Towards Learning Spatio-Temporal Data Stream Relationships for Failure Detection in Avionics

Spatio-temporal data streams are often related in complex ways, for example, while the airspeed that an aircraft attains in cruise phase depends on the weight it carries, it also depends on many other factors. Some of these factors are controllable such as engine inputs or the airframe’s angle of attack, while others contextual, such as air density, or turbulence. It is therefore critical to develop failure models that can help recognize errors in the data, such as an incorrect fuel quantity, a malfunctioning pitot-static system, or other abnormal flight conditions. In this paper, we extend our PILOTS programming language [1] to support machine learning techniques that will help data scientists: (1) create parameterized failure models from data and (2) continuously train a statistical model as new evidence (data) arrives. The linear regression approach learns parameters of a linear model to minimize least squares error for given training data. The Bayesian approach classifies operating modes according to supervised offline training and can discover new statistically significant modes online. As shown in Tuninter 1153 simulation result, dynamic Bayes classifier finds discrete error states on the fly while the error signatures approach requires every error state predefined. Using synthetic data, we compare the accuracy, response time, and adaptability of these machine learning techniques. Future dynamic data driven applications systems (DDDAS) using machine learning can identify complex dynamic data-driven failure models, which will in turn enable more accurate flight planning and control for emergency conditions.