Research On Performance Prediction Of Hot Rolled Products Based On Deep Learning

In recent years,hot-rolled steel has been widely used in many industries and fields such as construction,bridges,pipelines,automobiles,ships,railways,construction machinery,pressure vessels and so on.These key areas are closely related to people's livelihood and public safety,thus requiring hot-rolled products must have good product quality,especially the tensile strength,yield strength and elongation and other mechanical properties.Traditional hot-rolled product quality inspection usually randomly selected a small amount of samples to measure the mechanical properties of the sample steel.However,this method of sampling tests can take a great deal of time and manpower.During the process of smelting and hot rolling,a series of complicated changes occur in the microstructure of the rolled steel,which directly determines the mechanical properties of the steel.Therefore,using chemical composition and hot rolling process parameters to predict mechanical properties has important theoretical significance and application value.Due to hot rolling production is a real-time dynamic process,will be subject to a variety of random factors,and contains a large number of process parameters.Therefore,Irvine's theory of mathematical models based on many approximate assumptions is only suitable for ideal environment and theoretical analysis.In recent years,with the advent and development of artificial neural network,it provides an effective technical tool for forecasting the performance of hot rolled products,and the prediction accuracy has been greatly improved.However,due to the limitation of artificial neural network structure and learning algorithm,the current massive real-time industrial production data cannot be effectively trained and studied and cannot meet the increasing demand of accuracy prediction of users.In this paper,we propose a performance prediction model for hot rolled products based on deep neural networks.The prediction model of mechanical properties from steel composition and hot rolling technology to mechanical properties is realized by using multilayer perceptron machine and convolution neural network.Utilizing more than 50,000 hot-rolled data from two hot-rolling plants and Google's TensorFlow platform to implement the algorithms and models.In this paper,we first prove the feasibility and superiority of deep feedforward neural network in comparison with shallow neural network in the prediction of steel performance through comparative experiments,and get better precision,and obtain a better accuracy.Aiming at the shortcomings that the feedforward neural network cannot mine the relationship between the characteristic parameters,a method of predicting the performance of the steel by convolutional neural network is proposed.Based on the analysis of the two models and the problems of over fitting in the experiment,this paper improves the network model by using various techniques such as Dropout and Batch Normalization.The experimental results verify the effectiveness and reliability of the proposed algorithm,which can effectively improve the accuracy of the performance prediction of hot rolled products.