Stealth structure multi-objective optimization method and system based on transfer learning and domain adaptation

Abstract

The application discloses a stealth structure multi-objective optimization method based on transfer learning and domain adaptation, which comprises the following steps: mapping stealth structure parameters to a unified feature space, and constructing a cross-domain parameter space model; training a domain adaptation transfer network model based on the cross-domain parameter space model; the domain adaptation transfer network model comprises a double-channel transfer module, which aligns the performance distribution of a source domain and a target domain through adversarial learning; optimizing the parameters of the trained domain adaptation transfer network model based on actual measurement data of the target domain; and obtaining an optimized stealth structure parameter combination which takes into account the stealth performance and mechanical performance based on the optimized domain adaptation transfer network model. The application further discloses a stealth structure multi-objective optimization system based on transfer learning and domain adaptation for implementing the above method. The application can realize efficient optimization of stealth structure parameters across scenes, and can be widely applied to the field of electromagnetic compatibility devices and the like, and has the advantages of high optimization efficiency and strong engineering applicability.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary fields of electromagnetic stealth, transfer learning and structural mechanics, and specifically relates to a multi-objective optimization method and system for stealth structures based on transfer learning and domain adaptation. Background Technology

[0002] In stealth structure design, parameter optimization needs to balance electromagnetic stealth performance (such as broadband reflectivity) and mechanical load-bearing capacity, but traditional methods have the following drawbacks:

[0003] Poor domain adaptability: Parameters optimized in the simulation environment may experience a significant decrease in stealth performance in actual manufacturing scenarios due to factors such as fluctuations in material properties and process errors. For example, the honeycomb structure parameters optimized in the simulation may reduce the -10dB bandwidth by more than 30% during actual 3D printing due to deviations in the material's dielectric constant.

[0004] Strong data dependence: Traditional deep learning optimization requires a large amount of target domain (real-world scenario) data, but electromagnetic simulation is costly (a single full-band simulation takes several hours), and the acquisition of actual engineering data is limited by confidentiality requirements, making it difficult to obtain sufficient labeled data.

[0005] Insufficient generalization ability: When existing algorithms are optimized for a single scenario (such as X-band stealth), they cannot be quickly migrated to other frequency bands (such as Ku band) or other structure types (such as migrating from a cellular structure to a woodpile structure), resulting in repeated optimization that is time-consuming and labor-intensive.

[0006] In existing technologies, some studies have attempted to use traditional transfer learning methods, but these methods do not consider the electromagnetic-mechanical coupling characteristics of the stealth structure parameter space. As a result, the transferred parameters often exhibit a "negative transfer" phenomenon (e.g., improved mechanical performance but a sharp drop in stealth bandwidth). Therefore, there is an urgent need for an intelligent optimization method that can adapt to cross-domain differences and reduce the dependence on target domain data to achieve efficient generalization optimization of stealth structure parameters. Summary of the Invention

[0007] The purpose of this invention is to overcome the above-mentioned defects and provide a multi-objective optimization method and system for stealth structures based on transfer learning and domain adaptation. This invention solves the technical problems of low optimization efficiency and poor generalization ability of stealth structure parameters in different scenarios. By constructing a knowledge transfer mechanism between the source domain (such as a simulation environment) and the target domain (such as an actual engineering scenario), this invention achieves efficient cross-scenario optimization of stealth structure parameters. It can be widely used in fields such as electromagnetic compatibility devices and has the advantages of high optimization efficiency and strong engineering applicability.

[0008] To achieve the above-mentioned objectives, the present invention provides the following technical solution:

[0009] This invention presents a multi-objective optimization method for stealth structures based on transfer learning and domain adaptation. By constructing a three-layer framework of "source domain knowledge extraction - cross-domain difference adaptation - target domain fine-tuning", it enables rapid transfer and optimization of stealth structure parameters in different scenarios, thereby improving the engineering applicability of the parameters.

[0010] This invention provides a multi-objective optimization method for stealth structures based on transfer learning and domain adaptation, comprising:

[0011] a. Cross-domain parameter space modeling: High-level features of source domain cloaking structure parameters are extracted through an autoencoder to construct a unified feature space;

[0012] b. Domain Adaptation Transfer Network Training: Design a dual-channel adaptation module for the electromagnetic and mechanical domains, and align the performance distribution of the source and target domains through adversarial learning;

[0013] c. Target domain lightweight fine-tuning: Utilizing a few-sample meta-learning algorithm, the transfer model parameters are optimized based on measured data from the target domain;

[0014] d. Cross-domain parameter optimization: Combine target domain constraints to generate optimized parameter combinations that take into account both stealth performance and mechanical performance.

