Detecting loss-of-coolant accidents without accident-specific data

This paper develops an automated fault detection tool to detect very small LOCAs in pressurized water reactors that would be difficult for operators to detect manually. One of the primary challenges with previous automated fault detection methods, which are data-driven, is that they require data from LOCAs; however, it may be difficult to capture real operational data from LOCA scenarios. This work uses a physics-inspired approach that equates the physical effects of a LOCA to changes in known variables. This approach enables the detection of very small LOCAs using data-driven approaches that use nominal operating data without the need for LOCA data. The approach combines data-driven modeling with control-theoretic estimation techniques to detect LOCAs and estimate their magnitudes in real-time. First, simulated process data for a variety of nominal operating conditions is collected using a generic pressurized water reactor simulator. Then, that data is used to train an artificial neural network regression model that captures the nonlinear plant dynamics. Finally, the regression model is used in a particle filter to detect the onset and estimate the magnitude of the leak. These methods are successfully verified using LOCA simulations that would be hard to manually distinguish from normal operating transients.