Data-driven classification prediction of hot strip flatness
The flatness is an important quality indicator of the strip shape in hot strip rolling.In order to further improve strip shape control level,and in view of the problems of hot-rolled strip flatness on detection lag and low prediction accuracy,a data-driven classification model of the hot-rolled strip flatness is developed.The 66 key parameters that have a significant impact on the strip flatness among the process parameters,equipment parameters and strip parameters are used as input variables.According to the actual measured spacing of the flatness meter in the rolling direction and the measurement channel width along the strip width direction,the strip coils are dispersed into a series of strip units,and generating flatness classification labels corresponding to different strip units,the flatness classification labels of the middle strip units along the strip width direction are actually used as output variables in this study.The method of data preprocessing is proposed to obtain reliable and high-quality modeling data for building the flatness classification model.DBN algorithm is used to construct the data-driven classification model of the hot-rolled strip flatness.Using cross entropy as the evaluation index,the influence of hyperparameters on the perfor-mance of DBN model is analyzed to determine the best one.Although the sample categories in testing set are unbal-anced,the accuracy rates of the three flatness categories,including the large-medium wave,the no-medium wave and micro-medium wave,are 100％,96.87％and 77.78％respectively,the wrong classification categories of flat-ness are mainly adjacent flatness categories,and their actual flatness values are very close.This method can accu-rately classify the finished hot-rolled strip flatness,and plays an important role in improving the calculation accuracy of the shape preset model and improving strip shape.
hot rollingstripstrip shape predictiondata-driven modelingdata preprocessingdeep belief networks
NETL
NSTL
万方数据
