Air Conflict Resolution Using Recurrent Autoencoder and Deep Reinforcement Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing air-traffic control systems are limited in managing air conflicts in three-dimensional airspace, particularly when dealing with multiple aircraft, diverse categories, and various possible actions, as they often require extensive computational resources or are restricted to two-dimensional spaces and specific scenarios.
Innovation Solution
A device and method utilizing a recurrent autoencoder to create a reduced-dimension representation of the airspace, combined with a deep-reinforcement-learning algorithm, to determine conflict-resolution actions that can include speed adjustments, altitude changes, direction alterations, and route modifications, allowing for generalizable solutions across different scenarios and geometries.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If geometric computations are used to ensure continuous decision-making, then decision-making capability is improved, but computing resources required increase intensively
Solution Approach 1:
The system pre-trains deep neural networks offline using historical air traffic data and conflict scenarios. This preliminary action allows the model to learn optimal resolution strategies beforehand, so that during actual operation, decisions can be made quickly with minimal real-time computing resources, resolving the contradiction between decision-making capability and computing resource consumption
Solution Approach 2:
The system dynamically adapts to different air traffic scenarios by using the trained neural network to process real-time situation data and generate context-appropriate resolution actions. The model can handle variable numbers of aircraft, different conflict geometries, and multiple resolution actions dynamically without requiring intensive real-time computation
2Use of energy by moving object
If reinforcement-learning algorithms with dense neural networks are used, then air conflicts can be resolved with minimized computing resources, but the solution is restricted to two-dimensional space and specific scenarios
Solution Approach 1:
The system transitions from two-dimensional conflict resolution to three-dimensional airspace management by incorporating altitude as an additional dimension. The deep neural network is trained to handle vertical separation, climb/descend actions, and 3D conflict geometries, enabling versatile resolution in multi-dimensional airspace while maintaining computational efficiency
Solution Approach 2:
The trained deep neural network serves as a universal model that can handle multiple types of air conflicts, different numbers of aircraft, various resolution actions (speed changes, direction changes, altitude changes), and different airspace geometries. This single model replaces multiple scenario-specific models, achieving both computational efficiency and broad adaptability
3Measurement precision
If neural networks are trained for each type of scenario, then resolution accuracy for specific scenarios is improved, but the system cannot generalize to new successions of waypoints
Solution Approach 1:
The system performs comprehensive offline training of the deep neural network using diverse historical data covering multiple scenario types, conflict geometries, and resolution strategies. This preliminary action creates a generalized model that can accurately handle new scenarios without retraining, resolving the contradiction between resolution accuracy and generalizability
Solution Approach 2:
The system changes the approach from training multiple scenario-specific models to training a single model with adaptable parameters. The neural network learns to generalize across different scenarios by adjusting its internal parameters based on the input situation, maintaining high resolution accuracy while achieving broad generalizability to new waypoints and conflict configurations
Data Source
AI summary
A device for managing air traffic, in an airspace includes a reference aircraft and at least one other aircraft, the device receiving a three-dimensional representation of the airspace at a time when an air conflict is detected between the reference aircraft and the at least one other aircraft, the device comprising an airspace-encoding unit configured to determine a reduced-dimension representation of the airspace by applying a recurrent autoencoder to the three-dimensional representation of the airspace at the air-conflict detection time; a decision-assisting unit configured to determine a conflict-resolution action to be implemented by the reference aircraft, the decision-assisting unit implementing a deep-reinforcement-learning algorithm to determine the action on the basis of the reduced-dimension representation of the airspace, of information relating to the reference aircraft and/or the at least one other aircraft, and of a geometry corresponding to the air conflict.