Comparative study of PCA approaches in process monitoring and fault detection

This paper suggests an alternative scaling approach to PCA analysis for monitoring industrial processes. It also compares performance of the proposed moving PCA (MPCA) and three other PCA-based approaches including conventional PCA, adaptive PCA and exponentially weighted PCA, on a well known simulation model of an industrial plant and on data obtained from a petrochemical plant over a period of X months. The result showed that MPCA, which uses the mean and standard deviation of a moving window for scaling purpose, appeared to outperform the other three methods in monitoring processes with/without changes in operating conditions/set-points. While a conventional PCA seemed to work satisfactorily with the Tennessee Eastman Process (TEP) simulation, its performance was much poorer on the industrial data set. This comparison demonstrates that a degree of adaptation in scaling parameters is necessary for PCA-based approaches, especially for processes with multi operating modes.