The application of machine learning for on-line monitoring Nuclear Power Plant performance

The aim of the paper is focused on the development of on-line monitoring strategy and predictive methodology to analyse the performance of the nuclear system and components. Advancing online monitoring is attracting a lot of interest at nuclear power plants operating today as it involves the transition from traditional monitoring techniques of nuclear power plant, gathering via manually recorded data sheets, to a full embrace of digitalization. In this research, a conceptual framework for the application of digital twin technology to primary nuclear power plant component prognosis and maintenance process is proposed in order to reduce its failure risk that could, in turn, affects plant operations and safety. The development of machine learning algorithm for automated diagnostics and prognostics that, for example, may allow the transition from time-based to condition-based maintenance of the nuclear plant, is totally new and innovative. No prior knowledge of machine learning for on-line monitoring of nuclear items performance in the open literature is known. The methodology uses big data from sensors and logical controllers for training machine learning algorithm to recognize anomalies or useful pattern before components failure. Due to the limited available data on primary nuclear components, digital twin concept is adopted in order to generate them for different plant conditions through numerical simulation. After that, the trained algorithm is capable to predict the performance of nuclear components anticipating or delaying the planned inspection for their repairing/replacing. This approach may support the plant condition-based predictive maintenance optimization and the development of the "digital twin model" for improved plant safety and availability.