Classification of NOTAMs for SWISS International Airlines using AI

Flying today is safer than ever, thanks in part to the high standards of communication.

Jean Coupon, Astrophysicist and former Data Science student at SIT Academy, worked with SWISS International Airlines in his three-week Capstone project to classify aviation warning messages using Artificial Intelligence. The communication system is called "Notice To Airmen" (NOTAM) and consists of short text messages sent by airspace officials to warn pilots of anticipated events that could disrupt a flight on route (closed runways, construction, closed airspace, etc.). Thousands of NOTAMs are sent out every day (and the number is growing). One of the reasons for this rapid increase is an ever lower safety threshold that triggers a new NOTAM, sometimes resulting in irrelevant NOTAMs. As a result, each pilot has to search and sort through an increasing number of messages, which could lead to a risk of missing the important ones. At bigger airlines, a NOTAM officer is responsible for pre-screening NOTAMs before they are issued to pilots.
 

The Data Science team at SWISS International Airlines and Jean worked to develop an automated NOTAM classifier to help identify the most important messages using Machine Learning and Natural Language Processing (NLP) to save time while ensuring a high level of safety.

The challenge in this project was to determine which messages were relevant from around 3,000 NOTAMs that a NOTAM officer receives per day. It has shown that approximately 50% of all incoming messages are important to communicate to the pilot (70-150 NOTAMs per flight).
 

Essentially, the NOTAMs were labeled in a first step with an unsupervised Machine Learning approach, which can be divided into three further steps:
 

Afterwards comes sorting by importance. Jean trained a model (supervised Machine Learning algorithm) to define an Artificial Intelligence system that sorts importance probabilities. The main goal was to support NOTAM officers by pointing out very important notes.
 

The final model which Jean has tested had an accuracy of 94% using a Neural Network with only one month data of NOTAMs.

In summary, a model has been developed to distinguish between NOTAMs messages that are relevant and irrelevant for the day of operations. In the future, this project will be continued as follows: more data will be evaluated to establish which clusters of messages can be classified with close to 100% accuracy. The remaining low-certainty messages (about 20 - 30% of the volume) will be still evaluated on a daily basis by the expert. A feedback loop from expert decisions further improves accuracy of the model. Other steps to be taken include industrialization of the solution and user feedback.