Neural network model-based optimization method and device, medium and program product

Abstract

The invention discloses an optimization method and device based on a neural network model, a medium and a program product, and the method comprises the steps: (1) carrying out the modeling of an automobile rear subframe based on SFE-Concept, constructing a first-order modal maximization mathematical model according to an adaptive penalty function, generating an optimization population based on Latin hypercube, and carrying out the optimization of a first-order modal maximization mathematical model; performing first-order modal and rigidity simulation analysis on the optimized population in an Isight multidisciplinary optimization design platform; (2) generating an optimal candidate sub-population and a successful design variable vector through differential evolution based on a cubic kernel radial basis function machine learning model; (3) training the Dropout neural network model to obtain evolution parameters; (4) updating evolution parameters based on the Dropout neural network model; and (5) updating and optimizing the population based on the evolution parameters, if a convergence condition is met, outputting an optimal rear subframe, otherwise, returning to the step (2). According to the method, the evolution parameters are adaptively adjusted according to the Dropout neural network model, and the adaptability to the modal optimization problem of the rear subframe of the automobile is high.

Description

Technical Field

[0001] The present invention relates to the fields of artificial intelligence and automobile manufacturing technology, and in particular to an optimization method, device, medium and program product based on a neural network model. Background Art

[0002] In automotive engineering, the rear subframe, a key load-bearing component of the chassis system, has a dynamic characteristic that directly impacts the vehicle's NVH (Noise, Vibration, and Harshness) performance. If the first-order mode of the rear subframe approaches the frequency of engine, drivetrain, or road excitation, resonance can occur, leading to increased structural fatigue, reduced ride comfort, and even safety hazards. Therefore, optimizing the rear subframe's first-order mode while meeting stiffness constraints is a key issue in improving overall vehicle performance.

[0003] Traditional approaches rely on evolutionary algorithms combined with finite element simulation models for iterative design. However, their core flaw lies in the fact that the evolutionary parameters in evolutionary algorithms are often manually preset and fixed throughout the entire design process, making them difficult to adapt to the dynamic demands of different optimization stages in a high-dimensional design space. This is particularly true given the geometric complexity of the rear subframe, which requires design variables encompassing multiple dimensions such as position, shape, and curvature. Fixed parameters lead to inefficient search during initial, wide-area exploration. Later, during localized development, parameter mismatches can easily cause the evolutionary algorithm to become trapped in a local optimum. Furthermore, the penalty function method used in traditional approaches superimposes constraints on the objective function through static weights, failing to accurately reflect the correlation between the degree of constraint violation and the optimization process. This often leads to excessive penalties or constraint relaxation, resulting in a disconnect between the objective function and engineering requirements.

[0004] Existing technologies attempt to incorporate machine learning models to aid optimization, such as using surrogate models to reduce simulation computational complexity. However, due to limited training data and noise interference, traditional neural networks are prone to overfitting, resulting in significant bias in predictions of new design variables, making it difficult to reliably guide optimization. Furthermore, these models primarily focus on target prediction and are not deeply coupled with evolutionary parameter adjustment mechanisms. The evolutionary process remains constrained by fixed rules, making it impossible to extract dynamic parameter adjustment patterns from historical data. This results in wasted computing resources and inefficient knowledge transfer. Summary of the Invention

[0005] To address the limitations of existing technologies or the need for improvement, this paper proposes a neural network-based optimization method, device, medium, and program product. This method addresses the geometric complexity of the rear subframe, the low adaptability of the evolutionary parameters of traditional evolutionary algorithms, and the design requirement of stiffness-constrained first-order modal improvement. The method researches and designs a neural network-based solution. This method adaptively adjusts the evolutionary parameters using a dropout neural network model and precisely reconciles the conflict between first-order modal improvement and stiffness constraints using an adaptive penalty function, improving the adaptability of the rear subframe modal optimization problem.

