Enterprise AI Analysis
Lightweight Fatigue Detection Using Physiological Signals via Knowledge Distillation for Usage-Based Insurance Systems
Authors: Hua Yang (Xi'an Vocational and Technical College), Haojie Zhang (Beijing Institute of Technology)
Abstract: Fatigue is a critical factor in road accidents, and its detection is essential for improving driver safety, particularly in the context of Usage-Based Insurance (UBI) and conditionally automated driving systems. In this paper, we propose a lightweight fatigue detection model based on physiological signals, including ECG, Electrodermal Activity (EDA), and respiration, which are directly correlated with the driver's autonomic nervous system. To address the challenge of deploying such systems on resource-constrained devices, we utilize a Knowledge Distillation (KD) approach, where a complex multi-branch teacher model transfers its knowledge to a compact student model. The student model, which processes concatenated multi-channel input, achieves high classification performance while maintaining computational efficiency. We demonstrate that the proposed model performs competitively compared to state-of-the-art deep learning architectures, but with significantly fewer parameters and reduced computational cost. Our results indicate that the proposed approach is well-suited for real-time fatigue detection in embedded systems, such as wearable devices and vehicle Electronic Control Units (ECUs), and can be integrated into UBI systems to enhance risk assessment and safety features.
Executive Impact: Unlocking Efficiency in Driver Safety AI
This research introduces a paradigm shift in real-time driver fatigue detection, enabling deployment on resource-constrained devices without sacrificing critical performance.
0%
Parameter Reduction
0%
FLOPS Reduction
0%
Model Size Reduction
0%
Achieved F1-Score
Deep Analysis & Enterprise Applications
Select a topic to dive deeper, then explore the specific findings from the research, rebuilt as interactive, enterprise-focused modules.
Proposed Framework
Experimental Design
Performance & Efficiency
Real-world Applications
Teacher-Student Model & Knowledge Distillation
The core innovation lies in a Knowledge Distillation (KD) framework. A complex, multi-branch Teacher Model (three ResNet34 networks for ECG, EDA, Respiration) extracts rich features. This knowledge is then transferred to a compact Student Model, designed with Efficient Feature Extraction Blocks (EFEB) for computational efficiency. The student model processes concatenated multi-channel signals, learning to mimic the teacher's nuanced decision-making with significantly fewer parameters.
KD allows the student model to learn from "soft targets" (teacher's probability distributions) alongside hard targets, providing a more comprehensive understanding of class relationships and uncertainties. This process enables the student to achieve comparable performance to the teacher while being drastically lighter for deployment on edge devices.
Dataset, Preprocessing, and Evaluation
The study utilizes the public AdVitam dataset, specifically Exp. 4, which focuses on driver fatigue induced by sleep deprivation in simulated rural and urban driving scenarios. Driver fatigue levels were assessed using the Karolinska Sleepiness Scale (KSS), with scores of 7 or higher indicating fatigue.
Physiological signals (ECG, EDA, respiration) were normalized to [-1, 1] and segmented into 30-second windows. Feature extraction employed Short-Time Fourier Transformation (STFT) to generate spectrograms. The dataset was split 80%-20% for training and testing, maintaining class distribution.
Models were trained for 150 epochs using Adam optimizer. Performance was rigorously evaluated using Sensitivity (Se), Specificity (Sp), Precision (Pr), Accuracy (Acc), F1-Score (F1), and Unweighted Average Recall (UAR), ensuring a comprehensive assessment of the model's effectiveness.
Competitive Performance with Superior Efficiency
The student model, optimized with KD (specifically at T=3.0 and α=0.5), achieved an F1-score of 93.02%, accuracy of 95.73%, and UAR of 94.92%. This performance is highly competitive with, and in some cases surpasses, many state-of-the-art deep learning architectures like ResNet18, SqueezeNet, ShuffleNetV2, and MobileNetV4, while also approaching the performance of the much larger Teacher model (VGG16).
