New energy pickup truck frame MOP method, device, medium and program product based on transformer

Abstract

The application discloses a new energy pickup truck frame MOP method, equipment, medium and program product based on a Transformer, and comprises the following steps: (1) obtaining the mass, torsional stiffness and first-order natural frequency of the new energy pickup truck frame through parameterized modeling, torsional static analysis and free modal analysis, and establishing a MOP model; (2) generating an initial population based on Latin hypercube sampling, generating a reference vector by using a simplex method, and decomposing a multi-objective problem into multiple sub-problems; (3) constructing a state matrix sequence according to the evolution state of each sub-problem; (4) inputting the state matrix sequence into a Transformer model, adaptively determining a mutation strategy and mutation parameters for each sub-problem by using the Transformer model, and generating a candidate child individual vector; (5) performing simulation evaluation on the candidate child individual vector after screening, updating a database and a population, returning to step (3), and obtaining a Pareto optimization structure meeting design requirements until the Pareto optimization structure meets the design requirements. The application can improve the multi-objective optimization efficiency of the new energy pickup truck frame.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and swarm intelligence, and more specifically, to a method, device, medium, and program product based on Transformer-based MOP (Multi-objective Optimization) for new energy pickup truck frames. Background Technology

[0002] Current vehicle frame structure optimization design typically relies on finite element simulation, engineering experience, and parametric calculations. Designers establish finite element models based on the geometric parameters of the frame's longitudinal beams, crossbeams, and connecting areas. They evaluate structural performance through static analysis, torsional static analysis, and modal analysis, and repeatedly adjust structural parameters based on simulation results. However, for frame optimization problems involving multiple design variables, multiple performance objectives, and computationally expensive finite element simulations, the above methods usually require extensive parametric calculations and finite element simulations, resulting in low optimization efficiency. Especially in the design of new energy pickup truck frames, there are coupling relationships among structural design variables such as longitudinal beam thickness, longitudinal beam section height, longitudinal beam section width, crossbeam thickness, crossbeam section height, crossbeam section width, number of crossbeams, and number of longitudinal beams. The influence of each structural design variable on mass, torsional stiffness, and first-order natural frequency varies, making it difficult to efficiently obtain a Pareto-optimal structure that balances lightweighting, torsional stiffness, and modal performance solely through parametric calculations.

[0003] To improve the optimization efficiency of complex structures, existing technologies have introduced methods such as response surface models, radial basis function models, genetic algorithms, and differential evolution algorithms to replace some high-cost finite element calculations or assist in searching the design space. These methods can reduce the number of precise simulations to some extent. However, traditional machine learning-assisted evolutionary optimization methods typically employ pre-defined mutation strategies and evolutionary parameters. Mutation strategies, scaling factors, and crossover probabilities largely depend on human experience, making it difficult to adaptively adjust search behavior based on the search status of different subproblems at different iteration stages. When a multi-objective optimization problem is decomposed into multiple subproblems using reference vectors, different subproblems may have different search directions, convergence degrees, neighborhood distribution states, and stagnation degrees. Some subproblems require enhanced global exploration capabilities, some require stronger local development capabilities, and some need to escape local optima while maintaining a convergence trend. If all subproblems use the same mutation strategy and evolutionary parameters, insufficient searching of some target regions, repeated searches, or search stagnation can easily occur, which is detrimental to the convergence and distribution of the Pareto solution set. Summary of the Invention

[0004] This invention proposes a Transformer-based method, device, medium, and program product for multi-objective optimization of new energy pickup truck frames. It aims to solve the problems of high cost of finite element simulation evaluation, difficulty in uniformly representing the different search states of subproblems corresponding to different reference vectors, difficulty in adapting fixed mutation strategies and evolution parameters to different subproblems and different search stages, and difficulty in balancing Pareto front convergence and distribution in candidate individual selection.

[0005] To achieve the above objectives, according to one aspect of the present invention, a Transformer-based method for the MOP (Model-Operated Platform) of a new energy pickup truck frame is provided, the method comprising: Step (1): Based on the structural design variables of the new energy pickup truck frame, establish a parametric three-dimensional model and a finite element model, and construct an MOP model with mass, torsional stiffness and first natural frequency as the objectives; Step (2): The initial population is generated by Latin hypercube sampling, the reference vector is generated by simplex method, and the MOP model is decomposed into multiple sub-problems. The initial population is evaluated by simulation, and a database is established and a radial basis function model is constructed. Step (3): Determine the ideal reference point and the nadir reference point, normalize the target value, associate the individual vectors in the population with the reference vector, determine the representative individual vectors of each subproblem and construct the state matrix sequence; Step (4): Use the Transformer model to process the state matrix sequence, determine the mutation strategy, scaling factor and crossover probability for each subproblem, and generate candidate offspring individual vectors; Step (5): Predict the target value of candidate offspring individual vectors using the radial basis function model, filter and deduplicate the individual vectors to be evaluated based on the positive improvement distance and negative deviation distance, perform simulation evaluation, update the database, population and radial basis function model, and return to step (3) until the Pareto optimization structure that meets the design requirements is output.

[0006] Furthermore, step (1) specifically includes: Step 1.1: Based on the structural layout, manufacturing process, component installation space, and adjustable range of structural parameters of the new energy pickup truck frame, determine the range of values ​​for longitudinal beam plate thickness, longitudinal beam section height, longitudinal beam section width, crossbeam plate thickness, crossbeam section height, crossbeam section width, number of crossbeams, and number of longitudinal beams, and construct a structural design variable vector and a multi-dimensional design space. Step 1.2: Based on the non-load-bearing chassis structure, material properties, load-bearing component layout and typical driving conditions of the new energy pickup truck frame, establish a parametric 3D model of the new energy pickup truck frame in CATIA, and obtain the frame mass based on the material density and the geometric volume of the parametric 3D model. Step 1.3: Import the parametric 3D model into HyperMesh for geometry cleanup, mid-surface extraction, shell element mesh generation, material property assignment, plate thickness property definition, connection relationship definition, load application, and boundary condition setting to obtain the finite element model of the new energy pickup truck frame. Step 1.4: Use OptiStruct to perform torsional static analysis on the finite element model, extract the vertical displacement of the left and right loading points, calculate the torsional load moment based on the single-sided vertical loading force and the distance between the left and right loading points, calculate the frame torsion angle from the displacement difference between the left and right loading points, and finally calculate the torsional stiffness. Step 1.5: Use OptiStruct to perform free modal analysis on the finite element model, solve the modal characteristic equations corresponding to the finite element stiffness matrix and mass matrix, obtain the circular frequencies and mode shapes of each order, and thus obtain the first natural frequency. Step 1.6: Based on the design variable vector of the new energy pickup truck frame, the frame mass, torsional stiffness, and first-order natural frequency, construct the MOP mathematical model, the specific expression of which is as follows: , In the above formula, X This represents the vector of design variables for the frame structure of new energy pickup trucks. St represents the multidimensional design space comprised of the value ranges of each structural design variable, and St represents the vector of structural design variables for the new energy pickup truck frame. X The multidimensional design space constrained by the range of values ​​of each structural design variable middle, Indicates the thickness of the longitudinal beam plate. Indicates the height of the longitudinal beam section. Indicates the width of the longitudinal beam section. Indicates the thickness of the crossbeam plate. Indicates the height of the beam section. Indicates the width of the beam cross section. Indicates the number of beams. Indicates the number of longitudinal beams. This indicates the density of the frame material for new energy pickup trucks. Represents a vector of structural design variables X The corresponding parametric model geometric volume, Represents a vector of structural design variables X The corresponding weight of the new energy pickup truck frame, This represents the unilateral vertical loading force applied under torsional static conditions. B Indicates the spacing between the left and right loading points. Represents a vector of structural design variables X The corresponding torsional load moment, This represents the vertical displacement of the loading point on the left side in the torsional static analysis. This represents the vertical displacement of the loading point on the right side in the torsional static analysis. Represents a vector of structural design variables X The corresponding frame torsion angle, Represents a vector of structural design variables X The corresponding torsional stiffness of the new energy pickup truck frame, Represents a vector of structural design variables X The corresponding finite element stiffness matrix, Represents a vector of structural design variables X The corresponding finite element mass matrix, Indicates the first q Mode shape, Indicates the first q Circular frequency, Indicates the first-order angular frequency. Represents a vector of structural design variables X The corresponding first-order natural frequency of the new energy pickup truck frame, This represents the objective function vector for the multi-objective optimization of the new energy pickup truck frame. This means transforming the objective of maximizing torsional stiffness into the objective of minimizing it. This means transforming the objective of maximizing the first-order natural frequency into a minimization objective; Find indicates that in the multidimensional design space... Finding structural design variable vectors X The optimal solution, min represents the optimization of the objective function vector for multiple objectives. Perform a minimization optimization.

