoing bandpass filtering (20–500 Hz, 4th-order Butterworth), 1 kHz downsampling, normalization, and 2-second window segmentation (non-silent data only).
• Noise Contaminants: Five types are synthesized: baseline wander (MIT-BIH NSTDB), powerline interference (58.5–61.5 Hz drifting sine), ECG (MIT-BIH NSR), motion artifacts (NSTDB and tapping datasets), and white Gaussian noise. Mixtures of three or all five contaminants are created, each with equal energy contribution.
• SNR Conditions: Training SNRs are $\tfrac{d}{16} \times 1024$5 dB; testing SNRs are $\tfrac{d}{16} \times 1024$6 dB.
• Optimization: The Adam optimizer is used with a decaying learning rate (0.01 for 3 epochs, 0.001 through 30 epochs, then 0.0001) and a batch size of 256. Early stopping is triggered if validation loss stagnates for 15 epochs. Evaluation uses four-fold subject-wise cross-validation (10 subjects per fold).

4. Performance Evaluation and Comparative Results

TrustEMG-Net is benchmarked against five classical denoising algorithms (IIR, TS+IIR, EMD, CEEMDAN, VMD) and four neural baselines (CNN, FCN, U-Net, and TrustEMG-Net without masking). Evaluation addresses both reconstruction fidelity and feature preservation using five metrics:

• SNR Improvement ($\tfrac{d}{16} \times 1024$7): Gain between output and input SNRs.
• RMSE: Root mean squared error between target and output.
• Percentage RMS Difference (PRD)
• ARV-RMSE: Root mean square error of the average rectified value over 200 ms windows.
• MF-RMSE: RMSE for mean frequency across 10–500 Hz as derived from the STFT.

Method	SNR_imp (dB)	RMSE ($\tfrac{d}{16} \times 1024$8)	PRD (%)	ARV-RMSE ($\tfrac{d}{16} \times 1024$9)	MF-RMSE (Hz)
TrustEMG-Net	13.64	2.18	48.44	8.72	16.63
CEEMDAN (next best)	10.68	2.95	67.00	11.08	32.14

Across all SNRs (–14 to 2 dB) and all contaminant types, TrustEMG-Net outperforms baselines, particularly for powerline interference (21.45 dB SNR improvement versus CEEMDAN’s 17.22 dB). On 3-/5-type contaminant mixtures, TrustEMG-Net consistently remains superior.

5. Ablation Analysis and Representation Masking Effects

Ablation studies assess the contribution of U-Net, Transformer direct mapping (DM), and representation masking (RM):

• U-Net alone surpasses CNN/FCN by 10–20% on SNR_imp and RMSE, underscoring the importance of skip connections.
• Inclusion of Transformer (DM) provides additional 2–3 dB SNR_imp over pure U-Net, capturing contextual global dependencies.
• RM Layer offers further gains, particularly for narrowband interference (e.g., PLI: DM = 20.90 dB SNR_imp, RM = 21.45 dB; 5-type mixture: DM = 12.91, RM = 13.09 dB).
• Mask analyses reveal the 1024-dimensional mask selectively suppresses features where interference is present, preserving those tied to sEMG; manual manipulation of mask weights recapitulates or distorts specific signal/noise elements, validating the mask’s interpretability.

6. Computational Considerations and Applications

• Efficiency: TrustEMG-Net comprises 25.1 million parameters, requires 2.84 GFLOPs per run, and achieves inference latency of ~3.5 ms on an RTX 3090 GPU (~13 ms on CPU). Execution time is lower than EMD/CEEMDAN/VMD.
• Applications:
  • Clinical gait and muscle-fatigue monitoring (leveraging ARV/MF indices in rehabilitation).
  • Prosthetics and human-computer interaction, where denoised sEMG improves gesture recognition or force estimation.
  • Speech prostheses and silent-speech interfaces requiring noise-robust processing.
  • Multimodal biosignal denoising where sEMG is accompanied by ECG, EEG, or audio.

TrustEMG-Net’s combination of local and global modeling, via U-Net and representation-masking Transformer, establishes a robust and generalizable paradigm for sEMG denoising under a broad array of noise sources and SNR scenarios (Wang et al., 2024).