Autonomous Aircraft Control Using Surveillance-Trained Flight Intent
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current training methods for autonomous aircraft onboard automation tools are limited by data availability and lack generality, as they are specific to small UAVs and do not account for uncertainty in training data, making it difficult to safely operate commercial aircraft autonomously.
Innovation Solution
A method and system that utilize historical surveillance data from commercial aircraft to infer navigation and guidance commands, applying machine learning algorithms to create a mapping function between aircraft states and actions, allowing for tailored autonomous operation using the Aircraft Intent Description Language (AIDL) format, enabling the onboard automation tool to command the aircraft optimally based on real-time sensor data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If data-driven training methods are used for autonomous aircraft, then training can be performed using available data, but the methods lack generality and are specific to small UAVs only
Solution Approach 1:
The patent creates a universal training framework that can be applied to any aircraft type (UAVs, commercial aircraft, freighters) rather than being specific to small vehicles. The system uses generic surveillance data from multiple sources (radar, ADS-B, QAR) and processes it through a unified architecture that adapts to different aircraft categories, making the training method universally applicable while maintaining reliability through standardized safety protocols
2Reliability
If current training systems are used, then training data can be processed, but uncertainty in training data is not accounted for
Solution Approach 1:
The patent introduces an intermediary layer between raw surveillance data and the autonomous operation system. This layer includes data validation modules, uncertainty quantification mechanisms, and filtering processes that handle uncertain data before it reaches the training algorithm. The intermediary architecture manages complexity by organizing data processing into distinct stages with clear interfaces
3Productivity
If autonomous operation is implemented for commercial aircraft, then crew costs are reduced, but safety must be ensured through proper training
Solution Approach 1:
The patent emphasizes preliminary training actions before autonomous operation is deployed. The system performs extensive offline training using historical surveillance data, simulates various flight scenarios, and validates the autonomous system's performance against safety criteria before actual deployment. This preliminary preparation ensures that when autonomous operation reduces crew costs, safety requirements are already met through thorough pre-training
4Quantity of substance
If surveillance data is used for training, then data availability is improved, but the data must be processed and generalized across different aircraft types
Solution Approach 1:
The patent segments the data processing system into modular components: data collection modules that gather surveillance data from various sources, preprocessing modules that clean and validate data, feature extraction modules that identify relevant patterns, and training modules that apply algorithms. This segmentation handles the complexity of processing diverse surveillance data while maximizing data availability for training different aircraft types
Data Source
AI summary
A method and system for autonomously operating an aircraft. The method comprises a pre-flight training step comprising: retrieving recorded surveillance data of a plurality of flights corresponding to at least one aircraft type and at least one route; inferring aircraft intent from the recorded surveillance data; computing reconstructed trajectories using the inferred aircraft intent; selecting a training dataset comprising aircraft intent and reconstructed trajectories of flights corresponding to a particular aircraft type and route; and applying a machine learning algorithm on the training dataset to obtain a mapping function between aircraft states and actions. The method further comprises a real-time control step executed during a flight of an aircraft, the real-time control step comprising: repeatedly retrieving onboard sensor data; obtaining real-time aircraft states from the onboard sensor data; determining actions associated to the real-time aircraft states using the mapping function; and executing the selected actions on the aircraft.