A novel fault detection system taking into account uncertainties in the reconstructed signals

Abstract

A typical fault detection (FD) system comprises: (1) a model that reconstructs the values of the measured signals in normal conditions, (2) a technique for the analysis of the differences (residuals) between the measured and reconstructed values, and (3) a decision strategy for defining when the monitored situation is to be detected as anomalous, i.e., reflecting a fault. Traditional techniques for this task, like threshold-based methods and the Sequential Probability Ratio Test (SPRT), show difficulties in setting their parameters and in providing information on the confidence of the FD system outcomes. In this context, the objective of the present work is to develop a novel, non-parametric, sequential decision strategy to decide whether the component is in normal or abnormal conditions that takes into account the quantified uncertainty on the reconstructions in the form of Prediction Intervals (PIs). The Auto-Associative Kernel Regression (AAKR) method is adopted to build the empirical model of signal reconstructions. The novel FD system has been tested using an artificial case study representing the monitoring of a component during typical start-up transients and it is validated using a real industrial case concerning 27 shut-down transients of a nuclear power plant (NPP) turbine. The obtained results show that the approach is able to guarantee low false and missing alarm rates and, hence, provide the decision makers with robust information for establishing whether a maintenance intervention is required or not.