[0015] The source domain is the electromagnetic simulation environment, and the target domain is the actual processing or application scenario. Cross-domain differences include material parameter fluctuations, process errors, etc.

[0016] The dual-channel migration module includes:

[0017] Electromagnetic domain adaptation module: Employs a domain adversarial neural network to align reflectivity feature distribution;

[0018] Mechanical Domain Adaptation Module: Suppresses inter-domain differences in mechanical performance characteristics through a gradient inversion layer.

[0019] The target domain lightweight fine-tuning uses the Model Independent Meta-Learning (MAML) algorithm, which requires only 5-10 sets of actual target domain data.

[0020] The loss function for multi-task joint optimization considers both reflectivity bandwidth and bending strength, and the weighting coefficients are determined through cross-validation.

[0021] A stealth structure cross-domain parameter optimization system includes:

[0022] a. Data modeling module: used for source domain data acquisition and feature space mapping;

[0023] b. Transfer learning module: Implements domain adaptation network training and multi-task optimization;

[0024] c. Fine-tuning and optimization module: Completes model adaptation and parameter generation based on target domain data;

[0025] d. Verification output module: Verify the engineering feasibility of the optimized parameters and output the results.

[0026] The transfer learning module includes an electromagnetic domain adversarial network and a mechanical domain gradient inversion layer, supporting cross-domain feature alignment.

[0027] The fine-tuning and optimization module integrates meta-learning algorithms, supporting rapid model adaptation in scenarios with few samples.

[0028] Specifically, the present invention provides a multi-objective optimization method for stealth structures based on transfer learning and domain adaptation, comprising:

[0029] S1 obtains a cross-domain parameter space model containing a unified feature space and a steganographic structure-feature mapping relationship based on the basic parameters of the steganographic structure and the source domain dataset.

[0030] S2 trains the domain adaptation transfer network model based on the cross-domain parameter space model to obtain a preliminarily adapted domain adaptation transfer network model for cross-domain scenarios. The domain adaptation transfer network model includes a dual-channel transfer module, which aligns the performance distribution of the source domain and the target domain through adversarial learning.

[0031] S3 is a domain adaptation transfer network model optimized and trained based on measured data from the target domain.

[0032] Based on the optimized domain-adaptive migration network model, S4 obtains an optimized combination of stealth structure parameters that balances stealth performance and mechanical performance.

[0033] The above steps can also be further described as follows:

[0034] S1 constructs a cross-domain parameter space model by mapping the stealth structure parameters to a unified feature space to eliminate parameter coupling interference. Its inputs are the basic parameters of the stealth structure (such as dielectric constant ε, magnetic permeability μ, etc.) and the source domain dataset (D_S={X_S,Y_S}). After processing, the output is a cross-domain parameter space model containing a unified feature space and parameter-feature mapping relationships.

[0035] S2 is used to train the domain adaptation transfer network model. By leveraging adversarial learning, the performance distributions of the source and target domains are aligned, enabling the network to have cross-domain mapping capabilities. The inputs are the cross-domain parameter space model output by S1 and the performance distribution data of the source and target domains. The final output is a pre-trained domain adaptation transfer network model that is initially adapted to the cross-domain scenario.

[0036] The core of S3 is to optimize the parameters of the transfer network model. It combines the measured data of the target domain to fine-tune the model to improve the accuracy of parameter prediction. The input is the trained model obtained from S2, the measured data of the target domain (D_T), and the bi-objective loss function. After processing, the output is an optimized domain-adaptive transfer network model that is accurately adapted to the target domain.

[0037] S4 aims to generate optimized structural parameters, filtering parameters according to target domain engineering constraints (such as thickness and weight limitations) to obtain a solution that balances both performance and mechanical properties. The input is the optimized model output by S3 and the target domain engineering constraints. The final output is an optimized combination of stealth structural parameters that balances stealth and mechanical performance, thus achieving the invention's objective.

[0038] Furthermore, the source domain is the electromagnetic simulation environment, the target domain is the actual processing or application scenario, and cross-domain differences include material parameter fluctuations or process errors.

