Abstract
Mental fatigue is a cumulative phenomenon that significantly impairs physiological activities, posing risks in ubiquitous edge computing environments like the Internet of Things. This study investigates fatigue induction through two levels of Stroop tasks and evaluates the resulting electroencephalogram (EEG) signals. Using wavelet packet decomposition and sample entropy, the research identifies key physiological markers. Results indicate that fatigue states exhibit higher $\alpha$ and $\theta$ wave relative energy and significantly lower $\beta$ wave energy. Notably, low-difficulty tasks induced more pronounced fatigue markers than high-difficulty tasks, suggesting that task engagement may inversely correlate with fatigue onset.
Introduction
Mental fatigue is a critical factor in high-stakes environments, including long-haul transportation and industrial monitoring. It manifests as a gradual decline in cognitive performance, leading to slower response times and coordination errors. To develop effective "ubiquitous edge computing" systems—such as real-time drowsiness detection in smart vehicles—objective physiological markers are essential. While subjective reports are common, they are often unreliable. This paper seeks to identify robust EEG features that signal the transition from alertness to fatigue under varying cognitive loads.
Methodology
The study employed a multi-faceted approach to analyze EEG data collected during cognitive performance tests.
Task Design
Participants performed two variations of the Stroop task:
- Low-Difficulty: Consistent colors and fonts (e.g., the word "Red" in red ink).
- High-Difficulty: Contradictory colors and fonts (e.g., the word "Red" in green ink), requiring higher cognitive interference resolution.
Signal Processing
- Wavelet Packet Decomposition (WPD): Used to decompose non-stationary EEG signals into specific frequency bands: $\alpha$ (8–13Hz), $\beta$ (14–30Hz), $\theta$ (4–8Hz), and $\delta$ (0.5–4Hz).
- Sample Entropy (SampEn): A non-linear algorithm used to measure signal complexity. High entropy reflects an alert, complex brain state, while lower entropy indicates the more rhythmic, predictable patterns associated with fatigue.
- Feature Extraction: The researchers focused on energy ratios, specifically $F = \alpha/\beta$ and $R = (\alpha + \theta)/\beta$, as primary fatigue indices.
Datasets and Experimental Setup
The study involved 10 healthy volunteers (6 male, 4 female) aged 19–25. EEG recordings were captured from 10 specific channels (FP1, FP2, F7, F8, T7, T8, C3, C4, O1, O2) based on the international 10-20 system.
Subjective fatigue scales and behavioral metrics (accuracy and reaction time) were recorded every five minutes to provide ground-truth validation for the physiological data.
Key Insights and Results
The Boredom Paradox
One of the most significant findings was that low-difficulty tasks induced fatigue more rapidly and noticeably than high-difficulty tasks. The physiological shifts in EEG were more pronounced in the "easy" group, suggesting that higher cognitive engagement might actually delay the onset of mental fatigue.
Signal Simplification and Frequency Shifting
As fatigue progressed, brain activity became less complex, evidenced by a universal decrease in Sample Entropy across all monitored brain regions. For instance, in the low-difficulty group, mean entropy in the O1 (occipital) region dropped from 1.438 to 1.208. Simultaneously, energy shifted from "fast" $\beta$ waves (alertness) to "slow" $\alpha$ and $\theta$ waves (drowsiness).
Statistical Significance
The ratios $\alpha/\beta$ and $(\alpha + \theta)/\beta$ increased significantly ($P < 0.05$) during the transition to fatigue. These physiological changes correlated strongly with behavioral declines, such as increased reaction times and decreased accuracy.
Conclusion
The ratios $\alpha/\beta$ and $(\alpha + \theta)/\beta$ serve as effective indicators for measuring mental fatigue. The study highlights that task difficulty is inversely related to fatigue generation to a certain extent. Future work will focus on expanding the sample size and utilizing Deep Learning techniques to develop automated, high-accuracy fatigue recognition systems suitable for low-power IoT edge devices.
Summary updated