[0006] To achieve the above objectives, in a first aspect, the present invention provides an optimization method based on a neural network model, the method comprising the following steps: (1) The finite element model of the rear subframe of the automobile is established by using the SFE-Concept software and the proportional vector method is used to record the position, shape and curvature parameters of the parts to construct the design variable vector. The adaptive penalty function is used to construct the fitness function of the first-order mode with stiffness constraint as the penalty term as the objective function, and the mathematical model of the first-order mode maximization of the rear subframe of the automobile is constructed; the value range of the design variable vector is determined in combination with the actual conditions of the rear subframe structure; the design variable vector is organized into an optimization population based on Latin hypercube sampling in the multidimensional design space composed of the design variable vectors, and each design variable vector corresponds to an individual vector in the optimization population; the optimization population is simulated and analyzed by the Isight multidisciplinary optimization design platform for the first-order mode and stiffness simulation, and the first-order mode and stiffness values ​​of all individual vectors in the optimization population are obtained, and the optimization population and its corresponding first-order mode and stiffness values ​​are stored in the historical database; (2) Generate the optimal candidate subpopulation and successful design variable vector through differential evolution operation based on cubic kernel radial basis function machine learning model, and store the successful design variable vector and the corresponding evolution parameters in the success database; (3) Construct a Dropout neural network model, using the successful design variable vector in the success database as input, its corresponding evolutionary parameter as the target, using the mean square error function as the loss function, and setting the hyperparameters of the optimizer, learning rate, and training cycle to train the Dropout neural network model until convergence; (4) Adjust the evolutionary parameters of the optimized population based on the output results of the trained Dropout neural network model; (5) Based on the adjusted evolutionary parameters, the optimized population is updated. If the cubic kernel radial basis function machine learning model reaches the convergence condition, the optimal rear subframe solution is output. Otherwise, return to step (2) until the convergence condition is reached.

[0007] Furthermore, step (1) specifically includes: The first step is to build a finite element model of the rear subframe using the SFE-Concept software and use the scaled vector method to record the part position, shape, and curvature parameters to construct the design variable vector. The scaled vector method coordinate transformation formula is as follows:

[0008] In the above formula, the angle The scaling direction and scaling value of the control point FV Controls the scaling along a specified direction, is the transformed y-axis, is the z-axis after transformation; The specific recording method of the design variable vector is as follows: Record the angle of the scale vector for each control point of multiple sections and scaling values FV Parameters, these two parameters are used as part shape parameters; Record the X / Y / Z coordinates of the base point as part position parameters; Record the tangent direction or weight parameter of the baseline control point as the part curvature parameter; In the second step, the adaptive penalty function is used to construct the first-order mode fitness function with stiffness constraint as the penalty term as the objective function, and the mathematical model of the first-order mode maximization of the rear subframe of the vehicle is constructed. The specific expression is as follows:

[0009] In the above formula, x represents the design variable vector of the rear subframe of the car, Represents the part position parameters, Represents the shape parameters of the part, represents the part curvature parameter, where the number of part position parameters is p , the number of part shape parameters is qp , the number of part curvature parameters is nq , Find indicates the design variable vector that needs to be optimized, Min indicates that the optimization direction is minimized, and St indicates the constraints on the design variable vector. is the fitness function with stiffness constraint as penalty function, The variable vector for designing the rear subframe of a car is x The first-order mode when is the stiffness constraint function, Represents the multi-dimensional design space composed of the design variable vectors of the rear subframe of the automobile, For the g The adaptive penalty factor of the generation is updated using the following update formula:

[0010] In the above formula, g is the current iteration number, cp Controls the update speed of the penalty factor, the value range is 2 to 10, To control the number of iterations, the value range is 0.1 to 0.8 ,in is the maximum number of iterations; The third step is to construct a multidimensional design space consisting of part shape, position, and curvature design variable vectors. The dimension of the design space is consistent with the dimension of the design variable vector. The number of design variable vectors that need to be sampled is determined based on the dimension of the multidimensional design space and the computing resources. N , in the multidimensional design space, according to Latin hypercube sampling, Individuals are used as the optimization population, and each design variable vector corresponds to an individual vector in the optimization population; In the fourth step, the Isight multidisciplinary optimization design platform, which integrates finite element model automatic reconstruction software and finite element analysis programs, performs first-order modal simulation and stiffness simulation analysis on the optimized population to obtain the first-order modal and stiffness values ​​of all individual vectors in the optimized population. The optimized population and its corresponding first-order modal and stiffness values ​​are stored in a historical database. The steps for the first-order modal simulation and stiffness simulation analysis are as follows: The first step is to set up the finite element model automatic reconstruction software to automatically reconstruct the finite element model according to the optimized population; In the second step, the physical parameters of the material are defined and batch injected into the finite element model through script-driven injection to ensure that the optimized population can be accurately mapped to the finite element model generated by the finite element model automatic reconstruction software; The third step is to call the finite element analysis program to perform modal simulation analysis on the optimized population to obtain the first-order mode; The fourth step is to use the finite element analysis program to perform stiffness simulation analysis on the optimized population and quantify the stiffness of each point on the rear subframe of the vehicle; Step 5: Analyze and calculate the results of the stiffness analysis module to obtain the stiffness of each point on the rear subframe of the vehicle; Step 6: Regularly rename the result analysis files according to the sample number.

