Collaborative optimization design method and device based on proxy model and defect early warning

Abstract

The invention relates to the technical field of transformers, in particular to a collaborative optimization design method and device based on an agent model and defect early warning. The collaborative optimization design method comprises the following steps: performing performance analysis and reliability analysis on a design vector based on a double-engine parallel analysis framework; constructing a multi-objective optimization problem based on the performance index and the risk probability corresponding to the design vector, and solving the multi-objective optimization problem by adopting a multi-objective evolutionary algorithm to generate a Pareto optimal solution set; and carrying out verification and decision making based on the Pareto optimal solution set, and outputting a final design scheme. According to the method, performance analysis and reliability analysis are carried out on design vectors through a double-engine parallel analysis framework, reliability is taken as a quantifiable optimization target, the reliability and a traditional performance target are incorporated into an optimization process, collaborative design and multi-target balance of electrical performance, thermal performance, mechanical performance and long-term operation reliability are achieved, and the design efficiency is improved. And the inherent quality of the product is improved from the source.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of transformers, in particular to a collaborative optimization design method and device based on a proxy model and defect early warning. BACKGROUND

[0002] In the design of transformers, personal experience of designers is usually heavily relied on. Designers need to first parameterize modeling according to fuzzy intentions such as natural language descriptions and sketches, and then rely on high-fidelity physical simulation software such as finite element method and finite volume method to iteratively analyze and verify the performance (such as electromagnetic, thermal, stress) of the design scheme. The existing problems are: Firstly, the conversion from fuzzy design intentions to accurate parameterized models is inefficient, highly dependent on manual interpretation and operation, and lacks automation and intelligence.

[0003] Secondly, the multi-physical field simulation based on numerical simulation is extremely expensive, especially when the design parameter space is large, a complete simulation may take several hours or even several days. This makes it almost impossible to directly embed the simulation into the optimization loop which needs thousands of iterations, severely restricting the optimization efficiency and prolonging the product development cycle.

[0004] Furthermore, the existing optimization process mostly focuses on performance indicators (such as efficiency, loss, temperature rise), but lacks the ability to prospectively predict and quantitatively evaluate potential manufacturing or operating defects (such as winding overheating, insulation damage) in the design stage. Existing defect analysis often relies on experimental or field operation data after product manufacturing is completed, which has serious lag, and once problems are found, it will result in huge rework cost and time delay. SUMMARY

[0005] The main purpose of the present application is to provide a collaborative optimization design method and device based on a proxy model and defect early warning, aiming to provide an intelligent design method and system that can deeply integrate design intention understanding, trade-off performance indicators and reliability, to fundamentally improve design efficiency.

[0006] To achieve the above purpose, the collaborative optimization design method based on a proxy model and defect early warning proposed by the present application is applied to transformer design, including the following steps: Step S1: Obtain the multi-modal design requirements of the user for designing a transformer; Step S2: Analyze the multi-modal design requirements based on a trained multi-modal large model to generate a parameterized design vector; Step S3: Perform performance analysis and reliability analysis on the design vector based on a dual-engine parallel analysis framework to obtain the performance indicators and risk probability corresponding to the design vector; Step S4: based on the performance index and risk probability corresponding to the design vector, a multi-objective optimization problem is constructed, and a multi-objective evolutionary algorithm is used to solve the multi-objective optimization problem to generate a Pareto optimal solution set; Step S5: verification and decision are made based on the Pareto optimal solution set, and a final design scheme is output.

[0007] In the above collaborative optimization design method based on the proxy model and defect early warning, the generation of the parameterized design vector includes the following steps of the multi-modal large model: extracting a feature vector from the multi-modal design requirement; fusing the feature vectors of different modalities to obtain a representation vector; mapping the representation vector to the key parameters of the parameterized model to obtain the design vector.

[0008] In the above collaborative optimization design method based on the proxy model and defect early warning, the double-engine parallel analysis framework includes a PINN proxy model and a defect risk prediction model, the input of the PINN proxy model is the design vector and the spatial coordinates, and the output is the physical field and the performance index; the input of the defect risk prediction model is the design vector, and the output is the risk probability of occurrence of various defects.

