Robust monitoring of industrial processes using process data with outliers and missing values

The quality of process data has great effect on the performance of data-driven process monitoring methods. Two main problems that reduce the quality of data are the contamination (i.e., outliers) and missing values. To address these problems, a two-step method called COF-GRE is developed for the robust estimation of multivariate location and scatter matrix in the presence of missing data and both casewise and cellwise outliers. The first step applies cellwise outlier filters (COF) to identify and remove the contaminated cells in a multivariate data matrix. In the second step the generalized Rock S-estimators (GRE) are used to handle casewise outliers and missing values simultaneously. Based on the COF-GRE method, a robust process monitoring method is then developed to achieve high-performance monitoring even if the reference data of industrial processes contain outliers and missing values. A fault detection index, called the robust T-2 statistic, is defined using the multivariate location and scatter matrix obtained by COF-GRE. A fault diagnosis method is also proposed to determine faulty variables and at the same time to compute their fault magnitudes. The effectiveness and advantages of the proposed methods are illustrated with a simulation example and an industrial case study.