بهبود آشكارسازي اهداف در رادار پالسي بر مبناي دسته‌بندي محيط با استفاده از يادگيري عميق

Target detection is the most important task of a radar system. In addition to the target signal, unwanted signals such as noise and various types of clutter are also received by the radar receiver, which causes the processing environment to become out of homogeneity and the target detection process becomes more difficult. Also, the presence of targets close to each other is another factor that causes the environment to become heterogeneous. The use of deep learning in the process of target detection can help to improve detection, and the main goal of this research is to improve target detection in pulse radar using neural network. Each of the CFAR (Constant False Alarm Rate) algorithms perform better in certain environments and considering the fact that the CA (Cell Averaging) method for homogeneous environments, the OS (Ordered Statistics) method for multi-Target environments and the GO (Greatest Of) method for clutter environments provide better results, A combined and intelligent method was used to detect the target in all kinds of environments. Due to the lack of easy access to the data bank of the real radar signal in different environmental conditions, in the first step of this research, a software was designed to simulate the signals received by the receiver and generate sufficient data for the training and eva‎luation of the proposed method, In which, radar parameters, environment type, types of clutters and targets with different radar cross sections and in different ranges can be selected and simulated. After simulating a sufficient number of required signals with dimensions of 3000 x 4000, the pre-processing process was performed on the signals and the new matrices, including the main signal and signals of adjacent angles with dimensions of 2958 x 3, were divided into 17 packages. In the second stage, after passing through the normalization block, the data were classified by the 5-layer MLP (Multi Layer Perceptron) network between three homogeneous, multi-Target and clutter environments with an accuracy of about 92.5%. 19370 training data have been used for training and eva‎luation of this network. In this process, according to the type of environment, the appropriate CFAR method is selected and applied on the data frame to check the presence or absence of the target in each of the cells. According to the results of implementing the proposed method on 23 different and mixed scenarios, the true detection rate (TPR) is 0.809, the false alarm rate (FPR) is 0.0019 and the precision criterion is 0.4009. Also, for the purpose of comparison, common CFAR methods were also implemented for the above scenarios. The comparison of eva‎luation results shows that the proposed method has performed better than other common CFAR methods in processing mixed scenarios (scenarios with homogeneous, multi-Target and clutter environments).

5