Learning models from avionics data streams

Abstract

In aviation, there are many values that are useful for pilots to know that cannot be measured from sensors and require calculation using various charts or other means to accurately estimate. Aircraft take-off distance requires knowing wind speed, pressure altitude, and temperature. However, it is possible for the inputs of these calculations to change during flight, or be calculated incorrectly by mistake, and it would be useful for pilots to know these values in real-time. PILOTS is a programming language for spatio-temporal data stream processing. We have added improved integration for machine learning algorithms as well as a linguistic abstraction for training these models. In data-driven systems, it can be useful to use distributed processes for computation. We have designed a declarative framework for federated learning and the aggregation of results from multiple related models within PILOTS. Furthermore, we built a model using PILOTS that is able to estimate weight in real-time during take-off of a fixed-wing aircraft using data available from the avionics. We evaluated the results of several models on accuracy and timeliness. Data was collected from the flight simulator X-Plane. Accidents such as the fatal crash of Cessna 172R N4207P could have been prevented using the weight estimation methods illustrated.