Application of Dimension Reduction and Clustering Methods for Detection of Faulty Operations in Process Systems

Widespread application of distributed control systems and measurement technologies in chemical plants are prerequisites for quality control and safety monitoring of processes. This consequently has prompted large-scale data acquisition and storage from different processes and various sectors of a plant. The collected data holds information about the behavior of the process which can further assist in developing data-driven methods for detection and diagnosis of process anomalies (faults). Considering the recognition data-driven modeling has received in the past decades, additional studies on incorporation of data mining techniques for knowledge discovery from chemical process data would seem to be a useful practice.Data mining methods can be used to identify different groups or classes present in a dataset, or in the context of process systems engineering, distinguish faulty states from normal process operation in historical datasets. Two key tools used in this practice are dimension reduction techniques and clustering methods. The former helps with feature extraction from the data and the latter detects groups within the data, therefore, a productive combination of these two tools can be promising in facilitating the fault detection and diagnosis applications. The first part of this research studies the performance of different combinations of dimension reduction techniques and clustering methods to evaluate their ability in detection process faults, and demonstrates the higher compatibility of some of these methods with others. While performing clustering, detecting the number of clusters present in the dataset is either a direct aim of the study, or it can be beneficial in choosing the most suitable parameters and/or labelling. Motivated by the first part of the research, performance of clustering methods on a dataset before and after different dimensionality reduction techniques is studied using internal metrics for clustering performance. Based on multi-objective optimization, an approach is proposed to detect the cluster numbers in an unsupervised manner which successfully presents the expected cluster numbers in three distinct applications. In summary, this research aims to assist data-driven fault detection practices in chemical processes by elucidating the synergy between two data mining techniques; dimension reduction and clustering methods. Also, the introduced approach to detect the expected cluster number in a dataset contributes to the progress of unsupervised state-isolation studies for any application and generic datasets