[0039] Furthermore, step S1 includes:

[0040] S1.1 Generate the source domain dataset, D_S={X_S, Y_S}, where X_S is the cloaking structure parameter vector and Y_S is the performance label;

[0041] S1.2 Construct an autoencoder for extracting high-level features of cloaked structure parameters and complete feature space mapping.

[0042] Furthermore, the stealth structure parameters include dielectric constant, permeability, honeycomb side length or structure thickness. The range of stealth structure parameters is: dielectric constant ε∈[2, 15], permeability μ∈[1, 5], honeycomb side length a∈[2mm, 10mm], and structure thickness d∈[5mm, 20mm].

[0043] Performance labels include reflectivity bandwidth or intensity.

[0044] Furthermore, in step S1.2, the autoencoder structure includes three fully connected layers to capture the coupling relationship between stealth structural parameters and stealth-mechanical properties.

[0045] Furthermore, in step S2, the dual-channel migration module includes an electromagnetic domain channel and a mechanical domain channel;

[0046] Electromagnetic domain channel: Adapting reflectivity feature distribution using a domain adversarial neural network;

[0047] Mechanical domain channel: suppresses inter-domain differences in mechanical performance characteristics through gradient inversion layer.

[0048] Furthermore, in step S3, the stealth-mechanical bi-target loss function is:

[0049] L = L_perf + αL_adv_E + βL_adv_M

[0050] L_perf = w1L_reflect + w2L_strength

[0051] Where L_reflect is the reflectivity bandwidth prediction loss, with a target bandwidth ≥ 15 GHz; L_strength is the bending strength prediction loss, with a target strength ≥ 100 MPa; L_adv_E is the electromagnetic domain adversarial loss, and L_adv_M is the mechanical domain adversarial loss.

[0052] w1 is the weighting coefficient for the reflection bandwidth prediction loss, w2 is the weighting coefficient for the bending strength prediction loss, α is the weighting coefficient for the electromagnetic domain countermeasure loss, and β is the weighting coefficient for the mechanical domain countermeasure loss; w1, w2, α, and β are determined through cross-validation.

[0053] Furthermore, step S4 includes:

[0054] S4.1 Collect 5-10 sets of measured data D_T={X_T, Y_T} in the target domain; where X_T is the actual processed honeycomb structure parameter, and Y_T is the measured reflectivity or measured bending strength;

[0055] S4.2 uses the Model Independent Meta-Learning (MAML) algorithm to adapt the target domain.

[0056] Furthermore, step S4.2 includes:

[0057] Initialize the parameters of the domain adaptation migration network model. These are pre-trained parameters;

[0058] For each target domain sample (x_i, y_i), calculate the gradient ▽θL(f_θ(x_i), y_i), and generate temporary parameters θ_i = - η▽θL; where:

[0059] ▽θL: refers to the gradient of the loss function L with respect to the model parameter θ, reflecting the trend of the loss value L as the parameter θ changes, and is used to indicate the direction of parameter adjustment (e.g., if the gradient is positive, the parameter needs to be decreased, and if it is negative, the parameter needs to be increased).

[0060] f_θ(x_i): represents the predicted output of the model constructed with parameter θ for the i-th set of input parameters x_i in the target domain (such as predicted reflectivity, bending strength).

[0061] y_i: The measured performance label (such as measured reflectivity, measured bending strength) corresponding to the i-th group of input parameters x_i in the target domain, which is the reference standard for model prediction;

[0062] η: Learning rate, used to control the magnitude of parameter updates and avoid excessive or insufficient parameter adjustment due to gradient fluctuations (η=0.01 in the document example).

[0063] θ_i: Temporary parameters generated based on the i-th set of target domain samples; these are the initial parameters. The intermediate parameters, adjusted in the opposite direction of the gradient, are used for subsequent meta-loss calculations.

[0064] Calculate the meta-loss L_meta = ΣL (f_θ_i (x_j), y_j), and then... Updated to - μ▽θL_meta.

[0065] L_meta: Meta-loss is the sum of prediction errors of the model under multiple sets of temporary parameters. It is used to evaluate the overall fit of temporary parameters to the target domain and ensure that the model has stable cross-domain performance after fine-tuning.

[0066] f_θ_i (x_j): The model constructed with temporary parameters θ_i, predicting the output of the j-th set of input parameters x_j in the target domain (such as predicting reflectivity or bending strength).

