Transformer structure optimization method and system based on generative adversarial network

Abstract

The invention provides a mutual inductor structure optimization method and system based on a generative adversarial network, and relates to the technical field of generative adversarial networks, and the method comprises the steps: carrying out the potential excitation training of the generative adversarial network; pre-optimizing the structure of the mutual inductor by using the generative adversarial network after potential excitation training, and mining the optimization potential of the mutual inductor in the pre-optimization process; performing potential transformation training on the generative adversarial network for the mined optimization potential; and performing formal optimization on the mutual inductor structure by using the generative adversarial network after potential transformation training. According to the method, the potential of the generative adversarial network for structural optimization of the mutual inductor is greatly improved, and the optimization accuracy, effect and efficiency are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of generative adversarial network, in particular to a mutual inductor structure optimization method and system based on generative adversarial network. BACKGROUND

[0002] As a key sensing device in power system, the structure design of mutual inductor directly determines the core performance indicators such as measurement accuracy, linearity, bandwidth, anti-interference ability and miniaturization. The traditional mutual inductor structure optimization method mainly relies on the experience design of engineers, physical prototype trial production and iterative optimization based on numerical simulation technology such as finite element analysis (FEA). These methods have significant limitations such as strong experience dependence and high cost.

[0003] In recent years, machine learning, especially deep learning technology, has shown great potential in engineering optimization field. As a powerful generative model, generative adversarial network (GAN) is tried to be applied to the field of structure optimization design because of its ability to learn complex data distribution and generate novel samples. The basic idea is to use the generator network to generate candidate structures, and the discriminator network to evaluate the structure performance, and through the adversarial training of the two to guide the generator to output structures with better performance.

[0004] However, directly applying existing GAN models designed for general image or data generation tasks to the structure optimization problem of mutual inductor with complex physical constraints (such as electromagnetic field distribution, material properties, geometric precision) often does not meet expectations. The core problem faced by existing technologies is: 1. Insufficient potential mining: The training target of standard GAN mainly focuses on the similarity (such as fidelity) between generated samples and real data distribution, and does not aim at specific physical field performance optimization targets (such as sensitivity, linearity, bandwidth of mutual inductor). The network is limited to general generation ability in the initial stage, and the optimization potential inherent in the network for specific physical problems is not effectively activated and guided.

[0005] 2. Insufficient optimization ability transformation: Even if the network shows certain optimization tendency (i.e. potential) during training, existing methods lack effective mechanisms to transform these abstract and scattered potential into stable and reliable specific structure optimization ability. And the training process is usually an end-to-end one-time process, lacking a special stage of capturing, analyzing and solidifying potential.

[0006] 3. Optimization efficiency and accuracy are limited: due to the lack of potential excitation and transformation, the existing GAN-based optimization method often has problems such as unclear optimization direction, slow convergence speed, large result fluctuation, and easy to fall into suboptimal solution when applied to the design of mutual inductor structure. The accuracy, stability and efficiency of the optimization process cannot meet the ideal requirements of engineering application, and the theoretical advantages of GAN in exploring high-dimensional design space and discovering non-intuitive optimal structure cannot be fully utilized.

[0007] There is an urgent need for an innovative technical solution that can deeply stimulate the inherent potential of the generative adversarial network in the optimization of the mutual inductor structure and effectively transform this abstract potential into actual and efficient optimization capability, thereby overcoming the key defects of the prior art such as insufficient optimization potential mining, insufficient capability transformation, and limited efficiency and accuracy, and ultimately achieving significant improvement in the performance of the mutual inductor. SUMMARY

[0008] The embodiments of the present application provide a mutual inductor structure optimization method and system based on a generative adversarial network to solve the technical problems in the background art.

[0009] In a first aspect, the embodiments of the present application provide a mutual inductor structure optimization method based on a generative adversarial network, which comprises: Potential excitation training of the generative adversarial network; Pre-optimization of the mutual inductor structure using the generative adversarial network after potential excitation training, and mining the optimization potential in the pre-optimization process; Potential transformation training of the generative adversarial network for the mined optimization potential; Formal optimization of the mutual inductor structure using the generative adversarial network after potential transformation training.

[0010] Optionally, the potential excitation training of the generative adversarial network comprises: Embedding the following rules into the control layer of the game process of the generator and the discriminator in the generative adversarial network: When the generator generates the mutual inductor optimization structure for the i-th time, a random generated disturbance with a strength weight of j is applied to it; where j = j0+ a i, j0 is the disturbance strength benchmark weight, a is the strength gain coefficient, and a > 0; When the discriminator discriminates the mutual inductor optimization structure generated by the generator for the i-th time, a rational discrimination constraint with a strength weight of k is applied to it; where k = k0+ β i, k0 is the constraint strength benchmark weight, β is the strength gain coefficient, and β > 0.

[0011] Optionally, the mining of the optimization potential in the pre-optimization process comprises: The final transformer optimization structure output by the potential excitation trained generative adversarial network in the pre-optimization process is compared with other transformer optimization structures output by the generative adversarial network in history to locate innovative optimization features in the final transformer optimization structure; The related process of generating innovative optimization features by the generator and the discriminator in the potential excitation trained generative adversarial network is traced back; The game behavior change is analyzed to determine the game behavior change; The game behavior change is classified as autonomous behavior to obtain a distribution of autonomous behavior; The autonomous optimization intention represented by the distribution of autonomous behavior is reversely deduced; Based on the autonomous optimization intention, the optimization potential is determined.

