Application of an open-set recognition method for detecting untrained accident scenarios in a nuclear power plant accident diagnosis model

Nuclear power plant (NPP) operators conduct event/accident diagnoses under specific circumstances to ensure the safety of the plant. As inappropriate or delayed diagnosis of an event or accident may result in severe consequences, fast and accurate diagnosis is considered as one of the most important role for operators. With the rapid progress in artificial intelligence, various NPP event/accident diagnosis models have been proposed to aid fast and accurate diagnostic actions of operators. However, most of the existing diagnosis models have common problem that they always deduce results among trained event/accident scenarios even when the supplied input belongs to an untrained scenarios, as they have been developed based on closed-set classification concept. Since this problem may induce confusion to the operators and result in severe consequences, it is necessary to introduce open set recognition method for event/accident diagnosis models to make them capable of detecting untrained scenarios. In this study, an open-set recognition (OSR) method, namely OpenMax, was applied to detect untrained scenarios in an NPP accident diagnosis model. The OpenMax method was selected owing to its simple but concrete concept and advantage that the method can be applied without modifications to the model. During the experiments, accident diagnosis models were developed by intentionally neglecting specific accident scenarios, and the OpenMax method was applied to determine whether the method is capable of detecting untrained scenarios. The obtained experimental results indicate that the appropriate application of the OpenMax method to the model enables the detection of data from untrained scenarios with high accuracy, without deteriorating the classification performance of trained scenarios significantly. The experiments also revealed that the OSR performance was significantly affected by model configurations and hyperparameters, highlighting the importance of appropriate setting of these factors.