Magnetic loss prediction method for soft magnetic composite material

Abstract

The invention discloses a magnetic loss prediction method for a soft magnetic composite material. The method comprises the following steps: acquiring data; a magnetic induction intensity peak value and a magnetic induction intensity root-mean-square are calculated through the magnetic induction intensity waveform or the discrete sampling points, and the waveform shape factor is constructed; constructing material factors: constructing the material factors based on material identification including material physical parameters or statistical characterization quantity, wherein the material factors are used for characterizing material differences; and multi-factor coupling prediction: on the basis of a temperature correction Steinmetz model based on temperature and frequency, introducing correction terms of a waveform shape factor and a material factor, and outputting predicted magnetic loss. According to the method, a temperature-corrected Steinmetz model is taken as a physical kernel, multi-dimensional generalization is carried out through waveform shape factors and material factors, and finally, a prediction model with high stability under the conditions of a wide temperature range, multiple frequencies, arbitrary waveforms and different materials is realized.

Description

Technical Field

[0001] This invention relates to the field of magnetic loss prediction, and more particularly to a method for predicting magnetic loss in soft magnetic composite materials. Background Technology

[0002] In devices such as motors, inductors / transformers, wireless charging, and power electronic converters, the energy conversion efficiency and temperature rise level of magnetic components largely depend on the magnetic loss performance of the core material. Soft magnetic composites (SMCs) are typically formed by coating soft magnetic powder particles with an insulating layer, followed by pressing and heat treatment. They feature high resistivity, three-dimensional isotropy, and ease of fabrication into complex structures, making them suitable for the design of magnetic devices operating at medium to high frequencies.

[0003] Magnetic loss generally comprises hysteresis loss, eddy current loss, and additional losses, and its magnitude is influenced by frequency f, magnetic flux density amplitude (or peak magnetic induction intensity Bmax), temperature T, material formulation / process, and excitation waveform shape. In practical applications, excitation often exhibits non-sinusoidal waveforms such as triangular waves, trapezoidal waves, and PWM; simultaneously, the operating temperature range of devices is wide, and batch-to-batch and process variations can cause parameter drift. Therefore, establishing a magnetic loss prediction method with high accuracy and strong generalization under conditions of "wide temperature range + multiple waveforms + multiple materials" is an important foundation for SMC engineering design and rapid material selection.

[0004] Defects and shortcomings of existing technology: 1) Traditional Steinmetz empirical models and some improved models are mainly designed for sinusoidal excitation. When faced with non-sinusoidal excitation such as triangular waves, trapezoidal waves, and PWM, additional waveform assumptions or separate modeling of each waveform are usually required, making it difficult to form a unified and transferable prediction framework. 2) The coefficients of the magnetic loss model change significantly with temperature. Existing temperature correction methods are prone to unstable fitting or increased extrapolation error under wide temperature range conditions, making it difficult to simultaneously ensure the prediction accuracy at different temperature points. 3) Different SMC formulations, insulation layers and molding / heat treatment processes lead to significant differences in losses. Existing methods often require recalibrating parameters for each material, and their cross-material generalization ability is insufficient. 4) Parameter identification often uses least squares or gradient-based local optimization. When the nonlinearity of the coupled model is high, it is easy to get trapped in local optima, resulting in insufficient model stability and reliability. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method for predicting magnetic loss in soft magnetic composite materials.

[0006] The objective of this invention is achieved through the following technical solution: A first aspect of the present invention provides a method for predicting magnetic loss in soft magnetic composite materials, comprising the following steps: Data Acquisition: Obtaining Temperature ,frequency Magnetic induction intensity waveform or discrete sampling points of magnetic induction intensity Material labeling; Constructing the waveform shape factor: based on the magnetic flux density waveform or discrete sampling points Calculate the peak value of magnetic flux density Root mean square of magnetic induction And construct waveform shape factor ; Constructing material factors: Constructing material factors based on material identifiers, including material properties or statistical characterization parameters. , used to characterize material differences; Multi-factor coupled prediction: in temperature-based and frequency Based on the temperature-corrected Steinmetz model, a waveform shape factor is introduced. With material factor The correction term outputs the predicted magnetic loss. .

