Semantic-based aircraft multi-objective optimization design method and system

Abstract

The application relates to the field of model design, and aims to improve the calculation efficiency of multi-target balance, and provides a semantic-based aircraft multi-target optimization design method and system, wherein original design parameters and design rules are acquired first, the acquired data are converted into semantic vectors and an adjacency matrix is acquired through a semantic embedding function, a semantic gene map is generated based on the semantic vectors and the adjacency matrix, a comprehensive target function is constructed, and optimization solving is carried out based on the semantic gene map; a target mapping function is created, and the optimized result is mapped into specific engineering targets based on the target mapping function. Through the introduction of semantic modeling and cognitive enhancement mechanism, aircraft structure optimization is promoted from traditional numerical driving to semantic driving, semantic space and cognitive reasoning are used to accelerate the prediction and search process, and global optimization and traceability are realized in combination with swarm intelligence evolution and cross-domain super network; not only the efficiency and robustness are improved, but also the engineering feasibility and compliance reliability of the design result are ensured.

Description

Technical Field

[0001] This invention relates to the field of model design, specifically a semantic-based multi-objective optimization design method and system for aircraft. Background Technology

[0002] As next-generation aircraft continue to evolve towards lightweight, intelligent, and multi-mission adaptability, their structural design is increasingly exhibiting a complex situation involving high-dimensional parameter coupling, strong constraint coordination, and process feasibility. Traditional design models relying on finite element simulation and engineering experience are becoming increasingly inadequate in the face of this high complexity. A multitude of variables, such as stringer cross-sectional shape, frame thickness and height, overall and local skin thickness, material configuration, and connection node style, constitute a vast design space. These parameters not only affect load-bearing capacity, weight distribution, and lifespan reliability, but also are closely related to manufacturability, safety margins, and airworthiness regulations, leading to a highly nonlinear design process accompanied by significant uncertainties.

[0003] In the parametric design workflow of mainstream 3D modeling software such as CATIA, existing parameter management and specification verification typically rely on the following methods:

[0004] (1) Multidisciplinary Design Optimization (MDO) Framework

[0005] Coupled solutions to aerodynamic, structural, and process constraints on a unified platform can improve optimization efficiency to some extent. However, this method heavily relies on high-fidelity finite element models and coupled simulations, resulting in enormous computational costs that make it difficult to support rapid iteration and large-scale engineering applications.

[0006] (2) Combining parametric modeling with heuristic algorithms

[0007] Typical examples include local search methods such as genetic algorithms and particle swarm optimization, which can find relatively optimal solutions within a small parameter space. However, these tools mostly rely on hard-coded logic with fixed rules, have limited semantic understanding of parameters, and are difficult to adapt to complex engineering constraints.

[0008] (3) Database-based experience-driven tools

[0009] Some companies establish knowledge bases or design manuals to retrieve and reuse common parameters in order to improve design consistency. However, this method still relies on human experience and preset rules, lacks semantic understanding and intelligent evolution capabilities, and is difficult to cope with the complex challenges of new tasks or new configurations. Summary of the Invention

[0010] To improve the computational efficiency of multi-objective balancing, this application provides a semantic-based multi-objective optimization design method and system for aircraft.

[0011] The technical solution adopted by the present invention to solve the above problems is:

[0012] Semantic-based multi-objective optimization design methods for aircraft include:

[0013] Step 1: Obtain the original design parameters and design rules;

[0014] Step 2: Convert the data obtained in Step 1 into semantic vectors and obtain the adjacency matrix using a semantic embedding function. Generate a semantic gene map based on the semantic vectors and the adjacency matrix.

[0015] Step 3: Construct a comprehensive objective function and optimize it based on the semantic gene map; create an objective mapping function and map the optimized result to a specific engineering objective based on the objective mapping function.

[0016] Furthermore, step 2 specifically involves:

[0017] According to the division of labor requirements in the engineering design, the data obtained in step 1 is divided into different subdomains;

[0018] Semantic vectors and adjacency matrices are obtained from each subdomain to generate a subdomain graph;

[0019] The subdomain graphs are merged based on the cross-domain fusion function to generate the final semantic gene map;

[0020] The calculation process of the cross-domain fusion function is as follows: merge the nodes in each subdomain graph; generate cross-domain edges based on semantic similarity and calculate the weight of each cross-domain edge.

