MDRD-Based Channel State Information Acquisition Scheme for Intelligent Reflecting Surface-Aided Wireless Communication Systems

Abstract

The performance gains of intelligent reflecting surface (IRS)-aided wireless communication systems are highly dependent on the precise acquisition of channel state information (CSI). In IRS-aided wireless communication systems that employ model-driven CSI acquisition schemes, it is typically difficult to adapt to changes in the environment or user positions. The inadequate accuracy of CSI acquisition severely restricts the beamforming gains that IRS can bring to the system. Benefiting from minimal reliance on channel models and effective online estimation, data-driven CSI acquisition approaches are emerging as a research hotspot in IRS-aided wireless communication systems. However, in IRS-aided wireless communication systems using data-driven CSI acquisition schemes, a standard neural network is usually adopted, rather than constructing specialized architectures tailored to data characteristics, leaving room for further performance improvements. For obtaining more accurate real-time CSI and maximizing the performance gains brought by IRS, the CSI acquisition challenge is modelled as a denoising task, and a multi-scale dense residual denoising convolutional neural network (MDRD)-based denoising scheme is introduced in this paper. The scheme repeatedly adopts denoising blocks based on residual structures with short-skip connections and multi-scale convolutional kernels for network construction. The MDRD denoising model, with parameters set during offline training, is subsequently applied in the online application phase for CSI acquisition. Simulation results indicate that the MDRD-based CSI acquisition scheme achieves a relative enhancement of at least 36.2 % in CSI estimation accuracy at the cost of an almost negligible performance loss in denoising speed, thereby effectively balancing the accuracy and real-time capabilities of CSI acquisition.