Multivariate Statistical Approaches to Fault Detection and Isolation

This paper presents an overview of the multivariate statistical approaches to process monitoring and FDI that have become widely applied in industry over the past decade. These approaches are based on latent variable methods such as Principal Component Analysis (PCA) and Projection to Latent Structures (PLS), and the models are built using data collected from routine plant operations during periods where only "common cause" variation is present. As a result of the nature of these data employed for modeling, the latent variable models are non-causal in nature. Information in the data is projected down into low dimension hyper-planes that summarize the covariance structure of the data under normal (common-cause) operation of the process. Plots based on statistics such as the scores, Hotelling's T2 and the multivariate residual magnitude (SPE), are then used to detect the occurrence of faults. Diagnosis of the faults relies upon using tools such as contribution plots or signal reconstruction that are based on examining the nature of the break-down in the common-cause covariance structure by interrogating the underlying PCA/PLS model.These approaches, based more around the concepts of multivariate statistical process control are fundamentally different from the analytical redundancy approaches used in much of the FDI literature. The latter approaches use causal models developed from theory or from identification experiments, that are more difficult to obtain, but are potentially capable of more incisive fault isolation. This paper will discuss some of the differences between these approaches and give the author's perspective on their complementary strengths and weaknesses.