[0012] Optionally, the game behavior change is classified as autonomous behavior to obtain a distribution of autonomous behavior, comprising: For each behavior in the game behavior change, the logical relationship between the behavior and other behaviors in the preset change range before and after the behavior is analyzed, a first feature vector of the logical relationship is constructed, the first feature vector is matched with a first standard vector representing that the behavior belongs to the autonomous behavior category, and if the matching degree exceeds a preset first threshold, the game behavior is selected as a first candidate autonomous behavior; The game behavior change is divided into sliding windows using a preset step length time window; For each behavior window obtained by sliding division, if the window contains a first candidate autonomous behavior, the behavior distribution of the contained first candidate autonomous behavior in the window is analyzed, a second feature vector of the behavior distribution is constructed, the second feature vector is matched with a second standard vector representing that the other behaviors in the window except the contained first candidate autonomous behavior belong to the autonomous behavior category, and if the matching degree exceeds a preset second threshold, the other behaviors are selected as second candidate autonomous behaviors; All first candidate autonomous behaviors and second candidate autonomous behaviors are merged to determine the final autonomous behaviors; The distribution of autonomous behaviors in the game behavior change is determined and is taken as the distribution of autonomous behaviors.

[0013] Optionally, the autonomous optimization intention represented by the distribution of autonomous behaviors is reversely deduced, comprising: The distribution of autonomous behaviors is input into an intention speculation knowledge graph for propagation; Based on the associated reasoning results triggered after the propagation of the distribution of autonomous behaviors, the corresponding autonomous optimization intention is deduced.

[0014] Optionally, the potential transformation training of the generative adversarial network is performed based on the mined optimization potential, comprising: construct a potential feature vector of the excavated optimization potential; map and inject the potential feature vector into a latent space of a generator of the generative adversarial network to form a potential conditioned latent input; train the generative adversarial network by using a dataset corresponding to a target field containing the optimization potential; wherein the generator receives the potential conditioned latent input to generate potential transformed data with the optimization potential characteristics; and the discriminator receives real data samples and the potential transformed data to determine the authenticity and whether the potential transformed data meets the optimization characteristics represented by the potential feature vector; During the training process, the potential matching loss between the potential transformed data and the potential feature vector is calculated, and the parameters of the generator are updated by back propagation, so that the generator outputs data matching the optimization potential, and the discriminator and the generator are optimized until the generative adversarial network reaches a Nash equilibrium state, so that the generator can stably output data meeting the optimization potential at this time.

[0015] Optionally, the generative adversarial network trained by the potential transformation is used to formally optimize the mutual inductor structure, comprising: inputting the mutual inductor structure into the generative adversarial network trained by the potential transformation for formal optimization.

[0016] Optionally, after the generative adversarial network trained by the potential transformation is used to formally optimize the mutual inductor structure, it further comprises: interacting with the user during the formal optimization process.

[0017] Optionally, the interaction with the user during the formal optimization process comprises: providing human-computer cooperation for the user and the generative adversarial network trained by the potential transformation to further optimize the mutual inductor structure.

[0018] In a second aspect, the embodiments of the present application provide a mutual inductor structure optimization system based on a generative adversarial network, comprising: a potential excitation training module for potential excitation training of the generative adversarial network; an optimization potential excavation module for pre-optimizing the mutual inductor structure by using the generative adversarial network trained by the potential excitation, and excavating the optimization potential of the mutual inductor structure during the pre-optimization process; a potential transformation training module for potential transformation training of the generative adversarial network with respect to the excavated optimization potential; a mutual inductor structure optimization module for formally optimizing the mutual inductor structure by using the generative adversarial network trained by the potential transformation.

[0019] The present application has the following advantages: The application firstly performs potential excitation training on the generative adversarial network to excite the potential of the transformer structure optimization; then performs pre-optimization on the transformer structure by using the generative adversarial network after the training, and excavates the optimization potential in the pre-optimization process; the excitation and excavation are cooperated to improve the optimization performance and capture efficiency; then performs potential transformation training on the generative adversarial network to transform the optimization potential into actual transformer structure optimization capability; finally, performs formal optimization on the transformer structure by using the generative adversarial network after the potential transformation training. The potential of the generative adversarial network used for the transformer structure optimization is greatly improved, and the optimization accuracy, effect and efficiency are significantly improved.

[0020] Other features and advantages of the present application will be set forth in the following description, and in part will become apparent to those skilled in the art from the description, or can be learned by practice of the present application. The objects and other advantages of the present application can be realized and attained by means of the instrumentalities particularly pointed out in the written description and claims hereof.