[0009] In the above collaborative optimization design method based on the proxy model and defect early warning, the construction and training of the PINN proxy model include the following steps: constructing a PINN proxy model with design vector and spatial coordinates as input, and physical field and performance index as output obtaining a data set obtained by simulation or experiment as supervision data for model training; constructing a composite loss function composed of a data loss term, a physical loss term and a boundary condition loss term; wherein the data loss term is to measure the difference between the predicted value of the PINN proxy model and the supervision data; the physical loss term is obtained by sampling a set of configuration points in the calculation domain, and the physical field predicted by the PINN proxy model is substituted into its corresponding control equation and the residual is calculated; the boundary condition loss term is obtained by sampling a set of points on the boundary, and calculating the difference between the predicted value of the PINN proxy model and the preset condition of the boundary; minimizing the composite loss function by an optimization algorithm to optimize the weight parameters of the PINN proxy model until the loss converges, and obtaining the trained PINN proxy model.

[0010] In the above collaborative optimization design method based on the proxy model and defect early warning, the construction and training of the defect risk prediction model include the following steps: obtaining a training data set containing a plurality of defect samples, each training sample including a defect image, a design vector corresponding to the defect sample and a defect label; extracting a defect feature vector from the defect image by using a pre-trained feature extractor; splicing the defect feature vector and a design vector corresponding to the defect sample to generate a fusion feature vector; constructing a classifier model comprising a Softmax output layer; inputting the fusion feature vector into the classification model to obtain an original score vector through forward propagation calculation; converting the original score vector into a defect probability distribution through the Softmax output layer to obtain the risk probability of each type of defect; maximizing the log-likelihood function as the target, optimizing the parameters of the classifier model through an optimization algorithm until convergence, and obtaining the trained classifier model as a defect risk prediction model.

[0011] In the above collaborative optimization design method based on the proxy model and defect warning, the multi-objective evolutionary algorithm is the NSGA-II algorithm, and the following steps are used to solve the multi-objective optimization problem: Step S41: randomly generating an initial population comprising a plurality of design vectors; Step S42: for each design vector in the current population, obtaining its corresponding performance index and risk probability through a double-engine parallel analysis framework; Step S43: based on the prediction results of all design vectors, performing non-dominated sorting, selection, crossover and mutation operations on the population to generate a new offspring population; Step S44: repeating the above steps S42 and S43 until the iteration termination condition is met, and outputting the final Pareto optimal solution set.

[0012] In the above collaborative optimization design method based on the proxy model and defect warning, the verification based on the Pareto optimal solution set comprises the following steps: inputting each non-dominated solution in the Pareto optimal solution set into the defect risk prediction model to obtain the risk probability of each type of defect for each non-dominated solution; comparing the risk probability of each type of defect for each non-dominated solution with a risk threshold, and selecting the non-dominated solution whose risk probability of each type of defect does not exceed the risk threshold as a design scheme to be decided; When the risk probability of at least one type of defect for the non-dominated solution is greater than the risk threshold, the design parameter leading to the defect is back calculated through the defect risk prediction model.

[0013] The application also discloses a collaborative optimization design system based on a proxy model and defect warning, which is applied to the above collaborative optimization design method based on a proxy model and defect warning; comprising: an acquisition module, an analysis module, a prediction module, an optimization module, and a verification and decision module. The acquisition module is configured to acquire a multi-modal design requirement of a transformer designed by a user, the multi-modal design requirement including any one or more of a natural language description text, a design sketch, or a list of key parameters; The analysis module is configured to analyze the multi-modal design requirement based on a trained multi-modal large model to generate a design vector and a corresponding three-dimensional model of the transformer; The prediction module is configured to perform performance analysis and reliability analysis on the design vector based on a double-engine parallel analysis framework to obtain a performance index and a risk probability corresponding to the design vector; The optimization module is configured to construct a multi-objective optimization problem based on the performance index and the risk probability corresponding to the design vector, and to solve the multi-objective optimization problem using a multi-objective evolutionary algorithm to generate a Pareto optimal solution set; The verification and decision module is configured to verify and decide based on the Pareto optimal solution set to output a final design scheme.

[0014] The application further discloses an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the method described above when executing the program.

[0015] The application further discloses a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and the program is executed by a processor to implement the method described above.

[0016] The technical solution provided by the application can include the following beneficial effects: The application realizes the collaborative design and multi-objective trade-off of electrical performance, thermal performance, mechanical performance, and long-term operation reliability by taking "reliability" as a quantifiable optimization target and incorporating it into the optimization process together with the traditional "performance" target, thereby improving the inherent quality of products from the source.

