Res-LDAMP: Residual Deep Unfolding for Robust Beamspace Channel Estimation

Abstract

Millimeter-wave (mmWave) massive multiple-input multiple-output (MIMO) enable high data rates but rely on accurate channel state information (CSI) for efficient beamforming and data transmission. Approximate message passing (AMP) and learned variants (LAMP/LDAMP) suffer from performance degradation under high sparsity levels and propagation environments. Therefore, we propose a novel deep unfolding framework that integrates model-driven LDAMP with data-driven residual learning to enhance beamspace channel estimation accuracy, called Res-LDAMP. The Res-LDAMP method embeds a residual denoising convolutional neural network (DnCNN) and connected network within the LDAMP iterative structure, enabling effective nonlinear signal recovery. Furthermore, we introduce a dual Onsager correction for leveraging original measurement and its conjugate to improve convergence and reduce estimation error. The simulations are conducted using the Saleh–Valenzuela (S-V) channel model and the DeepMIMO dataset under signal-to-noise ratio (SNR). The results demonstrate that Res-LDAMP outperforms AMP, LAMP, and LDAMP. Improved channel estimation accuracy leads to enhance beam selection performance and reduced sum-rate degradation. The findings show the combination of deep unfolding and residual learning for next-generation mmWave and beyond wireless communication systems.