[0021] The technical solutions of the present application will be further described in detail below with the help of the accompanying drawings and embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0022] The accompanying drawings are included to provide a further understanding of the present application, and constitute a part of the specification, illustrate the present application together with the embodiments, and are used to explain the present application, and do not constitute a limitation on the present application. In the drawings: Figure 1 A schematic diagram of the transformer structure optimization method based on the generative adversarial network in the embodiments of the present application; Figure 2 A schematic diagram of the transformer structure optimization system based on the generative adversarial network in the embodiments of the present application. DETAILED DESCRIPTION

[0023] The preferred embodiments of the present application will be described below in conjunction with the accompanying drawings, and it should be understood that the preferred embodiments described herein are only used to illustrate and explain the present application, and do not limit the present application.

[0024] The research and development idea of the present application is to unlock the inherent ability of the generative adversarial network in the transformer structure optimization through potential excitation training, deeply excavate the optimization potential contained therein by using the pre-optimization process, implement directional potential transformation training on the excavated potential to solidify the abstract potential into the actual optimization capability of the network, and finally generate high-performance transformer structure through formal optimization, form a whole-chain optimization mechanism from potential excitation to capability solidification and then to structure generation, and realize the optimization and improvement of the performance of the transformer.

[0025] The embodiments of the present application provide a transformer structure optimization method based on a generative adversarial network, as shown in Figure 1 , comprising: S1, potential excitation training is performed on the generative adversarial network.

[0026] In this step, potential excitation training refers to training to excite the generative adversarial network to exhibit the potential of transformer structure optimization.

[0027] S2, the generative adversarial network after potential excitation training is used to pre-optimize the transformer structure, and the optimization potential is excavated in the pre-optimization process.

[0028] In this step, pre-optimization refers to the process of using the generative adversarial network after potential excitation training to preliminarily optimize the transformer structure. Optimization potential refers to the potential of transformer structure optimization exhibited by the generative adversarial network after potential excitation training in the pre-optimization process.

[0029] S3, potential transformation training is performed on the generative adversarial network for the excavated optimization potential.

[0030] In this step, potential transformation training refers to training to enable the generative adversarial network to transform the optimization potential into actual transformer structure optimization capability.

[0031] S4, the generative adversarial network after potential transformation training is used to formally optimize the transformer structure.

[0032] In this step, formal optimization refers to the process of using the generative adversarial network after potential transformation training to finally optimize the transformer structure, and the optimization result is displayed to the user.

[0033] In summary, steps S1 to S4 first perform potential excitation training on the generative adversarial network to excite it to exhibit the potential of transformer structure optimization; then use the generative adversarial network after this training to pre-optimize the transformer structure, and excavate its optimization potential in the pre-optimization process; excitation and excavation cooperate to improve optimization performance capture efficiency; then perform potential transformation training on the generative adversarial network to transform its optimization potential into actual transformer structure optimization capability; finally, the generative adversarial network after potential transformation training is used to formally optimize the transformer structure. The potential of the generative adversarial network for transformer structure optimization is greatly improved, and the optimization accuracy, effect and efficiency are significantly improved.

[0034] In some embodiments, the S1, potential excitation training is performed on the generative adversarial network, comprising: S11, embedding rules into the control layer of the game process between the generator and the discriminator in the generative adversarial network.

[0035] In this step, the generative adversarial network is composed of a generator and a discriminator. The generator learns to generate new samples (such as transformer structures) that simulate the distribution of real data, while the discriminator distinguishes between generated samples and real samples. Through adversarial training, both the generator and the discriminator continuously improve the generation quality, forming a closed loop of game play from generation to discrimination to improvement. The control layer can be set to control the game process of the two.

[0036] The above rules include: S111, when the generator generates the transformer optimization structure for the i-th time, a random generated disturbance with intensity weight j is applied to it; where j = j0 + a • i, j0 is the disturbance intensity reference weight, a is the intensity gain coefficient, and a > 0.

[0037] In this step, the random generated disturbance refers to the interference that makes the generator more random when generating the transformer optimization structure. The intensity weight represents the intensity of the random generated disturbance. As i increases from 1 to N (N is the total number of times the generator generates the transformer optimization structure), the corresponding intensity weight is calculated through the iterative incremental formula. The disturbance intensity reference weight and the intensity gain coefficient in the formula can be set according to actual needs, and each determines the reference and gain size of the iterative increment.

[0038] Specifically, a specific example is provided: the random generated disturbance is defined as a mechanism for injecting Gaussian noise (random fluctuations conforming to a normal distribution) into the latent space vector of the generator (the random input signal received by the generator), which is used to force the exploration of the transformer structure design space in the non-convergent area (the design space that the model has not been fully optimized); the variance scaling factor of the noise distribution of the Gaussian noise (the coefficient controlling the amplitude of the noise fluctuation) is defined as the intensity weight; the exploration ability (the ability to find new structures) of the generator and the convergence stability (the ability to maintain effective optimization) are dynamically balanced through iterative increment.

[0039] For example: the control layer superimposes a disturbance term Δz ~ N(0, j • σ^2I) (subject to a normal distribution with mean 0 and covariance j • σ^2 of the unit matrix I) before the generator input noise vector z ∈ R^d (d-dimensional real random vector). Where σ^2 is the reference noise variance (such as σ = 0.2, indicating the basic noise intensity); set j0 = 0.3, a = 0.05, so that the intensity weight j = 0.35 (noise variance is expanded by 0.35 times) at the first iteration, and the intensity weight j = 1.3 (noise variance is expanded by 1.3 times) at the 20th iteration. This mechanism forces the generator to focus on the existing structure distribution (such as a ring-shaped core) in the early training period (when i is small), and gradually explore unconventional topologies (such as a multi-gap segmented core) in the later period (when i increases), avoiding falling into a local optimal solution (only finding a suboptimal structure).