[0017] In addition, multi-modal interaction enables inexperienced engineers to complete complex designs based on intuitive descriptions; at the same time, the experience of experienced engineers and a large number of fault cases are solidified into algorithms and models to form "digital assets" that can be permanently inherited and iterated. The application overcomes the over-reliance on experts in the prior art and solves the problem of fluctuation in design quality caused by the "talent gap", thereby realizing the standardization and sustainable improvement of design capability. BRIEF DESCRIPTION OF DRAWINGS

[0018] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description only show some of the embodiments of the present application, and for those skilled in the art, other drawings can be obtained from the structures shown in the drawings without creative labor.

[0019] Figure 1 The flowchart of the collaborative optimization design method of the present application is shown in the figure. Figure 2 The schematic diagram of the collaborative optimization design system of the present application is shown in the figure. Figure 3 The schematic diagram of the frame of the electronic device of the present application is shown in the figure. DETAILED DESCRIPTION

[0020] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, not all. 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 the present application.

[0021] It should be noted that all directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present application are only used to explain the relative positional relationship, movement condition, etc. between components in a certain posture (as shown in the drawings), and if the certain posture changes, the directional indications will also change accordingly.

[0022] In the present application, unless otherwise explicitly specified and limited, the terms "connection", "fixation", etc. should be understood in a broad sense. For example, "fixation" can be fixed connection, or detachable connection, or integral; can be mechanical connection, or electrical connection; can be direct connection, or indirect connection through an intermediate medium; can be internal connection of two elements, or interaction relationship between two elements, unless otherwise explicitly limited. For those skilled in the art, the specific meanings of the above terms in the present application can be understood according to the specific circumstances.

[0023] In addition, the description such as "first", "second" and the like in the present application is only for the purpose of description and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the technical features indicated. Therefore, the features defined as "first", "second" can be explicitly or implicitly included at least one of the features. In addition, the meaning of "and / or" appearing throughout the text is to include three parallel schemes, for example, "A and / or B" includes A scheme, or B scheme, or A and B scheme. In addition, the technical solutions of each embodiment can be combined with each other, but it must be based on the realization of the technical personnel in the art, when the combination of technical solutions appears contradictory or cannot be realized, it should be considered that the combination of such technical solutions does not exist, nor is it within the protection scope required by the present application.

[0024] The embodiment of the present application is described below Figure 1 The method for collaborative optimization design based on agent model and defect early warning of the embodiment of the present application is applied to transformer design, and includes the following steps: Step S1: obtaining the multi-modal design requirements of the user designed transformer, the multi-modal design requirements including any one or more of natural language description text, design sketch or key parameter list; Step S2: based on the trained multi-modal large model, the multi-modal design requirements are analyzed to generate a parameterized design vector and a corresponding transformer three-dimensional model; Step S3: based on the double-engine parallel analysis framework, the performance analysis and reliability analysis of the design vector are performed to obtain the performance index and risk probability corresponding to the design vector; Step S4: based on the performance index and risk probability corresponding to the design vector, a multi-objective optimization problem is constructed, and a multi-objective evolutionary algorithm is used to solve the multi-objective optimization problem to generate a Pareto optimal solution set; Step S5: verification and decision are made based on the Pareto optimal solution set, and the final design scheme is output.

[0025] The present application realizes the quantification of "reliability" as an optimization target through the double-engine parallel analysis framework for performance analysis and reliability analysis of the design vector, and integrates the traditional "performance" target into the optimization process, realizes the collaborative design and multi-objective trade-off of electrical performance, thermal performance, mechanical performance and long-term operation reliability, and improves the inherent quality of the product from the source.

[0026] In addition, multi-modal interaction enables inexperienced engineers to complete complex design based on intuitive description; at the same time, the experience of experienced engineers and a large number of fault cases are solidified into algorithms and models to form "digital assets" that can be permanently inherited and iterated. Overcome the existing over-reliance on experts, solve the design quality fluctuation problem caused by "talent gap", and realize the standardization and sustainable improvement of design capability.

[0027] wherein the generating the parameterized design vector comprises the following steps: extracting feature vectors in the multi-modal design requirements; specifically, the multi-modal large model can be CLIP or GPT-4V based on the Transformer architecture. According to the natural language description text T, the design sketch I sketch or the list of key parameters P list , the text feature f t =TextEncoder(T) can be extracted, the image feature f i =ImageEncoder(I sketch ), and the parameter feature f p =Embedding(P list ).

[0028] fusing feature vectors of different modalities to obtain a representation vector f intent ; f inten t=Transformer_Fusion([f t ;f i ;f p ]).