[0011] Furthermore, the specific steps of step (2) are as follows: The first step is to use differential evolution to generate each individual vector in the optimized population. N candidate mutation individuals, where the differential evolution operation formula is as follows:

[0012] In the above formula, Represents the first vector generated for the current individualj candidate variant individuals, It represents the first order after all individuals in the optimized population are sorted according to the feasibility rule. p % An individual vector randomly selected from the individuals, represents the current individual vector, and represents two individuals randomly selected from the optimized population; F represents the scaling factor, which controls the magnitude of individual variation; In the second step, for each individual vector in the optimized population, a binomial crossover operation is performed to generate N candidate offspring individuals; In the third step, all individual vectors in the historical database and the corresponding stiffness and first-order modal values ​​are used to construct a machine learning model based on the cubic kernel radial basis function; The fourth step is to use the cubic kernel radial basis function machine learning model to predict the corresponding vector of each individual N The first-order mode and stiffness values ​​of candidate offspring individuals are obtained, and the feasibility rule is used to select the optimal candidate offspring individual and successful design variable vector corresponding to each individual vector; Step 5: Input the optimal candidate offspring individuals into the Isight multidisciplinary optimization design platform for modal analysis and stiffness analysis to obtain the first-order modal and stiffness values ​​of each optimal candidate offspring individual, and calculate the fitness of the optimal candidate offspring individual based on the fitness function; In the sixth step, the successful design variable vector, its corresponding scaling factor and crossover probability are used as evolution parameters and stored in the success database.

[0013] Furthermore, step (3) specifically includes: The first step is to build a Dropout neural network model to learn the data in the success database, as follows: The successfully designed variable vector is first processed by the batch normalization layer for data normalization and then input into the Dropout neural network model; Then, a 32-node fully connected layer is connected to realize the dimensionality increase mapping of the feature space, and the nonlinear feature expression capability is introduced through the ReLU activation function; Configure a Dropout layer with a probability of 0.5 to enhance model generalization; Then, a 16-node fully connected layer is used for feature compression, and the ReLU activation function is used to introduce nonlinear feature expression capabilities; The fully connected layer connected to 8 nodes continues to perform feature compression and uses the ReLU activation function to introduce nonlinear feature expression capabilities; Output the evolution parameters through a 2-node fully connected layer with a Sigmoid activation function; In the second step, we select the mean square error function as the loss function and set the hyperparameters of the optimizer, learning rate, and training cycle. The third step is to use the successful design variable vector in the success database as the input of the Dropout neural network model, and use its corresponding scaling factor and crossover probability as the target to construct the dataset for Dropout neural network model learning; The fourth step is to train the Dropout neural network model with the constructed data set and check the convergence of the Dropout neural network model. If it does not converge, return to the second step to adjust the hyperparameters until the Dropout neural network model converges.

[0014] Furthermore, the specific steps of step (4) are as follows: In the first step, the individual vectors of the optimized population are input into the trained Dropout neural network model to obtain the updated scaling factor and crossover probability evolution parameters; In the second step, the individual vectors of the optimized population are matched with the corresponding updated scaling factors and crossover probability evolution parameters and then saved.

[0015] Furthermore, step (5) specifically includes: In the first step, the population is optimized based on the updated scaling factor and crossover probability evolution parameters of the Dropout neural network model; In the second step, the updated optimized population is input into the Isight multidisciplinary optimization design platform for stiffness simulation analysis and first-order modal simulation analysis to obtain the stiffness and first-order modal values ​​and calculate the fitness of each individual vector through the fitness function; The third step is to determine whether the first-order mode and stiffness of the individual vector with the smallest fitness and the design variable vector meet the design indicators of the actual working conditions of the rear subframe of the automobile. If the design indicators are met, the design variable vector is output as the optimal rear subframe solution. Otherwise, return to step (2) and enter the next iterative loop until the design indicators are met.

[0016] In a second aspect, the present invention provides a computer device comprising a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the aforementioned optimization method based on a neural network model.

[0017] In a third aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the aforementioned optimization method based on a neural network model.

[0018] In a fourth aspect, the present invention provides a computer program product, comprising a computer program, which, when executed by a processor, implements the steps of the aforementioned optimization method based on a neural network model.

