PM Prediction Based on Time-Frequency Separation Feature Extraction

Huanming Zhang, Bo Lin, Feifei Gao

Abstract: One of the challenges in wireless communications is the pilot and feedback overhead. In this letter, we design a deep learning based time-frequency separation feature extraction network (TFNET) to predict the precoding matrix (PM) for massive multi-input multi-output (MIMO) systems. Specifically, we first design a feature extraction network to separately extract temporal and frequency features, which yields better prediction accuracy compared to standard neural network modules. Secondly, we utilize only a subset of past time slots and frequency bands to predict the current PM, which reduces the complexity of neural networks. Simulation results demonstrate that the proposed prediction method requires 75% pilot cost and achieves 113.72% prediction accuracy compared to the baseline.