[0007] Furthermore, step (2) specifically includes: Step 2.1: Determine the initial number of individual vectors based on the number of structural design variables, the number of optimization objectives, and the preset simulation evaluation budget. N p Latin hypercube sampling is used to generate [the data] in a multidimensional design space. N p The initial individual vectors form the initial population; Step 2.2: Based on the number of optimization objectives, generate the target space consisting of three objectives using the simplex method. N v Reference vectors, and using N v The multi-objective optimization problem of the new energy pickup truck frame is decomposed into several reference vectors. N v The issue of size; Step 2.3: Perform mass calculation, torsional static analysis and free modal analysis on each initial individual vector in the initial population to obtain the corresponding target values ​​of mass, torsional stiffness and first natural frequency, and store the initial individual vector and its target values ​​in the database; Step 2.4: Construct a radial basis function model using all individual vector information in the database.

[0008] Furthermore, step (3) specifically includes: Step 3.1: Determine the ideal reference point based on the database and the target value in the current population. and Nadir Reference Point and using ideal reference points and Nadir Reference Point The target values ​​corresponding to mass, torsional stiffness, and first natural frequency are normalized to obtain the normalized target values ​​of each body vector in the current population. The ideal reference point and the nadir reference point are determined by the following formula: , , In the above formula, The ideal reference point is at the th j The values ​​of each target, This indicates that the nadir reference point is at the 1st ytterbium position. j The values ​​of each target, DB For database, x represents i For individual vectors in the database, Represents the first in the database i Individual vector x i The j One target value, This represents minimizing the individual vector x in the database. i The j One target value, This means maximizing the individual vector x in the database. i The j One target value, m It is the target number; Step 3.2: Calculate the perpendicular distance between the normalized target value of each body vector in the current population and each reference vector. Establish an association relationship between each body vector and the reference vector with the smallest perpendicular distance, and determine the associated body set for each reference vector corresponding to the subproblem based on the association relationship. For subproblems with associated body sets, calculate the Chebyshev aggregation value of each body vector in the associated body set under the subproblem, and select the body vector with the smallest Chebyshev aggregation value as the representative body for the subproblem. For subproblems without associated body sets, first select the body vector with the smallest Chebyshev aggregation value under the current subproblem from the representative bodies of the neighboring subproblems as the representative body. If there is no representative body in the neighboring subproblems, then select the body vector with the smallest Chebyshev aggregation value from the current population as the representative body. The Chebyshev aggregation value is calculated by the following formula: , In the above formula, It is a Chebyshev aggregation function constructed based on an ideal reference point. For the first i Individual vectors, Is the reference vector at the th j The weight of each objective; Step 3.3: For each subproblem, a state matrix is ​​formed sequentially around the reference vector and representative individuals. The data of different dimensions in the state matrix are then normalized. The state matrix consists of the reference vector direction, the target state of the representative individual, the neighborhood search state, the variable perturbation response, and the historical improvement state. The reference vector direction is represented by the components of the subproblem's reference vector in the three target directions: mass, torsional stiffness, and first-order natural frequency. The target state of the representative individual is represented by the normalized target value of the representative individual, the perpendicular distance between the representative individual and the reference vector, and the Chebyshev aggregation value of the representative individual under the current subproblem. The neighborhood search state is represented by the distribution of Chebyshev aggregation values ​​of representative individuals in the neighborhood subproblems, the number of individuals associated with the current subproblem, and the dispersion of the neighborhood target space. The variable perturbation response is represented by the changes in the target value and Chebyshev aggregation value of the representative individual's structural design variables before and after local perturbations. The historical improvement state is represented by the historical success rate, number of stalls, and the most recent improvement amount of the subproblem in previous iterations. Step 3.4: Arrange the state matrices of each subproblem in the order of the reference vector numbers to form a state matrix sequence.

[0009] Furthermore, step (4) specifically includes: Step 4.1: Normalize the state variables in the state matrix sequence, and then classify them according to their feature sources. j The state characteristics of each sub-problem are divided into a reference vector feature group, a target search feature group, a neighborhood distribution feature group, a variable response feature group, and a historical evolution feature group. j State characteristics of each sub-problem Represented as: , In the above formula, Indicates the first j The reference vector feature set of each sub-problem Indicates the first j The target search feature set for each sub-problem Indicates the first j The neighborhood distribution characteristics of each sub-problem Indicates the first j Variable response feature set of sub-problems Indicates the first jThe historical evolutionary characteristics of individual sub-problems; Step 4.2: Using the feature mapping module, feature mapping is performed on the reference vector feature group, target search feature group, neighborhood distribution feature group, variable response feature group, and historical evolution feature group, respectively. The mapped features are then fused and superimposed with positional codes representing the reference vector numbering order to obtain the [number of features]. j Initial feature vectors for each subproblem: , In the above formula, Indicates the first j The sub-problems are input into the initial feature vector before the Transformer encoder. Indicates the feature group identifier. w , o , n , v , h These correspond to the identifiers of the reference vector feature group, target search feature group, neighborhood distribution feature group, variable response feature group, and historical evolution feature group, respectively. Indicates the first j The first sub-problem After normalization, the state feature vector of the class feature group is input to the feature mapping layer. Indicates the first j The first sub-problem The feature vector obtained after feature mapping of the class feature group Indicates the first j The feature vectors obtained after feature mapping of the reference vector feature set of each subproblem Indicates the first j The feature vector obtained after feature mapping of the target search feature set of each sub-problem Indicates the first j The feature vector is obtained by feature mapping of the neighborhood distribution feature set of each subproblem. Indicates the first j The feature vector obtained after feature mapping of the variable response feature set of each subproblem Indicates the first j The feature vector is obtained by feature mapping of the historical evolution feature set of each sub-problem. and They represent the first The mapping weight matrix and bias terms corresponding to the feature groups of each class Represents a non-linear activation function. and These represent the weight matrix and bias term of the fusion mapping layer, respectively. Indicates the first j The positional encoding corresponding to each sub-problem; Step 4.3: The initial feature vectors of each sub-problem are arranged into an input feature sequence according to the reference vector numbering order, and then input into the Transformer encoder. The contextual relationships between different sub-problems are extracted through a multi-head self-attention mechanism and a feedforward network. The first step of the Transformer encoder... l The layer updates its features according to the following formula: , In the above formula, l Indicates the layer number of the Transformer encoder. Indicates the first l The input feature sequence of the layer encoder, i.e., the first layer... The context feature sequence output by the layer encoder, Representation layer normalization, Indicates to The feature sequence obtained after layer normalization r This indicates the attention head number in the multi-head self-attention module, and , R Indicates the number of attention heads. , , They represent the first l The first layer encoder r The query mapping weight matrix, key mapping weight matrix, and value mapping weight matrix corresponding to each attention head. Indicates the first l The first layer encoder r Attention features of the output of each attention head. Represents the normalized exponential function, Indicates the first l The attention features output by the first attention head in the layer encoder. Indicates the first l The first layer encoder R Attention features output by each attention head, superscript Indicates matrix transpose. d h The feature dimension of a single attention head. , and They represent the first l The first layer encoder r The query matrix, key matrix, and value matrix corresponding to each attention head. This indicates that the output features of each attention head are concatenated. and They represent the first l The weight matrix and bias vector of the multi-head self-attention output mapping in a layer encoder, with subscripts OIndicates output mapping, Indicates the first l The feature sequence obtained by residually concatenating the output features of the multi-head self-attention module in the layer encoder with the input feature sequence. Indicates the first l Output features of the multi-head self-attention module in the layer encoder Indicates the first l The feedforward network feature processing function in the layer encoder, Indicates the first l The contextual features output by the layer Transformer encoder. and They represent the first l The weight matrix and bias vector of the first linear mapping layer in the position-by-position feedforward network of the layer encoder. This indicates a preset nonlinear activation function. and They represent the first l The weight matrix and bias vector of the second linear mapping layer in the position-by-position feedforward network of the layer encoder; After the Transformer encoder with a preset number of layers updates the features sequentially, the final context features corresponding to each sub-problem are obtained. Step 4.4: Input the final context features of each sub-problem into the output decision module, and obtain the mutation policy class probability, scaling factor class probability, and crossover probability class probability through the policy classification branch, scaling factor classification branch, and crossover probability classification branch, respectively. , In the above formula, Indicates the first j The final context features obtained by passing each sub-question through the Transformer encoder Represents the probability of the mutation strategy category. Represents the scaling factor class probability. Represents the cross probability and class probability. , , These represent the weight matrices for the strategy classification branch, scaling factor classification branch, and crossover probability classification branch, respectively. , , These represent the bias terms for the strategy classification branch, scaling factor classification branch, and cross-probability classification branch, respectively. Step 4.5, determine the first [probability] based on the principle of maximizing probability. j Mutation strategy and scaling factor for each sub-problem F j and crossover probability CR jscaling factor F The crossover probability is determined from the preset candidate scaling factor set {0.1, 0.3, 0.5, 0.7, 0.9}. CR Determined from the preset candidate set of crossover probabilities {0.1, 0.3, 0.5, 0.7, 0.9}; Step 4.6, regarding the first j For each subproblem, a parent candidate pool is constructed based on the representative individual vector of the subproblem and the representative individual vectors of its neighboring subproblems, and the parent individual vectors required to generate the mutation vector are selected from the parent candidate pool. Step 4.7, based on the mutation strategy and scaling factor determined in Step 4.5. F j Using the parent individual vector selected in step 4.6, the first generation is generated. j The mutation vector of each sub-problem Considering the search requirements of different sub-problems at different search stages after the multi-objective optimization problem is decomposed into multiple sub-problems by the reference vector, the mutation strategy includes a first type of mutation strategy focusing on global exploration, a second type of mutation strategy focusing on local development, and a third type of mutation strategy that takes into account both global exploration and local development, as detailed below: For subproblems that are still in the early stages of search or facing search stagnation, requiring extensive exploration in the design space to locate potential superior regions, a first-type mutation strategy is designed. This strategy focuses on global exploration, using random individuals as base points and employing a long-range random perturbation factor with power-law tail characteristics to enable the trial vector to generate random displacements in the search space. This allows for rapid coverage of design regions that have not yet been fully searched. The first-type mutation strategy can be determined by the following formula: , In the above formula, This represents the vector of the first parent individual randomly selected from the parent candidate pool. This represents the vector of the second parent individual randomly selected from the parent candidate pool. This represents the vector of the third parent individual randomly selected from the parent candidate pool. This represents the vector of the fourth parent individual randomly selected from the parent candidate pool. This represents the vector of the fifth parent individual randomly selected from the parent candidate pool. F j For the Transformer model, the first j The perturbation scaling factor of the output of each subproblem. This represents the amplitude adjustment coefficient for the Levy flight differential disturbance term. The value is 0.01. This represents the random step size of the Levy flight generated using the Mantegna algorithm. The distribution parameters representing the random step size of Levy's flight path are... The value is 1.5, therefore, the Levy flight random step size Determined by the following formula: , in, and Let be mutually independent random variables with first and second normal distributions, satisfying: , in, This indicates that the mean is 0 and the variance is 0. The normal distribution Represents a random variable with the first normal distribution. standard deviation Represents a random variable with a second normal distribution. The absolute value, Represents the gamma function. Represents the sine function. π Represents pi; For subproblems where a relatively good solution has been found, and further refinement is needed in the vicinity to improve convergence accuracy, a second type of mutation strategy is designed. This strategy focuses on local refinement and takes the optimal individual vector with the smallest current aggregation function value in the neighborhood of the subproblem as the search starting point. The improved second type of mutation strategy, which moves along the direction of the random parent difference vector, can be determined by the following formula: , In the above formula, For the first j The optimal individual vector with the smallest current aggregation function value within the neighborhood of each subproblem. Indicates the first j The mutated individual vectors of each subproblem The scaling factor is the Cauchy perturbation. The value is 0.01. This represents a standard Cauchy distribution random number with a location parameter of 0 and a scale parameter of 1. For some subproblems where convergence has been established, a balance needs to be struck between maintaining the convergence trend and escaping local optima. A third type of mutation strategy is designed, focusing on the dynamic balance between exploration and development. Starting from the current individual vector, it includes a convergence direction component pointing to the optimal individual. A centralized random perturbation factor with exponentially decaying tails is used to controllably scale the auxiliary component. This injects search diversity in a controlled manner while maintaining the general direction of convergence towards the optimal solution, achieving gradual and robust improvement. The third type of mutation strategy can be determined by the following formula: , In the above formula, For the current individual vector, is the Gaussian perturbation scaling factor. The value is 0.01. These are random numbers distributed according to a standard normal distribution. Step 4.8, according to the... j The crossover probability corresponding to each sub-problem CR j , for the j The mutation vector of each sub-problem Perform a crossover operation with the parent individual vector to generate a crossover individual vector. ; for cross individual vectors Each structural design variable is checked for boundaries to ensure that the longitudinal beam plate thickness, longitudinal beam section height, longitudinal beam section width, transverse beam plate thickness, transverse beam section height, transverse beam section width, number of transverse beams, and number of longitudinal beams are all within the value range corresponding to the multidimensional design space. Specifically, when a structural design variable is less than its lower limit, it is corrected to the lower limit; when it is greater than its upper limit, it is corrected to the upper limit; and when it is between its lower and upper limits, its value remains unchanged. This boundary processing yields the first... j Candidate offspring vectors for each subproblem.

