Method for predicting residual field of ultra-thin silicon steel sheet under complex excitation based on neural network
Patent Information
- Application Number
- CN202611233869.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-14
- Publication Date
- 2026-10-30
AI Technical Summary
然而，这类方法在实际应用中存在显著缺陷
[0016]本发明技术效果：本发明公开了基于神经网络预测复杂激励下超薄硅钢片剩余场的方法，通过将磁滞特性、热力学不可逆及温度特性三大物理约束以损失函数形式嵌入神经网络训练，有效克服了纯数据驱动模型缺乏物理意义和传统解析式模型精度不足的缺陷。该方法使得神经网络模型在预测复杂激励下的剩余场时，其输出结果严格遵循材料固有的磁化规律与热力学定律，显著提升了预测结果的物理可信度与模型的可解释性。同时，该模型具备优异的泛化能力，能够实现对训练数据外任意复杂激励波形下剩余场的准确、可靠预测，为高频变压器等设备的铁心材料选型与损耗精确评估提供了有效的技术工具。
Smart Images
Figure CN123017294A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of hysteresis model research technology for ferromagnetic materials, and particularly relates to a method for predicting the residual field of ultrathin silicon steel sheets under complex excitation based on neural networks. Background Technology
[0002] With the increasing complexity of operating frequencies and excitation waveforms in power electronic equipment such as high-frequency transformers, ultra-thin silicon steel sheets have gained widespread application due to their low loss characteristics in order to reduce core losses. However, accurately predicting the core losses of silicon steel sheets, especially under complex non-sinusoidal excitation, remains a technical challenge. Based on the consistency between the classic Bertotteti loss separation theory and magnetic field separation technology, the total core loss can be decomposed into three parts: hysteresis loss, eddy current loss, and residual loss. Currently, relatively mature methods exist for predicting hysteresis loss and eddy current loss with relative accuracy. However, under actual high-frequency complex operating conditions, due to excitation waveform distortion, the hysteresis loop of the material will exhibit a complex shape containing multiple small loops, making the accurate prediction of residual loss a key issue restricting the accuracy of total loss calculation.
[0003] Currently, methods for predicting the residual field mainly rely on mathematical models based on specific analytical expressions. However, these methods have significant drawbacks in practical applications. First, analytical models typically contain several statistical parameters that are difficult to identify and calibrate accurately through experiments, which poses a significant challenge to model building and practical application. Second, and more importantly, when faced with arbitrarily complex excitation scenarios containing higher harmonics, the overall prediction accuracy of existing analytical models is not high, failing to meet the engineering requirements for high-precision design and optimization.
[0004] In recent years, although data-driven methods such as neural networks have been attempted for predicting material properties, in predicting residual fields, simple mathematical modeling without clear physical guidance can turn a model into a "black box." Such models not only lack clear physical meaning in their predictions, making them unreliable for engineers, but also exhibit weak generalization ability, with a significant drop in reliability under excitation conditions not covered by the training data. Therefore, in accurately predicting the residual field of ultrathin silicon steel sheets under complex excitations, constructing a model that possesses high accuracy, physical interpretability, and strong generalization ability has become a pressing challenge for those skilled in the art. Summary of the Invention
[0005] To address the aforementioned technical problems, this invention proposes a method for predicting the residual field of ultrathin silicon steel sheets under complex excitation based on neural networks. This method can accurately predict the residual field of ultrathin silicon steel sheets under arbitrary complex excitation with small prediction errors.
[0006] To achieve the above objectives, this invention provides a method for predicting the residual field of ultrathin silicon steel sheets under complex excitations based on neural networks, comprising: A neural network model for predicting the residual field is constructed, wherein the input parameters of the neural network model include the harmonic characteristics of the excitation and the peak value of the magnetic induction intensity; Physical constraints reflecting the nonlinear magnetization behavior of ultrathin silicon steel sheets are defined, and these physical constraints are constructed as loss functions. During the training process of the neural network model, the loss function corresponding to the physical constraints is included in the total loss function to guide the network training to conform to physical laws. Using the trained neural network model, the corresponding residual field prediction value is output based on the input excitation parameters.
