Source-linked AI summary
Deep Learning-based CSI Feedback Approach for Time-varying Massive MIMO Channels
Tianqi Wang, Chao-Kai Wen, Shi Jin, Geoffrey Ye Li
TL;DR
Large antenna arrays make FDD CSI feedback costly, motivating methods that reduce overhead without losing recovery quality. The paper extends CsiNet with LSTM to learn spatial structure and temporal correlation, and reports superior recovery, robustness to compression-ratio reduction, and real-time reconstruction.
Problem
Large antenna arrays in FDD massive MIMO create excessive CSI feedback overhead, while existing approaches face quantization, sparsity-model mismatch, or ignored temporal correlation.
Method
CsiNet-LSTM extends CsiNet with an LSTM network that extracts temporal correlation from sequences of channel features for CSI compression and reconstruction.
Results
CsiNet-LSTM achieves the lowest NMSE across tested compression ratios, loses only 8% indoors and 10% outdoors as compression ratio decreases, and supports real-time reconstruction.
Takeaways & Limitations
Learning spatial structure together with temporal correlation provides a favorable trade-off among compression ratio, recovery quality, and complexity for time-varying massive MIMO CSI feedback.
Abstract
from arXiv · showhide
Massive multiple-input multiple-output (MIMO) systems rely on channel state information (CSI) feedback to perform precoding and achieve performance gain in frequency division duplex (FDD) networks. However, the huge number of antennas poses a challenge to conventional CSI feedback reduction methods and leads to excessive feedback overhead. In this article, we develop a real-time CSI feedback architecture, called CsiNet-long short-term memory (LSTM), by extending a novel deep learning (DL)-based CSI sensing and recovery network. CsiNet-LSTM considerably enhances recovery quality and improves trade-off between compression ratio (CR) and complexity by directly learning spatial structures combined with time correlation from training samples of time-varying massive MIMO channels. Simulation results demonstrate that CsiNet- LSTM outperforms existing compressive sensing-based and DLbased methods and is remarkably robust to CR reduction.
I. INTRODUCTION
FDD massive MIMO requires CSI feedback for precoding, but large antenna arrays create excessive overhead and challenge conventional reduction methods. CsiNet-LSTM extends deep-learning CSI feedback with LSTM-based temporal modeling, improving recovery quality and the compression–complexity trade-off.
- Motivation: Large antenna arrays make conventional FDD CSI feedback reduction difficult and increase feedback overhead.Quantization errors affect CSI-sensitive applications, while codebook design becomes more difficult as antenna counts grow.
- Prior approaches: Compressive sensing methods exploit spatial and temporal CSI correlation but can struggle when CSI is only approximately sparse or models mismatch.Approximate sparsity and imperfect sparsity models hinder compression and recovery.
- Prior approaches: Slow-channel methods reuse previously retained CSI, but estimated information is difficult to update in real time and degrades in fast-changing channels.The approach also reaches a limit in reducing feedback overhead as channels change more rapidly.
- Prior approaches: CsiNet learns channel structures with an autoencoder-like encoder–decoder but reconstructs time-varying CSI independently and ignores temporal correlation.Its one-off feedforward decoder improves over compressive sensing methods but does not model channel evolution.
- Proposed approach: CsiNet-LSTM extends CsiNet with an LSTM network to extract time correlation alongside CNN-based CSI compression and initial recovery.The architecture is motivated by recurrent convolutional networks for spatial-temporal representation.
- Reported results: CsiNet-LSTM achieves the best recovery quality, outperforms state-of-the-art compressive-sensing methods in complexity, and remains robust as compression ratio decreases.The authors report real-time and extensible feedback without considerably increasing overhead compared with CsiNet.
II. SYSTEM MODEL
The system models time-varying FDD massive MIMO-OFDM channels and reduces CSI feedback by exploiting angular-delay sparsity and temporal correlation. Adjacent transformed channel matrices are grouped within the channel coherence time for correlated reconstruction.
- System model: The considered system is an FDD massive MIMO-OFDM downlink with Nc subcarriers, Nt transmit antennas in a uniform linear array, and a single-antenna user.The channel varies over time because of user-equipment mobility.
- System model: The CSI matrix Ht contains frequency-domain channel vectors across Nc subcarriers and is continuously estimated and fed back to track channel variation.The base station uses received downlink CSI to design beamforming or precoding vectors.
- Angular-delay sparsity: A 2D-DFT transforms Ht into an approximately sparse angular-delay matrix H′t, leveraging limited multipath delay and large-array angular structure.For finite Nt, the transformed channel remains only approximately sparse, which challenges conventional compressive sensing.
- Angular-delay sparsity: The method retains the first N′c rows of the transformed matrix to reduce parameters, while avoiding a sparsity-prior constraint in the proposed deep-learning feedback architecture.The retained rows capture the non-zero delay-domain structure used for sensing and recovery.
- Temporal correlation: User motion produces Doppler spread and time-varying channels, so CSI within the coherence interval Δt is treated as correlated.The coherence interval depends on maximum movement velocity, carrier frequency, and the velocity of light.
