Data-driven methods for process analysis
The thesis is concerned with the development of data-driven methods for fault diagnosis of plant-wide disturbances. Industrial plants producing large quantities of liquid or gaseous chemicals run continuously and under tight cost, safety, quality and environmental constraints. Any unwanted variability in the form of disturbances in the production process affects one of these constraints. Worse still, disturbances can spread and cause large parts of the process to be upset. Detection and diagnosis of disturbances is therefore an important subject for chemical companies. Chemical processes are well equipped with modern instrumentation technology so that measurements of process variables such as flow, temperature or pressure are abundantly available. The research has used time series analysis for process measurements in a novel way. In particular, it focuses on measures to decide about cause and effect of processes variables to address the question whether A causes B or B causes A. Knowing the causal relation ship finds the fault propagation path in case of a disturbance and eventually traces the disturbance back to the root cause. Three different measures are investigated and developed for the application to chemical process data: one straight forward algorithm based on the cross-correlation function and two statistics based on nearest neighbours methods and transfer entropy. Together with the automatic generation of causal maps these approaches lead to a breakthrough in fault diagnosis. Guidelines for the parameters of the methods are tailored to signatures caused by disturbances common in chemical processes. A significance level is introduced for automatic implementation of industrial applications. Case studies with process disturbances, in particular those from a three months placement with Eastman Chemical Company, are analysed with the developed tools. The results are compared and recommendations of choosing the best method for a data set are generalised from results of the case studies.