Source-linked AI summary
Indirect Estimation of SINR via SSB and CSI-RS RSRP in 5G NR
Leonardo Spampinato, Mahamadou Togola, Matteo Bernabè, Azim Akhtarshenas, Lorenzo Mario Amorosa, David López-Pérez
TL;DR
Beam-based NR makes UE downlink SINR difficult to predict from reference-signal measurements, although such forecasts matter for proactive network control. The paper trains supervised models on SSB and CSI-RS RSRP measurements and finds that selecting active CSI-RS beams improves prediction accuracy while reducing input dimensionality.
Problem
Predicting UE downlink SINR is difficult because it depends on beam choices, neighboring-cell activity, and interference, while data-channel beams differ from SSB beams.
Source: §I · p. 1Method
The paper formulates average downlink SINR prediction as supervised learning from SSB and CSI-RS RSRP measurements, comparing feature representations that include active-beam filtering.
Source: B. Contribution · p. 2, A. SINR Predictor Input Feature · p. 4, 3) Top-K Active · p. 4Results
CSI-RS Top-K Active filtering outperforms unfiltered and strength-based inputs, and Kcsi = 16 with Kssb = 32 minimizes estimation error while limiting communication overhead.
Source: C. Results Discussion · pp. 5–6, §V · p. 6Takeaways & Limitations
Activity-aware reference-signal features support proactive radio-metric estimation in the evaluated setting.
Source: §V · p. 6
Abstract
from arXiv · showhide
Predicting user equipment (UE) performance is essential for proactive network control, resource management, and digital twin sandboxes. However, the inherent flexibility and complexity of beam-based 5G new radio (NR) networks make accurate performance forecasting highly challenging. This paper proposes a data-driven approach to predict the average downlink signal-to-interference-plus-noise ratio (SINR) relying exclusively on standardized reference-signal measurements, namely synchronization signal block (SSB) and channel state information-reference signal (CSI-RS) reference signal received power (RSRP). We formulate this prediction as a supervised learning problem and evaluate various input feature representations using a third generation partnership project (3GPP)-compliant synthetic dataset. Our analysis reveals that filtering measurements based on active CSI-RS beams significantly enhances prediction accuracy while reducing input dimensionality. This activity-aware strategy demonstrates the strong viability of machine learning models for proactive network optimization.
I. INTRODUCTION
The paper addresses the difficulty of predicting UE downlink SINR in flexible, beam-based NR networks and proposes supervised estimation from standardized reference-signal measurements. It evaluates activity-aware feature selection using a 3GPP-compliant synthetic dataset and reports improved accuracy with reduced input dimensionality.
I. INTRODUCTION
- Predicting data-channel SINR is difficult because it depends on neighboring-cell interference and data scheduling, while SSB beams are not the beams used for data transmission.CSI-RS measurements provide a more direct indication of beam-domain conditions relevant to data scheduling.Source: §I · p. 1
A. Related Work
- The paper frames UE performance prediction under prospective serving configurations as relevant to mobility management and energy-saving operation.These settings require estimating performance under radio conditions that may not yet be directly observable.Source: A. Related Work · p. 1
B. Contribution
- The study formulates downlink SINR estimation from reference-signal measurements as supervised learning and compares unfiltered, Top-K strongest, and Top-K active CSI-RS inputs.Its synthetic dataset includes SSB-RSRP, CSI-RS-RSRP, CSI-RS beam activity, and PRB-level SINR in a multi-cell mMIMO deployment.Source: B. Contribution · p. 2
- The authors show that CSI-RS beam-activity-aware feature selection significantly improves prediction accuracy while reducing the dimensionality of the input representation used by the learning model.The evaluation uses a 3GPP-compliant synthetic dataset containing SSB-RSRP, CSI-RS-RSRP, CSI-RS beam activity information, and PRB-level SINR values in a multi-cell mMIMO deployment.Source: B. Contribution · p. 2
II. SYSTEM MODEL
- The system model considers a sub-6 GHz FR1 network at 3.5 GHz and specifies a 100 MHz channel bandwidth corresponding to 273 PRBs at 30 kHz subcarrier spacing.The system-model passage identifies 3GPP-adopted models.Source: §II · p. 2
1) Network and User Layout:
The simulated network uses a 19-site, two-tier hexagonal layout with uniformly distributed single-antenna UEs. Its cells use planar antenna arrays, and its channel model combines large- and small-scale fading.