[0007] Furthermore, the process between data acquisition and waveform shape factor construction also includes: Data preprocessing: Outlier removal, missing value handling, and necessary normalization / standardization are performed on the samples to improve the robustness of the model.

[0008] Further, the construction of the waveform shape factor includes: Magnetic flux density waveform for each sample or discrete sampling points , i=1…n, calculate the peak value of magnetic induction intensity Root mean square of magnetic induction : ; ; Among them, the magnetic induction intensity waveform This represents a continuous waveform function showing the change of magnetic flux density with time t over one excitation cycle. (Magnetic flux density waveform) The value included at sampling time ti That is, discrete sampling points ; Define the waveform shape factor. : ; The waveform shape factor Used to unify the morphological differences of non-sinusoidal waveforms, including sine waves, triangle waves, trapezoidal waves, and PWM.

[0009] Further, the construction material factors include: Material factors are constructed using at least one measurable material property parameter, specifically including material resistivity. ,density Relative permeability Coercivity One or more of them, and normalized: ; In the formula, Indicates the reference resistivity. This represents the normalized resistivity of the material. Indicates the reference relative permeability. This represents the normalized permeability. Indicates reference coercivity. This represents the normalized coercivity. Indicates the reference density. This represents the normalized density; Construction material factor F: ; In the formula, ~ These are preset or unidentified weighting coefficients. If some measurable material properties are unavailable, the corresponding parts are removed.

[0010] Furthermore, in the multi-factor coupled prediction, based on temperature and frequency The temperature-corrected Steinmetz model specifically includes: Based on the Steinmetz equation, a temperature correction term is introduced to make the model coefficients change with temperature: , , ; In the formula, k This represents the loss coefficient as a function of temperature. The frequency index represents the change with temperature. The magnetic flux density index represents how much magnetic flux density changes with temperature. , , , , , These are preset or parameters to be identified; Based on temperature and frequency Temperature-corrected Steinmetz model for: ; In the formula, This represents the peak value of the magnetic flux density.

[0011] Furthermore, in the multi-factor coupled prediction, a waveform shape factor is introduced. With material factor The correction term outputs the predicted magnetic loss. ,include: ; In the formula, , The coupling coefficient is obtained by fitting the training data.

[0012] Furthermore, the predicted magnetic loss Global parameter optimization is required, and simulated annealing is used to globally optimize and identify the coupling coefficients.

[0013] Furthermore, the global optimization of the parameters corresponds to the objective function as follows: ; in, This represents the magnetic loss value obtained by experimental measurement of the i-th sample under the corresponding temperature Ti, frequency fi, magnetic induction intensity waveform Bi(t) or discrete sampling point Bi, and material conditions; This indicates that the predicted magnetic loss value is obtained by substituting the input variables of the i-th sample into the magnetic loss prediction model and using the set of parameters to be optimized, θ. This represents the set of parameters to be optimized, which includes at least... , ,for , , , , , , ~ , can be preset according to actual conditions or used as parameters to be optimized.

[0014] The beneficial effects of this invention are: In an exemplary embodiment of the present invention, a temperature-corrected Steinmetz model is used as the physical kernel, and multi-dimensional generalization is performed through waveform shape factor and material factor, ultimately realizing a prediction model with high stability over a wide temperature range, multiple frequencies, arbitrary waveforms, and different materials. Attached Figure Description

[0015] Figure 1 A flowchart illustrating a magnetic loss prediction method for soft magnetic composite materials, provided as an exemplary embodiment of the present invention. Detailed Implementation

[0016] The technical solution of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Furthermore, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0017] See Figure 1 , Figure 1 A flowchart illustrating an exemplary embodiment of the present invention provides a method for predicting magnetic loss in soft magnetic composite materials, comprising the following steps: Data Acquisition: Obtaining Temperature ,frequency Magnetic induction intensity waveform or discrete sampling points of magnetic induction intensity Material labeling; Constructing the waveform shape factor: based on the magnetic flux density waveform or discrete sampling points Calculate the peak value of magnetic flux density Root mean square of magnetic induction And construct waveform shape factor ; Constructing material factors: Constructing material factors based on material identifiers, including material properties or statistical characterization parameters. , used to characterize material differences; Multi-factor coupled prediction: in temperature-based and frequency Based on the temperature-corrected Steinmetz model, a waveform shape factor is introduced. With material factor The correction term outputs the predicted magnetic loss. .

