Millimeter Wave Beam Tracking Scheme with Neural Network Embedded Kalman Filter

With an ever-increasing demand for wireless access capacity, the Millimeter-wave (mmWave) has become a promising technology for the forthcoming B5G/6G communications. Unfortunately, consistent high accuracy beam tracking is still a challenging task due to the narrow beam of mmWave. To address this challenge, we propose an intelligent neural network (NN) embedded Kalman filter (KF) scheme. The KF filter can provide robust beam tracking but with a penalty of accuracy loss, if the state transition function deviates from the changing of beam direction in practical scenarios. In our proposed scheme, the acquisition of the state transition function in the KF is converted to a neural network training process by introducing variance Bayesian inference. Therefore, the proposed scheme is a hybrid model-data-driven solution, which improves the prediction accuracy while inheriting the KF robustness. To verify the effectiveness of our proposed solution, we implement the proposed scheme on a software-defined radio system and conduct field testing. Experimental results illustrate that the proposed algorithm outperforms the existing beam tracking methods in terms of mean square error of beam angle prediction.