Study On Physical Properties Prediction Model Based On Neural Network For Multi-component Mold Fluxes

Mould fluxes play very important roles in continuous casting of steel. Reasonable properties and compositions are necessary requirements for designing and developing mould fluxes. Up to now, all the designing processes mainly have depended upon abundant experimental experience of researchers. The established mathematical model between compositions and properties of mould fluxes were based on experimental results. The accuracy of the model was usually limited in predicting properties because of nonlinear relationship between properties and compositions and limited samples.Neural network(NN) can approach every nonlinear relationship, extract knowledge, which reflects inherent law and feature, from unobvious background and imperfect data, solve problems by the obtained knowledge. To overcome the disadvantage of the empirical models, BP NN was used to construct a series of innovative models to forecast melting temperature and viscosity of mould fluxes.To establish the BP NN model, CaO-SiO2-Na2O-B2O3-Al2O3-CaF2-Li2O-MnO -MgO slag system were employed in the study. The compositions of mould fluxes were designed according to the principle of blending regression design with bound, which could reduce and avoid blind-spot of slag system and guarantee the extensiveness and uniform distribution of data. The effects of compositions on properties of multi-component slag system were obtained according to experimental melting temperature and viscosity of 233 groups of mould fluxes. Base on the data, BP neural network prediction models were established, which were used to forecast melting temperature and viscosity. Prediction accuracy and efficiency were improved through optimizing unit number of hidden layer and learning rate. With simple construction and good extension ability, the models are suited for mould fluxes with multi-components and wide range of compositions. For the same data, average relative predication errors of the models were 8.25% and 0.36% for viscosity and melting temperature respectively, the errors were much less than the errors of 63.2% and 2.8% made by nonlinear regression formula. Therefore, the NN models can be used to predict the molten properties of mould fluxes and meet the requirement of mould fluxes development and design.