Fault detection based on Bayesian network and missing data imputation for building energy systems

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

“…The main techniques used for disposing MV of industrial processes include deletion, robust model − and MVI. − The deletion directly ignores the samples with MV, which is simple but leads to the loss of valuable information on incomplete samples. When the number of samples with MV is large, the data analysis result tends to be unreliable.…”

“…The above-mentioned achievements − have provided excellent MVI performance in the industrial processes; however, they all belong to the centralized models without selected features/variables. For the PWIP characterized by long processes and multiple procedures, the centralized models may not be the most suitable because they may involve the variables that are irrelevant or have little correlations with the variables that need to be imputed.…”

“…Therefore, the MVI for online samples is another important research topic for industrial processes. In terms of the online samples MVI, the foregoing achievements either directly neglect the incomplete samples − ,, or require a sufficient number of complete samples in the training data set. − However, acquiring a sufficient number of complete samples becomes intractable when the MV are part of the inherent structure of the study object . To the best of the authors’ knowledge, during 2006 to 2019, there is almost no study on the MVI that has considered both the training data set and online samples simultaneously when performing the missing data experiments for a specific missing rate .…”

“…The main contributions of this paper are summarized as follows: Directed toward the complicated correlations among the variables of PWIP, an enhanced feature subset selection algorithm is proposed by employing the time shift correlation and defining a new selection criterion, so that the feature subset with strong correlations to each variable with MV is selected. By applying VBPCA to the resulting feature subsets, a better imputation effect is achieved. The existing MVI achievements for industrial processes either neglect the MVI of online samples − ,, or require a sufficient number of complete samples in the training data set. − To the best of the authors’ knowledge to date, the reported literature on MVI for both training data set and online samples is scarce. On account of this, we first propose a novel DSCVBPCA technique for the MVI of the training data set and then apply the MW model and the modified DSCVBPCA to the online samples MVI. The designed experiments on the numerical example and gold hydrometallurgy are much closer to the actual situations of industrial processes.…”

“…The existing MVI achievements for industrial processes either neglect the MVI of online samples − ,, or require a sufficient number of complete samples in the training data set. − To the best of the authors’ knowledge to date, the reported literature on MVI for both training data set and online samples is scarce. On account of this, we first propose a novel DSCVBPCA technique for the MVI of the training data set and then apply the MW model and the modified DSCVBPCA to the online samples MVI.…”