Cost-Effective Federated Learning-Based Approach for SINR Prediction in Cellular-Connected UAVs

Abstract

This study introduces a novel approach to empower cellular-connected unmanned aerial vehicles (UAVs) in predicting signal quality. The proposed prediction model leverages data collected by the UAVs, addressing privacy concerns and ensuring effectiveness, while taking into account the constraints of UAVs. A unique three-step approach is proposed, which integrates a detailed physical ray-tracing (RT) method, deep learning, and federated learning (FL) for continuous learning and field adaptation. A dual input feature fusion convolutional neural network (DIFF-CNN) model is proposed, which is pretrained on RT data and fine-tuned using data collected by the UAVs via FL. The proposed model demonstrates superior performance and robustness to data sparsity compared to traditional machine learning algorithms. Notably, the model achieves a root mean squared error of 0.837 dB and an R-squared of 97.7% for signal-to-interference-plus-noise ratio (SINR) prediction after the fine-tuning step in the fixed-altitude scenario, but performance drops with uniform altitude distribution, highlighting the impact of flying height on fine-tuning. The research indicates that the proposed approach can enhance performance while reducing training rounds by 35% to 90%, thus mitigating FL overheads. Future research could explore efficiency gains by using different pretrained models tailored to specific flying heights.