A method for predicting magnetic core loss based on causal representation learning
By employing a causal representation learning method, and utilizing the causal relationship between magnetic flux density and material type, implicit causal representation variables are extracted. This addresses the issues of insufficient accuracy and inadequate generalization ability of existing magnetic core loss models under high frequency and high power density conditions, thus achieving high-precision and widely applicable magnetic core loss prediction.
Patent Information
- Application Number
- CN202411735630.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-29
- Publication Date
- 2025-10-31
- Estimated Expiration
- 2044-11-29
AI Technical Summary
Existing core loss models struggle to provide high-precision predictions in high-frequency and high-power-density applications, and traditional machine learning models have limited generalization ability when dealing with material category variables, making them unable to effectively handle unseen new materials.
The causal representation learning method is adopted. By acquiring the magnetic core dataset and dividing it into training and test datasets according to material categories, a latent causal representation extraction model and a magnetic core loss prediction model are constructed. Using magnetic flux density and material category as inputs, latent causal representation variables are extracted and iteratively trained to finally achieve the prediction of magnetic core loss.
It improves the accuracy and efficiency of core loss prediction, enhances the model's generalization ability, can effectively handle new materials that have not been seen before, and is suitable for diverse working conditions.
Smart Images
Figure CN119673341B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of machine learning technology and manufacturing, and more specifically, to a method for predicting magnetic core loss based on causal representation learning. Background Technology
[0002] Magnetic core loss is a critical factor in manufacturing, affecting product performance, efficiency, and reliability. With the continuous advancement of manufacturing, the industry demands higher accuracy and applicability from magnetic core loss models, especially in high-frequency and high-power-density applications, where precise loss models become particularly important. However, existing magnetic core material loss models still fall short of practical application requirements and struggle to comprehensively cover diverse operating conditions.
[0003] In the field of core loss research, numerous core loss models exist. For example, mathematical models such as the Priscilla model and the Giles-Atherton model provide detailed descriptions of hysteresis phenomena and can simulate complex magnetization processes and secondary loops. However, these models consume significant time and computational resources to solve complex nonlinear equations or process large amounts of measurement data. Time-domain approximation (TDA) models can handle sinusoidal and non-sinusoidal flux densities and are suitable for linear systems, but this method is only effective for linear systems, and its accuracy for PWM signals is limited at frequencies above 400Hz. Loss separation models (LSM) can decompose core losses into hysteresis losses, eddy current losses, and anomalous losses, making them suitable for various applications. Although LSM can calculate core losses under DC-biased square waves, this model only provides average information and cannot calculate core losses under harmonic excitation. Empirical models such as the Steinmetz equation and its variants (including the modified Steinmetz equation, the generalized Steinmetz equation, and the improved generalized Steinmetz equation) can provide rapid estimates of core losses based on experimental data. However, these models typically assume sinusoidal excitation and do not consider DC bias, resulting in limited accuracy for non-sinusoidal excitation. Therefore, the industry currently lacks a core loss model that is both widely applicable and provides high-precision prediction results.
[0004] Neural networks can fit the effects of different operating conditions and core materials on core loss. However, due to the lack of interpretability of deep neural networks, it is often difficult to interpret and generalize trained models. This is especially true when considering the material category feature, where the generalization ability of traditional machine learning models is limited. Material category, as an unordered categorical variable, differs from continuous features such as temperature, making it difficult to find a function to simulate its impact on core loss. Whenever a new material is input into the model, the model faces a completely new material category variable orthogonal to previous material features. This makes it difficult to quantify and generalize such categorical features, failing to meet the generalization requirements of the current problem. Current core loss model construction faces two major problems: 1) how core materials affect core loss; and 2) how to ensure reliable prediction results for core materials that the model has not been trained on. Summary of the Invention
[0005] This invention provides a core loss prediction method based on causal representation learning to improve the accuracy and efficiency of core loss prediction.
