Neural Network based Real-time UAV Detection and Analysis by Sound

Abstract

In this paper, we present a real-time artificial intelligence system for drone detection on multiple locations. With ensemble machine learning on multiple regional clients and a central neural network server, the users can easily monitor the drone's appearance based on its motor sound data. The clients perform FFT on the sampled real-time data and detect drones using Plotted Image Machine Learning (PIL) with sending the detected audio sample to the server. The PIL uses image data from the visualized FFT graph to detect robust points, and compares the average image similarity with a reference FFT template associated with a target of interest. The server visualizes each client's detection status including machine learning result, and trains Artificial Neural Network (ANN) with extensive regional samples from clients. Afterwards, the server tests its ANN model whenever the client reports drone detection. The accuracy rate of client's PIL test is 83% and server's ANN test accuracy rate is 86%. The major deliverables of this work are a software package framework one may use to train its ANN model with various sound samples from different places to make a generalized drone detection model.