Machine Learning Based Fourier Phase Retrieval for Planar Near-Field Antenna Measurements

The success and efficiency of many classical iterative plane-to-plane based phase retrieval algorithms is to a large extent dependent upon the fidelity of the initializing, i.e. guiding, phase estimation [1], [2]. This is especially so when using these techniques to recover the phase of active electronically scanned array antennas such as those employed within beam-steering mm-wave Massive MIMO antenna systems intended for 5G New Radio applications where the performance of the algorithm, and its ability to not become trapped within one of the (possibly many) local minima, is particularly dependent upon the quality of the initializing guess where access to a phase reference is not always convenient, or even possible. Many traditional phase recovery iterative Fourier methods employ simulation or passive measurement supported phase initialization [1], however this information is not always readily available, or in the measurement may require a destructive, invasive, examination of the device under test (DUT). In this work we address this issue by presenting a proof of concept which employs a machine learning based neural network [3] to estimate the initializing phase function based on the assessment of the measured amplitude only near-field pattern. Here, we show that there is sufficient information contained within the difference between the two near-field amplitude only scans to be able to determine the antenna beam steering characteristics. A simplified beam steering case with electronic scanning in one, or more, scanning axes is demonstrated and verifies the power of the novel method, as well as illustrating its inherent resilience to noise within the amplitude only measurements, and verification of the robustness of the approach thereby extending the range of measurement applications for which this class of iterative Fourier algorithms may be successfully deployed [4].