[0010] Furthermore, step (5) specifically includes: Step 5.1: Use the constructed radial basis function model to predict the mass, torsional stiffness and first natural frequency of each candidate offspring individual vector, obtain the prediction target value corresponding to each candidate offspring individual vector, and perform non-dominated sorting according to the prediction target value to obtain the prediction non-dominated candidate set. Step 5.2: Calculate the positive improvement distance and negative deviation distance of each candidate offspring individual vector in the predicted non-dominated candidate set relative to the Pareto reference set in the current database. The positive improvement distance is used to measure the potential improvement of the candidate offspring individual vector in the target space relative to the current Pareto reference set, and the negative deviation distance is used to measure the degree of deviation of the candidate offspring individual vector from the current Pareto reference set. Step 5.3: Select an individual vector with frontier expansion capability based on the positive improvement distance, select an individual vector with convergence improvement capability based on the negative deviation distance, and remove duplicates from the selected individual vectors to obtain the individual vectors to be evaluated that need to be accurately simulated. Step 5.4: Perform mass calculation, torsional static analysis and free modal analysis on the individual vector to be evaluated to obtain its mass, torsional stiffness and first natural frequency target values. Store the individual vector to be evaluated and its target values ​​in the database and update the current population and radial basis function model. Step 5.5: Determine whether the Pareto set in the database meets the design requirements for frame mass, torsional stiffness and first natural frequency. If it does, output the Pareto optimized structure of the new energy pickup truck frame corresponding to each volume vector in the Pareto set. Otherwise, return to step (3) until a Pareto optimized structure that meets the design requirements is obtained.

[0011] In a second aspect, the present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the aforementioned Transformer-based new energy pickup truck frame MOP method.

[0012] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the aforementioned Transformer-based new energy pickup truck frame MOP method.

[0013] Fourthly, the present invention provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the aforementioned Transformer-based new energy pickup truck frame MOP method.

[0014] In summary, compared with the prior art, the present invention provides a Transformer-based method, device, medium, and program product for MOP (Model-Operated Platform) of new energy pickup trucks, which overcomes the limitations of the prior art in the following ways: 1. Considering the differences in target search direction, neighborhood convergence state, variable perturbation response, and historical improvement state of subproblems corresponding to different reference vectors in the multi-objective optimization of new energy pickup truck frames, this invention constructs a state matrix sequence for each subproblem, so that the current search state of different subproblems can be structurally represented, providing a basis for the Transformer model to make state-aware decisions; 2. To address the problem that traditional evolutionary methods use fixed mutation strategies for each subproblem, leading to insufficient or repeated searches in some areas, this invention uses the Transformer model to determine corresponding mutation strategies for each subproblem, enabling different subproblems to choose mutation strategies that focus on global exploration, local development, or a combination of both, based on their search status. 3. Addressing the scaling factor in traditional evolutionary methods F and crossover probability CR Addressing the problem of relying on manual experience for scaling settings, which is difficult to adapt to different search stages, this invention utilizes the Transformer model to determine the scaling factor for each sub-problem. F and crossover probability CRThis allows the mutation magnitude and crossover strength to adaptively adjust with changes in the state of the subproblem, thereby improving the relevance and quality of candidate offspring individual vectors; 4. To address the high computational costs of evaluation processes such as weight calculation, torsional static analysis, and free modal analysis for new energy pickup truck frames, this invention utilizes a radial basis function model to predict the target values ​​of candidate offspring individual vectors. Based on positive improvement distance and negative deviation distance, the individual vectors to be evaluated are screened and deduplicated. Only the candidate offspring individual vectors to be evaluated are subjected to accurate simulation evaluation, thereby reducing unnecessary finite element simulation evaluation, while taking into account the expansion capability and convergence improvement capability of the Pareto front. 5. This invention can optimize the multi-objective structural optimization problem of new energy pickup truck frames involving complex simulations, improve the optimization efficiency of complex simulations, and can perform synergistic trade-off optimization for multiple objectives such as lightweighting, torsional performance and modal performance to obtain Pareto-optimal structures that meet different design requirements. This is beneficial for the optimization application of new energy pickup truck frames and other complex structures, and has practicality. Attached Figure Description

[0015] Figure 1 A simplified flowchart of a Transformer-based MOP method for new energy pickup truck frames provided by this invention. Detailed Implementation

[0016] To more clearly illustrate the objectives, technical solutions, and advantages of this invention, we will describe them in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention. Furthermore, the technical features in the following embodiments can be combined with each other as long as they do not conflict with each other.

