Distributed Missing Values Imputation Schemes for Plant-Wide Industrial Process Using Variational Bayesian Principal Component Analysis

Abstract

Industrial process data often contains missing values due to network transmission errors and sensor failures, etc. Unlike some fields such as biology and climatic science, missing values imputation (MVI) for online data is necessary for industrial processes, because most of the data-based intelligent decision support systems demand a complete training data set and online samples. To the best of the authors' knowledge to date, limited results on MVI for both a training data set and online samples are reported. How to achieve a comprehensive and highprecision MVI scheme in line with industrial reality is still an open problem. Directed toward the complicated correlations among variables of plant-wide industrial process (PWIP), we first propose an improved feature subset selection algorithm based on the time shift correlation and a newly defined selection criterion. Second, for each variable with missing values, the feature subset with strong correlations to this variable is selected. In this way, the time-lagged correlations both within and across variables are made full use of. Through applying variational Bayesian principal component analysis (VBPCA) into the resulting feature subsets, a novel distributed shift correlation-based VBPCA (DSCVBPCA) technique is developed to achieve a better imputation effect. Thereafter, the moving window strategy and the modified DSCVBPCA with tactfully set parameters are integrated to accomplish the MVI for online samples. Finally, the experiments much closer to the actual situations of PWIP are conducted on a numerical example and gold hydrometallurgy, indicating that the proposed imputation scheme can be a promising alternative for the MVI of industrial processes.

Data-Driven Process Monitoring and Fault Diagnosis: A Comprehensive Survey

“…Poor data quality: To address problems related to poor data quality, such as missing or spurious observations, a class of techniques called robust PCA has been developed. Pioneering work by Walczak and Massart [143] and Xie et al [144] laid the foundation for this field, which was further developed with applications and extensions [145][146][147][148][149][150][151].…”