Modeling Research On Rolling Process Of Cold Rolling Mill Based On Machine Learning

During the rolling process,the initial setting of the rolling force determines the setting of the rolling schedule.And affect the quality and output of the product,the accurate prediction of rolling force can reduce the length of the strip head and tail.The strip rolling process has the characteristics of multivariable,nonlinear and strong coupling.If the traditional mechanism model is used to derive and predict the rolling force,it is not only suitable for narrow surface,but also has large error value,which cannot meet the requirements of today's on-site multi-standard products.Flexible manufacturing requirements.In order to improve the calculation accuracy of the rolling force of the cold rolling mill,based on the basic theory of cold rolling,the rolling force is statistically modeled based on the machine learning algorithm.In order to improve the prediction accuracy of rolling force,three prediction models are established according to the network depth and learning mode.A shallow machine learning model is established for the defects of the traditional mechanism model and the difficulty of selecting parameters.The rolling force is modeled by the support vector machine model based on structural risk minimization.Considering the problem that the SVM parameters are difficult to select,the improved genetic algorithm is used to optimize the model parameters of the support vector machine.Elite strategy and adaptive genetic operator are added to the improved genetic algorithm to increase the network's convergence ability and local search ability.For shallow networks such as support vector machines,the expression ability is limited by the depth of the network and cannot effectively complete the fitting of complex functions.And in the rolling production process,a large amount of rolling force related data can be extracted.Therefore,in order to satisfy the prediction under the rolling force big data set,a deep network model is established to predict the rolling force.Considering the existence of gradient dispersion in deep networks,it is difficult to train the model.This paper will adopt a semi-supervised method to train the deep network model.In order to extract more effective model features,this paper uses the denoising self-encoder to complete the unsupervised training of the model.In order to get the deep network out of unsupervised learning and speed up the network training,in order to meet the requirements of rolling force for rapid modeling,and further improve the accuracy of the model.A new deep network structure is established,which uses the small batch gradient descent as the basic learning method of the network to improve the problems of random gradient descent and batch gradient descent.The Batch Normalization structure is used to optimize the forward propagation of the network and stabilize the input of each layer.Distribution;use Adam random optimization algorithm to optimize network back propagation,provide adaptive learning rate for different parameters in gradient update,and improve gradient update.In order to further solve the gradient dispersion problem of the network,Relu is selected as the activation function of the network.At the same time,according to the data collected on site,using PYTHON language for simulation experiments,the obtained rolling force model is based on conventional shallow model and semi-supervised model in terms of accuracy and modeling speed.