Data-Driven Detection of En-route Convective Weather Avoidance and Development of a Weather Assessment Model
Author(s)
Price, Rachel E.
DownloadThesis PDF (10.51Mb)
Advisor
Hansman, R. John
Terms of use
Metadata
Show full item recordAbstract
Convective weather has significant impacts on aviation operations around the world. It reduces available airspace capacity, and avoidance of hazardous convective systems increases the workload for both pilots and air traffic controllers. Optimizing airspace use in the presence of constantly evolving convective activity is an ongoing research effort. Today, convective weather avoidance models aid controllers in safely and efficiently managing their airspace. These models were built on past examples of aircraft-weather interactions drawn from archived weather observation data and air traffic data. However, the development of existing models was limited by the significant amount of work required to assess convective weather impacts on the final trajectory of each aircraft. This work explores data-driven approaches to weather avoidance modeling and introduces an image-based modeling paradigm.
At present, there is limited ability to examine historical aviation operations and automatically identify when aircraft deviated due to convective weather blocking the intended route. In particular, retrospectively assessing aircraft-weather interactions without the aid of human subject matter experts poses a difficult challenge. Each interaction is different, and simple heuristics have proven to be ineffective at classifying weather encounters. A key contribution of this work is the development of a deviation detection model, which takes in information about a flight’s flown route and its intended route as well as local weather observation data and assigns a behavioral classification. Individual cases are classified as non-deviations, deviations for tactical weather avoidance, or deviations unlikely to be caused by the tactical weather situation. This classifier was developed using a supervised machine learning approach, with training data labeled by crowd-sourced subject matter experts, generally airline pilots. By leveraging transfer learning techniques, a small labeled dataset was sufficient to train the deviation detection classifier.
After the deviation detection classifier was developed, it was used to label a large set of aircraft-weather interactions. The goal of this step was to identify interactions containing useful information on weather avoidance decisions and techniques, which was previously a process carried out by human subject matter experts. This included both non-deviations where the aircraft penetrated convective weather and tactical weather deviations where the aircraft maneuvered to avoid convective weather. The large set of non-deviations and tactical weather deviations was used to develop a deviation prediction model. For the deviation prediction model, only the intended route and local weather observations are considered. The model outputs a prediction as to whether the intended route will be flown or if there will be a deviation for convective weather.
The predictive model developed in this work outperforms existing convective weather avoidance models, and was able to do so based off of an entirely machine-labeled training dataset. This demonstrated the value of using a small dataset to effectively label a much larger dataset, as well as the image classification framework used for developing the deviation detection classifier. Reducing reliance on human data curation and labeling makes the creation of more specialized predictive models feasible and allows future research to leverage techniques requiring larger datasets than were previously possible.
Date issued
2023-09Department
Massachusetts Institute of Technology. Department of Aeronautics and AstronauticsPublisher
Massachusetts Institute of Technology