[0018] Specifically, in this exemplary embodiment, by introducing waveform shape factor and material factor, the limitations of traditional magnetic loss prediction models are fundamentally changed: (1) Traditional models are essentially “curve fitters” for specific waveforms (usually ideal sine waves). Once the actual excitation becomes a square wave, triangular wave, or a PWM waveform containing harmonics, the model predictions will deviate significantly, and engineers will have to perform a large number of experiments to recalibrate for each new waveform, or use a completely black-box compensation algorithm. In this exemplary embodiment, however, the waveform shape factor transforms the geometric characteristics of the waveform into a continuous, computable physical quantity. This gives the model the ability to generalize across waveform dimensions.

[0019] For a calibrated model, when the input changes from a sine wave to a triangular wave or a non-ideal square wave, it is only necessary to calculate the shape factor of the new waveform and substitute it into the formula. Reliable prediction results can be obtained without retraining or adjusting the core parameters. This is equivalent to enabling the model to understand how waveform shape affects loss, thus allowing it to smoothly handle various waveforms and converge prediction errors under different waveforms to a relatively stable range.

[0020] (2) The Steinmetz model coefficients are different for each material grade and even each production batch. Developing new magnetic materials or changing suppliers requires a lot of time and resources to remeasure across the entire frequency, magnetic flux density, and temperature range to fit a completely new set of parameters, which is a lengthy and costly process. In this exemplary embodiment, however, the material factor decouples the model’s dependence on the material from the implicit, global coefficients and explicitly associates it with the intrinsic physical properties of the material.

[0021] When faced with a new material, engineers no longer need to start from scratch. They only need to measure a few key physical properties of the material, calculate the material factor, and then substitute it into an existing, validated model framework. This is essentially a physics-based interpolation or extrapolation. While initial predictions may require calibration for entirely new material families with vastly different physical properties, it can significantly reduce the amount of experimental data required for different formulations or batches within the same material system. This drastically reduces the validation cycle and cost of material development, enabling effective knowledge transfer from known to unknown materials.

[0022] (3) Regarding temperature: The coupling of temperature and materials: The material factor determines how temperature affects the intrinsic properties of a material, while the temperature correction term quantifies the amplification or reduction of this change in the final loss. For example, the change in material resistivity at high temperatures significantly affects eddy current losses, and the model can accurately capture this through the coupling of the two.

[0023] Coupling of Temperature and Waveform: Temperature affects the saturation flux density of the magnetic core, thus indirectly altering the degree of waveform shear distortion under high magnetic flux density excitation. When dealing with nonlinear waveforms, the temperature term in the model provides a basis for correcting this indirect effect.

[0024] Unified prediction output: Regardless of whether engineers input high-frequency square waves driving new materials at high temperatures or low-frequency sine waves driving mature materials at low temperatures, the model can perform continuous and smooth calculations within a unified framework, outputting stable prediction values. This completely changes the fragmented approach of switching between different models or empirical coefficients for different operating conditions.

[0025] This method uses the temperature-corrected Steinmetz model as its physical kernel and performs multi-dimensional generalization through waveform shape factor and material factor, ultimately achieving a prediction model with high stability over a wide temperature range, multiple frequencies, arbitrary waveforms, and different materials.

[0026] The following will describe each exemplary embodiment in detail: More preferably, in an exemplary embodiment, the process further includes: between data acquisition and constructing the waveform shape factor. Data preprocessing: Outlier removal, missing value handling, and necessary normalization / standardization are performed on the samples to improve the robustness of the model.

