Robust Model-Based Fault Detection Using Monte-Carlo Methods and Highest Density Regions

Abstract

One of the major challenges in model based fault detection is the robust design of thresholds for the analytical redundancy relations. Those relations are residuals which differ from a zero in case of a fault and are equal to zero in the fault free case. In real world applications, however, these residuals usually differ from zero even in the fault free case due to, e.g. measurement errors and model uncertainties. This paper proposes a method based on Monte-Carlo simulations of possible residuals taking into account uncertainties as a-priori probability distributions. The statistical analysis of the resulting residual's probability distribution using posterior highest density regions enables a likelihood-based decision about the occurrence of a fault. The presented method is demonstrated and evaluated using a nonlinear physical model of an air cooling system of an unmanned aerial vehicle.