Granted Patent B2

US 12,652,550 · App. 18/181,447 · Granted Jun 9, 2026

Inventors: Daoud Burghal (San Jose, CA); Yang Li (Plano, TX); Pranav Madadi (Sunnyvale, CA); Jeongho Jeon (San Jose, CA); Joonyoung Cho (Portland, OR); Jianzhong Zhang (Dallas, TX)

Assignee: Samsung Electronics Co., Ltd.

H04W24/02H04L1/0073

Abstract

Claims (54)

1 . A method comprising:

receiving a pilot signal or a measurement report from a user equipment (UE) or a base station (BS);

updating a CSI buffer with channel state information (CSI) obtained from the pilot signal or the measurement report, the CSI buffer configured to store previous uplink or downlink channel estimates;

providing at least a portion of the CSI buffer to a CSI predictor comprising an artificial intelligence (AI) model that utilizes one or more weight sharing mechanisms, the AI model comprising a sequence of layers; and

predicting temporal CSI using the CSI predictor,

wherein predicting the temporal CSI using the CSI predictor comprises:

for each layer of the AI model:

performing circular padding along multiple dimensions representing angles of the CSI buffer to generate a padded signal;

providing the padded signal to that layer; and

performing at least one nonlinear activation function using that layer.

2 . The method of claim 1 , wherein the CSI predictor is configured to predict the temporal CSI per path while maintaining one or more correlations in the AI model between one or more paths in one or more channels.

3 . The method of claim 2 , wherein the CSI predictor is configured to use weight sharing to predict the temporal CSI per path.

4 . The method of claim 1 , further comprising:

preprocessing the CSI before updating the CSI buffer, wherein preprocessing the CSI comprises:

transforming channel information of the CSI into an angle delay domain using at least one of: a Fourier transform, Fourier beams, a parameter extraction algorithm, or an AI based algorithm.

5 . The method of claim 1 , wherein the layers of the AI model are separable and modular such that execution of the AI model includes skipping execution of one or more layers.

6 . The method of claim 1 , wherein a final layer in the sequence of layers is configured to be adapted based on a speed or a signal-to-noise ratio (SNR) of the UE or the BS.

7 . The method of claim 1 , wherein the CSI predictor is configured to perform a denoising task.

8 . The method of claim 1 , wherein the CSI predictor is configured to perform frequency extrapolation using the CSI buffer to predict the temporal CSI on multiple frequencies.

9 . A device comprising:

a transceiver configured to receive a pilot signal or a measurement report from a user equipment (UE) or a base station (BS); and

a processor operably connected to the transceiver, the processor configured to:

update a CSI buffer with channel state information (CSI) obtained from the pilot signal or the measurement report, the CSI buffer configured to store previous uplink or downlink channel estimates;

provide at least a portion of the CSI buffer to a CSI predictor comprising an artificial intelligence (AI) model that utilizes one or more weight sharing mechanisms, the AI model comprising a sequence of layers; and

predict temporal CSI using the CSI predictor,

wherein to predict the temporal CSI using the CSI predictor, the processor is further configured to:

for each layer of the AI model:

perform circular padding along multiple dimensions representing angles of the CSI buffer to generate a padded signal;

provide the padded signal to that layer; and

perform at least one nonlinear activation function using that layer.

10 . The device of claim 9 , wherein the CSI predictor is configured to predict the temporal CSI per path while maintaining one or more correlations in the AI model between one or more paths in one or more channels.

11 . The device of claim 10 , wherein the CSI predictor is configured to use weight sharing to predict the temporal CSI per path.

12 . The device of claim 9 , wherein the processor is further configured to:

preprocess the CSI before updating the CSI buffer, wherein to preprocess the CSI, the processor is configured to:

transform channel information of the CSI into an angle delay domain using at least one of: a Fourier transform, Fourier beams, a parameter extraction algorithm, or an AI based algorithm.

13 . The device of claim 9 , wherein the layers of the AI model are separable and modular such that execution of the AI model includes skipping execution of one or more layers.

14 . The device of claim 9 , wherein a final layer in the sequence of layers is configured to be adapted based on a speed or a signal-to-noise ratio (SNR) of the UE or the BS.

15 . The device of claim 9 , wherein the CSI predictor is configured to perform a denoising task.

16 . The device of claim 9 , wherein the CSI predictor is configured to perform frequency extrapolation using the CSI buffer to predict the temporal CSI on multiple frequencies.

17 . A non-transitory computer readable medium comprising program code that, when executed by a processor of a device, causes the device to:

receive a pilot signal or a measurement report from a user equipment (UE) or a base station (BS);

update a CSI buffer with channel state information (CSI) obtained from the pilot signal or the measurement report, the CSI buffer configured to store previous uplink or downlink channel estimates;

provide at least a portion of the CSI buffer to a CSI predictor comprising an artificial intelligence (AI) model that utilizes one or more weight sharing mechanisms, the AI model comprising a sequence of layers; and

predict temporal CSI using the CSI predictor,

wherein to predict the temporal CSI using the CSI predictor, the program code, when executed by the processor of the device, further causes the device to:

for each layer of the AI model:

perform circular padding along multiple dimensions representing angles of the CSI buffer to generate a padded signal;

provide the padded signal to that layer; and

perform at least one nonlinear activation function using that layer.

18 . The non-transitory computer readable medium of claim 17 , wherein the CSI predictor is configured to predict the temporal CSI per path while maintaining one or more correlations in the AI model between one or more paths in one or more channels.

19 . The non-transitory computer readable medium of claim 17 , wherein the CSI predictor is configured to use weight sharing to predict the temporal CSI per path.

20 . The non-transitory computer readable medium of claim 17 , wherein the program code, when executed by the processor of the device, further causes the device to:

preprocess the CSI before updating the CSI buffer, wherein to preprocess the CSI, the program code, when executed by the processor of the device, further causes the device to:

transform channel information of the CSI into an angle delay domain using at least one of: a Fourier transform, Fourier beams, a parameter extraction algorithm, or an AI based algorithm.

Assignments (1)

Continuity (2)

References Cited (16)