Quantification of Signal Reconstruction Uncertainty in Fault Detection Systems

Abstract

In Condition-Based Maintenance (CBM), Fault Detection (FD) systems monitor the health state of the components and aid the operator to decide whether a maintenance intervention is necessary. A FD system is a decision-aid tool typically based on i) a reconstruction model that estimates (reconstructs) the values of measurable signals in normal conditions, and ii) an analyzer of the differences (residuals) between the measured and reconstructed values: abnormal conditions are detected when residuals are statistically significant. The performance of the reconstruction model is influenced by several sources of uncertainty which can influence the operator decision: 1) measurement errors, 2) intrinsic stochasticity of the physical process, 3) uncertainty on the settings of the model parameters, and 4) uncertainty on the model output due to incompleteness of the training data. The objective of the present work is the quantification of the overall uncertainty affecting the model reconstructions. The proposed novel approach for uncertainty quantification relies on the estimation of Prediction Intervals (PIs) by using Order Statistics (OS) fora pre-defined confidence level. The proposed approach is verified with respect to an artificial case study; the obtained results show that the approach is able to guarantee the desired level of confidence on the correctness of the detection and provide the decision maker with the required information for establishing whether a maintenance intervention is necessary.