Characterizing Wake Vortices of Landing Aircraft Using Artificial Neural Networks and LiDAR Measurements

Abstract

The characterization of wake vortices with Light Detection and Ranging (LiDAR) instruments is commonly facilitated using analytical algorithms such as the Radial Velocities (RV) method. However, these can either not be employed for all LiDAR types, require time-consuming semi-automatic processing, or lack accuracy requirements for fast-time hazard prediction at airports. The approach taken in this paper employs Artificial Neural Networks (ANNs) for the estimation of the location and strength of the primary wake vortices trailing behind landing aircraft, going beyond the qualitative wake vortex identification of previous literature. Custom Multilayer Perceptron (MLP) and Convolutional Neural Network (CNN) architectures are generated, and compared to state of the art LiDAR processing algorithms. For this, LiDAR measurements taken at Vienna International Airport that were processed with the RV method are utilized for supervised training of the networks. In addition, feature engineering is performed, allowing to increase the performance of the ANNs by mitigating crosswind effects, optimizing measurement grid positions, and minimizing interfering boundary layer effects. Results indicate the superior performance of the custom CNNs over the custom MLPs in nearly all characterization parameters, while the evaluation speed of a single LiDAR scan turns out to be substantially faster compared to the current state of the art RV method. The custom CNN architecture results in circulation errors as low as 26 m^2/s and localization errors as low as 13 m. A hazard prediction reliability of up to 91% is obtained, given the accuracy of the RV method which constitutes a natural limit of the performance capabilities of ANNs.