[0067] x_j and y_j: are the j-th group of measured structural parameters (such as the cell edge length) and the corresponding measured performance labels (such as measured bending strength) in the target domain, respectively, and serve as the reference benchmark for model prediction.

[0068] μ: The learning rate of meta-learning, which controls the global parameter update magnitude (μ=0.001 in the document example) to avoid excessive parameter adjustment leading to model instability.

[0069] ▽θL_meta: Meta-loss L_meta relative to initial parameters The gradient indicates The optimization direction is to ensure that the updated parameters are more suitable for the target domain.

[0070] A stealth structure multi-objective optimization system based on transfer learning and domain adaptation is provided to implement the aforementioned stealth structure multi-objective optimization method based on transfer learning and domain adaptation. The system includes a data modeling module, a transfer learning module, a fine-tuning optimization module, and a verification output module.

[0071] The data modeling module is used for source domain data acquisition and feature space mapping;

[0072] The transfer learning module is used to implement domain adaptation network training and multi-task optimization; the transfer learning module includes an electromagnetic domain adversarial network and a mechanical domain gradient reversal layer;

[0073] The fine-tuning and optimization module is used to complete model adaptation and parameter generation based on target domain data; the fine-tuning and optimization module integrates meta-learning algorithms;

[0074] The verification output module is used to verify the engineering feasibility of the optimized parameters and output the results.

[0075] Compared with the prior art, the present invention has at least one of the following advantages:

[0076] (1) This invention has cross-domain generalization capability: through domain adaptation technology, the performance retention rate of simulation-optimized parameters in actual scenarios is increased to over 90%, which is 40% higher than that of traditional methods;

[0077] (2) The present invention has high data efficiency: the target domain only requires 5-10 sets of measured data to complete the optimization, which reduces the sample requirement by 90% compared with full data training;

[0078] (3) The present invention can achieve multi-objective collaborative optimization: it takes into account both electromagnetic stealth and mechanical performance, avoiding the problem of single-objective optimization leading to the deterioration of another objective in traditional methods. Attached Figure Description

[0079] Figure 1 This is a flowchart of a multi-objective optimization method for stealth structures based on transfer learning and domain adaptation, according to the present invention. Detailed Implementation

[0080] The features and advantages of the present invention will become clearer and more apparent from the following detailed description.

[0081] The term “exemplary” as used herein means “serving as an example, embodiment, or illustration.” Any embodiment illustrated herein as “exemplary” is not necessarily to be construed as superior to or better than other embodiments. Although various aspects of embodiments are shown in the accompanying drawings, the drawings are not necessarily drawn to scale unless specifically indicated otherwise.

[0082] This invention proposes a multi-objective optimization method for stealth structures based on transfer learning and domain adaptation. This method achieves efficient cross-scenario optimization of stealth structure parameters by constructing a knowledge transfer mechanism between the source domain (simulation environment) and the target domain (actual engineering scenario). This includes steps such as cross-domain parameter space modeling, domain-adaptive transfer network training, and lightweight fine-tuning of the target domain. This invention solves the problems of poor cross-domain adaptability and strong data dependence of traditional methods. Parameter optimization can be achieved with only a small amount of target domain data, making it widely applicable in fields such as electromagnetic compatibility devices. It has the advantages of high optimization efficiency and strong engineering applicability.

[0083] To achieve the above objectives, the present invention adopts the following technical solution, the overall route of which is as follows: Figure 1 As shown:

[0084] Cross-domain parameter space modeling: The stealth structure parameters (such as material dielectric constant and lattice size) are mapped to a unified feature space. High-level features of source domain (such as CST electromagnetic simulation) data are extracted through autoencoders to capture the coupling relationship between parameters and stealth-mechanical properties.

[0085] Domain Adaptive Transfer Network: A dual-channel transfer module is designed, comprising electromagnetic domain adaptation and mechanical domain adaptation. The former aligns the reflectivity distributions of the source and target domains (e.g., in actual manufacturing scenarios) through adversarial training, reducing performance deviations caused by material parameter fluctuations. The latter utilizes a gradient reversal layer to suppress inter-domain differences in mechanical performance characteristics, ensuring the stability of the structure's load-bearing capacity.