[0007] Optionally, the physical constraints include hysteresis constraints, thermodynamic irreversibility constraints, and temperature constraints.
[0008] Optionally, the hysteresis constraint is used to limit the peak value of the magnetic flux density predicted by the model from not exceeding the inherent saturation magnetic flux density of the material.
[0009] Optionally, the thermodynamic irreversible constraint is used to ensure that the energy dissipation of the magnetization process described by the model is non-negative, so as to conform to the law of irreversible magnetization.
[0010] Optionally, the temperature characteristic constraint is used to force the model to conform to the physical law that the residual field strength decreases as the temperature increases.
[0011] Optionally, the neural network model includes a first hidden layer and a second hidden layer; the first hidden layer is used to learn the nonlinear relationship between the input parameters and the magnetic field strength; the second hidden layer is used to incorporate the physical constraints.
[0012] Optionally, the training process of the neural network model further includes: acquiring hysteresis loop data of ultrathin silicon steel sheets measured under complex excitation containing different harmonic orders and harmonic contents; and calculating the measured value of the residual field from the hysteresis loop data based on the loss separation theory, as a training label.
[0013] Optionally, the method further includes a verification step, which compares the predicted residual field value output by the model with the measured residual field value to assess the prediction accuracy based on relative error.
[0014] An electronic device, the electronic device comprising: a processor and a memory storing computer program instructions; When the processor executes the computer program instructions, it implements the method for predicting the residual field of ultrathin silicon steel sheets under complex excitation based on neural networks.
[0015] A computer storage medium storing computer program instructions, which, when executed by a processor, implement the method for predicting the residual field of an ultrathin silicon steel sheet under complex excitation based on a neural network.
[0016] Technical advantages of this invention: This invention discloses a method for predicting the residual field of ultrathin silicon steel sheets under complex excitation based on neural networks. By embedding three major physical constraints—hysteresis characteristics, thermodynamic irreversibility, and temperature characteristics—into the neural network training as a loss function, it effectively overcomes the shortcomings of purely data-driven models lacking physical meaning and traditional analytical models lacking accuracy. This method ensures that the output of the neural network model, when predicting the residual field under complex excitation, strictly follows the inherent magnetization and thermodynamic laws of the material, significantly improving the physical reliability of the prediction results and the interpretability of the model. Simultaneously, this model possesses excellent generalization ability, enabling accurate and reliable prediction of the residual field under arbitrary complex excitation waveforms outside the training data, providing an effective technical tool for the selection of core materials and accurate loss assessment of equipment such as high-frequency transformers. Attached Figure Description
[0017] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart illustrating the method for predicting the residual field of an ultrathin silicon steel sheet under complex excitation based on a neural network, according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the neural network design according to an embodiment of the present invention; Figure 3 This is a schematic diagram of an experimental platform for measuring the magnetic properties of ultrathin silicon steel according to an embodiment of the present invention; Figure 4 This is a comparison chart of the predicted and measured residual field values of the ultrathin silicon steel material in an embodiment of the present invention, where (a) represents the 3rd harmonic with a harmonic content of 17%. B max =1.9T, (b) is the 5th harmonic, 5% harmonic content. B max =0.4T. Detailed Implementation
[0018] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.
[0019] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.
[0020] like Figure 1 As shown, this embodiment provides a method for predicting the residual field of ultrathin silicon steel sheets under complex excitation based on neural networks, including: A neural network model for predicting the residual field is constructed, wherein the input parameters of the neural network model include the harmonic characteristics of the excitation and the peak value of the magnetic induction intensity; Physical constraints reflecting the nonlinear magnetization behavior of ultrathin silicon steel sheets are defined, and these physical constraints are constructed as loss functions. During the training process of the neural network model, the loss function corresponding to the physical constraints is included in the total loss function to guide the network training to conform to physical laws. Using the trained neural network model, the corresponding residual field prediction value is output based on the input excitation parameters.