- Temporal correlation: T adjacent angular-delay channel matrices form a correlated group when 0 ≤ δt · T ≤ Δt, allowing current feedback and previous channel information to be combined.Each complex-valued transformed channel is compressed into an M-dimensional real-valued codeword, and a decoder with memory uses prior reconstructions.
- Recovery: The reconstructed angular-delay channels are converted back to the original spatial-frequency channel matrix using an inverse 2D-DFT.This completes the recovery pipeline after temporal information is incorporated.
III. CSINET-LSTM
CsiNet-LSTM extends CsiNet to model both angular-delay channel structure and temporal correlation. Its architecture uses feature extraction followed by correlation representation and final reconstruction.
- Architecture motivation: CsiNet-LSTM addresses CsiNet’s low-compression-ratio resolution loss by adding temporal correlation modeling to angular-delay sparsity processing.The design is motivated by the analogy between channel structure and temporal correlation and video spatial-temporal features.
- Architecture: The proposed architecture has two stages: angular-delay domain feature extraction, followed by correlation representation and final reconstruction.The stages correspond to spatial feature learning and temporal sequence processing.
1) Angular-delay domain feature extraction:
The feature-extraction stage uses CsiNet encoders and decoders with different compression ratios to represent a channel group efficiently. A high-compression representation initializes recovery, while lower-compression representations exploit correlation in later channels and share parameters.
- Angular-delay domain feature extraction: A high-CR CsiNet processes the first transformed channel to generate an M1 × 1 codeword that preserves sufficient structure for high-resolution recovery.The first channel receives a larger representation than the remaining channels.
- Angular-delay domain feature extraction: A low-CR CsiNet processes the remaining T − 1 channel matrices into M2 × 1 codewords, with M1 > M2 because correlated channels require less information.The codewords are concatenated with the first high-CR codeword before low-CR decoding.
- Angular-delay domain feature extraction: The CsiNet outputs are represented as two N′c × Nt feature matrices in the angular-delay domain.These extracted features provide the inputs to subsequent temporal correlation processing.
- Angular-delay domain feature extraction: All low-CR CsiNet instances share weights and bias, allowing the architecture to rescale when channel-group length T changes.In implementation, one low-CR CsiNet is reused T − 1 times rather than duplicated.
2) Correlation representation and final reconstruction:
CsiNet-LSTM extends CsiNet decoders with LSTMs that use sequences of decoder outputs to extract time correlation and perform final reconstruction. High- and low-compression encoders provide codewords that are combined before reconstruction, while the network is trained end to end with MSE.
- LSTMs receive length-T sequences from CsiNet decoders to extract time correlation and produce the final reconstruction.Each LSTM has 2N′cNt hidden units, matching the output dimension, and the outputs are reshaped into two N′ matrices.
- At each time step, the LSTMs merge previous inputs with current inputs and update correlation information to improve low-CR recovery quality.The reconstructed spatial-frequency CSI is obtained through inverse 2D-DFT, and the LSTMs reset after each channel group.
- All encoder and decoder parameters are learned end to end using ADAM optimization and mean-squared error loss.The parameters are denoted by Θ = {Θen, Θde}; the loss uses the Euclidean norm over the training samples.
- The first channel matrix is compressed with high CR, while the remaining T −1 matrices use lower-dimensional codewords that are combined with the first codeword.The low-CR codewords are concatenated with the first codeword before processing at the base station.
IV. SIMULATION RESULTS AND ANALYSIS
Simulations compare CsiNet-LSTM with CS-based methods and CsiNet across recovery quality, beamforming gain, runtime, and compression ratios. CsiNet-LSTM achieves strong low-CR recovery, robustness as CR decreases, and real-time reconstruction capability.
- Recovery performance: CsiNet-LSTM achieves the lowest NMSE at all compression ratios and outperforms CsiNet especially at low CR.The comparison includes NMSE, cosine similarity, and runtime.
- Recovery performance: CsiNet and CsiNet-LSTM maintain adequate beamforming gain at low CRs, whereas the CS-based methods fail to work.The comparison covers LASSO, TVAL3, and BM3D-AMP.
- Robustness to CR reduction: 8% and 10% performance loss are reported for CsiNet-LSTM as CR decreases in indoor and outdoor scenarios, respectively.The scenarios use COST 2100 time-varying MIMO channels with distinct UE velocities.
- Complexity and runtime: Approximately thousandfold faster execution is reported for DL-based methods than CS-based methods.CsiNet-LSTM is slightly less time-efficient than CsiNet, but its runtime remains shorter than the feedback interval δt = 0.04 s.
- Complexity and runtime: CsiNet-LSTM enables real-time reconstruction while improving NMSE and ρ relative to CsiNet.The reported runtime is considerably shorter than the feedback interval.
V. CONCLUSION
The proposed CsiNet-LSTM framework extends CsiNet with LSTM to provide real-time, end-to-end CSI feedback for time-varying massive MIMO channels. It uses time correlation and channel structure to balance compression ratio, recovery quality, and complexity.
- CsiNet-LSTM extends the deep learning-based CsiNet framework with LSTM for real-time, end-to-end CSI feedback.
- The framework utilizes time correlation and structure properties of time-varying massive MIMO channels.
- CsiNet-LSTM achieves a trade-off among compression ratio, recovery quality, and complexity.