Abstract:

Indoor beam selection method based on depth learning can significantly improve the beam matching probability and search efficiency, but this method requires large data sets to adjust a large number of trainable parameters, which leads to additional system overhead. To solve this problem, a migration learning technology is combined to enable the target scene neural network to obtain matching accuracy similar to that of large data sets in the form of small data sets, thus reducing the impact of data set size on the matching results in the beam selection method based on depth learning. Firstly, a large dataset is used to fully train the source neural network in a source scene, so that the network parameters can fully contain the channel state information and environment information. Then, the source neural network parameters are used to initialize the neural network in the target scene to varying degrees, so that the neural network can still obtain good beam matching performance after being trained in a small dataset. The simulation results show that for indoor beam selection scenarios, in the case of limited data sets, using the migration learning method for beam selection can also achieve high matching accuracy.

Key words: millimeter wave, indoor environment, beam search, transfer learning