Direction finding of sources around an array is considered one of the most important topics in the field of array signal processing and has brought many researchers. The reason for this can be known the wide application of the Direction of Arrival (DOA) estimation in the fields such as radar, sonar, wireless communication, medicine, etc. In recent years, researchers have made efforts to improve angle detection algorithms, which has led to an increase in the accuracy of estimating the DOA and improving the angular resolution of targets. On the other hand, the use of sparse array makes it possible to detect more targets with a smaller number of sensors than a uniform array while reducing system costs. Therefore, the importance of how to arrange the sensors of an array becomes more apparent than before. In the literature of array signal processing, the number of targets that are identified unambiguously is called the Degree of Freedom (DOF) of the array. Today, with the growth and evolution of communication technologies and the advancement of electronic science in the construction of powerful hardware, access to large amounts of data and their fast processing has become possible. So, the use of Deep Learning (DL) in direction finding has brought significant success. Deep Learning is a branch of machine learning that uses neural network tools to make machines intelligent and learn. The methods based on deep learning have better performance and comprehensiveness than the conventional methods of estimating the DOA. Therefore, deep learning has provided a pristine and suitable field to improve the performance of the arrival angle estimation and consider more stringent conditions such as low Signal to Noise Ratio (SNR) and multipath. In this thesis, while examining and eva‎luating the proposed methods for DOA estimation using deep learning and comparing with classical DOA estimation methods, a Convolutional Neural Network (CNN) is expanded by considering several different sparse arrangements. Also, the effect of non-calibration of the array and parameters like the number of layers on the performance of the CNN has been studied and the proposal of using a more efficient cost function in the field of DL-based angle detection to improve network training is proposed. Finally, a new mathematical framework for arranging the array and achieving a high DOF is provided quickly.

محمود مدرس هاشمي

جلال ذهبي

محمدرضا تابان , محمدعلي خسروي فرد