1) Network and User Layout:
- The outdoor urban network comprises 19 sites in a two-tier hexagonal grid, with three sectors per site and uniformly distributed single-antenna UEs.The set U denotes all active UEs.Source: 1) Network and User Layout · p. 2
2) Cell Antenna Array Unit:
- Each cell uses a 32-element vertically polarized uniform planar array arranged in four rows and eight columns, with one antenna per UE.The elements are individually connected to transceivers and spaced at half a carrier wavelength.Source: 2) Cell Antenna Array Unit · p. 2
3) Channel Model:
- The channel model adopts 3GPP urban-macro propagation with large-scale fading and spatially correlated shadowing across UEs.LoS and NLoS shadow-fading standard deviations are 4 dB and 6 dB, respectively.Source: 3) Channel Model · pp. 2–3
- Small-scale fading is modeled as a Rician channel combining deterministic LoS and stochastic NLoS components, with the LoS phase described using a plane-wave approximation.The wave vector represents phase variation along three spatial directions.Source: 3) Channel Model · p. 3
A. NR Initial Access
During initial access, cells transmit SSBs to support synchronization and cell selection. Multiple SSB beams cover spatial directions, and UEs measure their RSRP to select a serving cell.
A. NR Initial Access
- SSBs carry reference signals for UE synchronization, cell selection, and establishing a connection for data transmission.NR can beamform multiple SSBs in different spatial directions.Source: A. NR Initial Access · p. 3
- In FR1, a cell can transmit eight SSB beams, which are sent sequentially through a predefined beam-sweeping pattern rather than simultaneously.Each beam is represented by a complex codeword generated via two-dimensional DFT.Source: A. NR Initial Access · p. 3
- Each UE measures the RSRP of transmitted SSB beams and selects the serving cell based on those measurements.The paper defines the RSRP of an SSB beam from a cell.Source: A. NR Initial Access · p. 3
2 P ssb
The passage identifies SSB beam transmission power and states that a UE selects the cell with the highest SSB-beam RSRP.
2 P ssb
- The SSB-beam RSRP expression includes the beam's transmission power.The passage identifies this power as P_ssb for beam s from cell c.Source: §2 · p. 3
- The RSRP expectation is taken over multiple channel realizations and the resource elements assigned to reference signals.Source: §2 · p. 3
- A UE selects its serving cell as the one yielding the highest SSB-beam RSRP.Source: §2 · p. 3
B. Data Precoding and PRB SINR
The system uses Type-I CSI-RS codebook feedback to select per-UE precoders, then computes each UE’s data SINR on each PRB from useful received power, interference, and noise. Its model assumes equal long-term resource allocation and single-layer transmission.
B. Data Precoding and PRB SINR
- Type-I CSI-RS codebook feedback supplies per-UE, per-PRB channel estimates; the PMI identifies the serving-cell beam with the largest received power for precoding.CSI-RS beams use a fixed codebook, and their RSRP is resolved per PRB.Source: B. Data Precoding and PRB SINR · p. 3, §2 · p. 3
2 P csi
- The model assumes PRBs are fully reused across beams, equally divided among UEs served per beam, and each UE is served by one beam.The resource-allocation assumption corresponds to equal long-term allocation, as in round-robin scheduling.Source: §2 · pp. 3–4
- Per-PRB data SINR accounts for useful power from the selected beam, intra-cell inter-beam interference, neighboring-cell interference, and receiver noise.The precoding codeword and transmit power assigned to each UE on each PRB determine the selected-beam contribution.Source: §2 · p. 4
III. PROBLEM FORMULATION
The paper estimates each UE’s average SINR across scheduled PRBs from SSB and CSI-RS beam-level RSRP reports. It trains a parameterized estimator by minimizing mean squared error against target average SINR across UEs.
III. PROBLEM FORMULATION
- The target is each UE’s average SINR across its scheduled PRBs, estimated from serving- and neighboring-cell SSB and CSI-RS beam-level RSRP reports.The setup assumes access to the complete per-beam measurement report, although practical deployments typically report only a subset.Source: §III · p. 4
- The estimator’s parameters are chosen to minimize mean squared error between estimated and target average SINR across UEs.Each UE’s input is its associated feature vector.Source: §III · p. 4
A. SINR Predictor Input Feature
The predictor’s input representation is a key design choice, with unfiltered and ranked Top-K alternatives constructed from beam measurement reports. Ranking strongest reports reduces input size and places strong serving and interfering beams first.
A. SINR Predictor Input Feature
- The predictor’s input feature vector is the design choice varied across the three representations: Unfiltered, Top-K Strongest, and Top-K Active.The section introduces these alternatives for constructing the UE-specific vector.Source: A. SINR Predictor Input Feature · p. 4
1) Unfiltered
- The Unfiltered representation concatenates all serving- and neighboring-cell reports in a fixed order for every user.It uses all SSB and CSI-RS measurement reports.Source: 1) Unfiltered · p. 4
2) Top-K Strongest
- Top-K Strongest sorts reports by RSRP magnitude and retains the selected number of strongest SSB and CSI-RS measurements.This reduces input cardinality from N_ssb + N_csi to K_ssb + K_csi and places typically strong serving and primary interfering beams first.Source: 2) Top-K Strongest · p. 4
3) Top-K Active:
Top-K Active filters CSI-RS measurements using whether a beam serves at least one UE, then selects the strongest remaining CSI-RS reports. SSB measurements are excluded from the activity mask because SSB beams are assumed to use orthogonal resources to data.
