Magnetic Core Loss Prediction Method Based on Latent Space Temporal Knowledge Coupling Network

Abstract

This application relates to the field of electromagnetic control technology, and particularly to a method for predicting core loss based on a latent space temporal knowledge coupling network. A sample data set is collected, and the sample data set is divided into feature categories to obtain inherent attribute features and temporal attribute features; conditional encoding is used to preprocess the inherent attributes, and combined with an extended knowledge gate to obtain an inherent feature representation; data dimensionality reduction is used to preprocess the temporal attributes, and combined with a bidirectional long short-term memory network to obtain a temporal feature representation; the output data of the coupling extended knowledge gate and the output data of the bidirectional long short-term memory network are concatenated to construct a core loss prediction model; the sample data set of the material to be predicted is input into the core loss prediction model to output the predicted flux loss prediction result of the material to be predicted. The present invention dynamically fuses the conditional matrix and the temporal feature matrix, and through transfer learning, improves the generalization ability of the model, ensuring high-precision prediction even under unknown materials and working conditions.

Description

Technical Field

[0001] This application relates to the field of electromagnetic control technology, and particularly to a method for predicting core loss based on a latent space temporal knowledge coupling network. Background Art

[0002] In the design of modern ultra-small resistors and magnetic components, the accuracy of the core loss model has a decisive impact on system performance. With the rapid development of renewable energy, electric vehicles, and high-efficiency power conversion systems, the demand for high-performance magnetic materials and accurate loss prediction models has become unprecedentedly urgent. Accurately predicting core loss can not only improve system efficiency, reduce energy loss, but also extend the equipment life and enhance reliability, which has important economic and social value in practical applications. However, the non-linear characteristics, frequency dependence, temperature sensitivity of core materials, and complex multi-variable influencing factors make loss modeling a very challenging task. Traditional loss models, such as the Steinmetz equation (SE), have been widely used due to their simplicity and computational efficiency. However, SE is only applicable to sinusoidal steady-state conditions, and its accuracy is significantly reduced for non-sinusoidal, pulsed, and high-frequency waveforms commonly found in modern power electronics. For this reason, researchers have made several improvements to SE and proposed models such as modified SE (MSE), generalized SE (GSE), and improved generalized SE (iGSE) in order to expand its scope of application.

[0003] Although a large number of studies have been devoted to the modeling of core loss, in the face of the strong causality and complexity of the excitation density change curve, existing models still lack generality and high precision. We propose a deep learning model of latent space temporal knowledge coupling network (Latended Temporal Yoked Fusion). Taking the inherent attributes and temporal attributes of the core as input features, using the long short-term memory network (LSTM) to process time series data, and combining with an extended knowledge gate mechanism to dynamically fuse the conditional matrix and the temporal feature matrix. Through transfer learning, we improve the generalization ability of the model to ensure high-precision prediction under unknown materials and working conditions. Summary of the Invention

[0004] To achieve the above object, this application provides the following technical solutions:

[0005] According to the first aspect of the present invention, the present invention claims a method for predicting core loss based on a latent space temporal knowledge coupling network, including:

[0006] Collect a sample data set, perform feature category division on the sample data set to obtain inherent attribute features and temporal attribute features;

[0007] Preprocess the inherent attributes using conditional encoding and combine with an extended knowledge gate to obtain an inherent feature representation; preprocess the temporal attributes using data dimensionality reduction and combine with a bidirectional long short-term memory network to obtain a temporal feature representation;

[0008] Concatenate and couple the output data of the extended knowledge gate and the output data of the bidirectional long short-term memory network to construct a core loss prediction model;

[0009] Input the sample data set of the material to be predicted into the core loss prediction model to output the predicted result of the magnetic flux loss of the material to be predicted.

[0010] Further, the acquisition of the sample data set further includes:

[0011] The acquired sample data set is the core loss measurement data of various magnetic materials under different frequency, temperature, and waveform conditions;

[0012] After acquiring the sample data set, it further includes:

[0013] Divide the sample data set into a training set and a test set according to a preset ratio.

