Spatial Prediction of Channel Signal Strength Map Using Deep Fully Convolutional Neural Network

In this paper, we propose a deep learning pipeline to predict the signal strength map of a wireless channel at unvisited locations over the space, based on very sparse channel samples. Our goal is to enable prediction in complex indoor environments where channels can experience drastic spatial variations. We show how to build a comprehensive large training dataset based only on simulated data that we generate through two complementary approaches: RF propagation simulation and probabilistic channel generation. By employing a fully convolutional neural network, we then develop a channel predictor that 1) does not require any environmental information (e.g., area map, transmitter location), 2) does not make any assumptions on the distribution or characteristics of the channel, and 3) can handle environments of different sizes. We conduct extensive evaluations on a large real-world indoor test set that consists of many different areas, as well as on a realistic ray-tracing test set consisting of 225 different environments. Our results show that our proposed approach can predict channel spatial details well, and further outperforms the state-of-the-art considerably in both accuracy and computation time.