Synthetic Runway Training Data for Low-Visibility AI Landing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current technologies face challenges in generating sufficient synthetic training data for deep-learning algorithms to recognize runways under degraded visibility conditions, due to limited real data availability and high costs associated with collecting sensor images from various meteorological conditions and aircraft approaches.
Innovation Solution
A method using a flight simulator and conditional generative adversarial neural networks (cGAN) to generate synthetic sensor images, which can be used alone or in conjunction with real data to create a large training database, simulating diverse conditions and enhancing the detection capacity of runway-recognition systems.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If real sensor images are collected from various meteorological conditions and aircraft approaches, then the training data diversity and quality are improved, but the cost and time required for data collection increase significantly
Solution Approach 1:
The patent uses flight simulators to create virtual copies of real-world landing scenarios, generating synthetic sensor images that replicate diverse meteorological conditions and aircraft approaches without requiring physical flight data collection. This copying approach maintains training data quality while eliminating the time and resource costs of actual data gathering missions
Solution Approach 2:
The system pre-generates comprehensive training datasets through flight simulations before actual deployment needs arise. By performing the data generation action in advance through virtual environments, the system avoids the need for time-consuming real-world data collection when the runway recognition system needs to be trained or updated
2Adaptability or versatility
If real sensor images are collected from various meteorological conditions and aircraft approaches, then the training data diversity is improved, but the costs associated with data collection increase
Solution Approach 1:
Virtual flight simulators create cost-free copies of diverse landing scenarios including fog, rain, and different approach angles. These synthetic images provide the same training diversity as real collected data but without the expenses of deploying aircraft, sensors, and personnel to capture actual footage under various conditions
Solution Approach 2:
The flight simulator allows systematic variation of environmental parameters (weather conditions, lighting, visibility) and flight parameters (approach angle, altitude, speed) to generate diverse training datasets. This parameter-based generation provides comprehensive adaptability coverage at a fraction of the cost of physically recreating each scenario with real sensor collection
3Device complexity
If conventional algorithms are used for runway identification, then the system simplicity is maintained, but the recognition reliability under poor meteorological conditions decreases
Solution Approach 1:
The system performs preliminary training of deep learning algorithms using extensive synthetic flight data that covers rare and extreme meteorological conditions. This pre-training preparation enables the algorithm to achieve high reliability under poor visibility conditions while maintaining reasonable operational simplicity during actual landing operations
Solution Approach 2:
The patent transitions from conventional algorithms with fixed parameters to deep learning algorithms with learned parameters through training on diverse synthetic data. This parameter learning approach enables adaptive recognition under varying meteorological conditions, significantly improving reliability while the trained model maintains computational efficiency during deployment
4Loss of time
If deep-learning algorithms are trained with limited real data, then the training time is reduced, but the detection precision and robustness decrease
Solution Approach 1:
The system generates large volumes of synthetic training data through flight simulator copies of real landing scenarios. This abundant synthetic data replaces the need for extensive real data collection, enabling comprehensive model training that achieves high detection precision without the time delays associated with gathering sufficient real-world examples across all relevant conditions
Solution Approach 2:
Comprehensive training datasets are generated in advance through flight simulations, covering the full range of possible landing conditions. This preliminary data preparation enables complete algorithm training to occur beforehand, achieving maximum detection precision while eliminating the need for time-consuming real-data collection during deployment phases
Data Source
AI summary
A computer-implemented method for generating synthetic training data for an artificial-intelligence machine, the method includes at least steps of: defining parameters for at least one approach scenario of an aircraft approaching a runway; using the parameters of the at least one scenario in a flight simulator to generate simulated flight data, the flight simulator being configured to simulate the aircraft in the phase of approach toward the runway and to simulate an associated automatic pilot; using the simulated flight data to generate a plurality of ground-truth images, the ground-truth images corresponding to various visibility conditions; and generating, from each ground-truth image, a plurality of simulated sensor images.