Physical-Principle Based Extended Attributes for Process Fault Detection

Abstract

Process monitoring is of importance to maintain process safety, reliability, performance and cost efficiency. This work presents a hybrid fault detection approach that combines process knowledge such as first-principles and process causal relations into data-driven fault detection techniques. The process knowledge is embedded into the process dataset as the form of extended attributes (ExAs). In this paper, we discuss the benefits of adding process knowledge into the process data, as well as the procedure of extracting ExAs from available process information such as piping and instrument digraph. Our proposed method was successfully tested on the Tennessee Eastman Process using two commonly utilized data-driven fault detection techniques: principle component analysis and its variant kernel PCA.