R1-2500643
discussion
Specification support for AI/ML for positioning accuracy enhancement
From Sony
Summary
Sony submits a contribution for RAN1 Meeting #120 focusing on AI/ML for NR Air Interface positioning accuracy enhancements, presenting 13 proposals across data collection, model input/output, inference consistency, and performance monitoring. The document argues for supporting Channel Impulse Response (CIR) reporting with configurable sizes to manage overhead, ensuring consistency between training and inference via reference signal characteristic association, and defining specific mechanisms for model validity indications and updates.
Position
Sony proposes supporting Channel Impulse Response (CIR) reporting for data collection, specifically advocating for configurable report sizes and measurement windows to mitigate signaling overhead. They require the association of data sample parts (Part A and Part B) via timing or location information and support model transfer from the LMF to UE/gNB. Regarding inference consistency, Sony proposes defining parameters for reference signal characteristics and associating these characteristics with both the trained model and the inference operation to prevent performance degradation. They further propose studying signaling procedures for this consistency and supporting AD-IE-Group2 provision for Case 1. For performance monitoring, Sony supports Options A-1, A-2, and A-3 for Case 1 label-based monitoring and requires indications of model validity to be exchanged between the LMF and NG-RAN nodes/UEs depending on the monitoring option (A or B).
Key proposals
- Proposal 1 (Data Collection): Support radio channel characteristics reporting in the form of Channel Impulse Response (CIR), including power, time, and phase information, for AI/ML positioning.
- Proposal 2 (Data Collection): Support configurable CIR measurement reports, such as report size and measurement window size, to reduce signaling overhead.
- Proposal 3 (Data Collection): Associate Part A and Part B of collected data samples using timing information (e.g., timestamps) or location information.
- Proposal 4 (Data Collection): Support model transfer operations from network nodes (e.g., LMF) to UE/gNB to facilitate AI/ML-assisted positioning.
- Proposal 5 (Model Input): Support sample-based time domain channel measurements as the AI/ML model input for Case 2b.
- Proposal 6 (Model Output): Support an indication in positioning measurement reports (Cases 2a and 3a) specifying whether the measurement is based on AI/ML computation or legacy methods.
- Proposal 7 (Model Inference): Define a set of parameters representing reference signal characteristics (e.g., DL-PRS configuration, received signal quality) for AI/ML positioning.
- Proposal 8 (Model Inference): Support the association of reference signal characteristics to both the trained AI/ML model and the inference operation to ensure consistency.
- Proposal 9 (Model Inference): Further study the signaling procedures required to ensure consistency between AI/ML training and inference.
- Proposal 10 (Model Inference): Support the LMF providing the UE with AD-IE-Group2 (NW-side additional condition) for Case 1.
- Proposal 11 (Model Performance Monitoring): Support Options A-1, A-2, and A-3 for label-based model performance monitoring metric calculation in Case 1.
- Proposal 12 (Model Performance Monitoring): For Case 3a with Option B, support the LMF providing an indication of AI/ML model validity to the NG-RAN node.
- Proposal 13 (Model Performance Monitoring): For Case 1 and Case 3a with Option A, support the UE or gNB providing an indication of AI/ML model validity to the AI/ML server/management (e.g., LMF).