Abstract

The ability to fly has become more and more important since the invention of the airplane. Not only the importance of commercial aviation has grown, but especially that of military aviation. There is a continuous demand for better, faster and more manoeuvrable airplanes. As a result, flying an airplane has become more and more complex and the danger of a pilot making a mistake has become bigger and bigger. To prevent a pilot from making such mistakes during a flight cockpits have been designed as pilot friendly as possible and a lot of the systems in the cockpit have been automated. Research is being done constantly to new cockpit systems. That research has focused on intelligent cockpit environments. Such environments usually they filter the information for the pilot and present him with only the most relevant information for the current situation to keep the workload of the pilot as low as possible. But they can also monitor the pilot and inform him if he makes a mistake.
This report tries to answer a question that is relevant for most of those intelligent systems: Is it possible to design a system that recognizes, in real time, the situation a pilot is in based on the actions of the pilot, the state of the airplane and information from the environment?
The system that is described focuses on detecting the current situation when flying an F-16, but it is designed to be able to work with other airplanes as well. In this report three possible models for the system are described: a finite state model, a probabilistic model and a causal model. A prototype has been built and tested which is based on the probabilistic model. This model uses a knowledge base containing data that describes a number of situations that the pilot might encounter. The system will convert the knowledge in the knowledge base to a set of rules that compare the information in the knowledge base with the state of the real world at a certain moment. Based on this comparison the rules will generate probabilities that the situations given in the knowledge base are occurring. These probabilities are combined using Bayesian belief networks, which will produce for every situation the probability that it is occurring and the probability that it is not occurring. The system will then decide which of the situations is most likely to be occurring.
The prototype was tested using Microsoft Flight Simulator 2002 to simulate a flight with an F-16. The program performed very well. During the tests it became clear though that the program still made some mistakes incidentally. After analysing those mistakes the conclusion was that the main cause for these mistakes was that the probabilistic model could not account for the relationships between actions or events over a period of time.
The final conclusion is that a situation recognition system based on a probabilistic model performs very well, but that the performance of the system would be even better if it was based on a causal model.