Surface electromyography (sEMG) fatigue estimation has been done through hand-crafted features and traditional machine learning so far. Determine whether end-to-end deep learning directly on sEMG signals has a superior cross-subject baseline and find an optimal model configuration. A well-prepared set of 13 individuals and Leave-One-Subject-Out (LOSO) cross-validation will be used for systematic empirical investigation. In the first round, we will compare XGBoost (a classic machine learning model) with multiple neural network architectures in terms of 2-second and 4-second observation windows: 1D CNN, multi-scale CNN, and LSTM-augmented CNNs. CNN-based models outperform XGBoost in all of the above indices. A single Raw 1D CNN performs better on 2s windows (Macro F1: 0.4798), but a Multi-scale CNN can capture extended patterns in 4s windows more effectively (Macro F1: 0.4786). Adding Recurrent Layers (LSTM) degrades cross-subject generalization. Given the excellent performance of CNN backbones, the second stage will study different input pre-processing techniques, such as rectified signals, envelope integrations and time-frequency representations (STFT, CWT). Counterintuitively, the raw sEMG signals perform better than all explicit preprocessing methods. Rectification and envelope extraction are forms of intervention that reduce the time-domain feature richness and lose some high-frequency MUAP (Motor Unit Action Potential) information. At the same time, time-frequency representations face problems of architectural mismatch and computational bottleneck during full LOSO validation. Based on the above results, we believe that a raw-signal 1D CNN or Multi-scale CNN employing unmodified sEMG windows can be used to perform cross-subject fatigue detection more effectively and simply.

Jun Ye. Deep Learning for Cross-Subject sEMG Fatigue Classification: Architectures, Window Lengths, and Input Representations. International Journal of Big Data Intelligent Technology (2026), Vol. 7, Issue 1: 163-169. https://doi.org/10.38007/IJBDIT.2026.070119.

[1] Chowdhury, R.H. et al. (2013). "Surface electromyography signal processing and classification techniques." Sensors.

[2] Jaramillo-Yánez, A. et al. (2020). "Real-Time Hand Gesture Recognition Using Surface Electromyography and Machine Learning: A Systematic Literature Review." Sensors.

[3] Li, W. et al. (2021). "Gesture Recognition Using Surface Electromyography and Deep Learning for Prostheses Hand: State-of-the-Art, Challenges, and Future." Frontiers in Neuroscience.

[4] Sul, J.H. et al. (2025). "Electromyography Signal Acquisition, Filtering, and Data Analysis for Exoskeleton Development." Sensors.

[5] Lira et al. (2024). "A Comprehensive Dataset of Surface Electromyography and Self-Perceived Fatigue Levels for Muscle Fatigue Analysis." Sensors (Dataset DOI: 10.3390/s24248081).