[0014] Further, the feature category division of the sample data set to obtain inherent attribute features and temporal attribute features further includes:

[0015] The inherent attribute features at least include:

[0016] Temperature, frequency, core material, excitation waveform;

[0017] The temporal attribute features at least include:

[0018] Magnetic flux density, magnetic field strength.

[0019] Further, the preprocessing of the inherent attributes using conditional encoding and combining with an extended knowledge gate to obtain an inherent feature representation; the preprocessing of the temporal attributes using data dimensionality reduction and combining with a bidirectional long short-term memory network to obtain a temporal feature representation further includes:

[0020] Perform maximum-minimum normalization on the inherent attribute features, and denormalize the core loss after the model prediction to obtain a representation with the same order of magnitude as the original data;

[0021] Adopt continuous conditional encoding and the generator added in the continuous conditional encoding, and use matrix multiplication through the extended knowledge gate to update the knowledge gate parameters when updating the gate at each gradient update to expand the dimension of the continuous conditional encoding result to obtain the output data of the extended knowledge gate;

[0022] The matrix multiplication is the recombination and projection of the input features, and the update gate dynamically adjusts the parameters of the knowledge gate of the model during training.

[0023] Further, preprocessing the inherent attributes using conditional encoding, and combining with an extended knowledge gate to obtain an inherent feature representation; preprocessing the temporal attributes using data dimensionality reduction, and combining with a bidirectional long short-term memory network to obtain a temporal feature representation, further including:

[0024] Using wavelet transform to filter out redundant noise in the temporal attribute features, and checking for the existence of outlier points. If there are any, remove the outliers and fill in the adjacent average values.

[0025] Using latent space dimensionality reduction to convert high-dimensional information into low-dimensional representation, and using a combination of LSTM and BiLSTM to extract the long-term and short-term dependence information of the temporal attribute features, and outputting the data output by the bidirectional long short-term memory network.

[0026] Further, after splicing and coupling the output data of the extended knowledge gate and the output data of the bidirectional long short-term memory network to construct a core loss prediction model, further including:

[0027] Evaluating the effectiveness of the core loss prediction model, and using the training set and the test set to evaluate the effect of the model and test the model results with MAE, RMSE, and R2 as benchmarks to obtain a loss curve.

[0028] Monitoring the changes in the loss curve and test metrics during the training process to ensure that the core loss prediction model does not exhibit overfitting or underfitting phenomena.

[0029] This application relates to the field of electromagnetic control technology, and particularly to a core loss prediction method based on a latent space temporal knowledge coupling network. A sample data set is collected, and the sample data set is divided into feature categories to obtain inherent attribute features and temporal attribute features; the inherent attribute features are preprocessed to obtain the output data of the extended knowledge gate; the temporal attribute features are preprocessed to obtain the output data of the bidirectional long short-term memory network; the output data of the extended knowledge gate and the output data of the bidirectional long short-term memory network are spliced and coupled to construct a core loss prediction model; the sample data set of the material to be predicted is input into the core loss prediction model to output the predicted result of the magnetic flux loss of the material to be predicted. The present invention uses a long short-term memory network to process time series data, and combines with an extended knowledge gate mechanism to dynamically fuse the conditional matrix and the temporal feature matrix, and through transfer learning, improves the generalization ability of the model, ensuring high-precision prediction under unknown materials and working conditions. Description of the Drawings

[0030] Figure 1 It is a working flowchart of the core loss prediction method based on the latent space temporal knowledge coupling network claimed in the embodiment of the present application;

[0031] Figure 2This is the second flowchart of the core loss prediction method based on the latent space temporal knowledge coupling network requested to be protected by the embodiments of this application;

[0032] Figure 3 This is the flowchart of the continuous conditional embedding encoding and continuous conditional encoding generator of the core loss prediction method based on the latent space temporal knowledge coupling network requested to be protected by the embodiments of this application;

[0033] Figure 4 This is the expanded knowledge gate structure diagram of the core loss prediction method based on the latent space temporal knowledge coupling network requested to be protected by the embodiments of this application;

[0034] Figure 5 This is the structure diagram of the temporal attribute block of the core loss prediction method based on the latent space temporal knowledge coupling network requested to be protected by the embodiments of this application;

