Long Short-Term Memory Based Millimeter Wave Beam Change Prediction via Real-World Data

Beam management (BM) is a key component in fifth generation (5G), 5G-Advanced, and sixth generation (6G) millimeter-wave (mmWave) operations. Due to mobility or rotation of user equipment (UE), frequent beam measurements and search are needed at the UE based on also frequently transmitted reference signals (RSs). To enable mmWave BM with lower power consumed at the UE and less RS overhead, this paper proposes a machine learning (ML) assisted beam change prediction scheme based on long short-term memory (LSTM) facilitated models. Particularly, the ML-models are equipped at base station (BS) and use time series of UE reported beam qualities as inputs, which are based on the latest 3rd Generation Partnership Project (3GPP)5G new radio (NR) standards. The model outputs prediction on whether the strongest beam will be different from the currently reported one until the next applicable measurement occasion. This allows the system to operate with an extended beam measurement cycle greater than commercially popular ones with very limited beam mismatch errors. Our proposed ML-models are trained and tested via data collected using a cellphone equipped with Snapdragon 5G-mmWave Modem chipset and a prototype BS, both compliant with 3GPP Release-15 5G standards. Evaluation results show that the proposed scheme can accurately predict beam change instances and reduce over half of the power consumed at the UE for BM, which also demonstrate feasibility of ML based BM for 5GAdvanced and 6G communications.