Distributed Split Learning for Map-Based Signal Strength Prediction Empowered by Deep Vision Transformer

This article focuses on predicting the received signal strength (RSS) of mobile users, which is a fundamental problem for improving the coverage of cellular networks. Traditional methods for RSS prediction are based on ray tracing or stochastic radio propagation models. However, the former requires detailed environmental information that may not be practically available, while the latter cannot use the environmental data to its full potential and is not accurate enough. To address these issues, we design a practical RSS prediction system utilizing the trajectory information of users and satellite maps around base stations without sharing raw data. Specifically, we propose a map-based deep neural network (DNN) for the RSS prediction, empowered by the deep vision transformer (DeepVIT). Furthermore, to avoid sharing raw data, a novel split learning (SL) framework is developed. It splits the map-based DNN into two parts, which are deployed on the server and the user sides, respectively. The proposed models are evaluated by eight real-world data sets provided by an Australian Telecom operator, including 307,500 data samples. Simulation results show that in terms of prediction accuracy, the proposed map-based DNN significantly outperforms the 3GPP Urban Macro model and the long short-term memory method in centralized scenarios, while its SL variant outperforms the sequential SL and federated learning (FL) approaches in distributed scenarios. In addition, the communication overhead and the on-device computational burden of the proposed SL model are much lower than that of the FL approach.