Abstract:
To reduce the time and power overhead of millimeter wave (mmWave) beam training, a beam feature selection and prediction algorithm based on communication scenarios is proposed. First, the optimal feature-beam is selected according to the power loss probability minimization criterion, and a feature-beam set (a subset of beam indices) is generated using the optimal beam probability. Second, to obtain the optimal beam probability for communication scenarios, the local learning based clustering algorithm with feature selection (LLC-fs) is adopted. Finally, since there is an implicit and nonlinear mapping relationship between the scene-based feature-beam set and the optimal beam, the DNN model is used to approximate the mapping, and the offline training model is used to realize the prediction from the feature-beam set to the optimal beam. The simulation results show that the optimal mmWave beam can be predicted online by using the offline-trained scene-based DNN model. The prediction performance can approach the exhaustive beam search algorithm and effectively reduce the overhead of beam search.