Transformer core permeability prediction method and system combined with transfer learning

Abstract

The application provides a method and system for predicting the magnetic permeability of a transformer core by combining transfer learning, and belongs to the field of predicting the magnetic permeability of a core. The method comprises the following steps: obtaining material basic information, retrieving and obtaining a near-neighbor material, obtaining a set of time-space near-neighbor material magnetic permeability test correlation data and a set of time-space near-neighbor material magnetic permeability; performing time-space synchronization analysis to obtain a time-space synchronization analysis result; obtaining a base class training data set and an adversarial class training data set, training a fine-tuned transformer core magnetic permeability predictor; performing transfer learning to obtain a target transformer core magnetic permeability predictor, and predicting the magnetic permeability of the target transformer core. The application solves the technical problems of insufficient training of the prediction model and low prediction accuracy caused by the lack of magnetic permeability data of the transformer core material in the prior art, and achieves the technical effect of improving the prediction accuracy of the magnetic permeability of the transformer core by fully utilizing the rich data of the near-neighbor material through transfer learning.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of core magnetic permeability prediction, and particularly relates to a mutual inductor core magnetic permeability prediction method and system combining transfer learning. BACKGROUND

[0002] The mutual inductor core is the core component of the mutual inductor in the power system, and its magnetic permeability characteristics directly affect the measurement accuracy and operation performance of the mutual inductor. Accurate prediction of the magnetic permeability of the mutual inductor core is of great significance for mutual inductor design optimization, performance evaluation and fault diagnosis.

[0003] Traditional mutual inductor core magnetic permeability prediction methods rely on a large amount of experimental data and empirical formulas. However, due to the wide variety of mutual inductor core materials, and the significant differences in the magnetic permeability characteristics of different materials under different working conditions, especially in the knee region of the magnetization curve, the core magnetic permeability presents nonlinear variation characteristics, making the magnetic permeability prediction extremely complex.

[0004] In the prior art, the magnetic permeability prediction of a specific mutual inductor core material requires collecting a large amount of test data of the material under various working conditions to train the prediction model. However, in actual applications, due to high experimental costs, long test periods, test condition limitations and other factors, it is often difficult to obtain sufficient target mutual inductor core magnetic permeability data, the training data of the prediction model is insufficient, resulting in insufficient training of the prediction model, low prediction accuracy, and inability to meet the engineering actual demand. SUMMARY

[0005] The present application provides a mutual inductor core magnetic permeability prediction method and system combining transfer learning to solve the technical problem of insufficient training of the prediction model and low prediction accuracy caused by the lack of mutual inductor core material magnetic permeability data in the prior art.

[0006] The technical solution of the present application to solve the above technical problems is as follows:

[0007] In a first aspect, the present application provides a mutual inductor core permeability prediction method combined with transfer learning, comprising: obtaining material basic information of a target mutual inductor core, and retrieving a near neighbor material based on the material basic information as an index, and retrieving permeability test data of the near neighbor material at different regions and different time points according to a preset test correlation index, to obtain a time-space near neighbor material permeability test correlation data set and a time-space near neighbor material permeability set; performing time-space synchronization analysis on the time-space near neighbor material permeability set and the time-space near neighbor material permeability test correlation data set, to obtain a time-space synchronization analysis result; distinguishing base class training data and adversarial class training data from the time-space synchronization analysis result, to obtain a base class training data set and an adversarial class training data set; training an initial mutual inductor core permeability predictor using the base class training data set, and performing hyperparameter adversarial fine-tuning on the trained initial mutual inductor core permeability predictor according to the adversarial class training data set, to obtain a fine-tuned mutual inductor core permeability predictor; performing transfer learning on the fine-tuned mutual inductor core permeability predictor, to obtain a target mutual inductor core permeability predictor, and performing target mutual inductor core permeability prediction using the target mutual inductor core permeability predictor.

[0008] In a second aspect, the present application provides a mutual inductor core permeability prediction system combined with transfer learning, comprising: a data acquisition module, configured to obtain material basic information of a target mutual inductor core, and retrieve a near neighbor material based on the material basic information as an index, and retrieve permeability test data of the near neighbor material at different regions and different time points according to a preset test correlation index, to obtain a time-space near neighbor material permeability test correlation data set and a time-space near neighbor material permeability set; a time-space analysis module, configured to perform time-space synchronization analysis on the time-space near neighbor material permeability set and the time-space near neighbor material permeability test correlation data set, to obtain a time-space synchronization analysis result; a data classification module, configured to distinguish base class training data and adversarial class training data from the time-space synchronization analysis result, to obtain a base class training data set and an adversarial class training data set; a model training module, configured to train an initial mutual inductor core permeability predictor using the base class training data set, and perform hyperparameter adversarial fine-tuning on the trained initial mutual inductor core permeability predictor according to the adversarial class training data set, to obtain a fine-tuned mutual inductor core permeability predictor; and a transfer prediction module, configured to perform transfer learning on the fine-tuned mutual inductor core permeability predictor, to obtain a target mutual inductor core permeability predictor, and perform target mutual inductor core permeability prediction using the target mutual inductor core permeability predictor.

[0009] The present application has the following beneficial effects:

[0010] The material basic information of the target mutual inductor core is acquired, and the neighboring material is retrieved based on the material basic information as an index, and the permeability test data of the neighboring material at different regions and different time points is retrieved according to a preset test correlation index, so as to obtain a time-space neighboring material permeability test correlation data set and a time-space neighboring material permeability set, thereby fully mining the similar neighboring material resources of the target mutual inductor core, expanding the data basis that can be used for training, and solving the problem of the data scarcity of the target mutual inductor core. The time-space neighboring material permeability set is subjected to time-space synchronization carding on the time-space neighboring material permeability test correlation data set, and a time-space synchronization carding result is obtained, which can eliminate the data deviation caused by the test environment difference, ensure that the test correlation data corresponding to the same permeability value has consistency, and improve the data quality. The time-space synchronization carding result is subjected to stage-based basic class training data and adversarial class training data distinction, and a basic class training data set and an adversarial class training data set are obtained, and through this data classification strategy, a data basis is provided for subsequent stage-based training and adversarial optimization, so that the predictor can be fully trained at two levels of basic learning and difficult sample optimization. The basic class training data set is used to train an initial mutual inductor core permeability predictor, and the initial mutual inductor core permeability predictor trained is subjected to hyperparameter adversarial fine-tuning according to the adversarial class training data set, and a fine-tuned mutual inductor core permeability predictor is obtained, and through the stage-based training strategy, the basic prediction ability is established first, and then the difficult sample optimization is performed on the adversarial class data, so as to enhance the generalization ability and robustness of the predictor. The fine-tuned mutual inductor core permeability predictor is subjected to transfer learning, a target mutual inductor core permeability predictor is obtained, and the target mutual inductor core permeability predictor is used for target mutual inductor core permeability prediction; through the transfer learning, the general model trained on the neighboring material is adapted to the target mutual inductor core, knowledge transfer from the data-rich neighboring material to the data-scarce target mutual inductor core is realized, and finally the high-precision permeability prediction ability for the target mutual inductor core is obtained.

[0011] Through the above technical solution, the rich data resources of the neighboring material are fully utilized, the data quality is ensured through time-space synchronization carding, the learning ability of the model is improved through the stage-based training and adversarial fine-tuning strategy, and finally the effective conversion from the general predictor to the special predictor for the target mutual inductor core is realized through the transfer learning. Both the fundamental problem of the data scarcity of the target mutual inductor core is solved, and the prediction accuracy of the prediction model under the complex nonlinear permeability characteristics is ensured, and the high-precision prediction of the mutual inductor core permeability is realized, thereby providing reliable technical support for the design optimization and performance evaluation of the mutual inductor. BRIEF DESCRIPTION OF DRAWINGS

[0012] Figure 1 A flowchart of a mutual inductor core permeability prediction method provided by the application combined with transfer learning is provided.

[0013] Figure 2 A structure schematic diagram of a mutual inductor core permeability prediction system combined with transfer learning provided by the present application is shown in the figure.

[0014] In the figure, the components represented by each number are as follows:

[0015] The data acquisition module 11, the space-time carding module 12, the data classification module 13, the model training module 14, and the transfer prediction module 15. DETAILED DESCRIPTION

[0016] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0017] In the description of the present application, the terms "first", "second" are used only for the purpose of description, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined with "first", "second" can explicitly or implicitly include one or more of the features. In the description of the present application, the meaning of "multiple" is two or more, unless otherwise specifically limited.

[0018] In the description of the present application, the term "for example" is used to indicate "as an example, illustration or explanation". Any embodiment described as "for example" in the present application is not necessarily interpreted as more preferred or more advantageous than other embodiments. The following description is given in order to enable any person skilled in the art to implement and use the present application. In the following description, details are listed for the purpose of explanation. It should be understood that those skilled in the art can realize the present application without using these specific details. In other examples, well-known structures and processes will not be described in detail to avoid unnecessary details making the description of the present application obscure. Therefore, the present application is not intended to be limited to the shown embodiments, but is consistent with the broadest scope in accordance with the principles and characteristics disclosed.

