Interpretable Unsupervised Feature Extraction and Learning of Abnormal System State Transitions in Aircraft Sensor Data

Abstract

Commercial aircrafts generate a huge amount of data during each flight by sampling hundreds of variables at different resolutions during all phases of flight. While having this enormous source of data is useful for learning of faulty system behavior, its huge dimensionality and size can be an impeding factor to such analysis.

To address this problem, we have devised a data-driven process that automatically extracts persistent, underlying latent states that can succinctly describe the data and thereby reduce its dimensionality, while preserving the most salient aspects important for fault or potential fault analysis. By analyzing how these latent states transition in time by computing a transition matrix for every leg, which we use as features, we can classify certain precursors which are indicative of a potential fault. Specifically, this is achieved by supervised and unsupervised learning of hundreds of latent state transitions for a given subsystem. Analysis of temporal dynamics of state transitions allows us to pinpoint at what time the sensor variables were behaving atypically in a flight leg, thus allowing airline maintainers to fix the faulty component quicker and avoid flight delays due to unplanned maintenance. We demonstrate our method of supervised and unsupervised classification over temporal dynamics of system state transition on two subsystems, the Fan Air Modulating Valve (FAMV) and the Flow Control Valve (FCV), and have obtained 100% true positive rate (for both systems) and a false positive rate of 0.05-0.08%.