Digital Twin Channel-Aided CSI Prediction: A Region-Adaptive Environment-Based Subspace Extraction Approach With Channel Similarity

Abstract

Low-overhead channel state information (CSI) acquisition is critical for multiple-input multiple-output (MIMO) systems. Digital twin channel (DTC), which reconstructs the 3D physical environment using environmental information, is a promising low-overhead CSI acquisition paradigm. However, most existing works use raw sensing data as environmental information, incurring prohibitive transmission and processing overheads that defeat its low-overhead purpose. The emerging wireless environmental information (WEI) concept addresses this by defining only channel-determining physical properties of objects and scatterers in the propagation environment. This letter proposes a novel adaptive environment-based channel subspace basis (AECB)-aided partial-to-whole CSI prediction method (AECB-CP), with AECB as an efficient WEI representation. Specifically, AECB is pre-extracted from channel similarity-guided adaptively partitioned grids, using fine grids in shadowed regions for accuracy and coarse grids in open areas for efficiency. The precomputed AECB is integrated into a neural network to reconstruct full CSI from partial pilots. Simulations show that AECB-CP reduces pilot overhead by 50% over the method without environmental information, while achieving gains in both prediction accuracy and storage cost compared with the fixed-grid baseline.