SVR optimization with soft computing algorithms for incipient SGTR diagnosis

• Published in: Annals of nuclear energy Vol. 121; pp. 89 - 100
• Main Authors: Ayodeji, Abiodun, Liu, Yong-kuo
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.11.2018
• Subjects: Fault diagnosis; Feature selection algorithms; Steam generator tube rupture; Support vector regression; Fault diagnosis; Support vector regression; Feature selection algorithms; Steam generator tube rupture
• ISSN: 0306-4549; 1873-2100
• DOI: 10.1016/j.anucene.2018.07.011

•Proposed a hybrid of N-16 method and optimized SVR for incipient fault diagnosis.•Details soft computing optimization techniques for SVR model.•Optimized SVR model performance evaluated on CNP300 ruptured steam generator tubes.•The proposed method diagnoses SGTR faster than conventional methods. Fault severity awareness and fault identification are some of the key steps to a successful diagnosis in nuclear power plants. Currently, faults such as leak detection are being done using the N-16 method. However, traditional leak monitors are not sensitive to small leak rate changes, hence cannot be used for low-level leak rate detection under incipient fault conditions and are limited to post-accident analysis of significant releases. In this work, we present a diverse and implementable data-driven Support Vector Regression (SVR) model whose capability compensates for the weaknesses in the already established N-16 methods in the nuclear plant. The method can be integrated with the conventional N-16 method to form a robust hybrid diagnostic system, effective for detecting both incipient and large leakage in the steam generator. The purpose of the SVR model is to estimate uncertain parameters that are sensitive to certain faults, and the parameter estimation efficiency is evaluated using the mean squared error values (MSE). To obtain efficient predictive model capable of supporting decision-making process and to further optimize the model, minimize false alarm rate and reduce computation cost, we also utilized Particle Swarm Optimization algorithm, Sequential Feature Selection algorithm, and Genetic Algorithm for feature selection purposes. To demonstrate the method and evaluate the predictive model, we simulated steam generator tube rupture (SGTR) faults with varying severity in the reactor coolant system of CNP300 NPP, with RELAP5/SCDAP Mod4.0 code. The SVR’s relative error (MSE) with and without feature selection algorithms were compared using different solver algorithms. The feature selection performance of the algorithms and the resulting SVR model fault diagnosis performance evaluation are discussed in this paper.