[0017] Please see Figure 1 This invention provides a Transformer-based MOP method for new energy pickup truck frames, which is applicable to the three-objective optimization problem of lightweighting, torsional stiffness, and first-order natural frequency of new energy pickup truck frames. Specifically, the method includes steps (1) to (5).

[0018] Step (1): Based on the structural design variables of the new energy pickup truck frame, establish a parametric three-dimensional model and a finite element model, and construct an MOP model with mass, torsional stiffness and first natural frequency as the objectives; The specific steps of step (1) are as follows: Step 1.1: Based on the structural layout, manufacturing process, component installation space, and adjustable range of structural parameters of the new energy pickup truck frame, determine the range of values ​​for longitudinal beam plate thickness, longitudinal beam section height, longitudinal beam section width, crossbeam plate thickness, crossbeam section height, crossbeam section width, number of crossbeams, and number of longitudinal beams, and construct a structural design variable vector and a multi-dimensional design space. Step 1.2: Based on the non-load-bearing chassis structure, material properties, load-bearing component layout and typical driving conditions of the new energy pickup truck frame, establish a parametric 3D model of the new energy pickup truck frame in CATIA, and obtain the frame mass based on the material density and the geometric volume of the parametric 3D model. Step 1.3: Import the parametric 3D model into HyperMesh for geometry cleanup, mid-surface extraction, shell element mesh generation, material property assignment, plate thickness property definition, connection relationship definition, load application, and boundary condition setting to obtain the finite element model of the new energy pickup truck frame. Step 1.4: Use OptiStruct to perform torsional static analysis on the finite element model, extract the vertical displacement of the left and right loading points, calculate the torsional load moment based on the single-sided vertical loading force and the distance between the left and right loading points, calculate the frame torsion angle from the displacement difference between the left and right loading points, and finally calculate the torsional stiffness. Step 1.5: Use OptiStruct to perform free modal analysis on the finite element model, solve the modal characteristic equations corresponding to the finite element stiffness matrix and mass matrix, obtain the circular frequencies and mode shapes of each order, and thus obtain the first natural frequency. Step 1.6: Based on the design variable vector of the new energy pickup truck frame, the frame mass, torsional stiffness, and first-order natural frequency, construct the MOP mathematical model, the specific expression of which is as follows: , In the above formula, X This represents the vector of design variables for the frame structure of new energy pickup trucks. St represents the multidimensional design space comprised of the value ranges of each structural design variable, and St represents the vector of structural design variables for the new energy pickup truck frame. X The multidimensional design space constrained by the range of values ​​of each structural design variable middle, Indicates the thickness of the longitudinal beam plate. Indicates the height of the longitudinal beam section. Indicates the width of the longitudinal beam section. Indicates the thickness of the crossbeam plate. Indicates the height of the beam section. Indicates the width of the beam cross section. Indicates the number of beams. Indicates the number of longitudinal beams. This indicates the density of the frame material for new energy pickup trucks. Represents a vector of structural design variables X The corresponding parametric model geometric volume, Represents a vector of structural design variables X The corresponding weight of the new energy pickup truck frame, This represents the unilateral vertical loading force applied under torsional static conditions. B Indicates the spacing between the left and right loading points. Represents a vector of structural design variables X The corresponding torsional load moment, This represents the vertical displacement of the loading point on the left side in the torsional static analysis. This represents the vertical displacement of the loading point on the right side in the torsional static analysis. Represents a vector of structural design variables X The corresponding frame torsion angle, Represents a vector of structural design variables X The corresponding torsional stiffness of the new energy pickup truck frame, Represents a vector of structural design variables X The corresponding finite element stiffness matrix, Represents a vector of structural design variables X The corresponding finite element mass matrix, Indicates the first q Mode shape, Indicates the first q Circular frequency, Indicates the first-order angular frequency. Represents a vector of structural design variables X The corresponding first-order natural frequency of the new energy pickup truck frame, This represents the objective function vector for the multi-objective optimization of the new energy pickup truck frame. This means transforming the objective of maximizing torsional stiffness into the objective of minimizing it. This means transforming the objective of maximizing the first-order natural frequency into a minimization objective; Find indicates that in the multidimensional design space... Finding structural design variable vectors X The optimal solution, min represents the optimization of the objective function vector for multiple objectives. Perform a minimization optimization.

[0019] Step (2): The initial population is generated by Latin hypercube sampling, the reference vector is generated by simplex method, and the MOP model is decomposed into multiple sub-problems. The initial population is evaluated by simulation, and a database is established and a radial basis function model is constructed. The specific steps for step (2) are as follows: Step 2.1: Determine the initial number of individual vectors based on the number of structural design variables, the number of optimization objectives, and the preset simulation evaluation budget. N p Latin hypercube sampling is used to generate [the data] in a multidimensional design space. N p The initial individual vectors form the initial population; Step 2.2: Based on the number of optimization objectives, generate the target space consisting of three objectives using the simplex method.N v Reference vectors, and using N v The multi-objective optimization problem of the new energy pickup truck frame is decomposed into several reference vectors. N v The issue of size; Step 2.3: Perform mass calculation, torsional static analysis and free modal analysis on each initial individual vector in the initial population to obtain the corresponding target values ​​of mass, torsional stiffness and first natural frequency, and store the initial individual vector and its target values ​​in the database; Step 2.4: Construct a radial basis function model using all individual vector information in the database.

[0020] Step (3): Determine the ideal reference point and the nadir reference point, normalize the target value, associate the individual vectors in the population with the reference vector, determine the representative individual vectors of each subproblem and construct the state matrix sequence; The specific steps for step (3) are as follows: Step 3.1: Determine the ideal reference point based on the database and the target value in the current population. and Nadir Reference Point and using ideal reference points and Nadir Reference Point The target values ​​corresponding to mass, torsional stiffness, and first natural frequency are normalized to obtain the normalized target values ​​of each body vector in the current population. The ideal reference point and the nadir reference point are determined by the following formula: , , In the above formula, The ideal reference point is at the th j The values ​​of each target, This indicates that the nadir reference point is at the 1st ytterbium position. j The values ​​of each target, DB For database, x represents i For individual vectors in the database, Represents the first in the database i Individual vector x i The j One target value, This represents minimizing the individual vector x in the database. i The j One target value, This means maximizing the individual vector x in the database. i The j One target value, m It is the target number; Step 3.2: Calculate the perpendicular distance between the normalized target value of each body vector in the current population and each reference vector. Establish an association relationship between each body vector and the reference vector with the smallest perpendicular distance, and determine the associated body set for each reference vector corresponding to the subproblem based on the association relationship. For subproblems with associated body sets, calculate the Chebyshev aggregation value of each body vector in the associated body set under the subproblem, and select the body vector with the smallest Chebyshev aggregation value as the representative body for the subproblem. For subproblems without associated body sets, first select the body vector with the smallest Chebyshev aggregation value under the current subproblem from the representative bodies of the neighboring subproblems as the representative body. If there is no representative body in the neighboring subproblems, then select the body vector with the smallest Chebyshev aggregation value from the current population as the representative body. The Chebyshev aggregation value is calculated by the following formula: , In the above formula, It is a Chebyshev aggregation function constructed based on an ideal reference point. For the first i Individual vectors, Is the reference vector at the th j The weight of each objective; Step 3.3: For each subproblem, a state matrix is ​​formed sequentially around the reference vector and representative individuals. The data of different dimensions in the state matrix are then normalized. The state matrix consists of the reference vector direction, the target state of the representative individual, the neighborhood search state, the variable perturbation response, and the historical improvement state. The reference vector direction is represented by the components of the subproblem's reference vector in the three target directions: mass, torsional stiffness, and first-order natural frequency. The target state of the representative individual is represented by the normalized target value of the representative individual, the perpendicular distance between the representative individual and the reference vector, and the Chebyshev aggregation value of the representative individual under the current subproblem. The neighborhood search state is represented by the distribution of Chebyshev aggregation values ​​of representative individuals in the neighborhood subproblems, the number of individuals associated with the current subproblem, and the dispersion of the neighborhood target space. The variable perturbation response is represented by the changes in the target value and Chebyshev aggregation value of the representative individual's structural design variables before and after local perturbations. The historical improvement state is represented by the historical success rate, number of stalls, and the most recent improvement amount of the subproblem in previous iterations. Step 3.4: Arrange the state matrices of each subproblem in the order of the reference vector numbers to form a state matrix sequence.