[0040] S112, when the discriminator discriminates the mutual inductor optimization structure generated by the generator for the i-th time, a rational discrimination constraint with a strength weight of k is applied to it; wherein k=k0+β⋅i, k0 is a constraint strength baseline weight, β is a strength gain coefficient, and β>0.

[0041] In this step, the rational discrimination constraint refers to a constraint that makes the discriminator more rational when it discriminates the mutual inductor optimization structure generated by the generator. The strength weight represents the strength of the rational discrimination constraint. As i increases from 1 to N, the corresponding strength weight is calculated by the iterative incremental formula. In the formula, the constraint strength baseline weight and the strength gain coefficient can be set in advance according to actual needs, and each determines the baseline and gain size of the iterative increment.

[0042] Specifically, a specific example is provided: the rational discrimination constraint is defined as a hard condition limit based on electromagnetic physical laws (equations describing the basic laws of electromagnetic fields), and the discriminator is forced to pay attention to the effective structure that meets Maxwell's equations (core control equations of electromagnetic fields) through a penalty term (an additional constraint term in the loss function); the strength weight is defined as the proportion of the physical loss term in the total loss of the discriminator (the strength coefficient of the physical constraint), which is iteratively increased to make the training later more strictly filter non-physical feasible solutions (structure designs that violate electromagnetic laws).

[0043] For example: introduce a physical regularization term in the discriminator loss function: L_physics=k•||∇×H-J-∂D / ∂t||2², where ∇×H is the magnetic field curl (representing the rate of change of the magnetic field), J is the current density (current intensity per unit area), ∂D / ∂t is the time rate of change of the electric displacement vector (representing the change of the electric field), and ||•||2² is the L2 norm square (L2 norm square); set k0=0.4 and β=0.03 to make the first iteration strength weight k=0.43 (physical constraint proportion 43%), and the 30th iteration strength weight k=1.3 (physical constraint proportion 130%, dominant discrimination logic).

[0044] In summary, steps S111 to S112 build a co-evolution framework in which the generator's exploration ability is exponentially enhanced (j>i) and the discriminator's physical review strength is linearly improved (k>i) through dynamic coupling of perturbation injection and physical constraint mechanism, so that the generative adversarial network can systematically break through the traditional mutual inductor structure bottleneck while meeting the basic laws of electromagnetic fields, providing strict mathematical and physical guarantees for potential excitation training. The accuracy, comprehensiveness and efficiency of the generative adversarial network potential excitation training are significantly improved, and the ability to excite the generative adversarial network to show the potential of mutual inductor structure optimization through training is improved.

[0045] In some embodiments, in the pre-optimization process in S2, the optimization potential is excavated, including: S21, compare the final transformer optimization structure output by the potential excitation trained generative adversarial network in the pre-optimization process with other transformer optimization structures output by the generative adversarial network in history to locate innovative optimization features in the final transformer optimization structure.

[0046] In this step, the final transformer refers to the transformer optimization structure generated by the final game between the generator and the discriminator in the potential excitation trained generative adversarial network. By comparing its features with other transformer optimization structures output by the generative adversarial network in history, those features that are innovative optimization compared with historical optimization work, i.e. innovative optimization features.

[0047] Specifically, for each structural feature (such as core cross-sectional area, air gap width, and winding turn density in the transformer structure) in the final transformer optimization structure, the feature is compared with the same type of feature in other transformer optimization structures. If the similarity is less than a preset threshold (such as 80%), the feature is regarded as an innovative optimization feature.

[0048] S22, trace the related process of the generator and the discriminator in the potential excitation trained generative adversarial network to generate innovative optimization features.

[0049] In this step, the related process refers to the process record of any feature generated by the game between the generator and the discriminator in the potential excitation trained generative adversarial network. It at least includes: the transformer optimization structure output by the generator containing the innovative optimization feature, the random generation disturbance applied when the generator outputs the transformer optimization structure containing the innovative optimization feature, the rational discrimination constraint applied when the discriminator discriminates the transformer optimization structure containing the innovative optimization feature, and the corresponding discrimination result, etc.

[0050] Specifically, the related process can be traced by querying the working log of the potential excitation trained generative adversarial network.

[0051] S23, behavior analysis of the related process to determine the change of game behavior.

[0052] In this step, the behavior analysis refers to the analysis of the change of game behavior between the generator and the discriminator in the related process over time. The game behavior refers to the behavior of the game between the generator and the discriminator, such as: the discrimination attention area of the discriminator in the transformer optimization structure generated by the generator, the discrimination result, and the punishment or reward applied to the generator, etc., and also such as: the behavior of the generator to improve the transformer optimization structure affected by the discriminator, etc. The determined change of game behavior can be represented in the form of time series.

