R1-2500606
discussion
Discussion on specification support for AIML based positioning accuracy enhancement
From NEC
Summary
NEC submits 38 proposals and 2 observations focusing on AI/ML-based positioning accuracy enhancements for NR Rel-19, prioritizing first-priority cases (Case 1, 3a, 3b) while deferring second-priority cases to Rel-20. The document addresses critical aspects of model input/output definitions, training data collection quality, timestamp synchronization, and consistency mechanisms between training and inference phases.
Position
NEC proposes deferring second-priority cases (2a/2b) to Rel-20 to focus Rel-19 efforts on first-priority cases (1, 3a, 3b). For model inputs, NEC requires the LMF to determine consistent sampling parameters (Nt', k) for enhanced measurements in Case 3b and supports reporting phase information from gNB to LMF. Regarding model outputs, NEC argues that LOS/NLOS indicators are unnecessary when the model output is timing/angle information, as these are derived from a 'virtual' LOS assumption. For training data, NEC proposes defining quality indicators for power information and data samples jointly, and using 'UTC time + SFN + slot index' for timestamps. To ensure consistency between training and inference, NEC supports using associated IDs and validity areas, while opposing the explicit provision of TRP location (Alternative 4) in favor of implicit methods (Alternatives 1 and 2).
Key proposals
- Proposal 1 (Sec 2): Focus on first priority cases in Rel-19 and address second priority cases in Rel-20.
- Proposal 2 (Sec 3.1): Limit the value range of Nt' (number of samples) for enhanced measurement to less than or equal to 9.
- Proposal 4 (Sec 3.1): LMF determines consistent sampling parameters (e.g., k) for enhanced measurements in Case 3b.
- Proposal 5 (Sec 3.2): Ensure reported measurements from path-based and enhanced measurements are not in the same reporting timing granularity.
- Proposal 9 (Sec 3.3): Support phase information reported from gNB to LMF as model input for Case 3b.
- Proposal 10 (Sec 4.1): No need to report LOS/NLOS indicator to LMF if the model output is timing/angle information for Case 3a.
- Proposal 12 (Sec 5.1): Introduce a quality indicator for power information in channel measurement for generating model input.
- Proposal 15 (Sec 5.1): Define a quality indicator for a data sample based on associated measurement and ground truth label jointly.
- Proposal 18 (Sec 5.2): Use 'UTC time + SFN + slot index' as a baseline for timestamp design.
- Proposal 20 (Sec 5.3): Match Part A and Part B of training data based on timestamps, with further study on eliminating matching errors for moving UEs.
- Proposal 22 (Sec 6): Support semi-supervised learning for AI/ML based positioning.
- Proposal 23 (Sec 7.1): Support both Option A (UE-side calculation) and Option B (LMF-side calculation) for model performance monitoring in Case 1.
- Proposal 27 (Sec 8.1): Support Alternative 1 and 2 for ensuring consistency regarding TRP location, excluding Alternative 4.
- Proposal 28 (Sec 8.2): Introduce associated ID as an implicit way to ensure consistency between training and inference.
- Proposal 35 (Sec 8.2): LMF determines k, Nt', Nt, and selection rules for sample-based measurement in Case 2b.