Transformer Neural Network CSI Prediction

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Current methods for channel state information (CSI) feedback in wireless communications, particularly in MIMO systems, consume significant communication resources and strain wireless networks due to the need for raw CSI feedback.

Innovation Solution

A transformer-based neural network is employed for CSI prediction, where a context vector is generated from a plurality of CSI measurements using an encoder, and one or more predicted CSI values are generated from this context vector using a decoder.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If raw CSI feedback is used in MIMO systems, then channel state information accuracy is improved, but communication resource consumption increases and network strain increases

Engineering Contradiction:

Improvechannel state information accuracyVSAvoidcommunication resource consumption

Core Design Contradiction:

Measurement precisionVSQuantity of substance

Solution Approach 1:

The patent extracts only the essential features from raw CSI measurements by using a neural network to predict compressed CSI values. Instead of transmitting complete raw CSI data, the system extracts key channel state information and reconstructs it at the receiver end, significantly reducing feedback overhead while maintaining adequate accuracy for MIMO operations

Inventive Principle:

Principle #2Taking out (Extraction)

Solution Approach 2:

The patent transforms CSI representation parameters by using neural network models to map raw CSI measurements to compressed predicted values. The system changes the parameter space from high-dimensional raw CSI to lower-dimensional predicted CSI, reducing resource consumption while preserving essential channel characteristics through learned transformations

Inventive Principle:

Principle #35Parameter changes

2Reliability

If traditional CSI feedback methods are used, then communication reliability is maintained, but system latency increases and performance deteriorates

Engineering Contradiction:

Improvecommunication reliabilityVSAvoidsystem latency

Core Design Contradiction:

ReliabilityVSLoss of time

Solution Approach 1:

The patent performs preliminary CSI prediction at the receiver before feedback transmission is needed. By pre-computing compressed CSI values using neural network models, the system prepares prediction results in advance, reducing the time required for CSI feedback and improving system response speed while maintaining reliability through model-based prediction

Inventive Principle:

Principle #10Preliminary action

Data Source

PatentUS20250184775A1Method and apparatus for channel state information (CSI) prediction

Publication Date: 2025.06.05 MEDIATEK INC

AI summary

This disclosure provides an apparatus and a method for channel state information (CSI) prediction. Processing circuitry of the apparatus obtains a plurality of CSI measurements. Each CSI measurement is measured at a different time instant. The processing circuitry generates a context vector from an encoder of a transformer based neural network, based on the plurality of CSI measurements being input to the encoder of the transformer based neural network. The processing circuitry generates one or more predicted CSI values from a decoder of the transformer based neural network, based on the context vector being input to the decoder of the transformer based neural network.