[0019] In summary, the optimization method based on a neural network model provided by the present invention has the following improvements over the limitations of the prior art: 1. Considering that the evolutionary parameters of the traditional evolutionary algorithm must be manually preset and fixed throughout the entire process, it cannot adapt to the dynamic requirements of different optimization stages in the high-dimensional design space of the rear subframe. This invention uses a dropout neural network model to autonomously learn parameter mapping rules from a successful database to achieve dynamic adjustment of the evolutionary parameters.

[0020] 2. Use an adaptive penalty function to dynamically link the adaptive penalty factor to the number of iterations. Initially, the adaptive penalty factor is relaxed to explore high-potential areas, while later, it is strengthened to guide feasible solutions. This effectively alleviates the conflict between excessive penalty and constraint relaxation in multi-objective optimization.

[0021] 3. Introducing the Dropout layer improves the robustness of the Dropout neural network model and reduces interference from noisy data. Combined with the Batch Normalization layer, the Dropout neural network model significantly improves its generalization ability for the rear subframe optimization problem, ensuring that the evolutionary direction is consistent with the actual performance target.

[0022] 4. Build a closed-loop "simulation-learning-optimization" framework: Using a dropout neural network model, the algorithm continuously refines the mapping patterns between successful design variable vectors and evolutionary parameters in the success database. This dynamically updates the evolutionary strategy and dynamically adjusts the adaptive penalty factor under the guidance of an adaptive penalty function. This mechanism enables the algorithm to autonomously respond to changes in the design space, reducing the number of simulations while improving the quality of the solution set, providing a scalable paradigm for complex engineering optimization. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 The present invention provides a flowchart of an optimization method based on a neural network model. DETAILED DESCRIPTION

[0024] In order to more clearly illustrate the purpose, technical solutions and advantages of the present invention, it will be described in detail with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are only used to illustrate the present invention and are not intended to limit it. In addition, as long as the technical features in the following embodiments do not conflict with each other, they can be combined with each other.

[0025] See also Figure 1 The present invention provides an optimization method based on a neural network model, which is applicable to the modal optimization problem of the rear subframe of an automobile. Specifically, the method mainly includes steps (1) to (5).

[0026] Step (1): Establish a finite element model of the rear subframe of the automobile using the SFE-Concept software and use the proportional vector method to record the position, shape, and curvature parameters of the parts to construct the design variable vector. Use the adaptive penalty function to construct the fitness function of the first-order mode with stiffness constraint as the penalty term as the objective function, and build a mathematical model for maximizing the first-order mode of the rear subframe of the automobile; determine the value range of the design variable vector based on the actual conditions of the rear subframe structure; organize the design variable vector into an optimization population based on Latin hypercube sampling in the multidimensional design space composed of the design variable vectors, and each design variable vector corresponds to an individual vector in the optimization population; use the Isight multidisciplinary optimization design platform to perform first-order modal simulation and stiffness simulation analysis on the optimization population, and obtain the first-order mode and stiffness values ​​of all individual vectors in the optimization population. The optimization population and its corresponding first-order mode and stiffness values ​​are stored in the historical database.

[0027] Wherein, step (1) specifically includes: The first step is to build a finite element model of the rear subframe using the SFE-Concept software and use the scaled vector method to record the part position, shape, and curvature parameters to construct the design variable vector. The scaled vector method coordinate transformation formula is as follows:

[0028] In the above formula, the angle The scaling direction and scaling value of the control point FV Controls the scaling along a specified direction, is the transformed y-axis, is the z-axis after transformation; The specific recording method of the design variable vector is as follows: Record the angle of the scale vector for each control point of multiple sections and scaling values FV Parameters, these two parameters are used as part shape parameters; Record the X / Y / Z coordinates of the base point as part position parameters; Record the tangent direction or weight parameter of the baseline control point as the part curvature parameter; In the second step, the adaptive penalty function is used to construct the first-order mode fitness function with stiffness constraint as the penalty term as the objective function, and the mathematical model of the first-order mode maximization of the rear subframe of the vehicle is constructed. The specific expression is as follows:

[0029] In the above formula, x represents the design variable vector of the rear subframe of the car, Represents the part position parameters, Represents the shape parameters of the part, represents the part curvature parameter, where the number of part position parameters is p , the number of part shape parameters is qp , the number of part curvature parameters is nq , Find indicates the design variable vector that needs to be optimized, Min indicates that the optimization direction is minimized, and St indicates the constraints on the design variable vector. is the fitness function with stiffness constraint as penalty function, The variable vector for designing the rear subframe of a car is x The first-order mode when is the stiffness constraint function, Represents the multi-dimensional design space composed of the design variable vectors of the rear subframe of the vehicle, For the g The adaptive penalty factor of the generation is updated using the following update formula:

