GNSS Antenna Pattern Prediction and Placement Optimization: A Prototype Method Using Machine Learning to Aid Complex Electromagnetic Simulations Validated on a Vehicle Model

Abstract

An antenna’s radiation pattern is dependent on its geometrical characteristics and its antenna’s surroundings, materials, and geometries. As such, to predict the antenna’s performance in complex environments, such as that of small antennas on large vehicles, it is necessary to obtain a model that represents such a full scenario, so that the simulation may be accomplished in the process of antenna design and placement. Due to the complex and electrically large nature of some electromagnetic problems, the detailed representation, even for a simplified model, may imply a large computational effort, both in terms of time and memory, is needed to perform the simulation. This paper evaluates how machine learning models can be used to mitigate the computational effort required to predict the behavior of antennas requiring complex modeling. It is proposed to start from a more simplified model of the electromagnetic structure to obtain a prediction for the correct solution, without needing to simulate the full structure in every iteration, and to combine this with prediction algorithms to obtain the solution of the full problem. The proposed solution uses convolutional neural networks (U-Net) of a certain accuracy to help with the correct placement of small antennas on autonomous vehicles. The standard approach requires the simulation of a full model at each test position, requiring high computational time and memory. With this new proposal, it is possible to analyze more positions and radiation patterns in a much shorter time, and with less memory, when compared with the solution from the full model. Along with this methodology for each simulation, a Bayesian optimizer is proposed to improve the search process for the best location, leading to a reduction in the required steps. This methodology was applied to support the correct positioning of a GNSS antenna with reference to a set of performance indicators required for autonomous vehicles, but it can be also applied to larger and more complex structures, allowing one to reduce the simulation time of a large electromagnetic structure and the search time for the optimum location.