Learning the CSI Recovery in FDD Systems

Wolfgang Utschick, Valentina Rizzello, Michael Joham, Zhengxiang Ma, Leonard Piazzi

Abstract

Obtaining accurate Channel State Information (CSI) in Frequency Division Duplex (FDD) systems is a significant challenge due to high feedback overhead. This paper presents a deep learning approach that enables a base station (BS) to reconstruct full downlink CSI from sparse feedback. By training a Convolutional Neural Network (CNN) exclusively on locally available uplink data, the proposed method bypasses the need for expensive downlink training samples while maintaining high reconstruction accuracy.

Introduction

In Massive MIMO systems, efficient beamforming relies on the base station's knowledge of the channel state. While Time Division Duplex (TDD) systems benefit from channel reciprocity, Frequency Division Duplex (FDD) systems—which dominate global cellular networks—do not. In FDD, the mobile terminal (MT) must measure the downlink channel and feed it back to the BS. As the number of antennas grows, this feedback overhead becomes a bottleneck, potentially negating the gains of Massive MIMO. The core challenge is reconstructing the full downlink channel from a highly compressed feedback report without access to ground-truth downlink data for training.

Proposed Method

The researchers treat the channel matrix (antennas vs. subcarriers) as a digital image and employ a Convolutional Neural Network (CNN) for reconstruction.

The UL-DL Conjecture and Training

The fundamental premise is that while signal phases differ between frequencies, the physical propagation environment remains constant. Thus, a model trained to reconstruct the uplink channel can generalize to the downlink. The BS performs centralized training using its own high-quality uplink measurements. It artificially sparsifies this data to simulate the feedback process, then trains the CNN to recover the original full matrix.

System model for downlink CSI recovery at the base station using partial feedback from the mobile terminal.

Dilated Convolutions

To handle sparse inputs effectively, the architecture utilizes dilated (atrous) convolutions. By skipping elements in the convolutional kernel, the network expands its receptive field, allowing it to capture long-range correlations across frequency gaps and fill in missing subcarrier information. This approach allows the network to maintain a manageable depth while effectively "seeing" across the gaps in sparse feedback.

CNN architecture details showing the sequence of dilated transposed convolutions used for CSI reconstruction.

Experimental Setup

The study utilized synthetic data from the QuaDRiGa channel simulator, focusing on an Urban Microcell (UMi) Non-Line-of-Sight scenario. The dataset comprised 300,000 samples. The hardware configuration included a 64-antenna array (8x8) at the BS and 160 subcarriers. Testing involved various frequency gaps between uplink (2.5 GHz) and downlink, ranging from 120 MHz to 480 MHz.

Results and Discussion

Statistical validation using Maximum Mean Discrepancy (MMD) confirmed that uplink and downlink channel distributions share a similar "statistical fingerprint," supporting the transfer learning approach.

Key findings include:

  • Superior Reconstruction: The CNN-based recovery significantly outperformed linear interpolation. Even with a 40x compression ratio (2.5% feedback), reconstruction remained highly accurate.
  • Sum-Rate Performance: Multi-user simulations showed per-user data rates nearly identical to those achieved with perfect CSI.
  • Robustness: Performance showed minimal degradation as the frequency gap increased, validating the robustness of the spatial physics learned by the model.

Conclusion and Future Work

While highly effective, the model is currently site-specific; a network trained for a microcell does not automatically generalize to a macrocell. Future work will explore "background training" where the BS adapts to its specific environment during operation. Additionally, researchers aim to optimize the feedback mask using techniques like Concrete Autoencoders and address the challenges of high-mobility environments where channels change rapidly.

Summary updated

BibTeX
@article{utschick2022learning,
  author = {Wolfgang Utschick and Valentina Rizzello and Michael Joham and Zhengxiang Ma and Leonard Piazzi},
  title = {{Learning the CSI Recovery in FDD Systems}},
  journal = {IEEE Trans. Wireless Communications},
  year = {2022},
  doi = {10.1109/twc.2022.3149946},
}
Figure