FatigueFormer: Static-Temporal Feature Fusion for Robust sEMG-Based Muscle Fatigue Recognition

Tong Zhang; Hong Guo; Shuangzhou Yan; Dongkai Weng; Jian Wang; Hongxin Zhang

cs.LGMar 27, 2026

FatigueFormer: Static-Temporal Feature Fusion for Robust sEMG-Based Muscle Fatigue Recognition

Authors: Tong Zhang, Hong Guo, Shuangzhou Yan, Dongkai Weng, Jian Wang, Hongxin Zhang

Organizations: Zhejiang University

Abstract

We present FatigueFormer, a semi-end-to-end framework that deliberately combines saliency-guided feature separation with deep temporal modeling to learn interpretable and generalizable muscle fatigue dynamics from surface electromyography (sEMG). Unlike prior approaches that struggle to maintain robustness across varying Maximum Voluntary Contraction (MVC) levels due to signal variability and low SNR, FatigueFormer employs parallel Transformer-based sequence encoders to separately capture static and temporal feature dynamics, fusing their complementary representations to improve performance stability across low- and high-MVC conditions. Evaluated on a self-collected dataset spanning 30 participants across four MVC levels (20-80%), it achieves state-of-the-art accuracy and strong generalization under mild-fatigue conditions. Beyond performance, FatigueFormer enables attention-based visualization of fatigue dynamics, revealing how feature groups and time windows contribute differently across varying MVC levels, offering interpretable insight into fatigue progression.

Figures & tables

Figure 1: Overview of sEMG characteristics and fatigue dynamics. (a) physiological sEMG and noise; (b) raw sEMG across MVC levels; (c) fatigue-sensitive features (e.g., MF, DET) decreasing with fatigue.

Method	Task	Input Type	Participants	Use Transformer
(Zabihi et al., 2023) [ 63 ]	Hand gesture	Raw sEMG	40	✓
(Dang et al, 2023) [ 13 ]	Fatigue Detection	Raw sEMG	20
(Rani et al, 2024) [ 51 ]	Hand gesture	Feature Map	8 + Not Reported + 6
(Putro et al., 2024) [ 50 ]	Finger joint est.	Feature Map	5	✓
(Eddy et al, 2024) [ 17 ]	Hand gesture	Raw sEMG	612
(Mehlman et al, 2025) [ 43 ]	Typing	Raw sEMG	108	✓

Table 1: Representative deep learning methods for sEMG-based tasks.

Figure 2: Overview of the proposed semi–end-to-end framework. Left: the sEMG feature extraction engine produces increasing- and decreasing-type descriptors, which are fed into the temporal and static encoders in parallel. Right: both encoders adopt a Transformer-based sequence module with similar structure but independent parameters. Their fused representations are finally used for fatigue recognition.

Figure 3: The feature tokenizer maps an

n

-dimensional statistical feature vector into an

(n{+}1)\!\times\!D

embedding sequence.

MVC Level	Relaxed	Exerted	Fatigued	Total
20%	6313	9867	16417	32597
40%	1475	3331	4318	9124
60%	775	1408	1955	4138
80%	402	813	1603	2818

Table 2: Sample distribution across fatigue levels under different MVC conditions in static contraction experiments.

MVC (%)	Method	Precision	Recall	F1-score
20	SVM	80.94	81.13	80.96
	CNN-SVM	77.79	77.98	77.64
	CLT	85.87	85.60	85.71
	AutoInt	85.86	85.47	85.59
	Ours	97.85±0.03	97.85±0.03	97.85±0.03
40	SVM	83.64	83.68	83.61

Table 3: Performance comparison under different MVC levels. Each cell reports mean precision, recall, and F1-score (%). “Ours” additionally reports standard deviation across five runs. Best results are highlighted in bold .

