R1-2500635
discussion
AI/ML specification support for beam management
From Lenovo
Summary
Lenovo submits 27 proposals for AI/ML-enabled beam management in NR Rel-19, addressing data collection, model inference, and performance monitoring for both UE-side and NW-side models. The document focuses on specification support for BM-Case 1 (spatial prediction) and BM-Case 2 (temporal prediction), including mechanisms for overhead reduction, consistency between training and inference, and lifecycle management.
Position
Lenovo proposes supporting UE-initiated beam management procedures for data collection to enable UE-side model training. They require combining an associated ID with performance monitoring to ensure consistency between training and inference, arguing that the associated ID alone is insufficient due to signaling overhead. For BM-Case 2, they propose specific overhead reduction techniques, including differential RSRP quantification relative to the global maximum and reporting unique beams with time-stamp indicators. They introduce the concept of AI Process Units (APUs) to manage UE hardware resources and propose refining CSI computation time to account for AI inference latency. For performance monitoring, they support Alt 2 and Alt 3 metrics, which rely on L1-RSRP differences, and propose event-triggered beam reports for hybrid monitoring scenarios.
Key proposals
- Proposal 1 (Data Collection): Support UE-initiated beam management procedures for data collection to facilitate UE-side model training.
- Proposal 2 (Data Collection): Support beam reports with variable size, particularly for NW-side model training, to manage overhead when reporting large numbers of beams.
- Proposal 3 (UE-side Inference Config): For UE-side models in BM-Case 1, support configuring only the resource set for Set B (measurement beams) for inference result reporting, rather than explicitly configuring Set A.
- Proposal 4 (Consistency): Ensure consistency between training and inference for UE-side models by combining an associated ID with performance monitoring, rather than relying solely on the associated ID.
- Proposal 5 (Dynamic Switching): Support dynamic switching between AI/ML-based beam prediction and non-AI/ML-based reporting, as well as switching between different AI/ML models/functionalities.
- Proposal 6 (Common Framework): Specify a common beam report configuration to support both BM-Case 1 and BM-Case 2 for UE-side inference.
- Proposal 8 (Prediction Window): Define the prediction window for BM-Case 2 using two modes: one based on the CSI reference resource and another where the offset is configured by RRC or determined by the UE.
- Proposal 9 (Overhead Reduction): Reduce UCI overhead for BM-Case 2 by supporting differential RSRP quantification relative to the largest value across all future instances, reporting unique beams with time-stamp indicators, and indicating whether to report all or unique beams.
- Proposal 11 (Resource Management): Introduce AI Process Units (APUs) for beam reports with AI/ML inference at the UE side to align hardware resource occupation.
- Proposal 15 (Processing Time): Refine UE CSI computation time for aperiodic reports with AI inference by adding specific AI process time components (TAI_CSI).
- Proposal 18 (NW-side Monitoring): Support Tx beam repetition for NW-side AI/ML model performance monitoring to allow the UE to report the best L1-RSRP of a Tx beam among all its Rx beams.
- Proposal 23 (Monitoring Linkage): For BM-Case 1, determine the linkage of RS resources for monitoring and inference results based on the relationship between monitoring occasions and CSI reference resources, CSI reports, or inference RS occasions.
- Proposal 25 (Performance Metrics): Support Alt 2 (L1-RSRP difference of Top K predicted beams) and Alt 3 (RSRP difference between predicted and measured L1-RSRP) as performance metrics for AI/ML model monitoring.
- Proposal 27 (Hybrid Monitoring): Support event-triggered beam reports for hybrid performance monitoring of UE-side AI/ML models.