抄録
A common challenge at nuclear power plants is to ensure that routinely-collected data is fully utilised. Data analytics provides an opportunity for improvements in prognostics and health monitoring by identifying correlations in related datasets without major capital investment. This paper describes work focused on improving the fuel defect identification process in Bruce Power’s eight CANDU nuclear reactors in the province of Ontario, Canada. The CANDU reactor comprises individually-pressurised horizontal channels which can be refuelled without taking the reactor offline. The detection and location of fuel defects is typically achieved using two systems: the first monitors the primary coolant for the presence of fission products and specific radionuclides, and is used to detect the presence of fuel defects within the core. The second system is deployed periodically and uses the emission of delayed neutrons to identify the channel containing defect fuel. In this paper we focus on improving the assessment of online delayed neutron monitoring data, with the aim of reducing the time period between initial detection of an in-core defect to identification and removal of the damaged fuel. The existing process is manually intensive, reliant on the judgement of a domain expert and operating experience demonstrates that this time period varies considerably. A first stage of investigation examines potential improvements to the current data processing system: developing and applying new analytic techniques in this area has shown promising results, with some fuel defects potentially identified several days earlier than the current system. Results from some short representative case studies are presented and further work will consider a larger volume of data. In addition, an extensive historical dataset is available which spans several years. In the second stage of investigation, the paper explores previously undocumented trends in this data and discusses the potential to produce correlations with other reactor parameters. The application of this knowledge can lead to opportunities in the use of Machine Learning algorithms to allow more accurate predictions to be made.