[0029] mapping the representation vector to the key parameters of the parameterized model to obtain the design vector. Specifically, a decoder network (such as a fully connected layer or another Transformer decoder) can be used to map the representation vector to the key parameters of the specific, drivable parameterized model to obtain the design vector X=(x1,x2,...,x n ), where x1,x2,...,x n represent key parameters such as core diameter, window height, number of turns, etc. And through the three-dimensional modeling plug-in and the simulation grid generation plug-in, the corresponding three-dimensional model and simulation grid are automatically generated in the CAD / CAE software.

[0030] Specifically, the dual-engine parallel analysis framework includes a PINN proxy model and a defect risk prediction model. The input of the PINN proxy model is the design vector and the spatial coordinates, and the output is the physical field and the performance index. The input of the defect risk prediction model is the design vector, and the output is the risk probability of occurrence of various defects.

[0031] The PINN proxy model quickly and physically accurately evaluates the performance problems of the design vector, ensuring that the design scheme meets the performance requirements. The defect risk prediction model converts the fuzzy concept of reliability into a quantifiable design goal to improve the reliability of the final design scheme.

[0032] Further specifically, constructing and training the PINN agent model includes the following steps: A PINN agent model N(X; θ) is constructed, which takes design vectors and spatial coordinates as input, and takes physical fields (such as magnetic flux density B, temperature field T, stress field σ) and performance indicators (such as total loss, maximum temperature rise) as output, where X represents the design vector, and θ represents the weight parameters of the PINN agent model; Obtain the simulation or experimental data set as the supervised data for model training; A composite loss function composed of a data loss term, a physical loss term and a boundary condition loss term is constructed; wherein the data loss term is to measure the difference between the predicted value of the PINN agent model and the supervised data; Specifically, the expression of the data loss term is: ; Wherein, represents the data loss value, represents the total number of training data points, represents the predicted value of the input , represents the corresponding simulation or experimental value; The physical loss term is obtained by sampling a set of configuration points in the calculation domain, substituting the physical field predicted by the PINN agent model into its corresponding control equation and calculating the residual; specifically, taking the steady-state heat conduction equation as an example, Np points are randomly sampled in the calculation domain Ω, then the expression of the physical loss term is: ; represents the physical loss value, represents the total number of randomly sampled points, represents the temperature predicted by the PINN agent model at point x j , k represents the thermal conductivity, represents the heat source at point x j .

[0033] The boundary condition loss term is obtained by sampling a set of points on the boundary, and calculating the difference between the predicted value of the PINN agent model and the preset condition of the boundary; specifically, the expression of the boundary condition loss term is: ; represents the boundary condition loss value, represents the number of sampled points on the boundary, represents the boundary condition operator.

[0034] Therefore, the expression of the total loss function is: ; wherein, represents the total loss value, represents a hyperparameter balancing the importance of the data loss term, represents a hyperparameter balancing the importance of the physical loss term, represents a hyperparameter balancing the importance of the boundary condition loss term.

[0035] The weight parameters θ of the PINN agent model are optimized by minimizing the composite loss function through an optimization algorithm (such as gradient descent) until the loss converges, obtaining the trained PINN agent model N(X; θ*), θ* representing the weight parameters of the PINN agent model after training.

[0036] Further specifically, constructing and training the defect risk prediction model comprises the following steps: obtaining a training data set containing a plurality of defect samples Each training sample includes a defect image , a design vector corresponding to the defect sample , and a defect label ; Specifically, it is worth noting that the defect image can be an infrared thermal image, and the defect label can be a winding overheating, insulation damage, mechanical failure, etc. Defect category.

[0037] The pre-trained feature extractor is used to extract the defect feature vector from the defect image; the feature extractor can be a convolutional neural network (ResNet) or a visual Transformer (ViT).

[0038] The defect feature vector and the design vector corresponding to the defect sample are spliced to generate a fusion feature vector; A classifier model containing a Softmax output layer is constructed; The fusion feature vector is input into the classification model, and the original score vector is calculated through forward propagation; wherein the original score vector refers to the vector without passing through the Softmax output layer.

[0039] The original score vector is converted into a defect probability distribution through the Softmax output layer, obtaining the risk probability of each type of defect occurrence; The parameters of the classifier model are optimized through an optimization algorithm to maximize the log-likelihood function until convergence, obtaining the trained classifier model as the defect risk prediction model. Specifically, the expression of the maximized log-likelihood function is: ; wherein, Pi represents the probability of occurrence of each type of defect for the design vector of the mth sample, and M represents the number of samples.

