Self-Healing Data Streams Using Multiple Models of Analytical Redundancy

Abstract

We have created a highly declarative programming language called PILOTS that enables error detection and estimation of correct data streams based on analytical redundancy (i.e., algebraic relationship between data streams). Data scientists are able to express their analytical redundancy models with the domain specific grammar of PILOTS and test their models with erroneous data streams. PILOTS has the ability to express a single analytical redundancy, and it has been successfully applied to data from aircraft accidents such as Air France flight 447 and Tuninter flight 1153 where only one simultaneous sensor type failure was observed. In this work, we extend PILOTS to support multiple models of analytical redundancy and improve situational awareness for multiple simultaneous sensor type failures. Motivated by the two recent accidents involving the Boeing 737 Max 8, which was potentially caused by a faulty angle of attack sensor, we focus on recovering angle of attack data streams under multiple sensor type failure scenarios. The simulation results show that multiple models of analytical redundancy enable us to detect failure modes that are not detectable with a single model.