# US 2025/0096872 A1 · Csi compression and feedback method based on deep learning

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2025/0096872 A1
- Document ID: US20250096872A1
- Status: Published application
- Kind: A1
- Application number: 18/969,330
- Filed: 2024-12-05
- Published: 2025-03-20
- Earliest claimed: 2022-09-01
- Assignees: Zhejiang University
- Applicants: Zhejiang University
- Inventors: Zhaoyang ZHANG; Zirui CHEN
- Web page: https://patentgrep.com/patent/US20250096872A1

## CPC

- H04B 7/0626: {Channel coefficients, e.g. channel state information [CSI]}
- H04L 5/0007: {the frequencies being orthogonal, e.g. OFDM(A) or DMT}
- H04L 41/16: using machine learning or artificial intelligence
- H04B: TRANSMISSION
- H04L: TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION (arrangements common to telegraphic and telephonic communication H04M)

## Abstract

A channel state information CSI compression and feedback solution based on deep learning. In a large-scale multiple-input-multiple-output MIMO system, CSI is very important for the base station to achieve high performance gain. In a frequency division duplex FDD mode, user equipment UE needs to estimate CSI and then feed it back to the base station. Effective CSI compression will greatly reduce the cost of feedback. The present disclosure uses a two-dimensional sequence-to-sequence 2D Seq2Seq structure neural network to realize effective CSI compression and reconstruction. The base station uses CSI data to train the parameters of the neural network, which, after training, are deployed at the client and the base station for effective CSI compression and reconstruction. The present disclosure can realize high-precision compression and reconstruction of CSI, and provides an efficient and stable channel state information feedback method for a MIMO system.

## Claims

### Claim 1

A CSI compression and feedback method based on deep learning, comprising:
step S1, establishing and training a two-dimensional sequence-to-sequence structure neural network;
sub-step S1.1, collecting, by a base station, CSI data from historical communication data as training data of the two-dimensional sequence-to-sequence structure neural network;
sub-step S1.2, building, by the base station, the two-dimensional sequence-to-sequence structure neural network, and randomly initializing parameters of the two-dimensional sequence-to-sequence structure neural network, wherein the two-dimensional sequence-to-sequence structure neural network comprises an encoder and a decoder, the encoder comprises first cells and first sub-blocks, the decoder comprises second cells and second sub-blocks, the encoder comprises one two-dimensional recurrent neural network and one fully connected network in a cascade connection, and the decoder comprises one fully connected network and one two-dimensional recurrent neural network in a cascade connection; inputting, by the encoder, the first sub-blocks divided from CSI into the first cells, concatenating outputs of the first cells into a first matrix according to an order of the first cells, and compressing the first matrix into a first vector; and reshaping, by the decoder, the compressed first vector into a second matrix, dividing the second matrix into the second sub-blocks, input the second sub-blocks into the second cells, and concatenating outputs of the second cells according to an order of the second cells to obtain reconstructed CSI; and
sub-step S1.3, training, by the base station, the two-dimensional sequence-to-sequence structure neural network in the sub-step S1.2 by using the collected CSI data, and continuously training the parameters of the two-dimensional sequence-to-sequence structure neural network by a gradient descent method until convergence;
step S2, distributing, by the base station, the encoder of the two-dimensional sequence-to-sequence structure neural network trained in the sub-step S1.3 to a client;
step S3, compressing, by the client, downlink CSI by using the encoder distributed in the step S2, and feeding the compressed downlink CSI back to the base station; and
step S4, reconstructing, by the base station, the compressed downlink CSI fed back by the client in the step S3 by using the decoder of the two-dimensional sequence-to-sequence structure neural network trained in the sub-step S1.3 to obtain complete downlink CSI.

### Claim 2 (depends on 1)

The CSI compression and feedback method based on deep learning according to claim 1, wherein the two-dimensional recurrent neural network uses a two-dimensional long short-term memory network.

### Claim 3 (depends on 2)

The CSI compression and feedback method based on deep learning according to claim 2, wherein a CSI division process in the sub-step S1.2 comprises:
dividing the CSI into two-dimensional sequences comprising several sub-blocks for Nt×Ne-dimensional multi input MIMO-OFDM CSI, where Nt represents a number of base station antennas and Ne represents a number of subcarriers.

### Claim 4 (depends on 3)

The CSI compression and feedback method based on deep learning according to claim 3, wherein a working process of the encoder comprises:
inputting each sub-block obtained after dividing the MIMO-OFDM CSI into a first cell corresponding to the two-dimensional long short-term memory network of the encoder, wherein an output size of the first cell in the two-dimensional long short-term memory network is same as a dimension of the first sub-blocks; concatenating the outputs of all the first cells into a first matrix with a dimension size of Nt×Ne according to the order of the first cells, and stretching the first matrix into a first vector with a length of Nt×Ne and inputting the first vector with the length of Nt×Ne into a fully connected layer to obtain a compressed vector hcom.

### Claim 5 (depends on 3)

The CSI compression and feedback method based on deep learning according to claim 3, wherein a working process of the decoder comprises:
inputting the compressed vector hcom compressed by the encoder into a fully connected layer in the decoder to obtain a second vector with a length of Nt×Ne, reconstructing the second vector with the length of Nt×Ne into a second matrix with a dimension size of Nt×Ne, dividing the second matrix into the second sub-blocks in a same way as the CSI division process, sequentially inputting the second sub-blocks into the second cells corresponding to the two-dimensional long short-term memory network of the decoder, and concatenating the outputs of all the second cells into the second matrix with the dimension size of Nt×Ne according to the order of the second cells to obtain reconstructed MIMO-OFDM CSI.

### Claim 6 (depends on 1)

The CSI compression and feedback method based on deep learning according to claim 1, wherein in a training method for the two-dimensional sequence-to-sequence structure neural network in the sub-step S1.3, a loss function Loss (Θ) of the two-dimensional sequence-to-sequence structure neural network is defined as follows:
Loss
⁢
(
Θ
)
=
1
N
⁢
∑
n
=
1
N

H
n
-
H
n

2
2
,
where Θ represents a parameter set of the two-dimensional sequence-to-sequence structure neural network, N represents a number of the training data, Hn represents a nth original training data, and Hn represents an output obtained by inputting Hn into the two-dimensional sequence-to-sequence structure neural network.

### Claim 7 (depends on 1)

The CSI compression and feedback method based on deep learning according to claim 1, wherein the two-dimensional sequence-to-sequence structure neural network in the sub-step S1.3 is trained based on a optimizer based on gradient descent, comprising: performing, by the optimizer, gradient descent by deriving the loss function and back propagating the gradient to generate a gradient corresponding to the parameters of the two-dimensional sequence-to-sequence structure neural network, to continuously optimize the parameters of the two-dimensional sequence-to-sequence structure neural network until the convergence.
