Deep Learning-based Signal Strength Prediction Using Geographical Images and Expert Knowledge

摘要

Methods for accurate prediction of radio signal quality parameters are crucial for optimization of mobile networks, and a necessity for future autonomous driving solutions. The power-distance relation of current empirical models struggles with describing the specific local geo-statistics that influence signal quality parameters. The use of empirical models commonly results in an over- or under-estimation of the signal quality parameters and require additional calibration studies. In this paper, we present a novel model-aided deep learning approach for path loss prediction, which implicitly extracts radio propagation characteristics from top-view geographical images of the receiver location. In a comprehensive evaluation campaign, we apply the proposed method on an extensive real-world data set consisting of five different scenarios and more than 125.000 individual measurements.It is found that 1) the novel approach reduces the average prediction error by up to 53 % in comparison to ray-tracing techniques, 2) A distance of 250 - 300 meters spanned by the images offer the necessary level of detail, 3) Predictions with a root-mean-squared error of ≈ 6 dB is achieved across inherently different data sources. In this paper, we present a novel model-aided deep learning approach for path loss prediction, which implicitly extracts radio propagation characteristics from top-view geographical images of the receiver location. In a comprehensive evaluation campaign, we apply the proposed method on an extensive real-world data set consisting of five different scenarios and more than 125.000 individual measurements.It is found that 1) the novel approach reduces the average prediction error by up to 53 % in comparison to ray-tracing techniques, 2) A distance of 250 - 300 meters spanned by the images offer the necessary level of detail, 3) Predictions with a root-mean-squared error of ≈ 6 dB is achieved across inherently different data sources. It is found that 1) the novel approach reduces the average prediction error by up to 53 % in comparison to ray-tracing techniques, 2) A distance of 250 - 300 meters spanned by the images offer the necessary level of detail, 3) Predictions with a root-mean-squared error of ≈ 6 dB is achieved across inherently different data sources.