Application of machine learning in aircraft noise monitoring: Determining disturbance sounds within aircraft noise events

Abstract

We developed a method for determining the inclusion of disturbance sounds in aircraft noise events using a convolutional neural network (CNN). Considering road-traffic noise as a disturbance sound, recognition models were developed for aircraft and road-traffic noise (hereafter referred to as the base model) and then combined. In addition, we developed a new model, the “frequency-split parallel model,” by advancing the base model. This approach involved inputting spectrograms split along the frequency axis. After verifying the accuracy of each model for single sound events of aircraft and road-traffic noises, the frequency-split parallel model was evaluated using superimposed data obtained around Narita International Airport and compared with the base model. The misidentification rate of road-traffic noise as aircraft noise decreased by approximately 57% compared to the base model, and 16 out of 17 measured superimposed events were correctly determined. These results confirm the effectiveness of the proposed method.