[Paper Review] Deep Learning-based CSI Feedback Approach for Time-varying Massive MIMO Channels

Tianqi Wang, Chao-Kai Wen+2 more|arXiv (Cornell University)|Jul 31, 2018

Full-Duplex Wireless Communications12 references4 citations

TL;DR

This paper proposes CsiNet-LSTM, a deep learning-based CSI feedback framework for time-varying massive MIMO channels that integrates long short-term memory (LSTM) networks with a CNN-based autoencoder to exploit temporal correlation and spatial structure. It achieves superior reconstruction accuracy and robustness to compression ratio reduction compared to existing compressive sensing and deep learning methods, with NMSE improvements of up to 94% and 10% in indoor and outdoor scenarios, respectively, while maintaining real-time feasibility via GPU acceleration.

ABSTRACT

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.

Motivation & Objective

• Address the high feedback overhead in FDD-based massive MIMO systems due to the large number of antennas and time-varying channel conditions.
• Overcome limitations of conventional compressive sensing and codebook-based methods, which suffer from model mismatch and high complexity.
• Leverage temporal correlation in time-varying channels to improve CSI recovery quality without increasing feedback overhead.
• Develop an end-to-end, real-time deep learning framework that jointly learns spatial structure and temporal dynamics for efficient CSI feedback.
• Achieve a better trade-off between compression ratio, reconstruction accuracy, and computational complexity than state-of-the-art methods.

Proposed method

• Extend the CsiNet autoencoder architecture with a long short-term memory (LSTM) network to model temporal correlation in time-varying massive MIMO channels.
• Use a 2D discrete Fourier transform (2D-DFT) to convert the CSI matrix into an approximately sparse representation in the angular-delay domain, enabling efficient compression.
• Apply a CNN-based encoder to compress the CSI into low-dimensional codewords, followed by an LSTM layer that processes sequences of codewords across time to refine predictions.
• Use a decoder network to reconstruct the CSI from the LSTM-processed features, with the entire system trained end-to-end via backpropagation.
• Integrate codeword concatenation across time to provide richer measurements for low-compression-ratio decoding, improving recovery quality.
• Train and test the model using real-world channel traces from the COST 2100 model under both indoor and outdoor mobility scenarios.

Experimental results

Research questions

• RQ1 Can an LSTM-enhanced deep learning architecture significantly improve CSI reconstruction quality in time-varying massive MIMO channels compared to conventional compressive sensing and standard deep learning methods?
• RQ2 To what extent does incorporating temporal correlation via LSTM reduce the performance degradation of CSI feedback under high compression ratios?
• RQ3 How does the proposed CsiNet-LSTM framework compare in terms of reconstruction accuracy, beamforming gain, and computational latency to existing CS-based and DL-based CSI feedback methods?
• RQ4 Does the integration of temporal dynamics through LSTMs enable real-time, low-overhead CSI feedback suitable for practical deployment in FDD massive MIMO systems?
• RQ5 How robust is the CsiNet-LSTM framework to variations in channel conditions and compression ratios across different propagation environments?

Key findings

• CsiNet-LSTM achieves the lowest normalized mean square error (NMSE) across all compression ratios, with a 94% and 10% improvement in NMSE reduction from 1/16 to 1/64 compression ratio in indoor and outdoor scenarios, respectively.
• The beamforming gain, measured by cosine similarity (ρ), reaches 0.99 at 1/16 compression ratio in indoor scenarios and 0.95 in outdoor scenarios, significantly outperforming CS-based methods that achieve less than 0.5.
• CsiNet-LSTM maintains high reconstruction quality even at low compression ratios, with only 8% and 10% performance loss in NMSE for indoor and outdoor environments, respectively, when CR is reduced from 1/16 to 1/64.
• The framework enables real-time reconstruction, with a runtime of only 0.0003 seconds per CSI matrix on GPU, well below the 0.04-second feedback interval, making it suitable for practical deployment.
• Compared to CsiNet, CsiNet-LSTM improves NMSE by multiple times, especially at low CRs, while maintaining a modest increase in latency (from 0.0001s to 0.0003s), demonstrating a favorable trade-off between accuracy and efficiency.
• The use of codeword concatenation across time improves recovery quality for subsequent channel matrices, with the average NMSE of matrices from t=2 to T being significantly lower than the first matrix (e.g., -14.74 dB vs. -8.35 dB in indoor scenarios).