[0035] Figure 6 This is the LTYF model loss curve diagram of the core loss prediction method based on the latent space temporal knowledge coupling network requested to be protected by the embodiments of this application;

[0036] Figure 7 This is the test index change curve diagram of the core loss prediction method based on the latent space temporal knowledge coupling network requested to be protected by the embodiments of this application;

[0037] Figure 8 This is the fitting effect diagram of the core loss prediction of the core loss prediction method based on the latent space temporal knowledge coupling network requested to be protected by the embodiments of this application. Detailed implementation manners

[0038] Next, the technical solutions in the embodiments of this application will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of this application. Obviously, the described embodiments are only a part of the embodiments of this application, rather than all the embodiments. Based on the embodiments in this application, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of this application.

[0039] The terms "first", "second", and "third" in this application are only used for descriptive purposes and cannot be construed as indicating or implying relative importance or implicitly specifying the quantity of the indicated technical features. Thus, features defined with "first", "second", and "third" may explicitly or implicitly include at least one such feature. In the description of this application, the meaning of "a plurality of" is at least two, such as two, three, etc., unless otherwise specifically defined. In the embodiments of this application, all directional indications (such as up, down, left, right, front, back...) are only used to explain the relative positional relationship and movement conditions between components in a specific posture (as shown in the drawings). If the specific posture changes, the directional indications will also change accordingly. In addition, the terms "comprise" and "have" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the listed steps or units, but optionally also includes unlisted steps or units, or optionally also includes other steps or units inherent to these processes, methods, products, or devices.

[0040] Reference to "embodiments" herein means that a particular feature, structure, or characteristic described in connection with the embodiments can be included in at least one embodiment of the application. The phrase appears in various places in the specification and does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment mutually exclusive of other embodiments. Those skilled in the art will explicitly and implicitly understand that the embodiments described herein can be combined with other embodiments.

[0041] In the prior art, although the improved models have improved the prediction accuracy to a certain extent, in terms of the cross-coupling effect, traditional models usually only consider a single variable, such as frequency or magnetic flux density amplitude, and lack a description of the interaction between multiple factors, making it difficult to accurately reflect the loss characteristics under actual working conditions. These models need to pre-determine the characteristic parameters of the material, such as the Steinmetz coefficient. However, these parameters can usually only be obtained through experimental measurement, which limits the generality and adaptability of the models. Under non-sinusoidal or complex waveform conditions, the accuracy of traditional models drops significantly, unable to meet the requirements of modern power electronics applications. With the development of power electronics technology, the working conditions of magnetic components are becoming increasingly complex, and traditional models are difficult to adjust and adapt in a timely manner, restricting their application in emerging fields. To solve the above problems, researchers have started to explore data-driven methods. Artificial neural networks (ANNs), due to their powerful non-linear mapping ability and adaptive learning characteristics, have become a powerful tool for magnetic core loss modeling. In early work, Kucuk used neural networks and genetic algorithms to predict the magnetic hysteresis loop of the magnetic core; Amoiralis et al. combined artificial intelligence technology with a hybrid finite element-boundary element method to achieve the global optimization design of transformers.

[0042] According to the first embodiment of the present invention, the present invention claims a method for predicting core loss based on a latent space time-series knowledge coupling network. Referring to Figure 1 , it includes:

[0043] Collect a sample data set, classify the feature categories of the sample data set, and obtain inherent attribute features and time-series attribute features;

[0044] Preprocess the inherent attributes using conditional encoding, and combine with an extended knowledge gate to obtain inherent feature representations; preprocess the time-series attributes using data dimensionality reduction, and combine with a bidirectional long short-term memory network to obtain time-series feature representations;

[0045] Stitch and couple the output data of the extended knowledge gate and the output data of the bidirectional long short-term memory network to construct a core loss prediction model;

[0046] Input the sample data set of the material to be predicted into the core loss prediction model to output the predicted flux loss result of the material to be predicted.

