An Adaptive CSI Feedback Model Based on BiLSTM for Massive MIMO-OFDM Systems

Authors:

Hongrui Shen、Long Zhao、Kan Zheng、Yuhua Cao、Pingzhi Fan

Paper:

https://arxiv.org/abs/2408.06359

Introduction

Massive multiple-input multiple-output (MIMO) technology is a cornerstone of 5G mobile communication due to its high spectrum efficiency. For effective downlink signal transmission, accurate downlink channel state information (CSI) is crucial. In frequency division duplex (FDD) systems, obtaining downlink CSI from uplink pilots is challenging, making CSI feedback a critical area of research. Traditional methods like compressive sensing (CS) have limitations, especially with the increasing number of antennas in massive MIMO systems. This paper proposes an adaptive bidirectional long short-term memory network (ABLNet) for CSI feedback, designed to handle various input CSI lengths and feedback bit numbers efficiently.

System Model

Massive MIMO-OFDM System Model

In a massive MIMO-OFDM system, the gNB is equipped with multiple transmit antennas, and each UE has multiple receive antennas. The system divides the downlink channels into subbands, each consisting of several subcarriers. The beamformer for each subband is derived from the eigenvector of the average correlation matrix of the subband’s channels.

DL-Based CSI Feedback Model

The CSI feedback model employs an autoencoder architecture to compress and recover the CSI eigenvector. The encoder compresses the CSI eigenvector into a codeword vector, which is then quantized into a bit stream. The bit stream is transmitted to the gNB, where it is dequantized and decoded to reconstruct the CSI eigenvector.

Adaptive CSI Feedback Model

Architecture of ABLNet

ABLNet consists of an encoder and a decoder, both leveraging BiLSTM layers for feature extraction. The encoder compresses the input CSI eigenvector, while the decoder reconstructs it from the received bit stream.

Encoder

1. Feature Extraction Block: Two BiLSTM layers extract features from the input CSI eigenvector.
2. Residual Block: A fully connected layer transforms the extracted features, with a residual structure to improve performance.
3. Compression Block: Layer normalization and a fully connected layer compress the features.
4. Non-linearity Block: A Sigmoid function and flatten operation transform the compressed features into a codeword vector.

Decoder

1. Recovery Block: Reshapes and processes the codeword vector through a fully connected layer.
2. Feature Extraction Block: Two BiLSTM layers extract features from the reshaped vector.
3. Residual Block: A fully connected layer and residual structure finalize the feature extraction.
4. Output Layer: Layer normalization produces the recovered CSI eigenvector.

ABLNet for Adjustable Subband Number

ABLNet can handle different subband numbers by padding the input CSI eigenvectors to a fixed length. The feedback bit number is proportional to the subband number, ensuring efficient feedback for varying input lengths.

FBCU for Adjustable Feedback Bit Number

The Feedback Bit Control Unit (FBCU) allows the feedback bit number to be adjusted by discarding parts of the codeword vector based on feedback overhead requirements. This flexibility enhances the model’s adaptability to different scenarios.

BNA Algorithm for Target SGCS

The Bit Number Adjusting (BNA) algorithm adjusts the feedback bit number to achieve a target square of generalized cosine similarity (SGCS) for each input CSI. This ensures consistent feedback performance while minimizing feedback overhead.

UE-First Separate Training for Different Manufacturers

To address the model protection problem between different manufacturers, a UE-first separate training approach is proposed. Each UE trains its encoder independently, and the gNB trains a general decoder using the feedback from all UEs. This approach reduces the complexity and storage requirements at the gNB.

Simulation Results and Analyses

Simulation Configuration

The simulations use the clustered delay line (CDL) channel model with different subband numbers. The ABLNet model structure is detailed, and the performance is compared with other models like EVCsiNet and Transformer.

Performance of Proposed ABLNet

ABLNet outperforms other models in terms of feedback performance and model complexity, especially for lower subband numbers. The model adapts well to different input lengths and feedback bit numbers.

Performance of ABLNet with FBCU

The FBCU enhances ABLNet’s adaptability, allowing it to handle varying feedback bit numbers efficiently. The performance is compared with models using multiple decoders, showing that ABLNet with FBCU maintains high SGCS performance.

Performance of BNA Algorithm

The BNA algorithm effectively adjusts the feedback bit number to achieve the target SGCS, improving the feedback performance and reducing the average feedback overhead.

Performance of Link-Level BLER

The link-level block error rate (BLER) performance of ABLNet with FBCU is compared with traditional and other DL-based CSI feedback schemes. ABLNet with FBCU shows superior BLER performance, demonstrating its effectiveness in practical scenarios.

Performance of UE-First Separate Training

The separate training approach achieves similar performance to joint training, even with different encoder structures and CSI distributions. This validates the feasibility of the proposed training strategy.

Conclusions

The proposed ABLNet with FBCU and BNA algorithm provides an adaptive and efficient solution for CSI feedback in massive MIMO-OFDM systems. The separate training approach ensures compatibility between different manufacturers, reducing complexity and improving performance. The experimental results demonstrate the effectiveness and robustness of the proposed methods.