[0021] Furthermore, the physical constraints include hysteresis constraints, thermodynamic irreversibility constraints, and temperature constraints.
[0022] Specifically, the implementation process of this embodiment includes: By considering three physical constraints—hysteresis constraint, thermodynamic irreversibility constraint, and temperature constraint—the nonlinear magnetization behavior of ultrathin silicon steel sheets under complex excitation is characterized, making the model prediction no longer a simple data-driven process, but giving the model certain physical meaning.
[0023] Furthermore, the hysteresis constraint is used to limit the peak value of the magnetic flux density predicted by the model from not exceeding the inherent saturation magnetic flux density of the material.
[0024] Specifically, the implementation process of this embodiment includes: This invention characterizes the nonlinear magnetization behavior of ultrathin silicon steel sheets under complex excitations by considering three physical constraints: hysteresis constraint, thermodynamic irreversibility constraint, and temperature constraint. Previous model predictions were merely mathematically driven and lacked physical meaning. This invention, through the introduction of these three physical constraints, achieves physical meaning in the model. Among these, the hysteresis constraint requires that the relationship between the magnetic field strength and magnetic induction intensity generated by the residual field during the prediction process must conform to the inherent properties of the ultrathin silicon steel sheet. In particular, the peak value of the predicted magnetic induction intensity under arbitrary complex excitations must be considered. B a It cannot exceed the material's inherent maximum magnetic flux density. B sThat is, the predicted peak value of magnetic flux density B a The following relationship must be satisfied: (1).
[0025] Furthermore, the thermodynamic irreversible constraint is used to ensure that the energy dissipation of the magnetization process described by the model is non-negative, so as to conform to the law of irreversible magnetization.
[0026] Specifically, the implementation process of this embodiment includes: Thermodynamic irreversible constraint: During the magnetization process of ultrathin silicon steel sheets, it is necessary to ensure that the energy consumed is greater than or equal to 0, that is, to ensure that the ultrathin silicon steel sheets will not return to the initial magnetization process; in addition, during the formation of the residual field, the irreversible law of magnetization must be ensured. The formation of the residual field of the material is that after the magnetic field strength returns to 0, due to the irreversible change in its internal magnetic domain structure, a part of the magnetization direction is permanently retained, that is, the residual magnetic induction intensity.
[0027] Furthermore, the temperature characteristic constraint is used to force the model to conform to the physical law that the residual field strength decreases as the temperature increases.
[0028] Specifically, the implementation process of this embodiment includes: Temperature-related constraints: The formation of the residual field is a result of irreversible reversal of the magnetic domain walls within the material. Increased temperature intensifies the movement of these domain walls within the silicon steel sheet, causing smaller domain walls to merge. Furthermore, increased temperature intensifies the thermal motion of atoms within the silicon steel sheet. This disrupts the ordered arrangement of magnetic moments within the domains, causing some magnetic moments to deviate from their original directions, thus weakening the overall residual magnetic field strength. Therefore, the residual field must conform to the physical law that increases in temperature lead to a decrease in the residual field.
[0029] Furthermore, the neural network model includes a first hidden layer and a second hidden layer; the first hidden layer is used to learn the nonlinear relationship between the input parameters and the magnetic field strength; the second hidden layer is used to incorporate the physical constraints.
[0030] Specifically, the implementation process of this embodiment includes: like Figure 2 The neural network shown is constructed and designed with the following input layers: harmonic order, harmonic content, and peak magnetic induction intensity. The relationship between the input layer and the magnetic field intensity is learned through the hidden layer 1. The three physical constraints considered in step 1 are then incorporated through the hidden layer 2. At the same time, a loss function is introduced to correct the predicted value. Finally, the model outputs the predicted value of the residual field.