[0006] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:
[0007] This invention proposes a core loss prediction method based on causal representation learning, comprising the following steps:
[0008] S1: Obtain the magnetic core dataset of magnetic components, and divide the magnetic core dataset into test dataset and training dataset according to the category of magnetic core material, and construct the initial magnetic core loss prediction model and the initial implicit causal characterization extraction model.
[0009] S2: Input the training dataset into the initial latent causal representation extraction model, set the first loss function for iterative training, and when the preset number of iterations is reached, obtain the trained latent causal representation extraction model. The trained latent causal representation extraction model outputs the latent causal representation variables corresponding to the training dataset.
[0010] S3: Integrate the training dataset and the corresponding latent causal representation variables to obtain the integrated training dataset;
[0011] S4: Input the integrated training dataset into the initial core loss prediction model, set the second loss function for iterative training, and obtain the trained core loss prediction model when the preset number of iterations is reached.
[0012] S5: Input the test dataset into the trained latent causal representation extraction model and the trained core loss prediction model for testing, and obtain the final latent causal representation extraction model and the final core loss prediction model.
[0013] S6: Obtain the core data of the magnetic element to be predicted, input it into the final implicit causal characterization extraction model, and obtain the implicit causal characterization variables corresponding to the core data of the magnetic element to be predicted; input the core data of the magnetic element to be predicted and its corresponding implicit causal characterization variables into the final core loss prediction model to obtain the predicted value of the core loss of the magnetic element to be predicted.
[0014] Preferably, in step S1, the magnetic core dataset includes magnetic core material M, magnetic flux density B(t), temperature T, frequency F, waveform W, and magnetic core loss Y.
[0015] Preferably, in step S1, obtaining the magnetic core dataset of the magnetic element and dividing the magnetic core dataset into a test dataset and a training dataset according to the category of magnetic core material includes dividing according to the category of magnetic core material, and sequentially using the magnetic core data of the same magnetic core material as the test dataset and the magnetic core data of other magnetic core materials as the training dataset.
[0016] Preferably, step S2 includes:
[0017] S2.1: Construct an initial latent causal representation extraction model, including an encoder and a decoder;
[0018] S2.2: Input the magnetic flux density and core material in the training dataset into the encoder of the initial latent causal representation extraction model in batches to obtain latent causal representation variables. Input the latent causal representation variables and core material into the decoder of the initial latent causal representation extraction model to obtain the reconstructed magnetic flux density.
[0019] S2.3: Construct a first loss function based on magnetic flux density and reconstructed magnetic flux density, and train the initial latent causal representation extraction model until the preset number of iterations is reached to obtain the trained latent causal representation extraction model.
[0020] Preferably, the first loss function in step S2.3 includes the KL divergence between the reconstruction error and the standard prior distribution and the approximate posterior distribution of the latent causal representation variable, specifically:
[0021] ELBO = L1 - L2
[0022] Where ELBO represents the first loss function, L1 represents the reconstruction error, and L2 represents the KL divergence between the standard prior and approximate posterior of the latent causal representation variable.
[0023] Preferably, the method for determining the reconstruction error includes:
[0024] L1 = E q(z|y,m) [logp(y|z)]
[0025] Where L1 represents the reconstruction error, q(z|y,m) is the approximate posterior distribution, and p(y|z) is the conditional prior distribution.
[0026] Preferably, the method for determining the KL divergence between the standard prior distribution and the approximate posterior distribution of the latent causal characterization variable includes:
[0027] L2=KL(p(z|m)||q(z|y,m))
[0028] Where L2 represents the KL divergence between the standard prior distribution and the approximate posterior distribution of the latent causal characterization variable, and KL(·) represents the discreteness function.
[0029] Preferably, step S4 includes:
[0030] S4.1: Input the integrated training dataset into the initial core loss prediction model in batches to obtain the core loss prediction value;
[0031] S4.2: Construct a second loss function based on the predicted core loss value and the actual core loss value, and train the initial core loss prediction model until the preset number of iterations is reached to obtain the trained core loss prediction model.
