R1-2409447
discussion
Discussion on AI/ML for Beam Management
From Quectel
Summary
Quectel presents 15 proposals regarding AI/ML-based beam management in NR, focusing on data collection frameworks for network and UE-side models, inference reporting configurations, and performance monitoring mechanisms. The document addresses specification impacts for Beam Management Case 1 (BM-Case1) and Case 2 (BM-Case2), proposing enhancements to CSI-ReportConfig structures, support for aperiodic CSI-RS, and unified monitoring metrics to ensure model accuracy and reduce overhead.
Position
Quectel proposes extending UE capability for RS measurement beyond 64 RS per resource set to support larger Set A configurations, utilizing bitmaps for Set B pattern selection. They support configuring two resource sets for Set A and Set B separately within a single CSI-ResourceConfigId for inference reporting. For BM-Case2, Quectel proposes supporting Aperiodic (AP) CSI-RS and including time-based predictions with confidence levels in the inference report. They argue for a unified mechanism for measurement reporting, data collection, and monitoring to ensure consistent performance evaluation. Additionally, they propose that the network must be notified of performance degradation in UE-sided monitoring to maintain synchronized management during model transitions.
Key proposals
- Proposal 1 (Sec 2.1): Supports multiple Set B or multiple Set B patterns for training data collection in NW-side models to improve beam prediction accuracy.
- Proposal 2 (Sec 2.1): Extends UE capability for RS measurement beyond 64 RS per resource set to support larger Set A configurations, using bitmaps for Set B pattern selection.
- Proposal 3 (Sec 2.2): Allows UE to autonomously trigger data collection and report relevant configurations to ensure consistent data for UE-sided model training.
- Proposal 4 (Sec 2.2): Enables UE to request preferred Set B patterns from gNB, which pre-configures multiple patterns to minimize RS overhead during training and inference.
- Proposal 5 (Sec 3.1): Configures two resource sets for Set A and Set B separately within one CSI-ResourceConfigId for inference reporting.
- Proposal 6 (Sec 3.1): Supports Aperiodic (AP) CSI-RS in BM-Case2 to facilitate UE-side model inference with flexible timing.
- Proposal 7 (Sec 3.1): Implements Periodic, Semi-persistent, and Aperiodic reporting in different scenarios for BM-Case1 and BM-Case2.
- Proposal 8 (Sec 3.2): Includes predicted beam IDs and L1-RSRP values for Top-K beams in inference reports, with optional differential RSRP reporting.
- Proposal 9 (Sec 3.2): Incorporates time-based predictions in inference reports to anticipate beam performance over multiple time instances for high-mobility environments.
- Proposal 10 (Sec 3.2.1.2): Uses existing CSI-ReportConfig structures to support AI/ML-based beam management in BM-Case1 by reporting top-K predicted beams from Set A.
- Proposal 11 (Sec 3.2.1.2): Extends CSI-ReportConfig for BM-Case2 to include time-based predictions, time instance information, and prediction confidence levels.
- Proposal 12 (Sec 4.2.1): Defines a unified mechanism for measurement reporting for inference, data collection for training, and monitoring to ensure consistent performance evaluation.
- Proposal 13 (Sec 4.2.1): Ensures the network is notified of performance degradation in UE-sided monitoring to prevent configuration issues during model transitions.
- Proposal 14 (Sec 4.3): Implements AI/ML-driven intelligent monitoring systems that dynamically adjust monitoring frequency based on real-time network conditions.
- Proposal 15 (Sec 4.3): Utilizes feedback-driven model retraining to continuously update AI/ML models based on performance monitoring data.