[0019] Embodiment one, as shown in the present application, provides a mutual inductor core permeability prediction method combined with transfer learning, which comprises: Figure 1

[0020] ​S1, acquire material basic information of a target transformer core, and retrieve neighboring materials indexed by the material basic information, and retrieve permeability test data of the neighboring materials at different regions and different time points according to a preset test correlation index, to obtain a spatiotemporal neighboring material permeability test correlation data set and a spatiotemporal neighboring material permeability set;

[0021] Specifically, first, material basic information of a target transformer core is acquired. The material basic information refers to basic attribute parameters of the target transformer core material, specifically including chemical composition (such as silicon content and carbon content in silicon steel sheets), physical property parameters (such as permeability range, saturation magnetic induction intensity, and coercive force), geometric size parameters (such as core diameter, thickness, and lamination number), manufacturing process information (such as heat treatment temperature, cold rolling process, and insulation coating type), and usage environment conditions (such as working temperature range and humidity requirement).

[0022] Then, the acquired material basic information is used as a retrieval index to perform similarity matching retrieval in a pre-established material database, to acquire neighboring materials having similar characteristics to the target transformer core. The neighboring materials refer to other transformer core materials having a certain similarity in material composition, physical properties, manufacturing process, and the like, to the target transformer core. This similarity matching retrieval process uses multi-dimensional feature vector comparison, to filter out materials having a similarity exceeding a preset threshold as neighboring materials, by calculating similarity scores of the target transformer core and candidate materials in each feature dimension.

[0023] Subsequently, data retrieval is performed on the neighboring materials according to a preset test correlation index. The preset test correlation index refers to a standard for evaluating and screening the quality and relevance of permeability test data, including consistency of test methods (such as whether the same test standard and test equipment are used), comparability of test environments (such as similarity of environmental factors such as temperature, humidity, and electromagnetic interference), and data reliability indicators (such as test repeatability and data integrity). Based on these preset test correlation indexes, permeability test data of the neighboring materials at different regions and different time points is retrieved. Different regions refer to different test sites or test laboratories, and due to differences in geographical location, equipment conditions, and environmental factors, test results of the same material at different regions may have deviations. Different time points refer to different test time points, and the magnetic permeability characteristics of the material may change in the time dimension due to factors such as aging, environmental changes, and test equipment calibration status.

[0024] Through the above retrieval process, two data sets are obtained: a spatiotemporal near-neighbor material magnetic permeability test correlation data set and a spatiotemporal near-neighbor material magnetic permeability set. The spatiotemporal near-neighbor material magnetic permeability test correlation data set contains magnetic permeability test data and its corresponding test environment information, test condition parameters, test timestamp, test location identifier, and other correlation data, providing complete context information for subsequent data quality assessment and synchronicity analysis. The spatiotemporal near-neighbor material magnetic permeability set is a preliminary filtered pure magnetic permeability numerical data, providing basic data support for transfer learning.

[0025] S2, spatiotemporal synchronization combing the spatiotemporal near-neighbor material magnetic permeability set to the spatiotemporal near-neighbor material magnetic permeability test correlation data set, obtaining a spatiotemporal synchronization combing result.

[0026] Specifically, spatiotemporal synchronization combing refers to a data processing process that analyzes and corrects the inconsistency of test data caused by time difference and space difference, ensuring that the test correlation data corresponding to the same magnetic permeability value has spatiotemporal consistency. Specifically, since the magnetic permeability test data of near-neighbor materials comes from different test environments and different test times, even the same magnetic permeability value, its corresponding test correlation data may have significant differences. Therefore, the spatiotemporal near-neighbor material magnetic permeability set is spatiotemporally synchronized to the spatiotemporal near-neighbor material magnetic permeability test correlation data set.

[0027] First, in the spatiotemporal dimension, time synchronization analysis is performed. Since the calibration state of the test equipment, the seasonal changes of environmental temperature and humidity, the aging degree of the material, and other factors change over time, the same magnetic permeability value of the same material measured at different time points may correspond to different test conditions. By establishing a time correction model, the correlation between test time and test conditions is analyzed, and the test deviation caused by time factors is identified and corrected. At the same time, in the spatial dimension, spatial synchronization analysis is performed. The equipment accuracy, environmental conditions, and operation specifications of different test locations differ, resulting in the same magnetic permeability value of the same material at different locations may correspond to different test correlation parameters. By establishing a spatial correction model, the correlation between test location and test result is analyzed, and the systematic deviation caused by regional factors is identified and eliminated.

[0028] Subsequently, for each magnetic permeability value in the spatiotemporal neighborhood material magnetic permeability set, its corresponding test correlation data is searched in the spatiotemporal neighborhood material magnetic permeability test correlation data set. For multiple sets of test correlation data corresponding to the same magnetic permeability value, through spatiotemporal synchronization correction, test correlation data incorrectly classified due to test environment differences is eliminated, and test correlation data that truly belongs to the magnetic permeability value is retained. The spatiotemporal synchronization correction uses a multi-dimensional similarity analysis method, comprehensively considers multiple test condition parameters such as test temperature, test humidity, test frequency, and magnetic field strength, calculates the matching degree between the test correlation data and the magnetic permeability value, sets a synchronization threshold, identifies test correlation data with a matching degree lower than the threshold as spatiotemporal asynchronous data, and eliminates or reclassifies the test correlation data.

[0029] Through the above spatiotemporal synchronization sorting process, a spatiotemporal synchronization sorting result is obtained. The spatiotemporal synchronization sorting result is a high-quality data set after spatiotemporal correction and data cleaning, ensuring the accurate correspondence between the magnetic permeability value and its corresponding test correlation data, eliminating data pollution caused by test environment deviation, and providing a reliable data basis for subsequent phased training data classification.

[0030] S3, distinguishing the phased base class training data and the adversarial class training data from the spatiotemporal synchronization sorting result, to obtain a base class training data set and an adversarial class training data set.

[0031] Specifically, the phased base class training data and the adversarial class training data distinction refers to a data processing process of classifying data in the spatiotemporal synchronization sorting result according to its role in model training according to the characteristic attributes and training targets of the training data. The classification strategy follows the phased training theory in transfer learning, aiming to improve the generalization ability and prediction accuracy of the model through differentiated training of different types of data.

[0032] First, the magnetic permeability data in the spatiotemporal synchronization result is analyzed for features. Specifically, by calculating the feature complexity, data distribution density, and similarity to the target transformer core of each magnetic permeability sample, the role of the sample in model training is evaluated. The feature complexity refers to the complexity of the test correlation data corresponding to the magnetic permeability sample, including the diversity of test conditions, the range of environmental parameters, etc. The data distribution density refers to the frequency of occurrence of the magnetic permeability value in the entire data set and the distribution of surrounding data points. The similarity index reflects the matching degree of the sample and the target transformer core material. Based on the above feature analysis, the data is divided into two main categories. The base class training data refers to the magnetic permeability data samples with typical features, high data quality, and strong correlation with the target transformer core. This type of data usually has low feature complexity, is concentrated in the data set, and can provide stable basic learning ability for the model. The base class training data is mainly used for the initial training stage of the model, helping the model establish a basic understanding and prediction framework for the magnetic permeability prediction problem. The adversarial class training data refers to magnetic permeability data samples with high feature complexity, relatively sparse data distribution, or certain noise interference. This type of data corresponds to the magnetic permeability performance under complex test environments, boundary conditions, or abnormal working conditions, and belongs to the "difficult samples" in the training process. The adversarial class training data is mainly used for the adversarial fine-tuning stage of the model, and through the learning of these difficult samples, the robustness and adaptability of the model to complex situations are enhanced.

[0033] In the data classification process, clustering analysis algorithm and threshold judgment mechanism are used. First, the spatiotemporal synchronization result is clustered for multi-dimensional features to identify the natural grouping pattern of the data. Then, complexity threshold, density threshold, and similarity threshold are set to divide the samples that meet the base class data features into the base class training data set and the samples with challenging features into the adversarial class training data set.

[0034] Through the above phased data classification process, the base class training data set and the adversarial class training data set are obtained. The base class training data set provides stable and reliable learning samples for the basic training of the subsequent model, and the adversarial class training data set provides challenging training samples for the adversarial optimization of the model. The two work together to lay a hierarchical data foundation for building a high-performance magnetic permeability prediction model.

