Research On Cold Rolling Force Modeling Method Based On Machine Learning

In the process of cold rolling,the prediction accuracy of rolling force indirectly affects the thickness and flatness of the finished products,and determines the stability of the rolling process and the quality of the final product.Accurate prediction of rolling force can effectively reduce the length of strip head and tail.The utilization rate of raw materials will be improved.However,the actual rolling process is complex and changeable,with characteristics of multi-variable,nonlinear and strong coupling.If the traditional mechanism model with too many assumptions is used to predict the rolling force,it will lead to a large error,reduce the yield,and it can not meet the requirements of high-precision production.In order to solve the inherent defects of the traditional rolling force mechanism model and obtain a more accurate rolling force prediction model,the modeling method is deeply studied,details are as follows.(1)The shallow model can not extract the deep-seated features hidden in the data,and its expression ability is poor,so it can not effectively complete the fitting of complex functions;the model also has the problem of drift,which leads to the low prediction accuracy.In order to meet the prediction requirements of rolling force on large samples and complex problems,a rolling force model based on cyclic self-coding network is established.Based on auto-encoder,this model combines the advantages of cyclic gating unit(Gate Recurrent Unit,GRU)network in sequence recursion,memory,parameter sharing,etc.,and forms a Recurrent auto-encoding network to extract features from input data,at the same time use the mini-batch gradient descent algorithm and Adam optimization algorithm to improve network performance.Finally,Gauss process regression is used to fit the extracted features to get the predicted output.The simulation results show that the model has high prediction accuracy and can meet the requirements of actual production.The simulation results exhibit that the prediction accuracy of the model can be up to 3%,and the rolling force can be predicted with high precision.(2)In the above research,the uncertainty caused by the environmental change of rolling production site is not considered,and the model may have the problem of drift when the production data fluctuates greatly.In order to further improve the prediction accuracy of rolling force,update the model online and avoid the problem of concept drift,a rolling force model based on LSTM-JITRVM(Long Short Term Memory-Just In Time Relevance Vector Machine)is proposed.Firstly,the cyclic auto-encoding network is used to extract the features of the input data.Then,local outlier factor algorithm is used to judge whether the test samples and their neighborhood points belong to the same distribution.In the case of the same distribution,the LSTM model is used for prediction.In the case of the different distribution,Just-in-time learning framework is used for online modeling.In the Just-in-time learning framework,the samples with high similarity were selected from the sample database to form local data sets,and the JITRVM model was established online for prediction.The simulation results show that the model can realize the high-precision prediction of rolling force.