Millimeter waves are complicated to model using conventional empirical methods since they are susceptible to complex environmental influences caused by various landscapes and building structures. We present a novel convolutional neural network (CNN)-based path loss modelling method based on a set of 28 GHz mmWave field measurements in suburban scenarios. The Enhanced Local Area Multi-Scanning algorithm (E-LAMS) that provides a CNN with path loss environmental information is proposed. A new CNN structure with four subnetworks and feature-sharing layers added between convolutional layers is also proposed. The proposed methods demonstrate their superior performance over empirical models and deterministic models in terms of accuracy and complexity. The root mean square error of 8.59 dB has been achieved in path loss prediction in the test scenarios.

An accurate and efficient path loss modeling method for millimeter wave communications plays a significant role in the large-scale deployment of a fifth-generation (5G) mobile communication system. Conventional path loss modeling methods such as deterministic methods, empirical methods, and machine learning-based methods, in practice, cannot achieve the desired level of performance in terms of accuracy. This paper proposes a novel path loss model for 5G communications in suburban scenarios using deep learning with dilated convolution and attention. Dilated convolution is used to alleviate the locality of feature extraction and capture the global information on input images. Attention rendered by global context blocks helps the attention-enhanced convolutional neural network (AE-CNN) model utilize the global information on inputs to extract essential features of propagation environments. A distance-embedded local area multi-scanning algorithm which generates input images that can improve learning the latent features with dependency on distance is proposed. The experimental results indicate that the AE-CNN model can outperform state-of-the-art deterministic and empirical methods in terms of root mean square error in test scenarios.