Quantification of predicted uncertainty for a data-based model

• Main Authors: Jangbom Chai, Taeyun Kim
• Format: Article
• Language: English
• Published: Elsevier 2021-03-01
• Series: Nuclear Engineering and Technology
• Subjects: Model uncertainty; Predicted uncertainty; Data-based model; Drift monitoring; Sensor
• Online Access: http://www.sciencedirect.com/science/article/pii/S173857332030797X

A data-based model, such as an AAKR model is widely used for monitoring the drifts of sensors in nuclear power plants. However, since a training dataset and a test dataset for a data-based model cannot be constructed with the data from all the possible states, the model uncertainty cannot be good enough to represent the uncertainty of estimations. In fact, the errors of estimation grow much bigger if the incoming data come from inexperienced states. To overcome this limitation of the model uncertainty, a new measure of uncertainty for a data-based model is developed and the predicted uncertainty is introduced. The predicted uncertainty is defined in every estimation according to the incoming data. In this paper, the AAKR model is used as a data-based model. The predicted uncertainty is similar in magnitude to the model uncertainty when the estimation is made for the incoming data from the experienced states but it goes bigger otherwise. The characteristics of the predicted model uncertainty are studied and the usefulness is demonstrated with the pressure signals measured in the flow-loop system. It is expected that the predicted uncertainty can quite reduce the false alarm by using the variable threshold instead of the fixed threshold.