Modeling Rough Rolling Process Of Hot Continues Rolling Slab Based On Machine Learning

At present,hot continuous rolling is the main strip production method,and rough rolling is the key link in hot strip rolling.In the rough rolling process,the slab rolling process involves many process control parameters,which directly or indirectly affect the product quality and rolling stability of the produced strip steel.Among these process parameters,there are two parameters that play a crucial role in the quality of the strip,namely the rough-rolled outlet temperature of the slab and the rough-rolled slab warp head.The outlet temperature of the rough slab rolling has an important effect on the subsequent finishing rolling,layer cooling and other processes,which directly affects the mechanical properties of the strip,the product thickness index and the rolling stability;the rough-rolled slab warp head is caused by the asymmetric deformation of the head after the slab is rolled by the roughing stand,which seriously affects the production condition and quality of the strip.Therefore,it is extremely important to establish an accurate slab roughing outlet temperature prediction model and warp head prediction model.In this thesis,based on the actual production process data of the steel mills,the prediction models of slab rough rolling outlet temperature based on Random ForestLong Short-Term Memory Neural Network(RF-LSTM)and hot continuous rolling slab warping head based on stacking ensemble learning are established respectively by using machine learning methods.A slab roughing outlet temperature prediction model based on RF-LSTM is proposed,which is aimed at the problem that the data dimension of the hot continuous rolling process is too high and it is difficult to accurately predict the slab temperature,.Firstly,the improved Random Forest(RF)algorithm is used to select the characteristic variables.It measures the contribution of each characteristic variable by analyzing the change of the prediction results of slab rough rolling outlet temperature,and the characteristic variables with larger contributions are selected;Secondly,in view of the time series characteristics of the hot continues rolling production process data,the Long Short-Term Memory(LSTM)neural network is used to predict the outlet temperature of the rough slab rolling.The experimental verification results show that the average absolute error and root mean square error of the billet temperature prediction before and after feature selection are reduced by 0.21? and 0.25?,respectively,and the accuracy of the prediction relative error within ±3.0% reaches 99.07%.A prediction model based on stacking ensemble learning is proposed to predict the degree of head warping in roughing process,which is aimed at the problem of head bending in roughing process of hot continuous rolling.First,preprocess the modeling data,including the processing of outliers and missing values,artificial combination of some features to make up for the lack of features,use of Pearson correlation coefficient and maximum information coefficient combined with manual experience to perform feature screening to improve the quality of modeling data.Secondly,stacking ensemble learning method is used to integrate five machine learning algorithms including Adaptive Boosting(Ada Boost),Random Forest(RF),Gradient Boosting Decision Tree(GBDT),e Xtreme Gradient Boosting(XGBoost),and Support Vector machine Regression(SVR)to establish a higher-level prediction model for the rough-rolled slab warp head.Moreover,the specific idea of model parameter optimization is given.Through experimental verification and comparative analysis,compared with the prediction model of the rough-rolled slab warp head builded by single machine learning algorithm,the prediction accuracy of stacking ensemble learning model which absorbed the advantages of each algorithm has been significantly improved.Comparing the GBDT model with the highest accuracy among the warped head prediction models constructed by a single algorithm,the average absolute error and root mean square error of Stacking integration model are reduced by 0.0358 cm and 0.0762 cm respectively,and the proportion of samples with absolute error within ± 2cm in the test set increased by 1.81% to 88.50%.The effectiveness of the proposed prediction model based on stacking ensemble learning is fully proved.