[0021] Furthermore, the subdomains include aerodynamics, structure, and process.

[0022] Furthermore, step 3 also includes: creating and training a performance prediction model, which is used to predict the performance corresponding to the current parameters; during the optimization process, the performance corresponding to the current parameters is obtained by calling the performance prediction model, and the performance includes weight, strength and process complexity.

[0023] Furthermore, during the optimization process, the candidate solutions with the lowest prediction confidence are subjected to high-fidelity verification, and the performance prediction model is adjusted based on the verification results.

[0024] Furthermore, the comprehensive objective function is expressed as:

[0025] , where x is the design parameter. For adaptive weights, For single-objective predicted values, As a penalty factor for cross-domain constraints, The term is a penalty for violating constraints. When the constraint is satisfied, the term is 0; when it is violated, the term is positive.

[0026] Furthermore, step 3 employs a population evolution algorithm for optimization.

[0027] Furthermore, step 3 also includes identifying and recording each optimization process.

[0028] A semantic-based multi-objective optimization design system for aircraft is used to implement a semantic-based multi-objective optimization design method for aircraft, including:

[0029] The data acquisition module is used to acquire the original design parameters and design rules;

[0030] The semantic gene map module converts the data obtained in step 1 into semantic vectors and obtains the adjacency matrix through a semantic embedding function, and generates a semantic gene map based on the semantic vectors and the adjacency matrix;

[0031] The computation module constructs a comprehensive objective function and optimizes it based on a semantic gene map.

[0032] The mapping module creates target mapping functions and maps the optimized results to specific engineering goals based on these functions.

[0033] Furthermore, the calculation module also includes: creating and training a performance prediction model, which is used to predict the performance corresponding to the current parameters; during the optimization process, the performance corresponding to the current parameters is obtained by calling the performance prediction model, and the performance includes weight, strength and process complexity.

[0034] The advantages of this invention compared to the prior art are:

[0035] By converting the original design parameters and rules into semantic vectors, a multi-dimensional structural gene map spanning geometry, materials, and processes is constructed, explicitly revealing the deep semantic mapping relationships between parameters. This mechanism can significantly reduce the complexity of data preparation and parameter alignment, providing a unified and reasonable semantic foundation for subsequent prediction and optimization. It avoids the drawbacks of information fragmentation and knowledge gaps in traditional methods and improves the efficiency of optimization computation.

[0036] By constructing predictive models, performance indicators such as weight, strength, and stability can be quickly extrapolated, and self-learning calibration can be achieved in the iteration process. While maintaining prediction accuracy, the number of high-fidelity calls and physical tests can be reduced, significantly shortening the verification cycle and improving design efficiency and prediction reliability.

[0037] The optimization algorithm employs a population evolutionary algorithm, introducing swarm intelligence and a gene drift mechanism into the global solution space. Through the interaction, recombination, and mutation of solutions, it maintains population diversity, thereby achieving continuous evolution and global exploration. This mechanism not only improves global search efficiency and convergence stability but also effectively avoids local optimum traps, making the optimization results more comprehensive and robust.

[0038] By constructing a cross-domain semantic supernet among aerodynamics, structure, and process, heterogeneous models can achieve information interaction and dynamic weight adjustment at the semantic level, thereby driving a symbiotic and collaborative relationship between disciplines. This mechanism can generate a globally optimal solution that balances lightweight design, structural strength, and process feasibility, effectively avoiding the risks of goal fragmentation and rework.

[0039] Through a closed-loop mechanism of prediction-optimization-verification-relearning, the method dynamically balances strength, weight, and process in a virtual space, gradually converging to a verifiable optimal solution while incorporating task constraints. This approach ensures that the optimization results not only meet theoretical optimality but also align with task-oriented requirements, resulting in higher engineering feasibility and usability.

