Abstract
Nonlinear relationships among different process variables are very common in nuclear reactors. When constructing a nonlinear model to for process monitoring, linear-based monitoring technique has been proved inefficient and problematic. Kernel principle component analysis (KPCA) combines the principle component analysis and kernel method to cope with fault diagnose in nonlinear system. Experiments are conducted using data samples from Generic Pressurized Water Reactor Simulator(GPWR) to simulate sensor fault and three types system faults. The nonlinear relationship among monitoring variables is analyzed. KPCA is used to capture the nonlinear relationships to distinguish the sensor fault and system fault, which shows the superiority of KPCA to detect a univariate fault. Subsequently, system faults are further analyzed as a classification problem.
Samples imbalance is another common problem existing in diagnose domains. When training samples of one class outnumber than other class, traditional classification algorithm performance may deteriorate. Random under-sampling boosting (RUSboost), a hybrid sampling and boosting algorithm is used to alleviate the imbalance problem. The performances of classifiers by four metrics is evaluated. When compared to Adaboost (contains only boosting), tree variants (base model and ensemble bagged trees) and Gaussian SVM, RUS boosting shows significantly high accuracy and robust performance in dealing with different ratio of training samples, which shows that RUSboost is an excellent technique to cope with imbalanced data in fault diagnose for nuclear power plants.
