CSI Autoencoder Adaptation Layer for Variable Antenna Configurations
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing AI-based CSI feedback solutions require consistent antenna port and sub-band configurations with the training dataset, limiting flexibility and generalization performance.
Innovation Solution
The proposed method includes a CSI feedback method using an adaptation layer-based CSI autoencoder, which processes CSI data with variable dimensions by designing an adaptation layer. This allows the AI-based CSI feedback solution to adapt to multiple physical resource configurations at a relatively small model update cost.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the CSI autoencoder is trained on a large-scale dataset with fixed antenna port and sub-band configurations, then the model achieves good CSI feedback performance on the training configuration, but the model cannot be used for CSI feedback when the actual configuration differs from the training dataset, reducing flexibility
Solution Approach 1:
The encoder is segmented into a first network (higher level) and an adaptation layer (lower level). The adaptation layer is specifically designed to handle configuration variations, while the first network focuses on the core CSI feedback function. This segmentation allows the model to maintain good performance on training configurations while adapting to different actual configurations through the adaptation layer.
Solution Approach 2:
The adaptation layer is designed to dynamically adjust to different configurations. The input dimension of the adaptation layer is the output dimension of the first network, and the output dimension is determined according to the length of the bit stream that needs to be fed back under the target configuration. This dynamic structure enables the model to adapt to various antenna port and sub-band configurations without requiring complete retraining.
2Adaptability or versatility
If multiple encoder models are deployed to support different physical resource configurations, then the system can adapt to various configurations, but the storage overhead and model update complexity increase significantly
Solution Approach 1:
The encoder is designed with a universal first network that can handle the core CSI processing function across all configurations. The adaptation layer serves as a configurable interface that adapts to different target configurations. This universal design allows a single encoder model to support multiple physical resource configurations, eliminating the need to deploy and maintain multiple separate encoder models for different configurations.
3Adaptability or versatility
If the model is updated frequently to adapt to configuration changes, then the system maintains good performance across different configurations, but the air interface overhead and network resource consumption increase
Solution Approach 1:
The adaptation layer is pre-designed with the capability to handle configuration variations. The input dimension is set to the output dimension of the first network, and the output dimension can be determined according to the target configuration requirements. This preliminary design eliminates the need for frequent model updates and reduces air interface overhead, as the adaptation layer can directly adapt to different configurations without requiring network resource consumption for model retraining and distribution.
Data Source
Figure 1~3
Figure 4~6
Figure 7
AI summary
There are a channel state information (CSI) feedback method, a terminal device and a network device, and the method includes: processing, by a terminal device, first input information to obtain target input information, where the first input information is CSI data obtained based on a target configuration, and a dimension of the target input information is the same as a dimension of CSI data obtained based on a reference configuration; encoding the target input information based on an encoder to obtain a target bit stream; and transmitting the target bit stream to a network device; where the encoder includes an adaptation layer, an input dimension of the adaptation layer is an output dimension of a first network, an output dimension of the adaptation layer is determined according to a length of a bit stream that needs to be fed back under the target configuration, and the first network is a higher level network of the adaptation layer in the encoder.