Two-Stage ML Channel State Feedback Compression

Solution Overview

Problem

Current wireless communication systems face inefficiencies in channel state feedback, leading to increased latency and power consumption due to the transmission of significant bits for channel state information (CSI) feedback, particularly when reporting both precoder channel matrix and singular vectors.

Innovation Solution

Implementing a two-stage machine learning-based approach using a non-Discrete Fourier Transform (DFT) codebook to compress channel state feedback, where a user equipment (UE) selects a non-DFT codebook, determines singular vectors, and compresses them using machine learning models to reduce the number of bits transmitted.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If traditional channel state feedback methods are used to report precoder channel matrix and singular vectors, then communication performance is maintained, but the number of feedback bits increases leading to increased latency and power consumption

Engineering Contradiction:

Improvechannel state feedback accuracyVSAvoidfeedback latency

Core Design Contradiction:

Measurement precisionVSLoss of time

Solution Approach 1:

The patent extracts only the essential channel state information by performing singular value decomposition and selecting only the dominant singular vectors that capture the most significant channel characteristics. This extraction approach removes redundant information while preserving the critical feedback needed for accurate channel state representation, thereby reducing feedback bits and latency without sacrificing communication performance.

Inventive Principle:

Principle #2Taking out (Extraction)

Solution Approach 2:

The patent segments the channel state feedback into two distinct parts: precoder channel matrix information and singular vector information. By dividing the feedback into separate reportable components with different priorities and accuracies, the system can optimize the feedback quantity for each segment, reducing overall feedback overhead while maintaining necessary precision for accurate channel state representation.

Inventive Principle:

Principle #1Segmentation

2Measurement precision

If traditional channel state feedback methods are used to report precoder channel matrix and singular vectors, then communication performance is maintained, but the number of feedback bits increases leading to increased power consumption

Engineering Contradiction:

Improvechannel state feedback accuracyVSAvoidpower consumption

Core Design Contradiction:

Measurement precisionVSUse of energy by moving object

Solution Approach 1:

The patent extracts only the essential channel state information by performing singular value decomposition and selecting only the dominant singular vectors that capture the most significant channel characteristics. This extraction approach removes redundant information while preserving the critical feedback needed for accurate channel state representation, thereby reducing feedback bits and latency without sacrificing communication performance.

Inventive Principle:

Principle #2Taking out (Extraction)

Solution Approach 2:

The patent applies partial action by reporting only the most significant singular vectors rather than all possible channel state parameters. By providing just enough feedback information to maintain communication performance without reporting excessive details, the system reduces the computational and transmission power required for channel state feedback while preserving adequate accuracy for effective communication.

Inventive Principle:

Principle #16Partial or excessive action

3Loss of time

If machine learning models are used to compress channel state feedback, then feedback bits are reduced, but system complexity increases

Engineering Contradiction:

Improvefeedback latencyVSAvoidmachine learning model complexity

Core Design Contradiction:

Loss of timeVSDevice complexity

Solution Approach 1:

The patent changes the parameters of the machine learning models by using pre-trained models with fixed architectures and optimizing only specific parameters like the number of singular vectors to report and codebook selection parameters. This approach reduces model complexity while maintaining compression effectiveness, as the models leverage pre-learned features rather than requiring complex real-time training and parameter optimization.

Inventive Principle:

Principle #35Parameter changes

Data Source

PatentUS20240421875A1Two stage machine learning based channel state feedback

Publication Date: 2024.12.19 QUALCOMM INC

AI summary

Methods, systems, and devices for wireless communication are described. A user equipment (UE) may select a non-Discrete Fourier Transform (non-DFT) codebook of a set of non-DFT codebooks associated with a channel state feedback message. The UE may determine a set of singular vectors associated with the non-DFT codebook based on a first machine learning model, where the set of singular vectors corresponds to a subspace associated with the non-DFT codebook. The UE may compress the non-DFT codebook, the set of singular vectors, or both, based on a second machine learning model. The UE may transmit the channel state feedback message including the compressed non-DFT codebook, the compressed set of singular vectors, or both.