Convolutional Neural Network for Prediction Method of Path Loss Characteristics considering Diffraction and Reflection in an Open-Square Environment

This paper proposes new path loss prediction model using CNN that is one of the famous method in fields such as image-recognition in recent years. The proposed model can solve problems with consideration for effect of diffraction and reflection which is not considered in previous model. To verify the model, ray-tracing simulation in Open-Square environment was conducted by multiple scenarios. One simulation result were used for verification of the proposed model, and other results were used for learning of CNN. As a result, estimation accuracy was improved RMS error about 1.5 dB compared with the previous model in Non-Line-of sight area.

1. Introduction

In recent years, researches using deep-learning is popular in fields such as image-recognition. In particular, the technique using convolutional neural network (CNN) is most popular, because its estimation accuracy was the highest by considering space using the matrix convolution. We have been studying estimation method using CNN in Open-Square environment. The environment is one of the high traffic environments under 5G consideration, e.g., the public squares in front of large stations [reference 1]. In reference 2, the estimation method using CNN is proposed, and the result of the study that it can consider spatial information such as buildings and obstacles around transmitting and receiving antenna were stated. In reference 3, although highly accurate estimation results were shown, consideration of influence of reflection and diffraction by obstacles in the Non-Line-of-Sight (NLoS) against the Line-of-Sight (Los) environment was a problem. In addition, in reference 3, they have been studying using the same model as the authors in macro-cell environment, and the estimation accuracy obtained is similar to the authors, but the effect of reflection and diffraction on estimation accuracy has not been taken into consideration. In order to improve the estimation accuracy of the proposed model, we try to consider the influence of reflection in the NLoS environment. In this paper, we report results for the recognition result of improved prediction model using new parameter for considering the influence.