[0047] Furthermore, the collection of the sample data set further includes:

[0048] The collected sample data set is the core loss measurement data of various magnetic materials under different frequency, temperature, and waveform conditions;

[0049] After collecting the sample data set, it further includes:

[0050] Divide the sample data set into a training set and a test set according to a preset ratio.

[0051] Among them, in this embodiment, the data is from the MagNet Database dataset, which is used to support the proposed core loss modeling method based on LTYF-LSTM. MagNet is a large open-source magnetic material database provided by the MagNet Challenge project in 2023. The core loss is measured by the four-wire voltage / current method, covering a variety of ferrite materials (such as Ferroxcube-3C90, 3C94, and N27, N30 of TDK, etc.). The dataset includes measurement data in the frequency range from 50 kHz to 450 kHz, the temperature range from 0 °C to 120 °C, and sine, triangle, and trapezoidal waveform conditions from 10 mT to 300 mT, which is used to verify the wide adaptability and robustness of the proposed model. In this study, material N87 was selected to verify the proposed data-driven core loss modeling method. The remaining materials were used for further in-depth analysis of the generalization and robustness of the model. Before the training process, the dataset of material N87 was randomly divided into a training set and a test set according to a ratio of 8:2.

[0052] The attribute records of workpiece samples in the dataset can be divided into inherent attributes and time-series attributes. We list each type of attribute record in Table 1. Among them, the frequency, temperature, magnetic flux density, and core loss attributes are subjected to max-min normalization to ensure that the overall values are not too large. After the model prediction is completed, the core loss is denormalized to obtain a representation with the same order of magnitude as the original data.

[0053] Table 1 Classification Table of Attribute Feature Categories of Workpiece Samples

[0054]

[0055] Furthermore, the feature category division of the sample dataset to obtain inherent attribute features and time-series attribute features further includes:

[0056] The inherent attribute features at least include:

[0057] Temperature, frequency, core material, excitation waveform;

[0058] The time-series attribute features at least include:

[0059] Magnetic flux density, magnetic field strength.

[0060] Among them, in this embodiment, for the inherent attributes, in the preprocessing part, we choose to adopt the continuous conditional coding method to avoid the masking phenomenon between feature data, ensure that each feature can effectively participate in model training, and at the same time use the generator added in the continuous conditional coding to ensure that the coding results have both strong robustness and diversity.

[0061] For the time-series attributes, wavelet transform is used to filter out redundant noise, and whether there are outlier points is checked. If there are, the outliers are removed and filled with the adjacent average value.

[0062] Existing industry models are required to meet two points, namely high-precision prediction and wide applicability.

[0063] For wide applicability, we propose continuous conditional coding. By using the continuous conditional coding generator to maintain the robustness of the coding results, the model can also perform well when predicting the attributes of unknown materials. At the same time, the regularization method is used in the model to avoid overfitting of the model to the existing data, and the idea of transfer learning is introduced to ensure good performance of the model under unknown conditional attributes.

[0064] For high precision, we use LSTM and BiLSTM to form a deeper network structure, giving full play to the ability of these two models to capture bidirectional dependencies in sequence data. At the same time, the continuous conditional coding results are processed using the extended knowledge gate, combined with the latent space data dimensionality reduction method, to expand the dimension of the high-complexity high-dimensional magnetic flux density, facilitating model learning and fitting results.

[0065] Therefore, we designed the model as Figure 2 shown. It mainly consists of two parts, namely the inherent attribute block and the temporal attribute block. For the inherent attributes in the features, continuous conditional encoding is used to process its discrete features, and extended knowledge gates are used to process the inherent attribute features. For the temporal features, the idea of latent space dimensionality reduction is used to convert high-dimensional information into low-dimensional representations. LSTM and BiLSTM are combined to extract long-term and short-term dependence information. Finally, a coupling network is used to integrate and output the inherent attribute features and the temporal attribute features to obtain the predicted value of the magnetic flux density.

[0066] Furthermore, preprocessing the inherent attributes using conditional encoding and combining extended knowledge gates to obtain the inherent feature representation; preprocessing the temporal attributes using data dimensionality reduction and combining bidirectional long short-term memory networks to obtain the temporal feature representation, further includes:

[0067] Performing max-min normalization on the inherent attribute features, and after the model prediction is completed, anti-normalizing the core loss to obtain a representation with the same order of magnitude as the original data.