[0032] Preferably, the method for determining the second loss function includes:
[0033]
[0034] Where L represents the second loss function, k represents the number of training datasets, and y i This represents the true value of the core loss for training data i. This represents the predicted core loss value for training data i.
[0035] This invention proposes a magnetic core loss prediction system based on causal representation learning, used to implement the aforementioned magnetic core loss prediction method based on causal representation learning, comprising:
[0036] The data partitioning and initial model building module is used to obtain the magnetic core dataset of magnetic components, and divide the magnetic core dataset into test dataset and training dataset according to the category of magnetic core material, and build the initial magnetic core loss prediction model and the initial implicit causal characterization extraction model.
[0037] The latent causal representation variable acquisition module is used to input the training dataset into the initial latent causal representation extraction model, set the first loss function for iterative training, and obtain the trained latent causal representation extraction model when the preset number of iterations is reached. The trained latent causal representation extraction model outputs the latent causal representation variables corresponding to the training dataset.
[0038] The data integration module is used to integrate the training dataset and the corresponding latent causal representation variables to obtain an integrated training dataset;
[0039] The core loss prediction model training module is used to input the integrated training dataset into the initial core loss prediction model, set the second loss function for iterative training, and obtain the trained core loss prediction model when the preset number of iterations is reached.
[0040] The core loss prediction model testing module is used to input the test dataset into the trained latent causal representation extraction model and the trained core loss prediction model for testing, and to obtain the final latent causal representation extraction model and the final core loss prediction model.
[0041] The prediction module is used to acquire the core data of the magnetic component to be predicted, input it into the final implicit causal characterization extraction model, and obtain the implicit causal characterization variables corresponding to the core data of the magnetic component to be predicted; input the core data of the magnetic component to be predicted and its corresponding implicit causal characterization variables into the final core loss prediction model to obtain the predicted value of the core loss of the magnetic component to be predicted.
[0042] Compared with the prior art, the beneficial effects of the technical solution of the present invention are:
[0043] This invention proposes a magnetic core loss prediction method based on causal representation learning. Magnetic flux density is used as input, and material type as an auxiliary variable. These are input into a latent causal representation extraction model to extract latent causal representation variables. Then, the learned latent causal representation variables, along with features such as magnetic flux density, temperature, frequency, and waveform, are input into the magnetic core loss prediction model to predict core loss. This invention improves the accuracy and efficiency of magnetic core loss prediction and enhances the model's generalization ability. Attached Figure Description
[0044] Figure 1 This is a flowchart illustrating the core loss prediction method based on causal representation learning described in Example 1.
[0045] Figure 2 This is a schematic diagram of the fully connected neural network structure described in Example 2;
[0046] Figure 3 This is a schematic diagram of the core loss prediction method based on causal representation learning described in Example 3. Detailed Implementation
[0047] The accompanying drawings are for illustrative purposes only and should not be construed as limiting the scope of this patent.
[0048] To better illustrate this embodiment, some parts in the accompanying drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions;
[0049] It will be understood by those skilled in the art that certain well-known structures and their descriptions may be omitted in the accompanying drawings.
[0050] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.
[0051] Example 1
[0052] This embodiment provides a core loss prediction method based on causal representation learning, such as... Figure 1 As shown, it includes the following steps:
[0053] S1: Obtain the magnetic core dataset of magnetic components, and divide the magnetic core dataset into test dataset and training dataset according to the category of magnetic core material, and construct the initial magnetic core loss prediction model and the initial implicit causal characterization extraction model.
[0054] S2: Input the training dataset into the initial latent causal representation extraction model, set the first loss function for iterative training, and when the preset number of iterations is reached, obtain the trained latent causal representation extraction model. The trained latent causal representation extraction model outputs the latent causal representation variables corresponding to the training dataset.
