A Data Augmentation Approach to 28GHz Path Loss Modeling Using CNNs

Abstract

Millimeter waves are easily influenced by the surrounding environment, making it difficult to predict path loss values for 28GHz communication systems. Recently, deep learning approaches have become popular mainly thanks to their superior performance in terms of prediction accuracy, generalizability as well as local adaptability. These deep learning approaches require a sufficient number of training data which often lacks variability with respect to the parameter values of base station configuration if not unavailable at all. This paper proposes to use the data augmentation approach to address these two issues by using a simulator to generate predicted data for the arbitrary values of base station parameters. It is shown that a Convolution Neural Network (CNN) trained with both measurement and augmented data outperforms a vanilla CNN model trained with measurement data only and that it can make accurate predictions for arbitrary base station configurations.