Analysis of commercial drone sounds and its identification

Abstract

The usage of quadcopter types of drones is now on mature and a practical stage and many major manufacturers are expanding its applications into various regions with it. Considerable characteristic of this type of flying object as its maneuverability and practicality is now being focused on how we control this among our urban life from the possibility of any offensive usage. Most of them are either small enough to avoid many current airborne detection methods and cheap enough to use them as disposable. In this paper, we tried to analyze the recorded sounds of a subset of commercial quadcopter types of drones and built a trained simple non-linear neural network filter to classify them among the given sound samples. We borrowed Mel-frequency cepstral coefficients as the well-known methodology of sound analysis process but including some of the parameter adjustments for this research, and applied LeNet neural network filter structure for the following classification test. To maintain the information of adjacent samples among the series of wave samples, 2-D spectrogram planning was applied as for the input signal preprocessing. Most of the frequencies from drones were observed as gathered around 3 to 5Khz, up to around 10Khz, and adjusted LeNet architecture could classify over 10 types of drone categories with over 95% of accuracy.