Concealed structure multi-target distributed optimization method based on federal learning

Abstract

The invention discloses a federated learning-based stealth structure multi-target distributed optimization method, which comprises the following steps of: constructing a star federated learning architecture which comprises a central server and N distributed nodes; each node carries out model training based on local data and uploads a gradient to the central server; the central server aligns electromagnetic domain and mechanical domain feature distribution of each node through domain adversarial training based on the gradient of each node to realize cross-domain parameter optimization; performing weighted aggregation on each domain result to obtain a global gradient; global model parameters of the central server are updated by using the global gradient and then issued to each node to complete a round of iteration; and the central server periodically acquires the generalization error of the local model of each node, and determines the communication priority of each node according to the generalization error. According to the invention, the data volume is compressed to 10MB / time through gradient transmission, and the communication efficiency is greatly improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of electromagnetic stealth, distributed machine learning and structural mechanics, and particularly relates to a multi-objective distributed optimization method for a stealth structure based on federated learning. BACKGROUND

[0002] The traditional parameter optimization of a stealth structure has the following core challenges: 1. Data leakage risk. The centralized optimization needs to store the electromagnetic simulation data and material test data of each unit, and if the sensitive data of the A-type unit and the B-type unit are mixed for training, there is a risk of sensitive information leakage. 2. Low efficiency of cross-domain optimization. The data of the stealth structure in different scenarios (such as an airplane and a ship) are significantly different in distribution, and the centralized model is difficult to adapt to the cross-domain parameter optimization requirements. For example, the stealth structure of an airplane requires lightweight (weight ≤ 0.3 kg), while the structure of a ship pays more attention to salt mist resistance (material corrosion resistance index ≥ 80), and the traditional method needs to retrain the model for different scenarios. 3. High communication cost. The amount of electromagnetic simulation data in the full frequency band is huge (about 1 GB of data is generated for a single simulation), and the centralized optimization needs to transmit the original data between the center node and each terminal, resulting in a communication bandwidth bottleneck.

[0003] In the prior art, some studies attempt to use distributed optimization, but do not consider the electromagnetic-mechanical coupling characteristics of the parameter space of the stealth structure, and lack an effective cross-domain migration mechanism. For example, the federated average (FedAvg) algorithm in the optimization of the stealth structure leads to a decrease of more than 50% in the convergence speed of the model due to the large difference in the data distribution of each node. Therefore, there is an urgent need for a distributed optimization method that takes into account data privacy, cross-domain adaptability and communication efficiency. SUMMARY

[0004] The purpose of the present application is to overcome the above-mentioned defects and provide a multi-objective distributed optimization method for a stealth structure based on federated learning, which solves the technical problems of privacy protection difficulty and poor cross-domain adaptability of the traditional parameter of the stealth structure. The present application compresses the data volume to 10 MB / time through gradient transmission, greatly improving the communication efficiency.

[0005] To achieve the above-mentioned application purposes, the present application provides the following technical solutions: The present application discloses a multi-objective distributed optimization method for a stealth structure based on federated learning, which realizes multi-node collaborative optimization through the construction of a three-layer framework of "local training - cross-domain migration - global aggregation". This method uses the federated learning architecture to enable the distributed nodes such as the A-type unit and the B-type unit to independently train the model based on local data, and only uploads the gradient parameters with added differential privacy noise. Through the domain adversarial neural network to align the electromagnetic domain features, the gradient inversion layer to suppress the differences in the mechanical domain, and the self-adaptive communication scheduling strategy to dynamically adjust the communication frequency of the nodes, the method solves the problems of data privacy leakage, poor cross-domain adaptability and high communication cost in the traditional centralized optimization.

[0006] The application is a stealth structure multi-objective distributed optimization method based on federated learning, comprising: a. Construct a federated learning architecture, including a central server and multiple distributed nodes, each node trains a model based on local data and uploads gradients to the central server; b. Add noise to the uploaded gradients using differential privacy technology to protect data privacy; c. Design a cross-domain migration aggregation strategy to align the electromagnetic domain and mechanical domain feature distribution of each node through domain adversarial training to achieve cross-domain parameter optimization; d. Construct a multi-objective joint loss function to cooperatively optimize stealth performance and mechanical performance locally at each node; e. Based on node data quality evaluation, dynamically adjust the communication frequency of each node and the central server.

[0007] The cross-domain migration aggregation strategy includes electromagnetic domain migration and mechanical domain migration, which aligns the domain features through domain adversarial neural networks and gradient reversal layers respectively.

[0008] The multi-objective joint loss function includes reflectivity bandwidth loss, mechanical strength loss, and domain adaptation loss, with weight coefficients of 0.5, 0.3, and 0.2 respectively.

