Visual transformer based unified framework for radio map estimation and optimized site selection

The quality and coverage of signals within cellular networks are determined by base stations and their locations. The deployment of a base station relies on fine-grained radio maps. However, previous work involves iterating through all possible base station locations and selecting the base station location with the maximum signal coverage as the optimal reference deployment location, resulting in repeated and large time-consuming tasks. This paper proposes a Visual Transformer (ViT) based radio map estimation method, termed TxSTrans (Transmitter Selection Transformer), unifying radio map estimation and optimized site selection. The TxSTrans can accept parallel base station data inputs and quickly recommend the optimal base station site within given candidates to maximize the signal coverage. TxSTrans contains a multi-modal feature fusion module for integrating building maps (environment) and given base station coordinates, an information fusion pyramid module for reconstructing a high-resolution radio map, and a recommendation module for finding the optimal base station site among candidates. A knowledge distillation learning method is also incorporated to train the parameters of the recommendation module to sense the best base station site among candidates. Simulation results show that TxSTrans facilitates radio map estimation and optimized site selection. The runtime of the base station siting is 4.62 times faster than the sequential search for the optimal base station site by RadioUnet.