Learning fault patterns from data in streaming applications

In sensor-based systems, spatio-temporal data streams are often related in non-trivial ways. For example in avionics, while the airspeed that an aircraft attains in cruise phase depends on the weight it carries, it also depends on many other factors such as engine inputs, angle of attack and air density. It is therefore a challenge to develop failure models that can help recognize errors in the data, such as an incorrect fuel quantity or an incorrect airspeed. In this paper, we learn failure models from data streams using two approaches: (1) create static parameterized functional models using regression and use them to detect error modes based on error signatures, and (2) create statistical models based on the näıve Bayes classifier that dynamically update themselves to detect new error modes online. We extend our own PILOTS system to evaluate the accuracy, response time, and adaptability of these learning techniques. While error signatures can be more accurate and responsive than Bayesian learning, the latter method adapts better due to its data-driven nature.