Aircraft model identification using convolutional neural network trained by those noises in a wide area around an airfield

We have developed an aircraft model identification system that uses a convolutional neural network (CNN). Our previous study used the CNN model to classify seven and eighteen models of jet aircraft by using training based on measured data taken directly under their flight paths near airports, reaching an accuracy of 98%. However, there were some limitations, such as the study results being obtained at only three sites, the measurements being taken only directly under the flight path, and the season being limited to winter. In this study, we examine if this method is also effective for identifying sound sources over a wider area not directly under the flight path. We conducted aircraft type identification by using the frequency characteristics of aircraft noise obtained at 90 measurement sites located within an area of approximately 30 km north-south and 6 km east-west, in summer and winter. We were able to achieve an identification rate of 98% for four types of sound sources. The results suggest that this system can identify measurement data over a wider area around airfields and can be applied to more practical issues, such as generating noise maps and validating predictions in the noise maps.