Can Large Language Models Explain Flight Safety Events? A Prior-Guided Semantic LLM-based Approach

Air travel is widely considered the safest way to move people across the globe, yet the industry never stops searching for ways to make it safer. A significant portion of aviation risks occurs during the final moments of a flight, specifically when an aircraft touches down on the runway. To understand what happens during these critical seconds, airlines rely on Quick Access Recorders, devices that constantly monitor hundreds of flight parameters, from the speed of the engines to the angle of the wings. For decades, researchers have used this data to spot dangerous patterns, but a major hurdle has remained: knowing that a flight was risky is not the same as understanding exactly why. Traditional computer models can flag a bad landing, but they often struggle to explain the chain of events in a way that human pilots and safety experts can easily grasp. They might point to a number that went too high, but they cannot tell a story about the pilot's actions or the weather conditions that led to that spike.

This gap between raw data and human understanding is what a new study published in the IEEE Transactions on Intelligent Transportation Systems aims to bridge. The researchers, led by a team from Chongqing University and other institutions, developed a system called FlightLLM. This system uses a Large Language Model, a type of artificial intelligence known for its ability to reason with words and generate human-like text, to analyze flight data. The goal was not just to predict if a landing would be rough, but to act as an expert witness that can look at the numbers and write a clear, logical explanation of the cause. The team focused their work on "hard landings," a specific type of safety event where an aircraft hits the ground with excessive force. By testing their system on real-world data from 704 flights of the Airbus A320, they demonstrated that it is possible to combine the pattern-recognition power of traditional statistics with the storytelling ability of modern AI to produce safety reports that are both accurate and deeply understandable.

The challenge in this field is that flight data looks very different from the text these AI models are trained to read. While a language model can easily understand a sentence about a pilot pulling back on the control stick, it struggles when fed a long list of raw numbers representing vertical speed or altitude. If you simply feed these numbers directly into the AI, it often gets confused, treating a number like a word it has never seen before, which leads to unreliable answers. To solve this, the researchers first acted as translators. They took the complex, continuous streams of numbers from the flight recorders and converted them into qualitative descriptions. Instead of feeding the AI a raw value like "vertical speed minus 800 feet per minute," the system first analyzed the data to determine if that speed was "slightly high," "extremely high," or "normal" based on historical patterns. This process transformed abstract math into a language the AI could reason with, allowing it to focus on the meaning of the data rather than getting lost in the digits.

However, even with this translation, the AI still faced a limitation: it is not naturally good at making precise classifications based on numbers. To fix this, the researchers introduced a "statistical expert" into the process. They used a traditional, highly reliable computer model to look at the same data first and make a preliminary judgment. This judgment was then passed to the language model as a hint, or a piece of prior knowledge. Think of it as a junior analyst consulting a senior engineer before writing a report; the AI doesn't blindly follow the hint, but it uses it to anchor its reasoning. If the AI and the statistical expert agree, the explanation becomes more confident. If they disagree, the AI is prompted to look closer at the evidence to resolve the conflict. This collaboration ensures that the final output is grounded in statistical reality while retaining the ability to explain the "why" in plain language.

The system also had to deal with the fact that dangerous hard landings are rare events, meaning there is very little data to teach the AI what they look like. To overcome this scarcity, the researchers gave the AI a form of context by showing it examples of both normal and hard landings that were very similar to the flight it was currently analyzing. By asking the AI to compare the current flight against these specific, similar examples, the system could spot subtle differences that might otherwise be missed. This approach allowed the AI to learn from a few examples rather than needing thousands, effectively teaching it to distinguish between a rough landing and a safe one by looking at the nuances of the data.

When the team tested FlightLLM on their dataset of 704 flights, the results were compelling. The system correctly identified hard landings with a high degree of accuracy, matching or exceeding the performance of more complex, traditional deep learning models. But the true breakthrough was in the explanations it generated. In one detailed case study, the system analyzed a specific hard landing and correctly identified that the pilot had waited too long to perform a crucial maneuver called a "flare," which is the gentle raising of the nose just before touching down. The AI explained that this delay, combined with a sudden, sharp correction of the aircraft's pitch and some turbulent wind, caused the plane to hit the ground too hard. It did not just list the numbers; it described the pilot's actions, the timing of the error, and even suggested what the pilot could have done differently to avoid the incident, such as initiating the flare maneuver earlier.

The study also included a careful look at what happens when parts of the system are removed. When the researchers tested the AI without the translation of numbers into words, its ability to make accurate predictions dropped significantly, proving that the language model needs the data to be in a format it can understand. Similarly, when the "statistical expert" hint was removed, the system became less reliable, showing that the AI benefits from the guidance of traditional models. These tests confirmed that each part of the system plays a vital role in creating a tool that is both smart and trustworthy.

Ultimately, this work suggests a new path for aviation safety. Instead of relying on black-box models that give a result without a reason, or on human experts who must spend hours manually interpreting complex charts, FlightLLM offers a way to get immediate, clear, and actionable insights. It bridges the divide between the cold precision of data and the nuanced understanding of human behavior. By turning flight data into a narrative that explains the causes of safety events, the system provides a valuable tool for training pilots and preventing future accidents. The researchers acknowledge that while their current system works well for hard landings, it would need to be adapted for other types of flight risks, and that further work could be done to fine-tune the AI specifically for aviation. However, the core finding stands: by guiding large language models with domain knowledge and statistical hints, it is possible to create an AI that not only sees the danger but can explain it in a way that helps keep the skies safe.