This paper proposes a pre-training method for a deep- neural-network (DNN) based received power prediction scheme leveraging transfer learning for millimeter-wave (mmWave) networks. The received power prediction scheme has been proposed for proactive network control, which accurately predicts the received power 500 ms ahead using depth- camera images and a DNN. However, the prediction scheme requires a large number of training datasets and computational resources to prepare the accurate prediction model. In this paper, we propose a pre-training method that reduces the preparation time by leveraging the use of 3D model simulations with signal propagation simulations and transfer learning. The proposed method generates a dataset for pre- training using computer simulations, and trains a prediction model. The pre-trained model is transferred and fine-tuned by using a dataset obtained in a place where the system is actually used so that the model fits to the place. The experimental results show that the computational time of the proposed scheme with an RMS error of less than 5 dB is reduced by 78% compared with the previous work when using the dataset obtained in 60 s.

This study demonstrates the feasibility of proactive received power prediction by leveraging spatiotemporal visual sensing information towards reliable millimeter-wave (mmWave) networks. As the received power on a mmWave link can attenuate aperiodically owing to human blockages, a long-term series of the future received power cannot be predicted by analyzing the received signals prior to the blockage occurring. We propose a novel mechanism that predicts the time series of received power from the next moment to as many as several hundred milliseconds ahead. The key idea is to leverage camera imagery and machine learning (ML). Time-sequential images may involve the spatial geometry and mobility of obstacles representing mmWave signal propagation. ML is used to construct a prediction model from a dataset of sequential images labeled with received power in several hundred milliseconds ahead of the time at which each image is obtained. The simulation and experimental evaluations conducted using IEEE 802.11ad devices and a depth camera demonstrated that the proposed mechanism employing convolutional long short-term memory predicted a time series of received power up to 500 ms ahead, with an inference time of less than 3 ms and a root-mean-square error of 3.4 dB.

This paper experimentally finds the optimum number of input images of a machine learning-based mmWave received signal strength (RSS) value prediction scheme from depth images. By modeling the relationships between time-sequential depth images and RSS values based on machine learning, it is possible to predict the future RSS values, and thereby, a predictive handover makes a moment of degradation of the RSS value avoidable. As prediction models of RSS value, three machine learning models are compared: the convolutional neural networks (CNN), the combination of CNN and convolutional long short-term memory (CNN+ConvLSTM), and random forest. As the number of input images increases, the prediction accuracy generally improves, however, too numerous input images may make the prediction accuracy worse because of over-fitting. Experimental results reveal that the number of input images that are input in order to predict the RSS value the most accurately is 16.