Machine Learning based Surrogate Assisted Multi-Objective Optimization of Continuous Casting Process

Optimization of industrial continuous casting process requires faster models tuned across various operating regimes. Data based modelling techniques like Support Vector Regression (SVR) are proven to be efficient modelling techniques as they are based on structural risk minimization principle. However, the hyper-parameters of SVR are usually tuned on trial-and-error basis without any rationale leading to inappropriate model. To generate an efficient model for the continuous casting process, we propose an algorithm for estimating the hyper-parameters of SVR by considering Root Mean Square Error (RMSE) of the model and sample size required for modelling as the conflicting objectives. Differing importance to various inputs under different conditions leads us to use different kernel parameters for different inputs during model development. Additionally, many kernels are explored to decipher the unknown nature of the continuous casting process. Simulation results show that the proposed algorithm could develop temperature and bulging models, with which the optimization of the casting process has been shown to be effective.