Optimization method for large insulating mandrel pultrusion process based on improved multi-objective algorithm

Abstract

The application discloses a large insulating core rod pultrusion process optimization method based on an improved multi-objective algorithm, comprising the following steps: constructing a thermal-flow-solid multi-physical field coupling simulation model of the large insulating core rod pultrusion process, and generating a process-performance dataset by using a hybrid adaptive sampling strategy; modeling by using a deep Gaussian process, and training to obtain a process-performance proxy model; introducing a dynamic feasible region constraint processing mechanism, and establishing a multi-objective optimization problem; solving by using a reference vector guided multi-objective optimization algorithm, and obtaining a front solution set; screening an optimal process parameter combination; constructing a digital twin online compensation mechanism, and performing migration fine tuning and rolling optimization. The application improves the global optimization precision of the large insulating core rod pultrusion process parameters and the dynamic anti-interference ability of the production process by using a whole-process process optimization method which integrates thermal-flow-solid multi-physical field coupling simulation, hybrid adaptive sampling, improved multi-objective optimization and digital twin online compensation.

Description

Technical Field

[0001] This application relates to the field of pultrusion process optimization technology for insulating core rods, and in particular to a method for optimizing the pultrusion process of large insulating core rods based on an improved multi-objective algorithm. Background Technology

[0002] Large insulating core rods are the core load-bearing and insulating components of critical power equipment in fields such as ultra-high voltage power transmission, rail transit, and new energy equipment. Their performance directly determines the operational reliability and safety of the power system. With the development of power equipment towards larger capacity, higher voltage, and longer lifespan, more stringent requirements are being placed on the diameter specifications, insulation strength, mechanical properties, and production efficiency of insulating core rods. Pultrusion molding, as the mainstream manufacturing process for insulating core rods, involves complex multi-physics coupling processes such as resin flow, heat transfer, curing reaction, and structural deformation. There are strong nonlinear correlations and mutual constraints among the process parameters. Traditional empirical trial-and-error methods are insufficient for achieving synergistic optimization of multiple performance indicators, easily leading to defects such as uneven curing inside the core rod, excessive residual stress, and decreased insulation performance, seriously affecting product quality and service life. Therefore, research is needed on the optimization technology of the pultrusion process for large insulating core rods.

[0003] In the prior art, Chinese patent CN114619688A discloses a method for preparing a preheated one-time pultrusion molding insulating core rod. The method includes: Step 1: preparing outer ring impregnation, middle ring impregnation, and inner ring impregnation, each comprising epoxy resin, curing agent, and curing accelerator; Step 2: separating glass fibers into three strands, impregnating them with the outer ring, middle ring, and inner ring respectively; Step 3: preheating the glass fibers impregnated with the inner ring using a preheater; Step 4: passing the three impregnated glass fibers through a preforming device, with the glass fibers impregnated with the outer ring located in the outermost layer, the glass fibers impregnated with the inner ring located in the center, and the glass fibers impregnated with the middle ring located in the middle layer; Step 5: placing the rod in a heating mold for continuous curing and molding.

[0004] However, the aforementioned existing technologies mainly rely on manual experience to determine process parameters, and have not established a system multi-parameter collaborative optimization mechanism covering multiple performance indicators. They also lack real-time status feedback and dynamic parameter adjustment capabilities during the production process. The global optimization accuracy of large insulating core rod pultrusion process parameters and the dynamic anti-interference capability of the production process need to be improved. Summary of the Invention

[0005] This application provides an optimization method for the pultrusion process of large insulating core rods based on an improved multi-objective algorithm, which addresses the problem that the global optimization accuracy of the pultrusion process parameters for large insulating core rods and the dynamic anti-interference capability of the production process need to be improved in the prior art.

[0006] On the one hand, this application provides an optimization method for the pultrusion process of large insulating core rods based on an improved multi-objective algorithm, including the following steps: Step 1: Construct a thermal-fluid-solid multiphysics coupled simulation model of the pultrusion process of large insulating core rods, and generate process-performance datasets by adopting a hybrid adaptive sampling strategy based on orthogonal experiment and prediction variance maximization.

