基于深度学习的电工钢片磁畴磁化过程预测与特征量提取  

Prediction of Magnetic Domains Magnetization Process and Feature Extraction of Electrical Steel Sheet Based on Deep Learning

作　　者：吴鑫 张艳丽[1] 王振 李梦星 姜伟[1] Wu Xin Zhang;Yanli Wang Zhen;Li Mengxing;Jiang Wei(Ministry of Education Key Laboratory of Special Motors and High Voltage Electrical Apparatus Shenyang University of Technology,Shenyang,110870,China)

机构地区：[1]沈阳工业大学教育部特种电机与高压电器重点实验室,沈阳110870

出　　处：《电工技术学报》2023年第9期2289-2298,共10页Transactions of China Electrotechnical Society

基　　金：国家自然科学基金资助项目(52277015)。

摘　　要：高效电工产品的研发离不开对低损耗高品质电工钢磁特性的准确表征,而电工钢磁畴结构及其动态磁化特征量的提取是实现从介观磁化机理到宏观特性模拟的桥梁。该文利用磁光克尔显微镜对外加磁场下取向电工钢片磁畴结构的动态演变过程进行了观测和研究。首先在不同方向的外加磁场下,获取了磁场由饱和到零再到反向饱和的磁化过程中电工钢磁畴结构动态演变的图像信息,分析了磁畴重组、湮灭、成核的过程;其次为了获取足够多的可以表征电工钢片磁化特征的磁畴图像信息,对比研究了基于深度学习理论的两种神经网络——基于卷积神经单元的长短期记忆神经网络(ConvLSTM)和基于卷积神经单元的门控循环单元(ConvGRU)对磁畴动态演变图像的预测;最后在此基础上,提出了以磁畴面积作为特征参量表征样片动态磁化过程及磁畴状态的方法。结果表明,所提取的特征量能够有效地表征样片的磁化程度及其内部磁畴的演变过程,为进一步实现基于磁畴磁化机理的宏观磁滞模型奠定基础。The development of high-efficiency electrical products is inseparable from the accurate characterization of magnetic properties of high-quality electrical steel with low loss.The extraction of magnetic domains structure describing the dynamic magnetization characteristic in an electrical steel is a bridge to realize the simulation from mesoscopic magnetization mechanism to macroscopic characteristics.Current research on the magnetic properties of electrical steel tends to focus on the characterization of macroscopic magnetic properties,while the magnetic properties of materials often depend on the internal domains structure of materials,and the study of reflecting the observed dynamic processes of magnetic domains into the model of macroscopic magnetic properties is still in its initial stage.In order to further investigate the mesoscopic hysteresis characteristics model,a magneto-optical Kerr microscope was used to observe and study the dynamic evolution process of magnetic domains structure in a grain-oriented electrical steel sheet under an external magnetic field.Firstly,the image information of dynamic evolution of magnetic domains structure in an electrical steel was obtained under the external magnetic field changing from saturation to zero to reverse saturation,and the process of magnetic domains reorganization,annihilation and nucleation was analyzed.Secondly,in order to obtain image information of magnetic domains enough to characterize the magnetization properties of electrical steel sheet,two kinds of neural networks based on deep learning theory such as ConvLSTM(convolutional long-short term memory)and ConvGRU(convolutional gate recurrent unit),were compared and studied on the prediction effects of dynamic magnetic domains evolution images.Finally,based on the experimentally observed and model-predicted magnetic domains images of electrical steel,a method was proposed to characterize the dynamic magnetization process and magnetic domains state of the sample by extracting the area of the magnetic doma

关 键 词：电工钢片 磁畴 深度学习 磁光克尔显微镜 

分 类 号：TM275[一般工业技术—材料科学与工程]