Nuclear Power Plant Accident Diagnosis Algorithm Including Novelty Detection Function Using LSTM

Diagnosis of the accident or transient at the nuclear power plants is performed under the judgment of operators based on the procedures. Although procedures given to operators, numerous and rapidly changing parameters are generated by measurements from a variety of indicators and alarms, thus, there can be difficulties or delays to interpret a situation. In order to deal with this problem, many approaches have suggested based on computerized algorithms or networks. Although those studies suggested methods to diagnose accidents, if an unknown (or untrained) accident is given, they cannot respond as they do not know about it. In this light, this study aims at developing an algorithm to diagnose the accidents including “don’t know” response. Long short term memory recurrent neural network and the auto encoder are applied for implementing the algorithm including novelty detection function. The algorithm is validated with various examples regarding untrained cases to demonstrate its feasibility.