Abstract
Fault detection methods based on empirical models such as principal component analysis (PCA) or support vector machines (SVM) are compared. As an example of realistic plants, absorption chiller system is chosen for comparisons. Several datasets whose distribution show different degrees of nonlinearity are generated based on an empirical model which describes heat exchanger efficiency. Sample numbers for model training is crucial on fault detection ability of models, especially for SVM. Nonlinear distribution of observed plant data deteriorates fault detection sensitivity of linear PCA model, whereas it does not affect SVM model.