Model Variant	MVC Level (%)
Model Variant	20	40	60	80
w/o Temporal Encoder	90.87	94.45	96.03	97.01
w/o Static Encoder	97.35	97.45	94.44	96.67
w/o Cross Attention	96.69	97.44	97.26	97.78
w/o Residual Path (MLP)	97.32	97.45	96.76	97.78
w/o Multi-scale Loss	97.75	97.56	97.26	95.92

Table 4: Ablation study of key components across different MVC levels. Each value reports the average F1-score (%) over five runs. Removing specific modules or supervision terms leads to performance degradation under different contraction levels, highlighting their complementary effects.

Model	Precision (%)	Recall (%)	F1-score (%)
AutoInt	53.32	52.69	52.01
CLT	56.24	53.58	53.47
Ours	61.15	61.08	60.41

Table 5: Cross-MVC + L5SO generalization performance. Each value reports the mean over all folds. Our semi end-to-end framework achieves the best generalization across unseen subjects.

Panel of Figure 4 — Figure 4: Temporal attention visualization across and within MVC levels. (a) Cross-MVC heatmap illustrates the evolution of attention intensity across time windows (W1–W5) under different contraction levels (20–80% MVC). (b–c) Representative attention distributions at low (20%) and high (80%) MVC show how the model reallocates its temporal focus across fatigue states ( relaxed , exerted , fatigued ).

Panel of Figure 5 — Figure 5: Static attention maps across MVC levels. (a) Self-attention maps show feature-group dependencies between increasing (INC) and decreasing (DEC) signals across four MVC levels (20%, 40%, 60%, 80%). (b) Cross-attention maps illustrate inter-group interactions ( DEC $\rightarrow$ INC and INC $\rightarrow$ DEC ) under the same MVC ordering.

Figure 6: Illustration of the experimental setup during sEMG data acquisition. The photo shows the participant’s forearm and the attached load used to control contraction intensity. For privacy protection, no identifiable information is visible, and the surface electrodes placed on the upper arm are not shown in this illustrative view. All photos were taken with informed consent.

RPE Score	Perceived Sensation	Fatigue State
6–7	Very, very light / Effortless	Relaxed
8–10	Very light / Easy	Relaxed
11	Fairly light	Exerted
12–14	Somewhat hard	Exerted
15	Hard	Exerted
16–18	Very hard	Fatigued

Table 6: RPE score–to–fatigue state mapping used in the experiment.

Category	Increasing Features	Decreasing Features
Time-domain features	AEMG, iEMG, RMS, MAV, MCV, DASDV	ZC, SSC, WA
Frequency-domain features	SMR, FSM2, TP	MPF, MF, MDF, IMPF, IMF, BSE
Time–frequency features	ERHL, IMNF, IMFB
Wavelet-based features	WIRM1551, WIRM1522, WIRE51, WIRW51, WEE
Nonlinear features	DET, ACC	AE, SE, LZC, FD, BE, WENT

Table 7: Categories of extracted sEMG features and their trend-based grouping (increasing vs. decreasing) based on statistical significance analysis.

Figure 7: Temporal evolution of eight representative sEMG features across fatigue progression. Solid curves show subject-averaged trajectories, and dashed lines denote linear regression fits. Features with positive slopes are categorized as increasing-type , whereas those with negative slopes are labeled decreasing-type . These empirical patterns validate the trend-based feature grouping used in the main paper.

Platform	Model	Arch.	Cores (L/P)	Freq.	L3	Best Proc/Time (s)	Speedup
Server	Xeon 8470Q	x86_64	208/104	2.0–3.8	97.5	24 / 23.68	43.88 $\times$
Server	Xeon 6430	x86_64	128/64	2.1–3.5	60	20 / 61.60	16.87 $\times$
Desktop	i7-11800H	x86_64	16/8	2.3–4.6	24	16 / 116.3	8.93 $\times$
Desktop	Apple M1	ARM64	8/8	3.2	12	8 / 85.41	12.16 $\times$

Table 8: Runtime performance of the Python-based feature extraction engine.