[0068] Using continuous conditional encoding and the generator added in the continuous conditional encoding, and through the extended knowledge gate using matrix multiplication, updating the knowledge gate parameters during each gradient update to expand the dimension of the continuous conditional encoding result to obtain the output data of the extended knowledge gate.

[0069] The matrix multiplication is the recombination and projection of the input features, and the update gate dynamically adjusts the parameters of the knowledge gate of the model during training.

[0070] Among them, in this embodiment, the continuous conditional encoding structure is as Figure 3 shown.

[0071] The continuous conditional encoding generator is an embedding encoding method for multi-condition and multi-dimensional feature sequences. It realizes efficient encoding and accurate modeling of complex data by fusing conditional information, position information, and semantic features. The continuous conditional encoding generator is based on the embedding design of conditional attribute features and variable attribute features, and uses to map to a more flexible representation space, enabling it to adapt to various conditional inputs and context environments when dealing with complex tasks, thereby enhancing the expression ability of the model.

[0072] The conditional encoding block first introduces a conditional and continuous variable embedding module as the first part of the encoding, embedding continuous variables such as waveform, material, temperature, and frequency into a high-dimensional space as prior information of the input features. Its output encoding result, as part of the main body of the conditional encoding, is part of the output encoding result, and its result will be input to the continuous conditional encoding generator module. Assuming the input condition is , where represents the A conditional variable, through a continuous embedding function , these conditions are mapped to an embedding space:

[0073] ;

[0074] Here, represents the encoding process, is the dimension of the embedding vector. The conditional embedding captures the high-dimensional correlations between external conditional variables and provides a basis for subsequent feature generation.

[0075] Subsequently, is introduced into the continuous conditional encoding generator to ensure that the model can capture the temporal characteristics of sequence data. The positional encoding is based on the embedding dimension transformation and is constructed in the form of sine and cosine functions. For the -th position in the sequence, its positional encoding is defined as:

[0076] ;

[0077] ;

[0078] Among them, is the frequency dimension index in the positional encoding, represents the position index in the sequence, is a non-linear transformation function implemented by a set of linear layer modules. Therefore, compared with the conventional positional encoding, it combines the advantages of conventional encoding and positional encoding, retains the global position information, and also retains the continuous feature variables.

[0079] Secondly, compared with the conventional positional encoding, another difference is the way it retains the temporal characteristics. The continuous conditional encoding uses the number of time steps as the number of encoding layers. Let the number of time steps be , then the continuous conditional encoding will encode layers, and the encoding results are accumulated and weighted. The encoding result of the current time step is set as , then it is obtained by adding the assumed previous layer feature encoding and the current layer feature encoding. Its calculation process is

[0080] ;

[0081] Through this dynamic weighting mechanism, the continuous conditional encoding generator can capture the joint influence of conditional variables and position information on feature generation.

[0082] It can effectively ensure that the model maximizes the processing and utilization of discrete features. The temperature, core material and excitation waveform recorded in the data set are all classic discrete features. Conventional encoding processing methods can only convert attributes from text to digital form to participate in training, and the conversion results are single and not flexible. When unknown temperatures, waveforms or materials appear, conventional encoding is likely to fail. Therefore, we use the encoding method of inherent attribute combinations and the conditional coding generator to increase the diversity of encoding results while ensuring high versatility. When the model is applied to unknown attributes, it can still complete the prediction task well.

[0083] Expand your knowledge Figure 4 As shown in , it further expands the dimension of the continuous conditional encoding result to obtain a more diverse result representation. It mainly uses matrix multiplication to expand the dimension, and uses the update gate to update the knowledge gate parameters at each gradient update, thereby improving the model's adaptability to data in different scenarios. The process of matrix multiplication can be regarded as a reorganization and projection of the input features, allowing the model to capture more complex feature interactions in high-dimensional space. The introduction of the update gate enables the model to dynamically adjust the parameters of the knowledge gate during the training process, so that it can better adapt to changes in input data and avoid overfitting problems. In this way, the model not only performs well on the original data distribution, but can also be flexibly applied under different data modes, showing stronger generalization ability and robustness. The extended knowledge gate output and update gate iteration formulas are shown in Equations 1 to 3.

