A Method for Constructing Industrial Simulation Proxy Models Based on Multi-Algorithm Collaboration

Abstract

This invention relates to the field of model building technology, and more particularly to a method for constructing an industrial simulation proxy model using multi-algorithm collaboration. The method involves constructing a pre-proxy model, determining parameter distribution representation values ​​and feature correlation values, calculating the optimization contribution of structural layers, adjusting the pre-proxy model, traversing several structural layers of the pre-proxy model, determining the dynamic redundancy of each structural layer, determining the actual sensitivity coefficient of each structural layer, determining the structural layer feature combination of each structural layer, determining the prior sensitivity coefficient of each structural layer based on a relational database constructed through structural layer sensitivity experiments, calculating the sensitivity for each structural layer based on the prior sensitivity coefficient and the actual sensitivity coefficient, determining the sensitivity ranking based on the sensitivity, and simplifying the structural layers based on the sensitivity ranking. This invention, by constructing a pre-proxy model and performing targeted processing, determines the industrial simulation proxy model, reducing the waste of computing power and improving the computational efficiency of the model.

Description

Technical Field

[0001] This invention relates to the field of model building technology, and in particular to a method for building industrial simulation proxy models using multi-algorithm collaboration. Background Technology

[0002] Proxy modeling technology approximates the input-output relationships of complex simulation models using mathematical models. It boasts advantages such as low computational cost and low complexity, and is widely used in engineering optimization, design analysis, and other fields. In particular, with the development of industrial simulation technology, multi-algorithm collaboration has become an important means to improve simulation accuracy and efficiency. Multi-algorithm collaborative industrial simulation proxy models can integrate the advantages of multiple algorithms to solve complex industrial problems, improve the accuracy and reliability of the model, and are suitable for nonlinear design spaces and design problems with long single-calculation times, such as 3D simulation. This helps avoid the problem of simulation models getting trapped in local extrema due to numerical noise, and improves the global search capability and efficiency of the algorithm.

[0003] Chinese Patent Publication No. CN118820696A discloses a method for constructing a federated evolutionary multi-agent-assisted feature selection model, comprising: collecting a gradient set; performing filtered feature selection using an XGBoost tree model for any participant; multiple participants uploading feature position indices with importance of 0 to a trusted third-party server to perform joint filtered feature selection; all participants obtaining a preliminary reduced feature subset RB; initializing the initial model of the regression agent and initializing relevant parameters of the PSO algorithm for any participant; performing a PSO-based agent-assisted federated evolutionary feature selection method on the obtained feature subset RB for any participant; and determining whether the current population of all participants has reached the termination condition of the algorithm. This invention can effectively solve the problem of high evaluation cost in existing technologies when handling high-dimensional gradient joint modeling in privacy-preserving scenarios with multiple participants.

[0004] Chinese Patent Publication No. CN114692347A discloses a method for constructing a temperature field surrogate model based on neural network architecture search technology. The method includes: generating training gradients according to component layout design requirements; preprocessing the training gradients; determining a neural network baseline model; setting search parameters based on the baseline model; constructing a super network model containing all possible paths; randomly selecting search parameters based on the super network model to obtain a sub-model; training the sub-model using the training gradients to update its model parameters; repeating the random selection of sub-models and training process until the number of training iterations reaches a preset value; searching for an approximately optimal sub-model using a multi-objective optimization algorithm based on the trained super network model; and retraining the approximately optimal sub-model using the training gradients to obtain a temperature field surrogate model. This invention can automatically design neural network structures according to component layout design requirements, resulting in a model with fewer parameters, higher prediction accuracy, and shorter prediction time.

[0005] However, the following problems still exist in the existing technology.

[0006] In the construction of industrial simulation proxy models, excessive algorithm involvement can generate redundant and useless data. This data can make the structure and logic of the entire industrial simulation proxy model more complex, increasing the difficulty of proxy model construction and the workload. Furthermore, in the process of simplifying the structural layers of the industrial simulation proxy model, some models are quite complex. Simplifying by traversing and analyzing the structural layers is inefficient and wastes computing power, resulting in low efficiency in the model construction process. Summary of the Invention

[0007] To address this issue, the present invention provides a method for constructing an industrial simulation proxy model using multiple algorithms in collaboration. This method aims to solve the problems that excessive algorithm involvement in the construction of industrial simulation proxy models leads to redundancy, increases the difficulty and workload of proxy model construction, complicates the model simplification process, wastes computing power, and results in low efficiency in the model construction process.