[0040] By recording the design evolution path and generating a unique identifier for each modification, the mechanism enables parameter evolution tracking and consistency comparison across each stage. This forms an end-to-end transparent chain of evidence, enhancing the credibility of the design process and providing reliable support for quality control and compliance review. Attached Figure Description

[0041] Figure 1 Flowchart of multi-objective optimization design method for aircraft. Detailed Implementation

[0042] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0043] Existing design processes heavily rely on high-fidelity modeling and physical experimental verification, resulting in lengthy design cycles, high costs, and difficulty in adapting to rapid iteration requirements. Modeling often remains at a single-disciplinary or numerical level, failing to reveal the deep semantic relationships between geometric, material, and process parameters, leading to fragmented optimization. Furthermore, during optimization, most algorithms are limited to searching local parameter spaces, lacking global consistency and evolutionary capabilities, and easily getting trapped in local optima. Moreover, balancing safety margins with lightweight design is difficult, resulting in either redundant and wasteful materials or complex and impractical designs. Therefore, this invention proposes a semantic-based multi-objective optimization design method for aircraft, such as... Figure 1 As shown, it includes:

[0044] Step 1: Obtain the original design parameters and design rules.

[0045] Taking aircraft design as an example, design parameters include: shape, thickness, height and external parameters, such as aerodynamic parameters: airfoil parameters, overall wing surface parameters, three-dimensional external parameters, etc.; design rules include: distribution pattern, material configuration, connection method and airworthiness standards, etc.

[0046] Step 2: Convert the data obtained in Step 1 into semantic vectors and obtain the adjacency matrix through the semantic embedding function. Generate a semantic gene map based on the semantic vectors and the adjacency matrix.

[0047] The original design parameters obtained in step 1 are used through a semantic embedding function. Transform into semantic vectors , , These are the original design parameters. The semantic embedding function can take the following forms depending on the data type of the parameters:

[0048] 1) Normalization + polynomial / Fourier basis expansion, applicable to continuous parameters such as thickness, height, and aspect ratio;

[0049] like , The original value of the i-th design parameter. This is the mean of the parameter in the training samples or prior data. Let be the standard deviation of the parameter, and T be the period of the Fourier expansion.

[0050] 2) Category embedding, applicable to discrete parameters such as material configuration and connection method;

[0051] like E is the embedding matrix, and R represents the d-dimensional real space in which the semantic vector resides.

[0052] 3) Function unrolling + neural encoder, applicable to functional parameters such as skin thickness distribution and aerodynamic airfoil shape;

[0053] like , A set of coefficient vectors obtained by expanding the function is used to approximate the function.

[0054] All semantic vectors form a set S, which constitutes the nodes of the graph. Combining these semantic vectors with the adjacency matrix A generates the semantic gene graph G=(S,A). The adjacency matrix A expresses the relationships between semantic nodes. and If there is a semantic relationship, then ,otherwise, .

[0055] According to the engineering design division of labor requirements, the data obtained in step 1 can be divided into three subdomains: aerodynamic, structural, and process. Aerodynamic data mainly includes information related to aerodynamic shape, lift, drag, and aerodynamic loads; structural data mainly includes information related to the girder, frame, skin thickness, and material strength; and process data mainly includes information related to assembly methods, processing feasibility, manufacturing complexity, and airworthiness standards. Different subdomains correspond to different single-objective functions: aerodynamics corresponds to weight, structure to strength, and process to process complexity. This division makes the optimization problem more hierarchical. Since semantic relationships cannot be directly obtained between these three types of data, local modeling is performed first, followed by global coupling for cross-domain fusion.

[0056] This embodiment uses a cross-domain fusion function. By merging the aerodynamic, structural, and process subgraphs into a semantic supernet Z, cross-domain edges enable different disciplinary objectives (weight, strength, and process) to interact and optimize on a single semantic supernet. ,in, This is a cross-domain fusion function; : Aerodynamic subgraph, mainly containing semantic vectors and their relationships related to aerodynamic shape, lift, drag, and aerodynamic load; : Structural sub-graph, mainly containing the semantic relationship between structural parameters and constraints such as stringers, frames, skin thickness, and material strength; The process sub-diagram mainly includes process-related semantic relationships such as assembly method, processing feasibility, manufacturing complexity, and airworthiness standards.

[0057] The operation logic is as follows:

[0058] 1) Node-level operations: Alignment / Merging , , , These are the node sets of the aerodynamic subgraph, structural subgraph, and process subgraph, respectively. This step merges the node sets of the three subgraphs into the same semantic space.