[0084] ;

[0085] in, It is the current moment The output result for the input sample is is a weight matrix that expands the input dimension and is used to capture the complex relationship between input features. is the input intrinsic attribute condition matrix at the current moment, is the bias vector used to adjust the output result of each sample. is the output of the update gate, is the Sigmoid activation function, which is used to control the update intensity of the knowledge base. is the weight matrix, which acts on the current input , is the bias vector, is the updated knowledge base weight matrix, is the knowledge base parameter at the current moment, is the initial knowledge base weight, is the Hadamard product, which represents element-wise multiplication.

[0086] Further, preprocess the inherent attributes using conditional encoding, and combine with an extended knowledge gate to obtain an inherent feature representation; preprocess the temporal attributes using data dimensionality reduction, and combine with a bidirectional long short-term memory network to obtain a temporal feature representation. It further includes:

[0087] Use wavelet transform to filter out the redundant noise of the temporal attribute features, and check whether there are outlier points. If there are, remove the outliers and fill in the adjacent average value;

[0088] Use latent space dimensionality reduction to convert high-dimensional information into low-dimensional representation, and use the combination of LSTM and BiLSTM to extract the long-term and short-term dependence information of the temporal attribute features, and output the data output by the bidirectional long short-term memory network.

[0089] Among them, in this embodiment, as Figure 5 shown, the temporal attribute block is mainly composed of the combination of LSTM and BiLSTM, and extracts features from the front and back directions of the time series. The bidirectional structure allows the model to not only capture historical information, but also enhance the feature representation by supplementing future information, thereby improving the decision-making ability for the current time step.

[0090] Both LSTM and BiLSTM are controlled by three gating mechanisms and a memory unit, namely the forget gate, input gate, output gate, and update memory unit. Their calculation formulas are as follows.

[0091] ;

[0092] Among them, is the output of the memory gate, is the hidden state of the previous time step, is the current input, is the weight matrix, is the bias term, is the sigmoid function.

[0093] ;

[0094] ;

[0095] Among them, is the output of the input gate, and are the input gate weight matrix and bias term, is the candidate new memory, and tanh is the hyperbolic tangent activation function.

[0096] ;

[0097] Among them, is the memory unit at the current moment, is the memory of the previous moment.

[0098] ;

[0099] ;

[0100] Among them, is the hidden state at the current moment.

[0101] Finally, the coupling network integrates the inherent attribute matrix and the temporal attribute matrix data, reduces the data dimension using a linear layer, and the model combines the inherent attributes and the temporal attributes to comprehensively judge the result and outputs the magnetic flux loss prediction result.

[0102] Furthermore, after splicing and coupling the output data of the extended knowledge gate and the output data of the bidirectional long short-term memory network to construct the core loss prediction model, it further includes:

[0103] Evaluate the effectiveness of the core loss prediction model, and use the training set and the test set to evaluate the model effect and test the model result with MAE, RMSE, and R2 as benchmarks to obtain the loss curve;

[0104] Monitor the changes of the loss curve and test metrics during the training process to ensure that the core loss prediction model does not have overfitting or underfitting phenomena.

[0105] Among them, in this embodiment, the hyperparameter design is shown in Table 2, and the test model result, and its loss curve is as Figure 6 shown. Generally speaking, the losses of the training set and the test set, and both curves show a rapid downward trend, indicating that the model learns quickly in this stage and quickly adjusts the parameters in the first few iterations, reducing the error. After 50 iterations, the loss value basically stabilizes, and the loss curves of the training set and the test set are relatively close, indicating that the learning effect of the model is good and there is no significant overfitting or underfitting phenomenon. Generally speaking, this loss curve shows that the model has been reasonably trained and finally achieved a stable effect. The loss values of the training set and the test set are consistent under a long number of iterations, indicating that the generalization ability of the model is good and it is applicable to data outside the training set. This trend conforms to the characteristic that the loss value gradually converges with the passage of training time in the deep learning model, demonstrating the convergence and stability of the model.

