Brownian Bridge-Based Diffusion Channel Denoising for ISAC Massive MIMO Systems

Abstract

Generative models represent a promising paradigm for enhanced channel estimation in 6G wireless communication systems. Recent studies have demonstrated that generative artificial intelligence (GAI)-based methods numerically provide superior channel estimation performance compared to conventional techniques such as least square (LS) and linear minimum mean square error (LMMSE) methods, as well as traditional deep learning (DL)-based channel estimation methods. Among these, diffusion models, including denoising diffusion probabilistic models (DDPMs), have garnered increasing attention for their ability to capture the underlying data distribution and provide high-quality estimates under noisy and limited observation conditions. In this work, we propose a diffusion model-based approach that aims to enhance the sensing channel estimation performance in integrated sensing and communication (ISAC) systems. Specifically, we treat the sensing channel matrix as an image and recast the channel estimation problem as a signal denoising task. To effectively capture the characteristics of the sensing channel, we employ a virtual channel matrix (VCM) model for initial processing. Additionally, to overcome the limitations of traditional DDPM architectures, particularly their requirement for a large number of time steps, we introduce a Brownian bridge (BB) process within the diffusion model. Our diffusion neural network architecture is meticulously designed to exploit the inherent properties of sensing channels in ISAC systems. Numerical results demonstrate the superior performance of our proposed channel estimation method compared to existing methods. Particularly, ablation experiments and analyses are conducted to verify the effectiveness of the proposed BB-based diffusion model design.