[0040] The training phase uses the defect images, the design vectors corresponding to the samples, and the defect labels to train the classifier model. The purpose is to let the classifier model learn the mapping from the parameters of the design vector to the image features, and the mapping from the image features to the defect labels. After training is completed, the classifier model can predict the risk probability of occurrence of different defects only according to the design parameters, without the need for image data. It is assumed that there is some kind of internal relationship between the design vector and the defect image features. For example, the parameters in the design vector (such as magnetic flux density, current density, etc.) will affect the physical field distribution (such as temperature field, magnetic field) of the transformer, and then affect the occurrence of defects (such as overheating, insulation aging). And the defect image (such as infrared thermal image) is the visualization of these physical field distributions. Therefore, the model can be trained to predict the defect features contained in the defect image from the design vector.

[0041] The present application solves how to systematically transform the defects, defect features, etc. "implicit knowledge" contained in the fault of the operation and maintenance image data into "explicit knowledge" available for the design stage, and establishes the mapping relationship between "defect features-design parameters", so that potential design weaknesses can be warned and avoided at the early stage of design.

[0042] For example, the mathematical model of the multi-objective optimization problem is: min X [F1(X),F2(X),...,F k (X)]; gi(X)≤0,i=1,2,...,p; Where F1(X), F2(X), F k (X) represent the performance index related objective function or reliability objective function respectively. The performance index related objective function can be a performance index related objective function such as total loss, maximum temperature, temperature rise, efficiency, etc., for example: F1(X)=N loss (X) (total loss), F2(X)=N temp (X) (maximum temperature). The reliability objective function can find the function of the highest defect risk probability or the weighted defect risk probability, for example F3(X)=maxP(l|X) (highest defect risk probability). gi(X) represents different engineering constraints, such as cost, volume, impedance, etc.

[0043] As some preferred embodiments, the multi-objective evolutionary algorithm is the NSGA-II algorithm, and the following steps are used to solve the multi-objective optimization problem: Step S41: randomly generate an initial population containing multiple design vectors; Step S42: For each design vector in the current population, the corresponding performance index and risk probability are obtained through the dual-engine parallel analysis framework. Step S43: Based on the prediction results of all design vectors, non-dominated sorting, selection, crossover and mutation operations are performed on the population to generate a new offspring population; wherein in the same non-dominated layer, the crowding degree of non-dominated solutions is calculated to maintain diversity.

[0044] Step S44: Repeat the above steps S42 and S43 until the iteration termination condition is met, and output the final Pareto optimal solution set.

[0045] In this way, the synergistic design and multi-objective trade-off of electrical performance, thermal performance, mechanical performance and long-term operation reliability are realized, and the performance and reliability of the product are considered.

[0046] Further, the verification based on the Pareto optimal solution set includes the following steps: Each non-dominated solution in the Pareto optimal solution set is a design vector. Each non-dominated solution in the Pareto optimal solution set is input into the defect risk prediction model to obtain the risk probability of each type of defect for each non-dominated solution; Compare the probability of each non-dominated solution appearing each type of defect with the risk threshold, and select the non-dominated solution with a risk probability of each type of defect not exceeding the risk threshold as the design scheme to be decided; the final design scheme can be manually selected from multiple design schemes to be decided. When outputting the design scheme to be decided, the early warning report of the design scheme can be output synchronously. The early warning report contains the defect type and the probability of occurrence of the defect to warn potential defects.

[0047] When the risk probability of at least one type of defect in the non-dominated solution is greater than the risk threshold, the design parameters leading to the occurrence of the defect are back calculated through the defect risk prediction model. To remind the engineer to find the adjusted design parameters. Specifically, the importance of each design parameter in the design vector can be calculated using the gradient method or SHAP value. According to the feature importance ranking, the few design parameters that contribute most to the high risk are identified. For example, when using the gradient method, suggestions for improvement can be made according to the gradient sign of the few design parameters that contribute most to the high risk. For example, if the gradient of a design parameter is positive, it means that increasing the value of the design parameter will increase the risk, so the value of this design parameter should be reduced. The adjusted design vector is input into the defect risk prediction model to recalculate the risk probability of the design vector and verify whether the risk is reduced.