[0106] Table 2 Hyperparameter Statistical Table

[0107]

[0108] As Figure 7 shown, we also plotted the change curves of the test metrics during the training process, including MAE, RMSE, Regarding the index changes, generally, the curve shows a steady downward trend, indicating that the overall performance of the model is gradually improving. The curve rapidly approaches above 0.95 from a relatively low level as the training progresses. It gradually tends to be stable and approaches 100%. This curve trend indicates that the fitting effect of the model is very good and it can accurately predict the core loss value.

[0109] Furthermore, we randomly selected 80 samples for core loss testing and compared the predicted values with the actual values of each sample to intuitively feel the fitting effect of the core loss. The results are as Figure 8 shown. To observe the overall fitting effect, we chose the horizontal axis as the sample points and used a line chart to observe the overall trend. Generally speaking, the prediction results of the model show a high degree of consistency with the true values of the core loss. Especially in multiple loss peak regions, the model's prediction can well capture the actual trend. Due to the emphasis on generalization in the model, the model has poor fitting ability for extreme cases, while it has a better fitting effect for more conventional conditions and shows good fitting ability in most core loss predictions.

[0110] Further, we utilized the idea of transfer learning and selected the material category 3:1 division method. We took one of the four materials (3C90, 3C94, 3E6, 3F4 corresponding to material one to four respectively) as the test set in turn, and the other three as the training set to observe the model's generalization ability when testing unknown materials. By continuously rotating the materials as the test set, we observed the transfer performance of the model among different materials from various angles to provide data support for the subsequent optimization of the model and the improvement of its generalization ability.

[0111] We listed the test results of the four combination methods in Table 3. We also selected MAE, RMSE, and R2 as quantitative evaluation indicators to judge its generalization ability. At the same time, we randomly selected samples and plotted the model fitting effect. There are certain differences in the test results of different material combinations, which reflects the differences in material characteristics and the transfer ability of the model when facing different materials. Among all the division methods, the change ranges of MAE and RMSE are relatively small, and the R2 values all remain at a level close to 100%, indicating that the model has good generalization performance when dealing with different materials. However, in some combinations, the error of the test set is slightly higher, indicating that there is still room for further optimization on specific materials.

[0112] Table 3 Core Loss Prediction Fitting Transfer Learning Test Table

[0113]

[0114] Furthermore, we modified the idea of transfer learning for waveform and temperature factors, and also conducted experiments on predicting and fitting core losses. A total of seven sets of test prediction results were listed. In each set, one result was used as the test set, and the remaining conditional factors were used as the training set, visually demonstrating the generalization ability of the model from various perspectives.

[0115] Therefore, the LTYF-LSTM designed by us is a core loss prediction model that can span different material types and working conditions. Its continuous conditional encoding has strong versatility, enabling the model to accurately predict core losses even under unknown working conditions or materials, greatly improving the accuracy and efficiency of magnetic component design.

[0116] To further verify the superiority of the model, the present invention examines the generalization ability of the model when facing unknown complex waveforms. The training set is restricted to be composed of sine and triangular flux density B(t) waveforms, while all the remaining trapezoidal waveforms are reserved for the test set to evaluate the error. Table 4 compares the prediction errors of each model for trapezoidal waveforms in N87 material under the above constraints. The results show that the method based on LTYF-LSTM can be generalized to unknown waveforms.

[0117] Table 4 Prediction error information of different models for unknown waveforms

[0118]

[0119] In several embodiments provided in the present application, it should be understood that the disclosed systems, devices, and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of units is only a logical function division, and there may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the displayed or discussed couplings or direct couplings or communication connections to each other can be through some interfaces, and the indirect couplings or communication connections of devices or units can be in electrical, mechanical, or other forms.

[0120] In addition, each functional unit in various embodiments of the present application can be integrated in a processing unit, or each unit can exist physically alone, or two or more units can be integrated in one unit. The above integrated units can be implemented in the form of hardware or in the form of software functional units. The above is only the implementation mode of the present application, and does not limit the patent scope of the present application. Any equivalent structure or equivalent process transformation made using the specification and drawings of the present application, or directly or indirectly applied in other related technical fields, is equally included in the patent protection scope of the present application.