[0007] Step 2: The process-performance dataset is modeled using a mixed kernel function sparse multi-output deep Gaussian process, and a process-performance surrogate model is obtained through variational sparse inference training.

[0008] Step 3: Based on the process-performance proxy model, with insulation strength, curing uniformity, residual stress and production efficiency as optimization objectives, a dynamic feasible domain constraint processing mechanism is introduced to establish a multi-objective optimization problem.

[0009] Step four: A multi-objective optimization algorithm guided by a reference vector with adaptive evolution of the covariance matrix is ​​used to solve the multi-objective optimization problem. An adaptive reference vector adjustment strategy based on the change of front curvature is used to obtain a uniformly distributed Pareto front solution set.

[0010] Step 5: Select the optimal combination of process parameters from the Pareto front solution set using fuzzy grey relational analysis based on dynamic preference weights.

[0011] Step six: Construct a digital twin online compensation mechanism, perform migration fine-tuning of the process-performance proxy model based on actual pultrusion feedback data, and perform rolling optimization of the optimal process parameter combination.

[0012] In one possible implementation, in step one, the thermo-fluid-solid multiphysics coupled simulation model includes the resin flow control equation, the temperature-curing degree field coupled control equation, and the stress-strain constitutive equation, and the temperature distribution, curing degree distribution, and thermal residual stress distribution during the pultrusion process are solved using the finite element method.

[0013] In one possible implementation, step one, the hybrid adaptive sampling strategy based on orthogonal experimentation and maximizing prediction variance includes: First, an initial process-performance sample point set is generated by orthogonal experimental design. Then, the prediction variance of the preliminary Gaussian process surrogate model trained with the initial process-performance sample point set is maximized as the addition criterion. Sample points are adaptively added in the process design space until the sample point coverage meets the preset convergence condition, thus obtaining the process-performance dataset.

[0014] In one possible implementation, in step two, the hybrid kernel function sparse multi-output deep Gaussian process is composed of several stacked Gaussian process layers, each Gaussian process layer adopting a hybrid kernel function and an induced point sparse approximation structure; the hybrid kernel function is formed by combining radial basis kernels, linear kernels and periodic kernels through adaptive weights.

[0015] In one possible implementation, in step two, the variational sparse inference training aims to minimize the variational lower bound of the multi-output marginal likelihood. It utilizes reparameterization techniques and stochastic gradient optimization to jointly learn the mixed kernel function weights, induced point locations, and variational parameters of all Gaussian process layers, thereby obtaining the process-performance surrogate model.

[0016] In one possible implementation, step three, establishing the multi-objective optimization problem, includes: The four optimization objectives are to maximize insulation strength, minimize curing non-uniformity, minimize residual stress, and maximize production efficiency. At the same time, a dynamic feasible domain constraint boundary is introduced that is related to the temperature field distribution. The constraint limit of the dynamic feasible domain constraint boundary is adjusted in real time according to the maximum temperature difference inside the mandrel output by the process-performance proxy model.

[0017] In one possible implementation, in step four, the reference vector-guided multi-objective optimization algorithm uses a covariance matrix adaptive evolution strategy to generate offspring populations during the evolution process, uses angle-penalized distance indicators for environmental selection, and adds or deletes reference vectors in the irregular front region of the target space through an adaptive reference vector adjustment strategy based on changes in front curvature.

[0018] In one possible implementation, step five of the fuzzy grey relational analysis method based on dynamic preference weights includes: A standardized decision matrix is ​​constructed for the Pareto front solution set. Then, the grey relational coefficient of each Pareto solution to the ideal solution is calculated using the preference weights that change dynamically according to production demand. The comprehensive grey relational degree is obtained by weighting the fuzzy membership degree. The solution corresponding to the maximum comprehensive grey relational degree is the optimal combination of process parameters.

[0019] In one possible implementation, in step six, the migration fine-tuning involves constructing a loss function that includes parameter regularization terms and using actual pultrusion feedback data to fine-tune some layer parameters of the process-performance proxy model, while retaining the predictive power based on historical data.

