Data-Driven Strip Crown Prediction And Fault Classification For Hot Rolled Strips

In this paper,a 2160 mm hot rolling production line was studied,and based on the collected production data,as well as the theory of strip crown control,a precise strip crown prediction model and a fault classification model were proposed using artificial intelligence technique.The main research work is as follows:(1)Combined with the actual technological background of the production line and principle of data transmission,variables that can influence the strip crown were selected by analyzing the control mode of strip crown.Based on the real production data,a processing scheme including data collecting,outlier removing and noise reducing,as well as the normalization for input features was proposed,which could help establish the strip crown prediction models.(2)Prediction models for hot rolled strips were established based on deep neural networks,MSE and R were used as indicators to study the influence of key hyperparameters on the prediction performance.MSE,R,MAE,MAPE and RMSE were used as the evaluation indexes for comparisions of the models,the optimal prediction model for strip crown was established based on deep learning,which RMSE was 2.06 ?m,MAE was 1.39 ?m,MAPE was 2.55,and 97.04%of samples had absolute errors less than 5 ?m.By fully considering the mechanism of hot rolling,influence rules of width,roll shifting,rolling force,bending force on strip crown were extracted to prove that the prediction model had compatibility with existing physical understanding.(3)Based on a deep belief network,a model used to classify the strip crown was developed.In the hot rolling process,the real production data has characteristics of multiple and imbalanced classes,SMOTE was used to tackle miniorities,and an improved active learning framework with a new developed selection strategy was proposed to improve the classification performance,and five benchmark datasets were used to validate the effectiveness of the framework.For the actual hot rolling data,experiments about the influence of the number of selected samples and radial basis function on the classification performance were studied,and G-mean and MAUC were used as the evaluation indexes,stable results were obtained by five times 5-flod cross validation.Compared with other typical classification algorithms,deep belief network with active learning and SMOTE method(AL-SMOTE-DBN)had the best performance,with F1-Macro of 0.7140,F1-Micro of 0.8179,g-mean of 0.8109 and standard deviation of 0.0507,MAUC of 0.9522 and standard deviation of 0.0241.Through comprehensive analysis of evaluation indexes,AL-SMOTE-DBN has good classification performance and stability.