LightEdu-Net: Noise-Resilient Multimodal Edge Intelligence for Student-State Monitoring in Resource-Limited...

Area of Science:

• Educational Technology
• Artificial Intelligence
• Computer Vision

Background:

• Deploying student-state monitoring in rural classrooms is challenging due to noisy sensors and constrained computing resources.
• Existing multimodal perception systems struggle with real-time processing on low-cost edge devices.

Purpose of the Study:

• To develop a noise-resilient, multimodal, real-time student-state recognition system for low-cost edge devices.
• To address the limitations of sensor noise and computational constraints in educational technology deployments.

Main Methods:

• Proposed LightEdu-Net, a lightweight Transformer-based multimodal network integrating visual, physiological, and environmental signals.
• Incorporated a sensor noise adaptive module (SNAM), cross-modal attention fusion module (CMAF), and edge-aware knowledge distillation (EAKD).
• Constructed a multimodal behavioral dataset from rural schools for training and evaluation.

Main Results:

• LightEdu-Net achieved 92.4% accuracy and 91.4% F1-score, outperforming baseline models.
• Demonstrated strong robustness to sensor noise, with accuracy dropping only 1.1% at a noise level of 0.3.
• Achieved real-time performance on edge devices like Jetson Nano (42.8 ms latency) and maintained accuracy on Raspberry Pi 4B and Intel NUC.

Conclusions:

• LightEdu-Net offers an effective solution for noise-resilient, real-time student-state recognition on edge devices in resource-constrained environments.
• Provides a low-cost, quantifiable mechanism for capturing learning indicators, supporting educational economics research in underdeveloped regions.