[0030] In the above formula, g is the current iteration number, cp Controls the update speed of the penalty factor, the value range is 2 to 10, To control the number of iterations, the value range is 0.1 to 0.8 ,in is the maximum number of iterations; The third step is to construct a multidimensional design space consisting of part shape, position, and curvature design variable vectors. The dimension of the design space is consistent with the dimension of the design variable vector. The number of design variable vectors that need to be sampled is determined based on the dimension of the multidimensional design space and the computing resources. N , in the multidimensional design space, according to Latin hypercube sampling, Individuals are used as the optimization population, and each design variable vector corresponds to an individual vector in the optimization population; In the fourth step, the Isight multidisciplinary optimization design platform, which integrates finite element model automatic reconstruction software and finite element analysis programs, performs first-order modal simulation and stiffness simulation analysis on the optimized population to obtain the first-order modal and stiffness values ​​of all individual vectors in the optimized population. The optimized population and its corresponding first-order modal and stiffness values ​​are stored in a historical database. The steps for the first-order modal simulation and stiffness simulation analysis are as follows: The first step is to set up the finite element model automatic reconstruction software to automatically reconstruct the finite element model according to the optimized population; In the second step, the physical parameters of the material are defined and batch injected into the finite element model through script-driven injection to ensure that the optimized population can be accurately mapped to the finite element model generated by the finite element model automatic reconstruction software; The third step is to call the finite element analysis program to perform modal simulation analysis on the optimized population to obtain the first-order mode; The fourth step is to use the finite element analysis program to perform stiffness simulation analysis on the optimized population and quantify the stiffness of each point on the rear subframe of the vehicle; Step 5: Analyze and calculate the results of the stiffness analysis module to obtain the stiffness of each point on the rear subframe of the vehicle; Step 6: Regularly rename the result analysis files according to the sample number.

[0031] Step (2): Generate the optimal candidate subpopulation and successful design variable vector through differential evolution operation based on cubic kernel radial basis function machine learning model, and store the successful design variable vector and the corresponding evolution parameters in the success database.

[0032] The specific steps of step (2) are as follows: The first step is to use differential evolution to generate each individual vector in the optimized population. N candidate mutation individuals, where the differential evolution operation formula is as follows:

[0033] In the above formula, Represents the first vector generated for the current individual j candidate variant individuals, It represents the first order after all individuals in the optimized population are sorted according to the feasibility rule. p % An individual vector randomly selected from the individuals, represents the current individual vector, and represents two individuals randomly selected from the optimized population; F represents the scaling factor, which controls the magnitude of individual variation; In the second step, for each individual vector in the optimized population, a binomial crossover operation is performed to generate N candidate offspring individuals; In the third step, all individual vectors in the historical database and the corresponding stiffness and first-order modal values ​​are used to construct a machine learning model based on the cubic kernel radial basis function; The fourth step is to use the cubic kernel radial basis function machine learning model to predict the corresponding vector of each individual N The first-order mode and stiffness values ​​of candidate offspring individuals are obtained, and the feasibility rule is used to select the optimal candidate offspring individual and successful design variable vector corresponding to each individual vector; Step 5: Input the optimal candidate offspring individuals into the Isight multidisciplinary optimization design platform for modal analysis and stiffness analysis to obtain the first-order modal and stiffness values ​​of each optimal candidate offspring individual, and calculate the fitness of the optimal candidate offspring individual based on the fitness function; In the sixth step, the successful design variable vector, its corresponding scaling factor and crossover probability are used as evolution parameters and stored in the success database.

[0034] Step (3): Construct a Dropout neural network model, take the successful design variable vector in the success database as input, use its corresponding evolution parameter as target, use the mean square error function as the loss function, set the hyperparameters of the optimizer, learning rate, and training cycle to train the Dropout neural network model until convergence.

