ML Intent Prediction for Air Traffic Collision Alerts
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current air traffic monitoring systems often generate unnecessary alerts due to their reliance on physics-based monitoring, which can overload air traffic controllers with extraneous information and desensitize them to actual collision risks, as they predict potential collisions that do not materialize.
Innovation Solution
The use of vehicle intent prediction models generated through machine learning algorithms analyzing past vehicle track data and contextual factors, such as weather and aircraft type, to predict future trajectories and reduce the number of alerts by distinguishing real collision risks from non-risk scenarios.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If physics-based monitoring is used to predict potential collisions, then the system can identify safety risks, but it generates excessive false alerts that overload controllers
Solution Approach 1:
The patent segments the monitoring system into two distinct components: a physics-based monitoring component that identifies potential collision risks, and a machine learning-based intent prediction component that filters false alerts. This segmentation allows each component to specialize in its strength while compensating for the other's weaknesses, resolving the contradiction between comprehensive risk identification and alert accuracy.
Solution Approach 2:
The machine learning model serves as an intermediary between the physics-based monitoring system and the air traffic controllers. It processes the raw collision risk predictions, adds contextual understanding through intent prediction, and outputs filtered, high-confidence alerts. This intermediary layer transforms the unreliable signal stream into actionable intelligence without losing the original system's risk detection capabilities.
2Reliability
If physics-based monitoring predicts all potential collisions, then comprehensive safety coverage is achieved, but controller attention is depleted by false alarms
Solution Approach 1:
Instead of alerting controllers to all potential collision risks (excessive action), the system uses the machine learning model to perform partial action by selectively filtering alerts. The ML model predicts vehicle intent and only generates alerts when there is high confidence of actual collision risk, thereby maintaining comprehensive safety coverage while reducing controller workload by eliminating false alarms.
3Reliability
If traditional monitoring systems alert on all predicted collisions, then no real risks are missed, but controllers become desensitized to actual dangers
Solution Approach 1:
The system implements feedback through the machine learning model that continuously learns from past collision incidents and refines its intent prediction accuracy. The ML model analyzes historical data to understand patterns of actual collision risks versus false alarms, and uses this feedback to improve its filtering decisions, thereby maintaining complete risk detection while progressively improving alert accuracy.
Data Source
AI summary
Method, system and computer program product for providing a predicted vehicle track and for providing alerts when two predicted vehicle tracks are closer than a threshold amount. A vehicle intent prediction model is generated based on past instance of tracks for a vehicle operation, known vehicle intent data for the past instances, and contextual factors, such as weather, airline operator, air vehicle type or configuration, day of the week, etc. for the past instances. The vehicle intent prediction model can be generated using one or more machine learning algorithms. A future vehicle trajectory for a current vehicle operation can be output by the vehicle intent prediction model using the current track and existing contextual factors for the current vehicle operation. In the event that two vehicles following their respective predicted vehicle future trajectories would be closer than a threshold distance, an alert can be provided.