What if an unstable aircraft approach could be identified early enough for a flight instructor to intervene?
That is the problem behind this open source project: AI Early Warning for Unstable Approaches in GA Flight Training.
The project uses public ADS-B data from the OpenSky Network around Daytona Beach International Airport (KDAB) to analyze general aviation training approaches.
It currently:
- Extracts individual approaches from flight tracks
- Filters out unreliable ADS-B reports and fragments
- Focuses on common training aircraft such as Cessna 152/172/182, Piper PA-28 and Diamond DA40/DA42
- Detects potential instability using sink rate and speed rules
- Creates a blind review sheet for flight instructors
- Compares automated flags against instructor judgment
The initial dataset covered 3 hours of flight activity:
105 approaches analyzed
90 training aircraft approaches
72 approaches selected for instructor review
7 approaches flagged by at least one rule
One Cessna 172S approach triggered both speed and sink-rate warnings, highlighting how data-driven methods could support instructor review.
Importantly, the project does not claim that every automated flag represents a genuinely unstable approach. Ground speed is affected by wind, ADS-B measurements have limitations, and instructor review is essential.
That is what makes the project interesting: the next step is validating and calibrating the rules against expert human judgment.
This is a good example of how open source aviation data can be turned into a practical safety research problem.
🔗 GitHub: https://github.com/samsuseelan/ga-unstable-approach-warning
Top comments (5)
This is a strong example of using real-world aviation data to support safety rather than simply replacing human judgment. The comparison between automated warnings and instructor assessment could provide valuable insights into how these systems should be calibrated.
Very interesting application of ADS-B data. Combining sink-rate and speed-based rules with instructor review creates a practical framework for evaluating whether automated approach warnings are actually useful in flight training.
This is a really impressive project! I love how it connects open-source ADS-B data, machine learning, and real-world aviation safety in a practical way. The focus on keeping the instructor in the loop rather than treating the AI as the final authority is especially valuable. Great work turning an important aviation safety problem into something that can actually be tested, reviewed, and improved.
The idea of early detection hinges entirely on how you define the threshold for "unstable" without triggering constant nuisance alerts. In my experience working with sensor data and predictive modeling, the biggest hurdle isn't just identifying the deviation, but managing the latency between the data spike and the instructor's ability to react meaningfully. If the system flags a trend too late, the instructor is just reacting to a crisis rather than preventing one. You have to balance sensitivity against the cognitive load placed on the pilot; if the intervention system is too chatty, it might actually contribute to the very instability it's trying to mitigate.
Some comments may only be visible to logged-in visitors. Sign in to view all comments.