Study On Magnetostriction Model Based On Improved Neural Network Algorithm Under Alternating Magnetization

The power transformer core and the motor stator core are stacked by the oriented silicon steel sheet and the non-oriented silicon steel sheet,respectively,and the magnetostriction in these two kinds of silicon steel will cause vibration and noise of the core,which may reducethe service life of the electrical devices and affect the human health.Therefore,the accurate measurement and simulation of the magnetostrictive properties in the silicon steel sheetsare essential work to study the vibration and noise of transformer and motor cores.Most of the existing models are based on the complex polynomial fitting method,and the shortcomings of them are complexformula derivation and many fitting parameters,in this paper,the magnetostriction of the oriented and non-oriented silicon steel sheets is measured under alternating magnetic field.Based on the magnetostrictive measurement data of the grain-oriented silicon steel sheet,the magnetostriction model based on the improved neural network algorithm is established and the accuracy of the model is verified by the experiment of one single-phase transformer core.The main research contents are summarized as:First of all,under the alternating magnetic field,the magnetostrictive characteristics of the oriented and non-oriented silicon steel sheets are measured by using the magnetostriction measurement system based on a triaxial strain gauge.Then the magnitude and direction of the magnetostrictive principal strain are calculated.Under the alternating magnetic field,similarities and differences of the magnetostrictionfor these two kinds of silicon steel sheets are compared and analyzed,and the anisotropy of the magnetostriction is verified.Secondly,based on magnetostrictive data of the grain-oriented silicon steel sheet under the alternating magnetic field,the the magnetostriction model based on traditional Back Propagation(BP)neural network model trained by the gradient descent method is established,in which the inputs are x-and y-components of the magnetic flux density waveforms while the outputsarethe magnitude and direction of magnetostrictive principle strain waveforms.The shortcomings of the traditional BP algorithm based on gradient descent method in the application of the magnetostriction model are discussed.The defect is that the computation speed is too slow and it is easy to fall into the local optimum.Then,the traditional BP algorithm based on gradient descent method is improved.The particle swarm optimization(PSO)algorithm is used to optimize the initial weights of BP neural network in order to avoid the problem of falling into local optimum.Meanwhile,the Levenberg Marquardt(LM)algorithm is used to train the network instead of the gradient descent method,and the computation speed of the network is improved greatly.After that,the improved BP neural network is used to model the magnetostriction and the range of precision of the model is given by 15 samples without being in the training set.Finally,the experimental model of a single-phase transformer core is designed and manufactured and the magnetostriction of three different positions on the core is measured.The calculated values are compared with the measured ones to verify the accuracy of the magnetostriction model under alternating magnetic field.This research will provide a reference data for future research.