Machine-Learning-Based Future Received Signal Strength Prediction Using Depth Images for mmWave Communications

This paper discusses a machine-learning (ML)-based future received signal strength (RSS) prediction scheme using depth camera images for millimeter-wave (mmWave) networks. The scheme provides the future RSS prediction of any mmWave links within the camera's view, including links where nodes are not transmitting frames. This enables network controllers to conduct network operations before line-of-sight path blockages degrade the RSS. Using the ML techniques, the prediction scheme automatically learns the relationships between the RSS values and the time-series depth images. We apply a powerful neural-network model, which is capable of learning both spatial and long-term temporal relationships: the Convolutional Neural Network + Convolutional Long Short-Term Memory (CNN+ConvLSTM) network. We generate RSS values and depth images simulating a real-life environment using ray-tracing software and three-dimensional computer graphics (3DCG) software, and then we evaluate the accuracy of the proposed scheme and reveal the impact of camera positions on the prediction accuracy. Moreover, we conduct experiments using commercially available IEEE 802.11ad devices and an RGB-D camera to demonstrate the feasibility of the scheme in a real-life environment. Simulation and experimental results show that the CNN+ConvLSTM network achieves a higher-accuracy in future RSS value prediction compared to other ML models. Moreover, the results confirm that the scheme is feasible for real-time prediction systems and show that the scheme predicts RSS values in 0.5 s future with RMS errors of less than 2.1 dB in the simulations and 3.5 dB in the experiments.