[0009] The adaptive communication scheduling strategy dynamically adjusts the communication frequency according to the generalization error of the local model of each node. The smaller the generalization error, the higher the communication priority.

[0010] The distributed nodes include A-type nodes, B-type nodes, and C-type nodes, which store stealth structure data in different scenarios.

[0011] A system for implementing the above method, comprising: a. Central server: for model initialization, cross-domain migration, and global model aggregation; b. Distributed nodes: for local data training and gradient calculation; c. Communication module: for model parameter and gradient transmission between the central server and distributed nodes, and to implement differential privacy protection; d. Domain adaptation module: for cross-domain feature alignment of electromagnetic domain and mechanical domain; e. Optimization scheduling module: for multi-objective joint optimization and adaptive communication scheduling.

[0012] More specifically, the application includes: A stealth structure multi-objective distributed optimization method based on federated learning, comprising: S1 Construct a star-shaped federated learning architecture, which includes a central server and N distributed nodes; S2 Each node trains the model based on local data and uploads the gradient to the center server; the center server aligns the electromagnetic domain and mechanical domain feature distribution of each node based on the gradient of each node through domain adversarial training, realizes cross-domain parameter optimization, and obtains the results of each domain; The results of each domain are weighted and aggregated to obtain the global gradient; The global gradient is used to update the model parameters of the center server, and the center server distributes the updated model parameters to each node to complete one round of iteration; Repeat step S2 until the parameter value predicted by the local model of each node meets the design requirements; S3 The center server regularly obtains the generalization error of the local model of each node, and determines the communication priority of each node and the center server according to the generalization error.

[0013] Further, each node trains the local model based on local data using a stealth-mechanical dual-objective loss function.

[0014] Further, the training process of each node includes: Each node receives the initial model parameters distributed by the center server; According to the design requirements of the high-speed aircraft skin, a stealth-mechanical dual-objective loss function is used to calculate the multi-objective loss in the local model, and an Adam optimizer is used to calculate the model parameter gradient in the process.

[0015] Further, when each node uploads the local model parameter gradient to the center server, a differential privacy method is used to add Laplace noise to the uploaded gradient.

[0016] Further, the source or type of local data of each node is different; The types of local data include electromagnetic simulation data, mechanical performance data, or salt spray corrosion test data.

[0017] Further, the local data includes normalized electromagnetic simulation data, mechanical performance data, or salt spray corrosion test data; Among them, the electromagnetic simulation data is mapped to the [0, 1] interval; the mechanical performance data is standardized by Z-score, and the topology type is encoded as a one-hot vector; the salt spray corrosion test data is normalized using the mean variance method.

[0018] Further, in step S2, different nodes correspond to different domains, and the method of aligning the electromagnetic domain and mechanical domain feature distribution of each node through domain adversarial training includes: Domain-invariant features are generated using a domain adversarial neural network to align the electromagnetic domain feature distribution of each node; The gradient reversal layer is used to align the mechanical domain feature distribution of each node.

[0019] Further, the stealth-mechanical dual-target loss function includes reflectivity bandwidth loss, mechanical strength loss and domain adaptation loss, and the optimal weight combination of each loss is determined through hyperparameter search.

[0020] Further, in step S3, the smaller the generalization error is, the higher the communication priority is.

[0021] A stealth structure multi-objective distributed optimization system based on federated learning is used to implement the above-mentioned stealth structure multi-objective distributed optimization method based on federated learning, and includes a center server, a distributed node, a communication module, a domain adaptation module and an optimization scheduling module. The center server is used for model initialization, cross-domain migration and global model aggregation. The distributed node is used for local data training and gradient calculation. The communication module is used for model parameter and gradient transmission between the center server and the distributed node, and realizes differential privacy protection. The domain adaptation module is used to realize cross-domain feature alignment of the electromagnetic domain and the mechanical domain. The optimization scheduling module is used for multi-objective joint optimization and adaptive communication scheduling.

[0022] Compared with the prior art, the present application has at least one of the following beneficial effects: (1) The present application completes model iteration only through gradient transmission, which fundamentally avoids the risk of data leakage. (2) The present application proposes an adaptive communication scheduling strategy based on data quality, which reduces communication overhead, and at the same time, accelerates model convergence through a cross-domain migration aggregation strategy, realizes multi-objective balance of privacy protection, cross-domain optimization and communication efficiency, and provides an efficient and secure technical paradigm for collaborative design of stealth structures in various fields. BRIEF DESCRIPTION OF DRAWINGS

[0023] Figure 1 A flowchart of the stealth structure multi-objective distributed optimization method based on federated learning of the present application is shown in the figure. DETAILED DESCRIPTION

[0024] The characteristics and advantages of the present application will become clearer and more explicit with the following detailed description of the present application.

