| Flight object detection is the basis of airspace safety,but the rapidly increased complication of aero electromagnetic environment has brought great challenge for reliable flight object detection.Radar cross section(RCS)is the widely used data for flight object detection due to the rich information reflecting the characteristics of flight objects.However,with the wide deployment of decoy and interfering technology,it is more chanllenging to detect the real object quickly and reliably.Classic methods,such as feature parameter detection,cumulant-based statistics methods,and machine learning-based methods,which requires high-resolution radar and the priori knowledge of feature parameters.In practical cases,these two requirements may not be supported,especially in antagonistic environments.Deep learning-based methods have attracted much research attention and present the art-of-the-state performance due to the excellent study and abstraction capability.However,the data-driven deep learning methods demand sufficient samples,which hinders its application due to the scarcity of measured data in confrontation environments.In this thesis a mechanism-driven deep learning method is studied for flying object detection from RCS signals which contains enough information to detect flight objects such as shape,material and motion characteristics.Firstly,an analytical modeling method is developed,which is based on kinematics for the featuring movement and electromagnetism for the received signals.Then,a generative adversarial network(GAN)-based deep learning model is designed to greatly reduce the complexity of dynamic RCS signal simulation and augment the RCS data,from which a TF-AM-GRU model is proposed to enhance the detection accuracy and reliability of flight objects from numerous decoy objects and interference.The main contributions are listed as follows:(1)Aiming at the ploblem of expensive cost of RCS measurement,a low-cost,repeatable and instructive analytical dynamic RCS data simulation method is proposed,which is based on kinematic mechanism modeling of the movement of flight objects.3D structure modeling of the flight object and electromagnetic simulation are firstly structured to generate static RCS signals.After kinematic modeling of the movement,the dynamic RCS signals are derived by sampling the static RCS signals in a form determined by the kinematic equations.(2)A deep learning method is proposed to generate RCS data with sufficient volumn and variety by data analysis and AC-GAN.On the basis of feature analysis from analytical modeling of RCS signals,a generative adversarial networks(GAN)model is designed to generate RCS samples.The validity of the RCS signals generated from GAN model is also testified.(3)A time-frequency attention mechasim aided deep learning model is proposed for flight object detection.The network uses the Attention Mechanism(AM)to improve the GRU’s ability of capturing data periodicity,and the Fast Fourier Transform(FFT)to process the input data to obtain the frequency domain features for recognition.The accuracy score and F1 score of the proposed network can improve about 10%comparing with some traditional networks. |
