AI-Based CSI Compression for Layer-Specific Payload Control
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In high-rank scenarios, existing channel state information (CSI) compression methods using AI models result in inconsistent accuracy and length of compression results across different layers, with some layers being inaccurately compressed or excessively long.
Innovation Solution
The terminal adjusts the encoding AI model or its output length for each layer based on first information to match the payload information, allowing flexible adjustment of compression results for different layers.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If a same encoding AI model is used to compress precoding matrices of different layers, then the compression process is simple and unified, but the accuracy of compression results for important layers is insufficient and the length of compression results for unimportant layers is excessively long
Solution Approach 1:
The patent divides the compression process into separate encoding AI models for different layer types (important layers vs. unimportant layers). This segmentation allows each model to be optimized specifically for its target layer category, improving compression accuracy for important layers while reducing the length of compression results for unimportant layers, thereby resolving the contradiction between unified process simplicity and differential performance requirements.
Solution Approach 2:
The patent applies different encoding AI models with different output lengths to different layers based on their importance. Important layers use a first encoding AI model with a first output length to ensure high compression accuracy, while unimportant layers use a second encoding AI model with a second output length to reduce compression result length. This local differentiation resolves the contradiction by tailoring the compression quality to the specific requirements of each layer type.
2Ease of operation
If the output length of encoding AI model is fixed, then the compression process is straightforward, but the reporting overheads cannot be optimized and accuracy for different layers is inconsistent
Solution Approach 1:
The patent introduces dynamic output length adjustment by configuring different output lengths for different layer types through network configuration information. The terminal adapts the output length of encoding AI models based on layer importance and network conditions, allowing the system to optimize reporting overhead efficiently while maintaining operational simplicity through automated configuration.
Solution Approach 2:
The patent changes the output length parameter of encoding AI models based on layer type and network configuration. By adjusting this key parameter dynamically, the system achieves optimized reporting overhead without sacrificing operational simplicity, as the changes are automatically managed through configuration messages from the network side.
Data Source
Figure 1~2
Figure 3~5
Figure 6~8
AI summary
This application discloses an information processing method, an information transmission method, an apparatus, a terminal, and a network-side device, and belongs to the field of communication technologies. The information processing method in embodiments of this application includes: The terminal obtains first information, where the first information indicates payload information of first channel characteristic information corresponding to at least one layer; and the terminal performs, based on the first information, first processing on channel information of a target layer by using a target first AI unit, to obtain the first channel characteristic information corresponding to the target layer, where the payload information of the first channel characteristic information corresponding to the target layer matches an output length of the target first AI unit, and the at least one layer includes the target layer.