[0021] Step (4): Use the Transformer model to process the state matrix sequence, determine the mutation strategy, scaling factor and crossover probability for each subproblem, and generate candidate offspring individual vectors; The specific steps for step (4) are as follows: Step 4.1: Normalize the state variables in the state matrix sequence, and then classify them according to their feature sources. j The state characteristics of each sub-problem are divided into a reference vector feature group, a target search feature group, a neighborhood distribution feature group, a variable response feature group, and a historical evolution feature group. j State characteristics of each sub-problem Represented as: , In the above formula, Indicates the first j The reference vector feature set of each sub-problem Indicates the first j The target search feature set for each sub-problem Indicates the first j The neighborhood distribution characteristics of each sub-problem Indicates the first j Variable response feature set of sub-problems Indicates the first j The historical evolutionary characteristics of individual sub-problems; Step 4.2: Using the feature mapping module, feature mapping is performed on the reference vector feature group, target search feature group, neighborhood distribution feature group, variable response feature group, and historical evolution feature group, respectively. The mapped features are then fused and superimposed with positional codes representing the reference vector numbering order to obtain the [number of features]. j Initial feature vectors for each subproblem: , In the above formula, Indicates the first j The sub-problems are input into the initial feature vector before the Transformer encoder. Indicates the feature group identifier. w , o , n , v , h These correspond to the identifiers of the reference vector feature group, target search feature group, neighborhood distribution feature group, variable response feature group, and historical evolution feature group, respectively. Indicates the first j The first sub-problem After normalization, the state feature vector of the class feature group is input to the feature mapping layer. Indicates the first j The first sub-problem The feature vector obtained after feature mapping of the class feature group Indicates the first j The feature vectors obtained after feature mapping of the reference vector feature set of each subproblem Indicates the first jThe feature vector obtained after feature mapping of the target search feature set of each sub-problem Indicates the first j The feature vector is obtained by feature mapping of the neighborhood distribution feature set of each subproblem. Indicates the first j The feature vector obtained after feature mapping of the variable response feature set of each subproblem Indicates the first j The feature vector is obtained by feature mapping of the historical evolution feature set of each sub-problem. and They represent the first The mapping weight matrix and bias terms corresponding to the feature groups of each class Represents a non-linear activation function. and These represent the weight matrix and bias term of the fusion mapping layer, respectively. Indicates the first j The positional encoding corresponding to each sub-problem; Step 4.3: The initial feature vectors of each sub-problem are arranged into an input feature sequence according to the reference vector numbering order, and then input into the Transformer encoder. The contextual relationships between different sub-problems are extracted through a multi-head self-attention mechanism and a feedforward network. The first step of the Transformer encoder... l The layer updates its features according to the following formula: , In the above formula, l Indicates the layer number of the Transformer encoder. Indicates the first l The input feature sequence of the layer encoder, i.e., the first layer... The context feature sequence output by the layer encoder, Representation layer normalization, Indicates to The feature sequence obtained after layer normalization r This indicates the attention head number in the multi-head self-attention module, and , R Indicates the number of attention heads. , , They represent the first l The first layer encoder r The query mapping weight matrix, key mapping weight matrix, and value mapping weight matrix corresponding to each attention head. Indicates the first l The first layer encoder r Attention features of the output of each attention head. Represents the normalized exponential function, Indicates the firstl The attention features output by the first attention head in the layer encoder. Indicates the first l The first layer encoder R Attention features output by each attention head, superscript Indicates matrix transpose. d h The feature dimension of a single attention head. , and They represent the first l The first layer encoder r The query matrix, key matrix, and value matrix corresponding to each attention head. This indicates that the output features of each attention head are concatenated. and They represent the first l The weight matrix and bias vector of the multi-head self-attention output mapping in a layer encoder, with subscripts O Indicates output mapping, Indicates the first l The feature sequence obtained by residually concatenating the output features of the multi-head self-attention module in the layer encoder with the input feature sequence. Indicates the first l Output features of the multi-head self-attention module in the layer encoder Indicates the first l The feedforward network feature processing function in the layer encoder, Indicates the first l The contextual features output by the layer Transformer encoder. and They represent the first l The weight matrix and bias vector of the first linear mapping layer in the position-by-position feedforward network of the layer encoder. This indicates a preset nonlinear activation function. and They represent the first l The weight matrix and bias vector of the second linear mapping layer in the position-by-position feedforward network of the layer encoder; After the Transformer encoder with a preset number of layers updates the features sequentially, the final context features corresponding to each sub-problem are obtained. Step 4.4: Input the final context features of each sub-problem into the output decision module, and obtain the mutation policy class probability, scaling factor class probability, and crossover probability class probability through the policy classification branch, scaling factor classification branch, and crossover probability classification branch, respectively. , In the above formula, Indicates the firstj The final context features obtained by passing each sub-question through the Transformer encoder Represents the probability of the mutation strategy category. Represents the scaling factor class probability. Represents the cross probability and class probability. , , These represent the weight matrices for the strategy classification branch, scaling factor classification branch, and crossover probability classification branch, respectively. , , These represent the bias terms for the strategy classification branch, scaling factor classification branch, and cross-probability classification branch, respectively. Step 4.5, determine the first [probability] based on the principle of maximizing probability. j Mutation strategy and scaling factor for each sub-problem F j and crossover probability CR j scaling factor F The crossover probability is determined from the preset candidate scaling factor set {0.1, 0.3, 0.5, 0.7, 0.9}. CR Determined from the preset candidate set of crossover probabilities {0.1, 0.3, 0.5, 0.7, 0.9}; Step 4.6, regarding the first j For each subproblem, a parent candidate pool is constructed based on the representative individual vector of the subproblem and the representative individual vectors of its neighboring subproblems, and the parent individual vectors required to generate the mutation vector are selected from the parent candidate pool. Step 4.7, based on the mutation strategy and scaling factor determined in Step 4.5. F j Using the parent individual vector selected in step 4.6, the first generation is generated. j The mutation vector of each sub-problem Considering the search requirements of different sub-problems at different search stages after the multi-objective optimization problem is decomposed into multiple sub-problems by the reference vector, the mutation strategy includes a first type of mutation strategy focusing on global exploration, a second type of mutation strategy focusing on local development, and a third type of mutation strategy that takes into account both global exploration and local development, as detailed below: For subproblems that are still in the early stages of search or facing search stagnation, requiring extensive exploration in the design space to locate potential superior regions, a first-type mutation strategy is designed. This strategy focuses on global exploration, using random individuals as base points and employing a long-range random perturbation factor with power-law tail characteristics to enable the trial vector to generate random displacements in the search space. This allows for rapid coverage of design regions that have not yet been fully searched. The first-type mutation strategy can be determined by the following formula: , In the above formula, This represents the vector of the first parent individual randomly selected from the parent candidate pool. This represents the vector of the second parent individual randomly selected from the parent candidate pool. This represents the vector of the third parent individual randomly selected from the parent candidate pool. This represents the vector of the fourth parent individual randomly selected from the parent candidate pool. This represents the vector of the fifth parent individual randomly selected from the parent candidate pool. F j For the Transformer model, the first j The perturbation scaling factor of the output of each subproblem. This represents the amplitude adjustment coefficient for the Levy flight differential disturbance term. The value is 0.01. This represents the random step size of the Levy flight generated using the Mantegna algorithm. The distribution parameters representing the random step size of Levy's flight path are... The value is 1.5, therefore, the Levy flight random step size Determined by the following formula: , in, and Let be mutually independent random variables with first and second normal distributions, satisfying: , in, This indicates that the mean is 0 and the variance is 0. The normal distribution Represents a random variable with the first normal distribution. standard deviation Represents a random variable with a second normal distribution. The absolute value, Represents the gamma function. Represents the sine function. π Represents pi; For subproblems where a relatively good solution has been found, and further refinement is needed in the vicinity to improve convergence accuracy, a second type of mutation strategy is designed. This strategy focuses on local refinement and takes the optimal individual vector with the smallest current aggregation function value in the neighborhood of the subproblem as the search starting point. The improved second type of mutation strategy, which moves along the direction of the random parent difference vector, can be determined by the following formula: , In the above formula, For the first j The optimal individual vector with the smallest current aggregation function value within the neighborhood of each subproblem. Indicates the first j The mutated individual vectors of each subproblem The scaling factor is the Cauchy perturbation. The value is 0.01. This represents a standard Cauchy distribution random number with a location parameter of 0 and a scale parameter of 1. For some subproblems where convergence has been established, a balance needs to be struck between maintaining the convergence trend and escaping local optima. A third type of mutation strategy is designed, focusing on the dynamic balance between exploration and development. Starting from the current individual vector, it includes a convergence direction component pointing to the optimal individual. A centralized random perturbation factor with exponentially decaying tails is used to controllably scale the auxiliary component. This injects search diversity in a controlled manner while maintaining the general direction of convergence towards the optimal solution, achieving gradual and robust improvement. The third type of mutation strategy can be determined by the following formula: , In the above formula, For the current individual vector, is the Gaussian perturbation scaling factor. The value is 0.01. These are random numbers distributed according to a standard normal distribution. Step 4.8, according to the... j The crossover probability corresponding to each sub-problem CR j , for the j The mutation vector of each sub-problem Perform a crossover operation with the parent individual vector to generate a crossover individual vector. ; for cross individual vectors Each structural design variable is checked for boundaries to ensure that the longitudinal beam plate thickness, longitudinal beam section height, longitudinal beam section width, transverse beam plate thickness, transverse beam section height, transverse beam section width, number of transverse beams, and number of longitudinal beams are all within the value range corresponding to the multidimensional design space. Specifically, when a structural design variable is less than its lower limit, it is corrected to the lower limit; when it is greater than its upper limit, it is corrected to the upper limit; and when it is between its lower and upper limits, its value remains unchanged. This boundary processing yields the first... j Candidate offspring vectors for each subproblem.