[0020] The rolling optimization re-triggers the multi-objective optimization problem in each feedback cycle with a fine-tuned surrogate model, updating the optimal combination of process parameters.

[0021] The optimization method for the pultrusion process of large insulating core rods based on an improved multi-objective algorithm in this application has the following advantages: By integrating thermal-fluid-solid multiphysics coupling simulation, hybrid adaptive sampling, improved multi-objective optimization, and digital twin online compensation into a full-process process optimization method, the global optimization accuracy of large insulating core rod pultrusion process parameters and the dynamic anti-interference capability of the production process are improved.

[0022] By employing a hybrid sampling strategy—first generating initial representative samples through orthogonal experiments, and then adaptively adding new samples based on the criterion of maximizing the prediction variance—the comprehensiveness and sampling efficiency of the process-performance dataset were improved, while the generation of redundant samples was reduced.

[0023] By employing a hybrid kernel function consisting of an adaptive weighted combination of radial basis kernel, linear kernel, and periodic kernel, and a multi-output deep Gaussian process structure with induced point sparse approximation, the fitting accuracy of the surrogate model for complex nonlinear coupling relationships of multiple performance indicators is improved.

[0024] By combining reparameterization techniques and stochastic gradient optimization with a variational sparse inference training method that aims to minimize the variational lower bound of the multi-output marginal likelihood, the training efficiency and multi-output prediction accuracy of deep Gaussian process surrogate models are improved.

[0025] By taking insulation strength, curing uniformity, and other factors as multiple objectives and introducing a dynamic feasible domain constraint that is adjusted in real time according to the maximum temperature difference inside the mandrel, the engineering feasibility of the multi-objective optimization solution is improved, and curing defects caused by process parameters exceeding the safety boundary are avoided.

[0026] By employing a multi-objective optimization algorithm that uses covariance matrix adaptive evolution to generate offspring, angle-penalized distance environment selection, and adaptive reference vector adjustment based on leading-edge curvature, the convergence speed of the algorithm and the uniformity of the Pareto solution set distribution in the target space are improved.

[0027] By constructing a standardized decision matrix and combining the fuzzy grey relational analysis method that calculates the grey relational coefficient and fuzzy membership degree weighted by dynamic preference weights, the flexibility of selecting optimal process parameters and their adaptability to actual production needs are improved.

[0028] By employing a loss function with parameterized regularization for migration fine-tuning and a rolling optimization mechanism that re-solves the problem every cycle based on feedback data, the adaptability of the surrogate model to actual production fluctuations and the online optimization accuracy of process parameters are improved. Attached Figure Description

[0029] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0030] Figure 1 This is a flowchart illustrating the optimization method for the pultrusion process of large insulating core rods based on an improved multi-objective algorithm, provided in an embodiment of this application. Detailed Implementation

[0031] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0032] like Figure 1 As shown in the embodiments of this application, an optimization method for the pultrusion process of large insulating core rods based on an improved multi-objective algorithm is provided, including the following steps: Step 1: Construct a thermal-fluid-solid multiphysics coupled simulation model of the pultrusion process of large insulating core rods, and generate process-performance datasets by adopting a hybrid adaptive sampling strategy based on orthogonal experiment and prediction variance maximization.

[0033] Step 2: The process-performance dataset is modeled using a mixed kernel function sparse multi-output deep Gaussian process, and a process-performance surrogate model is obtained through variational sparse inference training.

[0034] Step 3: Based on the process-performance proxy model, with insulation strength, curing uniformity, residual stress and production efficiency as optimization objectives, a dynamic feasible domain constraint processing mechanism is introduced to establish a multi-objective optimization problem.

[0035] Step four: A multi-objective optimization algorithm guided by a reference vector with adaptive evolution of the covariance matrix is ​​used to solve the multi-objective optimization problem. An adaptive reference vector adjustment strategy based on the change of front curvature is used to obtain a uniformly distributed Pareto front solution set.

[0036] Step 5: Select the optimal combination of process parameters from the Pareto front solution set using fuzzy grey relational analysis based on dynamic preference weights.

[0037] Step six: Construct a digital twin online compensation mechanism, perform migration fine-tuning of the process-performance proxy model based on actual pultrusion feedback data, and perform rolling optimization of the optimal process parameter combination.

