Deep Learning-Based CSI Feedback Approach for Time-Varying Massive MIMO Channels

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.

Deep Learning-Based Downlink Channel Prediction for FDD Massive MIMO System

“…Let NMSE Dt (k) represent the NMSE of the direct-transfer algorithm evaluated on the k-th target Update the parameter Ω using unbiased 1 st and 2 nd moment vectors: Ω ← Ω − ημ 1 /( √ν 1 + ε) 12 end 13 Testing stage 14 Initialize NMSE: NMSE Nt ← 0 15 for k = 1, · · · , K T do 16 Generate the testing dataset D Te (k) for T T (k) 17 Predict the downlink CSI base on D Te (k) and Ω using Eq. (8) 18 Calculate NMSE Nt (k) using Eq. (14) 19 NMSE Nt ← NMSE Nt + NMSE Nt (k)/K T 20 end environment that can be obtained by testing the network on the dataset D Te (k) using Eqs.…”

“…Specifically, Ω S,k is initialized as the current network parameter Ω, and is then updated with G Tr steps of gradient descents. Since an overlarge gradient leads to the instability of gradient descents, we limit the values of the source-task-specific gradient ∇ Ω S,k Loss DTrSup(k) (Ω S,k ) into a certain range and obtain the truncated source-task-specific gradient υ S,k as [42] [υ S,k ] p =min Υ, ∇ Ω S,k Loss DTrSup(k) (Ω S,k ) p , p = 1, · · · , len(υ S,k ), (18) where Υ is the upper threshold of the gradient. Generally, an overlarge Υ may lead to large fluctuations of the loss function, while a too small Υ distorts the direction of the update, resulting in too early convergence.…”

Deep Convolutional Compression For Massive MIMO CSI Feedback

“…The feature encoder extracts key features from the CSI matrix to obtain a lower dimensional representation, which is subsequently converted into a discrete-valued vector by applying scalar quantization. While previous works simply send the 32-bit scalar quantized version of the feature vector as the CSI feedback [11], [12], [14], we have observed that the autoencoder structure does not produce uniformly distributed feature values, and hence, can be further compressed. To further reduce the required feedback, we employ an entropy encoder; in particular, we use the context-adaptive binary arithmetic coding (CABAC) technique [26], which outputs a variable-length bit stream.…”

“…Several autoencoder-based CSI reduction techniques [11], [12], [14] focus on dimensionality reduction by direct application of the autoencoder architecture. These works are based on the assumption that reducing the dimension of the CSI matrix to be fed back to the BS would result in reduced feedback overhead.…”

“…ii) Many of the existing DL-based architectures for CSI compression mainly focus on di-mensionality reduction by direct application of the autoencoder architecture and do not consider further compression of the CSI at a bit level [11], [12], [14]. DeepCMC includes quantization and entropy coding blocks within its architecture to directly convert the channel gain matrix into bits for subsequent communication.…”