[0055] S3: Integrate the training dataset and the corresponding latent causal representation variables to obtain the integrated training dataset;
[0056] S4: Input the integrated training dataset into the initial core loss prediction model, set the second loss function for iterative training, and obtain the trained core loss prediction model when the preset number of iterations is reached.
[0057] S5: Input the test dataset into the trained latent causal representation extraction model and the trained core loss prediction model for testing, and obtain the final latent causal representation extraction model and the final core loss prediction model.
[0058] S6: Obtain the core data of the magnetic element to be predicted, input it into the final implicit causal characterization extraction model, and obtain the implicit causal characterization variables corresponding to the core data of the magnetic element to be predicted; input the core data of the magnetic element to be predicted and its corresponding implicit causal characterization variables into the final core loss prediction model to obtain the predicted value of the core loss of the magnetic element to be predicted.
[0059] In the specific implementation process, the magnetic core dataset is first divided into a test dataset and a training dataset according to the category of the magnetic core material. One category is fixed as the test set, and the remaining categories are used as the training set. Next, a latent causal representation extraction model is trained, and the latent causal representation variables corresponding to the model output training dataset are extracted. The training dataset and the corresponding latent causal representation variables are then integrated. The integrated training dataset is then input into the initial magnetic core loss prediction model, and a second loss function is set for iterative training. When a preset number of iterations is reached, the trained magnetic core loss prediction model is obtained. Finally, the test dataset is input into the trained latent causal representation extraction model and the trained magnetic core loss prediction model for testing, resulting in the final latent causal representation extraction model and the final magnetic core loss prediction model.
[0060] Example 2
[0061] This embodiment provides a core loss prediction method based on causal representation learning, including the following steps:
[0062] S1: Obtain the magnetic core dataset of magnetic components, and divide the magnetic core dataset into test dataset and training dataset according to the category of magnetic core material, and construct the initial magnetic core loss prediction model and the initial implicit causal characterization extraction model.
[0063] The magnetic core dataset includes core material M, magnetic flux density B(t), temperature T, frequency F, waveform W, and core loss Y. Core material M and waveform W are discrete variables. Core material M includes four possible values: material 1 (M1), material 2 (M2), material 3 (M3), and material 4 (M4). Waveform W includes three possible values: sine wave, trapezoidal wave, and triangular wave. Magnetic flux density B(t), temperature T, frequency F, and core loss Y are continuous variables, with core loss Y being the target variable.
[0064] This embodiment analyzes the potential relationships between categorical variables and other variables from the perspective of causal representation learning. Different materials have different physical properties, such as area, shape, and magnetic properties. These physical properties affect magnetic flux density, which in turn affects core loss. Therefore, the physical properties of these core materials can be considered as latent causal representation variables between core material category and magnetic flux density, resulting in different effects of magnetic flux density on core loss under different core materials. Based on the category of core material, core data of the same core material are used as the test dataset, and core data of other core materials are used as the training dataset.
[0065] An initial core loss prediction model and an initial latent causal characterization extraction model are constructed. Both the initial core loss prediction model and the initial latent causal characterization extraction model employ fully connected neural networks, such as... Figure 2The diagram shows a fully connected neural network structure, including an input layer, a first hidden layer, a first ReLU activation function, a second hidden layer, a second ReLU activation function, a third hidden layer, a third ReLU activation function, and an output layer connected in sequence. The parameters of the fully connected neural network are set, including the number of network layers, activation functions, number of iterations, and training batch size. In this embodiment, the number of iterations is set to 100, and the training batch size is set to 64.
[0066] S2: Input the training dataset into the initial latent causal representation extraction model, set the first loss function for iterative training, and when the preset number of iterations is reached, obtain the trained latent causal representation extraction model. The trained latent causal representation extraction model outputs the latent causal representation variables corresponding to the training dataset.
