Serhii Sokolskyi. Acoustic detection of unmanned aerial vehicles – The qualifying scientific paper for the degree of Philosophy Doctor consists of an introduction, four chapters, conclusions, a list of references (50 sources), 3 appendices, and contains 70 figures and 9 tables. The aim of the thesis is to improve the efficiency of the system for detecting and localizing small unmanned aerial vehicles (UAV) and processing of their acoustic signals registered by the input sensor.
Research tasks:
1) Analyzing the main methods of detection and localization of UAVs;
2) Creation and accumulation of a database of audio recordings of acoustic signatures of different models of UAVs;
3) Development of a mathematical model for identifying the type of UAV, using a database of audio recordings of their acoustic signatures;
4) Development of a method for digital processing of acoustic signals using a convolutional neural network (CNN) of deep learning with the Adam optimizer;
5) Based on the proposed method, develop software for processing acoustic signals for the presence of UAVs signatures;
6) Creating an acoustic detector for UAVs;
7) Checking the effectiveness of the developed software and hardware for detecting of UAVs.
The object is the process of acoustic detection of the UAV. The subject is the methods of realization of the system of transmission and processing of acoustic signals of the UAV.
The scientific novelty of the obtained results is that:
1) For the first time, a mathematical model of the UAV identifier is proposed. The scientific novelty of this model is that it is based on the use of a database of audio recordings of acoustic signatures of drones and makes it possible to reduce the error in classifying their model.
2) For the first time, a method of digital processing of acoustic signals of drones is developed, the scientific novelty of which is that it is based on the theory of artificial intelligence and uses a CNN of deep learning with the Adam optimizer to increase the speed and efficiency of the identification of the type of small unmanned aerial vehicles.
3) The method of calculating the number of mel-filters and their frequency range is improved, which, unlike the existing ones, takes into account the desired frequency resolution and makes it possible to increase the speed and accuracy of obtaining the main representations of the UAV audio signal when using the technique of weighting the FFT coefficients using banks of mel-filters.
The practical significance of the obtained results is that:
1) On the basis of the developed mathematical model, a database of audio signals of UAV models “Mavic 2 Pro”, “Mavic 3” and “FPV”.
2) On the basis of the developed method and methodology, software for fast and efficient processing of audio signals of small UAV using a CNN of deep learning with the Adam optimizer is written.
3) An acoustic detector of small UAV with an effective object detection distance of up to 200 meters is created.
4) Relevant universal recommendations for further improvement of the small UAV detection system are provided.
According to the results of the thesis, 6 scientific papers are published: 3 articles in professional editions of Ukraine, 2 of which are in periodicals included in the international scientometric database WEB OF SCIENCE, and 3 abstracts in the proceedings of international scientific and technical conferences.
The introductory part confirms the relevance, formulates the aim, tasks, and methods of research, also provides information about scientific novelty, as well as the practical significance of the obtained results. In the first chapter, a comparative and critical analysis of the potential possibilities of the main methods of detecting small UAV, such as optical, radar and acoustic, is carried out. Each of the methods calculates the theoretical maximum detection range of the DJI Mavic 3 drone. In the second chapter, we consider the implementation of a simple acoustic detector with a single microphone. It allows detection of the sound emitted by the drone's engines and blades. Based on the test results, a database of audio files of the noise of the DJI Mavic 3 quadcopter is created. The spectra of the received audio recordings of the radiation of the quadcopter are important features in the development of the algorithm for the classification of drones. The third chapter describes the process of developing an algorithm for efficient processing and classification of UAV audio signals using a deep learning CNN, building the architecture, and theoretically evaluating of its performance. In the fourth chapter, practical testing of the finished detector model is carried out. The effective distance of UAV detection by the algorithm is 200 m.
Keywords: drone, small unmanned aerial vehicle, UAV, target detection, acoustic location, controlled area, signal recorder, detection, spectrum, signal processing, machine learning, neural network, deep learning.