[0027] Specifically, in this exemplary embodiment, the training samples are cleaned and outlier removed (preferably using the IQR method), and necessary normalization / standardization is performed. The input quantities are then standardized and given uniform units.

[0028] This approach improves the model's generalization ability and predictive stability. Raw experimental data inevitably contains measurement errors, random interference (such as transient fluctuations in testing equipment, environmental electromagnetic interference), and even gross errors. If these noises and outliers are directly used for training, they contaminate the model, causing it to learn incorrect patterns to fit these errors, resulting in inaccurate predictions when faced with new, clean data—that is, overfitting or poor stability. However, by using statistical methods such as IQR (interquartile range) to remove outliers, supplemented by missing value handling, the purity of the training dataset is ensured. The model learns the true physical laws governing the effects of materials, waveforms, and temperature on losses, rather than data noise. This directly improves the model's generalization accuracy and predictive stability on unknown data, making its output more reliable.

[0029] Furthermore, it ensures the comparability of multi-source data and model convergence: input features (such as temperature T, frequency f, peak value, resistivity ρ) have different dimensions and numerical ranges (e.g., frequency of 10⁻⁶). 3 -10 6 Hz, resistivity 10 -3 -10 0(Ω·m). This order-of-magnitude difference leads to difficulties in model training and distortion of feature weights. Therefore, it is necessary to normalize / standardize the input features (such as Z-score normalization or Min-Max normalization). This can accelerate and stabilize training: all features are placed on the same scale, allowing the optimization algorithm to find the optimal solution more efficiently and smoothly; it fairly reflects feature contributions: the model can more accurately assess the true impact weight of each physical quantity (such as material resistivity and waveform peak value) on loss, enhancing the physical consistency of the model.

[0030] More preferably, in an exemplary embodiment, the construction of the waveform shape factor includes: Magnetic flux density waveform for each sample or discrete sampling points , i=1…n, calculate the peak value of magnetic induction intensity Root mean square of magnetic induction : ; ; Among them, the magnetic induction intensity waveform This represents a continuous waveform function showing the change of magnetic flux density with time t over one excitation cycle. (Magnetic flux density waveform) The value included at sampling time ti That is, discrete sampling points ; Define the waveform shape factor. : ; The waveform shape factor Used to unify the morphological differences of non-sinusoidal waveforms, including sine waves, triangle waves, trapezoidal waves, and PWM.

[0031] Specifically, in this exemplary embodiment, the parameterization of waveform influence is implemented. The model no longer needs to treat waveform type as a discrete, unordered category label (such as one-hot encoded sine or trigonometric), but instead introduces a continuous, physically meaningful numerical feature, namely the waveform shape factor M. This enables the model to learn the continuous functional relationship between loss and M.

[0032] It should be noted that some existing technologies employ a combination of "empirical models + machine learning + neural network compensation + weighted summaries," rather than continuously quantifying waveform morphology using a single empirical formula. The predicted value is obtained by combining the two prediction results and the compensation term according to their weights. Under this approach, waveforms often appear as "operating condition characteristics / category variables," making it difficult to express "continuous changes within the same type of waveform" (e.g., morphological differences caused by PWM duty cycle changes, distortion, and noise). This can lead to instability in generalization under "unseen waveforms / distorted waveforms." For example, treating materials and waveforms as discrete variables and processing them sequentially is essentially still "classification / numbering," not continuous morphological quantification.

[0033] In this exemplary embodiment, instead of assigning numbers to the waveforms, statistical quantities such as peak value and root mean square are calculated based on the magnetic induction intensity waveform B(t) or discrete sampling points, and a waveform shape factor M is constructed to uniformly quantify the morphological differences of waveforms such as sine, triangular, trapezoidal, and PWM. Subsequently, M is coupled as a correction term into the single prediction formula of the temperature correction Steinmetz (instead of building an additional "waveform classifier / compensator" and then weighting it).

[0034] This approach has the following advantages: (1) Enhanced continuous generalization capability of waveform dimension: As long as waveform data can be collected, M can be calculated and substituted into the same formula, so that different waveforms can be continuously connected under the same model, reducing the risk of "retraining / recalibrating when encountering new waveforms".