[0048] In a specific embodiment, a 10kV / 400V distribution transformer is designed as an example: A multi-modal design requirement is obtained through natural language and key parameter input. The input natural language description text is: "Design a 10kV / 400V distribution transformer with high efficiency, moderate cost, and good reliability, capacity 1000kVA, temperature rise limit 75K." The input key parameter list is: Rated capacity: 1000 kVA; Voltage combination: 10±2×2.5% / 0.4 kV; Connection group label: Dyn11; Short-circuit impedance: 4-6% (optimizable); Design constraints: No-load loss ≤ 1500 W; Load loss ≤ 10000 W; Top oil temperature rise ≤ 75 K.

[0049] Based on the trained multi-modal large model, the multi-modal design requirement is analyzed. Through the multi-modal large model, the key intentions of "high efficiency" (corresponding to low loss), "moderate cost" (corresponding to material usage control), and "high reliability" are identified from the natural language. Then the high-level intentions are mapped to the bottom layer of the core design variable vector X=(x1,x2,...,x n ); x1 is the core diameter (mm), x2 is the magnetic flux density (T), X3 is the low-voltage winding current density (A / mm²), X4 is the high-voltage winding current density (A / mm²), and X5 is the number of heat sinks. For example, through the above steps, an initial design vector X= [240, 1.65, 3.2, 2.8, 28] is obtained; and a transformer three-dimensional model is automatically generated in the CAD system.

[0050] The initial design vector X= [240, 1.65, 3.2, 2.8, 28] is input into the PINN proxy model, and the performance index Y= [P0=1400W, P k =9800W, ΔT oil =68K, Cost=￥45,000] is output by the PINN proxy model, where P0 is the no-load loss, P k is the load loss, ΔT oil is the top oil temperature rise, and Cost is the material cost estimate.

[0051] The initial design vector X = [240, 1.65, 3.2, 2.8, 28] is input into the defect risk prediction model, and the defect risk prediction model outputs the risk probability P = [P1=15%, P2=8%, P3=25%], where P1 represents the local overheating risk probability, P2 represents the short-circuit mechanical force risk probability, and P3 represents the insulation aging acceleration risk probability.

[0052] Optimization problem modeling: Objective 1 (minimization): total loss F1(X) = P0+P k ; Objective 2 (minimization): cost F2(X) = Cost; Objective 3 (minimization): maximum defect risk F3(X) = max(P l ,P2,P3); Constraints: ΔT oil ≤ 75K, short-circuit impedance in 4-6%.

[0053] Solve the optimization problem by NSGA-II, search in thousands of design points, get the Pareto optimal solution set, which includes the following three design schemes: Scheme one: Design vector X is: [260, 1.70, 3.0, 2.6, 30]; Performance index Y is: [P0=1350W, P k =9500W, Cost=￥48,500]; Risk probability P is: [P1=8%, P2=5%, P3=30%]; Characteristics: highest efficiency, but high cost, slightly increased insulation risk.

[0054] Scheme two: Design vector X is: [245, 1.63, 3.1, 2.7, 28]; Performance index Y is: [P0=1380W, P k =9650W, Cost=￥45,800]; Risk probability P is: [P1=10%, P2=6%, P3=18%]; Characteristics: balanced performance and cost, controllable risk.

[0055] Scheme three: Design vector X is: [230, 1.58, 3.4, 3.0, 26]; Performance index Y is: [P0=1450W, P k= 9900W, Cost = ￥43,000]; Risk probabilities P are: [P1 = 22%, P2 = 10%, P3 = 12%]; Characteristics: lowest cost, but high loss and overheating risk.

[0056] The second artificial selection scheme is chosen as the final scheme because it achieves the best balance among efficiency, cost, and reliability. The automatic output optimization design scheme report is: The parameters in the final design vector X are: core diameter 245mm, magnetic flux density 1.63T, low-voltage winding current density 3.1 A / mm², high-voltage winding current density 2.7 A / mm², and 28 groups of heat sinks.

[0057] The scheme predicts performance: no-load loss 1380W, load loss 9650W, efficiency >99.5%, cost ￥45,800, and meets all constraints.

[0058] Defect warning and design guidance report: Overall evaluation: the design scheme passes, local overheating risk probability 10%, short-circuit mechanical force risk probability 6%, and insulation aging acceleration risk probability 18%. The defect risks are within the acceptable range (<20% threshold).

