Pengfei Xue’s research while affiliated with Beijing Jiaotong University and other places

To improve the accuracy and generalization of channel modeling in complex scenarios, an explainable neural network (XNN)-enabled self-learning channel modeling approach is proposed. With the help of model visualization and feature importance analysis, it is shown that the XNN channel model can reveal the intrinsic relationship between the channel characteristics and system parameters. The output and input of the channel model can be represented by mathematical expressions, making the channel model more transparent and credible. The self-learning optimization training (SLOT) algorithm enables fine-tuning and self-optimization of the channel model to ensure scenario adaptation. Specifically, when predicting the path loss, the simulation results show that the root mean square error (RMSE) is consistently less than the predefined error threshold in various test scenarios at different buildings.

To accommodate the rapid change of radio propagation environment for mobile communication scenarios, millimeter-wave beamforming requires instantaneous channel state information (CSI) to update its operational parameters in real time, resulting in heavy system overhead. As the number of antennas increases, the system overhead associated with beam management will increase dramatically. To address this overarching problem, a neural network-aided millimeter-wave beamforming algorithm is proposed in this paper. A new parameter, referred to as “beam adjustment interval”, is proposed to evaluate the beamforming performance. It is defined as the maximum time duration in which the signal-to-interference-plus-noise ratio (SINR) of the user equipment can be maintained above the predefined threshold. Besides, a predictive method of beam adjustment to maximize the beam adjustment interval is developed, which considers the SINR not only at the current location but also future possible locations. Simulation results show that the proposed algorithm can significantly increase beam adjustment interval and reduce the total number of beam adjustments for the moving user equipment, thus reducing the system overhead 41.4% on average over 10 randomly generated test traces.