Research On Flatness Of Cold Rolling Strip Predictive Algorithm Based On Data Driven

The delay of flatness measurement in cold rolling brings out difficulty for the flatness control system to generate control signals in time,which leads to the descend of strip quality.Therefore,the prediction of strip flatness has become a hot research topic.Since the production process of cold rolling has accumulated a large number of cold rolling data,mining the potential rolling law from cold rolling data for flatness prediction by data mining technology is significant to improve the precision of flatness control.In this thesis,the researches of flatness recognition and prediction are described in detail,the production process of cold strip mill is systematically analyzed and the methods of flatness recognition and predictive modeling are studied.In flatness recognition,considered dimensional variation of flatness data caused by different strip specifications and the variable structure of recognition model,the Euclidean distance between the measured flatness value and the basic flatness pattern was taken as the input of the recognition model.A flatness recognition model based on multi-layer perceptron(MLP)and a flatness recognition model based on long short term memory(LSTM)were proposed.In flatness prediction,a flatness prediction model based on deep belief network(DBN)is proposed.The main work of this thesis is as follows.The flatness recognition model based on MLP was proposed to solve the problem caused by the limited nonlinear fitting accuracy of shallow neural network in traditional methods.To overcome the problem of insufficient precision of nonlinear fitting,the method of deepening network layers was adopted,which provides a new idea for the accurate flatness recognition.In order to make full use of the dependences among Euclidean distances to further improve the accuracy of flatness recognition,a recognition model based on LSTM was proposed.Considering 8 Euclidean distances step-by-step,the higher order features and dependencies between Euclidean distances were extracted,which reduces unnecessary nonlinear fitting and enhances the anti-interference ability of recognition model.Simulation results show that the proposed method can recognize the strip flatness accurately.To solve the detection delay in the low-speed rolling process,a predictive model of strip flatness based on DBN was proposed.Considering the rolling process with numerous variables,a feature selection method based on mutual information was proposed,which can obtain the feature subset with high correlation.The low accuracy problem was caused by high dimension input with redundancy feature variable,which can be solved by contrast divergence algorithm.On the premise of retaining effective information,DBN decreased input dimensions by the unsupervised training process.Meanwhile,supervised training improved prediction accuracy.Simulation results show that the proposed method is more accurate than other data-driven models.