Crucially, the proposed model boasts dramatically superior efficiency:
- Parameters: 321.63 K (compared to 64.62 M for ResNet34, a 99.5% reduction)
- FLOPS: 235.46 M (compared to 44.18 G for ResNet34, a 99.4% reduction)
- Total Size: 81.92 MB (compared to 872.77 MB for ResNet34, a 90.6% reduction)
This demonstrates that the model effectively balances high classification performance with minimal computational cost, making it ideal for real-time applications on constrained devices.
Integration into UBI Systems & Embedded Devices
This lightweight fatigue detection model is specifically designed for integration into Usage-Based Insurance (UBI) systems. By providing objective, real-time driver state monitoring, it enhances risk assessment beyond traditional telematics, leading to more accurate premiums and improved safety features.
The computational efficiency achieved through Knowledge Distillation makes the model highly suitable for deployment on resource-constrained embedded systems, such as wearable devices (smartwatches/bands) and vehicle Electronic Control Units (ECUs). This paves the way for ubiquitous, continuous driver fatigue monitoring without demanding significant hardware upgrades.
While promising, future work includes testing on diverse real-world driving data, refining fatigue labeling beyond binary thresholds, and implementing a full end-to-end deployment pipeline to assess latency and energy consumption on actual hardware.
Enterprise Process Flow: Real-Time Fatigue Detection System
Application Context (UBI Insurance System)
→
Data Input & Preprocessing
→
Proposed Framework (Teacher-Student Learning & Knowledge Distillation)
→
Output & Deployment (Real-Time IoT/Edge)
93.0% F1-Score
Achieved by the lightweight student model, comparable to SOTA, with massive efficiency gains.
| Model | F1-Score (%) | Accuracy (%) | Parameters (K) | Total Size (MB) |
|---|---|---|---|---|
| Teacher (VGG16) | 95.70 | 97.36 | 48,890 | 353.89 |
| Student (Our Model with KD) | 93.02 | 95.73 | 321.63 | 81.92 |
| ResNet34 | 93.76 | 96.17 | 64,620 | 872.77 |
| MobileNetV4 | 87.10 | 92.65 | 27,260 | 183.94 |
Case Study: Integrating Lightweight AI for UBI & Driver Safety
A leading automotive insurance provider aims to enhance its Usage-Based Insurance (UBI) offering by incorporating real-time driver state monitoring. Traditional deep learning models for fatigue detection were too resource-intensive for integration into vehicle ECUs or wearable devices, requiring significant computational power and increasing costs.
By adopting the Lightweight Fatigue Detection model with Knowledge Distillation, the provider can deploy an AI solution that achieves a 93.02% F1-score, comparable to much larger models, but with 99.5% fewer parameters and 90.6% smaller model size. This enables seamless, real-time fatigue detection directly on edge devices.
This integration leads to a more comprehensive risk assessment for UBI, allowing for dynamic premium adjustments and immediate safety alerts. Drivers benefit from enhanced safety features, while the insurer gains a competitive edge through advanced, cost-effective technology.
Calculate Your Potential AI Impact
Estimate the efficiency gains and cost savings your enterprise could achieve by leveraging optimized AI solutions.
Annual Savings
$0
Hours Reclaimed Annually
0
Our AI Implementation Roadmap
A clear, phased approach to integrating advanced AI into your operations, ensuring measurable results and seamless adoption.
Phase 1: Discovery & Strategy
In-depth analysis of current workflows, identification of AI opportunities, and tailored strategy development. Define clear KPIs and ROI objectives.
Phase 2: Pilot & Development
Rapid prototyping and development of a Proof-of-Concept. Iterative testing and refinement based on real-world data and feedback.
Phase 3: Integration & Scaling
Seamless integration of AI solutions into existing systems. Comprehensive training for your teams and strategic scaling across the enterprise.
Phase 4: Optimization & Support
Continuous monitoring, performance optimization, and ongoing support to ensure long-term success and adaptation to evolving needs.
Ready to Transform Your Enterprise with AI?
Unlock the full potential of artificial intelligence with a tailored strategy designed for your business. Schedule a complimentary consultation with our experts today.
Ready to Get Started?
Book Your Free Consultation.
Let's Discuss Your AI Strategy!
Lets Discuss Your Needs
Select a Date
Sun
Mon
Tue
Wed
Thu
Fri
Sat