[0035] S4, training an initial transformer core magnetic permeability predictor using the base class training data set, and performing hyperparameter adversarial fine-tuning on the trained initial transformer core magnetic permeability predictor according to the adversarial class training data set to obtain a fine-tuned transformer core magnetic permeability predictor;

[0036] Specifically, the training process of the initial mutual inductor core permeability predictor adopts a supervised learning algorithm, taking the base-class training data set as the training sample. First, the network architecture of the mutual inductor core permeability predictor is established, which includes an input layer, multiple hidden layers and an output layer. The input layer receives the permeability test correlation data, including test temperature, test frequency, magnetic field strength, material composition and other characteristic parameters; the hidden layer learns the complex mapping relationship between the input features and the permeability through nonlinear transformation and feature extraction; and the output layer outputs the predicted permeability value.

[0037] In the basic training stage, the network architecture of the mutual inductor core permeability predictor is trained using the base-class training data set. Since the base-class training data has good data quality and low feature complexity, it can quickly converge on this data set and establish a basic understanding framework for permeability prediction problems. The training process adopts a gradient descent algorithm, which continuously optimizes model parameters by minimizing the loss function between predicted values and true values. After sufficient basic training, the initial mutual inductor core permeability predictor is obtained, which has the basic ability to predict the permeability under normal working conditions.

[0038] Then, the initial mutual inductor core permeability predictor is fine-tuned with hyperparameters based on the adversarial training data set. Hyperparameter adversarial fine-tuning refers to the process of adjusting the hyperparameter configuration of the model and further optimizing the model using challenging adversarial training data. Hyperparameters include learning rate, regularization coefficient, network layer number, neuron number and other key parameters that affect model performance. In the adversarial fine-tuning process, first analyze the feature distribution of the adversarial training data, identify the differences and complexity of these data relative to the base-class data. Based on these analysis results, dynamically adjust the hyperparameter settings of the model. For example, for adversarial samples with high feature complexity, the learning rate may be reduced to avoid overfitting, or the regularization strength may be increased to improve the model's generalization ability. Adversarial fine-tuning adopts a contrastive learning strategy, which enhances the model's ability to identify difficult samples by constructing positive sample pairs and negative sample pairs. Similar samples in the adversarial training data form positive sample pairs, and samples with large differences form negative sample pairs. The model is trained to correctly distinguish these samples in the feature space, thereby improving the model's adaptability and robustness to complex situations. In the fine-tuning process, a progressive training strategy is adopted, gradually increasing the proportion of adversarial data for training, so that the model can smoothly transition from basic prediction ability to prediction ability in complex scenarios. At the same time, set early stopping mechanism and performance monitoring indicators to prevent overfitting of the model on adversarial data, and ensure that the model maintains basic prediction ability while obtaining stronger generalization performance.

[0039] Through the above hyperparameter adversarial fine-tuning process, a fine-tuned mutual inductor core permeability predictor is obtained. The fine-tuned mutual inductor core permeability predictor not only inherits the prediction ability of the initial mutual inductor core permeability predictor under normal working conditions, but also obtains the prediction ability for complex working conditions and boundary conditions through the training of the adversarial data, has stronger robustness and adaptability, and lays a high-quality model foundation for subsequent transfer learning.

[0040] S5, performing transfer learning on the fine-tuned mutual inductor core permeability predictor to obtain a target mutual inductor core permeability predictor, and using the target mutual inductor core permeability predictor to perform target mutual inductor core permeability prediction.

[0041] Specifically, transfer learning refers to migrating the knowledge and prediction ability of the fine-tuned mutual inductor core permeability predictor trained on the near-neighbor material data to the target mutual inductor core material. Since the fine-tuned mutual inductor core permeability predictor is trained based on the rich data of the near-neighbor material, it has a deep understanding of the permeability prediction problem and generalization ability, and through transfer learning, these prediction abilities can be effectively adapted to the target mutual inductor core with scarce data.

[0042] In the transfer learning process, first, limited permeability test data of the target mutual inductor core material is obtained. These data are usually small in quantity, but highly representative and targeted for the target mutual inductor core. These target mutual inductor core data are used as fine-tuning samples for transfer learning to guide the fine-tuned mutual inductor core permeability predictor to transform from a general predictor to a target mutual inductor core dedicated predictor. Specifically, the structure of the bottom feature extraction network of the fine-tuned mutual inductor core permeability predictor is kept unchanged, and these bottom networks have learned the general ability to extract permeability-related features from the test correlation data, and the focus is on adjusting the high-level decision network and the output layer parameters of the model to adapt to the specific permeability characteristics of the target mutual inductor core. In the parameter adjustment process, a small learning rate gradient update strategy is adopted to avoid significantly modifying the trained model parameters and prevent damaging the useful knowledge learned by the model on the near-neighbor material. In the validation stage of transfer learning, evaluation strategies such as cross-validation and leave-one-out method are used to evaluate the model performance using limited data of the target mutual inductor core, monitor the prediction error, convergence speed, and generalization performance of the target mutual inductor core, and ensure the effectiveness of the transfer learning process.

[0043] Through the above transfer learning process, a target mutual inductor core permeability predictor is obtained. The target mutual inductor core permeability predictor is optimized specifically for the target mutual inductor core material, retains the general permeability prediction knowledge learned from the near-neighbor material data, and adapts to the specific permeability characteristics of the target mutual inductor core, and can achieve high-precision permeability prediction under the condition of scarce target mutual inductor core data.

[0044] Finally, the target transformer core permeability prediction is performed by using the target transformer core permeability predictor. In actual application, the user inputs the test correlation data of the target transformer core (such as working temperature, frequency, magnetic field strength, etc.), and the target transformer core permeability predictor outputs the predicted value of the permeability under the corresponding working condition, thereby providing accurate and reliable technical support for the design optimization, performance evaluation and fault diagnosis of the transformer.

[0045] Further, the preset test correlation indexes include magnetic field strength, magnetic flux density, frequency, temperature, material characteristics and process parameters.

[0046] Specifically, the magnetic field strength refers to the strength value of the external magnetic field applied to the transformer core material. The magnetic field strength is one of the core parameters that affect the core permeability, and especially in the knee region of the magnetization curve, a slight change in the magnetic field strength will cause a significant change in the permeability. By recording the magnetic permeability test data under different magnetic field strengths, a nonlinear mapping relationship between the magnetic field strength and the permeability is established.

[0047] The magnetic flux density refers to the magnetic flux passing through per unit area. The magnetic flux density and the magnetic field strength jointly determine the magnetization state of the material, and is an important index for evaluating the magnetic performance of the core. In the data preparation stage of transfer learning, a vector magnetic property tester or a ring sample measurement device is preferably used to measure the BH curve of the material under different working conditions, that is, the relationship curve between the magnetic flux density B and the magnetic field strength H, and the effective permeability μeff is obtained through mathematical conversion.

[0048] The frequency refers to the frequency parameter of the alternating magnetic field. The frequency has a significant impact on the permeability of the core material, mainly reflected in the frequency-related physical phenomena such as eddy current loss and hysteresis loss. The magnetic permeability test data under different frequencies can reflect the change rule of the magnetic performance of the material under different working frequencies, thereby providing data support for the design of the transformer for wideband applications.

[0049] The temperature refers to the temperature condition of the test environment, measured in degrees Celsius (℃). The temperature is an important environmental factor affecting the permeability, and the permeability of the core material will change with the temperature, especially in high temperature environment, the magnetic performance of the material may be significantly degraded. By collecting the magnetic permeability test data under different temperature conditions, the prediction accuracy of the predictor under different working temperatures is improved.

[0050] Material properties refer to the inherent physical and chemical characteristic parameters of the core material of the transformer, including but not limited to the chemical composition of the material (such as the silicon content and carbon content in silicon steel sheets), the crystal structure, the grain size, the material density, the resistivity and other basic physical parameters. These material properties directly determine the basis of the magnetic performance of the material and are an important basis for material similarity matching and near-neighbor material retrieval.

[0051] Process parameters refer to the process condition parameters involved in the manufacturing and processing of the core of the transformer. Specifically, they include annealing temperature, annealing time, winding tension parameters and air gap parameters. Among them, the annealing temperature and annealing time affect the crystal structure and internal stress distribution of the material, and thus affect the permeability characteristics; the winding tension parameter refers to the mechanical tension applied during core winding, and excessive tension will introduce mechanical stress, leading to a decrease in permeability; the air gap parameter refers to the intentionally set air gap size in the core, and the presence of air gap will significantly affect the equivalent permeability of the core.

[0052] In the data preparation and feature construction phase, the above-mentioned preset test correlation indicators are used as input variables for feature construction, with effective permeability as the output variable, to establish a mapping relationship from multi-dimensional input features to permeability prediction values, providing a complete feature engineering basis for transfer learning.

