Strip Crown Prediction in Hot Rolling Process Using Random Forest

This paper established a model based on an ensemble method to predict strip crown using 230,000 coils of data obtained from a hot rolling line. Before modeling, a specific method was proposed to reduce noise and remove outliers, and a new dataset of 5657 samples was generated. Parameter tuning considering mean squared error (MSE) was carried out to establish three machine learning models including support vector machine (SVC), regression tree (RT), and random forest (RF). Determination coefficient (R2), mean absolute error (MAE) and root mean squared error (RMSE) were used as indicators to evaluate the prediction of models. Results showed that the RF had the best performance with the highest R2 of 0.707, as well as the lowest RMSE of 5.66 μm. Moreover, an additional method that repeated the three models 100 times was developed, and box plots were used to visualize the distribution of R2, MAE and RMSE. RF can correct for decision trees to reduce overfitting to their training set, improving the generalization, and in this paper, the trained RF which had stable performance is considered as the most recommended model. After that, for RF, rankings of rolling process variable were validated to make a comparison with the existing physical understanding.