A Deep Learning-Based Indoor Radio Estimation Method Driven by 2.4 GHz Ray-Tracing Data

Abstract

This paper presents a novel method for estimating received signal strength (RSS) in indoor radio propagation using a deep learning approach. The proposed method utilizes a training dataset comprised of imitated real-world indoor environments and radio-map images generated through 2.4 GHz ray-tracing. Additionally, we introduce a convolutional neural network (CNN) named Radio Residual UNet (RadioResUNet) to facilitate the training and prediction of indoor radio propagation. To assess the feasibility and effectiveness of this deep learning network for indoor radio estimation, we compare the RSS obtained from practical wireless equipment with that obtained by RadioResUNet in two indoor environments: an anechoic chamber and an office floor. Furthermore, we explore the prediction outcomes achieved using different loss functions, including mean squared error (MSE), binary cross-entropy (BCE), and dice binary cross-entropy (Dice_BCE), across varying dataset sizes. The results reveal that the proposed deep learning-based radio estimation method exhibits estimation discrepancies of 4.25 dB and 5.4 dB compared to practical measurements in real-world environments of the anechoic chamber and the office floor, respectively. These results indicate a performance that is comparable to the indoor propagation model of ITU-R P.1238. Additionally, we introduce an indoor radio estimation tool that utilizes the deep learning network of RadioResUNet to predict radio propagation in a target area with minimal input.

Keywords

• Convolutional neural network (CNN)
• deep learning-based radio estimation
• received signal strength (RSS)
• indoor radio propagation
• RadioResUNet
• ray-tracing