Why explain aviation incidents?
Flight safety directly depends on how accurately we understand the causes of abnormal situations. It is not enough to detect a hazardous event like a hard landing — we need to figure out which crew actions and flight parameters led to that outcome. Without this, it is difficult to develop recommendations and prevent repeated mistakes.
Traditional explainable AI methods, such as feature importance maps, show which factors influenced a decision but require serious domain knowledge from the analyst. Language models can build coherent textual descriptions, but working with aviation data poses challenges: numerical sensor readings and textual reports do not align well, models are weak at classification, and labeled data for fine-tuning is scarce.
What is FlightLLM?
The study authors propose the FlightLLM approach, which combines the strengths of statistical algorithms and large language models. The goal is not only to determine what happened but also to explain why. A numerical "expert" handles parameter analysis, while the language model generates an explanation understandable to humans.
The solution consists of several blocks. First, feature engineering: statistical characteristics and physically meaningful indicators are added to sensor data to bridge the gap between numerical and textual representations. Then, a semantic discretization module converts numerical patterns into qualitative descriptions like "sharp decrease in vertical speed." Additionally, CatBoost is used as a statistical expert — its prediction is fed into the prompt and serves as a reference point for the language model.
How does the explanation process work?
The system receives flight recordings as input. After feature extraction and semantic discretization, the numerical data takes on a more "human" form. Next, CatBoost produces a preliminary assessment of the event, which, along with a structured prompt and aviation knowledge, is passed to the large language model. The prompts are designed so that the model sees both correct and erroneous examples — this is a contrastive few-shot strategy. As a result, the model produces an explanation that directly points to likely causes rather than merely listing correlating features.
What was FlightLLM tested on?
Experiments were conducted on real data: 704 A320 flight samples. A hard landing was chosen as the target event.
The results showed that FlightLLM achieves competitive classification accuracy while providing direct and logical explanations of incident causes. In other words, the method not only matches specialized classifiers but also adds interpretation that is crucial for flight debriefing.
Why this matters
Explainability in aviation is not a luxury but a necessity. Pilots and analysts gain a tool that helps turn raw data into understandable event scenarios. In the future, such systems could become part of standard incident investigation procedures and crew training.
