Condition-Based Maintenance Modelling and Decision Making taking into account Uncertainties

International audienceOver the last few decades, Condition Monitoring (CM) techniques have been strongly developed in terms of measurement devices, data processing and management capabilities. These developments have encouraged industries like nuclear, oil & gas, automotive and chemical to apply Condition-Based Maintenance (CBM) for increasing system availability, reducing maintenance costs, minimizing unscheduled shutdowns and increasing safety. CBM resorts to Fault Detection (FD) system, which delivers information on the health state of the monitored component through an alarm system interface. On the basis of the received information, the operator decides if it is necessary to perform a maintenance action or if it is possible to postpone it. In this context, one of the objectives of the European Union Project INNovation through Human Factors in risk analysis and management (INNHF, www.innhf.eu) is the development of an integrated framework of CBM modelling and Decision Making taking into account uncertainties in both the FD system and the decision making. The work here presented focuses on the first aspect of the analysis of the uncertainties that are originated in the fault detection phase. The fault detection system is considered to be based on i) an empirical model (typically called signal reconstruction model) that estimates (reconstructs) the values of measurable signals such as temperature, pressure and flow rate, in normal conditions, and on ii) statistical method to analyze the difference between the measured and reconstructed values: the presence of an abnormal condition is concluded, when the difference between the measured and reconstructed values is (statistically) significant and, then, an alarm signal is triggered. A signal reconstruction model has typically to deal with the following sources of uncertainty: 1) sensor measurement errors; 2) model uncertainties due to incorrect parameter settings or simplifying hypothesis and assumptions; 3) intrinsic stochasticity of the physical processes occurring in the plant. All these uncertainties can reduce the accuracy of the conclusions drawn from FD, and may lead the operator to perform wrong actions. In this work, a novel approach for the quantification of the effects of uncertainty on the signal reconstruction is proposed; the model considered for the reconstruction of the signals in normal condition is the Auto-Associative Kernel Regression (AAKR) model. The approach, then, relies on the estimate of the prediction intervals of the signal reconstructions with a predefined confidence level. The approach is based on the definition of a scale factor and on the use of order statistics. The proposed method has been tested using an artificial case study and will be validated using real data containing signals measured during normal operation of an healthy plant component