[0059] 2) Edge-level operations: Cross-domain joins Let Aa, As, and Am be the sets of edges in the aerodynamic subgraph, structural subgraph, and process subgraph, respectively, and let Across be the set of cross-domain edges. Cross-domain edges are new connections added by calculating semantic similarity or based on knowledge rules. Taking similarity as an example, we first calculate the similarity between nodes: , For similarity function, , Semantic vectors from any two different subdomains of aerodynamics, structure, and process; if the similarity score Then a new cross-domain edge will be added. The preset values ​​are used; then the normalized weights of each cross-domain edge are calculated: k represents the node Indexes of all connected candidate neighbor nodes.

[0060] Semantic gene mapping is not only a data representation, but also a unified mapping of the implicit coupling relationships between geometry, materials and processes into a knowledge network. This allows subsequent optimization to be carried out on a unified semantic basis, avoiding the problem of "parameter fragmentation" in traditional methods.

[0061] Step 3: Construct a comprehensive objective function and optimize it based on the semantic gene map; create an objective mapping function and map the optimized result to a specific engineering objective based on the objective mapping function.

[0062] In aircraft design, weight, strength, and manufacturing process objectives often conflict with each other. Therefore, this embodiment achieves a balance by constructing a comprehensive objective function. The comprehensive objective function is expressed as: In the formula, x is the design parameter. To achieve adaptive weighting, the weights are dynamically adjusted based on the importance of different subdomains in the optimization process; For single-domain target prediction values, This is a penalty factor for cross-domain constraints. When certain constraints are violated, the penalty intensity will be dynamically adjusted according to the cross-domain relationship. The penalty for violating the constraint is that the term is 0 when the constraint is satisfied, and positive when the constraint is violated, triggering the penalty.

[0063] The construction method of the comprehensive objective function directly injects the dynamic relationships of the cross-domain semantic hypernet into the weights and penalty factors, enabling the objective function to adaptively adjust with the design state, thereby achieving multi-disciplinary collaborative optimization. Unlike the fixed-weight coupling method of traditional MDO, this invention achieves dynamic weight adjustment through the semantic hypernet, driving the symbiotic evolution between different disciplines and generating a globally optimal solution that balances lightweight, strength, and manufacturability.

[0064] Aircraft structure optimization often requires extensive high-fidelity simulations or experiments, resulting in extremely high computational costs. It's impractical to call these simulations for every optimization iteration. Therefore, this embodiment creates a performance prediction model to reduce computational costs during optimization iterations, enabling faster searches while maintaining sufficient accuracy. The semantic vector is input into the performance prediction model to obtain the prediction results. By iteratively training to minimize the error L between the predicted and true values, we achieve learning and calibration simultaneously. The prediction model is represented as: , θ: Predictive performance output, representing the structural performance inferred by the model under given parameters, including weight estimation, strength prediction, and manufacturing complexity. These metrics are strongly coupled (lightweighting often weakens strength, while increasing strength may increase manufacturing difficulty). Using a unified prediction model, shared feature representations can be utilized to constrain each other during training, reducing overfitting. θ: The set of model parameters, including weights, biases, etc., is continuously updated during training to minimize prediction error. y: True performance value, derived from high-fidelity simulation, experimental data, or engineering measurements; L: Loss function, in this embodiment, the squared error norm is used. .

[0065] During the optimization iteration from k to k+1:

[0066] 1) Based on the current parameters Semantic embedding is obtained ;

[0067] 2) Model prediction performance ;

[0068] 3) Perform high-fidelity verification on some candidate solutions with the lowest prediction confidence, that is, run finite element experiments on the same solution to obtain the true value. ;

[0069] 4) Update the loss function: ;

[0070] 5) Adjust θ.

[0071] In this embodiment, θ is corrected with a small amount of new data in each optimization iteration, so there is no need to collect a large amount of high-fidelity data from scratch, thereby reducing the computational cost.

[0072] The prediction results are merely "raw estimates" from the model output; they need to be transformed into an engineering-comparable and optimizable objective mapping function form. This is achieved using the objective mapping function. The prediction results are mapped to specific indicators such as weight, strength, or process.

[0073] ;

[0074] Extract the weight index from the prediction results. W: is the baseline design weight or maximum permissible weight;

[0075] Extract the intensity / stability index from the prediction results, corresponding to f2(x). σ: Predicted maximum stress Allowable stress of the material Predicted buckling load, Critical buckling load.

