Application of Kernel Density Estimation to Achieve Automated Near Real-Time Antenna Pattern Data Processing and Analysis in an Anechoic Chamber

The Benefield Anechoic Facility (BAF) at Edwards Air Force Base is the world's largest known anechoic chamber. Due to its unmatched size and equipment inventory, the BAF hosts far-field pattern measurements at all azimuth angles and multiple simultaneous elevations of installed antennas on large aircraft typically across the BAF's operating frequency range of 0.1-18 GHz. Antenna tests at the BAF rapidly produce copious data, which often require immediate analysis to allow system owners to make relevant improvements. Historically, the BAF had accomplished quality assurance manually. Analysis was accomplished post-test by customers and the BAF team. Today, the BAF team has developed scripts that use kernel density estimation and basic machine learning to automatically check incoming data and highlight anomalies for review. During a 2019 test of installed antennas on a B-1B bomber, the BAF team used these scripts to process antenna patterns in near real-time and bring unusual results to the customer's attention fast enough to allow modifications to be applied and re-tested during the same test event-highly significant as aircraft and BAF schedule times are limited and may be a one-time opportunity to gather required data. This paper will explore the algorithm used to evaluate antenna patterns, as well as the expected characteristics of patterns that enable the selection of relevant data. Development of this algorithm found that using kernel density estimation to calculate the maxima in a pattern's distribution of gain values, then performing this recursively over the main lobe, can identify problems such as incorrect switching, mismatched transmission lines, and multipath. Algorithm optimization was achieved using generated data, then the algorithm was implemented as a prototype during the B-1B test by searching for data that deviated from a sample pattern. Finally, this paper will discuss the application and impact of this algorithm during a live test. All plotted data and numerical results included in this paper are based on training data generated in Python to represent generic directive antenna patterns, and do not correspond to the real-world conditions or results of the B-1B test.