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

Abstract

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