[0067] S2.1: Construct an initial latent causal representation extraction model, including an encoder and a decoder;
[0068] S2.2: Input the magnetic flux density and core material in the training dataset into the encoder of the initial latent causal representation extraction model in batches to obtain latent causal representation variables. Input the latent causal representation variables and core material into the decoder of the initial latent causal representation extraction model to obtain the reconstructed magnetic flux density.
[0069] S2.3: Construct a first loss function based on magnetic flux density and reconstructed magnetic flux density, and train the initial latent causal representation extraction model until the preset number of iterations is reached to obtain the trained latent causal representation extraction model.
[0070] In this embodiment, the construction parameters of the encoder and decoder are determined by the parameters of the fully connected neural network. The training generates a model that includes the encoding process and the reconstruction process during decoding. The magnetic flux density b(t) is used as the input to the encoder, and the magnetic core material is used as the input supervision signal m. The output latent causal characterization variable z has the following form:
[0071] z = f enc (b(t),m)
[0072] The decoder outputs the reconstructed result based on the supervision signal m and the latent causal representation variable z from the encoder output. Its form is:
[0073]
[0074] The first loss function includes the KL divergence between the reconstruction error and the standard prior and approximate posterior distributions of the latent causal representation variables, specifically:
[0075] ELBO = L1 - L2
[0076] Where ELBO represents the first loss function, L1 represents the reconstruction error, and L2 represents the KL divergence between the standard prior and approximate posterior of the latent causal representation variable.
[0077] Among them, the reconstruction error measurement model is based on the given implicit causal representation variables. Reconstructing input data The specific capabilities are:
[0078] L1 = E q(z|y,m) [logp(y|z)]
[0079] Where L1 represents the reconstruction error, q(z|y,m) is the approximate posterior distribution, and p(y|z) is the conditional prior distribution. The KL divergence between the standard prior distribution and the approximate posterior distribution of the latent causal characterization variable is specifically as follows:
[0080] L2=KL(p(z|m)||q(z|y,m))
[0081] Where L2 represents the KL divergence between the standard prior distribution and the approximate posterior distribution of the latent causal characterization variable, and KL(·) represents the discreteness function.
[0082] The latent causal representation extraction model primarily uses magnetic flux density as input and material category as an auxiliary variable to learn potential latent causal representations. As training progresses, once the loss function of the latent causal representation extraction model converges, the learned latent causal representation variables are extracted and used as part of the input.
[0083] S3: Integrate the training dataset and the corresponding latent causal representation variables to obtain the integrated training dataset;
[0084] Data such as magnetic flux density b(t), temperature t, frequency f, and waveform w are integrated with implicit causal characterization variables and input together into the initial core loss prediction model.
[0085] S4: Input the integrated training dataset into the initial core loss prediction model, set the second loss function for iterative training, and obtain the trained core loss prediction model when the preset number of iterations is reached.
[0086] Step S4 is as follows:
[0087] S4.1: Input the integrated training dataset into the initial core loss prediction model in batches to obtain the core loss prediction value;
[0088] S4.2: Construct a second loss function based on the predicted core loss value and the actual core loss value, and train the initial core loss prediction model until the preset number of iterations is reached to obtain the trained core loss prediction model.
[0089] The second loss function is:
[0090]
[0091] Where L represents the second loss function, k represents the number of training datasets, and y i This represents the true value of the core loss for training data i. This represents the predicted core loss value for training data i.
[0092] S5: Input the test dataset into the trained latent causal representation extraction model and the trained core loss prediction model for testing, and obtain the final latent causal representation extraction model and the final core loss prediction model.
[0093] The test dataset is input into the trained latent causal representation extraction model and the trained core loss prediction model for testing. The predicted core loss value of the test data is output. If the error between the predicted core loss value and the true value of the test data is greater than the preset value, the model parameters are modified and the process is returned to step S2 for retraining until the error is less than the preset value. At this point, the final latent causal representation extraction model and the final core loss prediction model are obtained.
