Just-in-time updating soft sensor model of endpoint carbon content and temperature in BOF steelmaking based on deep residual supervised autoencoder

Abstract

The key of Basic oxygen furnace (BOF) endpoint control is to achieve accurate forecast of the endpoint carbon content and temperature. Due to the highly nonlinear and complex distribution characteristics of process data, the predictions of these two indicators are inaccurate. Although deep learning can extract abstract features of complex data, existing methods are pre-trained in an unsupervised manner. For soft sensor applications, the key is to highlight important information and extract features related to the target variables for prediction. Thus, this paper proposes a feature extraction model based on deep residual supervised autoencoder (DRSupAE). By incorporating important information in the feature of previous hidden layers into deeper feature learning in the form of residual connection, and each residual supervised autoencoder (RSupAE) utilizes the target variables to guide feature extraction. In this way, the stacking of new deep networks not only enhances the impact of important target-related information in layer-wise training, but also captures features related to target variables. Moreover, for the time-varying phenomenon of the BOF steelmaking process, a strategy of just-in-time updating regression network (JITRN) is proposed, which is used to solve the problem that the static regression model cannot adapt to the frequent changes of sample characteristics and causes the prediction performance to decline. An experimental study on BOF steelmaking process data is provided to demonstrate the effectiveness of the proposed method.