Accounting for modelling errors in model-based diagnosis by using Gaussian process models

Imperfections in process models, if ignored, may affect reliability of the diagnostic system, for example through excessive false alarm rates. The problem tackled in this paper is how to handle unmodelled effects, caused by imperfect nominal models. We propose a covariance model, which builds on the idea that bias in nominal model can be described by a stochastic process. The Gaussian process model is used to capture the discrepancy between reality and nominal model, hence resulting in a refined model of the plant. With unmodelled dynamics as a second-order stochastic process, fault detection problem reduces to the problem of statistical decision making. The onset of a fault is inferred by comparing the statistical pattern of the residuals, collected under current operating mode with the pattern in the nominal (fault-free) condition. Major novelty of the approach resides in employing Jensen-Renyi divergence as a means to express the "distance" between the two corresponding ensembles of distributions. The ideas of the approach and their potentials are demonstrated on a simulated solid oxide fuel cell system.