基于深度置信网络算法的面向铁磁材料旋转磁滞损耗的矢量磁滞模型  

Vector Hysteresis Model for Rotational Hysteresis Loss of Ferromagnetic Materials Based on Deep Belief Network Algorithm

作　　者：马阳阳 李永建[1] 孙鹤 杨明[1] 窦润田 Ma Yangyang;Li Yongjian;Sun He;Yang Ming;Dou Runtian(State Key Laboratory of Reliability and Intelligence of Electrical Equipment Hebei University of Technology,Tianjin 300130,China;State Gird Cangzhou Electric Power Supply Company,Cangzhou 061000,China)

机构地区：[1]省部共建电工装备可靠性与智能化国家重点实验室(河北工业大学),天津300130 [2]国网河北省电力有限公司沧州供电分公司,沧州061000

出　　处：《电工技术学报》2023年第15期4063-4075,共13页Transactions of China Electrotechnical Society

基　　金：国家自然科学基金重点项目(52130710);国家自然科学基金项目(51777055,51977122);河北省自然科学基金创新群体项目(E2020202142)资助。

摘　　要：铁磁材料磁滞建模是电气工程领域的基础性理论研究之一。该文基于深度置信网络(DBN)算法结合磁滞算子空间理论提出一种矢量磁滞模型。在模型结构中,引入郎之万函数作为映射函数对磁滞数据进行输入转换计算。利用多个磁滞算子构建算子空间生成高维算子数据,算子空间的数据输出作为DBN模型的输入,结合DBN算法表征算子数据与模型输出的非线性关系。利用样本的磁感应强度数据和生成的算子数据训练模型,获得模型参数。通过仿真表明构建的模型可以有效地描述铁磁材料在旋转磁化情况下的非线性特性和各项异性。同时,结合磁损分离理论改进磁损模型中相应的损耗系数,构建动态磁损计算模型,并将磁滞模型获得的数据应用于动态损耗计算。仿真表明,构建的磁滞模型可以有效地表征铁磁材料的实际磁化特性和损耗情况。The silicon steel sheet is the core material of electrical equipment,and its magnetization characteristics directly affect the operation mechanism of equipment.So,the hysteresis modeling of ferromagnetic materials is one of the basic theoretical studies in the field of electrical engineering.In this paper,a vector hysteresis model is proposed based on the deep belief network(DBN)algorithm and hysteresis operator space theory.The structure of the model consists of three parts:input mapping function,operator space and DBN model.In this paper,the Langevin function is used as the input mapping function to calculate the input mapping of hysteresis data,so that the data can adapt to the characteristics of hysteresis operator in the subsequent structure and can reflect the saturation characteristics of hysteresis phenomenon.Hysteresis operators in multiple directions in H space construct a hysteresis operator space.And the magnetization trajectory of the material mapped by Langevin function is projected in all directions on H space.The high-dimensional hysteresis operator data is generated by calculating hysteresis operators in all directions.Then the output of the operator space is taken as the input of the DBN model.In the construction of vector hysteresis model,DBN model is mainly used to characterize the nonlinear relationship between the high-dimensional vector data output by the operator and the magnetic induction data of the material.The parameters of the vector hysteresis model are obtained by training the magnetic induction data of training samples and the operator data generated by the training samples.The model parameters are mainly obtained by training DBN parameters.And the training process of DBN mainly consists of two parts:(1)The CD algorithm is used to the pre-training of the RBM in each layer,then the RBMs are stacked to obtain the preliminary optimization parameters of the model.(2)The parameters obtained by pre-training are taken as initial values,and the Nadam optimizer is used for global parameter

关 键 词：磁滞模型 深度置信网络算法 磁滞算子 磁滞损耗 

分 类 号：TM15[电气工程—电工理论与新技术]