Conflict Prediction Using Generative Model from ADS-B Data

Conflict detection is one of the fundamental elements in air traffic control. The output of conflict detection is the important input for conflict resolution which is an important task in maintaining safety and efficiency for flights. Even though there are tools to support air traffic controller in detect potential conflict, their quality and accuracy are still limited since the difficulty in encapsulating uncertainty in predicting flight trajectory. To tackle that challenge, several studies focus on applying probabilistic approaches by modelling aircraft dynamic and uncertainty. Those approaches share common limitations such as their assumption about uncertainty distribution and high computational cost for locating and computing the probability of conflict. In this work, we propose a data-driven approach to model the uncertainty of flight trajectory combining with technique for quickly locating position with highest conflict probability. Firstly, we use generative models to capture the complex trajectory pattern and its uncertainty directly from ADS-B data. Those model are used to predict aircraft position in future with uncertainty given current aircraft information. To compute the time at which two aircraft have highest probability of conflict, Monte Carlo method will be applied. Since conflict is a rare event so that the cost will be expensive, Bayesian Optimization algorithm is applied to quickly identify that time. Final the performance and computational cost of our approach is compared with Monte Carlo method using simulated data.