[0035] Among them, step (3) specifically includes: The first step is to build a Dropout neural network model to learn the data in the success database, as follows: The successfully designed variable vector is first processed by the batch normalization layer for data normalization and then input into the Dropout neural network model; Then, a 32-node fully connected layer is connected to realize the dimensionality increase mapping of the feature space, and the nonlinear feature expression capability is introduced through the ReLU activation function; Configure a Dropout layer with a probability of 0.5 to enhance model generalization; Then, a 16-node fully connected layer is used for feature compression, and the ReLU activation function is used to introduce nonlinear feature expression capabilities; The fully connected layer connected to 8 nodes continues to perform feature compression and uses the ReLU activation function to introduce nonlinear feature expression capabilities; Output the evolution parameters through a 2-node fully connected layer with a Sigmoid activation function; In the second step, we select the mean square error function as the loss function and set the hyperparameters of the optimizer, learning rate, and training cycle. The third step is to use the successful design variable vector in the success database as the input of the Dropout neural network model, and use its corresponding scaling factor and crossover probability as the target to construct the dataset for Dropout neural network model learning; The fourth step is to train the Dropout neural network model with the constructed data set and check the convergence of the Dropout neural network model. If it does not converge, return to the second step to adjust the hyperparameters until the Dropout neural network model converges.

[0036] Step (4): Adjust the evolutionary parameters of the optimized population based on the output results of the trained Dropout neural network model.

[0037] The specific steps of step (4) are as follows: In the first step, the individual vectors of the optimized population are input into the trained Dropout neural network model to obtain the updated scaling factor and crossover probability evolution parameters; In the second step, the individual vectors of the optimized population are matched with the corresponding updated scaling factors and crossover probability evolution parameters and then saved.

[0038] Step (5): Update the optimized population based on the adjusted evolutionary parameters. If the cubic kernel radial basis function machine learning model reaches the convergence condition, the optimal rear subframe solution is output. Otherwise, return to step (2) until the convergence condition is reached.

[0039] Wherein, step (5) specifically includes: In the first step, the population is optimized based on the updated scaling factor and crossover probability evolution parameters of the Dropout neural network model; In the second step, the updated optimized population is input into the Isight multidisciplinary optimization design platform for stiffness simulation analysis and first-order modal simulation analysis to obtain the stiffness and first-order modal values ​​and calculate the fitness of each individual vector through the fitness function; The third step is to determine whether the first-order mode and stiffness of the individual vector with the smallest fitness and the design variable vector meet the design indicators of the actual working conditions of the rear subframe of the automobile. If the design indicators are met, the design variable vector is output as the optimal rear subframe solution. Otherwise, return to step (2) and enter the next iterative loop until the design indicators are met.

[0040] This embodiment uses the benchmark function in the CEC2010 test suite to illustrate the optimization performance of the optimization method based on the neural network model provided by this embodiment. The expression of the benchmark function in the CEC2010 test suite is as follows: , In this embodiment is the objective function, is the constraint function, Indicates the i design variables, It is i The offset value of the design variable, The values ​​of are shown in Table 1 below.

[0041] Table 1 Value table of

[0042] To further illustrate this embodiment, a neural network model-based optimization method in this embodiment was compared with a classical differential evolution algorithm. The classical differential evolution algorithm employed a feasibility rule as its constraint processing mechanism. The maximum number of simulation evaluations in this embodiment was set to 1000, the number of design variable vectors was set to 40, the Adam optimizer was used, the learning rate was 0.001, and the training period was 40. The experimental results are shown in Table 2. The comparison was performed using the average, standard deviation, and minimum values ​​of 20 independent runs. Given the same number of simulations, the method in this embodiment significantly outperformed the classical differential evolution algorithm, demonstrating its ability to effectively solve the automotive rear subframe modal optimization problem.

[0043] Table 2 Comparison of optimization results of different methods

[0044] The present invention provides an optimization method based on a neural network model. The method adaptively adjusts the evolution parameters through the Dropout neural network model and accurately coordinates the conflict between the first-order modal improvement and the stiffness constraint through an adaptive penalty function, providing a systematic solution to the modal optimization problem of the automobile rear subframe.

[0045] In a second aspect, the present invention provides a computer device comprising a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the aforementioned optimization method based on a neural network model.

[0046] In a third aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the aforementioned optimization method based on a neural network model.

[0047] In a fourth aspect, the present invention provides a computer program product, comprising a computer program, which, when executed by a processor, implements the aforementioned optimization method steps based on a neural network model.

[0048] It will be easily understood by those skilled in the art that the above are only 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 in the scope of protection of the present invention.