[0035] (2) Unified modeling: Avoid keeping the “waveform effect” at the level of categorical variables or black box compensation, but instead enter into an interpretable empirical model structure, which is conducive to explaining “how waveform changes affect loss”.

[0036] More preferably, in an exemplary embodiment, the construction material factor includes: Material factors are constructed using at least one measurable material property parameter, specifically including material resistivity. ,density Relative permeability Coercivity One or more of them, and normalized: ; In the formula, Indicates the reference resistivity. This represents the normalized resistivity of the material. Indicates the reference relative permeability. This represents the normalized permeability. Indicates reference coercivity. This represents the normalized coercivity. Indicates the reference density. This represents the normalized density; Construction material factor F: ; In the formula, ~ These are preset or unidentified weighting coefficients. If some measurable material properties are unavailable, the corresponding parts are removed.

[0037] Specifically, in this exemplary embodiment, feature selection is first performed: intrinsic physical property parameters directly related to the physical mechanism of magnetic loss are selected. resistivity : Dominant eddy current loss. The higher the resistivity, the lower the eddy current loss. Coercivity The dominant hysteresis loss. Lower coercivity results in a narrower hysteresis loop and lower hysteresis loss. Relative permeability. Density affects magnetic field distribution and magnetization difficulty, and is related to hysteresis and eddy current losses. The mass of magnetic material per unit volume is directly related to the total power loss density. A linear weighted synthesis is then performed. Physical meaning: This formula projects the multidimensional material characteristic space onto a one-dimensional material factor F, with weighting coefficients... ~ This determines the relative importance of each physical property parameter to the final loss.

[0038] This approach achieves dimensionality reduction and quantification of material characteristics: complex material differences are represented by a single number, the material factor F. In the prediction model, only the material factor F needs to be used as an input feature to distinguish different materials, greatly simplifying the model structure.

[0039] It should be noted that, similar to the waveform shape factor, this exemplary embodiment introduces a material factor F: constructed from at least one measurable material property parameter (such as resistivity, density, relative permeability, coercivity, etc.), and obtained by normalization and combination; the weight can be preset or to be identified. Like M, F is coupled as a correction term into the same temperature-corrected Steinmetz empirical model, forming a single coupled formula.

[0040] The technical effects achieved include: (1) Improved cross-material migration / generalization ability: Material differences are no longer remembered by "numbering", but driven by physical property parameters, which enables the model to make interpolation predictions for new materials / new batches (at least significantly reducing the pressure of "re-calibrating the entire batch every time a material is changed").

[0041] (2) Reduced engineering costs: When updating materials, physical property parameters can be measured first and losses can be quickly estimated, reducing reliance on large-scale loss experimental data.

[0042] In a specific exemplary embodiment, the material identification sources include: (i) obtaining the material's physical properties (e.g., resistivity ρ, density d, relative permeability μr, coercivity Hc) by looking up a table or testing to construct the material factor F; and / or (ii) grouping samples according to the material identification and calculating statistical characterizations of the material (e.g., mean frequency fi within the material group, average peak value of the material, etc.) to construct the material factor F.

[0043] More preferably, in an exemplary embodiment, the multi-factor coupled prediction is based on temperature. and frequency The temperature-corrected Steinmetz model specifically includes: Based on the Steinmetz equation, a temperature correction term is introduced to make the model coefficients change with temperature: , , ; In the formula, k This represents the loss coefficient as a function of temperature. The frequency index represents the change with temperature. The magnetic flux density index represents how much magnetic flux density changes with temperature. , , , , , These are preset or parameters to be identified; Based on temperature and frequency Temperature-corrected Steinmetz model for: ; In the formula, This represents the peak value of the magnetic flux density.

[0044] Specifically, in the prior art, only k is temperature-corrected, while this exemplary embodiment allows all three core indices α and β to vary with temperature, which is crucial and physically realistic.

