Analysis and implementation of certain multivariate statistical process monitoring Tools for fault detection and isolation (FDI) task in a laboratory scale shell-tube heat exchanger

Abstract

Early detection and diagnosis of faulty events in industrial processes can represent economic, social and environmental profits. When the process has a great quantity of sensors or actuators, the Fault Detection and Isolation (FDI) task is very difficult. Advanced statistical based FDI methods are extensively used for fault detection and isolation purposes. In this work, three multivariate statistical techniques such as neural network based Principal Component Analysis (PCA), neural network-based Fischer Discriminant Analysis (FDA) and Correspondence Analysis (CA) was applied to the multivariate data extracted from laboratory scale shell and tube heat exchanger. Performance metric such as detection delay, estimated time of occurrence of fault, misclassification rate was computed for those three techniques for the detection and isolation of biases in sensors and actuators. Correspondence Analysis was proven to perform better when compared to PCA and FDA. CA was observed to perform FDI with minimal detection delay (which is less than or equal to 7 seconds) and lower misclassification rate (which is less than or equal to 6%) in case of both sensor & actuator faults. PCA and FDA showed significant detection delay and missed alarm rate for single and multiple fault identification.