[0022] Step (5): Predict the target value of candidate offspring individual vectors using the radial basis function model, filter and deduplicate the individual vectors to be evaluated based on the positive improvement distance and negative deviation distance, perform simulation evaluation, update the database, population and radial basis function model, and return to step (3) until the Pareto optimization structure that meets the design requirements is output. Step (5) The specific steps are as follows: Step 5.1: Use the constructed radial basis function model to predict the mass, torsional stiffness and first natural frequency of each candidate offspring individual vector, obtain the prediction target value corresponding to each candidate offspring individual vector, and perform non-dominated sorting according to the prediction target value to obtain the prediction non-dominated candidate set. Step 5.2: Calculate the positive improvement distance and negative deviation distance of each candidate offspring individual vector in the predicted non-dominated candidate set relative to the Pareto reference set in the current database. The positive improvement distance is used to measure the potential improvement of the candidate offspring individual vector in the target space relative to the current Pareto reference set, and the negative deviation distance is used to measure the degree of deviation of the candidate offspring individual vector from the current Pareto reference set. Step 5.3: Select an individual vector with frontier expansion capability based on the positive improvement distance, select an individual vector with convergence improvement capability based on the negative deviation distance, and remove duplicates from the selected individual vectors to obtain the individual vectors to be evaluated that need to be accurately simulated. Step 5.4: Perform mass calculation, torsional static analysis and free modal analysis on the individual vector to be evaluated to obtain its mass, torsional stiffness and first natural frequency target values. Store the individual vector to be evaluated and its target values ​​in the database and update the current population and radial basis function model. Step 5.5: Determine whether the Pareto set in the database meets the design requirements for frame mass, torsional stiffness and first natural frequency. If it does, output the Pareto optimized structure of the new energy pickup truck frame corresponding to each volume vector in the Pareto set. Otherwise, return to step (3) until a Pareto optimized structure that meets the design requirements is obtained.

[0023] Example 1 uses the benchmark function DTLZ1 to illustrate the optimization performance of a Transformer-based MOP method for new energy pickup truck frames. The expression of the benchmark function DTLZ1, which has three objectives, is as follows: , In the above formula, Represents a vector of design variables. This represents the first design variable. This represents the second design variable. x i Indicates the first i One design variable, n Indicates the number of design variables, and , Indicates by the first m The first to the second n A vector of variables, composed of several design variables, used to calculate the distance function.m Indicates the number of objective functions. This indicates the number of design variables involved in the distance function calculation, where , This represents the first objective function value. This represents the value of the second objective function. This represents the value of the third objective function. This represents a distance function used to describe the degree to which the design variable vector is far from the Pareto optimal region. π It represents pi (π).

[0024] In this embodiment, the number of design variables n Set to 30, number of objective functions m Setting it to 3 satisfies the Pareto optimal frontier. ,in The first Pareto optimal frontier is represented by the... j The objective function value.

[0025] The above benchmark test function DTLZ1 is processed through steps (1) to (5) of the MOP method for new energy pickup truck frames based on Transformer provided by this invention to obtain the experimental results.

[0026] To illustrate this embodiment in more detail, the Transformer-based MOP method for new energy pickup truck frames in this embodiment is compared with another excellent self-organizing agent-assisted non-dominated sorting differential evolution algorithm (SSDE algorithm). The maximum number of simulation evaluations in this embodiment is set to 500. The experimental results are shown in Table 1. The comparison method is the average IGD value after 25 independent runs. Under the same number of simulations, the method in this embodiment is significantly better than the SSDE algorithm. It can be considered that the method in this embodiment can effectively solve the multi-objective optimization design problem of lightweighting, torsional stiffness, and first-order natural frequency of new energy pickup truck frames.

[0027] Table 1. Comparison of optimization results of different algorithms

[0028] The present invention provides a Transformer-based method for MOP (Model-Oriented Programming) of new energy pickup truck frames. This method constructs a state matrix sequence representing the current search state of a sub-problem and then uses a Transformer model to adaptively determine the mutation strategy and scaling factor. F and crossover probability CR By combining the radial basis function model and the dual-distance screening criterion, candidate offspring vectors for simulation evaluation are determined, providing a systematic solution for multi-objective collaborative optimization of lightweighting, torsional performance, and modal performance of new energy pickup truck frames.

[0029] In a second aspect, the present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of a Transformer-based new energy pickup truck frame MOP method of the foregoing embodiments.

[0030] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of a Transformer-based new energy pickup truck frame MOP method of the foregoing embodiments.

[0031] Fourthly, the present invention provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the Transformer-based MOP method for new energy pickup truck frames described in the foregoing embodiments.

[0032] Those skilled in the art will readily understand that the above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A Transformer-based method for MOP (Model-Operated Platform) of new energy pickup truck frames, characterized in that, The method includes: Step (1): Based on the structural design variables of the new energy pickup truck frame, establish a parametric three-dimensional model and a finite element model, and construct an MOP model with mass, torsional stiffness and first natural frequency as the objectives; Step (2): The initial population is generated by Latin hypercube sampling, the reference vector is generated by simplex method, and the MOP model is decomposed into multiple sub-problems. The initial population is evaluated by simulation, and a database is established and a radial basis function model is constructed. Step (3): Determine the ideal reference point and the nadir reference point, normalize the target value, associate the individual vectors in the current population with the reference vector, determine the representative individual vectors of each subproblem and construct the state matrix sequence; Step (4): Use the Transformer model to process the state matrix sequence, determine the mutation strategy, scaling factor and crossover probability for each subproblem, and generate candidate offspring individual vectors; Step (5): Predict the target value of candidate offspring individual vectors using the radial basis function model, filter and deduplicate based on the positive improvement distance and negative deviation distance to obtain the individual vectors to be evaluated for simulation evaluation, update the database, population and radial basis function model, and return to step (3) until the Pareto optimization structure that meets the design requirements is output.

