Research On Parameter Forward Modelling Of Permanent Magnetic Field Based On Neural Network

With the rapid development of national technology and industry,the quality of metallic material are required more and more highly by heavy industries such as iron-making&steel-making,nuclear industry,ship-making,energy transmission,etc.Under high temperature and pressure environment,tremendous accidents,such as rupture,explosion,would be caused by metal defects.Thus,the lossless detection technologies for ferromagnetic material have been hot and important research field.Therefore,it is quite urgent and necessary to develop a method which could not only guarantee the accuracy,but also reduce the computations' demand.Based on the BP(back propagation)neural network's attributes of nonlinear mapping,self-study,self-adaptive and parallel computing,the BP neural network method is adopted in this study instead of Finite Element Method(FEM)to explore the solution of forward question.Furthermore,taking the advantage of Genetic Algorithm(GA)and Support Vector Machine(SVM),the shortcomings of BP neural network have been overcame.As a result,the parameter forward modelling of permanent magnetic field based on neural network is accomplished.The main work of this study are summarized as follows:Firstly,in this paper,the principle of Maxwell equations and several forward methods have been briefly introduced.Besides,the applicability of the solution of the magnetic forward question has been analyzed.Therefore,the theoretical basis for the design of the parameter forward modelling could be provided.Furthermore,the principle and the implementation steps of FEM has been outlined as well.Afterwards,the applicability and adaptability of FEM numerical method have been examined by modeling and analyzing the two-demotion and three-demotion of leakage magnetic field with faults.On this basic of this,the experimental data for parameter forward modelling of permanent magnetic field based on neural network could be obtained.Secondly,in order to reduce the computation complexity and improve the solution speed of FEM numerical method.A new method has been proposed in this paper,which is based on BPNN.The optimum parameters of network structure have been determined by experimental analysis,and the net could be off-line trained and stored by utilizing magnetic strength characterization data.Consequently,on the premise of precision and efficiency of solution,this new method could improve the solution speed obviously,and acquire better effect.Thirdly,the globally-searching genetic algorithm has been released in this study to overcome the shortcomings of BPNN,which are easy to fall into local minimum point.This new algorithm has been proved by experiment that it could improve the speed of network training and success rate.Furthermore,this algorithm could improve the prediction accuracy by utilizing special GA-BPNN separated by SVM.Last but not the least,in this paper,different kinds of interpolation methods have been researched.The inference and calculation method with single-peak which is based on cubic spline interpolation,and the magnetic intensity magnetic intensity inference and calculation method with double-peaks which is based on cubic Hermite have been developed in this paper.In addition to this,the flow of BPNN forward method has been systemized,and the algorism has been examined by simulation data.As a conclusion,the final result could meet the requirement of high computation speed and high accuracy,and this new forward algorism has certain feasibility,which could be used in future research.