Decision Versus Identification Issues in Fault Detection/Isolation for Predictive Maintenance

Abstract The purpose of this paper is to emphasize the key role of decision issues and their connection with identification issues in statistical methods for fault detection and diagnosis (or isolation). We first describe a general approach to the design of statistical decision rules for fault detection and isolation in industrial plants, starting from a parametric characterization of the plant We then report the main performances of the resulting algorithms for two applications: vibration monitoring for mechanical structures or machines, monitoring the combustion chambers in power plant gas turbines. Finally we discuss decision versus identification issues. We outline why fault detection cannot be reduced neither to repeated identification of the parameters nor to innovation monitoring. We explain how to decide that the discrepancy between the identified parameter values and the reference ones is significant, especially in the presence of noises and disturbances on the system. We also outline that the early warning and isolation of small faults can be obtained even with biased identified parametric models.