Abstract:Vacuum consumable melting is one of the important methods to produce high-quality special steel. However, volatile elements are easy to burn and cause fluctuations in ingot composition during the melting process, which affects the quality of the final ingot. In order to accurately control the composition fluctuation of Mn before and after smelting, the Pearson correlation coefficient method and mutual information method were used to reduce the dimension of production data, and 12 features were selected as model input. After data standardization and super-parameter optimization, a prediction model for the end point Mn content of vacuum consumable ingot Gradient Boosting Decision Tree(GBDT) was established, and compared with the Decision Tree (DT), Random Forest (RF) and adaptive improvement (Adaboost) prediction model. The study has found that the melting current, feeding current, and vacuum degree among the process parameters have the greatest impact on the end point Mn content. After dimensionality reduction, the prediction error of each model has decreased, and the GBDT model has the largest reduction, with its root mean square error 0.031 46, and average absolute error 0.025 51. The error of GBDT model is the lowest, with the percentage of heats within the error range of ±0.06%, ±0.04%, and ±0.02% being 96%, 78%, and 44%, respectively. The prediction effect within the overall error range is good, which has certain guiding significance for actual production.
BAI Y,LI M,CHENG C,et al. Study on microstructure and in situ tensile deformation behavior of Fe-25Mn-xAl-8Ni-C alloy prepared by vacuum arc melting[J]. Metals,2021,11(5):814.
ZHANG Z,JUNG C. GBDT-MO:Gradient-boosted decision trees for multiple outputs[J]. IEEE Transactions on Neural Networks and Learning Systems,2021,32(7):3156.