Process fault detection and isolation based on symbolic data or interval-valued principal component analysis

Abstract

Principal component analysis (PCA) is a well-known data-driven method that has extensively been used as a fault detection technique for the last three decades. Aside from the non-linearity property, processes today are associated with measurement uncertainties and dynamic properties. The standard PCA method cannot acknowledge these uncertainty and/or dynamic features, let alone incorporate them into the fault detection model. The dynamic PCA has been proposed in the literature to take care of the dynamic properties of real processes in an effort to build a robust fault detection model. On the other end of the spectrum, multiple variants of PCA methods have been developed for interval-valued data. Since interval data, which are a type of symbolic data, are capable of modeling measurement errors and uncertainties, these proposed interval PCA methods prove helpful for modeling systems with sensor imprecision and uncertainties. Still, they cannot handle dynamic properties as the dynamic PCA method did. In this thesis, different interval PCA methods have been investigated to detect faults in real processes. Being capable of acknowledging measurement uncertainties, these interval PCA methods produce better performance than their classical counterpart. Three of these interval PCA methods have been extended to include dynamic properties—a treat that existing methods in the literature did not accomplish. Included in this manuscript is an extension of the combined index to the intervalvalued case where it has been implemented and tested with common interval-valued PCA methods. Moreover, the contribution plot isolation method has also been extended to the interval-valued case for the purpose of isolating faulty variables when using interval PCA methods. Real data from a cement plant and a grid-connected photovoltaic system have been used to apply and test the proposed techniques