Artificial Neural Networks for Individual Tracking and Characterization of Wake Vortices in LiDAR Measurements

Abstract

The main restricting factor of airport capacity are aircraft separations. These separations exist to avoid potentially hazardous wake vortex encounters (WVEs). Especially during the final approach, this hazard is of great concern since aircraft follow the same glide path. The severity of wake vortex encounters depends on the generating and the encountering aircraft, thus dynamic pairwise aircraft separations, which adapt depending on prevailing weather conditions, are desired. Light Detection and Ranging (LiDAR) scans are suggested for monitoring Wake Vortex Advisory Systems due to their fast-time strength and location characterization of wake vortices. The first approaches to automating such characterizations were made with multi-layer perceptrons (MLP) and convolutional neural networks (CNN). Those were shown to be sufficient for wake vortex characterization but could not yet compete with traditional methods in terms of accuracy. For that reason, this work proposes a machine-learning pipeline that uses bounding box predictions by a YOLOv4 network to restrict the input to single vortices for the following CNN to achieve higher accuracy. The LiDAR scans used for training contain radial velocity measurements made at Vienna International Airport. After preprocessing and testing feature engineering, those LiDAR scans are transformed into images as required input for YOLOv4. Afterward, the bounding box predictions are used to cut out individual vortices from the original scans. The individual vortices are then used to train a CNN to enhance localization and the vortex strength estimation further. The evaluation shows that a prediction pipeline is superior to a single CNN approach. The localization error was decreased by more than 90% and the vortex strength estimation by up to 31% to a localization error as low as 2.87 m and a vortex strength error as low as 20.88 m²/s. Furthermore, the precision of detecting hazardous wake vortices was increased by 7.51% to gain a precision of 96.11%. This pipeline can be executed while maintaining a sufficiently low computation time.