Semiparametric PCA and bayesian network based process fault diagnosis technique

Abstract

AbstractSemiparametric Principal Component Analysis has advantages over Principal Component Analysis (PCA), as it can deal with nonlinear and non‐monotonic correlation and non‐Gaussian distribution process data. In Semiparametric PCA the distance correlation coefficient matrix is used to replace the covariance matrix, and a semi‐parametric Gaussian transformation is used to allow variables to follow multivariate Gaussian distribution. To reduce the cost of monitoring and alarm flooding, a fault diagnosis technique, which combines Semiparametric PCA and Bayesian Network (BN), is proposed here. In the first stage, Semiparametric PCA is used to find the fault in monitored variables. Considering the interaction of process variables and historical process data, a Bayesian network is developed in the second stage. Considering Semiparametric PCA outcome as evidence, the Bayesian network applies deductive and abductive reasoning to update and analysis, which assist in determining the true root cause(s) and fault propagation pathway. The implementation and applicability of the proposed methodology are demonstrated using three process systems.