Optimising approach to designing kernel PCA model for diagnosis purposes with and without a priori known data reflecting faulty states

Abstract

Fault detection plays an important role in advanced control of complex dynamic systems since precise information about system condition enables efficient control. Data driven methods of fault detection give the chance to monitor the plant state purely based on gathered measurements. However, they especially nonlinear, still suffer from a lack of efficient and effective learning methods. In this paper we propose the two stages learning algorithm for designing the kernel Principal Component Analysis (kPCA) model parameters in two cases: with access to data reflecting the faulty states of the plant and without such data. The method is explained on simple testing example and verified in the case study showing the efficiency of detecting the leakages in drinking water distribution systems.