R1-2500590
discussion
Discussion on specification support for beam management
From NEC
Summary
NEC presents 46 proposals and 15 observations regarding AI/ML beam management for NR Rel-19, focusing on lifecycle management (LCM) signaling, model inference configurations, and performance monitoring. The document addresses specific technical gaps in UE capability reporting, resource efficiency for Set A/B measurements, and the integration of AI/ML predictions with legacy TCI states and beam failure recovery procedures.
Position
NEC proposes that UE capability reporting must include detailed timing conditions for Set B measurements and Set A predictions to enable correct model selection. They require that P/SP NZP-CSI-RS resources for Set A be available for PDSCH rate matching during inference to prevent throughput degradation. NEC supports the use of an associated ID to ensure consistency of NW-side additional conditions across multiple cells, rather than restricting them to a single cell. They propose separating CPU counting for AI/ML inference (ACPU) from legacy CSI processing and define specific performance monitoring metrics, including RSRP difference (Alt 3) and probability reporting (Alt 4). Furthermore, NEC argues for enhancing the TCI framework to support predicted beams via QCL references to similar measured patterns and proposes using single MAC CE/DCI commands to activate multiple future TCI states for BM-Case 2.
Key proposals
- Proposal 2 (Sec 2.1): The timing information of Set B measurements and Set A predictions should be specified as conditions for BM-Case1 and BM-Case2 and comprised in UE capability reporting.
- Proposal 4 (Sec 2.1): For the consistency of NW-side additional conditions across training and inference, an associated ID shall indicate site/cell specific variables like antenna configuration, allowing consistency across multiple cells.
- Proposal 7 (Sec 2.2): The configured P/SP NZP-CSI-RS resources for Set A should be available for other channels/signals (e.g., PDSCH) at least during model inference to preserve system throughput.
- Proposal 11 (Sec 2.2): For BM-Case1 and BM-Case2, the value of K (number of predicted beams to report) should be determined and provided to the NW by the UE, rather than being fixed by NW configuration.
- Proposal 14 (Sec 2.2): For BM-Case2, observation and prediction windows should be explicitly configured for P/SP beam reports, requiring measurement/reporting only in the observation window.
- Proposal 19 (Sec 2.2): Support variable Set B for model inference, either selected from pre-configured patterns or as a subset of measured beams Set C based on defined criteria.
- Proposal 24 (Sec 2.3): Support event-triggered or UE-initiated performance monitoring of the current AI/ML model if beam failure instance indications or out-of-sync indications are received.
- Proposal 27 (Sec 2.3): Support Alt 3 for performance monitoring: reporting the RSRP difference between predicted RSRP and measured L1-RSRP for corresponding beams in the monitoring resource set.
- Proposal 32 (Sec 2.3): Define performance metrics for a set of samples using prediction accuracy per instance, where accuracy is 1 if all beams are predicted correctly, or a ratio of correct beams if partial accuracy is allowed.
- Proposal 38 (Sec 2.4): Support separately counted CPUs for inference (named ACPUs) distinct from legacy CSI processing CPUs, allowing UE to report capabilities for both sets independently.
- Proposal 42 (Sec 3.1): Enhance unified TCI states to indicate predicted beams by configuring RS in associated Set A as the QCL reference signal, or using a measured beam pattern similar to the predicted beam if unmeasured.
- Proposal 43 (Sec 3.1): For BM-Case 2, support using one MAC CE or DCI to activate/indicate multiple future TCI states along with their corresponding time periods.
- Proposal 45 (Sec 3.2): Support beam failure detection based on time-domain prediction for BM-Case 2, leveraging predicted beam quality to trigger early switching or avoid false alarms.