💻 computer science

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

This paper proposes an energy-efficient Spiking Neural Network framework with a novel quantization-compatible training scheme (SDH) that achieves robust, real-time muscle fatigue detection from sEMG signals with significantly lower energy consumption and superior noise resilience compared to existing deep learning models.

Original authors: Kaiwen Tang, Jiaqi Dong, Zhanglu Yan, Weng-Fai Wong

Published 2026-07-14

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Kaiwen Tang, Jiaqi Dong, Zhanglu Yan, Weng-Fai Wong

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine your muscles are like a marathon runner. At the start, they are fresh and fast, but as they keep going, they get tired, their signals get messy, and they start to stumble. Detecting exactly when that "tired" switch flips is crucial for athletes, workers, and anyone wearing a smartwatch that wants to keep you safe. But here's the catch: the devices we wear are tiny, battery-powered, and can't handle the heavy lifting of the super-smart computers we usually use to analyze these signals.

Enter a new team of digital detectives: Spiking Neural Networks (SNNs).

Think of a traditional computer brain (like the ones in your phone) as a noisy factory where every single machine is humming and working 24/7, even when there's nothing to do. It's loud, it's hot, and it eats up a lot of battery. In contrast, the SNNs proposed in this paper are like a group of whispering spies. They only "speak" (or fire a signal) when something important actually happens. If nothing is going on, they stay silent, saving massive amounts of energy.

The Problem with the Old Guard

The researchers tried using the usual suspects—standard deep learning models—to spot muscle fatigue. They found these models were too hungry for power and too picky about data. They also struggled when the signals got messy, which happens constantly when you move around (like when your arm swings and the sensor wiggles). The paper argues that these heavy, energy-guzzling models are simply not the right fit for the tiny, battery-powered gadgets we wear on our wrists.

The New Solution: A "Spiking" Detective

The team built a system that mimics how real neurons in your brain work. Instead of constantly churning numbers, these digital neurons wait for a specific threshold to be crossed before they "spike" (send a signal). This makes them incredibly efficient.

But there was a snag: these spiking networks can be a bit fragile when the data is noisy. To fix this, the authors invented a special training recipe called SDH. Imagine you are teaching a spy to recognize a friend in a crowd. If you only show them the friend in perfect lighting, they might fail when it rains. The SDH method is like training the spy in the rain, the wind, and the mud all at once. It uses a mix of three different "training drills" to make sure the network stays calm and accurate even when the muscle signals are jittery or distorted by movement.

The Results: Fast, Quiet, and Tough

The team tested their system on two public datasets (collections of muscle signals from real people doing exercises). Here is what they found:

Accuracy: Their spiking models were just as good, or even better, than the heavy-duty models. On one dataset, they hit an F1 score of 90.14%, beating the best non-spiking models.
Noise Resistance: When they threw seven different types of "noise" at the system (simulating everything from electrical interference to the jitters of a moving arm), the spiking models stayed steady. They didn't panic when the data got messy.
Energy Savings: This is the big one. Because these networks only work when they need to, and because the authors squeezed the data down to tiny 3-bit and 4-bit numbers (like compressing a high-res photo into a tiny icon), the energy savings were massive. They estimated the new models use up to 201.77 times less energy than the standard models.

To put that in perspective: if you were to run this on a typical smartwatch with a 4000 J battery, the paper's simulations suggest it could keep detecting muscle fatigue continuously for about 56.83 hours. That's more than two full days of non-stop monitoring on a single charge!

What They Didn't Do (And Why It Matters)

It's important to note what this paper didn't do. They didn't just say, "It works!" They explicitly tested against a wide range of competitors and found that older methods often failed under noisy conditions. They also didn't claim to have solved the problem for every single medical scenario; their results are based on simulations and tests on specific public datasets (SUE and SPE). They didn't test this on a live human patient in a hospital yet, but they did prove that the math works and the energy savings are real in their simulated environment.

The Bottom Line

This paper suggests that by switching from "always-on" computer brains to "whispering spy" brains, we can finally put powerful muscle-fatigue detectors on our wrists without killing the battery. It's a step toward wearable tech that is smart enough to know when you're tired, but quiet enough to let you keep moving.

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

1. Problem Statement

Detecting muscle fatigue via surface electromyography (sEMG) is critical for sports optimization, rehabilitation, and occupational safety. However, deploying reliable fatigue detection on wearable devices faces three primary challenges:

Computational Constraints: Existing deep learning models (e.g., LSTM, Transformers) often incur high latency and energy costs, making them unsuitable for power-constrained edge devices.
Data Limitations: Models must achieve high accuracy with limited training data, as collecting large-scale, labeled sEMG datasets is difficult.
Noise Robustness: sEMG signals are highly susceptible to noise from motion artifacts, electrode-skin impedance variations, and environmental factors, yet few existing methods explicitly address robustness to these perturbations in real-time, on-device scenarios.