[0038] For example, in step one, the thermal-fluid-solid multiphysics coupling simulation model includes the resin flow control equation, the temperature-curing degree field coupling control equation, and the stress-strain constitutive equation, and uses the finite element method to solve for the temperature distribution, curing degree distribution, and thermal residual stress distribution during the pultrusion process.

[0039] Specifically, in this embodiment, in step one, the resin flow control equation is an incompressible Navier-Stokes equation combined with a Darcy flow term; the temperature-curing degree field coupling control equation is a transient heat conduction equation including a reaction exothermic source term and a curing kinetic equation; and the stress-strain constitutive equation is a linear thermoelastic constitutive relation. The temperature, curing degree, and residual stress at any position of the mandrel during the entire pultrusion process are obtained using finite element software through sequential coupling iterative solutions.

[0040] For example, in step one, the hybrid adaptive sampling strategy based on orthogonal experimentation and prediction variance maximization includes: First, an initial process-performance sample point set is generated by orthogonal experimental design. Then, the prediction variance of the preliminary Gaussian process surrogate model trained with the initial process-performance sample point set is maximized as the addition criterion. Sample points are adaptively added in the process design space until the sample point coverage meets the preset convergence condition, thus obtaining the process-performance dataset.

[0041] Specifically, in this embodiment, the preliminary Gaussian process surrogate model uses a squared exponential kernel to predict the variance. Represented as ,in This is a vector of process parameter combinations. For kernel function, Let be the column vector of covariance between test points and training sample points. The training sample covariance matrix. Adaptively added sample points are... When the number of newly added sample points reaches a preset threshold Or the maximum prediction variance is below the threshold When sampling stops, the accumulated input-output pairs constitute the process-performance dataset.

[0042] For example, in step two, the hybrid kernel function sparse multi-output deep Gaussian process is composed of several stacked Gaussian process layers, and each Gaussian process layer adopts a hybrid kernel function and an induced point sparse approximation structure; the hybrid kernel function is formed by combining radial basis kernels, linear kernels and periodic kernels through adaptive weights.

[0043] Specifically, in this embodiment, in step two, the hybrid kernel function... for: .

[0044] in, , , For adaptive weights, the sum is 1; Radial base core; It is a linear kernel; It is a periodic kernel.

[0045] A deep Gaussian process has a total of L Layer, number l The output of the layer is The interlayer transformation is as follows: .

[0046] Among them, when hour, This is the input process parameter matrix. Each Gaussian process layer uses... M One inducement point Z l and the corresponding induced output Perform a sparse approximation.

[0047] For example, in step two, the variational sparse inference training aims to minimize the variational lower bound of the multi-output marginal likelihood. It uses reparameterization techniques and stochastic gradient optimization to jointly learn the mixed kernel function weights, induced point positions, and variational parameters of all Gaussian process layers to obtain the process-performance surrogate model.

[0048] Specifically, in this embodiment, the variational lower bound of variational sparse inference training is... for: .

[0049] in, The number of output performance indicators (insulation strength, curing unevenness, residual stress, production efficiency). For the first The true value of each performance metric For the first Layer The potential function value of each output. The joint variational distribution of the latent functions of all layers. Indicates variational distribution The expectations below For the first Variational Gaussian distribution of layer-induced output, As a prior distribution, This represents the Kullback-Leibler divergence between the variational Gaussian distribution and the prior distribution. By maximizing... By jointly optimizing all hybrid kernel weights, length scales, induced point positions, and variational parameters, a process-performance surrogate model is obtained after training.

[0050] For example, in step three, establishing the multi-objective optimization problem includes: The four optimization objectives are to maximize insulation strength, minimize curing non-uniformity, minimize residual stress, and maximize production efficiency. At the same time, a dynamic feasible domain constraint boundary is introduced that is related to the temperature field distribution. The constraint limit of the dynamic feasible domain constraint boundary is adjusted in real time according to the maximum temperature difference inside the mandrel output by the process-performance proxy model.

[0051] Specifically, in this embodiment, in step three, the multi-objective optimization problem is expressed as: , .

