Prediction Of Mechanical Properties Of Hot Rolled Steel Based On RBF Neural Network

Hot rolled strip steel is heated by high temperature rolled steel,it has good weldability and plasticity.It is widely used in household such as stairs,slide proof ladders,ladder treads,and other industrial fields,such as automobile wheel steel strip,hot rolled steel plate,mechanical structure of important parts of crane.Mechanical properties are an important part of reflecting the quality of steel,tensile strength is the major factor that determines mechanical properties,it refers to the critical value for the transition of metal from uniform plastic deformation to local concentrated plastic deformation,it is also the maximum load-bearing capacity of metal under static tension conditions,that is the fracture resistance of steel.In the hot rolling process of steel,the production process and composition elements are complicated,the prediction accuracy obtained by the traditional statistical modeling method is not perfect,so how to establish a reasonable prediction model to predict the tensile strength of steel based on the existing production data is of great significance to improve the quality of products,reduce production costs and better meet the market demand.RBF neural network has strong nonlinear fitting ability,can map arbitrary complex nonlinear relations,and the learning rules are simple,it has a strong robustness.A prediction model of steel mechanical properties based on RBF neural network is studied in this paper.First,we respectively propose a generalized RBF neural network model based on composite quantile regression(Generalized radial basis function neural network composite quantile regression,CQR-RBFNN)and a generalized RBF neural network model based on composite expectile regression(Generalized radial basis function neural network composite expectile regression,CER-GRBFNN),these two models combine the non-linear fitting ability of RBF neural network and the interpretability of quantile regression and expectile regression for data.Secondly,in order to solve the parameters of the above two models,we combine the classical gradient descent algorithm and cuckoo algorithm,and the specific steps are as follows.(1)Aiming at the disadvantage that the center of the generalized radial basis function neural network model is not easy to select,we propose an improved canopy +kmeans unsupervised algorithm to select the center.(2)For the gradient descent algorithm easy to fall into the local optimal situation,we incorporate the cuckoo search algorithm into the gradient descent algorithm,that the cuckoo algorithm is used to determine the initial value of the parameters,and then the gradient descent algorithm is used to solve the final parameters,and the iterative cycle is carried out until convergence.Finally,we collected 3,534 pieces of data from the hot-rolled steel production line of a large domestic steel company,we use the tensorflow framework in Python 3.5 after cleaning and standardizing the data,and respectively use CQR-GRBFNN and CER-GRBFNN to modeling,the results show that our method is effective,it has high prediction accuracy and good generalization ability.