[0053] Further, the spatiotemporal near-neighbor material permeability set is subjected to spatiotemporal synchronization sorting on the spatiotemporal near-neighbor material permeability test correlation data set, to obtain a spatiotemporal synchronization sorting result, including:

[0054] S21, dispersively screening the spatiotemporal near-neighbor material permeability set to determine a plurality of dispersively screened spatiotemporal near-neighbor material permeabilities;

[0055] S22, aggregating the spatiotemporal near-neighbor material permeability set based on the plurality of dispersively screened spatiotemporal near-neighbor material permeabilities to obtain a plurality of aggregated spatiotemporal near-neighbor material permeability sets;

[0056] S23, mapping and sorting the spatiotemporal near-neighbor material permeability test correlation data set based on the plurality of aggregated spatiotemporal near-neighbor material permeability sets to obtain a plurality of aggregated spatiotemporal near-neighbor material permeability test correlation data sets, mapping and correlating the plurality of aggregated spatiotemporal near-neighbor material permeability test correlation data sets and the plurality of aggregated spatiotemporal near-neighbor material permeability sets to obtain the spatiotemporal synchronization sorting result.

[0057] In an implementable embodiment, firstly, the dispersion screening is performed on the spatiotemporal near-neighbor material permeability set to determine a plurality of dispersion screening spatiotemporal near-neighbor material permeabilities. The dispersion screening refers to a data screening process of identifying and retaining representative and reliable permeability values by analyzing the distribution characteristics of the permeability data in the feature space. Since the spatiotemporal near-neighbor material permeability data is derived from different test environments and test times, the reliability of the data cannot be directly identified, and there are abnormal data caused by factors such as test errors, equipment deviations, and environmental interference. Specifically, a statistical analysis method is used to evaluate the dispersion of the spatiotemporal near-neighbor material permeability set. The data density and distribution uniformity of each permeability value in its neighborhood range are calculated to identify those permeability values with good dispersion and representativeness in the feature space. The permeability values with good dispersion usually have the following characteristics: there are a sufficient number of similar data points around them, forming stable data clusters; they maintain an appropriate distance from other permeability values to avoid data overlap; and they show consistency in multiple tests with small test variances. By setting the dispersion threshold and clustering radius parameters, the permeability values that meet the dispersion requirements are screened to form a plurality of dispersion screening spatiotemporal near-neighbor material permeabilities. These screened permeability values have high data quality and reliability, providing a reliable data basis for subsequent aggregation operations.

[0058] Subsequently, the spatiotemporal near-neighbor material permeability set is aggregated based on the plurality of dispersion screening spatiotemporal near-neighbor material permeabilities to obtain a plurality of aggregated spatiotemporal near-neighbor material permeability sets. The aggregation process refers to a data organization process of classifying and integrating similar permeability data with dispersion screening spatiotemporal near-neighbor material permeabilities as the center points. Specifically, a similarity-based aggregation algorithm is used to calculate the similarity of other permeability values in the spatiotemporal near-neighbor material permeability set to each dispersion screening spatiotemporal near-neighbor material permeability as the aggregation center. The similarity calculation considers multiple dimensions such as the difference between permeability values, the proximity of corresponding test conditions, and the spatiotemporal correlation of data sources. For each aggregation center, a similarity threshold is set, and the permeability data with a similarity exceeding the threshold is classified into the aggregated set corresponding to the aggregation center. In this way, the originally dispersed permeability data is organized into a plurality of aggregated spatiotemporal near-neighbor material permeability sets with inherent consistency. The permeability data that is not aggregated by any aggregation center will be identified as outlier data or noise data and will be excluded from the subsequent processing flow.

[0059] Afterwards, the plurality of aggregated spatiotemporal near-neighbor material permeability test correlation data sets are mapped and sorted based on the plurality of aggregated spatiotemporal near-neighbor material permeability sets, to obtain a plurality of aggregated spatiotemporal near-neighbor material permeability test correlation data sets. Mapping and sorting refers to a data processing process of reorganizing and quality verifying the test correlation data according to the permeability similarity principle. The mapping and sorting process is based on the spatiotemporal material permeability similarity screening principle. This principle shows that if a plurality of permeability values have similarity and are aggregated into the same set, then the test correlation data corresponding to these permeability values should also have similarity. Based on this principle, some accidental abnormal test correlation data can be identified and removed, and the overall quality of the data set can be improved. For each permeability value in the aggregated spatiotemporal near-neighbor material permeability set, its corresponding test correlation data is searched in the spatiotemporal near-neighbor material permeability test correlation data set. For a plurality of test correlation data in the same aggregated set, the similarity and consistency of the characteristic parameters (such as temperature, frequency, magnetic field strength, etc.) are calculated. If a set of test correlation data is significantly different from other data in the same set, it is identified as accidental data and removed. After mapping and sorting, a plurality of aggregated spatiotemporal near-neighbor material permeability test correlation data sets are obtained. The data in each aggregated spatiotemporal near-neighbor material permeability test correlation data set has good internal consistency, and the data pollution caused by test environment differences and accidental factors is eliminated.

[0060] Afterwards, the plurality of aggregated spatiotemporal near-neighbor material permeability test correlation data sets and the plurality of aggregated spatiotemporal near-neighbor material permeability sets are mapped and correlated to establish an accurate correspondence between the permeability values and their corresponding test correlation data, to obtain a spatiotemporal synchronization sorting result. The spatiotemporal synchronization sorting result ensures the spatiotemporal consistency and reliability of the data, and provides a high-quality data basis for subsequent phased training data classification.

[0061] Further, the spatiotemporal near-neighbor material permeability set is dispersedly screened to determine a plurality of dispersedly screened spatiotemporal near-neighbor material permeabilities, comprising:

[0062] S211, a plurality of initial spatiotemporal near-neighbor material permeabilities are randomly extracted from the spatiotemporal near-neighbor material permeability set.

[0063] S212, it is judged whether the similarity of any two initial spatiotemporal near-neighbor material permeabilities in the plurality of initial spatiotemporal near-neighbor material permeabilities is less than or equal to a preset threshold value, and if so, the plurality of initial spatiotemporal near-neighbor material permeabilities are verified by dispersed screening and used as the plurality of dispersedly screened spatiotemporal near-neighbor material permeabilities.

[0064] In a preferred embodiment, first, a plurality of initial spatiotemporal near-neighbor material permeabilities are randomly extracted from the set of spatiotemporal near-neighbor material permeabilities. Random extraction refers to a data sampling process in which a random sampling method is used to select a number of spatiotemporal near-neighbor material permeabilities from the set of spatiotemporal near-neighbor material permeabilities as candidate objects for dispersion screening. Specifically, first, a random extraction quantity parameter is determined, which is set according to the overall size of the set of spatiotemporal near-neighbor material permeabilities and the desired number of dispersion screening results. In general, the extraction quantity should ensure sufficient representativeness while avoiding excessive computational complexity. A uniform random sampling method is used to ensure that each spatiotemporal near-neighbor material permeability in the set of spatiotemporal near-neighbor material permeabilities has an equal probability of being selected, avoiding sampling bias. In the random extraction process, each spatiotemporal near-neighbor material permeability is assigned a random number identifier, sorted by random number size, and the top-ranked number of permeability values are selected as initial spatiotemporal near-neighbor material permeabilities, resulting in a plurality of initial spatiotemporal near-neighbor material permeabilities. This random extraction method can effectively avoid human selection bias and ensure the objectivity and reliability of the subsequent dispersion verification process.

[0065] Subsequently, it is determined whether the similarity between any two of the plurality of initial spatiotemporal near-neighbor material permeabilities is less than or equal to a predetermined threshold value. If so, the plurality of initial spatiotemporal near-neighbor material permeabilities are verified by dispersion screening as the plurality of dispersion screening spatiotemporal near-neighbor material permeabilities. Similarity calculation refers to a numerical calculation process that evaluates the closeness between two permeability values. A multidimensional similarity evaluation method is used, which considers not only the difference between the permeability values themselves, but also the similarity of their corresponding test condition parameters. For example, the similarity calculation formula takes into account the absolute value of the difference between the permeability values, the difference in test frequency, the difference in test temperature, the difference in magnetic field strength, and other dimensions, and obtains a comprehensive similarity score through weighted averaging. The plurality of initial spatiotemporal near-neighbor material permeabilities are compared pairwise, and the similarity between any two permeability values is calculated. The predetermined threshold value is a pre-set upper limit of similarity according to the dispersion requirement, used to determine whether two permeability values are sufficiently similar. When the similarity between two permeability values is less than or equal to the predetermined threshold value, it indicates that the two permeability values have sufficient difference, meeting the dispersion requirement; when the similarity is greater than the predetermined threshold value, it indicates that the two permeability values are too similar and do not meet the dispersion condition.