[0045] α(T): Frequency exponent as a function of temperature. This accurately describes the ratio of hysteresis loss to eddy current loss at different temperatures. For example, changes in resistivity at high temperatures may alter the frequency dependence of eddy current loss. β(T): Magnetic flux density exponent as a function of temperature. This reflects the nonlinear effect of temperature on the magnetization process, especially near saturation. k(T): Combined loss coefficient, capturing other unmodeled temperature effects.

[0046] More preferably, in an exemplary embodiment, a waveform shape factor is introduced into the multi-factor coupled prediction. With material factor The correction term outputs the predicted magnetic loss. ,include: ; In the formula, , The coupling coefficient is obtained by fitting the training data.

[0047] Specifically, in this exemplary embodiment, the five core factors affecting magnetic loss—temperature, frequency, magnetic flux density, waveform, and material—are unified, achieving a leap from precise fitting under specific operating conditions to continuous generalization prediction across all operating conditions. It also possesses interpolation and extrapolation capabilities for new waveforms, new materials, and new temperature points. Robustness is ensured through data preprocessing, interpretability is provided through a modular white-box model, and data dependency and implementation costs are reduced through staged parameter fitting.

[0048] More preferably, in an exemplary embodiment, the predicted magnetic loss Global parameter optimization is required, and simulated annealing is used to globally optimize and identify the coupling coefficients.

[0049] Specifically, after completing the innovative design of the model architecture (temperature correction Steinmetz + waveform shape factor + material factor), the model is no longer a simple linear or low-order nonlinear equation. The introduction of waveform factor M and material factor F, coupled with the basic model in a product form, results in a large number of local optima in the parameter space of the model. Using local search methods such as gradient descent, the results will heavily depend on the initial guess values ​​and are very likely to converge to a locally optimal but globally poor solution. This will lead to: (1) Model instability: two training sessions may result in models with huge performance differences; (2) Weak generalization ability: under locally optimal parameters, the model may only remember the noise of the training data and not learn the real physical laws, resulting in poor prediction ability for new data.

[0050] Therefore, to avoid the fitting of multiple coupled factors getting stuck in local optima, this exemplary embodiment preferably uses simulated annealing algorithm for parameter identification. This mechanism allows the algorithm to escape the current local optimum trap, thereby exploring a larger parameter space and ultimately converging to the global optimum or a region close to the global optimum with a higher probability. In other words, it can reduce the risk of local optima and improve the fitting stability and generalization ability.

[0051] Improved fit stability: Because SA is more likely to find the global optimum, the training results of the model are highly reproducible. Regardless of the initial parameters, it will eventually converge stably to the same set of high-performance parameters. This gives the model the reliability of an industry-standard tool.

[0052] Improving generalization ability: Globally optimal parameters mean that the model captures the true physical laws behind the data to the greatest extent possible, rather than overfitting to the noise in the training set. Therefore, when the model encounters unseen temperature points, new waveforms, or new materials, its predictions based on physical consistency will be more accurate and reliable. This is the fundamental guarantee for achieving generalized predictions over a wide temperature range, multiple waveforms, and multiple materials.

[0053] More preferably, in an exemplary embodiment, the parameters are globally optimized, and the corresponding objective function is: ; in, This represents the magnetic loss value obtained by experimental measurement of the i-th sample under the corresponding temperature Ti, frequency fi, magnetic induction intensity waveform Bi(t) or discrete sampling point Bi, and material conditions; This indicates that the predicted magnetic loss value is obtained by substituting the input variables of the i-th sample into the magnetic loss prediction model and using the set of parameters to be optimized, θ. This represents the set of parameters to be optimized, which includes at least... , ,for , , , , , , ~ , can be preset according to actual conditions or used as parameters to be optimized.

[0054] The simulated annealing process includes: (1) Initialization: Set the initial temperature Temp=Temp0, set the cooling coefficient r and the number of iterations N at each temperature, give the initial parameter solution θ=θ0, and calculate the objective function value E(θ); at the same time, set the current optimal solution to θ. ∗ =θ, the optimal objective function value is set to E. ∗ =E(θ).