[0052] in, For process parameter vectors, For mold temperature, For pultrusion speed, This refers to the post-curing temperature; For insulation strength, To address the issue of uneven curing, For the maximum residual stress, For production efficiency; For the first A standard constraint. For dynamic constraint boundary functions, The maximum temperature difference inside the mandrel, predicted by the surrogate model. The dynamic constraint boundary function is set to... ,in For the first The coefficient of the first-order term of temperature difference under a conventional constraint, For the first The temperature difference constant term of the conventional constraint allows the constraint boundary to adaptively relax or tighten as the temperature uniformity changes, avoiding excessive constraint in infeasible process regions.

[0053] For example, in step four, the reference vector-guided multi-objective optimization algorithm uses a covariance matrix adaptive evolution strategy to generate offspring populations during the evolution process, uses angle penalty distance index for environmental selection, and adds or deletes reference vectors in the irregular front region of the target space through an adaptive reference vector adjustment strategy based on front curvature changes.

[0054] Specifically, in this embodiment, in step four, the covariance matrix adaptive evolution strategy generates offspring in the following way: parent individuals through Generate offspring ,in For global step size, and These are the covariance matrices. The orthogonal and diagonal matrices obtained from eigenvalue decomposition .

[0055] Environmental selection uses angle penalty distance index ,in, For individuals Angle penalty distance value, For individuals The Euclidean distance from the target vector to the ideal point. For individuals The angle between the target vector and the associated reference vector, For follow A monotonically increasing penalty function.

[0056] The adaptive reference vector adjustment strategy based on front curvature changes is as follows: calculate the local curvature of the current non-dominated front; when the local curvature exceeds a preset upper bound, insert a new reference vector in the corresponding region. ,in, For the new reference vector, The objective value for the non-dominated solution with the maximum congestion distance in this region is... As the ideal point of the current non-dominant frontier, The L2 norm is represented. When the local curvature is less than the preset lower bound and the density of reference vectors in the region is too large, redundant reference vectors are deleted to keep the total number of reference vectors constant, thereby obtaining a uniformly distributed Pareto front solution set on the irregular front of the target space.

[0057] For example, in step five, the fuzzy grey relational analysis method based on dynamic preference weights includes: A standardized decision matrix is ​​constructed for the Pareto front solution set. Then, the grey relational coefficient of each Pareto solution to the ideal solution is calculated using the preference weights that change dynamically according to production demand. The comprehensive grey relational degree is obtained by weighting the fuzzy membership degree. The solution corresponding to the maximum comprehensive grey relational degree is the optimal combination of process parameters.

[0058] Specifically, in this embodiment, in step five, it is assumed that the Pareto front solution set includes... There are solutions, each solution has... One target value, standardized decision matrix ,in For the first The solution is the first The normalized value of each objective. Calculate the grey relational coefficient: .

[0059] in, Indicates the first The solution is the first The grey relational coefficient of each target. For the absolute difference of all objectives, solve for the whole set. The minimum value, For the absolute difference of all objectives, solve for the whole set. The maximum value, For the first The ideal value for a target =0.5 is the resolution coefficient.

[0060] The dynamic preference weights are determined by the fuzzy membership function: .

[0061] in, For the first Fuzzy membership degree of each target For the first The reachability preference parameters of the current state of each objective. To set the first in real time according to the production task Parameters of demand preferences for each objective For the first Sensitivity factor for each target For the first The final dynamic preference weights for each objective To optimize the number of objectives, The sum of the fuzzy membership degrees of all targets. (Comprehensive grey relational analysis) ,choose The solution with the largest value is taken as the optimal combination of process parameters.

[0062] For example, in step six, the migration fine-tuning involves constructing a loss function that includes parameter regularization terms and using actual pultrusion feedback data to fine-tune some layer parameters of the process-performance proxy model, while retaining the predictive ability based on historical data.

[0063] The rolling optimization re-triggers the multi-objective optimization problem in each feedback cycle with a fine-tuned surrogate model, updating the optimal combination of process parameters.

[0064] Specifically, in this embodiment, in step six, the migration fine-tuning loss function is: .

