Ecoflight Monitoring - new aircraft noise monitoring system with neural network-based data filtering

Abstract

Reliable measurements and further evaluation of noise environment in the field, as well as refinement of an airfield noise zone are directly impacted by the completeness and accuracy of registered individual noise events created by aircrafts flying near monitoring locations. Also, monitoring stations are often located in densely-populated areas that feature other noise sources. Detecting aircraft flight events becomes even more complicated when there is no location data for aircrafts that lack ADS-B transponders. The Moscow region, where most aircraft noise monitoring projects are run, is unique in that there are 8 closely located airfields used both for regular passenger flights and flights of experimental aircrafts not equipped with transponders. Each of the 50 noise monitoring stations can simultaneously register noise events that correspond to flight of aircrafts at different airfields. Aircraft flyover events were detected on the noise time line using an artificial neural network (ANN) that allows to identify aircraft audibility periods at the monitoring location using the 1/3 octave band of noise registered in real time for all measurement locations. This article describes the development of an ANN architecture based on the YOLO v4 convolutional neural network, as well as requirements and mechanisms for normalizing data fed to the ANN that filters data provided by the Ecoflight Monitoring aircraft noise monitoring system.