[0094] S6: Obtain the core data of the magnetic element to be predicted, input it into the final implicit causal characterization extraction model, and obtain the implicit causal characterization variables corresponding to the core data of the magnetic element to be predicted; input the core data of the magnetic element to be predicted and its corresponding implicit causal characterization variables into the final core loss prediction model to obtain the predicted value of the core loss of the magnetic element to be predicted.
[0095] In the specific implementation process, this embodiment adopts a batch reading of the training dataset, instead of directly using complete data samples as input. All data are trained in batches, with one material used as the test set and the other materials used as the training set each time. Specifically, for example, for a dataset of four magnetic core material categories, the first batch uses all data corresponding to material 1 (M1) as the test set, and all data corresponding to material 2 (M2), material 3 (M3), and material 4 (M4) as the training set; the second batch uses all data corresponding to material 2 (M2) as the test set, and all data corresponding to material 1 (M1), material 3 (M3), and material 4 (M4) as the training set; the third batch uses all data corresponding to material 3 (M3) as the test set, and all data corresponding to material 1 (M1), material 2 (M2), and material 4 (M4) as the training set; the fourth batch uses all data corresponding to material 4 (M4) as the test set, and all data corresponding to material 1 (M1), material 2 (M2), and material 3 (M3) as the training set. Each training run of all batches of data constitutes one iteration, and the above process is repeated.
[0096] In this embodiment, magnetic flux density is used as input, and core material is used as an auxiliary variable. These are input into the latent causal characterization extraction model, which learns the conditional distribution of latent causal variables, characterizing and eliminating the confounding effect of core material category on magnetic flux density. Subsequently, the learned latent causal characterization variables, along with features such as magnetic flux density, temperature, frequency, and waveform, are input into the core loss prediction model to predict core loss. Traditional neural networks struggle to achieve effective generalization when dealing with categorical variables (such as material category). However, the latent causal characterization extraction model, by introducing material category as an auxiliary variable, successfully extracts the latent causal characterization variables between material category and magnetic flux density. This approach enables the model to handle inputs of different core materials, especially unseen new materials, achieving good generalization ability.
[0097] This embodiment successfully combines the discriminative power of the latent causal representation extraction model with the powerful modeling capabilities of neural networks by first training the latent causal representation extraction model and then inputting the learned latent causal representation variables into the core loss prediction model for final prediction. This two-stage training strategy not only improves the model's generalization ability but also effectively handles the complex interaction between material type and magnetic flux density.
[0098] The core loss prediction method based on causal representation learning described in this embodiment demonstrates strong predictive ability for unseen materials. Through cross-validation and data partitioning, all data for a specific core material is used as the test set, while data for other materials are used as the training set. The implicit causal representation extraction model still performs excellently on unseen new materials. This robustness to new materials validates the model's practical application potential, making it widely applicable in tasks requiring the prediction of new material performance.
[0099] Example 3
[0100] This embodiment provides a magnetic core loss prediction system based on causal representation learning, used to implement the magnetic core loss prediction method based on causal representation learning described in Embodiment 1 or Embodiment 2, such as... Figure 3 As shown, it includes:
[0101] The data partitioning and initial model building module is used to obtain the magnetic core dataset of magnetic components, and divide the magnetic core dataset into test dataset and training dataset according to the category of magnetic core material, and build the initial magnetic core loss prediction model and the initial implicit causal characterization extraction model.
[0102] The latent causal representation variable acquisition module is used to input the training dataset into the initial latent causal representation extraction model, set the first loss function for iterative training, and obtain the trained latent causal representation extraction model when the preset number of iterations is reached. The trained latent causal representation extraction model outputs the latent causal representation variables corresponding to the training dataset.