2. The method as described in claim 1, characterized in that, Step (1) specifically includes: Step 1.1: Determine the structural design variables and their value ranges based on the structural layout, manufacturing process, component installation space, and adjustable range of structural parameters of the new energy pickup truck frame. Construct a structural design variable vector and a multi-dimensional design space. The structural design variables include longitudinal beam plate thickness, longitudinal beam section height, longitudinal beam section width, crossbeam plate thickness, crossbeam section height, crossbeam section width, number of crossbeams, and number of longitudinal beams. Step 1.2: Based on the non-load-bearing chassis structure, material properties, load-bearing component layout, and typical driving conditions of the new energy pickup truck frame, establish a parametric three-dimensional model in CATIA, and obtain the mass based on the material density and the geometric volume of the parametric three-dimensional model. Step 1.3: Import the parametric 3D model into HyperMesh for geometry cleanup, mid-surface extraction, shell element mesh generation, material property assignment, plate thickness property definition, connection relationship definition, load application, and boundary condition setting to obtain the finite element model; Step 1.4: Use OptiStruct to perform torsional static analysis on the finite element model, extract the vertical displacement of the left and right loading points, calculate the torsional load moment based on the single-sided vertical loading force and the distance between the left and right loading points, calculate the frame torsion angle from the displacement difference between the left and right loading points, and finally calculate the torsional stiffness. Step 1.5: Use OptiStruct to perform free modal analysis on the finite element model, solve the modal characteristic equations corresponding to the finite element stiffness matrix and mass matrix, obtain the circular frequencies and mode shapes of each order, and thus obtain the first natural frequency. Step 1.6: Based on the design variable vectors, mass, torsional stiffness, and first-order natural frequency of the new energy pickup truck frame, construct the MOP mathematical model, the specific expression of which is as follows: ， In the above formula, X This represents the vector of design variables for the frame structure of new energy pickup trucks. St represents the multidimensional design space comprised of the value ranges of each structural design variable, and St represents the vector of structural design variables for the new energy pickup truck frame. X The multidimensional design space constrained by the range of values ​​of each structural design variable middle, Indicates the thickness of the longitudinal beam plate. Indicates the height of the longitudinal beam section. Indicates the width of the longitudinal beam section. Indicates the thickness of the crossbeam plate. Indicates the height of the beam section. Indicates the width of the beam cross section. Indicates the number of beams. Indicates the number of longitudinal beams. This indicates the density of the frame material for new energy pickup trucks. Represents a vector of structural design variables X The corresponding parametric model geometric volume, Represents a vector of structural design variables X The corresponding weight of the new energy pickup truck frame, This represents the unilateral vertical loading force applied under torsional static conditions. B Indicates the spacing between the left and right loading points. Represents a vector of structural design variables X The corresponding torsional load moment, This represents the vertical displacement of the loading point on the left side in the torsional static analysis. This represents the vertical displacement of the loading point on the right side in the torsional static analysis. Represents a vector of structural design variables X The corresponding frame torsion angle, Represents a vector of structural design variables X The corresponding torsional stiffness of the new energy pickup truck frame, Represents a vector of structural design variables X The corresponding finite element stiffness matrix, Represents a vector of structural design variables X The corresponding finite element mass matrix, Indicates the first q Mode shape, Indicates the first q Circular frequency, Indicates the first-order angular frequency. Represents a vector of structural design variables X The corresponding first-order natural frequency of the new energy pickup truck frame, This represents the objective function vector for the multi-objective optimization of the new energy pickup truck frame. This means transforming the objective of maximizing torsional stiffness into the objective of minimizing it. This means transforming the objective of maximizing the first-order natural frequency into a minimization objective; Find indicates that in the multidimensional design space... Finding structural design variable vectors X The optimal solution, min represents the optimization of the objective function vector for multiple objectives. Perform a minimization optimization.

3. The method as described in claim 1, characterized in that, Step (2) specifically includes: Step 2.1: Determine the initial number of individual vectors based on the number of structural design variables, the number of objectives, and the preset simulation evaluation budget. N p Latin hypercube sampling is used to generate [the data] in a multidimensional design space. N p The initial individual vectors form the initial population; Step 2.2: Based on the number of targets, generate the target space consisting of three targets using the simplex method. N v Reference vectors, and using N v The MOP model is decomposed into several reference vectors. N v The issue of size; Step 2.3: Perform mass calculation, torsional static analysis and free modal analysis on each initial individual vector in the initial population to obtain the target values ​​corresponding to the mass, torsional stiffness and first natural frequency of each initial individual vector, and store the initial individual vector and its target value in the database. Step 2.4: Construct a radial basis function model using all individual vector information in the database.

4. The method as described in claim 1, characterized in that, Step (3) specifically includes: Step 3.1: Determine the ideal reference point based on the database and the target value in the current population. and Nadir Reference Point and using ideal reference points and Nadir Reference Point The target values ​​corresponding to mass, torsional stiffness, and first natural frequency are normalized to obtain the normalized target values ​​of each body vector in the current population. The ideal reference point and the nadir reference point are determined by the following formula: ， ， In the above formula, The ideal reference point is at the th j The values ​​of each target, This indicates that the nadir reference point is at the 1st ytterbium position. j The values ​​of each target, DB For database, x represents i For individual vectors in the database, Represents the first in the database i Individual vector x i The j One target value, This represents minimizing the individual vector x in the database. i The j One target value, This means maximizing the individual vector x in the database. i The j One target value, m It is the target number; Step 3.2: Calculate the perpendicular distance between the normalized target value of each body vector in the current population and each reference vector. Establish an association relationship between each body vector and the reference vector with the smallest perpendicular distance, and determine the associated body set for each reference vector corresponding to the subproblem based on the association relationship. For subproblems with associated body sets, calculate the Chebyshev aggregation value of each body vector in the associated body set under the subproblem, and select the body vector with the smallest Chebyshev aggregation value as the representative body vector of the subproblem. For subproblems without associated body sets, first select the body vector with the smallest Chebyshev aggregation value under the current subproblem from the representative body vectors of the neighboring subproblems as the representative body vector. If there is no representative body vector in the neighboring subproblems, then select the body vector with the smallest Chebyshev aggregation value from the current population as the representative body vector. The Chebyshev aggregation value is calculated by the following formula: ， In the above formula, It is a Chebyshev aggregation function constructed based on an ideal reference point. It is a reference vector. Is the reference vector at the th j The weight of each objective; Step 3.3: For each subproblem, a state matrix is ​​formed sequentially around the reference vector and representative individual vector. The data of different dimensions in the state matrix are then normalized. The state matrix consists of the reference vector direction, the target state of the representative individual vector, the neighborhood search state, the variable perturbation response, and the historical improvement state. The reference vector direction is represented by the components of the reference vector in the three target directions of mass, torsional stiffness, and first-order natural frequency. The target state of the representative individual vector is represented by the normalized target value of the representative individual vector, the perpendicular distance between the representative individual vector and the reference vector, and the Chebyshev aggregation value of the representative individual vector under the current subproblem. The neighborhood search state is represented by the distribution of Chebyshev aggregation values ​​of the representative individual vectors in the neighborhood subproblems, the number of individuals associated with the current subproblem, and the dispersion of the neighborhood target space. The variable perturbation response is represented by the changes in the target value and Chebyshev aggregation value of the structural design variables of the representative individual vector before and after local perturbation. The historical improvement state is represented by the historical success rate, number of stalls, and the most recent improvement amount in the previous iterations of the subproblem. Step 3.4: Arrange the state matrices of each subproblem in the order of the reference vector numbers to form a state matrix sequence.

