Multi-Label Classification with Embedded Feature Selection for Complex Abnormal Event Diagnosis

A nuclear power plant is the largest electrical power generation system composed of hundreds of components. When an abnormal situation occurs in a nuclear power plant, operators have to perform an appropriate diagnosis to alleviate the plant state. This abnormal event diagnosis process is based on the alarms and symptoms described in the abnormal operating procedures. However, when two or more abnormal events occur simultaneously, the plant parameters may show complex changes unlike the alarms and symptoms described. Abnormal event diagnosis models can be helpful greatly to operators when they can provide diagnostic information in more difficult situations such as these. In this study, the diagnostic performance of the existing artificial neural network model was improved by applying embedded feature selection to classify complex abnormal events. An embedded feature selection uses the feature importance of parameters used when a pre-prepared machine learning classifier trains a dataset. The parameters selected through this method only the characteristic parameters for each event so that the artificial neural network model can efficiently perform diagnosis. These results enable the abnormal state diagnosis model to provide diagnostic information to operators even in complex situations. In conclusions, this approach can increase the applicability of the diagnostic model using artificial neural networks to the actual operator support system for safer actual nuclear power plants.

Keywords: Nuclear power plant, Abnormal event diagnosis, Convolutional neural network, Multi-label classification, Embedded feature selection.

Download PDF