Airport surface movement prediction and safety assessment with spatial–temporal graph convolutional

Problem: Airport surface movement and safety assessment

Motivation: A predictive model with the ability to forecast the trajectories of ground objects will significantly enhance the situational awareness of airport ground controllers, thus mitigating the risk of object collisions on the airport surface. However, till now, no study has attempted to forecast the trajectories of the objects on the airport surface as a whole from a machine learning perspective; much of the previous work concentrates on the prediction of individual trajectories. 

Framework: Raw data processing→Graph-based representations of ground object movements→Trajectory forecasts of all ground objects→Separation distance-based safety assessment

Raw data processing：

SWIM Terminal Data Distribution System (STDDS)，provides detailed information on the movement of ground objects on the airport surface.

Data format: eXtensible Markup Language (XML) format-The eXtensible Markup Language (XML) is a versatile markup language that is designed to store and transport data. It uses a tag-based structure to define elements and attributes within a document. XML documents are both human-readable and machine-readable, making them widely used for representing structured data. The format is defined by the World Wide Web Consortium (W3C) and has become a standard for data interchange between applications.

Processing tools: Apache Spark-an open-source distributed computing system that provides a fast and general-purpose cluster-computing framework for big data processing and analytics.

Graph representations of airport objects:

Convolution in graph neural network (STG-CNN):

Trajectory forecasts of all ground objects: 

Specifically, TXP-CNN operates on the temporal dimension of the graph embedding 𝑽̂ , treats the graph embedding as 𝑇 channels of images, and expands it to forecast the future positions of the objects.

Since our goal is to predict the object trajectory over the next 𝑇̂ time steps, 𝑇̂ convolutional matrices are constructed with each single convolutional matrix corresponding to one single time step in the forecast horizon.

Collision safety assessment on the airport surface:

bi-variate Gaussian distribution in order to represent the uncertainty in model prediction, it will be an area not just a point. Mento Carlo sampling draws points from ares.

Experiment:

generalisability：peak，off-peak