Multi-objective Optimization Based Recursive Feature Elimination for Process Monitoring

Process monitoring helps to estimate the quality of the end products, equipment health parameters, and operational reliability of chemical processes. This is an area in which data-driven approaches are widely used by academic and industrial practitioners. With the ever-increasing complexities in process industries, there is an increased thrust in developing the process monitoring methods of generic nature which are capable of handling the inherent nonlinear characteristics of the chemical process. This demanded the employment of complex data-driven model paradigms in the process monitoring framework. To circumvent the issues related to high-dimensional process data, a large body of these process monitoring algorithms extract only relevant features during the training. Further, model complexity is another important issue that needs to be accounted while employing these methods. In this work, an optimization-based features selection method for process monitoring is proposed, that simultaneously trades-off between the optimal feature selection and the resulting model complexity, by means of solving a multi-objective optimization problem. Particularly, this paper focuses on combining neural network architecture with recursive feature elimination and genetic algorithm to obtain an improved identification accuracy while reducing the number of variables to be measured continuously in the process plant. The efficacy of the proposed approach was validated using a basic numerical case and tested upon the operational data collected from the benchmark Tennessee Eastman plant data, and steel plates manufacturing case studies.