Model Variant	MVC Level (%)				Total parameters
Model Variant	20	40	60	80	Total parameters
Embedding dim = 128	96.30	97.95	98.13	96.50	2,825,353
Embedding dim = 512	98.26	98.12	99.01	97.39	44,331,529
Temporal window = 3	94.50	96.25	96.70	96.40	11,155,209
Temporal window = 7	98.72	98.11	98.48	98.12	11,156,233
Full model (D = 256, T = 5)	97.85	97.98	98.66	98.60	11,155,721

Table 9: Ablation results on embedding dimension

D

and temporal window length

T

. We report F1-scores (%) across MVC levels and total parameter counts, illustrating the trade-off between model capacity and efficiency.

Model	Class	MVC Level (%)
Model	Class	20	40	60	80
CLT	Relaxed	90.00	89.00	84.00	78.00
	Exerted	77.00	82.00	82.00	66.00
	Fatigued	89.00	90.00	92.00	86.00
AutoInt	Relaxed	87.44	87.91	95.52	81.25
	Exerted	78.88	86.87	92.31	86.75

Table 10: Per-class F1-score (%) across MVC levels for three methods.

Explore similar work

Efficient and Robust Spiking Neural Networks for sEMG-Based Muscle Fatigue Detection
Jul 13, 2026Kaiwen Tang, Jiaqi Dong, Zhanglu Yan +1Energy-Efficient ML Spiking Neural Networks
University of Singapore
Cross-subject Muscle Fatigue Detection via Adversarial and Supervised Contrastive Learning with Inception-Attention Network
Apr 3, 2026Zitao Lin, Chang Zhu, Wei Meng
School of Information Engineering, Wuhan University of Technology, Wuhan 430070, China
Estimating Human Muscular Fatigue in Dynamic Collaborative Robotic Tasks with Learning-Based Models
Feb 17, 2026Feras Kiki, Pouya P. Niaz, Alireza Madani +1Human-Robot Collaboration
F. Kiki, P. P. Niaz, A. Madani, and C. Basdogan (corresponding author) are with the Robotics and Mechatronics Laboratory (RML) and the KUIS AI Center, Koc University, Sariyer, Istanbul 34450, Turkey.

Jul 13, 2026J/K move · Enter open · S save

Efficient and Robust Spiking Neural Networks for sEMG-Based Muscle Fatigue Detection

Kaiwen Tang, Jiaqi Dong, Zhanglu Yan, Weng-Fai Wong

University of Singapore

Energy-Efficient ML Spiking Neural Networks Noise Robustness Neural Network Quantization

Open paper AskPDF arXiv

Detecting muscle fatigue via surface electromyography (sEMG) is essential for applications in sports, rehabilitation, and wearable health monitoring. Accurate and timely detection of fatigue is crucial for preventing injuries, optimizing physical performance, and ensuring user safety during prolonged activity. However, existing deep learning models are often unsuitable for this task due to their high computational cost and dependence on large-scale data. In this work, we propose an energy-efficient framework for muscle fatigue detection based on Spiking Neural Networks (SNNs), which exploit sparse, event-driven computation and temporal modeling. We further introduce a quantization-compatible training scheme (SDH) that combines multiple regularization terms to improve robustness under noisy conditions. Evaluated on two public sEMG datasets against a broad set of baselines and under seven noise conditions including physically motivated perturbations, our quantized SNNs match or exceed strong baselines while remaining more stable under diverse noise and reducing estimated energy consumption by up to 201.77x. These results demonstrate the framework's strong potential for real-time deployment in low-power wearable systems.

Abstract

Explore similar work

Efficient and Robust Spiking Neural Networks for sEMG-Based Muscle Fatigue Detection

Cross-subject Muscle Fatigue Detection via Adversarial and Supervised Contrastive Learning with Inception-Attention Network

Estimating Human Muscular Fatigue in Dynamic Collaborative Robotic Tasks with Learning-Based Models