FAULT DETECTION AND DIAGNOSIS OF PROCESS DATA

Abstract

The thesis is concerned with the development of a novel data-driven method for fault detection and diagnosis using process data. This method involves the use of univariate charts for fault detection and the concepts of cross-correlation and Granger causality for fault diagnosis. The performance of this method is demonstrated by applying it on benchmark examples such as CSTR and Tennessee Eastman process. Established methods such as Principal Component Analysis (PCA) and Dynamic PCA are implemented and used for the purpose of comparison. Different investigations are conducted in each method with the aim of improving the performance. Guidelines for the practical implementation of on-line frameworks for each of the methods used are provided. The methods used for calculating performance measures are discussed. The best results obtained from each method are compared against one another and insights are reported.