Faults prediction and monitoring of complex processes using an ensemble of machine learning regression models: application to the Tennessee Eastman Process

Abstract

Modern industrial installations are a source of big amount of data, these data serve as a means of monitoring and control and can also be used for the prediction of the parameters characterizing the supervised process and thus the anomalies. For this, there are several machine learning regression models that can be considered in order to select the best prediction tool. Our study consists in selecting the best means of prediction of the different operating parameters of the Tennessee Eastman (TEP) process. Indeed, this consists in selecting the best model that can ensure an efficient and cost-effective prognosis and predictive monitoring system. Four regression models were considered during our comparative study: the Support Vector Regression (SVR), Gaussian Process Regression (GPR), Decision Tree Regression (DTR) and Ensemble of Learners approaches for Regression (ELR). As evaluation criteria, we chose the MSE regression error by re-substitution, the MSE regression loss for the cross-validation kernel, the optimization of the hyper-parameters and the training time.