Speaker
Description
Real-time and accurate isotope identification is critical for nuclear safety. Traditional methods rely on continuously transmitting raw spectral data from edge nodes to a centralized server, imposing severe bandwidth constraints and privacy vulnerabilities. While Federated Learning (FL) offers a decentralized alternative for this scenario, its deployment in radiation networks is hindered by feature misalignment caused by spectral drift and latency constraints on edge devices. To address these challenges, we propose a Physics-Aware Hardware-Efficient Federated Learning (PAHE-FL) framework. First, we embed a physics-aware preprocessing module that utilizes the ubiquitous K-40 background signature for unsupervised gain stabilization, ensuring spectral consistency across edge devices. Second, to tackle non-IID environmental interference, we design a decoupled training strategy where the feature extractor is aggregated globally to learn universal characteristics, while classifier heads are updated locally to adapt to specific backgrounds. Finally, the model is optimized via post-training quantization to enable deployment on resource-constrained microcontrollers. Experimental results on STM32 devices demonstrate that PAHE-FL achieves millisecond-level latency and superior robustness, significantly outperforming standard approaches in dynamic environments.
| Minioral | Yes |
|---|---|
| IEEE Member | No |
| Are you a student? | Yes |