[0025] The word "exemplary" used herein means "serving as an example, embodiment or illustration". Any embodiment described herein as "exemplary" is not necessarily to be construed as preferred or advantageous over other embodiments. Although various aspects of the embodiments are shown in the drawings, the drawings are not necessarily drawn to scale unless specifically indicated.

[0026] The application is a stealth structure cross-domain multi-objective distributed optimization method based on a federated learning framework, which solves the problems of data privacy protection difficulty and poor cross-domain adaptability in traditional centralized optimization by constructing a multi-node collaborative optimization mechanism.

[0027] The application provides a stealth structure multi-objective distributed optimization method based on federated learning, which realizes multi-node collaborative optimization by constructing a three-layer framework of "local training - cross-domain migration - global aggregation", and improves the cross-domain adaptability and optimization efficiency of stealth structure parameters.

[0028] To achieve the above purpose, the application adopts the following technical solutions, and the specific flow chart is as shown in Figure 1 Firstly, a star-shaped federated learning architecture is constructed, including a center server and N distributed nodes, each node performs model training based on local data, only uploads model parameter gradient instead of original data, and differential privacy technology is used to add Laplace noise to the uploaded gradient to protect data privacy.

[0029] Then, a cross-domain migration aggregation strategy is designed, the reflectivity feature distribution of each node is aligned through domain adversarial training, domain invariant features are generated by using domain adversarial neural network, the difference in electromagnetic data distribution of different scenes (air and sea) is solved, the domain difference of mechanical performance features of each node is suppressed by using gradient reversal layer to ensure the consistency of structure bearing capacity, and the update of global model parameters is realized through a specific aggregation algorithm.

[0030] Then, a stealth-mechanical dual-objective loss function is constructed, which is optimized at each node, wherein the loss function includes reflectivity bandwidth loss, mechanical strength loss and domain adaptation loss, the optimal weight combination is determined through hyperparameter search to realize collaborative optimization of multi-objective.

[0031] Finally, the communication frequency is dynamically adjusted according to the data quality of the node, the generalization error of the local model on the validation set is calculated, the communication priority is determined according to the error size, efficient communication scheduling is realized, and the cross-domain optimization efficiency is improved while protecting the data privacy.

[0032] ​This invention achieves distributed optimization of stealth structure parameters through a federated learning framework. While protecting the data privacy of each node, it solves the problems of poor cross-domain adaptability and high communication costs in traditional centralized optimization. Specifically, differential privacy technology ensures that sensitive data from different units does not need to be centrally stored, and model iteration is completed only through gradient transmission, fundamentally avoiding the risk of data leakage. An adaptive communication scheduling strategy based on data quality reduces communication overhead, while a cross-domain migration aggregation strategy accelerates model convergence, achieving a multi-objective balance between privacy protection, cross-domain optimization, and communication efficiency. This provides an efficient and secure technical paradigm for the collaborative design of stealth structures.

[0033] Example: The method proposed in this invention is implemented in accordance with the following four steps: Step 1: Deployment of Federated Learning Nodes and Data Preprocessing 1.1 Node Architecture Setup: Construct a star topology consisting of one central server and three distributed nodes, with each node deployed in physical isolation. Node 1 (Class A Unit): Configured with a high-performance computing server to store electromagnetic simulation data of the aircraft skin, including 1000 sets of data such as 2-40GHz full-band reflectivity curves, dielectric constant (ε=3-12), and structural thickness (d=5-15mm), in the format of *.out files output by CST simulation.

[0034] Node 2 (Type B Unit): Deploy a material testing workstation to store mechanical property data of carbon fiber / epoxy resin composites, including 800 sets of test data such as flexural strength (S=80-150MPa), tensile strength, and density, which are derived from actual measurements by a universal testing machine.

[0035] Node 3 (Class C Unit): Configure an environmental simulation laboratory server to store salt spray corrosion test data of ship decks, including 600 sets of engineering data such as material corrosion resistance index (0-100), humidity tolerance range, and temperature cycling performance.

[0036] 1.2 Data Standardization Processing: Each node normalizes its local data: electromagnetic reflectivity values ​​are mapped to the [0,1] interval, mechanical parameters are standardized using Z-score, and topological types are encoded as one-hot vectors (e.g., honeycomb = 100, woodpile = 010, custom = 001). Node 1 adds access control permissions to the data based on the national security standard GB / T 45652-2025 "Security Specification for Pre-training and Optimization Training Data of Generative Artificial Intelligence in Network Security Technology", ensuring that feature values ​​can only be read during model training, and the raw data is not transmitted externally.

