Domain Adaptation for Fault Detection in Civil Nuclear Plants

Abstract

Recent domain adaptation approaches have been shown to generalise well between distant data domains achieving high performance in machine fault detection through time series classification. An interesting aspect of this transfer-learning inspired approach, is that the algorithm need not be exposed to fault data from the target domain during training. This promotes the application of these methods to environments in which fault data is unfeasible to obtain, such as the detection of loss-of-coolant accidents (LOCA) in nuclear power plants (NPPs). A LOCA is a failure mode of a nuclear reactor in which coolant is lost due to a physical break in the primary coolant circuit. If undetected, or not managed effectively, a LOCA can result in reactor core damage. Three high-fidelity physics based models were created with divergent behaviour that represent different data domains. The first model is used to generate source domain data by simulating labelled training data under both nominal and LOCA conditions. The second and third models act as surrogates of real plants and are used to generate target domain data, i.e. to simulate nominal data for training and LOCA condition data for validation. Several deep-learning feature encoders (with varying levels of connectivity) were applied to this LOCA detection problem. Among these, a ’Baseline’ encoder was used to quantify the improvement that domain adaptation techniques make to LOCA detection performance under large domain divergences. Classification accuracy for each model is explored within the context of LOCA break size and location within each plant model. The proposed method for LOCA detection demonstrates how the dependence upon sparse accident-specific data can be alleviated through the use of domain adaptation. Detection capability of the LOCA condition is maintained even when no data examples are available in the target domain.