[0086] Multi-task joint optimization: Construct a stealth-mechanical dual-objective loss function, simultaneously optimize reflectivity bandwidth (objective: ≥15GHz) and bending strength (objective: ≥100MPa), and dynamically adjust parameter weights through an attention mechanism.

[0087] Lightweight fine-tuning of the target domain: The transfer model is fine-tuned using a small amount of target domain data (such as 5-10 sets of actual test samples) to quickly adapt to parameter shifts in specific scenarios through meta-learning.

[0088] Example:

[0089] The method proposed in this invention is implemented in accordance with the following four steps:

[0090] Step 1: Cross-domain parameter space modeling

[0091] 1.1 Source Domain Data Acquisition: Generate multiple sets of stealth structure parameter samples in the source domain (electromagnetic simulation environment):

[0092] Parameter range: dielectric constant ε∈[2, 15], magnetic permeability μ∈[1, 5], cell side length a∈[2mm,10mm], structural thickness d∈[5mm, 20mm].

[0093] Simulation verification: The reflectivity of each set of parameters was simulated in the 2-40GHz range using the CST Microwave Studio, and the bending strength was calculated simultaneously using ANSYS to form a source domain dataset D_S={X_S, Y_S}, where X_S is the parameter vector and Y_S is the performance label (reflectivity bandwidth, intensity).

[0094] 1.2 Feature Space Mapping: Construct an autoencoder AE(X) = (encoder E(X), decoder D(E(X))), and extract high-level features of the parameters through unsupervised learning:

[0095] Encoder structure: 3 fully connected layers (number of neurons: 128→64→32), capturing the electromagnetic-mechanical coupling relationship between parameters.

[0096] Loss function: , where λ is the sparsity regularization coefficient (0.01).

[0097] L_ae: The total loss function of the autoencoder (AE), used to measure the accuracy of the autoencoder in feature extraction and reconstruction of cloaked structure parameters. The smaller the value, the better the model's handling of the parameters.

[0098] X: The original cloaked structure parameter vector (such as a vector composed of dielectric constant, cell side length, etc.) of the input autoencoder is the original data for model learning.

[0099] E(X): The result of the encoder (E) processing the original parameter X, that is, the high-level features of the extracted stealth structure parameters, which are used to capture the electromagnetic-mechanical coupling relationship between parameters.

[0100] D(E(X)): The reconstruction result of the high-level feature E(X) by the decoder (D), that is, the extracted feature is restored to a parameter vector with the same dimension as the original parameter X, which is used to evaluate the effectiveness of the feature.

[0101] The mean squared error term calculates the difference between the original parameter X and the reconstructed parameter D(E(X)). The smaller the difference, the stronger the decoder's ability to restore the features.

[0102] λ: Sparsity regularization coefficient, with a value of 0.01 in the document. Its function is to control the sparsity of features, avoid the model from over-relying on some parameter features, and improve the generalization ability of features.

[0103] L1 regularization term: calculates the sum of the absolute values ​​of high-level features E(X). When combined with λ, it can help the encoder extract more concise and key features and reduce redundant information.

[0104] Step 2: Training the Domain Adaptation Transfer Network

[0105] 2.1 Dual-channel migration module design:

[0106] Electromagnetic Domain Channel: A Domain Adversarial Neural Network (DANN) is employed. The discriminator D_E distinguishes the reflectivity features of the source and target domains, forcing the feature extractor F_E to generate domain-invariant features. The adversarial loss is given by L_adv_E = -E [log(D_E (F_E (X_S)))] - E [log (1-D_E (F_E (X_T)))], where X_T represents the target domain parameter samples. In other words, the electromagnetic domain loss L_adv_E consists of two parts. The first part, -E [log (D_E (F_E (X_S)))], involves the discriminator D_E identifying the reflectivity features of the source domain parameter X_S as source domain features after the feature extractor F_E obtains them. The loss is calculated using expectation and logarithm to ensure accurate classification of the source domain features. The latter part, -E [log (1-D_E (F_E (X_T)))], is used to extract features from the target domain parameter X_T via F_E. This makes it as difficult as possible for the discriminator D_E to identify its domain attributes (i.e., misclassify it as the source domain). The loss is calculated using expectation and logarithm to align the feature distributions of the target domain and the source domain.

[0107] Mechanical domain channel: The domain difference of mechanical performance features is suppressed by gradient inversion layer (GRL), and the loss function is L_adv_M = -L_mech (F_M (X), Y_M), where L_mech is the mechanical performance prediction loss.