[0055] Among them, θ′ represents a new solution obtained by perturbing the current solution θ in the neighborhood; ΔE represents the change in the objective function, ΔE = E(θ′)−E(θ); θ ∗ represents the optimal solution obtained by the algorithm search (or the currently recorded optimal solution).

[0056] (2) Neighborhood perturbation generation: In each iteration, randomly perturb the current solution θ to obtain a candidate solution θ′. The perturbation method can be: adding a random perturbation amount to one or more parameter components in the parameter set. For example, θ′ = θ + δ, where δ is a random vector satisfying a preset range; or randomly selecting a certain parameter component θ j and setting θ j′ = θ j + δ j , and keeping the remaining components unchanged; the perturbation amplitude can gradually decrease as the temperature decreases. If the parameter exceeds the preset value range after perturbation, truncate it to the boundary or resample it.

[0057] (3) Calculate the energy difference: Calculate ΔE = E(θ′)−E(θ).

[0058] (4) Accept or reject according to the Metropolis criterion: If ΔE ≤ 0, accept the new solution and set θ = θ′; if ΔE > 0, accept the new solution with a probability p = exp(−ΔE / Temp), that is, generate a uniform random number u ∈ (0,1), and when u < p, accept and set θ = θ′, otherwise reject and keep θ unchanged.

[0059] (5) Update the optimal solution: If E(θ) < E ∗ , then update θ ∗ = θ, and update E ∗ = E(θ).

[0060] (6) Cooling iteration: After completing N iterations at the current temperature, update the temperature according to the cooling strategy. For example, Temp ← r⋅Temp; repeat steps (2) to (6) until the termination condition is met (such as Temp being lower than the threshold, the change in the objective function being less than the threshold, or reaching the maximum number of iterations).

[0061] (7) Output: Output the optimal parameter solution θ ∗ , and obtain the final parameters of the magnetic loss prediction model accordingly.

[0062] Specifically, the simulated annealing SA of some existing technologies is used to optimize the internal hyperparameter (smoothing factor σ) of a black-box neural network (GRNN). The purpose is to make the black-box model itself train better without changing its black-box nature. It does not involve any empirical formula parameters with physical meanings.

[0063] In this exemplary embodiment, simulated annealing (SA) is used to optimize all interpretable parameters of a white-box physical empirical model. These parameters include the predicted magnetic loss of the temperature-corrected Steinmetz model. In , It can include basic model parameters and / or the weights of F, etc. (i.e.) , , , , , , ~ (Parameters such as these can be preset or used as parameters to be optimized based on actual conditions). In other words, the optimization process in this exemplary embodiment essentially involves using data to calibrate the parameters of a physical model, and the optimized model remains fully interpretable.

[0064] Existing local optimization methods (such as gradient descent and Levenberg-Marquardt) suffer from a significant drawback: they heavily rely on guessing initial parameters. If the initial values ​​are poor, they are highly susceptible to getting trapped in local optima, meaning they find a set of parameters that makes the model appear to fit the training set well, but fail to capture the true global physical laws. For the coupled model with numerous parameters and strong nonlinearity, as exemplified in this model, the risk of local optima is extremely high. This leads to poor model stability: training with different initial values ​​may result in vastly different models; and low reliability: predictions may deviate significantly in regions not covered by the training data (during generalization).

[0065] Therefore, in this exemplary embodiment, SA (Automatic Optimization) escapes local optima by introducing a mechanism that accepts poor solutions with a certain probability, enabling the optimization process to escape the valleys of local optima and thus having a greater chance of finding a globally optimal or near-global optimal parameter set. Furthermore, it is suitable for complex models: SA does not rely on the gradient information of the objective function, making it very suitable for solving parameter identification problems of complex empirical models with nonlinearity, multiple peaks, and parameter coupling, as in this invention. Ultimately, it improves fitting stability: because it is more likely to find a globally optimal solution, the parameter sets obtained from multiple independent training sessions will be very similar, significantly improving the model's repeatability and stability.

