A Novel Robust Broad Nonlinear Representation CVA Method for Monitoring Blast Furnace Iron-making Process

Abstract

Blast furnace iron-making process is one of the most crucial parts in iron and steel industry. Due to many interference factors and a series of complex physical and chemical reactions, abnormal furnace conditions often occur. The nonlinear characteristics hidden in the blast furnace data and the existence of large noise and outliers make it difficult to establish an effective monitoring model. In this paper, a novel robust broad nonlinear representation canonical variate analysis (RBNCVA) method is proposed to overcome the above problems. First, a feature extraction strategy is developed to extract robust broad nonlinear features based on stacked denoising autoencoder (SDAE). The robust broad nonlinear features can assist the model to cope with the complex nonlinearity and resist the interference of noise and outliers. Then canonical variate analysis (CVA) method is used to analyze the relationship between past and future feature vectors. Subsequently, control limits are computed through probability density functions defined by kernel density estimation. Finally, the practical blast furnace data is adopted to validate the effectiveness and robustness of the proposed method.