Watson–Watt Direction Finding With Transfer Learning-Based Calibration

Abstract

A compression of an entire 4-element Watson-Watt direction finding (DF) process, including calibration, into a single neural network based system is introduced. First, a Watson-Watt DF architecture is approximated using neural networks, assuming idealized patterns. Next, the transfer learning is applied on these patterns to demonstrate that the conventional, often costly look-up-table calibration can be replaced at low cost and overhead. Finally, several neural networks are trained, and their performance is analyzed as a function of what data is used in the transfer learning process, and where it was collected. The proposed DF and calibration procedure is implemented with a four element array to achieve reduction in the maximum and root mean square DF errors by about 45% and 39%, respectively.