[0053] S24, class autonomous behavior discrimination of the change of game behavior to obtain a class autonomous behavior distribution.

[0054] In this step, the autonomous behavior refers to a game behavior similar to artificial autonomous optimization of the mutual inductor structure. The autonomous behavior distribution refers to the time distribution of all autonomous behaviors in the game behavior change.

[0055] S25, reverse derivation of the autonomous optimization intention represented by the autonomous behavior distribution.

[0056] In this step, the autonomous optimization intention refers to the intention of the potential excitation trained generative adversarial network to autonomously generate an optimized mutual inductor structure. The autonomous behavior distribution is generated due to this intention, and therefore, the autonomous optimization intention represented by the autonomous behavior distribution can be reversely derived.

[0057] S26, determining the optimization potential based on the autonomous optimization intention.

[0058] In this step, the autonomous optimization intention reflects the potential of the potential excitation trained generative adversarial network to autonomously generate an optimized mutual inductor structure, and therefore, the optimization potential can be determined therefrom. For example, if the autonomous optimization intention is to optimize the magnetic performance of the mutual inductor structure, the optimization potential is determined to be the magnetic performance optimization potential.

[0059] In summary, steps S21 to S26 locate the innovative features through feature comparison, trace the related game process, analyze the game behavior, distinguish the autonomous behavior, reversely derive the autonomous optimization intention, and determine the optimization potential, forming a verifiable potential mining closed loop, greatly improving the accuracy of mining optimization potential in the pre-optimization process, and significantly improving the applicability of the mined optimization potential for targeted potential transformation training of the generative adversarial network.

[0060] In some embodiments, the S24, the autonomous behavior of the game behavior change is distinguished, and the autonomous behavior distribution is obtained, including: S241, for each behavior in the game behavior change, analyzing the logical relationship between the behavior and other behaviors in the preset change range before and after the behavior, constructing a first feature vector of the logical relationship, matching the first feature vector with a first standard vector representing that the behavior belongs to the autonomous behavior category, and if the matching degree exceeds a preset first threshold, the game behavior is taken as a first selected autonomous behavior.

[0061] In this step, the preset change range refers to a range preset for selecting logical connection analysis, for example, 3 behaviors before and after the behavior in the sequence of game behavior changes. The change range can be set as needed, such as setting a larger change range if it is expected that the system analyzes more logical connections between behaviors and the behavior. The logical connection refers to the logical relationship between different behaviors, such as the causal relationship, gradient transmission relationship, etc. The first standard vector is constructed according to the logical relationship between the continuous strategy of manually optimizing the mutual inductor structure. The first feature vector is constructed based on the analysis of the logical type, and the first feature vector is matched with the first standard vector. If the matching degree exceeds the preset first threshold, it indicates that the behavior before and the logical relationship are highly similar to the continuous strategy of manually optimizing the mutual inductor structure, and are used as the first selected autonomous behavior.

[0062] Specifically, a specific example is provided: the current behavior is that the generator optimizes the core waist compensation slot, the analyzed logical connection is that the generator adjusts the winding to make the discriminator optimize the compensation slot (hereinafter referred to as causal relationship A) and the permeability is improved after the discriminator optimizes the compensation slot (hereinafter referred to as gradient transmission relationship B), and the first feature vector constructed is [causal relationship A, gradient transmission relationship B]; the continuous strategy of manually optimizing the compensation slot of the mutual inductor structure is extracted in advance, the logical relationship of the strategy is that the winding is adjusted to make the compensation slot optimized and the permeability improved (hereinafter referred to as logical relationship C), and the second feature vector constructed is [logical relationship C]. The cosine similarity between the first feature vector and the second feature vector is 80%, which exceeds the preset first threshold of 65%, and the current behavior is used as the selected autonomous behavior.

[0063] S242, slidingly segmenting the game behavior changes using a preset step length time window.

[0064] In this step, the preset step length is a number preset to make the time window contain continuous behaviors at a time, for example, the step length is set to 5 behaviors. When slidingly segmenting, the time window is slid in the sequence of game behavior changes. After each sliding, the behavior originally at the first position exits the window, and a new behavior enters the window. The behavior window formed after each sliding is used as the sliding segmentation result.

[0065] Specifically, a specific example is provided: the preset step length is set to 5 behaviors. The first sliding is that the time window contains the first to fifth behaviors in the sequence of game behavior changes, forming a behavior window. The second sliding is that the time window contains the second to sixth behaviors in the sequence, and so on.

[0066] S243, for each behavior window obtained by sliding division, if the window contains a first candidate autonomous behavior, analyze the behavior distribution of the contained first candidate autonomous behavior in the window, construct a second feature vector of the behavior distribution, match the second feature vector with a preset second standard vector representing that the behaviors other than the contained first candidate autonomous behavior in the window belong to the autonomous behavior category, and if the matching degree exceeds a preset second threshold, the other behaviors are taken as second candidate autonomous behaviors.