[0037] Step 2: Local Model Training and Gradient Generation Take node 1 as an example, the local training process is as follows: 2.1 Model initialization: receive the initial neural network parameters issued by the central server, and the model architecture is a 3-layer fully connected network.

[0038] 2.2 Loss function calculation: according to the design requirements of the aircraft skin (target bandwidth B_target=15GHz, bending strength S_target=100MPa), calculate the multi-objective loss.

[0039] 2.3 Gradient generation and privacy protection: use the Adam optimizer to calculate the parameter gradient, and add Laplace noise with scale parameter λ=0.5.

[0040] Step three: cross-domain migration and global model aggregation After the central server receives the gradients of each node, the following operations are performed: 3.1 Domain grouping and feature alignment: divide the gradients into 3 groups (A, B, and C domains) according to node type, and use domain adversarial neural network (DANN) to align electromagnetic domain features (reflectivity distribution). Use gradient reversal layer (GRL) to suppress domain differences for mechanical domain features (strength distribution) to ensure consistency of mechanical performance in different scenarios.

[0041] 3.2 Global aggregation algorithm: according to the human scoring, the results of each domain are weighted and aggregated, and the updated global gradient is used to update the model parameters, which are then sent to each node to complete a round of iteration. The original gradient has a serious "domain difference", such as the A domain is electromagnetic data (reflectivity 0-1), the B domain is mechanical data (strength 80-150MPa), and the feature dimension, value range, and physical meaning are completely different. Direct aggregation will lead to the global model being unable to adapt to the requirements of any domain. "Alignment" essentially translates the gradients of different domains into the same language, allowing different domains to have the same form. The results of each domain (A, B, and C domains) in this step are aligned gradient data output by each domain after step 3.1 domain grouping and feature alignment. It is not the original gradient, but the "standardized gradient" after cross-domain adaptation. Specifically: A domain result: electromagnetic simulation domain gradient (reflectivity related) → after DANN alignment, the gradient data without domain feature bias; B domain result: mechanical test domain gradient (strength related) → after GRL suppression of domain differences, the mechanical gradient data that can adapt to different scenarios; C domain result: corrosion engineering domain gradient (corrosion resistance, temperature and humidity tolerance related) → similarly, after feature alignment, the gradient data that can be reused across scenarios.

[0042] Step four: adaptive communication scheduling and iterative optimization 4.1 Node priority evaluation: the central server periodically calculates the generalization error of each node (using the cross-domain validation set): Node 1 (Class A unit): e1=0.08 (high priority, communicate once every 5 rounds) Node 2 (Class B unit): e2=0.15 (medium priority, communicate once every 10 rounds) Node 3 (Class C unit): e3=0.22 (medium priority, communicate once every 10 rounds) 4.2 Multi-round iterative optimization: Round 1: generate initial parameters (ε=7.5, d=8mm, topology = honeycomb), node 1 simulation shows bandwidth 13.2GHz (not up to standard), trigger communication.

[0043] Round 5: through domain adaptation adjustment, node 1 parameter optimization ε=8.2, d=7.5mm, bandwidth increased to 14.5GHz, strength 102MPa.

[0044] Round 8: global model converges, cross-domain test set average bandwidth 14.8GHz, strength 102MPa, cross-domain attenuation rate 8.7%.

[0045] Step five: verification of optimization results 5.1 Global model testing: use cross-domain test set (including aircraft, ship, vehicle stealth structure data) to evaluate the model: Electromagnetic performance: average reflectivity bandwidth 14.8GHz (target 15GHz) Mechanical performance: average bending strength 102MPa (target 100MPa) Domain adaptability: cross-domain performance attenuation rate <10% 5.2 Engineering verification: node 1 optimizes certain type of aircraft stealth structure parameters based on global model: Optimized parameters: dielectric constant ε=8.2, thickness d=7.5mm, honeycomb side length a=5mm Actual performance: bandwidth 14.9GHz, bending strength 105MPa, test meets design requirements.

[0046] In summary, in the present invention, the local model inputs structure parameters (such as ε, d), outputs performance prediction values (bandwidth, strength, etc.), and gradients; the global model inputs aligned gradients of each node, outputs updated model parameters, and then issues them to the local model for use. Through the above process, some structure parameters (such as dielectric constant reflecting electromagnetic performance, thickness d reflecting mechanical performance) reflecting the electromagnetic performance (such as bandwidth) and mechanical performance (such as bending strength) of the stealth aircraft are optimized Through the above steps, the application realizes a stealth structure multi-objective distributed optimization method based on federated learning, directly generates innovative parameter combinations from performance requirements, breaks through the efficiency and innovation bottleneck of traditional design methods, and provides a new paradigm for the rapid iteration of stealth structures.

