LAUSR

Abstract The molten steel temperature in ladle furnace is a significant variable, but it is hard to be measured by real-time detection, which has some bad effects on productions. Soft sensors are alternative and effective techniques to solve this issue. In this paper, the soft sensor of the molten steel temperature established by the Modified Maximum Entropy based Pruned Bootstrap Feature Subsets Ensemble (MMEP-BFSE) method is proposed. Although the Bootstrap Feature Subsets Ensemble (BFSE) temperature model is prominent in the precision and the forecasting speed on the large-scale and noisy data, its main drawback is too many sub-models required to combine, which is not always feasible for applications. To alleviate this drawback, the Modified Maximum Entropy based Pruning (MMEP) approach is presented, in which a subset of sub-models that better approximates the complete ensemble is find based on the maximum Renyi entropy and the trade-off parameter between the precision and the diversity of sub-models. Then, the soft sensor of the temperature based on the MMEP-BFSE is established on the practical data. Experiments show that the proposed soft sensor outperforms the others in the precision, and meets the precision requirements. Sub-models of the BFSE temperature model are substantially pruned with improved generalization by the MMEP approach.