Source-linked AI summary
Data-Driven Case Study of gNB Placement Optimization in a Private Indoor 5G Testbed
Diogo de O. Soares, Victor F. Monteiro, Fco. Rodrigo P. Cavalcanti, Vicente A. de Sousa, J. Pedro B. Lima
TL;DR
Indoor radio planning is difficult because structural obstacles limit the accuracy of standardized propagation models for specific environments. The paper trains a LightGBM model on measured RSRP using distance and wall count, then searches gNB placements under coverage and max-min objectives. The framework identifies configurations that trade larger covered area against better worst-case and cell-edge conditions, with prediction error near 7 dB in the studied office scenario.
Problem
Standardized propagation models may inadequately represent specific indoor deployment environments, while measurement-driven approaches often lack validation in operational wireless settings.
Method
A LightGBM propagation model trained on experimental RSRP measurements using distance and wall count is integrated with combinatorial search to evaluate alternative gNB placements.
Results
The framework produces a persistent trade-off: coverage maximization serves a larger area, while max-min optimization improves 10th-percentile and worst-case RSRP; the model achieves RMSE close to 7 dB.
Takeaways & Limitations
The workflow supports indoor radio planning by systematically evaluating deployments according to either covered-area or challenging-region service objectives.
Abstract
from arXiv · showhide
Accurate radio planning is a fundamental requirement for the deployment of wireless networks in indoor environments, where signal propagation is strongly affected by walls, partitions, and other structural obstacles. Despite the availability of standardized propagation models, their ability to represent the characteristics of specific deployment scenarios is often limited, motivating the use of measurement-driven approaches. In this context, this paper presents a data-driven case study of next generation NodeB (gNB) placement optimization in an office using measurements collected from an experimental fifth generation (5G) testbed. A propagation model is trained from reference signal received power (RSRP) measurements using distance and wall count as input features and integrated with a combinatorial search framework. The proposed workflow is used to evaluate alternative deployment strategies under different optimization criteria. Results indicate that satisfactory indoor coverage and improved cell-edge conditions can be achieved with a small number of gNBs.
I. INTRODUCTION
Indoor private 5G deployment requires accurate radio planning because walls and structural obstacles strongly affect propagation, while standardized models may inadequately represent specific environments. This study addresses that gap with measurements from a private 5G testbed and a data-driven gNB placement workflow.
- Indoor private 5G deployment depends on accurate radio planning because walls, partitions, furniture, and other obstacles strongly influence signal propagation.
- Standardized indoor propagation models are widely used, but their accuracy may be limited for specific deployment scenarios.
- Measurement-driven models can learn propagation relationships from empirical radio data, addressing effects that analytical formulations may not fully represent.
- The study trains a LightGBM propagation model from RSRP measurements collected in a private 5G office testbed using geometry- and obstacle-based features.
- The trained model is combined with combinatorial search to evaluate gNB placements for coverage maximization and max-min signal optimization.
II. EXPERIMENTAL SETUP
The study uses a standalone indoor 5G testbed and two office measurement campaigns designed to capture complementary obstruction and visibility conditions. The resulting dataset combines measurements across diverse distances and propagation environments.
- The radio access link operates at 3.7 GHz in 5G NR band n78, enabling RSRP collection from the experimental deployment.
- The office scenario spans approximately 35.64 × 10.00 m and includes interconnected spaces separated by internal partitions.
- The two campaigns place the gNB near obstacles and in a more open area, respectively, capturing obstructed links and stronger line-of-sight conditions.
- The combined campaigns produced 280 RSRP samples, including 147 from the first campaign and 133 from the second.
- A two-dimensional floor-plan model maps 39 wall segments and identifies obstacles intersecting direct gNB–UE paths.
III. DATA-DRIVEN PROPAGATION MODELING
The propagation model represents each gNB–UE pair using distance and wall count, then learns the relationship between these features and measured RSRP with LightGBM. This representation supports estimates for previously unobserved gNB locations.