[0103] The data integration module is used to integrate the training dataset and the corresponding latent causal representation variables to obtain an integrated training dataset;
[0104] The core loss prediction model training module is used to input the integrated training dataset into the initial core loss prediction model, set the second loss function for iterative training, and obtain the trained core loss prediction model when the preset number of iterations is reached.
[0105] The core loss prediction model testing module is used to input the test dataset into the trained latent causal representation extraction model and the trained core loss prediction model for testing, and to obtain the final latent causal representation extraction model and the final core loss prediction model.
[0106] The prediction module is used to acquire the core data of the magnetic component to be predicted, input it into the final implicit causal characterization extraction model, and obtain the implicit causal characterization variables corresponding to the core data of the magnetic component to be predicted; input the core data of the magnetic component to be predicted and its corresponding implicit causal characterization variables into the final core loss prediction model to obtain the predicted value of the core loss of the magnetic component to be predicted.
[0107] The same or similar labels correspond to the same or similar parts;
[0108] The terms used to describe positional relationships in the accompanying drawings are for illustrative purposes only and should not be construed as limiting this patent.
[0109] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art can make other variations or modifications based on the above description. It is neither necessary nor possible to exhaustively describe all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.
Claims
1. A method for predicting magnetic core loss based on causal representation learning, characterized in that, Includes the following steps: S1: Obtain the magnetic core dataset of magnetic components, and divide the magnetic core dataset into test dataset and training dataset according to the category of magnetic core material, and construct the initial magnetic core loss prediction model and the initial implicit causal characterization extraction model. S2: Input the training dataset into the initial latent causal representation extraction model, set the first loss function for iterative training, and when the preset number of iterations is reached, obtain the trained latent causal representation extraction model. The trained latent causal representation extraction model outputs the latent causal representation variables corresponding to the training dataset. S2.1: Construct an initial latent causal representation extraction model, including an encoder and a decoder; S2.2: Input the magnetic flux density and core material in the training dataset into the encoder of the initial latent causal representation extraction model in batches to obtain latent causal representation variables. Input the latent causal representation variables and core material into the decoder of the initial latent causal representation extraction model to obtain the reconstructed magnetic flux density. S2.3: Construct the first loss function based on magnetic flux density and reconstructed magnetic flux density, train the initial latent causal representation extraction model until the preset number of iterations is reached, and obtain the trained latent causal representation extraction model. The parameters for building the encoder and decoder are determined by the parameters of the fully connected neural network. The training generates a model that includes the encoding and reconstruction processes during decoding. The magnetic flux density b(t) is used as the input to the encoder, and the magnetic core material is used as the input supervision signal m. The output is a latent causal representation variable z, in the form: z=f enc (b(t),m) The decoder outputs the reconstructed result based on the supervision signal m and the latent causal representation variable z from the encoder output. Its form is: S3: Integrate the training dataset and the corresponding latent causal representation variables to obtain the integrated training dataset; The magnetic flux density b(t), temperature t, frequency f, waveform w, and implicit causal characterization variables are integrated and input together into the initial core loss prediction model; S4: Input the integrated training dataset into the initial core loss prediction model, set the second loss function for iterative training, and obtain the trained core loss prediction model when the preset number of iterations is reached. S5: Input the test dataset into the trained latent causal representation extraction model and the trained core loss prediction model for testing, and obtain the final latent causal representation extraction model and the final core loss prediction model. S6: Obtain the core data of the magnetic element to be predicted, input it into the final implicit causal characterization extraction model, and obtain the implicit causal characterization variables corresponding to the core data of the magnetic element to be predicted; input the core data of the magnetic element to be predicted and its corresponding implicit causal characterization variables into the final core loss prediction model to obtain the predicted value of the core loss of the magnetic element to be predicted.
2. The core loss prediction method based on causal representation learning according to claim 1, characterized in that, In step S1, the magnetic core dataset includes magnetic core material M, magnetic flux density B(t), temperature T, frequency F, waveform W, and magnetic core loss Y.