[0066] The dispersion screening verification refers to a verification process of confirming whether the magnetic permeability of any two initial spatiotemporal neighbor materials meets the dispersion requirement through the similarity comparison. The similarity of any two initial spatiotemporal neighbor material magnetic permeabilities in the plurality of initial spatiotemporal neighbor material magnetic permeabilities is checked one by one. Only when all the pairwise similarities are less than or equal to the preset threshold, the two initial spatiotemporal neighbor material magnetic permeabilities can pass the dispersion screening verification. If the two initial spatiotemporal neighbor material magnetic permeabilities fail to pass the verification (i.e., there is a magnetic permeability pair with a similarity greater than the preset threshold), the random extraction will be performed again until the two initial spatiotemporal neighbor material magnetic permeabilities meeting the dispersion requirement are found. Through this iterative verification mechanism, it is ensured that the plurality of dispersion screening spatiotemporal neighbor material magnetic permeabilities finally selected have good dispersion and representativeness in the feature space.

[0067] After the above dispersion screening process, the plurality of dispersion screening spatiotemporal neighbor material magnetic permeabilities are obtained, which have appropriate differences between each other, avoid data redundancy, provide high-quality aggregation center points for subsequent aggregation operations, and ensure the effectiveness and reliability of the aggregation results.

[0068] Further, the plurality of dispersion screening spatiotemporal neighbor material magnetic permeabilities are used to aggregate the set of spatiotemporal neighbor material magnetic permeabilities, and a plurality of aggregated spatiotemporal neighbor material magnetic permeability sets are obtained, including:

[0069] S221, taking the plurality of dispersion screening spatiotemporal neighbor material magnetic permeabilities as centers and taking a preset aggregation bandwidth as a radius, a plurality of initial dispersion neighborhoods are constructed;

[0070] S222, the plurality of initial dispersion neighborhoods are diffused again based on one fourth of the preset aggregation bandwidth, and a plurality of diffused dispersion neighborhoods are obtained.

[0071] S223, it is judged whether the neighborhood density of the plurality of diffused dispersion neighborhoods is greater than or equal to the neighborhood density of the plurality of initial dispersion neighborhoods. If yes, the plurality of diffused dispersion neighborhoods are diffused again according to one fourth of the preset aggregation bandwidth until a preset diffusion number is met, and the plurality of aggregated spatiotemporal neighbor material magnetic permeability sets are obtained.

[0072] In a preferred embodiment, first, a plurality of initial dispersion neighborhoods are constructed, each centered on a dispersionally screened spatiotemporal proximate material permeability and with a pre-set aggregation bandwidth as the radius. The pre-set aggregation bandwidth refers to a threshold parameter used to define the range of permeability similarity, usually expressed in the form of a percentage, for example, 80%. Having the aggregation bandwidth as the radius means that the permeability values that meet the requirement of the aggregation bandwidth in terms of similarity with the center permeability are included in the same neighborhood range. The similarity of each dispersionally screened spatiotemporal proximate material permeability with all other permeability values in the set of spatiotemporal proximate material permeabilities is calculated. When the similarity of a permeability value with the center permeability meets the requirement of the pre-set aggregation bandwidth, the permeability value is added to the initial dispersion neighborhood centered on the center permeability. In this way, an initial neighborhood range containing similar permeability values is constructed for each dispersionally screened permeability, forming a plurality of initial dispersion neighborhoods.

[0073] Subsequently, the plurality of initial dispersion neighborhoods are diffused again based on one-fourth of the pre-set aggregation bandwidth, obtaining a plurality of diffused dispersion neighborhoods. The diffusion process refers to the process of expanding the neighborhood range by lowering the similarity requirement to include more potentially relevant permeability values. The one-fourth diffusion strategy of the pre-set aggregation bandwidth aims to avoid excessive exclusion of data. Specifically, if the pre-set aggregation bandwidth is 80%, one-fourth diffusion means reducing the similarity requirement by 20% (i.e., one-fourth of 80%), so that the permeability values that originally have a similarity between 60% and 80% and fail to enter the initial neighborhood have the opportunity to be included in the diffused neighborhood. This gradual diffusion strategy can maximize the use of available permeability data while ensuring data quality, avoiding the loss of useful data due to overly strict screening. The similarity range is recalculated for each initial dispersion neighborhood, reducing the similarity requirement by one-fourth of the original pre-set aggregation bandwidth, and re-screening the permeability values in the set of spatiotemporal proximate material permeabilities. The permeability values that meet the new similarity requirement are included in the corresponding diffused dispersion neighborhood, thereby obtaining a plurality of diffused dispersion neighborhoods.

[0074] Afterwards, it is judged whether the neighborhood density of the plurality of diffusion dispersion neighborhoods is greater than or equal to the neighborhood density of the plurality of initial dispersion neighborhoods. If yes, the plurality of diffusion dispersion neighborhoods is continuously diffused according to one fourth of the preset aggregation bandwidth until a preset diffusion number is met, and the plurality of aggregated spatiotemporal near-neighbor material magnetic permeability sets is obtained. The neighborhood density refers to the number of magnetic permeability values contained in the neighborhood, and is used to evaluate the data richness of the neighborhood. The number of magnetic permeability values in each initial dispersion neighborhood and the corresponding diffusion dispersion neighborhood is respectively counted, and the effectiveness of the diffusion operation is judged by comparing the neighborhood densities of the two. When the neighborhood density of the diffusion dispersion neighborhood is greater than or equal to the neighborhood density of the initial dispersion neighborhood, it indicates that the diffusion operation successfully incorporates more relevant magnetic permeability values, and improves the data density of the neighborhood. In this case, the next round of diffusion operation is continuously performed, and the similarity requirement is again reduced according to one fourth of the preset aggregation bandwidth, so as to further expand the neighborhood range. The above diffusion process is repeatedly executed until the preset diffusion number is met. The preset diffusion number is a preset upper limit of the diffusion round number according to the data characteristics and the aggregation quality requirement, and is used to prevent excessive diffusion from causing neighborhood boundary blurring and data quality degradation. By controlling the diffusion number, a balance between data utilization and aggregation quality can be achieved.

[0075] After the above aggregation process, the plurality of aggregated spatiotemporal near-neighbor material magnetic permeability sets is finally obtained. Each aggregated set is centered on a dispersion screening magnetic permeability, and contains other magnetic permeability values with different degrees of similarity to the center magnetic permeability. These aggregated sets not only ensure the relevance of the internal data, but also maximize the use of available data through the progressive diffusion strategy, thereby providing a structured and high-quality data basis for subsequent mapping and combing operations.

[0076] Further, the spatiotemporal synchronization combing result is distinguished into base class training data and adversarial class training data in stages to obtain a base class training data set and an adversarial class training data set, including:

[0077] S31, the number of the plurality of aggregated spatiotemporal near-neighbor material magnetic permeability test correlation data sets in the spatiotemporal synchronization combing result is counted, and a plurality of aggregation numbers is obtained.

[0078] S32, the aggregated spatiotemporal near-neighbor material magnetic permeability test correlation data set corresponding to the aggregation number less than or equal to the preset number threshold in the plurality of aggregation numbers, and the corresponding aggregated spatiotemporal near-neighbor material magnetic permeability set are added to the base class training data set.

[0079] S33, the aggregated spatiotemporal near-neighbor material magnetic permeability test correlation data set corresponding to the aggregation number greater than the preset number threshold in the plurality of aggregation numbers, and the corresponding aggregated spatiotemporal near-neighbor material magnetic permeability set are added to the adversarial class training data set.

[0080] In a preferred embodiment, first, the number of the plurality of aggregated spatiotemporal proximate material magnetic permeability test correlation data sets in the spatiotemporal synchronization combing result is traversed to obtain a plurality of aggregation numbers. Traversing statistics refers to a data statistical process of counting each aggregated spatiotemporal proximate material magnetic permeability test correlation data set in the spatiotemporal synchronization combing result one by one. All aggregated spatiotemporal proximate material magnetic permeability test correlation data sets in the spatiotemporal synchronization combing result are traversed and accessed, and the number of test correlation data entries contained in each aggregation set is counted. The aggregation number reflects the data richness and representativeness of each aggregation set. The aggregation set with a smaller number usually corresponds to a relatively sparse data distribution and a relatively simple feature interval of magnetic permeability; the aggregation set with a larger number corresponds to a dense data distribution and a complex and diverse feature interval of magnetic permeability. By counting the number of each aggregation set, a plurality of aggregation numbers are obtained, which provides a quantitative basis for distinguishing between base class data and adversarial class data in the subsequent stage. The aggregation number is used as an important index to evaluate the data complexity and training difficulty, and lays a data foundation for the implementation of the staged training strategy.