[0059] As shown in Figure 2 The second aspect of the present application also discloses a collaborative optimization design system based on a proxy model and defect warning, which is applied to the collaborative optimization design method based on the proxy model and the defect warning; comprising: an acquisition module 210, an analysis module 220, a prediction module 230, an optimization module 240, and a verification and decision module 250; the acquisition module 210 is used to acquire the multi-modal design requirements of the user for designing a transformer, and the multi-modal design requirements include any one or more of natural language description text, design sketch, or key parameter list; the analysis module 220 analyzes the multi-modal design requirements based on a trained multi-modal large model to generate a design vector and a corresponding transformer three-dimensional model; the prediction module 230 performs performance analysis and reliability analysis on the design vector based on a double-engine parallel analysis framework to obtain the performance index and risk probability corresponding to the design vector; the optimization module 240 constructs a multi-objective optimization problem based on the performance index and risk probability corresponding to the design vector, and solves the multi-objective optimization problem by using a multi-objective evolutionary algorithm to generate a Pareto optimal solution set; and the verification and decision module 250 verifies and decides based on the Pareto optimal solution set to output a final design scheme.

[0060] As shown in Figure 3As shown, another aspect of the present application discloses an electronic device 300, comprising: a processor 301 and a memory 302. Wherein, the processor 301 and the memory 302 are connected, such as connected through a bus 303. Further, the electronic device 300 can also comprise a transceiver 304. It should be noted that the transceiver 304 is not limited to one in actual application, and the structure of the electronic device 300 does not constitute a limitation on the embodiments of the present application. Wherein, the processor 301 is applied to the embodiments of the present application, and is used to realize the collaborative optimization design method based on the agent model and defect warning. The processor 301 can be a CPU, a general processor, a DSP, an ASIC, an FPGA or other programmable logic devices, transistor logic devices, hardware components or any combination thereof. It can realize or execute various exemplary logical blocks, modules and circuits described in combination with the disclosure of the present application. The processor 301 can also be a combination of computing functions, such as one or more microprocessor combinations, combinations of DSP and microprocessor, etc. The bus 303 can include a channel for transmitting information between the above components. The bus 303 can be a PCI bus or an EISA bus, etc. The bus 303 can be divided into an address bus, a data bus, a control bus, etc. For the sake of representation, Figure 3 The memory 302 can be a ROM or other types of static storage devices that can store static information and instructions, a RAM or other types of dynamic storage devices that can store information and instructions, and can also be an EEPROM, a CD-ROM or other optical disk storage, an optical disk storage (including a compact disk, a laser disk, an optical disk, a digital versatile disk, a Blu-ray disk, etc.), a magnetic disk storage medium or other magnetic storage device, or any other medium capable of carrying or storing desired program codes in the form of instructions or data structures and capable of being accessed by a computer, but not limited to this. The memory 302 is used to store the application program code for executing the scheme of the present application, and is controlled by the processor 301 to execute. The processor 301 is used to execute the application program code stored in the memory 302 to realize the steps of the collaborative optimization design method based on the agent model and defect warning provided by the present application.

[0061] On the other hand, the embodiments of the present application provide a storage medium having a computer program stored thereon, which is executed by a processor to perform the steps of the collaborative optimization design method based on the agent model and defect warning as executed by the server above.

[0062] The above description is only the preferred embodiments of the present application, and does not limit the patent scope of the present application. Any equivalent structural transformation made in the concept of the present application, or direct / indirect application in other related technical fields, is included in the patent protection scope of the present application.

Claims

1. A collaborative optimization design method based on surrogate model and defect early warning, applied to transformer design, characterized in that, Includes the following steps: Step S1: Obtain the user's multi-mode design requirements for the transformer; Step S2: Based on the trained multimodal large model, analyze the multimodal design requirements and generate parameterized design vectors; Step S3: Perform performance and reliability analysis on the design vector based on the dual-engine parallel analysis framework to obtain the performance index and risk probability corresponding to the design vector; Step S4: Based on the performance index and risk probability corresponding to the design vector, construct a multi-objective optimization problem, and use a multi-objective evolutionary algorithm to solve the multi-objective optimization problem to generate a Pareto optimal solution set; Step S5: Verify and make decisions based on the Pareto optimal solution set, and output the final design scheme.

2. The collaborative optimization design method based on surrogate model and defect early warning as described in claim 1, characterized in that: The generation of parameterized design vectors involves performing the following steps for a multimodal large model: Extract feature vectors from multimodal design requirements; The feature vectors of different modalities are fused to obtain the representation vector; The representation vector is mapped to the key parameters of the parameterized model to obtain the design vector.

3. The collaborative optimization design method based on surrogate model and defect early warning as described in claim 1, characterized in that: The dual-engine parallel analysis framework includes a PINN surrogate model and a defect risk prediction model. The PINN surrogate model takes a design vector and spatial coordinates as input and outputs a physical field and performance indicators as output. The defect risk prediction model takes a design vector as input and outputs the probability of various defects.