[0031] Hidden layer 1 consists of 128 neurons. Through deep learning, it establishes a nonlinear relationship between the three input parameters—harmonic content, harmonic order, and peak magnetic flux density—and the magnetic field strength. Hidden layer 2 consists of 80 neurons, incorporating the three physical constraints considered in step 1. A corresponding loss function is introduced for each physical constraint to reduce prediction error. The loss function for the hysteresis constraint is expressed as follows: (2); Among them, the model predicts y The value, through the tanh function, is mapped to the inherent properties of the material. B s on, and the predicted B a should not be greater than B To reduce the impact of the excess portion on the prediction results, s is introduced. L 1. Loss function The value represents the weighting coefficient, which is 50% in this invention. The ReLU function retains the value only if the value within the parentheses is positive; otherwise, it outputs 0. Equation (2) can be used to predict the residual field. B The correction of a.
[0032] The mathematical function of the thermodynamic irreversibility constraint is expressed as: (3); Here, ∮ represents the closed-loop integral, i.e., the loop integral. Furthermore, based on the consistency between Bertitto's loss separation theory and field separation techniques, the loss field can be divided into the hysteresis field. H hy eddy field H cl and remaining fields H ex And hysteresis field H hy The eddy field can be predicted by the Preisach resignation model. H cl It can be calculated using the corresponding field strength formula. Therefore, the hysteresis field... H hy and vortex field H cl Numerical integration of the hysteresis loop shows that when the value of equation (3) is greater than or equal to 0, it indicates that energy dissipation has occurred during the magnetization process, ensuring that the material will not return to its initial magnetization state. This is achieved by introducing an irreversible multiplication factor into its loss function. L In 2, it is represented as: (4); in, H The change in the loss field, This represents the weighting coefficient, which is 30% in this invention. When the material does not dissipate energy during magnetization, then it can be achieved through… L 2. Suppress the loss function.
[0033] Loss function constrained by temperature characteristics L 3 is represented as: (5); In the formula, represent L 3. The weight of the loss function is 20% in this invention. Represents the rate of change of the residual field with temperature. L 3 is used to punish when B a increases with increasing temperature, reducing the error in the model's prediction of the residual field.
[0034] Furthermore, the training process of the neural network model also includes: acquiring hysteresis loop data of ultrathin silicon steel sheets measured under complex excitation containing different harmonic orders and harmonic contents; and calculating the measured value of the residual field from the hysteresis loop data based on the loss separation theory, which is used as the training label.
[0035] Specifically, the implementation process of this embodiment includes: Based on the constructed neural network model, the residual field prediction under arbitrary complex excitation is realized, and the prediction results are compared with the experimental measurement and calculation results. Using the BROCKHAUS magnetic property measurement platform, the hysteresis loop data of ultrathin silicon steel under complex excitation containing 3rd, 5th and 17th harmonics with harmonic content of 5%, 17% and 30% are measured, and the residual field data of the corresponding working conditions are calculated by equation (6): (6); in, H mea This represents the experimentally measured magnetic field strength. H ex-cal This represents the residual field calculated based on experimental measurement results.
[0036] Furthermore, the method also includes a verification step, which compares the predicted residual field value output by the neural network model with the measured residual field value to assess the prediction accuracy based on the relative error.
[0037] Specifically, the implementation process of this embodiment includes: The neural network model of this invention is used to predict the corresponding residual field. H ex-image The accuracy of the model's prediction of the residual field is expressed by the relative error as follows: (7).
[0038] An electronic device, the electronic device comprising: a processor and a memory storing computer program instructions; When the processor executes the computer program instructions, it implements the method for predicting the residual field of ultrathin silicon steel sheets under complex excitation based on neural networks.
[0039] A computer storage medium storing computer program instructions, which, when executed by a processor, implement the method for predicting the residual field of an ultrathin silicon steel sheet under complex excitation based on a neural network.
[0040] Combining the neural network model of this invention with Figure 3 The experimental platform measured the data and obtained the following results: Figure 4 The graph showing the comparison between predicted and measured residual field values is shown. Figure 4 (a) represents the 3rd harmonic, with a harmonic content of 17%. B max =1.9T, Figure 4 (b) represents the 5th harmonic, with a harmonic content of 5%. B max =0.4T. The results show that this embodiment can accurately simulate the residual field of ultrathin silicon steel material under complex excitation. By comparing with the measured loss value, the calculation error is kept at about 4%.