- The model uses physically meaningful descriptors to avoid overdependence on absolute coordinates and reduce the risk of spatial overfitting.
- Each gNB–UE pair is represented by Euclidean distance and the number of walls intersecting the direct propagation path.
- The target variable is the measured RSRP associated with each gNB–UE pair.
- LightGBM learns a nonlinear mapping from distance and wall count to RSRP using experimentally collected measurements.
- The dataset is split randomly into 70% training and 30% validation subsets.
- Because it relies on distance and obstacle information, the model can estimate signal levels for unobserved gNB locations without additional measurements.
IV. GNB PLACEMENT OPTIMIZATION
The placement framework searches combinations of candidate gNB locations and evaluates them using coverage maximization or max-min signal optimization. These criteria explicitly trade served-area expansion against improvement of the weakest conditions.
- The trained model estimates RSRP for arbitrary candidate gNB locations, enabling deployment evaluation without additional measurement campaigns.
- For a fixed number of gNBs, combinatorial search selects the candidate configuration maximizing a predefined objective.
- Inter-cell interference is excluded by assuming orthogonal frequency resources for neighboring cells.
- Candidate locations are uniformly distributed over the floor plan at 1 m spatial resolution, with predicted RSRP evaluated across the grid.
- With multiple gNBs, each location uses the strongest available predicted signal, producing a composite coverage map.
- Coverage Maximization: Coverage maximization maximizes the percentage of locations whose RSRP exceeds the threshold τ, fixed here at −95 dBm.
- Max–Min Optimization: The max-min criterion maximizes the weakest signal level across the environment to improve challenging cell-edge conditions.
V. RESULTS AND DISCUSSION
The evaluation compares LightGBM with alternative regressors and examines gNB placements optimized either for covered area or disadvantaged locations. Results show a persistent trade-off: coverage maximization serves more locations, while max–min improves lower-tail and worst-case conditions, especially with more gNBs.
- Propagation-model evaluation: LightGBM achieved the best overall predictive performance among the evaluated models, despite KNN obtaining a slightly lower MAE.LightGBM yielded the lowest RMSE and MSE values and was selected for the placement analysis.
- Single-gNB placement: With one gNB, coverage maximization produced stronger signal over a larger contiguous area, whereas max–min generated a more balanced distribution.Neither criterion provided satisfactory coverage across the entire floor plan with a single gNB.
- Four-gNB placement: With four gNBs, coverage maximization expanded the area above the target RSRP threshold, while max–min targeted poorly covered regions.The resulting max–min placement produced a more homogeneous signal distribution, whereas some highly attenuated zones remained under coverage maximization.
- Optimization trade-off: The coverage-maximization criterion increased adequately served locations, while max–min prioritized the most challenging propagation conditions.This trade-off may favor max–min where service continuity across the operational area matters more than maximizing already well-served coverage.
- CDF analysis: In the four-gNB case, max–min shifted the lower tail toward higher RSRP values, with the effect particularly visible at the 10th percentile.The single-gNB scenarios showed very similar lower-tail behavior because one deployed gNB offers limited placement flexibility.
- Optimization trade-off: Across one-to-four-gNB scenarios, coverage maximization achieved the largest coverage ratio, whereas max–min consistently improved the 10th-percentile and worst-case RSRP metrics.From two deployed gNBs onward, lower-tail improvements became substantial as additional placement flexibility became available.
VI. CONCLUSION
The paper presents a measurement-driven framework for indoor gNB placement optimization and evaluates coverage-maximization and max–min criteria. Results reveal a trade-off between expanding coverage and improving cell-edge performance.
- A LightGBM-based propagation model trained on experimental RSRP measurements evaluates alternative indoor gNB deployment configurations.
- The framework compares coverage-maximization and max–min optimization criteria for indoor radio planning.
- The methodology reveals a trade-off between coverage expansion and cell-edge performance.
- Future work includes optimization formulations incorporating interference and capacity constraints.