[0121] The specific embodiments of the invention have been described in detail above, but they are only examples, and the present application is not limited to the specific embodiments described above. For those skilled in the art, any equivalent modifications or substitutions to the invention are also within the scope of the present application. Therefore, all equivalent transformations, modifications, improvements, etc. made without departing from the spirit and principles of the present application should be covered within the scope of the present application.

Claims

1. A method for predicting core loss based on a latent space temporal knowledge coupling network, characterized in that Including: Collect a sample data set, perform feature category division on the sample data set to obtain inherent attribute features and time series attribute features; Preprocess the inherent attributes using conditional encoding, and combine with an extended knowledge gate to obtain inherent feature representations; Preprocess the time series attributes using data dimensionality reduction, and combine with a bidirectional long short-term memory network to obtain time series feature representations; Stitch and couple the output data of the extended knowledge gate and the output data of the bidirectional long short-term memory network to construct a core loss prediction model; Input the sample data set of the material to be predicted into the core loss prediction model to output the predicted flux loss result of the material to be predicted.

2. The magnetic core loss prediction method based on the latent space time series knowledge coupling network according to claim 1, wherein The collection of the sample data set further includes: The collected sample data set is the core loss measurement data of various magnetic materials under different frequencies, temperatures, and waveform conditions; After collecting the sample data set, it further includes: Divide the sample data set into a training set and a test set according to a preset ratio.

3. The magnetic core loss prediction method based on the latent space time series knowledge coupling network according to claim 1, characterized in that The feature category division of the sample data set to obtain inherent attribute features and time series attribute features further includes: The inherent attribute features at least include: Temperature, frequency, core material, excitation waveform; The time series attribute features at least include: Magnetic flux density, magnetic field strength.

4. The magnetic core loss prediction method based on the latent space time-series knowledge coupling network according to claim 1, wherein, Preprocess the inherent attributes using conditional encoding, and combine with an extended knowledge gate to obtain inherent feature representations; Preprocess the time series attributes using data dimensionality reduction, and combine with a bidirectional long short-term memory network to obtain time series feature representations, further includes: Perform maximum-minimum normalization on the inherent attribute features, and denormalize the core loss after the model prediction to obtain a representation with the same order of magnitude as the original data; Adopt continuous conditional encoding and the generator added in the continuous conditional encoding, use matrix multiplication through the extended knowledge gate, and update the knowledge gate parameters when updating the gate each time to expand the dimension of the continuous conditional encoding result to obtain the output data of the extended knowledge gate; The matrix multiplication is the recombination and projection of the input features, and the update gate dynamically adjusts the parameters of the knowledge gate of the model during training.

5. The magnetic core loss prediction method based on the latent space temporal knowledge coupling network according to claim 1, wherein Preprocess the inherent attributes using conditional encoding, and combine with an extended knowledge gate to obtain inherent feature representations; Preprocess the time series attributes using data dimensionality reduction, and combine with a bidirectional long short-term memory network to obtain time series feature representations, further includes: Use wavelet transform to filter out the redundant noise of the time series attribute features, and check whether there are outlier points. If so, remove the outliers and fill in the adjacent average values; Adopt latent space dimensionality reduction to convert high-dimensional information into low-dimensional representations, and use the combination of LSTM and BiLSTM to extract the long-term and short-term dependence information of the time series attribute features, and output the output data of the bidirectional long short-term memory network.

6. The magnetic core loss prediction method based on the latent space temporal knowledge coupling network according to claim 1, wherein After stitching and coupling the output data of the extended knowledge gate and the output data of the bidirectional long short-term memory network to construct a core loss prediction model, it further includes: Evaluate the effectiveness of the core loss prediction model, and use MAE, RMSE, and R2 as benchmarks for the training set and the test set to evaluate the effect of the model and test the model results to obtain a loss curve; Monitor the changes of the loss curve and test metrics during training to ensure that the core loss prediction model does not exhibit overfitting or underfitting phenomena.

Images