[0081] Subsequently, the aggregated spatiotemporal proximate material magnetic permeability test correlation data set corresponding to the aggregation number less than or equal to the preset number threshold in the plurality of aggregation numbers, and the corresponding aggregated spatiotemporal proximate material magnetic permeability set are added to the base class training data set. The preset number threshold is a number boundary preset according to the training data complexity evaluation, which is used as a judgment standard for distinguishing between base class data and adversarial class data. When the aggregation number of an aggregated spatiotemporal proximate material magnetic permeability test correlation data set is less than or equal to the preset number threshold, it indicates that the data scale of the aggregation set is relatively small, the data distribution is relatively simple, and the feature complexity is low. Such data usually has good data consistency and low training difficulty, and is suitable for being used as the base training data of the model. The base class training data is characterized in that it can provide stable and reliable learning samples for the model and help the model establish a basic understanding framework for the magnetic permeability prediction problem. The aggregated spatiotemporal proximate material magnetic permeability test correlation data set and the corresponding aggregated spatiotemporal proximate material magnetic permeability set that meet the condition are added to the base class training data set at the same time. This paired addition ensures the complete correspondence between the magnetic permeability value and its test correlation data, and provides a structurally complete training sample for subsequent model training.

[0082] Subsequently, the aggregated spatiotemporal neighbor material magnetic permeability test correlation data set corresponding to the number of aggregates greater than the preset number threshold and the corresponding aggregated spatiotemporal neighbor material magnetic permeability set are added to the adversarial training data set. When the number of aggregates of a certain aggregated spatiotemporal neighbor material magnetic permeability test correlation data set is greater than the preset number threshold, it indicates that the aggregate set contains a large amount of test correlation data, the data distribution is complex, and the feature diversity is high. Such data usually correspond to complex test environments, boundary conditions or magnetic permeability performance under special conditions, and belong to "difficult samples" in the training process. The adversarial training data is characterized by high feature complexity and training challenge. These data often contain more environmental variables, a wider parameter range, and more complex nonlinear relationships, and the model needs to have stronger learning ability and generalization ability to accurately predict. By adding these complex data as adversarial training data, the model's ability to adapt to complex situations can be improved during the adversarial fine-tuning phase. Similarly, the aggregated spatiotemporal neighbor material magnetic permeability test correlation data set and its corresponding aggregated spatiotemporal neighbor material magnetic permeability set that meet the conditions are added to the adversarial training data set in pairs to ensure the integrity and consistency of the data.

[0083] Through the above phased data differentiation process, the spatiotemporal synchronization results are classified according to data complexity and training difficulty, obtaining the base class training data set and the adversarial class training data set. The base class training data set provides a stable and reliable learning foundation for the initial training of the model, and the adversarial class training data set provides challenging optimization samples for the adversarial fine-tuning of the model. The two work together to provide layered training data support for building a high-performance magnetic permeability prediction model.

[0084] Further, according to the adversarial training data set, the hyperparameters of the initial mutual inductor core magnetic permeability predictor are fine-tuned to obtain a fine-tuned mutual inductor core magnetic permeability predictor, including:

[0085] S41, load the hyperparameters of the initial mutual inductor core magnetic permeability predictor to obtain a to-be-adjusted hyperparameter, wherein the to-be-adjusted hyperparameter includes a learning rate, a regularization coefficient, a batch size, and an iteration round;

[0086] S42, train the initial mutual inductor core magnetic permeability predictor using the adversarial training data set, and fine-tune the to-be-adjusted hyperparameters according to the training output result until convergence, to obtain the fine-tuned mutual inductor core magnetic permeability predictor.

[0087] In a preferred embodiment, first, the hyperparameters of the initial transformer core permeability predictor are loaded to obtain the to-be-adjusted hyperparameters, wherein the to-be-adjusted hyperparameters include a learning rate, a regularization coefficient, a batch size, and an iteration round. Hyperparameter loading refers to the process of extracting the training configuration parameters of the initial transformer core permeability predictor from the trained model, which control the training behavior and performance of the model. The learning rate refers to the step size of each parameter update in the gradient descent optimization process of the model, usually expressed in decimal form, such as 0.001 or 0.01. The learning rate directly affects the convergence speed and training stability of the model: too large a learning rate may cause the training process to oscillate or diverge, and too small a learning rate may cause the convergence speed to be too slow or fall into a local optimum. In the adversarial fine-tuning stage, since the adversarial training data has higher complexity, the learning rate needs to be adjusted to adapt to the new training difficulty. The regularization coefficient refers to the parameter used to control the model complexity and prevent overfitting. The size of the regularization coefficient determines the degree of penalty of the model on the parameter size: a larger regularization coefficient will produce a stronger regularization effect, which helps to improve the generalization ability of the model; a smaller regularization coefficient allows the model to fit more complex data patterns. The batch size refers to the number of training samples processed simultaneously in each training iteration, expressed in integer form, such as 32, 64, or 128. The batch size affects the stability and computational efficiency of the training: a larger batch size can provide more stable gradient estimates, but requires more computational resources; a smaller batch size has higher computational efficiency, but the gradient estimate has more noise. The iteration round refers to the number of complete cycles of the model training on the entire training data set, also known as the number of training periods. The iteration round determines the training length and learning sufficiency of the model: too few iteration rounds may result in insufficient model training, and too many iteration rounds may result in overfitting. In the adversarial fine-tuning process, the iteration round needs to be dynamically adjusted according to the characteristics of the adversarial data and the model convergence.

[0088] Subsequently, the initial mutual inductor core permeability predictor is trained using an adversarial training data set, and the to-be-adjusted hyperparameters are fine-tuned according to the training output results until convergence is obtained, and a fine-tuned mutual inductor core permeability predictor is obtained. The adversarial training process refers to a model optimization process of further training the initial mutual inductor core permeability predictor using adversarial training data with high complexity and challenges. The adversarial training data set is input into the initial mutual inductor core permeability predictor, the prediction result is calculated through forward propagation, then the gradient of the loss function is calculated through the back propagation algorithm, and the model parameters are updated. During the training process, the training output results, including training loss, validation loss, prediction accuracy, convergence speed and other key indicators, are continuously monitored. The training output results reflect the training state and performance of the model under the current hyperparameter configuration, providing feedback information for hyperparameter fine-tuning. The to-be-adjusted hyperparameters are adaptively fine-tuned according to the training output results. Specifically, when the training loss decreases slowly, the learning rate is appropriately increased to speed up the convergence; when overfitting occurs, the regularization coefficient is increased or the learning rate is reduced; when the training is unstable, the batch size is adjusted or the iteration strategy is modified. This feedback-based hyperparameter fine-tuning mechanism can ensure that the training effect of the model on adversarial data is optimized.

[0089] Convergence refers to the process of determining whether the training has reached a stable state by monitoring the trend of the training indicators. The convergence condition is set, such as the change amplitude of the loss function in consecutive iterations being less than a preset threshold, or the performance indicators on the validation set no longer significantly improving. When the convergence condition is met, the training process is stopped to ensure that the model has learned the features of adversarial data sufficiently and has avoided performance degradation caused by overtraining.

[0090] Through the above hyperparameter adversarial fine-tuning process, a fine-tuned mutual inductor core permeability predictor is obtained. The fine-tuned mutual inductor core permeability predictor has the prediction ability for complex working conditions and boundary conditions through the training of adversarial data on the basis of maintaining the original basic prediction ability, and has stronger robustness and adaptability, providing a high-quality pre-training model basis for subsequent transfer learning.

[0091] Further, the fine-tuned mutual inductor core permeability predictor is subjected to transfer learning to obtain a target mutual inductor core permeability predictor, comprising:

[0092] S51, performing permeability testing on the target mutual inductor core for a preset number of times to obtain a target mutual inductor core permeability test correlation data set and a target mutual inductor core permeability set;

[0093] S52, based on the target transformer core permeability test correlation data set and the target transformer core permeability set, the fine-tuning transformer core permeability predictor is migrated to learn to update the hyperparameters with the optimization goal of minimizing prediction error, and the target transformer core permeability predictor is obtained.

[0094] In a preferred embodiment, first, the target transformer core is subjected to permeability test for a preset number of tests, and a target transformer core permeability test correlation data set and a target transformer core permeability set are obtained. The preset number of tests refers to the number of permeability tests that are pre-set according to the data requirements of transfer learning and test cost. Since the magnetic permeability data of the target transformer core material is relatively scarce, it is necessary to obtain the basic magnetic permeability characteristic data of the material through limited experimental tests. The permeability test process is strictly carried out according to the standard test protocol to ensure the accuracy and reliability of the test data. During the test process, different test condition parameters are controlled, including magnetic field strength, test frequency, environmental temperature, magnetic flux density and other key parameters, to obtain the magnetic permeability performance of the target transformer core under various working conditions.

[0095] In each test, the magnetic permeability value and the corresponding test correlation data are recorded simultaneously. The target transformer core permeability set contains the magnetic permeability values of the target transformer core under different test conditions, which directly reflect the magnetic performance characteristics of the target transformer core. The target transformer core permeability test correlation data set contains complete test environment information corresponding to each magnetic permeability value, such as test temperature, humidity, magnetic field strength, frequency setting, test equipment parameters, etc.