[0108] 2.2 Multi-task joint optimization: The dual-objective loss function is constructed as follows,

[0109] L = L_perf + αL_adv_E + βL_adv_M

[0110] L_perf = w1L_reflect + w2L_strength

[0111] Where L_reflect is the reflectivity bandwidth prediction loss (mean square error), with a target bandwidth ≥ 15 GHz. L_adv_E is the electrical domain loss, and L_adv_M is the mechanical domain loss.

[0112] L_strength is the bending strength prediction loss (mean square error), with a target strength ≥ 100 MPa. w1=0.6, w2=0.4, α=0.2, β=0.2 (weighting coefficients are determined through cross-validation).

[0113] Step 3: Lightweight fine-tuning of the target domain

[0114] 3.1 Small Sample Data Collection: Collect 5-10 sets of measured data D_T={X_T, Y_T} in the target domain (e.g., an actual 3D printing scenario), including:

[0115] Actual parameters of the honeycomb structure (including material errors: dielectric constant fluctuation ±5%, dimensional error ±0.1mm).

[0116] The reflectivity was measured by the bow-shaped reflection method, and the bending strength was measured by the universal testing machine.

[0117] 3.2 Meta-learning Fine-tuning: Employing the Model-Independent Meta-Learning (MAML) algorithm for rapid adaptation to the target domain:

[0118] 1) Initialize the migration model parameters These are the pre-trained parameters.

[0119] 2) For each target domain sample (x_i, y_i), calculate the gradient ▽θL (f_θ(x_i), y_i), and generate temporary parameters θ_i = - η▽θL.

[0120] 3) Calculate the meta-loss L_meta = ΣL (f_θ_i (x_j), y_j), and update the global parameters. ← - μ▽θL_meta.

[0121] The learning rate η = 0.01, μ = 0.001, and the number of iterations is 50-100.

[0122] Step 4: Cross-domain parameter optimization and verification

[0123] 4.1 Optimization of Parameter Generation: The training method predicts the optimal parameters in the target domain using a transfer model.

[0124] Input: Target domain constraints (e.g., thickness ≤ 10mm, weight ≤ 0.5kg).

[0125] Output: Optimized parameter combination (e.g., ε=8.5, μ=2.3, a=6mm, d=8mm), predicted reflectivity bandwidth 16.2GHz, bending strength 112MPa.

[0126] 4.2 Engineering Validation: Samples were prepared by 3D printing. The measured reflectivity bandwidth was 15.8 GHz (with a prediction error of 2.5%) and the bending strength was 108 MPa (with a prediction error of 3.6%), validating the effectiveness of cross-domain optimization.

[0127] Through the above steps, this invention achieves cross-domain intelligent optimization of stealth structure parameters, solves the problem of performance degradation of traditional methods in different scenarios, and significantly improves the efficiency and engineering applicability of parameter optimization.

[0128] The present invention has been described in detail above with reference to specific embodiments and exemplary examples; however, these descriptions should not be construed as limiting the present invention. Those skilled in the art will understand that various equivalent substitutions, modifications, or improvements can be made to the technical solutions and embodiments of the present invention without departing from the spirit and scope of the invention, and all such modifications and improvements fall within the scope of the present invention. The scope of protection of the present invention is defined by the appended claims.

[0129] The contents not described in detail in this specification are common knowledge to those skilled in the art.

