Dynamic Decentralized Monitoring for Large-Scale Industrial Processes Using Multiblock Canonical Variate Analysis Based Regression

Decentralized monitoring methods, which divide the process variables into several blocks and perform local for each sub-block, have been gaining increasing attention in large-scale plant-wide due to complexity of their processes. In such dynamic nature data is a relevant issue not usually managed. Here, new data-driven distributed scheme proposed deal with this issue, integrating regression automatically blocks, multivariate statistical analysis (Canonical Variate Analysis, CVA) monitoring, Bayesian inference achieve decision making. By constructing sub-blocks using regression, it possible identify most commonly associated every block. Three methods are proposed: LASSO (Least Absolute Shrinkage Selection Operator), forces coefficients less towards zero; Elastic-net, robust method that compromise between Ridge Lasso regression; and, finally, non-linear based on Multilayer Perceptron Network (MLP). Then, CVA model implemented sub-block consider characteristics industrial processes provides global fault detection. The Tennessee Eastman benchmark validates efficiency feasibility regarding some state-of-the-art methods.