Candidate: Ahmed Sayed
Date: November 27, 2023
Time: 11:00 AM - 12:00 PM
Place: Remote Attendance
Supervisor(s): Ramahi, Omar - Shaker, George
Abstract:
Unmanned Air Vehicles (UAVs), commonly known as drones, have become widely available across the globe; they can be used for many useful applications such as search and rescue, surveillance, traffic monitoring, weather monitoring, firefighting, videography, agriculture, and delivery services. However, they can also be used for illegal activities such as drug smuggling, interference with airplanes, privacy invasion, and terrorist attacks. Therefore, it is crucial to identify and classify UAVs for safety and security purposes. For this objective, detection and classification of UAVs is important. Radar systems, for their advantages, are considered a promising solution for the detection and classification of UAVs compared to the other UAV detection systems. Radar systems work day and night and in all weather conditions, they can detect autonomous flights, multiple UAVs, and they can classify different type of UAVs when empowered by Machine Learning (ML) algorithms. Generating radar datasets that contain UAVs information to train ML algorithms through real measurements is costly and time consuming. In addition, these datasets are limited to the used radar parameters, the types and number of UAVs, and the background environment where these measurements are taken. By simulating all these parameters, synthetic radar datasets that contain UAVs information can be generated without many constraints. Furthermore, in the state-of-the-art literature, ML algorithms were trained and tested on radar UAVs datasets that were generated by taking radar measurements of hovering and/or pitching UAVs only. This leads to a degradation in ML accuracy to classify these UAVs when the trained ML algorithms are tested on different datasets that contain different motions of the UAVs.
In this thesis, a new method to generate synthetic radar datasets that contain UAVs information is introduced. This method is based on using 3D full-wave Electromagnetic (EM) Computer-Aided Design (CAD) tools to model the UAVs, the radars, and the background environments. The novelty of this method is that full-wave EM CAD tools are used to simulate the motion of UAVs and perform time-based full-wave analysis. Using this new method, a large number of complex scenarios can be modeled and investigated within a short time, and with low cost. The proposed method is validated, and a ML algorithm is applied to it, yielding an accuracy exceeds 97%. Additionally, the proposed methodology makes it easier to analyze attempts to deceive radar systems, which has the potential to improve the accuracy and effectiveness of UAV detection and classification using radar systems. Furthermore, an investigation on the effect of the control systems of four different UAVs on their range-Doppler signatures is conducted. Different UAVs datasets containing range-Doppler information of different motions of these UAVs are generated using a full-wave EM CAD tool. The range-Doppler signatures are demonstrated to be highly dependent on the mechanical control information of these UAVs. The mechanical control information for the four UAVs is employed using eight different ML algorithms to investigate their impact on classifying the four UAVs. A Mechanical Control-Based Machine Learning (MCML) algorithm is introduced to classify these UAVs. The proposed MCML method overcomes the degraded classification accuracy in case the mechanical control information of UAVs was not taken into consideration, yielding an accuracy exceed 90%. Overall, the proposed MCML method demonstrated superior accuracy when compared to state-of-the-art works employing radars for UAV classification.
In addition, the importance of using Multiple-Input Multiple-Output (MIMO) radars in the field of detection and classification of UAVs is explored. A Single-Input Single-Output (SISO) radar system is used to investigate the impact of its antenna’s Field of View (FOV) on the classification of UAVs. It is shown that the classification accuracy decreases with the increase of the relative angle between the UAVs and the radar’s antenna. However, it is shown that MIMO radars increase the classification accuracy significantly. For example, it is demonstrated that with an antenna beamwidth of 30o, the classification accuracy of a Convolutional Neural Network (CNN) algorithm increases from 13.4%, using SISO radar system to 60.83% using MIMO radar system, at a relative angle of 80o between the antennas and the UAVs. Also, MIMO radar systems provide a high level of accuracy in the area of multi-UAV classification. MIMO radars are used to count the number of UAVs in a scene and enable their simultaneous classification. Although using MIMO radar systems enhance the detection of UAVs compared to SISO or Single-Input Multiple-Output (SIMO) radar systems, MIMO radars struggle to detect and classify multiple UAVs at wide relative angles. Beamforming is considered one of the significant advancements in the field of antenna technology and signal processing for the detection and classification of UAVs. By using radar beamforming, the detection probability of UAVs is increased by electronically guiding the antenna beam in a certain direction. Additionally, radar beamforming improves spatial resolution of UAVs and raises the Signal to Noise Ratio (SNR), ensuring the detection and classification of small UAVs. To address this issue, the thesis proposes a Range-Doppler Integration while Steering (RDIwS) beamforming-based method, which further enhances UAV detection and classification, particularly in scenarios involving multiple UAVs and wider relative angles. The method involves steering the antenna beam to various angles, thereby concentrating the radar’s energy on targets within these angles. Subsequently, the range-Doppler maps obtained within each sector of steered angles are integrated to generate a range-Doppler map for the entire FOV. The RDIwS beamforming-based method improves the SNR of radar targets, subsequently boosting the radar’s detection probability for multiple targets at varying angles. For a scenario with four UAVs positioned at different angles and distances, with a SNR of −30 dB and a false alarm probability of 10−5 , the RDIwS method achieves a detection probability of 75%, as opposed to 5% for steering-only beamforming and no detection in the MIMO radar case.