3) Top-K Active
- A CSI-RS beam is marked active when it serves at least one UE, and the mask filters CSI-RS measurements before the Top-K Active feature is formed.The feature ranks the masked CSI-RS vector and retains its strongest entries alongside SSB reports.Source: 3) Top-K Active · p. 4
- A strong interfering-beam measurement need not represent realized interference when that CSI-RS beam is inactive.Activity filtering distinguishes measured beam strength from whether the beam transmits data.Source: 3) Top-K Active · p. 4
- SSB transmissions are excluded from the activity mask because the setup assumes their resources are orthogonal to data transmissions.Under this assumption, SSB beams do not generate data-channel interference.Source: 3) Top-K Active · p. 4
IV. NUMERICAL RESULTS
The numerical evaluation uses synthetic UMa simulations generated with Giulia. The dataset pipeline selects serving cells and beams, schedules PRBs, and labels each UE by its average scheduled-PRB SINR.
IV. NUMERICAL RESULTS
- The dataset was generated with Giulia, configured to follow the 3GPP UMa model in TR38.901.Source: §IV · p. 4
A. Synthetic Dataset Generation
- Across N independent simulations, users are placed uniformly at random and connect to the cell with the strongest measured SSB-RSRP.Source: A. Synthetic Dataset Generation · p. 4
- Beam refinement selects the strongest serving-cell CSI-RS RSRP, whose beam is used for data transmission as cells schedule UEs across PRBs and beams.Source: A. Synthetic Dataset Generation · p. 5
- For each user, the dataset records all-cell SSB and CSI-RS RSRPs, scheduled PRBs, per-PRB SINRs, and the CSI-RS activity mask; the target is average SINR over scheduled PRBs.Source: A. Synthetic Dataset Generation · p. 5
B. SINR Predictor Training
The predictor is trained on a simulated dataset, and Top-K Active inputs outperform the other feature alternatives. The results identify useful input sizes and show where prediction errors are greater.
B. SINR Predictor Training
- The estimator is a three-layer MLP trained to minimize mini-batch MSE, with early stopping after 100 epochs without improved validation loss.The dataset is split 70/15/15; features are standardized using training-set statistics, and training is capped at 500 epochs.Source: B. SINR Predictor Training · p. 5
C. Results Discussion
- Top-K Active outperforms Unfiltered, Top-K Strongest, and No SSB; the unfiltered input fails to train an effective predictor.Top-K Active supplies ordered RSRP values for active beams, while Top-K Strongest cannot indicate which strong beams affect SINR; removing SSBs leads to earlier overfitting and worse performance.Source: C. Results Discussion · p. 5, Figure 2 · p. 5
- Prediction errors are higher in the lower-SINR regime than at high SINR, and the growing gap between median and mean absolute errors indicates strong outliers.The SINR distribution has an approximately 1 dB mean and 6 dB standard deviation.Source: C. Results Discussion · pp. 5–6, Figure 3 · p. 5
- Kcsi = 16 jointly minimizes test RMSE and inter-cell communication overhead; smaller values omit active interferers, while larger values yield diminishing returns.Figure 4 compares test RMSE across the number of ordered top-active CSI-RS RSRP inputs.Source: C. Results Discussion · p. 6, Figure 4 · p. 6
- With SSB-only inputs, prediction error ranges up to 2.9 dB and reaches a 2.4 dB minimum at Kssb = 32.CSI-RS beams carry data and thus play a more prominent role in SINR prediction; increasing Kssb beyond the optimum adds measurement and feedback burden with diminishing returns.Source: C. Results Discussion · p. 6, Figure 5 · p. 6
V. CONCLUSIONS
The conclusion presents supervised learning from standardized SSB and CSI-RS RSRP as a way to predict average downlink SINR. It highlights activity-aware filtering and tuned input sizes, while identifying real-world measurements and digital-twin control as future work.
V. CONCLUSIONS
- The paper introduces supervised prediction of average downlink SINR using reported standardized SSB and CSI-RS RSRP measurements.Source: §V · p. 6
- Top-K Active filtering outperforms unfiltered and strength-based inputs, with Kcsi = 16 and Kssb = 32 identified as input sizes that minimize error while limiting communication overhead.Source: §V · p. 6
- Future work will use real-world measurements and digital-twin sandboxing to simulate and optimize closed-loop control policies, including predictive handover and dynamic energy saving.Source: §V · p. 6