[0065] in, The loss function value is used for migration fine-tuning. These are the parameters to be fine-tuned in the process-performance surrogate model. For the first A vector of true performance metrics for each actual sample. For the first A vector of process parameters for each actual sample. For parameters The time-based proxy model in The predicted output vector at that point, To fine-tune the model parameters, The regularization coefficient is . This is to increase the actual data sample size.

[0066] Rolling optimization refers to replacing the original surrogate model with a fine-tuned surrogate model after each feedback cycle, resolving the multi-objective optimization problem described in steps three and four, obtaining the updated optimal combination of process parameters, and then sending it to the pultrusion equipment for execution to form a closed loop.

[0067] The embodiments of this application improve the global optimization accuracy of the pultrusion process parameters of large insulating core rods and enhance the dynamic anti-interference capability of the production process by integrating thermal-fluid-solid multiphysics coupling simulation, hybrid adaptive sampling, improved multi-objective optimization, and digital twin online compensation into a whole-process process optimization method.

[0068] By employing a hybrid sampling strategy—first generating initial representative samples through orthogonal experiments, and then adaptively adding new samples based on the criterion of maximizing the prediction variance—the comprehensiveness and sampling efficiency of the process-performance dataset were improved, while the generation of redundant samples was reduced.

[0069] By employing a hybrid kernel function consisting of an adaptive weighted combination of radial basis kernel, linear kernel, and periodic kernel, and a multi-output deep Gaussian process structure with induced point sparse approximation, the fitting accuracy of the surrogate model for complex nonlinear coupling relationships of multiple performance indicators is improved.

[0070] By combining reparameterization techniques and stochastic gradient optimization with a variational sparse inference training method that aims to minimize the variational lower bound of the multi-output marginal likelihood, the training efficiency and multi-output prediction accuracy of deep Gaussian process surrogate models are improved.

[0071] By taking insulation strength, curing uniformity, and other factors as multiple objectives and introducing a dynamic feasible domain constraint that is adjusted in real time according to the maximum temperature difference inside the mandrel, the engineering feasibility of the multi-objective optimization solution is improved, and curing defects caused by process parameters exceeding the safety boundary are avoided.

[0072] By employing a multi-objective optimization algorithm that uses covariance matrix adaptive evolution to generate offspring, angle-penalized distance environment selection, and adaptive reference vector adjustment based on leading-edge curvature, the convergence speed of the algorithm and the uniformity of the Pareto solution set distribution in the target space are improved.

[0073] By constructing a standardized decision matrix and combining the fuzzy grey relational analysis method that calculates the grey relational coefficient and fuzzy membership degree weighted by dynamic preference weights, the flexibility of selecting optimal process parameters and their adaptability to actual production needs are improved.

[0074] By employing a loss function with parameterized regularization for migration fine-tuning and a rolling optimization mechanism that re-solves the problem every cycle based on feedback data, the adaptability of the surrogate model to actual production fluctuations and the online optimization accuracy of process parameters are improved.

[0075] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0076] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. An optimization method for the pultrusion process of large insulating core rods based on an improved multi-objective algorithm, characterized in that, Includes the following steps: Step 1: Construct a thermal-fluid-solid multiphysics coupled simulation model of the pultrusion process of large insulating core rods, and generate process-performance datasets by adopting a hybrid adaptive sampling strategy based on orthogonal experiment and prediction variance maximization. Step 2: Model the process-performance dataset using a mixed kernel function sparse multi-output deep Gaussian process, and obtain the process-performance proxy model through variational sparse inference training. Step 3: Based on the process-performance proxy model, with insulation strength, curing uniformity, residual stress and production efficiency as optimization objectives, a dynamic feasible domain constraint processing mechanism is introduced to establish a multi-objective optimization problem. Step four: A multi-objective optimization algorithm guided by a reference vector with adaptive evolution of the covariance matrix is ​​used to solve the multi-objective optimization problem, and an adaptive reference vector adjustment strategy based on the change of front curvature is used to obtain a uniformly distributed Pareto front solution set. Step 5: Select the optimal combination of process parameters from the Pareto front solution set using fuzzy grey relational analysis based on dynamic preference weights; Step six: Construct a digital twin online compensation mechanism, perform migration fine-tuning of the process-performance proxy model based on actual pultrusion feedback data, and perform rolling optimization of the optimal process parameter combination.

