In this paper, we describe a novel 3-D Luneburg lens using radial holes. A Luneburg lens can suppress the side lobe of the antenna and increase the gain. The prototype Luneburg lens was fabricated using a 30mm radius dielectric sphere with uniform permittivity by 2.1 and many holes by 1mm radius, called radial holes drilled vertically on the sphere surface. The relative permittivity distribution of the Luneburg lens is realized by changing the number and depth of five types of holes. To confirm the effectiveness of the fabricated Luneburg lens, the far-field radiation pattern at 25GHz was measured using a flanged rectangular waveguide as a wave source. It achieved a high gain of 22.4dBi. The measured radiation pattern generally agreed well with the theoretical radiation pattern of the Luneburg lens, and the validity of the Luneburg lens with radial holes at the quasi-millimeter wave was confirmed. It can be used for beam steering in 5G/6G base stations.
Studies are being conducted on analysis methods for multilayer networks. If each layer network of the original multilayer network can be reconstructed from the single-layer network obtained by aggregating the multilayer network, a more detailed analysis of the multilayer network will become possible by using conventional network analysis methods. In this study, we investigate the extent to which the original multilayer network can be reconstructed from the single-layer network with the aggregated multilayer network. Our findings include that semisupervised learning of graph neural networks can reconstruct more than 50% of an original multilayer network from the single-layer network with the aggregated multilayer network.
This study is motivated by the demand for an efficient deep learning-based model that helps us predict the future link quality for intelligent decision-making systems. In this letter, we propose a transfer learning-based approach to predict millimeter-wave future received power in an indoor environment. The model is pre-trained using formulation-aid generated data and fine-tuned using measured data. The proposed framework reduces more than 30% in root-mean-square error and 6.5% in accuracy with high training speed compared to the baseline training from scratch.