[0096] Through the preset number of tests, a basic magnetic permeability database of the target transformer core material is obtained. Although the number of these data is relatively limited, they are highly targeted and representative, providing a real data basis for the target domain for subsequent transfer learning.

[0097] Subsequently, based on the target transformer core magnetic permeability test correlation data set and the target transformer core magnetic permeability set, the fine-tuned transformer core magnetic permeability predictor is subjected to transfer learning, hyperparameter updating is performed with the optimization target of minimizing prediction error, and the target transformer core magnetic permeability predictor is obtained. The transfer learning process refers to the model adaptation process of transferring the knowledge of the fine-tuned transformer core magnetic permeability predictor trained on the near-neighbor material to the target transformer core, keeping the bottom feature extraction capability of the fine-tuned transformer core magnetic permeability predictor unchanged, and focusing on adjusting the high-level decision network to adapt to the specific magnetic permeability characteristics of the target transformer core. The target transformer core magnetic permeability test correlation data set is taken as the input feature, and the target transformer core magnetic permeability set is taken as the true label to construct the training data pair. In the transfer learning process, the small batch gradient descent algorithm is adopted, the predicted value is calculated through forward propagation, and then the prediction error between the predicted value and the true magnetic permeability value is calculated. The prediction error minimization refers to the optimization process of continuously optimizing and adjusting the model parameters so that the prediction error of the model on the target transformer core data reaches a minimum value. For example, the mean square error loss function is used as the optimization target, which can effectively measure the difference between the predicted value and the true value. The calculation formula of the loss function comprehensively considers the prediction errors of all test samples, and the prediction accuracy of the model is improved by minimizing the overall prediction error. In the hyperparameter updating process, an adaptive learning rate adjustment strategy is adopted. Since the target transformer core data is relatively small, a relatively small learning rate is set to avoid the loss of learned knowledge caused by the drastic change of model parameters. At the same time, the early stopping mechanism and regularization constraint can be introduced to prevent the model from overfitting on limited target data. The transfer learning process adopts a progressive training strategy. First, the model is fine-tuned using a small learning rate, and the convergence of the model on the target data is observed. With the training, the training parameters are dynamically adjusted according to the convergence speed and prediction accuracy to ensure the stability and effectiveness of the transfer learning process.

[0098] Through the above transfer learning process, the target transformer core magnetic permeability predictor is obtained. The target transformer core magnetic permeability predictor is optimized specifically for the target transformer core material, which not only inherits the powerful prediction ability and adaptability to complex working conditions of the fine-tuned predictor, but also obtains precise prediction ability for the magnetic permeability characteristics of the specific material through training of the target transformer core data, realizes effective conversion from a general prediction model to a target transformer core special prediction model, and provides a high-precision and high-reliability technical solution for the magnetic permeability prediction of the target transformer core.

[0099] Embodiment two, as shown in Figure 2 the same inventive concept as the transformer core magnetic permeability prediction method provided in embodiment one, the present application embodiment also provides a transformer core magnetic permeability prediction system combining transfer learning, comprising:

[0100] The data acquisition module 11 is configured to acquire material basic information of a target mutual inductor core, search for neighboring materials based on the material basic information, search for permeability test data of the neighboring materials at different regions and different time points according to a preset test correlation index, and obtain a time-space neighboring material permeability test correlation data set and a time-space neighboring material permeability set.

[0101] The time-space combing module 12 is configured to perform time-space synchronization combing on the time-space neighboring material permeability set and the time-space neighboring material permeability test correlation data set, and obtain a time-space synchronization combing result.

[0102] The data classification module 13 is configured to distinguish base class training data and adversarial class training data from the time-space synchronization combing result, and obtain a base class training data set and an adversarial class training data set.

[0103] The model training module 14 is configured to train an initial mutual inductor core permeability predictor by using the base class training data set, perform hyperparameter adversarial fine-tuning on the trained initial mutual inductor core permeability predictor according to the adversarial class training data set, and obtain a fine-tuned mutual inductor core permeability predictor.

[0104] The migration prediction module 15 is configured to perform migration learning on the fine-tuned mutual inductor core permeability predictor, obtain a target mutual inductor core permeability predictor, and perform target mutual inductor core permeability prediction by using the target mutual inductor core permeability predictor.

[0105] Further, the preset test correlation index includes magnetic field strength, magnetic flux density, frequency, temperature, material characteristics and process parameters.

[0106] Further, the execution steps of the time-space combing module 12 include:

[0107] The time-space neighboring material permeability set is subjected to dispersion screening to determine a plurality of dispersion screening time-space neighboring material permeabilities.

[0108] The time-space neighboring material permeability set is aggregated based on the plurality of dispersion screening time-space neighboring material permeabilities to obtain a plurality of aggregated time-space neighboring material permeability sets.

[0109] The time-space neighboring material permeability test correlation data set is mapped and combed based on the plurality of aggregated time-space neighboring material permeability sets to obtain a plurality of aggregated time-space neighboring material permeability test correlation data sets, the plurality of aggregated time-space neighboring material permeability test correlation data sets and the plurality of aggregated time-space neighboring material permeability sets are mapped and associated, and the time-space synchronization combing result is obtained.

[0110] Further, the execution steps of the space-time combing module 12 further include:

[0111] Randomly extracting a plurality of initial space-time near-neighbor material permeabilities from the space-time near-neighbor material permeability set;

[0112] Determining whether the similarity between any two initial space-time near-neighbor material permeabilities is less than or equal to a preset threshold value, and if so, verifying the plurality of initial space-time near-neighbor material permeabilities through dispersion screening to obtain a plurality of dispersion screening space-time near-neighbor material permeabilities.

[0113] Further, the execution steps of the space-time combing module 12 further include:

[0114] Centering the plurality of dispersion screening space-time near-neighbor material permeabilities and constructing a plurality of initial dispersion neighborhoods with a preset aggregation bandwidth as the radius;

[0115] Further diffusing the plurality of initial dispersion neighborhoods based on one-fourth of the preset aggregation bandwidth to obtain a plurality of diffused dispersion neighborhoods;

[0116] Determining whether the neighborhood density of the plurality of diffused dispersion neighborhoods is greater than or equal to the neighborhood density of the plurality of initial dispersion neighborhoods, and if so, continuing to diffuse the plurality of diffused dispersion neighborhoods based on one-fourth of the preset aggregation bandwidth until a preset diffusion number is met to obtain the plurality of aggregated space-time near-neighbor material permeability sets.

[0117] Further, the execution steps of the data classification module 13 include:

[0118] Iteratively counting the number of a plurality of aggregated space-time near-neighbor material permeability test correlation data sets in the space-time synchrony combing result to obtain a plurality of aggregation numbers;

[0119] Adding the aggregated space-time near-neighbor material permeability test correlation data set corresponding to the aggregation number less than or equal to a preset number threshold value and the corresponding aggregated space-time near-neighbor material permeability set to the base class training data set;

[0120] Adding the aggregated space-time near-neighbor material permeability test correlation data set corresponding to the aggregation number greater than the preset number threshold value and the corresponding aggregated space-time near-neighbor material permeability set to the adversarial class training data set.

[0121] Further, the execution steps of the model training module 14 include:

[0122] Loading the hyperparameters of the initial mutual inductor core permeability predictor to obtain a to-be-adjusted hyperparameter, wherein the to-be-adjusted hyperparameter includes a learning rate, a regularization coefficient, a batch size, and an iteration round.

[0123] The initial mutual inductor core permeability predictor is trained by using the set of adversarial training data, and the to-be-adjusted hyperparameters are fine-tuned according to a training output result until convergence is achieved, so as to obtain the fine-tuned mutual inductor core permeability predictor.

[0124] Further, the execution steps of the migration prediction module 15 include:

[0125] The target mutual inductor core is subjected to permeability tests for a preset number of times, so as to obtain a set of target mutual inductor core permeability test correlation data and a set of target mutual inductor core permeability data;

[0126] Based on the set of target mutual inductor core permeability test correlation data and the set of target mutual inductor core permeability data, the fine-tuned mutual inductor core permeability predictor is subjected to migration learning, hyperparameter updating is performed with the optimization target of minimizing prediction error, and a target mutual inductor core permeability predictor is obtained.

[0127] It should be noted that in the above embodiments, the description of each embodiment has its own emphasis, and the parts not described in detail in a certain embodiment can be referred to the related description of other embodiments.

[0128] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0129] The present application is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of flows and / or blocks in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded computer, or other programmable data processing devices to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices produce a device that implements the functions specified in the flowcharts and / or block diagrams. Figure 1 The device that implements the functions specified in one flow or multiple flows and / or blocks Figure 1 The device that implements the functions specified in one flow or multiple flows and / or blocks

[0130] These computer program instructions can also be stored in a computer- readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instructions which implement the Figure 1 function specified in the flow or flows and / or blocks Figure 1 of the block or blocks.

