Year: 2026
Type: dissertation
Abstract: Indoor radio propagation modeling is essential for wireless system design and planning. Traditionalapproaches face a fundamental tradeoff: empirical models are fast but inaccurate, while deterministic methods like ray-tracing provide higher accuracy at significant computational cost. Machine learning models offer a promising middle ground, combining efficient runtimes with sufficient accuracy for practical applications. This thesis presents two novel machine learning approaches for indoor propagation modeling. The first is a U-Net-based model incorporating physics-inspired features to capture complete propagation environments, including transmitter characteristics, antenna patterns, and wall properties. A compact three-feature representation achieves an MAE of 3.81 dB on real building geometries in under 1 second per prediction. The model introduces a sliding box sampling strategy enabling predictions on arbitrarily-sized geometries without interpolation-based distortion, handling environments from small rooms to 60 m × 60 m buildings. The model also demonstrates robustness to wall height variation (MAE 3.1–3.42 dB across 10–50% variation), enabling accurate predictions in true 3D environments. The second contribution is a diffusion-based model enabling arbitrary in-painting of RSS maps through unconditional generation paired with the RePaint algorithm. Unlike previous approaches requiring separate training for each masking pattern, this single model handles any partial input configuration, facilitating hybrid workflows combining ML predictions with measured data or highfidelity simulations. The model achieves an MAE of 2.59 dB for complete generation and 1.99 dB with 25% uniformly distributed samples. While requiring 7–8 seconds per prediction, it offers unique in-painting capabilities enabling seamless integration with measurement campaigns and full-wave simulations. Both models significantly outperform ray-tracing in speed while maintaining suitable accuracy for network planning. Together, these contributions advance ML-based propagation modeling by addressing critical gaps in feature design, geometric flexibility, height variation handling, and integration with existing techniques.
Source: TSpace (University of Toronto)
Author: Xu, Lingxing
Language: English
Subfield: Electrical and Electronic Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7204014011