Presentation Information
[B-1A-15]Effect of Training Data Density on the Accuracy of LSTM-Based Received Signal Level Prediction
◎Kosuke Ishii1, Tetsuro Imai1, Toshiki Hozen2, Kazuma Tomimoto2, Ryo Yamaguchi2, Shumpei Tabuchi2 (1. Tokyo Denki University, 2. SoftBank Corp.)
Keywords:
6G,LSTM,Human Body Blockage
For the utilization of high-frequency bands such as millimeter-wave bands in the sixth-generation mobile communication systems (6G), research on propagation prediction using machine learning (LSTM) and prediction of received signal level fluctuations caused by human blockage is being actively conducted. However, the effects of the time resolution (sample density) and frequency components required for the training data on prediction accuracy have not yet been clarified. In this paper, using propagation data measured at 2000 samples/s in an indoor 29.7 GHz band, we individually evaluated the effects of the sampling rate and high-frequency components of the training data on the prediction performance of LSTM. The evaluation results revealed that the degradation of prediction accuracy associated with a decrease in the sampling rate is caused not by the loss of high-frequency components, but by the lack of time resolution (insufficient sample density) relative to the variations in the received signal level. Furthermore, we showed that the training data must be acquired with a sample density of at least 1200 samples/s to achieve sufficient prediction accuracy under the conditions of this experiment.