[0067] In this step, the behavior distribution refers to the position distribution of the first candidate autonomous behavior in each behavior in the window, the ratio of the first candidate autonomous behavior to the total number of behaviors in the window, etc. The second standard vector is constructed in advance according to the standard behavior distribution that can represent that the behaviors other than the contained first candidate autonomous behavior in the window belong to the autonomous behavior category, for example: if the position distribution of the first candidate autonomous behavior in each behavior in the window indicates that the first and last behaviors are the first candidate autonomous behaviors, and the ratio of the first candidate autonomous behavior to the total number of behaviors in the window exceeds 0.6, it means that each behavior in the behavior window is autonomous behavior at the beginning and end, and the autonomous behavior accounts for a large proportion in it, so the behaviors other than the contained first candidate autonomous behavior belong to the autonomous behavior category. The second feature vector of the behavior distribution is constructed, and if the matching degree between the second feature vector and the second standard vector exceeds the preset second threshold, the other behaviors are taken as second candidate autonomous behaviors.

[0068] S244, merging all the first candidate autonomous behaviors and the second candidate autonomous behaviors to determine the final autonomous behaviors.

[0069] S245, determining the distribution of the autonomous behaviors in the change of the game behaviors, and taking it as the autonomous behavior distribution.

[0070] In this step, the above distribution refers to the time change distribution of the autonomous behaviors in the change of the game behaviors, etc.

[0071] To sum up, steps S241 to S245 construct a first feature vector by analyzing the logical connection between the behavior in the game behavior change and other behaviors in the preset change range before and after it, introduce a first standard vector, and preliminarily determine a first candidate class autonomous behavior based on the matching degree of the two. Then, the game behavior change is divided into sliding windows using a preset step size. For the behavior window obtained by sliding division, a second feature vector is constructed by analyzing the behavior distribution of the first candidate class autonomous behavior, a second standard vector is introduced, and a second candidate class autonomous behavior is determined in depth based on the matching degree of the two. Finally, all the first candidate class autonomous behaviors and the second candidate class autonomous behaviors are merged to obtain a class autonomous behavior, and the distribution of the class autonomous behavior is analyzed to obtain a class autonomous behavior distribution. The accuracy, comprehensiveness and efficiency of determining the class autonomous behavior distribution are greatly improved, and the accuracy of using the class autonomous behavior distribution as a subsequent reverse derivation of autonomous optimization intention is improved.

[0072] In some embodiments, the S25, the autonomous optimization intention represented by the class autonomous behavior distribution includes: S251, inputting the class autonomous behavior distribution into an intention speculation knowledge graph for propagation.

[0073] In this step, the intention speculation knowledge graph is constructed in advance according to the autonomous optimization intentions represented by different class autonomous behavior distributions. Specifically, a tree structure can be used for representation. For each autonomous behavior distribution and the autonomous optimization intention represented thereby, the type of the autonomous behavior distribution is set at the root node, and the behaviors in the autonomous behavior distribution are set at the successive branch nodes in sequence, and finally the autonomous optimization intention represented thereby is set at the leaf node. When the class autonomous behavior distribution is input into the intention speculation knowledge graph for propagation, each behavior in the class autonomous behavior distribution will correspond to different branch nodes. If each branch node on the same path has a corresponding behavior, the autonomous optimization intention in the leaf node of the path is taken as the associated reasoning result triggered after the propagation.

[0074] S252, deriving the corresponding autonomous optimization intention based on the associated reasoning result triggered after the propagation of the class autonomous behavior distribution.

[0075] In this step, the corresponding autonomous optimization intention can be directly derived based on the associated reasoning result triggered after the propagation of the class autonomous behavior distribution.

[0076] To sum up, steps S251 to S252 input the class autonomous behavior distribution into the intention speculation knowledge graph for propagation, derive the corresponding autonomous optimization intention based on the associated reasoning result triggered after the propagation of the class autonomous behavior distribution, and improve the accuracy of reverse derivation of the autonomous optimization intention represented by the class autonomous behavior distribution. In some embodiments, the S3, the potential transformation training of the generative adversarial network for the mined optimization potential includes: S31, construct a potential feature vector of the excavated optimization potential.

[0077] In this step, the key features of the excavated optimization potential are extracted using PCA feature extraction technology, including at least load capacity, temperature adaptability, frequency response, etc. These key features can represent various potential directions that may occur during the optimization process of the transformer, and together form a potential feature vector. The potential feature vector can also be mapped to a high-dimensional space, so that each potential feature can reflect different optimization possibilities. For example, changes in the thickness and material of the structure may affect the performance of the transformer under certain conditions, and these changes will be expressed by the potential feature vector.

[0078] S32, map and inject the potential feature vector into the latent space of the generator of the generative adversarial network to form a potential conditioned latent input.

[0079] In this step, the key is how to effectively pass the potential feature vector extracted in the previous step to the generative adversarial network. Specifically, the potential feature vector is injected as a conditional input into the latent space of the generator. This process not only considers the potential feature itself, but also considers its actual impact on structural optimization. The potential feature vector will push the generator to explore the possible optimization space. In order to ensure the influence of the feature vector on the generation process, the potential feature will be mapped through a specific network layer (such as a fully connected layer or a convolutional layer) to form a latent space that the generator can understand and process. These potential conditioned latent inputs will guide the generator to move in the direction of optimization potential during the generation process. With the continuous injection of conditional input, the generator begins to adjust its internal parameters so that it can more accurately respond to potential features during the optimization process. The potential feature vector affects the structural characteristics of the generator output, making the generated transformer design more in line with optimization requirements.