[0076] Extract the process complexity index from the prediction results. C a Assembly complexity (e.g., number of assembly steps, process difficulty); C m : Processing complexity (e.g., material machinability, precision requirements); C c Complexity of airworthiness review; Weighting coefficients (which can be determined by expert experience or data training).

[0077] : The optimization objective function driven by the prediction model represents the estimated value of the i-th type of performance index (where i=1,2,3, corresponding to weight, strength / stability and process complexity, respectively) given the design parameters x and the model state θ.

[0078] In this embodiment, a population evolution algorithm is used for optimization, allowing candidate solutions to interact, recombine, and undergo gene drift, maintaining solution diversity and preventing getting trapped in local optima. Compared to traditional genetic algorithms, it can utilize semantic information, making the search process faster and more reliable. The specific process is as follows:

[0079] 1) Initialize the semantic population

[0080] Based on semantic vector set Several feasible designs were randomly selected as the initial population. , where each individual s j This corresponds to a potential design solution.

[0081] 2) Fitness evaluation

[0082] The performance estimates of individuals are quickly calculated using a prediction model and a target mapping function, and the fitness of the population is scored by combining the comprehensive objective function J(x;θ,Z).

[0083] 3) Interaction of solutions

[0084] Under the constraints of the semantic adjacency matrix A, for two parent semantic vectors , Interact with: This ensures the semantic rationality of the combined solution while inheriting the advantageous features of the parent generation.

[0085] 4) Recombination of solutions

[0086] By recombining semantic fragments from different individuals across subdomains, such as inheriting shape parameters from aerodynamic subgraphs, material strength from structural subgraphs, and assembly methods from process subgraphs, new candidate solutions can be obtained. Where Ma, Ms, and Mm are subfield selection masks. : Semantic vector of the aerodynamic subdomain; : Semantic vector of a structural subdomain; : Semantic vector of the process subdomain Multiply corresponding elements of two matrices (or vectors) of the same shape.

[0087] 5) Gene drift

[0088] Introduce graph-based perturbations to the semantic representation of individuals: s: The semantic vector of the current candidate solution. New candidate solutions, The coefficients controlling the perturbation amplitude (step size or learning rate); Δ is the perturbation vector, which follows a Gaussian distribution; the covariance matrix Σ(A) comes from the semantic adjacency matrix and represents the correlation between parameters.

[0089] 6) Maintain population diversity and iterative updates

[0090] Through the combined effects of interaction, recombination, and drift, the population maintains its diversity. The evolutionary iteration process is as follows: ,in This represents a semantic evolution operation. As the number of iterations increases, the solution gradually approaches the global optimum.

[0091] Furthermore, the optimized solution must meet the requirements of the specific task, such as lifetime and safety margin. This step involves verification and selection through convergence conditions and task-related performance constraints to ensure that the solution is not only theoretically optimal but also usable in practical tasks. The optimized solution must satisfy the task constraints: ,in, This represents task-related performance (such as lifetime or safety margin), determined by the prediction model. The results obtained from the target mapping function Γ(·) are the task-related parts, such as fatigue life, structural safety margin, and task reliability index, where δ is the allowable threshold.

[0092] The convergence condition can be defined as: or ;

[0093] The solution in the k-th iteration comes from a population evolution algorithm, representing the design solution of each generation during the evolution process, gradually approaching the optimal solution. η is a preset value; in this way, the system not only converges to the theoretical optimal solution, but also ensures that the solution meets the task requirements, thus improving the actual usability of the results.

[0094] This invention elevates aircraft structure optimization from traditional numerical-driven to semantic-driven by introducing semantic modeling and cognitive enhancement mechanisms. It accelerates the prediction and search process by utilizing semantic space and cognitive reasoning, and achieves global optimality and traceability by combining swarm intelligence evolution and cross-domain hypernet. This paradigm shift not only significantly improves efficiency and robustness, but also ensures the engineering feasibility and compliance credibility of the design results.

[0095] To facilitate traceability, the entire evolution path of the optimization process is fully recorded, with each parameter update having a unique identifier that allows for tracing back to its source, forming an end-to-end transparent evolution chain. For airworthiness reviews and lifecycle quality management, this mechanism provides a complete and reliable chain of evidence.