Claims

1. An optimization method based on a neural network model, characterized in that: The method comprises: (1) The finite element model of the rear subframe of the automobile is established by using the SFE-Concept software and the proportional vector method is used to record the position, shape and curvature parameters of the parts to construct the design variable vector. The adaptive penalty function is used to construct the fitness function of the first-order mode with stiffness constraint as the penalty term as the objective function, and the mathematical model of the first-order mode maximization of the rear subframe of the automobile is constructed; the value range of the design variable vector is determined in combination with the actual conditions of the rear subframe structure; the design variable vector is organized into an optimization population based on Latin hypercube sampling in the multidimensional design space composed of the design variable vectors, and each design variable vector corresponds to an individual vector in the optimization population; the optimization population is simulated and analyzed by the Isight multidisciplinary optimization design platform for the first-order mode and stiffness simulation, and the first-order mode and stiffness values ​​of all individual vectors in the optimization population are obtained, and the optimization population and its corresponding first-order mode and stiffness values ​​are stored in the historical database; (2) Generate the optimal candidate subpopulation and successful design variable vector through differential evolution operation based on cubic kernel radial basis function machine learning model, and store the successful design variable vector and the corresponding evolution parameters in the success database; (3) Construct a Dropout neural network model, using the successful design variable vector in the success database as input, its corresponding evolutionary parameter as the target, using the mean square error function as the loss function, and setting the hyperparameters of the optimizer, learning rate, and training cycle to train the Dropout neural network model until convergence; (4) Adjust the evolutionary parameters of the optimized population based on the output results of the trained Dropout neural network model; (5) Based on the adjusted evolutionary parameters, the optimized population is updated. If the cubic kernel radial basis function machine learning model reaches the convergence condition, the optimal rear subframe solution is output. Otherwise, return to step (2) until the convergence condition is reached.

2. The method according to claim 1, wherein Step (1) specifically includes: The first step is to build a finite element model of the rear subframe using the SFE-Concept software and use the scaled vector method to record the part position, shape, and curvature parameters to construct the design variable vector. The scaled vector method coordinate transformation formula is as follows: In the above formula, the angle The scaling direction and scaling value of the control point FV Controls the scaling along a specified direction, is the transformed y-axis, is the z-axis after transformation; The specific recording method of the design variable vector is as follows: Record the angle of the scale vector for each control point of multiple sections and scaling values FV Parameters, these two parameters are used as part shape parameters; Record the X / Y / Z coordinates of the base point as part position parameters; Record the tangent direction or weight parameter of the baseline control point as the part curvature parameter; In the second step, the adaptive penalty function is used to construct the first-order mode fitness function with stiffness constraint as the penalty term as the objective function, and the mathematical model of the first-order mode maximization of the rear subframe of the vehicle is constructed. The specific expression is as follows: In the above formula, x represents the design variable vector of the rear subframe of the car, Represents the part position parameters, Represents the shape parameters of the part, represents the part curvature parameter, where the number of part position parameters is p , the number of part shape parameters is qp , the number of part curvature parameters is nq , Find indicates the design variable vector that needs to be optimized, Min indicates that the optimization direction is minimized, and St indicates the constraints on the design variable vector. is the fitness function with stiffness constraint as penalty function, The variable vector for designing the rear subframe of a car is x The first-order mode when is the stiffness constraint function, Represents the multi-dimensional design space composed of the design variable vectors of the rear subframe of the vehicle, For the g The adaptive penalty factor of the generation is updated using the following update formula: In the above formula, g is the current iteration number, cp Controls the update speed of the penalty factor, the value range is 2 to 10, To control the number of iterations, the value range is 0.1 to 0.8 ,in is the maximum number of iterations; The third step is to construct a multidimensional design space consisting of part shape, position, and curvature design variable vectors. The dimension of the design space is consistent with the dimension of the design variable vector. The number of design variable vectors that need to be sampled is determined based on the dimension of the multidimensional design space and the computing resources. N , in the multidimensional design space, according to Latin hypercube sampling, Individuals are used as the optimization population, and each design variable vector corresponds to an individual vector in the optimization population; In the fourth step, the Isight multidisciplinary optimization design platform, which integrates finite element model automatic reconstruction software and finite element analysis programs, performs first-order modal simulation and stiffness simulation analysis on the optimized population to obtain the first-order modal and stiffness values ​​of all individual vectors in the optimized population. The optimized population and its corresponding first-order modal and stiffness values ​​are stored in a historical database. The steps for the first-order modal simulation and stiffness simulation analysis are as follows: The first step is to set up the finite element model automatic reconstruction software to automatically reconstruct the finite element model according to the optimized population; In the second step, the physical parameters of the material are defined and batch injected into the finite element model through script-driven injection to ensure that the optimized population can be accurately mapped to the finite element model generated by the finite element model automatic reconstruction software; The third step is to call the finite element analysis program to perform modal simulation analysis on the optimized population to obtain the first-order mode; The fourth step is to use the finite element analysis program to perform stiffness simulation analysis on the optimized population and quantify the stiffness of each point on the rear subframe of the vehicle; Step 5: Analyze and calculate the results of the stiffness analysis module to obtain the stiffness of each point on the rear subframe of the vehicle; Step 6: Regularly rename the result analysis files according to the sample number.

