Advanced polarization estimation method with machine learning

Nowadays, the importance of electronic warfare systems is increasing. Electronic support systems, which are one of the elements of Electronic Warfare systems, passively use the electromagnetic spectrum. These systems, which analyze the emitted signals, are able to detect the directions, positions, identities, and content of communication between the enemy and friendly elements. In this way, electronic support systems play a major role in extracting intelligence data. With the development of technology, systems that analyze radar or communication signals are expanding their capabilities. Electronic warfare and radar applications try to increase target direction and position finding performance by using signal polarization information as well as increasing reconnaissance capability. These factors have led to the incorporation of polarization information of the signal into existing systems. Depending on the location or direction of the platform where the receiver or transmitter is located, polarization mismatch between the transmitter and receiver may occur. In order for target signals to be received at high power, for successful direction finding performance, or for an effective countermeasure against electronic warfare systems, polarization mismatch in the receiver and transmitter must be prevented. Therefore, detection of the polarization of the emitted signal in electronic support systems plays a critical role in identification, direction and position finding systems. In this study, an approach that estimate the polarization of the incoming signal through deep learning is proposed. In the proposed approach, first of all, the covariance matrix is calculated, such as algorithms that perform classical angle of arrival and polarization estimation. The input images of the algorithm were obtained from the real, imaginer and phase values of the calculated covariance matrix. Thanks to convolutional neural networks, features are extracted from the input images and polarization estimation is performed thanks to fully connected layers. In the study, linear polarization was considered at certain resolutions. Linear polarization is divided into ten different classes, the first class being horizontal polarization and the last class being vertical polarization. The model, which adopts the supervised learning model and includes convolutional neural networks and fully connected layers, has yielded faster results than classical algorithms. The proposed approach has revealed the potential to be used in real-time systems. Within the scope of polarization estimation performance, the algorithm that estimates neighboring classes even with erroneous estimates has given more successful results compared to the classical MUSIC algorithm, especially at low SNR values. In addition, within the scope of the thesis, a hybrid algorithm is proposed for the estimation of the angle of arrrival. The proposed hybrid algorithm primarily performs polarization estimation with a convolutional neural network based algorithm. Then, the classical Multiple Signal Classification (MUSIC) algorithm is used to estimate the angle of arrival. As a result of polarization estimation with convolutional neural networks, the classical MUSIC algorithm estimated the arrival angle faster, since it will search directly only spatial angles without scanning polarization parameters. Since the proposed hybrid algorithm estimates polarization more successfully at low SNR values, it has given more successful results compared to the classical MUSIC algorithm in the scope of angle of arrival estimation.