[0096] Correspondingly, the present invention also provides a semantic-based multi-objective optimization design system for aircraft, used to implement a semantic-based multi-objective optimization design method for aircraft, including:

[0097] The data acquisition module is used to acquire the original design parameters and design rules;

[0098] The semantic gene map module converts the data obtained in step 1 into semantic vectors and obtains the adjacency matrix through a semantic embedding function, and generates a semantic gene map based on the semantic vectors and the adjacency matrix;

[0099] The computation module constructs a comprehensive objective function and optimizes it based on a semantic gene map.

[0100] The mapping module creates target mapping functions and maps the optimized results to specific engineering goals based on these functions.

[0101] Furthermore, the calculation module also includes: creating and training a performance prediction model, which is used to predict the performance corresponding to the current parameters; during the optimization process, the performance corresponding to the current parameters is obtained by calling the performance prediction model, and the performance includes weight, strength and process complexity.

Claims

1. A semantic-based multi-objective optimization design method for aircraft, characterized in that, include: Step 1: Obtain the original design parameters and design rules; Step 2: Convert the data obtained in Step 1 into semantic vectors and obtain the adjacency matrix using a semantic embedding function. Generate a semantic gene map based on the semantic vectors and the adjacency matrix. Step 3: Construct a comprehensive objective function and optimize it based on the semantic gene map; create an objective mapping function and map the optimized result to a specific engineering objective based on the objective mapping function; Step 2 is as follows: According to the division of labor requirements in the engineering design, the data obtained in step 1 is divided into different subdomains; Semantic vectors and adjacency matrices are obtained from each subdomain to generate a subdomain graph; The subdomain graphs are merged based on the cross-domain fusion function to generate the final semantic gene map; The calculation process of the cross-domain fusion function is as follows: merge the nodes in each subdomain graph; generate cross-domain edges based on semantic similarity and calculate the weight of each cross-domain edge.

2. The semantic-based multi-objective optimization design method for aircraft according to claim 1, characterized in that, The subdomains include aerodynamics, structure, and process.

3. The semantic-based multi-objective optimization design method for aircraft according to claim 2, characterized in that, Step 3 also includes: creating and training a performance prediction model, which is used to predict the performance corresponding to the current parameters; during the optimization process, the performance corresponding to the current parameters is obtained by calling the performance prediction model, and the performance includes weight, strength and process complexity.

4. The semantic-based multi-objective optimization design method for aircraft according to claim 3, characterized in that, During the optimization process, the candidate solutions with the lowest prediction confidence are subjected to high-fidelity verification, and the performance prediction model is adjusted based on the verification results.

5. The semantic-based multi-objective optimization design method for aircraft according to claim 3, characterized in that, The overall objective function is expressed as: x is the design parameter. For adaptive weights, For single-objective predicted values, As a penalty factor for cross-domain constraints, The term is a penalty for violating constraints. When the constraint is satisfied, the term is 0; when it is violated, the term is positive.

6. The semantic-based multi-objective optimization design method for aircraft according to claim 1, characterized in that, Step 3 uses a population evolution algorithm for optimization.

7. The semantic-based multi-objective optimization design method for aircraft according to claim 1, characterized in that, Step 3 also includes identifying and recording each optimization process.

8. A semantic-based multi-objective optimization design system for aircraft, used to implement the semantic-based multi-objective optimization design method for aircraft as described in any one of claims 1-7, characterized in that, include: The data acquisition module is used to acquire the original design parameters and design rules; The semantic gene map module converts the data obtained in step 1 into semantic vectors and obtains the adjacency matrix through a semantic embedding function, and generates a semantic gene map based on the semantic vectors and the adjacency matrix; The computation module constructs a comprehensive objective function and optimizes it based on a semantic gene map. The mapping module creates target mapping functions and maps the optimized results to specific engineering goals based on these functions.

9. The semantic-based multi-objective optimization design system for aircraft according to claim 8, characterized in that, The calculation module also includes: creating and training a performance prediction model, which is used to predict the performance corresponding to the current parameters; during the optimization process, the performance corresponding to the current parameters is obtained by calling the performance prediction model, and the performance includes weight, strength and process complexity.

Images