Multilevel Knowledge Distillation with U-Net for Resource-constrained Antenna Gain Prediction on IoT Edge Devices

Inference with Knowledge Distillation (U-LINK), lightweight neural network, Internet of Things (IoT)Deploying deep learning models for antenna gain prediction on IoT sensing nodes and edge gateways poses significant challenges due to severe constraints on memory, computation, and power.In sensor-driven IoT systems, reliable wireless transmission is crucial for maintaining the quality of sensing data, and antenna gain significantly impacts the communication stability between distributed sensors and edge gateways.We present U-LINK, a lightweight three-layer U-Net architecture with multilevel knowledge distillation optimized for resource-constrained devices with 2 GB of RAM.Using physics-informed augmentation, which expands 1,267 antenna designs to 12,670 samples while preserving electromagnetic validity (reciprocity, radiation efficiency, and power conservation), the proposed framework enables real-time antenna gain adaptation to support reliable sensing data transmission.Experimental results showed that U-LINK achieves R 2 = 0.964 (p < 0.001) with a 73.8% memory reduction (1,850 MB → 485 MB), a 73% latency reduction (45.2 ms → 12.4 ms), and a 67% power reduction (8.5 W → 2.8 W) compared with the teacher model.The student model maintains an R 2 = 0.98 correlation with teacher predictions (p < 0.001, Cohen's d = 2.85), enabling real-time on-device antenna optimization for environmental, agricultural, unmanned aerial vehicle or drone-based and intelligent infrastructure sensing.Cross-platform validation on three edge devices demonstrates robust performance (coefficient of variation CV = 0.10%).By allowing antenna gain to be adaptively optimized directly on sensor nodes or edge gateways, without relying on cloud-based electromagnetic simulation, U-LINK provides a practical solution for integrating intelligent antenna optimization into next-generation IoT sensing systems.Synergistic multilevel distillation integrating output, feature, and skip connection knowledge achieves +4.6% R 2 improvement over baseline distillation (p < 0.001, Cohen's d = 2.87), confirming effective knowledge transfer under aggressive compression.