Physics-Informed Data Augmentation for CSI Data-Limited Deep Learning-Driven Communication Tasks

A variety of deep learning (DL)-driven wireless communication tasks rely on large volumes of channel state information (CSI) data for training. However, obtaining sufficient high-quality CSI data is often challenging, which limits the effectiveness and generalizability of DL-based communication methods. In this paper, we aim to design an efficient data augmentation (DA) method to augment CSI dataset for massive multiple-input multiple-output (MIMO) systems from a physics-informed perspective. We first point out that the CSI data of a massive MIMO system has a special indexed physical characteristic (IPC), which reveals a correspondence between the two-dimensional (2-D) indices in angular domain CSI and the physical characteristics of the communication environment. Leveraging such IPC, we propose a physics-informed data augmentation (PIDA) method that transforms angular domain CSI to generate physically meaningful CSI data aligned with actual physical characteristics. Experimental results show that the proposed PIDA achieves outstanding performance on CSI feedback and beamforming tasks, even with very limited CSI data, and performs robustly across various types of deep neural networks (DNNs).