Channel Fingerprint Construction for Massive MIMO: A Deep Conditional Generative Approach

TL;DR

The paper tackles the challenge of constructing ultra-fine channel fingerprints (CF) for massive MIMO when only coarse CF data are economically feasible to collect. It introduces CF twins and a conditional diffusion model (CGDM) that learns the conditional distribution $p(G_{HR}|G_{LR})$ via ELBO-based variational training, using $G_{LR}$ as side information to iteratively refine HR CF from Gaussian noise. A lightweight variant, LiCGDM, is developed through one-shot layer pruning and multi-objective knowledge distillation to maintain performance with fewer parameters. Experimental results on QuaDRiGa-generated data show CGDM achieving competitive SR CF reconstruction and strong zero-shot generalization to unseen magnification factors, with LiCGDM providing practical deployment benefits. Overall, the CF twin framework enables environment-aware wireless design by efficiently bridging coarse sensing data and fine-grained channel knowledge.

Abstract

Accurate channel state information (CSI) acquisition for massive multiple-input multiple-output (MIMO) systems is essential for future mobile communication networks. Channel fingerprint (CF), also referred to as channel knowledge map, is a key enabler for intelligent environment-aware communication and can facilitate CSI acquisition. However, due to the cost limitations of practical sensing nodes and test vehicles, the resulting CF is typically coarse-grained, making it insufficient for wireless transceiver design. In this work, we introduce the concept of CF twins and design a conditional generative diffusion model (CGDM) with strong implicit prior learning capabilities as the computational core of the CF twin to establish the connection between coarse- and fine-grained CFs. Specifically, we employ a variational inference technique to derive the evidence lower bound (ELBO) for the log-marginal distribution of the observed fine-grained CF conditioned on the coarse-grained CF, enabling the CGDM to learn the complicated distribution of the target data. During the denoising neural network optimization, the coarse-grained CF is introduced as side information to accurately guide the conditioned generation of the CGDM. To make the proposed CGDM lightweight, we further leverage the additivity of network layers and introduce a one-shot pruning approach along with a multi-objective knowledge distillation technique. Experimental results show that the proposed approach exhibits significant improvement in reconstruction performance compared to the baselines. Additionally, zero-shot testing on reconstruction tasks with different magnification factors further demonstrates the scalability and generalization ability of the proposed approach.

Figures (7)

• Figure 1: Schematic diagram of the CF twin: CGDM functions as the core computational unit of the CF twin, reconstructing fine-grained CF to optimize wireless transmission technologies and network planning.
• Figure 2: The mechanism of CGDM for generating HR CF consists of a Gaussian diffusion process (without learnable parameters) and an iterative refinement process based on LR CF. Specifically, the pink arrow indicates the direction of the Gaussian diffusion process, which progressively adds noise to the HR CF. The blue arrow indicates the direction of the iterative refinement process, which utilizes the implicit prior learned during training and is conditioned on the source information (LR CF) to generate the HR CF.
• Figure 3: Diagram and key modules of the CGDM architecture. Specifically, the network architecture of CGDM consists of three primary stages: Dn (substages 1-5), Mid (substage 6), and Up (substages 7-12). The blocks included in each stage are illustrated in the subfigures (a), (b), (c), (e), (f), and (g). (d) shows the network architecture of the proposed CGDM. Additionally, the red and purple arrows represent the embedding of the time constant $t$ and the skip connections, respectively. Taking the $\times4$ HR CF reconstruction task as an example, the LR CF with a size of $32\times32\times3$ is upsampled to the target resolution (i.e., $128\times128\times3$), and concatenated with noise of the same resolution along the channel dimension to form the input, resulting in a size of $128\times128\times6$. The number $2c_{in}$ of input channels is expanded to $c_1$, representing the number of base channels in the latent space, after passing through substage 0. Note that within the same substage, the height $h$ and width $w$ of the feature maps remain unchanged, while the number $c$ of channels at different substages is controlled by the channel number multiplier $\bar{\eta} = {c_1}:{c_2}:{c_3}:{c_4}:{c_5}$. The specific values of $c_1$, $\bar{\eta}$, and $N_{\rm{RA}}$ will be discussed in Subsection \ref{['Experiment Results']}.
• Figure 4: The layout of the massive MIMO-OFDM system.
• Figure 5: Convergence analysis of the CGDM under different hyperparameter settings: (a) represents different base channels $c_1$ in the feature maps within the latent space; (b) shows different channel number multipliers $\bar{\eta}$; (c) illustrates varying numbers $N_{\rm{RA}}$ of integrated Res$^+$ and self-attention blocks.