[0080] S33, train the generative adversarial network using a dataset containing optimization potential corresponding target domain; wherein the generator receives the potential conditioned latent input and generates potential transformation data with optimization potential characteristics; the discriminator receives real data samples and potential transformation data to determine their authenticity and whether they meet the optimization characteristics represented by the potential feature vector.

[0081] In this step, a set of target domain data containing rich optimization potential features is selected. For example, the data set can contain performance parameters, structural characteristics of various different mutual inductor designs, and their optimized results. These data will be used to train the generator. The generator attempts to generate data output (such as optimized mutual inductor design) that meets the optimization potential after receiving the conditional input of the potential feature vector. At the same time, the discriminator is responsible for evaluating whether the generated data meets the real data distribution and the optimization target described by the potential feature vector. The discriminator not only judges the authenticity of the generated data, but also determines whether it meets the optimization standard of the potential feature. The generator and the discriminator engage in mutual game through adversarial training. The generator continuously improves the quality and diversity of its output data, and the discriminator continuously improves its ability to distinguish whether the generated data meets the potential feature.

[0082] In the training process, the potential matching loss between the potential transformation data and the potential feature vector is calculated, and the parameters of the generator are updated by back propagation, so that the generator output matches the optimization potential, and the discriminator and the generator are optimized until the generative adversarial network reaches a Nash equilibrium state, so that the generator can stably output data that meets the optimization potential at this time.

[0083] In this step, first, the matching loss between the generated data and the target potential feature vector is calculated. This loss reflects the degree of consistency between the generated data and the potential feature vector. Ideally, the generated data should be highly consistent with the optimization direction represented by the potential feature vector. Using the calculated potential matching loss, the parameters of the generator are updated by back propagation. By continuously adjusting the weights of the generator, the data generated by the generator gradually approaches the optimization design that meets the potential feature. At the same time, the parameters of the discriminator also need to be optimized to ensure that it can more accurately evaluate the quality of the generated data. As the training deepens, the abilities of the generator and the discriminator are continuously improved, and finally the generator can stably generate data that meets the optimization potential. Through repeated optimization, the generator and the discriminator eventually reach a Nash equilibrium state. In this state, the generator can continuously generate high-quality optimization structures, and the discriminator is difficult to determine whether the generated data is real data, indicating that the generator has been able to stably output designs that meet the optimization potential requirements.

[0084] In summary, steps S31 to S34 will enable the generative adversarial network to transform the optimization potential into actual mutual inductor structure optimization capability through training. During the training process, the injection of potential feature vectors, the calculation of potential transformation loss, and the adversarial optimization of the generator and the discriminator, together promote the effective implementation of this process, ensuring that the generated mutual inductor structure has high performance.

[0085] In some embodiments, the S4, using the potential transformation trained generative adversarial network, performs formal optimization on the transformer structure, comprising: inputting the transformer structure into the potential transformation trained generative adversarial network to perform formal optimization.

[0086] In the formal optimization of the transformer structure, the transformer structure is directly inputted into the potential transformation trained generative adversarial network.

[0087] In some embodiments, after the S4, using the potential transformation trained generative adversarial network, performs formal optimization on the transformer structure, further comprising: S5, interacting with the user during the formal optimization process.

[0088] The S5, interacting with the user during the formal optimization process, comprises: providing the user with the potential transformation trained generative adversarial network for further optimization of the transformer structure in human-machine collaboration.

[0089] In the interaction with the user during the formal optimization process, the user can be provided with the potential transformation trained generative adversarial network for further optimization of the transformer structure in human-machine collaboration.

[0090] The embodiments of the present application provide a transformer structure optimization system based on a generative adversarial network, as shown in Figure 2 comprising: a potential excitation training module 1 for performing potential excitation training on a generative adversarial network; an optimization potential mining module 2 for performing pre-optimization on a transformer structure using the potential excitation trained generative adversarial network, and mining optimization potential during the pre-optimization process; a potential transformation training module 3 for performing potential transformation training on the generative adversarial network for the mined optimization potential; a transformer structure optimization module 4 for performing formal optimization on the transformer structure using the potential transformation trained generative adversarial network.

[0091] Obviously, those skilled in the art can make various modifications and variations to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalent technologies, the present application also intends to include these modifications and variations.

Claims

1. A method for transformer structure optimization based on a generative adversarial network, characterized in that, The method comprises the following steps: Potential excitation training is performed on the generative adversarial network; The generative adversarial network after the potential excitation training is used to pre-optimize the structure of the mutual inductor, and the optimization potential of the mutual inductor is excavated in the pre-optimization process; Potential transformation training is performed on the generative adversarial network according to the excavated optimization potential; The generative adversarial network after the potential transformation training is used to formally optimize the structure of the mutual inductor.