2. The optimization method for large insulating core rod pultrusion process based on improved multi-objective algorithm according to claim 1, characterized in that, In step one, the thermal-fluid-solid multiphysics coupling simulation model includes the resin flow control equation, the temperature-curing degree field coupling control equation, and the stress-strain constitutive equation. The temperature distribution, curing degree distribution, and thermal residual stress distribution during the pultrusion process are solved using the finite element method.

3. The optimization method for large insulating core rod pultrusion process based on an improved multi-objective algorithm according to claim 1, characterized in that, In step one, the hybrid adaptive sampling strategy based on orthogonal experimentation and prediction variance maximization includes: First, an initial process-performance sample point set is generated by orthogonal experimental design. Then, the prediction variance of the preliminary Gaussian process surrogate model trained with the initial process-performance sample point set is maximized as the addition criterion. Sample points are adaptively added in the process design space until the sample point coverage meets the preset convergence condition, thus obtaining the process-performance dataset.

4. The optimization method for large insulating core rod pultrusion process based on improved multi-objective algorithm according to claim 1, characterized in that, In step two, the hybrid kernel function sparse multi-output deep Gaussian process is composed of several stacked Gaussian process layers, and each Gaussian process layer adopts a hybrid kernel function and an induced point sparse approximation structure; the hybrid kernel function is formed by combining radial basis kernel, linear kernel and periodic kernel through adaptive weights.

5. The optimization method for the pultrusion process of large insulating core rods based on an improved multi-objective algorithm according to claim 1, characterized in that, In step two, the variational sparse inference training aims to minimize the variational lower bound of the multi-output marginal likelihood. It uses reparameterization techniques and stochastic gradient optimization to jointly learn the mixed kernel function weights, induced point locations, and variational parameters of all Gaussian process layers to obtain the process-performance surrogate model.

6. The optimization method for large insulating core rod pultrusion process based on improved multi-objective algorithm according to claim 1, characterized in that, In step three, establishing the multi-objective optimization problem includes: The four optimization objectives are to maximize insulation strength, minimize curing non-uniformity, minimize residual stress, and maximize production efficiency. At the same time, a dynamic feasible domain constraint boundary is introduced that is related to the temperature field distribution. The constraint limit of the dynamic feasible domain constraint boundary is adjusted in real time according to the maximum temperature difference inside the mandrel output by the process-performance proxy model.

7. The optimization method for large insulating core rod pultrusion process based on improved multi-objective algorithm according to claim 1, characterized in that, In step four, the reference vector-guided multi-objective optimization algorithm uses a covariance matrix adaptive evolution strategy to generate offspring populations during the evolution process, uses angle penalty distance index for environmental selection, and adds or deletes reference vectors in the irregular front region of the target space through an adaptive reference vector adjustment strategy based on front curvature changes.

8. The optimization method for large insulating core rod pultrusion process based on improved multi-objective algorithm according to claim 1, characterized in that, Step five, the fuzzy grey relational analysis method based on dynamic preference weights, includes: A standardized decision matrix is ​​constructed for the Pareto front solution set. Then, the grey relational coefficient of each Pareto solution to the ideal solution is calculated using the preference weights that change dynamically according to production demand. The comprehensive grey relational degree is obtained by weighting the fuzzy membership degree. The solution corresponding to the maximum comprehensive grey relational degree is the optimal combination of process parameters.

9. The optimization method for the pultrusion process of large insulating core rods based on an improved multi-objective algorithm according to claim 1, characterized in that, In step six, the migration fine-tuning involves constructing a loss function that includes a parameter regularization term and using actual pultrusion feedback data to fine-tune some layer parameters of the process-performance proxy model, while retaining the predictive ability based on historical data. The rolling optimization re-triggers the multi-objective optimization problem in each feedback cycle with a fine-tuned surrogate model, updating the optimal combination of process parameters.

Images