Abstract
Deep learning based channel state information (CSI) feedback in frequency division duplex systems has drawn much attention in both academia and industry. In this paper, we focus on integrating the Type-II codebook in the beyond fifth-generation (B5G) wireless systems with deep learning to enhance the performance of CSI feedback. In contrast to its counterpart in Release 16, the Type-II codebook in Release 17 (R17) exploits the angular-delay-domain partial reciprocity between uplink and downlink channels and selects part of angular-delay-domain ports for measuring and feeding back the downlink CSI, where the performance of the conventional deep learning methods is limited due to the deficiency of sparse structures. To address this issue, we propose the new paradigm of adopting deep learning to improve the performance of R17 Type-II codebook. Firstly, considering the relatively low signal-to-noise ratio of uplink channels, deep learning is utilized to refine the selection of the dominant angular-delay-domain ports, where the focal loss is harnessed to solve the class imbalance problem. Secondly, we propose to reconstruct the downlink CSI by way of deep learning based on the feedback of R17 Type-II codebook at the base station, where the information of sparse structures can be effectively leveraged. Finally, a weighted shortcut module is designed to facilitate the accurate reconstruction, and a two-stage loss function with the combination of the mean squared error and sum rate is proposed for adapting to actual multi-user scenarios. Simulation results demonstrate that our proposed angular-delay-domain port selection and CSI reconstruction paradigm can improve the sum rate performance by more than 10% compared with the traditional R17 Type-II codebook and deep learning benchmarks.
Introduction
In modern massive MIMO (Multiple-Input Multiple-Output) wireless systems, the Base Station (BS) requires accurate Downlink (DL) Channel State Information (CSI) to manage interference and maximize data rates. In Frequency Division Duplex (FDD) systems, the uplink and downlink operate on different frequencies, preventing the BS from directly inferring the downlink channel from uplink measurements. Consequently, the User Equipment (UE) must measure the channel and feed that information back to the BS—a process that often consumes significant signaling overhead.
The 3GPP Release 17 (R17) Type-II codebook attempts to mitigate this overhead by leveraging "partial reciprocity." This principle suggests that while frequencies differ, the physical paths—specifically the angles and delays—of the signal remain similar. By selecting a subset of "ports" in the angular-delay domain for measurement, the system reduces feedback volume. However, two major hurdles remain: noisy port selection due to low uplink Signal-to-Noise Ratio (SNR) and reconstruction errors caused by quantization during feedback. This paper proposes a deep learning (DL) framework to address these limitations, creating a new paradigm for R17-compliant CSI feedback.
Proposed Methodology
The proposed framework integrates Deep Learning into the R17 workflow at two critical stages: port selection and channel reconstruction.
1. DL-Based Port Selection
Traditional methods rely on simple power-based thresholds to select angular-delay ports, which are highly susceptible to noise in the uplink. The authors propose using a Convolutional Neural Network (CNN) to identify the optimal ports from noisy data.
- Circular Padding: To respect the cyclic nature of the Discrete Fourier Transform (DFT) bases used in wireless signals, the CNN employs circular padding. This ensures that correlations at the edges of the angular-delay domain are preserved rather than buried by standard zero-padding.
- Focal Loss: Because only a small fraction of available ports are useful (sparsity), the training process faces a "class imbalance" problem. The use of focal loss allows the model to focus on hard-to-classify useful ports, improving selection accuracy in low-SNR environments.
2. DL-Based CSI Reconstruction
Once the quantized feedback is received at the BS, a second CNN is employed to reconstruct the full channel. This stage introduces two key innovations:
- Weighted Shortcut Module: Recognizing that the feedback is already a coarse representation of the truth, the network utilizes a weighted shortcut. This allows the model to learn the residual (the difference between the feedback and the actual channel), making the learning process more efficient and accurate.
- Two-Stage Loss Function: The model is initially trained using Mean Squared Error (MSE). It is then fine-tuned using a combined loss function of MSE and Negative Sum Rate. This ensures the reconstruction is optimized for actual data throughput and multi-user interference reduction, rather than just mathematical similarity.
Experimental Setup
The study evaluates the proposed paradigm using the 3GPP 38.901 channel model in an Urban Macro-cell (UMa) scenario. The dataset comprises 102,400 training samples and 2,560 validation samples, covering both Line-of-Sight (LOS) and Non-Line-of-Sight (NLOS) conditions to ensure environmental robustness.
Key Results and Insights
Performance Gains
The proposed method achieves a significant performance boost, improving the average sum rate by more than 10% compared to both the traditional R17 Type-II codebook and existing deep learning benchmarks like EVCsiNet.
Robustness to Noise
The DL-based port selection demonstrates remarkable resilience. Even at an uplink SNR of -5 dB, the system achieves a normalized CSI power ($P_N$) of over 93%. This performance nearly matches what would be achieved with a perfect, noise-free uplink, validating the CNN's ability to extract signal structure from noise.
Architectural Advantages
A core insight of this work is the strategic placement of the "intelligence" at the Base Station. While traditional DL feedback models (like CsiNet) place the encoder at the UE, the irregular nature of R17 feedback makes this difficult. By performing reconstruction at the BS, the system can map coefficients back to their original physical ports, effectively leveraging the sparse structure of the channel that CNNs are optimized to process.
Conclusion
This research introduces a robust paradigm for enhancing R17 Type-II codebooks through deep learning. By refining port selection and optimizing reconstruction with a sum-rate-oriented loss function, the framework significantly improves spectral efficiency in FDD massive MIMO systems. Future work will focus on reducing the computational complexity for real-time BS implementation and exploring the model's generalizability across diverse environments such as indoor and rural cells.
Summary updated