CSI Feedback Evaluation for AI Model Reliability in New Radio
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In New Radio systems, AI/ML-based CSI compression feedback algorithms may perform poorly under varying channel conditions, leading to incorrect CSI information at the base station and adversely affecting downlink transmission performance, necessitating a method to evaluate and monitor their performance.
Innovation Solution
A performance evaluation method and apparatus that assesses the performance of AI/ML models by comparing first and second CSI metrics, using various evaluation methods and metrics such as relative relationships, ratios, and thresholds to determine the effectiveness of CSI feedback algorithms.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If AI/ML-based CSI compression feedback algorithms are used, then CSI feedback overhead is reduced and feedback accuracy is improved, but system performance becomes unpredictable and may deteriorate under varying channel conditions
Solution Approach 1:
The patent implements a performance evaluation mechanism that continuously monitors CSI feedback quality by comparing model-generated CSI with reference CSI. When performance degradation is detected through metrics like NMSE or correlation coefficients, the system triggers retraining or model switching, creating a closed-loop feedback system that maintains reliability while using compression algorithms
Solution Approach 2:
The system performs preliminary performance evaluation and channel condition assessment before deploying AI/ML-based CSI feedback. By evaluating channel variability and comparing it against training conditions in advance, the system proactively determines whether the compressed feedback algorithm will be reliable, preventing performance deterioration before it occurs
2Measurement precision
If model-based CSI compression feedback is deployed, then feedback accuracy is improved under certain conditions, but the system cannot adapt to varying channel conditions
Solution Approach 1:
The patent implements dynamic model selection and adaptive retraining mechanisms. The system continuously evaluates channel conditions and selects the most appropriate pre-trained model from multiple candidates, or triggers retraining when channel characteristics diverge significantly from training conditions. This dynamic adaptation maintains high CSI feedback accuracy across varying channel environments
Solution Approach 2:
The system changes operational parameters by switching between different AI/ML models with varying complexity and compression ratios based on channel conditions. When channel variability increases, the system may switch to models trained on more diverse data or adjust compression parameters to maintain accuracy, thereby adapting to different channel scenarios while preserving measurement precision
3Loss of information
If AI/ML models are used for CSI compression, then feedback overhead is reduced, but evaluation and monitoring of model performance becomes necessary and complex
Solution Approach 1:
The patent introduces reference CSI as an intermediary for performance evaluation. Instead of complex direct verification of model accuracy, the system uses pre-computed reference CSI (generated through traditional methods or ground truth) as a mediator to indirectly assess model performance through simple metric comparisons like NMSE or correlation, simplifying the evaluation process while maintaining effectiveness
Solution Approach 2:
The system creates simplified copies of the performance evaluation functionality at both terminal and network sides. Rather than implementing complex centralized evaluation, identical lightweight evaluation modules are deployed locally, each independently computing simple metrics and making decisions, thereby reducing overall system complexity while enabling comprehensive monitoring
Data Source
Figure 1~3
Figure 4~5
Figure 6~7
AI summary
Disclosed are a performance evaluation method and apparatus, and a readable storage medium, which relate to the technical field of communications. The method includes: evaluating performance of a target object based on a first CSI and a second CSI, where the target object includes a target model and/or a first method; the target model includes one or more of a first model and a second model; the first method includes a method for determining and/or feeding back CSI; and the first CSI is obtained based on the first model or the second model.