Research On Rolling Force Prediction Modeling Of Tandem Cold Rolling Mill Based On Deep Learning

In the iron and steel cold rolling production process,the thickness of the final product and the flatness of the plate shape were determined by the prediction accuracy of the rolling force.Accurate rolling force prediction can reduce the strip head and tail length effectively.The yield of raw materials will be improved.The rolling environment is complex and changeable.the parameters are coupled with each other.It is difficult to establish an accurate mathematical model.The traditional rolling force mechanism model introduces many assumptions to simplify the derivation process and cannot guarantee the final prediction accuracy.In order to improve the rolling force prediction accuracy of cold rolling mills,a rolling force model is established based on the basic rolling theory and deep learning method.the rolling force prediction modeling is studied.Taking the rolling force mechanism model Bland-Ford-Hill formula as the theoretical background,the main parameter variables that affect the rolling force are analyzed and determined.Some unmeasured data is calculated by the mechanism model.These data are adopted as the input data of the neural network to predict the rolling force.The rolling force prediction model established by the data-driven method,which can avoid the excessive exploration of rolling mechanism and obtain a more realistic model.The main works are as follows:(1)The fitting ability of machine learning methods is limited by its own network structure,which is difficult to obtain high prediction accuracy.Aiming at this problem,a rolling force prediction model based on a semi-supervised deep network was established.Firstly,a high-level feature representation of input data was extracted layer by layer using a stacked denoise autoencoder to solve the problem of difficult training of deep networks.To improve the validity of feature extraction,according to the degree of correlation between the input value and the target value,different proportions were applied to the feature loss function of each dimension to form a proportional loss stack denoise autoencoder.Then,the deep network was initialized using the features extracted by proportional loss stack denoise autoencoder to predict the target value.The simulation results verify the effectiveness of the method.(2)Aiming the problem that the parameters of the fixed model cannot be adjusted with the change of the production environment,which leads to the decrease of prediction accuracy,the just-in-time-learning learning method was adopted by selecting some similar samples and establishing a local model of the current operating point dynamically.However,due to the lack of sufficient process knowledge,it is difficult to select suitable samples to construct an accurate prediction model using only a single similarity measure.An ensemble just-in-time-learning method based on multi-weighted similarity measures was proposed.Firstly,different weighted similarity measure was used to select relevant samples.Then,the local model of the deep network was constructed to estimate the target value of the query data.The batch normalization method was used to stabilize the forward output interval.The small batch gradient descent algorithm and Adam optimization algorithm was introduced to speed up the model convergence.Finally,the ensemble learning strategy was adopted to integrate the output results of each partial model.On this basis,the cumulative similarity factor was introduced to optimize the number of samples in local modeling.The similarity threshold is set to update the local model adaptively.The simulation results show that the model can achieve high-precision prediction of rolling force in a short time.Finally,summarizes the whole paper.