[0047] The application is described in detail above in combination with specific embodiments and exemplary examples, but these descriptions cannot be understood as limiting the application.The person skilled in the art understands that the technical solutions and embodiments of the application can be variously replaced, modified or improved without departing from the spirit and scope of the application, and these all fall within the scope of the application.The protection scope of the application is subject to the appended claims.

[0048] The contents not described in detail in the specification of the application are the known technology of the person skilled in the art.

Claims

1. A multi-objective distributed optimization method for stealth structures based on federated learning, characterized in that, include: S1 constructs a star-shaped federated learning architecture, which includes a central server and N distributed nodes; Each node in S2 trains its model based on local data and uploads its gradients to the central server. The central server then uses domain adversarial training to align the electromagnetic and mechanical domain feature distributions of each node, thereby achieving cross-domain parameter optimization and obtaining results for each domain. The results from each domain are weighted and aggregated to obtain the global gradient. The model parameters of the central server are updated using global gradients. The central server then distributes the updated model parameters to each node, completing one iteration. Repeat step S2 until the parameter values ​​predicted by the local model at each node meet the design requirements; The S3 central server periodically obtains the generalization error of the local model of each node and determines the communication priority between each node and the central server based on the generalization error.

2. The method for multi-objective distributed optimization of cloaked structures based on federated learning according to claim 1, characterized in that, Each node trains its local model using a stealth-mechanics dual-objective loss function based on local data.

3. The method for multi-objective distributed optimization of stealth structures based on federated learning according to claim 2, characterized in that, The training process for each node includes: Each node receives the initial model parameters sent by the central server; To address the design requirements of high-speed aircraft skin, a stealth-mechanical dual-objective loss function is used in the local model to calculate multi-objective losses, and the Adam optimizer is used to calculate the gradient of model parameters during this process.

4. The method for multi-objective distributed optimization of stealth structures based on federated learning according to claim 3, characterized in that, When each node uploads the gradient of its local model parameters to the central server, it simultaneously adds Laplacian noise to the uploaded gradient using a differential privacy method.

5. The method for multi-objective distributed optimization of stealth structures based on federated learning according to claim 1, characterized in that, The source or type of local data differs for each node; Local data can be categorized into electromagnetic simulation data, mechanical property data, or salt spray corrosion test data.

6. The method for multi-objective distributed optimization of stealth structures based on federated learning according to claim 5, characterized in that, Local data includes normalized electromagnetic simulation data, mechanical property data, or salt spray corrosion test data; Among them, electromagnetic simulation data are mapped to the [0,1] interval; mechanical performance data are standardized by Z-score and topological type is encoded as a one-hot vector; salt spray corrosion test data are normalized using the mean-variance method.

7. The method for multi-objective distributed optimization of stealth structures based on federated learning according to claim 6, characterized in that, In step S2, different nodes correspond to different domains. The method of aligning the electromagnetic and mechanical domain feature distributions of each node through domain adversarial training includes: Domain adversarial neural networks are used to generate domain-invariant features and align the electromagnetic domain feature distributions of each node. The mechanical domain feature distribution of each node is aligned using a gradient inversion layer.

8. The method for multi-objective distributed optimization of cloaked structures based on federated learning according to claim 1, characterized in that, The stealth-mechanical dual-objective loss function includes reflectivity bandwidth loss, mechanical strength loss, and domain adaptation loss. The optimal weight combination of each loss is determined through hyperparameter search.

9. A multi-objective distributed optimization method for stealth structures based on federated learning according to claim 6, characterized in that, In step S3, the smaller the generalization error, the higher the communication priority.

10. A multi-objective distributed optimization method for stealth structures based on federated learning according to any one of claims 1-9, characterized in that, A stealth structure multi-objective distributed optimization system based on federated learning is adopted. The stealth structure multi-objective distributed optimization system based on federated learning includes a central server, distributed nodes, a communication module, a domain adaptation module and an optimization scheduling module. Central server: used for model initialization, cross-domain migration, and global model aggregation; Distributed nodes: used for local data training and gradient calculation; Communication module: used for the transmission of model parameters and gradients between the central server and distributed nodes, and to implement differential privacy protection; Domain Adaptation Module: Used to achieve cross-domain feature alignment between the electromagnetic and mechanical domains; Optimized scheduling module: used for multi-objective joint optimization and adaptive communication scheduling.

Images