2. Methodology

The authors propose a compact, quantized Spiking Neural Network (SNN) framework designed for real-time, on-device muscle fatigue detection. The pipeline consists of four main stages:

A. Feature Extraction

Raw sEMG signals are smoothed and divided into short-term windows (0.25s for SUE, 7.5s for SPE). From these, 11 fatigue-related features (frequency-domain, time-domain, spectral flux, entropy) are extracted. These are aggregated into overlapping mid-term windows (2s for SUE, 60s for SPE) to capture long-term trends, resulting in a final input vector of 132 dimensions.

B. SNN Construction and Training

ANN-to-SNN Conversion: To address the non-differentiable nature of spiking neurons, the authors first train a traditional Artificial Neural Network (ANN) with a four-layer fully connected architecture (132-128-32-1). This ANN is then converted into an SNN using a structure-preserving pipeline.
Encoding: Inputs are encoded into spike trains using an Average Spike Generation (ASG) variant of the Leaky Integrate-and-Fire (IF) model.
Robustness Training (SDH): A novel training scheme named SDH is introduced to enhance robustness against noise. It combines three regularization terms into a composite loss function:
1. SHAQ (Safe Haven Activation Quantization): Pushes activation values toward the midpoint of quantization intervals to reduce sensitivity to small perturbations.
2. DQ (Defensive Quantization): Penalizes large weights to limit noise propagation across layers and reduce feature correlation.
3. Huber Loss: Suppresses the influence of transient spikes and outliers by treating large residuals linearly rather than quadratically.
  The total loss is $L = L_{BCE} + \lambda_1 L_{act} + \lambda_2 L_{orth} + \lambda_3 L_{huber}$ .

C. Quantization and Post-Processing

Quantization: The trained models undergo post-training quantization (PTQ) to 3-bit (QSN3) and 4-bit (QSN4) weights, significantly reducing memory footprint without retraining.
QAF (Quick Adapt Filter): A post-processing module is applied to suppress short-term classification noise. It utilizes a sliding window, a median filter, and an adaptive threshold based on an Exponentially Weighted Moving Average (EMA). It also includes a "fast channel" mechanism to trigger immediate fatigue states upon strong signal detection.

D. Hardware Implementation

The authors designed a custom Compute Unit (CU) and scheduler for neuromorphic hardware. The design minimizes power by using a single CU for all computations, separating immutable parameters (weights) from activations in SRAM, and utilizing a local membrane-potential buffer to avoid repeated memory accesses.

3. Key Contributions

Energy-Efficient Paradigm: The paper establishes SNNs as a viable paradigm for muscle fatigue detection, demonstrating that spike-based models can reduce estimated energy consumption by up to 201.77× compared to full-precision ANN baselines.
Robust Training Scheme (SDH): The introduction of the SDH scheme, which integrates SHAQ, DQ, and Huber loss, significantly boosts model resilience under both synthetic noise and physically motivated perturbations (e.g., motion artifacts, baseline drift).
Comprehensive Evaluation: The framework is evaluated on two public sEMG datasets (SUE and SPE) against a broad set of baselines (including SVM, Random Forest, LSTM, and Transformers) under seven distinct noise conditions.

4. Experimental Results

Accuracy: On the SUE dataset, the proposed SNN achieved an F1 score of 87.93% (with QAF), outperforming the best non-spiking baseline (LSTM) which scored 87.36%. On the SPE dataset, the SNN reached 90.14%, surpassing the best baseline (MLP) at 87.00%.
Quantization Performance: The quantized models (QSN3 and QSN4) maintained performance comparable to their full-precision counterparts, with QSN4 achieving 87.84% (SUE) and 90.00% (SPE).
Noise Robustness: Under seven noise types (Gaussian, Poisson, Uniform, Multiplicative Gain, Motion Artifacts, Power-line, and Drift), the proposed models remained consistently stable. Notably, the SDH-trained models outperformed variants trained without robustness enhancements or with only SHAQ/DQ.
Energy Efficiency:
- The estimated energy consumption was reduced by 1.89×–5.13× for standard SNNs and 52.07×–201.77× for quantized SNNs compared to the baseline MLP.
- Simulations on a low-power reconfigurable ASIC neuromorphic system estimated a continuous operation time of 56.83 hours on a typical smartwatch battery (4000 J), demonstrating suitability for long-term wearable deployment.

5. Significance and Claims

The authors claim that this work presents the first energy-efficient, edge-device-compatible SNN solution specifically tailored for sEMG-based muscle fatigue detection with enhanced noise resilience. The significance of the work lies in:

Feasibility for Wearables: By drastically reducing energy consumption and computational overhead, the framework enables real-time, privacy-preserving fatigue monitoring on resource-constrained devices without relying on cloud computation.
Robustness in Unconstrained Environments: The SDH training scheme addresses the critical gap of noise susceptibility in real-world sEMG acquisition, ensuring reliability even with motion artifacts and signal variability.
Practical Deployment: The combination of aggressive quantization (3/4-bit) and specialized post-processing (QAF) bridges the gap between theoretical SNN efficiency and practical, high-accuracy deployment in sports and occupational safety contexts.

The paper concludes that this approach lays a foundation for practical wearable systems capable of real-time muscle fatigue monitoring, balancing high predictive accuracy with the strict energy and robustness requirements of edge computing.