ProxyFAUG: Proximity-based Fingerprint Augmentation

Grigorios G. Anagnostopoulos, Alexandros Kalousis

Abstract

Fingerprint-based localization is a critical technology for IoT devices in environments where GPS is unavailable or energy-prohibitive. However, the efficacy of machine learning models in this domain is often limited by the high cost of collecting dense training data. We present ProxyFAUG, a proximity-based fingerprint augmentation method that leverages genetic algorithm operators—crossover and mutation—to synthesize training data from spatially proximate signal snapshots. Evaluated on a large-scale urban Sigfox dataset, ProxyFAUG demonstrates a 40% reduction in median positioning error, effectively filling gaps in signal space without additional field surveys.

Introduction

Localization via signal fingerprinting involves mapping Received Signal Strength Indicators (RSSI) from various base stations to specific geographic coordinates. While effective, creating these maps is a labor-intensive process. Traditional data augmentation techniques often assume the availability of multiple signal readings at identical coordinates to model noise. In real-world scenarios, such as "drive-by" data collection, datasets frequently contain only a single snapshot per location.

ProxyFAUG addresses this limitation by relaxing the requirement for exact coordinate matches. Instead, it utilizes spatial proximity to identify parent fingerprints for augmentation, enabling the generation of plausible synthetic data points between existing observations.

Methodology

ProxyFAUG treats signal fingerprints as "chromosomes" and applies a genetic-inspired pipeline to generate new training samples:

  1. Proximity Clustering: The algorithm identifies pairs of fingerprints recorded within a defined distance $r$ of each other.
  2. Crossover: A synthetic "child" fingerprint is created by randomly selecting feature values (RSSI from specific base stations) from either of the two parent fingerprints.
  3. Mutation: To account for natural signal fluctuations, a mutation operator with probability $p_m$ replaces feature values with a random value within the range defined by the two parents.
  4. Midpoint Labeling: The resulting augmented fingerprint is assigned a geographic label corresponding to the midpoint between the two parent locations.
Visualization of the crossover and mutation process generating augmented fingerprints (Aug1, Aug2) from parents A and B.

This combinatorial approach explores the potential signal profiles that could be observed in the physical space between known data points, creating a denser training mesh.

Experimental Evaluation

The method was evaluated using a public Sigfox dataset from Antwerp, Belgium, featuring 84 base stations in an urban outdoor environment. The original training set of approximately 10,000 points, collected by postal service vehicles, was expanded to over 68,000 points using ProxyFAUG. A k-Nearest Neighbors (kNN) positioning model was then trained on both the original and augmented datasets for comparison.

Results and Discussion

The application of ProxyFAUG yielded significant improvements in localization accuracy, particularly for typical use cases. The median positioning error on an independent test set dropped from 108 meters to 65 meters, representing a 40% improvement. While the mean error saw a more modest 6% reduction (from 298m to 280m), the performance gains in the lower error quartiles were substantial.

Cumulative Distribution Function (CDF) of positioning error comparing the original dataset performance against the ProxyFAUG-augmented dataset.

As shown in the CDF, the augmented model (blue) consistently outperforms the original model (red) across the majority of the error distribution. This indicates that the method is highly effective at making "typical" localizations more accurate, even if extreme outliers—often caused by signal aliasing in complex urban environments—remain a challenge.

Conclusion

ProxyFAUG provides a robust framework for enhancing fingerprint-based localization without the need for additional site surveys. By leveraging the spatial correlation of nearby signals, it successfully synthesizes high-quality training data that improves model reliability. Future research will focus on dynamic parameter tuning to adapt to varying data densities and testing the method's generalizability across other wireless protocols such as LoRaWAN and WiFi.

Summary updated

BibTeX
@inproceedings{anagnostopoulos2021proxyfaug,
  author = {Grigorios G. Anagnostopoulos and Alexandros Kalousis},
  title = {{ProxyFAUG: Proximity-based Fingerprint Augmentation}},
  booktitle = {IPIN},
  year = {2021},
  doi = {10.1109/ipin51156.2021.9662590},
}
Figure