Abstract
This work addresses the challenge of radio frequency (RF) geolocation in complex multipath environments. By utilizing 3D electromagnetic propagation modeling to characterize specific environments, researchers can generate path-loss radio maps for use in received signal strength indicator (RSSI) fingerprinting. This data-based approach incorporates site-specific information, allowing it to outperform traditional model-based methods like Time-Difference of Arrival (TDOA) or generalized propagation models. Validated through both synthetic simulations and physical testbeds, the results demonstrate that simulation-informed fingerprinting significantly improves geolocation accuracy, particularly in indoor settings where multipath interference is prevalent.
Introduction
Locating an "uncooperative" radio frequency (RF) emitter—one where the transmit power, time, and location are unknown—is a significant challenge in indoor environments. In these spaces, signals do not travel in straight lines; instead, they undergo multipath interference, bouncing off walls and diffracting around corners.
Traditional "model-based" methods, such as Time Difference of Arrival (TDOA) or simple path-loss models, rely on generalized assumptions about signal propagation that often fail in complex office layouts. An alternative is "data-based" fingerprinting, which compares received signals against a pre-measured "radio map" of the building. However, creating these maps via physical site surveys is resource-intensive and time-consuming. This research explores whether 3D electromagnetic simulation can bridge this gap by generating high-fidelity training data without the need for extensive manual labor.
Methodology
The study evaluates two primary categories of geolocation:
- Model-Based Methods: These include TDOA (estimating location via signal time-lags), Differential Received Signal Strength (DRSS) Circles, and RSSI Gaussian Joint Estimation.
- Data-Based Methods: These include Weighted K-Nearest Neighbor (WKNN) and Convolutional Neural Networks (CNN). These methods treat the received signal strengths across multiple nodes as a unique "fingerprint" to be matched against a database.
The core innovation is the application of 3D Ray Tracing, specifically using Remcom’s Wireless InSite, to simulate the environment. By importing CAD models of buildings and assigning material properties to walls and furniture, the team generated synthetic "complex impulse responses." This enabled the creation of massive datasets of signal fingerprints for every square meter of a facility, which were then used to train the WKNN and CNN models.
Experimental Setup
The study utilized three primary datasets to validate the approach:
- Synthetic Office Environment: Based on a typical indoor office model from NYU, featuring 28 receiver locations and 2,400 simulated transmitter locations.
- Physical Testbed 1: A real-world indoor space with 6 receivers and 100 measured locations used for validation.
- Physical Testbed 2: A more complex real-world space with 8 receivers and 500 measured locations.
For the physical testbeds, the researchers compared models trained on measured data (from a physical survey) versus simulated data (from the 3D model).
Results and Discussion
Performance Comparison
Data-driven methods consistently outperformed model-based methods because they account for the unique geometry of the multipath environment rather than assuming a generic decay curve. In the large-scale simulation with 28 receivers, the CNN achieved the lowest Root Mean Square (RMS) error at 2.7 meters, followed closely by KNN at 3.6 meters. For comparison, the standard TDOA method resulted in an error of 6.6 meters.
Simulation as a Surrogate
In Testbed 1, the KNN model trained on simulated data actually performed better than the one trained on measured data. This is likely because the simulation could generate a much denser grid of training points than a human could feasibly measure by hand. Even in Testbed 2, where the simulation had slight discrepancies from reality (likely due to inaccurate material properties or layout details), the simulation-informed model still beat the best model-based approach (RSSI Gauss).
Hardware and Scalability
The research also quantified the impact of hardware density. While increasing the number of receivers from 4 to 28 significantly reduced error (e.g., CNN error dropping from 8.1m to 2.7m), the most substantial gains occurred with the addition of the first few nodes, indicating a point of diminishing returns for hardware deployment.
Conclusion and Future Work
This work has shown that data-based approaches are more accurate than model-based ones at the expense of requiring site-specific training data. Simulation-informed RF geolocation provides a viable path forward for deploying accurate spectral monitoring solutions without the cost of manual site surveys.
Future work will focus on improving simulation fidelity by refining electromagnetic material properties in CAD models and exploring open-source alternatives like Nvidia Sionna. Additionally, the team is investigating the use of the Xilinx RFSoC to run complex geolocation algorithms directly on receiver nodes, potentially incorporating Angle of Arrival (AoA) data for even higher precision.
Summary updated