2. The generator-adversary network based transformer structure optimization method of claim 1, wherein, The potential excitation training of the generative adversarial network comprises the following steps: The following rules are embedded in the control layer of the generative adversarial network to control the game process between the generator and the discriminator: When the generator generates the structure of the mutual inductor for the i-th time, a random generated disturbance with a strength weight of j is applied to the generator; wherein j = j0 + a i, j0 is a disturbance strength reference weight, and a is a strength gain coefficient, and a > 0; When the discriminator discriminates the structure of the mutual inductor generated by the generator for the i-th time, a rational discrimination constraint with a strength weight of k is applied to the generator; wherein k = k0 + β i, k0 is a constraint strength reference weight, and β is a strength gain coefficient, and β > 0.

3. The generator-adversarial network based transformer structure optimization method of claim 1, wherein, The optimization potential of the mutual inductor is excavated in the pre-optimization process, comprising the following steps: The final mutual inductor optimization structure output by the generative adversarial network in the pre-optimization process is compared with other mutual inductor optimization structures output by the generative adversarial network in history to locate innovative optimization features in the final mutual inductor optimization structure; The related process of the generator and the discriminator in the generative adversarial network to generate innovative optimization features is traced back; The behavior analysis is performed on the related process to determine the change of the game behavior; The game behavior distribution is obtained by performing the class autonomous behavior discrimination on the change of the game behavior; The autonomous optimization intention represented by the class autonomous behavior distribution is reversely deduced; The optimization potential is determined based on the autonomous optimization intention.

4. The generator-adversarial network based transformer structure optimization method of claim 3, wherein, The game behavior distribution is obtained by performing the class autonomous behavior discrimination on the change of the game behavior, comprising the following steps: For each behavior in the change of the game behavior, the logical relationship between the behavior and other behaviors in the preset change range before and after the behavior is analyzed, a first feature vector of the logical relationship is constructed, the first feature vector is matched with a first standard vector representing that the behavior belongs to a class autonomous behavior category, if the matching degree exceeds a preset first threshold, the game behavior is taken as a first selected class autonomous behavior; The game behavior change is divided by using a time window with a preset step length; For the behavior window obtained by each sliding division, if the window contains the first selected class autonomous behavior, the behavior distribution of the first selected class autonomous behavior in the window is analyzed, a second feature vector of the behavior distribution is constructed, the second feature vector is matched with a second standard vector representing that other behaviors in the window except the first selected class autonomous behavior belong to the class autonomous behavior category, if the matching degree exceeds a preset second threshold, the other behaviors are taken as second selected class autonomous behaviors; All the first selected class autonomous behaviors and the second selected class autonomous behaviors are merged to determine the final class autonomous behaviors; The distribution of the class autonomous behaviors in the change of the game behavior is determined and taken as the class autonomous behavior distribution.

5. The generator-adversarial network-based mutual inductor structure optimization method of claim 3, wherein, The autonomous optimization intention represented by the class autonomous behavior distribution is reversely deduced, comprising the following steps: The autonomous behavior distribution is input into the intention inference knowledge graph for propagation; Based on the associated inference results triggered after the propagation of the autonomous behavior distribution, the corresponding autonomous optimization intention is derived.

6. The generator-adversarial network based transformer structure optimization method of claim 1, wherein, The potential transformation training of the generative adversarial network is performed on the basis of the mined optimization potential, including: A potential feature vector of the mined optimization potential is constructed; The potential feature vector is mapped and injected into the latent space of the generator of the generative adversarial network to form a potential conditioned latent input; The generative adversarial network is trained by using a data set containing the optimization potential corresponding to the target field; wherein the generator receives the potential conditioned latent input and generates potential transformation data with optimization potential characteristics; the discriminator receives real data samples and potential transformation data to determine their authenticity and whether they meet the optimization characteristics represented by the potential feature vector; During the training process, the potential matching loss between the potential transformation data and the potential feature vector is calculated, and the parameters of the generator are updated by back propagation, so that the generator outputs the matching optimization potential, and the discriminator and the generator are optimized until the generative adversarial network reaches a Nash equilibrium state, so that the generator can stably output data meeting the optimization potential at this time.

7. The generator-adversarial network based transformer structure optimization method of claim 1, wherein, The formal optimization of the transformer structure is performed by using the generative adversarial network trained by the potential transformation, including: The transformer structure is input into the generative adversarial network trained by the potential transformation for formal optimization.

8. The generator-adversarial network based transformer structure optimization method of claim 1, wherein, After the formal optimization of the transformer structure is performed by using the generative adversarial network trained by the potential transformation, it further includes: Interaction with the user during the formal optimization process.

9. The generator-adversarial network based transformer structure optimization method of claim 8, wherein, The interaction with the user during the formal optimization process includes: the user and the generative adversarial network trained by the potential transformation cooperatively optimize the transformer structure for further optimization.

10. A transformer structure optimization system based on a generative adversarial network, characterized in that, It includes: A potential excitation training module for potential excitation training of the generative adversarial network; An optimization potential mining module for pre-optimization of the transformer structure by using the generative adversarial network trained by the potential excitation, and mining of the optimization potential of the transformer structure during the pre-optimization process; A potential transformation training module for potential transformation training of the generative adversarial network on the basis of the mined optimization potential; A transformer structure optimization module for formal optimization of the transformer structure by using the generative adversarial network trained by the potential transformation.

Images