Kernel approaches for fault detection and classification in PARR-2

Abstract

Safety and reliability of nuclear power plants is of utmost importance. For that purpose, modern fault detection and classification (FDC) techniques are being devised to compliment the existing hardware redundancy and limit checking techniques. Among these modern techniques, Fisher discriminant analysis (FDA) and support vector machines (SVM) have been shown to be successful for FDC of nuclear reactors. By considering the fact that both FDA and SVM are basically established for linear systems and nuclear reactors are highly nonlinear processes, it becomes more intuitive to utilize some nonlinear FDC technique. To this end, application of kernel based non-linear approaches including kernel FDA (KFDA) and kernel SVM (KSVM) is proposed in this paper for fault detection and classification in Pakistan Research Reactor-2. Control rod withdrawal and accidental external reactivity insertion faults are manually executed at PARR-2, and training data is collected from the reactor based on which KFDA and KSVM models are developed. The online data is subsequently tested using the developed models which resulted into reliable fault classification.