[0066] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art can make other variations or modifications based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A method for predicting magnetic loss in soft magnetic composite materials, characterized in that: Includes the following steps: Data Acquisition: Obtaining Temperature ,frequency Magnetic induction intensity waveform or discrete sampling points of magnetic induction intensity Material labeling; Constructing the waveform shape factor: based on the magnetic flux density waveform or discrete sampling points Calculate the peak value of magnetic flux density Root mean square of magnetic induction And construct waveform shape factor ; Constructing material factors: Constructing material factors based on material identifiers, including material properties or statistical characterization parameters. , used to characterize material differences; Multi-factor coupled prediction: in temperature-based and frequency Based on the temperature-corrected Steinmetz model, a waveform shape factor is introduced. With material factor The correction term outputs the predicted magnetic loss. .

2. The magnetic loss prediction method for soft magnetic composite materials according to claim 1, characterized in that: Between data acquisition and waveform shape factor construction, the following is also included: Data preprocessing: Outlier removal, missing value handling, and necessary normalization / standardization are performed on the samples to improve the robustness of the model.

3. The magnetic loss prediction method for soft magnetic composite materials according to claim 1, characterized in that: The waveform shape factor includes: Magnetic flux density waveform for each sample or discrete sampling points , i=1…n, calculate the peak value of magnetic induction intensity Root mean square of magnetic induction : ； ； Among them, the magnetic induction intensity waveform This represents a continuous waveform function showing the change of magnetic flux density with time t over one excitation cycle. (Magnetic flux density waveform) The value included at sampling time ti That is, discrete sampling points ; Define the waveform shape factor. : ； The waveform shape factor Used to unify the morphological differences of non-sinusoidal waveforms, including sine waves, triangle waves, trapezoidal waves, and PWM.

4. The magnetic loss prediction method for soft magnetic composite materials according to claim 3, characterized in that: The construction material factors include: Material factors are constructed using at least one measurable material property parameter, specifically including material resistivity. ,density Relative permeability Coercivity One or more of them, and normalized: ； In the formula, Indicates the reference resistivity. This represents the normalized resistivity of the material. Indicates the reference relative permeability. This represents the normalized permeability. Indicates reference coercivity. This represents the normalized coercivity. Indicates the reference density. This represents the normalized density; Construction material factor F: ； In the formula, ~ These are preset or unidentified weighting coefficients. If some measurable material properties are unavailable, the corresponding parts are removed.

5. The magnetic loss prediction method for soft magnetic composite materials according to claim 4, characterized in that: In the multi-factor coupled prediction, temperature is the basis. and frequency The temperature-corrected Steinmetz model specifically includes: Based on the Steinmetz equation, a temperature correction term is introduced to make the model coefficients change with temperature: ， ， ； In the formula, k This represents the loss coefficient as a function of temperature. The frequency index represents the change with temperature. The magnetic flux density index represents how much magnetic flux density changes with temperature. , , , , , These are preset or parameters to be identified; Based on temperature and frequency Temperature-corrected Steinmetz model for: ； In the formula, This represents the peak value of the magnetic flux density.

6. The magnetic loss prediction method for soft magnetic composite materials according to claim 5, characterized in that: In the multi-factor coupled prediction, a waveform shape factor is introduced. With material factor The correction term outputs the predicted magnetic loss. ,include: ； In the formula, , The coupling coefficient is obtained by fitting the training data.

7. The magnetic loss prediction method for soft magnetic composite materials according to claim 6, characterized in that: The predicted magnetic loss Global parameter optimization is required, and simulated annealing is used to globally optimize and identify the coupling coefficients.

8. The magnetic loss prediction method for soft magnetic composite materials according to claim 7, characterized in that: The objective function for the global optimization of the parameters is: ； in, This represents the magnetic loss value obtained by experimental measurement of the i-th sample under the corresponding temperature Ti, frequency fi, magnetic induction intensity waveform Bi(t) or discrete sampling point Bi, and material conditions; This indicates that the predicted magnetic loss value is obtained by substituting the input variables of the i-th sample into the magnetic loss prediction model and using the set of parameters to be optimized, θ. This represents the set of parameters to be optimized, which includes at least... , ,for , , , , , , ~ , can be preset according to actual conditions or used as parameters to be optimized.

Images