Transfer Learning Deep EDC-Tabnet for Predicting End-Point of BOF Steelmaking Process with Small Samples

The basic oxygen furnace (BOF) steelmaking processes with complex multiphase reactions can be modeled by data-driven technique, but it is particularly difficult to predict the end-point quality indicators of steel grades with small sampled data. A tree structure deep network with transfer learning is the first attempt to be designed based on Tabnet by fusing Encoding and Decoding blocks (EDC-Tabnet). The obtained steel grades containing large sample data and small sample data are taken as the source domain and the target domain, respectively. The Encoding block utilizes the decision separation mechanism of each step to achieve feature selection and extraction from the input data. The features extracted by the Encoding block are passed through a cascade structure to the Decoding block for deeper biased feature extraction. In the transfer learning process, the improved Tabnet network is firstly trained through source domain data to obtain a pre-trained model. An output adaptive layer is added following the EDC-Tabnet to match the target domain for improving the prediction accuracy. Finally, based on an actual BOF production process, the effectiveness of the proposed EDC-Tabnet with transfer learning for predicting the end-point of Q256 steel with small samples is verified by comparing with the existing methods.