Analyzing high-dimensional thresholds for fault detection and diagnosis using active learning and Bayesian statistical modeling
Yuning He · Annual Conference of the PHM Society · 2015
Many Fault Detection and Diagnosis (FDD) systems use dis- crete models for fault detection and analysis. Complex indus- trial systems generally have hundreds of sensors, which are used to provide data to the FDD system. Usually, the FDD wrapper code discretizes each sensor value individually and ignores any non-linearities as well as correlations between different sensor signals. This can easily lead to overly con- servative threshold settings potentially resulting in many false alarms. In this paper, we describe an advanced statistical framework that uses Bayesian dynamic modeling and active learning techniques to detect and characterize a threshold surface and shape in a high-dimensional space. The use of active learning techniques can drastically reduce the effort to study threshold surfaces. Automated Bayesian modeling of complex thres- hold surfaces has the potential to improve quality and perfor- mance of traditional wrapper code, which often uses hyper- cube thresholds.