Estimation of Electromagnetic Field Strength: Experiments Using Vision Transformers

초록

The commercialization of fifth-generation (5G) mobile communication technology has enabled ultrafast data transmission and massive connectivity, while simultaneously introducing new challenges, such as exposure to electromagnetic fields (EMF). In particular, the use of massive multiple-input and multiple-output antennas and dense small cell-based 5G networks has increased the risk of EMF concentration in specific areas, thereby increasing the possibility of exceeding EMF regulations. Traditionally, EMF evaluations have relied on measurement-based approaches. However, the rapid deployment of 5G base stations and complex propagation environments pose significant challenges to these methods. Moreover, calculation-based approaches are limited to specific scenarios. Therefore, more accurate and efficient evaluation methods for EMF measurements are required. This study aims to overcome the limitations of existing evaluation methods by applying various vision transformer-based models, that have demonstrated outstanding performance in the field of image processing. Using physically-informed data representations, this study systematically applied notable vision transformer-based models, including the Vision Transformer, Swin Transformer, SegFormer, HRFormer, and Mask2Former for EMF prediction. The performance of the models was compared using a diverse array of metrics, including regression, classification, and perceptual quality analyses. Among these models, HRFormer achieved the best performance, recording a masked MAE of 5.45% and outperforming a baseline U-Net model by approximately 31%, while yielding the most perceptually robust estimations. To the best of our knowledge, this is the first comprehensive comparative study of ViT architectures applied to physically informed EMF estimation. The outcomes of this study are expected to contribute to a more accurate and efficient EMF management system by addressing the limitations of traditional evaluation methods.

키워드