5. The method as described in claim 1, characterized in that, Step (4) specifically includes: Step 4.1: Normalize the state variables in the state matrix sequence, and then classify them according to their feature sources. j The state characteristics of each sub-problem are divided into a reference vector feature group, a target search feature group, a neighborhood distribution feature group, a variable response feature group, and a historical evolution feature group. j State characteristics of each sub-problem Represented as: ， In the above formula, Indicates the first j The reference vector feature set of each sub-problem Indicates the first j The target search feature set for each sub-problem Indicates the first j The neighborhood distribution characteristics of each sub-problem Indicates the first j Variable response feature set of sub-problems Indicates the first j The historical evolutionary characteristics of individual sub-problems; Step 4.2: Using the feature mapping module, feature mapping is performed on the reference vector feature group, target search feature group, neighborhood distribution feature group, variable response feature group, and historical evolution feature group, respectively. The mapped features are then fused and superimposed with positional codes representing the reference vector numbering order to obtain the [number of features]. j Initial feature vectors for each subproblem: ， In the above formula, Indicates the first j The sub-problems are input into the initial feature vector before the Transformer encoder. Indicates the feature group identifier. w , o , n , v , h These correspond to the identifiers of the reference vector feature group, target search feature group, neighborhood distribution feature group, variable response feature group, and historical evolution feature group, respectively. Indicates the first j The first sub-problem After normalization, the state feature vector of the class feature group is input to the feature mapping layer. Indicates the first j The first sub-problem The feature vector obtained after feature mapping of the class feature group Indicates the first j The feature vectors obtained after feature mapping of the reference vector feature set of each subproblem Indicates the first j The feature vector obtained after feature mapping of the target search feature set of each sub-problem Indicates the first j The feature vector is obtained by feature mapping of the neighborhood distribution feature set of each subproblem. Indicates the first j The feature vector obtained after feature mapping of the variable response feature set of each subproblem Indicates the first j The feature vector is obtained by feature mapping of the historical evolution feature set of each sub-problem. and They represent the first The mapping weight matrix and bias terms corresponding to the feature groups of each class Represents a non-linear activation function. and These represent the weight matrix and bias term of the fusion mapping layer, respectively. Indicates the first j The positional encoding corresponding to each sub-problem; Step 4.3: The initial feature vectors of each sub-problem are arranged into an input feature sequence according to the reference vector numbering order, and then input into the Transformer encoder. The contextual relationships between different sub-problems are extracted through a multi-head self-attention mechanism and a feedforward network. The first step of the Transformer encoder... l The layer updates its features according to the following formula: ， In the above formula, l Indicates the layer number of the Transformer encoder. Indicates the first l The input feature sequence of the layer encoder, i.e., the first layer... The context feature sequence output by the layer encoder, Representation layer normalization, Indicates to The feature sequence obtained after layer normalization r This indicates the attention head number in the multi-head self-attention module, and , R Indicates the number of attention heads. , , They represent the first l The first layer encoder r The query mapping weight matrix, key mapping weight matrix, and value mapping weight matrix corresponding to each attention head. Indicates the first l The first layer encoder r Attention features of the output of each attention head. Represents the normalized exponential function, Indicates the first l The attention features output by the first attention head in the layer encoder. Indicates the first l The first layer encoder R Attention features output by each attention head, superscript Indicates matrix transpose. d h The feature dimension of a single attention head. , and They represent the first l The first layer encoder r The query matrix, key matrix, and value matrix corresponding to each attention head. This indicates that the output features of each attention head are concatenated. and They represent the first l The weight matrix and bias vector of the multi-head self-attention output mapping in a layer encoder, with subscripts O Indicates output mapping, Indicates the first l The feature sequence obtained by residually concatenating the output features of the multi-head self-attention module in the layer encoder with the input feature sequence. Indicates the first l Output features of the multi-head self-attention module in the layer encoder Indicates the first l The feedforward network feature processing function in the layer encoder, Indicates the first l The contextual features output by the layer Transformer encoder. and They represent the first l The weight matrix and bias vector of the first linear mapping layer in the position-by-position feedforward network of the layer encoder. This indicates a preset nonlinear activation function. and They represent the first l The weight matrix and bias vector of the second linear mapping layer in the position-by-position feedforward network of the layer encoder; After the Transformer encoder with a preset number of layers updates the features sequentially, the final context features corresponding to each sub-problem are obtained. Step 4.4: Input the final context features of each sub-problem into the output decision module, and obtain the mutation policy class probability, scaling factor class probability, and crossover probability class probability through the policy classification branch, scaling factor classification branch, and crossover probability classification branch, respectively. ， In the above formula, Indicates the first j The final context features obtained by passing each sub-question through the Transformer encoder Represents the probability of the mutation strategy category. Represents the scaling factor class probability. Represents the cross probability and class probability. , , These represent the weight matrices for the strategy classification branch, scaling factor classification branch, and crossover probability classification branch, respectively. , , These represent the bias terms for the strategy classification branch, scaling factor classification branch, and cross-probability classification branch, respectively. Step 4.5, determine the first [probability] based on the principle of maximizing probability. j Mutation strategy and scaling factor for each sub-problem F j and crossover probability CR j scaling factor F The crossover probability is determined from the preset candidate scaling factor set {0.1, 0.3, 0.5, 0.7, 0.9}. CR Determined from the preset candidate set of crossover probabilities {0.1, 0.3, 0.5, 0.7, 0.9}; Step 4.6, regarding the first j For each subproblem, a parent candidate pool is constructed based on the representative individual vector of the subproblem and the representative individual vectors of its neighboring subproblems, and the parent individual vectors required to generate the mutation vector are selected from the parent candidate pool. Step 4.7, based on the mutation strategy and scaling factor determined in Step 4.

5. F j Using the parent individual vector selected in step 4.6, the first generation is generated. j The mutation vector of each sub-problem Considering the search requirements of different sub-problems at different search stages after the multi-objective optimization problem is decomposed into multiple sub-problems by the reference vector, the mutation strategy includes a first type of mutation strategy focusing on global exploration, a second type of mutation strategy focusing on local development, and a third type of mutation strategy that takes into account both global exploration and local development, as detailed below: For subproblems that are still in the early stages of search or facing search stagnation, requiring extensive exploration in the design space to locate potential superior regions, a first-type mutation strategy is designed. This strategy focuses on global exploration, using random individuals as base points and employing a long-range random perturbation factor with power-law tail characteristics to enable the trial vector to generate random displacements in the search space. This allows for rapid coverage of design regions that have not yet been fully searched. The first-type mutation strategy can be determined by the following formula: ， In the above formula, This represents the vector of the first parent individual randomly selected from the parent candidate pool. This represents the vector of the second parent individual randomly selected from the parent candidate pool. This represents the vector of the third parent individual randomly selected from the parent candidate pool. This represents the vector of the fourth parent individual randomly selected from the parent candidate pool. This represents the vector of the fifth parent individual randomly selected from the parent candidate pool. F j For the Transformer model, the first j The perturbation scaling factor of the output of each subproblem. This represents the amplitude adjustment coefficient for the Levy flight differential disturbance term. The value is 0.

01. This represents the random step size of the Levy flight generated using the Mantegna algorithm. The distribution parameters representing the random step size of Levy's flight path are... The value is 1.5, therefore, the Levy flight random step size Determined by the following formula: ， in, and Let be mutually independent random variables with first and second normal distributions, satisfying: ， in, This indicates that the mean is 0 and the variance is 0. The normal distribution Represents a random variable with the first normal distribution. standard deviation Represents a random variable with a second normal distribution. The absolute value, Represents the gamma function. Represents the sine function. π Represents pi; For subproblems where a relatively good solution has been found, and further refinement is needed in the vicinity to improve convergence accuracy, a second type of mutation strategy is designed. This strategy focuses on local refinement and takes the optimal individual vector with the smallest current aggregation function value in the neighborhood of the subproblem as the search starting point. The improved second type of mutation strategy, which moves along the direction of the random parent difference vector, can be determined by the following formula: ， In the above formula, For the first j The optimal individual vector with the smallest current aggregation function value within the neighborhood of each subproblem. Indicates the first j The mutated individual vectors of each subproblem The scaling factor is the Cauchy perturbation. The value is 0.

01. This represents a standard Cauchy distribution random number with a location parameter of 0 and a scale parameter of 1. For some subproblems where convergence has been established, a balance needs to be struck between maintaining the convergence trend and escaping local optima. A third type of mutation strategy is designed, focusing on the dynamic balance between exploration and development. Starting from the current individual vector, it includes a convergence direction component pointing to the optimal individual. A centralized random perturbation factor with exponentially decaying tails is used to controllably scale the auxiliary component. This injects search diversity in a controlled manner while maintaining the general direction of convergence towards the optimal solution, achieving gradual and robust improvement. The third type of mutation strategy can be determined by the following formula: ， In the above formula, For the current individual vector, is the Gaussian perturbation scaling factor. The value is 0.

01. These are random numbers distributed according to a standard normal distribution. Step 4.8, according to the... j The crossover probability corresponding to each sub-problem CR j , for the j The mutation vector of each sub-problem Perform a crossover operation with the parent individual vector to generate a crossover individual vector. ; for cross individual vectors Each structural design variable is checked for boundaries to ensure that the longitudinal beam thickness, longitudinal beam section height, longitudinal beam section width, transverse beam thickness, transverse beam section height, transverse beam section width, number of transverse beams, and number of longitudinal beams all fall within the corresponding value range of the multidimensional design space. Specifically, when the value of a structural design variable is less than its lower limit, the value is adjusted to its lower limit; when the value is greater than its upper limit, the value is adjusted to its upper limit; and when the value is between its lower and upper limits, the value remains unchanged. This boundary processing yields the first... j Candidate offspring vectors for each subproblem.

6. The method as described in claim 1, characterized in that, Step (5) specifically includes: Step 5.1: Use the constructed radial basis function model to predict the mass, torsional stiffness and first natural frequency of each candidate offspring individual vector, obtain the prediction target value corresponding to each candidate offspring individual vector, and perform non-dominated sorting according to the prediction target value to obtain the prediction non-dominated candidate set. Step 5.2: Calculate the positive improvement distance and negative deviation distance of each candidate offspring individual vector in the predicted non-dominated candidate set relative to the Pareto reference set in the current database. The positive improvement distance is used to measure the potential improvement of the candidate offspring individual vector in the target space relative to the Pareto reference set in the current database, and the negative deviation distance is used to measure the degree of deviation of the candidate offspring individual vector from the Pareto reference set in the current database. Step 5.3: Select an individual vector with frontier expansion capability based on the positive improvement distance, select an individual vector with convergence improvement capability based on the negative deviation distance, and remove duplicates from the selected individual vectors to obtain the individual vectors to be evaluated that need to be accurately simulated. Step 5.4: Perform mass calculation, torsional static analysis and free modal analysis on the individual vector to be evaluated to obtain the target values ​​corresponding to its mass, torsional stiffness and first natural frequency. Store the individual vector to be evaluated and its target values ​​in the database and update the current population and radial basis function model. Step 5.5: Determine whether the Pareto reference set in the current database meets the design requirements for frame mass, torsional stiffness and first natural frequency. If it does, output the Pareto optimized structure of the new energy pickup truck frame corresponding to each volume vector in the Pareto reference set. Otherwise, return to step (3) until a Pareto optimized structure that meets the design requirements is obtained.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1-6.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1-6.

9. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1-6.

Images