3. The core loss prediction method based on causal representation learning according to claim 2, characterized in that, In step S1, obtaining the magnetic core dataset of the magnetic element and dividing the magnetic core dataset into a test dataset and a training dataset according to the category of magnetic core material includes dividing according to the category of magnetic core material, and sequentially using the magnetic core data of the same magnetic core material as the test dataset and the magnetic core data of other magnetic core materials as the training dataset.
4. The core loss prediction method based on causal representation learning according to claim 1, characterized in that, The first loss function in step S2.3 includes the KL divergence between the reconstruction error and the standard prior distribution and the approximate posterior distribution of the latent causal representation variable, specifically: ELBO = L1 - L2 Where ELBO represents the first loss function, L1 represents the reconstruction error, and L2 represents the KL divergence between the standard prior and approximate posterior of the latent causal representation variable.
5. The core loss prediction method based on causal representation learning according to claim 4, characterized in that, The method for determining the reconstruction error includes: L1=E q(z|y,m) [logp(y|z)] Where L1 represents the reconstruction error, q(z|y,m) is the approximate posterior distribution, and p(y|z) is the conditional prior distribution.
6. The core loss prediction method based on causal representation learning according to claim 4, characterized in that, The methods for determining the KL divergence between the standard prior distribution and the approximate posterior distribution of the implicit causal characterization variable include: L2=KL(p(z|m)||q(z|y,m)) Where L2 represents the KL divergence between the standard prior distribution and the approximate posterior distribution of the latent causal characterization variable, and KL(·) represents the discreteness function.
7. The core loss prediction method based on causal representation learning according to claim 4, characterized in that, Step S4 includes: S4.1: Input the integrated training dataset into the initial core loss prediction model in batches to obtain the core loss prediction value; S4.2: Construct a second loss function based on the predicted core loss value and the actual core loss value, and train the initial core loss prediction model until the preset number of iterations is reached to obtain the trained core loss prediction model.
8. The core loss prediction method based on causal representation learning according to claim 7, characterized in that, The methods for determining the second loss function include: Where L represents the second loss function, k represents the number of training datasets, and y i This represents the true value of the core loss for training data i. This represents the predicted core loss value for training data i.
9. A magnetic core loss prediction system based on causal representation learning, used to implement the magnetic core loss prediction method based on causal representation learning as described in any one of claims 1-8, characterized in that, include: The data partitioning and initial model building module is used to obtain the magnetic core dataset of magnetic components, and divide the magnetic core dataset into test dataset and training dataset according to the category of magnetic core material, and build the initial magnetic core loss prediction model and the initial implicit causal characterization extraction model. The latent causal representation variable acquisition module is used to input the training dataset into the initial latent causal representation extraction model, set the first loss function for iterative training, and obtain the trained latent causal representation extraction model when the preset number of iterations is reached. The trained latent causal representation extraction model outputs the latent causal representation variables corresponding to the training dataset. The data integration module is used to integrate the training dataset and the corresponding latent causal representation variables to obtain an integrated training dataset; The core loss prediction model training module is used to input the integrated training dataset into the initial core loss prediction model, set the second loss function for iterative training, and obtain the trained core loss prediction model when the preset number of iterations is reached. The core loss prediction model testing module is used to input the test dataset into the trained latent causal representation extraction model and the trained core loss prediction model for testing, and to obtain the final latent causal representation extraction model and the final core loss prediction model. The prediction module is used to acquire the core data of the magnetic component to be predicted, input it into the final implicit causal characterization extraction model, and obtain the implicit causal characterization variables corresponding to the core data of the magnetic component to be predicted; input the core data of the magnetic component to be predicted and its corresponding implicit causal characterization variables into the final core loss prediction model to obtain the predicted value of the core loss of the magnetic component to be predicted.
Citation Information
Patent Citations
Free shape metasurface construction method, system and device and readable storage medium
CN115859794A
Ammonia arc thruster service life evaluation method and system
CN118468734A