Soft-sensor modeling of silicon content in hot metal based on sparse robust LS-SVR and multi-objective optimization

To solve the problem that the parameter of silicon content ([Si]) in hot mental is difficult to be directly detected and obtained by manual analysis with large time delay, a method of sparse and robust least squares support vector regression (R-S-LS-SVR) was proposed to establish a dynamic model of[Si] with the help of the multi-objective genetic optimization of model parame-ters. First, owing to the issue that the Lagrange multiplier of the standard least squares support vector machine (LS-SVR) is directly proportional to the error term and solves the lack of sparsity, the maximal independent set of sample data in the feature space mapping set was extracted to realize the sparse of the training sample set and reduce the computational complexity of modeling. Next, in view of the problem that the standard least squares support vector machine has no regularization term, a method to improve the modeling ro-bustness was proposed by introducing the IGGⅢ weighting function into the obtained sparse least squares support vector regression (S-LS-SVR) model. Last, the multi-objective evaluation index that synthesizes the modeling residue and the estimated trend was presented to compensate for the deficiency of the single root mean square error (RMSE) index. Based on those, an on-line soft sensor model of hot metal[Si] with the optimal parameters was obtained by using the multi-objective genetic algorithm (NSGA-Ⅱ) with the non-dominated sort and elitist strategy. Industrial verification and analysis show the effectiveness and superiority of the proposed method.