[0041] This invention discloses a method for predicting the residual field of ultrathin silicon steel sheets under complex excitations based on neural networks. By embedding three major physical constraints—hysteresis characteristics, thermodynamic irreversibility, and temperature characteristics—into the neural network training as a loss function, it effectively overcomes the shortcomings of purely data-driven models lacking physical meaning and traditional analytical models lacking accuracy. This method ensures that the output of the neural network model, when predicting the residual field under complex excitations, strictly follows the inherent magnetization and thermodynamic laws of the material, significantly improving the physical reliability of the prediction results and the interpretability of the model. Simultaneously, the model possesses excellent generalization ability, enabling accurate and reliable prediction of the residual field under arbitrary complex excitation waveforms outside the training data, providing an effective technical tool for the selection of core materials and accurate loss assessment of equipment such as high-frequency transformers.
[0042] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.
Claims
1. A method for predicting the residual field of ultrathin silicon steel sheets under complex excitation based on neural networks, characterized in that, include: A neural network model for predicting the residual field is constructed, wherein the input parameters of the neural network model include the harmonic characteristics of the excitation and the peak value of the magnetic induction intensity; Physical constraints reflecting the nonlinear magnetization behavior of ultrathin silicon steel sheets are defined, and these physical constraints are constructed as loss functions. During the training process of the neural network model, the loss function corresponding to the physical constraints is included in the total loss function to guide the network training to conform to physical laws. Using the trained neural network model, the corresponding residual field prediction value is output based on the input excitation parameters.
2. The method for predicting the residual field of ultrathin silicon steel sheets under complex excitation based on neural networks as described in claim 1, characterized in that, The physical constraints include hysteresis constraints, thermodynamic irreversibility constraints, and temperature constraints.
3. The method for predicting the residual field of ultrathin silicon steel sheets under complex excitation based on neural networks as described in claim 2, characterized in that, The hysteresis constraint is used to limit the peak value of the magnetic flux density predicted by the model to not exceed the material's inherent saturation magnetic flux density.
4. The method for predicting the residual field of ultrathin silicon steel sheets under complex excitation based on neural networks as described in claim 2, characterized in that, The thermodynamic irreversible constraint is used to ensure that the energy dissipation of the magnetization process described by the model is non-negative, so as to conform to the law of irreversible magnetization.
5. The method for predicting the residual field of ultrathin silicon steel sheets under complex excitation based on neural networks as described in claim 2, characterized in that, The temperature characteristic constraint is used to force the model to conform to the physical law that the residual field strength decreases as the temperature increases.
6. The method for predicting the residual field of ultrathin silicon steel sheets under complex excitation based on neural networks as described in claim 1, characterized in that, The neural network model includes a first hidden layer and a second hidden layer; the first hidden layer is used to learn the nonlinear relationship between the input parameters and the magnetic field strength; the second hidden layer is used to incorporate the physical constraints.
7. The method for predicting the residual field of ultrathin silicon steel sheets under complex excitation based on neural networks as described in claim 1, characterized in that, The training process of the neural network model also includes: acquiring hysteresis loop data of ultrathin silicon steel sheets measured under complex excitation containing different harmonic orders and harmonic contents; and calculating the measured value of the residual field from the hysteresis loop data based on the loss separation theory, which is used as the training label.
8. The method for predicting the residual field of ultrathin silicon steel sheets under complex excitation based on neural networks as described in claim 1, characterized in that, The method also includes a verification step, which compares the predicted residual field value output by the model with the measured residual field value to assess the prediction accuracy based on the relative error.
9. An electronic device, characterized in that, The electronic device includes: a processor and a memory storing computer program instructions; When the processor executes the computer program instructions, it implements the method for predicting the residual field of ultrathin silicon steel sheets under complex excitation based on neural networks as described in any one of claims 1-8.
10. A computer storage medium, characterized in that, The computer storage medium stores computer program instructions, which, when executed by a processor, implement the method for predicting the residual field of an ultrathin silicon steel sheet under complex excitation based on a neural network as described in any one of claims 1-8.