“Reverse Engineering” of Antenna Arrays

Sammanfattning

One of the main objectives of a radar system is estimating the direction of an incoming signal, called direction of arrival (DoA) estimation. Precise knowledge of the positions of the antenna array elements in the receiver is a prerequisite to obtaining accurate DoA estimation. There are established methods for calibration of the antenna array parameters, however, these have been shown to lack robustness. Furthermore, with the recent buzz around machine learning, researchers are looking at the viability of various data driven estimation approaches as alternative methods to traditional DoA estimation techniques. To the best of my knowledge, no machine learning approaches have yet been developed to do the opposite, using the incoming signal together with the DoA to estimate hardware parameters of the antenna array, which is what this thesis has focused on. This thesis used a preprocessed version of a received antenna signal as the input to a neural network, in order to estimate the antenna positions of the receiving array. Both a multilayer perceptron (MLP) and a convolutional neural network (CNN) were trained and evaluated in several tests. These neural networks were trained using a supervised approach, where the inputs were labeled with the true positions of the array elements. Both the training and the evaluation were performed on simulated data. The tests were focused on testing the generalization performance and robustness of the networks by studying the estimation accuracy while certain characteristics of the received signals were changed. The neural networks were also tested on a real dataset. The results were similar for the two network types, with both showing promising estimation accuracy. However, the tests on the real data showed some signs of unstable estimates. One of the reasons for the unimpressive results on the real data could be due to the training not being broad enough.