4. The collaborative optimization design method based on surrogate model and defect early warning as described in claim 3, characterized in that: The construction and training of the PINN agent model includes the following steps: Construct a system that takes design vectors and spatial coordinates as input, and physical fields and performance indicators as input. Use datasets obtained from simulations or experiments as supervised data for model training; A composite loss function is constructed, consisting of a data loss term, a physical loss term, and a boundary condition loss term. The data loss term measures the difference between the PINN surrogate model's predicted values ​​and the supervised data. The physical loss term is obtained by sampling a set of configuration points within the computational domain, substituting the physical fields predicted by the PINN surrogate model into its corresponding governing equations, and calculating the residuals. The boundary condition loss term is obtained by sampling a set of points on the boundary and calculating the difference between the PINN surrogate model's predicted values ​​and the preset boundary conditions. The composite loss function is minimized by an optimization algorithm to optimize the weight parameters of the PINN proxy model until the loss converges, thus obtaining the trained PINN proxy model.

5. The collaborative optimization design method based on surrogate model and defect early warning as described in claim 3, characterized in that: The construction and training of the defect risk prediction model includes the following steps: Obtain a training dataset containing multiple defect samples. Each training sample includes a defect image, a design vector corresponding to the defect sample, and a defect label. Defect feature vectors are extracted from defect images using a pre-trained feature extractor. The defect feature vector is concatenated with the design vector corresponding to the defect sample to generate a fused feature vector; Construct a classifier model that includes a Softmax output layer; The fused feature vector is input into the classification model, and the original score vector is obtained through forward propagation. The original score vector is converted into a defect probability distribution through the Softmax output layer to obtain the risk probability of various defects. With the goal of maximizing the log-likelihood function, the parameters of the classifier model are optimized through an optimization algorithm until convergence, resulting in a trained classifier model used as a defect risk prediction model.

6. The collaborative optimization design method based on surrogate model and defect early warning as described in claim 1, characterized in that: The multi-objective evolutionary algorithm is the NSGA-II algorithm, which solves the multi-objective optimization problem in the following steps: Step S41: Randomly generate an initial population containing multiple design vectors; Step S42: For each design vector in the current population, obtain its corresponding performance index and risk probability through a dual-engine parallel analysis framework; Step S43: Based on the prediction results of all design vectors, perform non-dominated sorting, selection, crossover and mutation operations on the population to generate a new offspring population; Step S44: Repeat steps S42 and S43 above until the iteration termination condition is met, and output the final Pareto optimal solution set.

7. The collaborative optimization design method based on surrogate model and defect early warning as described in claim 3, characterized in that: Verification based on the Pareto optimal solution set includes the following steps: Each nondominated solution in the Pareto optimal solution set is input into the defect risk prediction model to obtain the risk probability of each type of defect occurring for each nondominated solution. Compare the risk probability of each type of defect occurring in each non-dominated solution with the risk threshold, and select the non-dominated solution whose risk probability of each type of defect does not exceed the risk threshold as the design scheme to be decided. If the probability of at least one type of defect occurring in a nondominated solution is greater than the risk threshold, the design parameters that cause the defect to occur can be deduced by using the defect risk prediction model.

8. A collaborative optimization design system based on surrogate model and defect early warning, characterized in that, The collaborative optimization design method based on surrogate model and defect early warning, as described in any one of claims 1-7, includes: an acquisition module, an analysis module, a prediction module, an optimization module, and a verification and decision-making module. The acquisition module is used to acquire the user's multimodal design requirements for the transformer design. The multimodal design requirements include any one or more of the following: natural language description text, design sketches, or a list of key parameters. The parsing module analyzes the multimodal design requirements based on the trained multimodal large model, generating design vectors and corresponding transformer 3D models; The prediction module performs performance and reliability analysis on the design vector based on a dual-engine parallel analysis framework, and obtains the performance index and risk probability corresponding to the design vector. The optimization module constructs a multi-objective optimization problem based on the performance index and risk probability corresponding to the design vector, and uses a multi-objective evolutionary algorithm to solve the multi-objective optimization problem to generate a Pareto optimal solution set. The verification and decision-making module performs verification and decision-making based on the Pareto optimal solution set, and outputs the final design scheme.

9. An electronic device, characterized in that: It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the method according to any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method described in any one of claims 1-7.

Images