Claims

1. A multi-objective optimization method for stealth structures based on transfer learning and domain adaptation, characterized in that, include: S1 obtains a cross-domain parameter space model containing a unified feature space and a steganographic structure-feature mapping relationship based on the basic parameters of the steganographic structure and the source domain dataset. S2 trains the domain adaptation transfer network model based on the cross-domain parameter space model to obtain a preliminarily adapted domain adaptation transfer network model for cross-domain scenarios. The domain adaptation transfer network model includes a dual-channel transfer module, which aligns the performance distribution of the source domain and the target domain through adversarial learning. S3 is a domain adaptation transfer network model optimized and trained based on measured data from the target domain. S4, based on the optimized domain-adaptive migration network model, obtains an optimized combination of stealth structure parameters that balances stealth performance and mechanical performance. Step S1 includes: S1.1 Generate the source domain dataset, D_S={X_S, Y_S}, where X_S is the cloaking structure parameter vector and Y_S is the performance label; S1.2 Construct an autoencoder for extracting high-level features of stealth structure parameters and complete feature space mapping; In step S1.2, the autoencoder structure includes three fully connected layers, which are used to capture the coupling relationship between stealth structural parameters and stealth-mechanical performance. In step S2, the dual-channel migration module includes an electromagnetic domain channel and a mechanical domain channel; Electromagnetic domain channel: Adapting reflectivity feature distribution using a domain adversarial neural network; Mechanical domain channel: suppresses inter-domain differences in mechanical performance characteristics through gradient inversion layer; In step S3, the stealth-mechanical bi-target loss function is: L = L_perf + αL_adv_E + βL_adv_M L_perf = w1L_reflect + w2L_strength Where L_reflect is the reflectivity bandwidth prediction loss, with a target bandwidth ≥ 15 GHz; L_strength is the bending strength prediction loss, with a target strength ≥ 100 MPa; L_adv_E is the electromagnetic domain adversarial loss, and L_adv_M is the mechanical domain adversarial loss; w1 is the weighting coefficient for the reflection bandwidth prediction loss, w2 is the weighting coefficient for the bending strength prediction loss, α is the weighting coefficient for the electromagnetic domain countermeasure loss, and β is the weighting coefficient for the mechanical domain countermeasure loss; w1, w2, α, and β are determined through cross-validation. Step S4 includes: S4.1 Collect 5-10 sets of measured data D_T={X_T, Y_T} in the target domain; where X_T is the actual processed honeycomb structure parameter, and Y_T is the measured reflectivity or measured bending strength; S4.2 employs the Model Independent Meta-Learning Algorithm (MAML) to adapt the target domain; Step S4.2 includes: Initialize the parameters of the domain adaptation migration network model. These are pre-trained parameters; For each set of target domain samples (x_i, y_i), calculate the gradient. θL (f_θ(x_i), y_i), generates temporary parameters θ_i= - η θL; where: θL refers to the gradient of the loss function L with respect to the model parameters θ; f_θ(x_i) represents the predicted output of the model built with parameters θ for the i-th set of input parameters x_i in the target domain; y_i represents the measured performance label corresponding to the i-th set of input parameters x_i in the target domain. The initial parameters of the domain adaptation transfer network are: η is the learning rate; θ_i is the temporary parameter generated based on the i-th set of target domain samples. Calculate the meta-loss L_meta = ΣL (f_θ_i (x_j), y_j), and then... Updated to - μ θL_meta; where L_meta represents the meta-loss, f_θ_i(x_j) represents the predicted output of the model constructed with temporary parameters θ_i for the j-th set of input parameters x_j in the target domain, and x_j and y_j: the j-th set of measured structural parameters and corresponding measured performance labels in the target domain, respectively. θL_meta is the meta-loss L_meta with respect to the initial parameters. The gradient.

2. The method for multi-objective optimization of stealth structures based on transfer learning and domain adaptation according to claim 1, characterized in that, The source domain is the electromagnetic simulation environment, and the target domain is the actual processing or application scenario. Cross-domain differences include material parameter fluctuations or process errors.

3. The method for multi-objective optimization of stealth structures based on transfer learning and domain adaptation according to claim 1, characterized in that, The stealth structure parameters include dielectric constant, permeability, honeycomb side length or structure thickness. The range of stealth structure parameters is: dielectric constant ε∈[2, 15], permeability μ∈[1, 5], honeycomb side length a∈[2mm, 10mm], and structure thickness d∈[5mm, 20mm]. Performance labels include reflectivity bandwidth or intensity.

4. A stealth structure multi-objective optimization system based on transfer learning and domain adaptation, characterized in that, The method for implementing the stealth structure multi-objective optimization method based on transfer learning and domain adaptation as described in any one of claims 1-3 includes a data modeling module, a transfer learning module, a fine-tuning optimization module, and a verification output module. The data modeling module is used for source domain data acquisition and feature space mapping; The transfer learning module is used to implement domain adaptation network training and multi-task optimization; the transfer learning module includes an electromagnetic domain adversarial network and a mechanical domain gradient reversal layer; The fine-tuning and optimization module is used to complete model adaptation and parameter generation based on target domain data; the fine-tuning and optimization module integrates meta-learning algorithms; The verification output module is used to verify the engineering feasibility of the optimized parameters and output the results.

Images