3. The method according to claim 1, wherein Step (2): The specific steps are as follows: The first step is to use differential evolution to generate each individual vector in the optimized population. N candidate mutation individuals, where the differential evolution operation formula is as follows: In the above formula, Represents the first vector generated for the current individual j candidate variant individuals, It represents the first order after all individuals in the optimized population are sorted according to the feasibility rule. p % An individual vector randomly selected from the individuals, represents the current individual vector, and represents two individuals randomly selected from the optimized population; F represents the scaling factor, which controls the magnitude of individual variation; In the second step, for each individual vector in the optimized population, a binomial crossover operation is performed to generate N candidate offspring individuals; In the third step, all individual vectors in the historical database and the corresponding stiffness and first-order modal values ​​are used to construct a machine learning model based on the cubic kernel radial basis function; The fourth step is to use the cubic kernel radial basis function machine learning model to predict the corresponding vector of each individual N The first-order mode and stiffness values ​​of candidate offspring individuals are obtained, and the feasibility rule is used to select the optimal candidate offspring individual and successful design variable vector corresponding to each individual vector; Step 5: Input the optimal candidate offspring individuals into the Isight multidisciplinary optimization design platform for modal analysis and stiffness analysis to obtain the first-order modal and stiffness values ​​of each optimal candidate offspring individual, and calculate the fitness of the optimal candidate offspring individual based on the fitness function; In the sixth step, the successful design variable vector, its corresponding scaling factor and crossover probability are used as evolution parameters and stored in the success database.

4. The method according to claim 1, wherein Step (3) specifically includes: The first step is to build a Dropout neural network model to learn the data in the success database, as follows: The successfully designed variable vector is first processed by the batch normalization layer for data normalization and then input into the Dropout neural network model; Then, a 32-node fully connected layer is connected to realize the dimensionality increase mapping of the feature space, and the nonlinear feature expression capability is introduced through the ReLU activation function; Configure a Dropout layer with a probability of 0.5 to enhance model generalization; Then, a 16-node fully connected layer is used for feature compression, and the ReLU activation function is used to introduce nonlinear feature expression capabilities; The fully connected layer connected to 8 nodes continues to perform feature compression and uses the ReLU activation function to introduce nonlinear feature expression capabilities; Output the evolution parameters through a 2-node fully connected layer with a Sigmoid activation function; In the second step, we select the mean square error function as the loss function and set the hyperparameters of the optimizer, learning rate, and training cycle. The third step is to use the successful design variable vector in the success database as the input of the Dropout neural network model, and use its corresponding scaling factor and crossover probability as the target to construct the dataset for Dropout neural network model learning; The fourth step is to train the Dropout neural network model with the constructed data set and check the convergence of the Dropout neural network model. If it does not converge, return to the second step to adjust the hyperparameters until the Dropout neural network model converges.

5. The method according to claim 1, wherein Step (4): The specific steps are as follows: In the first step, the individual vectors of the optimized population are input into the trained Dropout neural network model to obtain the updated scaling factor and crossover probability evolution parameters; In the second step, the individual vectors of the optimized population are matched with the corresponding updated scaling factors and crossover probability evolution parameters and then saved.

6. The method according to claim 5, wherein Step (5) specifically includes: In the first step, the population is optimized based on the updated scaling factor and crossover probability evolution parameters of the Dropout neural network model; In the second step, the updated optimized population is input into the Isight multidisciplinary optimization design platform for stiffness simulation analysis and first-order modal simulation analysis to obtain the stiffness and first-order modal values ​​and calculate the fitness of each individual vector through the fitness function; The third step is to determine whether the first-order mode and stiffness of the individual vector with the smallest fitness and the design variable vector meet the design indicators of the actual working conditions of the rear subframe of the automobile. If the design indicators are met, the design variable vector is output as the optimal rear subframe solution. Otherwise, return to step (2) and enter the next iterative loop until the design indicators are met.

7. A computer device comprising a memory, a processor, and a computer program stored in the memory, characterized in that: The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 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, the steps of the method according to any one of claims 1 to 6 are implemented.

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

Images