Assessing Interference with Regression Analysis Techniques

Jonathan Swindell*, Carson Slater, Samuel Hussey, Charles Baylis, Robert J. Marks II

Overview

Paper

This work proposes using regression analysis to predict aggregate interference in next-generation Dynamic Spectrum Access (DSA) systems, which must operate in real-time. The generalized additive model performs comparably to other machine learning methods but is much faster to train, making it suitable for dynamic spectrum environments.

---

Background

A synthetic dataset comprised of transmit power spectral densities (PSDs), distances between devices, and subsequent PSDs after propagation losses were generated by simulating a simple wireless network on a spectral brokering platform. In these simulations, Spectrum Consumption Models (SCMs) were used to articulate various aspects of spectrum use. For our purposes, SCMs are data structures in which transmitter models capture the origin, frequency, strength, and direction of emissions, while receiver models define interference thresholds and frequencies of reception. The dataset was captured by varying the location of a radiometer model across the network’s 1 km2 area while a set of three 5G base stations remained fixed in place. The center operating frequency of the radiometer was chosen to be 23.84 GHz with a 200 MHz channel width. For simplicity, frequencies of the out-of-band transmissions from the adjacent 5G-NR n258 band were modeled using the same center and bandwidth. To determine the extent of interference at each location, each emission is scaled by free space path loss in the appropriate direction. The interference contribution from each base station is then summed to provide an aggregate interference target value. The final dataset consists of out-of-band transmission PSDs at each radiometer location (ranging from -70 to -40 dBW/200 MHz), distances between base stations and radiometer (ranging approximately from 50 to 1650 meters), the individual interfering powers, and the aggregated interfering powers. This data is used to model the interference between a single radiometer base-station pair. Thus, this data is constructed by concatenating the data from each radiometer base station pair in the simulation. Fig.1 illustrates an instance in which emissions of similar powers experiencing path losses are incident upon the radiometer location. This scenario was simulated approximately 20000 times, with the radiometer location being adjusted consistently each iteration.

---

Methodology

Previously, other methods employed to model aggregate interference in Dynamic Spectrum Access (DSA) systems featured machine learning models, of which are not the most parsimoious. This approach simply considers employing a generalized linear model to predict this interference. Machine learning generally exploits correlations within data to develop accurate predictions, but can be hard to interpret. This is why many ML models are referred to as `black box models.’ Regression models, although rudimentary, could prove useful as they are interpretable and relatively computationally inexpensive.

After exploring some different models, we found one of the simplest, yet most accurate models was the following:

$$\hat{R}_{PSD} = \hat{\beta}_0 + \hat{\beta}_1 T_{PSD} + \hat{\beta}_2 f(D).$$

Where the $R_{PSD}$ and $T_{PSD}$ are the recieved and transmitted power spectral density, and $f(D)$ is a transformation of the distance from the transmitter to the reciever. Using this simple, trivial yet powerful model, we were able to obtain performance metrics that are comparable to other machine learning models.

---

Analysis

Below is a table of important files that are required to replicated and tinker with our anaylsis. Most of the analysis was done in R/RStudio, using a select set of packages in the tidymodels metapackage. Additionally, surface plots were made in MatLab.

File	Description
analysis.qmd	A Quarto markdown file containing all of the distilled analysis.
analysis.pdf	A .pdf rendering of analysis.qmd which is much more readable.
MamdaniTuningData.csv	The simulated data used to build, train, and test the model.

: List of files in this repository that are relevant for reproducibility.

Citation

If you find our paper useful, please cite using this BibTeX:

@INPROCEEDINGS{10619025,
  author={Swindell, Jonathan E. and Slater, Carson and Hussey, Samuel and Baylis, Charles and Marks, Robert J.},
  booktitle={2024 IEEE Texas Symposium on Wireless and Microwave Circuits and Systems (WMCS)}, 
  title={Assessing Interference with Regression Analysis Techniques}, 
  year={2024},
  volume={},
  number={},
  pages={1-6},
  keywords={Training;Analytical models;Aggregates;Noise;Interference;Machine learning;Data models;aggregate interference;spectrum management;dynamic spectrum access;regression analysis;linear regression;generalized additive models},
  doi={10.1109/WMCS62019.2024.10619025}}