Detection and Modeling of high-dimensional Thresholds for Fault Detection and Diagnosis

Many Fault Detection and Diagnosis (FDD) systems use discrete models for detection and reasoning. To obtain categorical values like ”oil pressure too high”, analog sensor values need to be discretized using a suitable threshold. This task is usually performed by the “wrapper code” of the FDD system. In practice, selecting the right threshold is very difficult, because it heavily influences the quality of diagnosis. In many cases, proper thresholding needs to be performed using non-linear highdimensional threshold surfaces to accommodate dependencies between system components and different sensors. Often those dependencies are complex and can not handled analytically. In this paper, we will describe a statistical modeling technique using hierarchical Bayesian methods for the detection of threshold surfaces using a low number of necessary simulation experiments.