[0131] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions that are executed on the computer or other programmable apparatus provide steps for implementing the Figure 1 function specified in the flow or flows and / or blocks Figure 1 of the block or blocks.

[0132] Although preferred embodiments of the application have been described, those skilled in the art will recognize that additional modifications and variations can be made thereto without departing from the spirit and scope of the application.

[0133] It will be apparent to those skilled in the art that various modifications and variations can be made to the present application without departing from the spirit or scope of the application. Thus, it is intended that the present application cover modifications and variations of this application provided they come within the scope of the appended claims and their equivalents.

Claims

1. A method for predicting the magnetic permeability of a transformer core in combination with transfer learning, characterized by, The method comprises: obtaining the material basic information of the target mutual inductor core, retrieving the neighboring materials based on the material basic information, retrieving the magnetic permeability test data of the neighboring materials at different regions and different time points according to a preset test correlation index, obtaining a time-space neighboring material magnetic permeability test correlation data set and a time-space neighboring material magnetic permeability set; performing time-space synchronization carding on the time-space neighboring material magnetic permeability set and the time-space neighboring material magnetic permeability test correlation data set to obtain a time-space synchronization carding result; distinguishing the time-space synchronization carding result into base class training data and adversarial class training data to obtain a base class training data set and an adversarial class training data set; training an initial mutual inductor core magnetic permeability predictor using the base class training data set, and performing hyperparameter adversarial fine-tuning on the trained initial mutual inductor core magnetic permeability predictor according to the adversarial class training data set to obtain a fine-tuned mutual inductor core magnetic permeability predictor; performing transfer learning on the fine-tuned mutual inductor core magnetic permeability predictor to obtain a target mutual inductor core magnetic permeability predictor, and using the target mutual inductor core magnetic permeability predictor to predict the magnetic permeability of the target mutual inductor core; wherein distinguishing the time-space synchronization carding result into base class training data and adversarial class training data to obtain a base class training data set and an adversarial class training data set comprises: traversing and counting the number of a plurality of aggregated time-space neighboring material magnetic permeability test correlation data sets in the time-space synchronization carding result to obtain a plurality of aggregation quantities, wherein each aggregation quantity is the number of test correlation data entries contained in each aggregated time-space neighboring material magnetic permeability test correlation data set; adding the aggregated time-space neighboring material magnetic permeability test correlation data set corresponding to the aggregation quantity less than or equal to a preset quantity threshold in the plurality of aggregation quantities, and the corresponding aggregated time-space neighboring material magnetic permeability set, into the base class training data set; adding the aggregated time-space neighboring material magnetic permeability test correlation data set corresponding to the aggregation quantity greater than the preset quantity threshold in the plurality of aggregation quantities, and the corresponding aggregated time-space neighboring material magnetic permeability set, into the adversarial class training data set.

2. The method for predicting the magnetic permeability of a transformer core incorporating transfer learning according to claim 1, wherein, The preset test correlation index includes magnetic field strength, magnetic flux density, frequency, temperature, material properties, and process parameters.

3. The method for predicting the magnetic permeability of a transformer core incorporating transfer learning according to claim 1, wherein, performing time-space synchronization carding on the time-space neighboring material magnetic permeability set to obtain a time-space synchronization carding result, comprising: performing dispersion screening on the time-space neighboring material magnetic permeability set to determine a plurality of dispersion screening time-space neighboring material magnetic permeabilities; aggregating the time-space neighboring material magnetic permeability set based on the plurality of dispersion screening time-space neighboring material magnetic permeabilities to obtain a plurality of aggregated time-space neighboring material magnetic permeability sets; Mapping and sorting the spatiotemporal neighbor material magnetic permeability test correlation data set based on the plurality of aggregated spatiotemporal neighbor material magnetic permeability sets, obtaining a plurality of aggregated spatiotemporal neighbor material magnetic permeability test correlation data sets, and mapping and correlating the plurality of aggregated spatiotemporal neighbor material magnetic permeability test correlation data sets and the plurality of aggregated spatiotemporal neighbor material magnetic permeability sets to obtain the spatiotemporal synchronization sorting result.

4. The method for predicting the magnetic permeability of a transformer core incorporating transfer learning according to claim 3, wherein, Dispersive screening the spatiotemporal neighbor material magnetic permeability set to determine a plurality of dispersively screened spatiotemporal neighbor material magnetic permeabilities, including: Randomly extracting a plurality of initial spatiotemporal neighbor material magnetic permeabilities from the spatiotemporal neighbor material magnetic permeability set; Determining whether the similarity between any two initial spatiotemporal neighbor material magnetic permeabilities of the plurality of initial spatiotemporal neighbor material magnetic permeabilities is less than or equal to a preset threshold, and if so, verifying the plurality of initial spatiotemporal neighbor material magnetic permeabilities through dispersive screening to obtain the plurality of dispersively screened spatiotemporal neighbor material magnetic permeabilities.

5. The method for predicting the magnetic permeability of a transformer core incorporating transfer learning according to claim 4, wherein, Aggregating the spatiotemporal neighbor material magnetic permeability set based on the plurality of dispersively screened spatiotemporal neighbor material magnetic permeabilities to obtain a plurality of aggregated spatiotemporal neighbor material magnetic permeability sets, including: Centering the plurality of dispersively screened spatiotemporal neighbor material magnetic permeabilities and constructing a plurality of initial dispersive neighborhoods with a preset aggregation bandwidth as the radius; Diffusing the plurality of initial dispersive neighborhoods based on one fourth of the preset aggregation bandwidth to obtain a plurality of diffused dispersive neighborhoods; Determining whether the neighborhood density of the plurality of diffused dispersive neighborhoods is greater than or equal to the neighborhood density of the plurality of initial dispersive neighborhoods, and if so, continuing to diffuse the plurality of diffused dispersive neighborhoods based on one fourth of the preset aggregation bandwidth until a preset diffusion number is met to obtain the plurality of aggregated spatiotemporal neighbor material magnetic permeability sets.

6. The method for predicting the magnetic permeability of a transformer core incorporating transfer learning according to claim 1, wherein, Hyperparameter adversarial fine-tuning an initial mutual inductor core magnetic permeability predictor trained based on an adversarial training data set to obtain a fine-tuned mutual inductor core magnetic permeability predictor, including: Loading the hyperparameters of the initial mutual inductor core magnetic permeability predictor to obtain to-be-adjusted hyperparameters, wherein the to-be-adjusted hyperparameters include learning rate, regularization coefficient, batch size, and iteration round; Training the initial mutual inductor core magnetic permeability predictor using the adversarial training data set and fine-tuning the to-be-adjusted hyperparameters based on the training output result until convergence is achieved to obtain the fine-tuned mutual inductor core magnetic permeability predictor.

7. The method for predicting the magnetic permeability of a transformer core incorporating transfer learning according to claim 1, wherein, Transfer learning the fine-tuned mutual inductor core magnetic permeability predictor to obtain a target mutual inductor core magnetic permeability predictor, including: Performing a magnetic permeability test on a target mutual inductor core for a preset number of times to obtain a target mutual inductor core magnetic permeability test correlation data set and a target mutual inductor core magnetic permeability set; Based on the target mutual inductor core magnetic permeability test correlation data set and the target mutual inductor core magnetic permeability set, transfer learning the fine-tuned mutual inductor core magnetic permeability predictor to update the hyperparameters with the optimization goal of minimizing prediction error to obtain the target mutual inductor core magnetic permeability predictor.

8. A system for predicting the magnetic permeability of a transformer core incorporating transfer learning, characterized by, A system for implementing the transformer core permeability prediction method combined with transfer learning as claimed in any one of claims 1 to 7, the system comprising: a data acquisition module configured to acquire material basic information of a target transformer core, retrieve a near neighbor material based on the material basic information, and retrieve permeability test data of the near neighbor material at different regions and different time points according to a preset test correlation index, to obtain a spatiotemporal near neighbor material permeability test correlation data set and a spatiotemporal near neighbor material permeability set; a spatiotemporal combing module configured to perform spatiotemporal synchronization combing on the spatiotemporal near neighbor material permeability set and the spatiotemporal near neighbor material permeability test correlation data set, to obtain a spatiotemporal synchronization combing result; a data classification module configured to distinguish base class training data and adversarial class training data from the spatiotemporal synchronization combing result, to obtain a base class training data set and an adversarial class training data set; a model training module configured to train an initial transformer core permeability predictor using the base class training data set, and perform hyperparameter adversarial fine-tuning on the trained initial transformer core permeability predictor according to the adversarial class training data set, to obtain a fine-tuned transformer core permeability predictor; a transfer prediction module configured to perform transfer learning on the fine-tuned transformer core permeability predictor, to obtain a target transformer core permeability predictor, and perform target transformer core permeability prediction using the target transformer core permeability predictor.

Images