[0008] To achieve the above objectives, this invention provides a method for constructing a multi-algorithm collaborative industrial simulation proxy model, comprising:

[0009] Construct a pre-proxy model, determine the parameter distribution representation value based on the parameters of each structural layer in the pre-proxy model, and determine the feature correlation value based on the function of each structural layer;

[0010] The structural layer optimization contribution is calculated based on the parameter distribution characterization value and feature correlation value to determine whether the pre-proxy model should be adjusted.

[0011] The pre-proxy model is traversed through several structural layers. The dynamic redundancy of each structural layer is determined based on the absolute value of the gradient. The actual sensitivity coefficient of the structural layer is determined by combining the feature correlation value of the corresponding structural layer.

[0012] Determine the combination of structural features for each structural layer, and determine the prior sensitivity coefficient of each structural layer based on the association database constructed through structural layer sensitivity experiments;

[0013] The sensitivity of each structural layer is calculated based on the prior sensitivity coefficient and the actual sensitivity coefficient. The sensitivity is then ranked according to the sensitivity, and the structural layers are simplified based on the sensitivity ranking.

[0014] Fixed pre-proxy model, completed construction;

[0015] The structural layer sensitivity experiment includes simplifying each structural layer of a known structural layer feature combination, recording the change in runtime memory of the simplified sample simulation proxy model, and storing the correlation between several structural layer feature combinations and prior sensitivity coefficients. The structural layer feature combination includes the numerical range to which several structural layer features belong.

[0016] Furthermore, the process of determining the parameter distribution representation value based on the parameters of each structural layer in the pre-proxy model includes,

[0017] Determine several structural layers of the pre-proxy model;

[0018] Determine the number of parameters corresponding to the plurality of structural layers;

[0019] The variance of the parameters is defined as the characteristic value of the parameter distribution.

[0020] Furthermore, the process of determining the feature correlation value based on the function of each of the structural layers includes,

[0021] Record the number of local minima and saddle points corresponding to functions in several structural layers of the pre-proxy model;

[0022] The feature correlation value is calculated based on the number of local minima and the number of saddle points.

[0023] Furthermore, the process of calculating the contribution of the structural layer optimization includes,

[0024] The ratio of the parameter distribution characterization value to the benchmark parameter distribution characterization value is determined as the first influencing factor;

[0025] The ratio of the feature correlation value to the benchmark feature correlation value is determined as the second influencing factor;

[0026] The weighted sum of the first and second impact factors is determined to be the contribution of the structural layer optimization.

[0027] Furthermore, the determination of whether to adjust the pre-proxy model, wherein,

[0028] If the optimization contribution of the structural layer is greater than the optimization contribution threshold of the structural layer, it is determined that the pre-proxy model should be adjusted. The pre-proxy model is traversed through several structural layers, and the dynamic redundancy of each structural layer is determined based on the absolute value of the gradient. The actual sensitivity coefficient of the structural layer is determined by combining the feature correlation value of the corresponding structural layer.

[0029] If the contribution of the structural layer optimization is less than or equal to the structural layer optimization contribution threshold, it is determined that the pre-proxy model will not be adjusted.

[0030] Furthermore, the process of determining the dynamic redundancy of each structural layer based on the absolute value of the gradient includes,

[0031] Calculate the absolute values ​​of the gradients of several learnable parameters of the pre-proxy model;

[0032] Determine the number of learnable parameters corresponding to gradient absolute values ​​being less than a gradient absolute value threshold;

[0033] The ratio of the stated quantity to the total number of learnable parameters is determined as the dynamic redundancy.

[0034] Furthermore, the process of determining the actual sensitivity coefficient of the structural layer includes,

[0035] The ratio of the dynamic redundancy to the baseline dynamic redundancy is determined as the dynamic redundancy influence factor;

[0036] The ratio of the correlation value of the corresponding structural layer to the baseline correlation value is determined as the correlation value influence factor.

[0037] The weighted sum of the dynamic redundancy influence factor and the correlation value influence factor is determined to be the actual sensitivity coefficient of the structural layer.

[0038] Furthermore, the process of determining the prior sensitivity coefficients of each structural layer based on the association database constructed through structural layer sensitivity experiments includes:

[0039] Determine the range of values ​​for each structural layer feature, and construct the combination of structural layer features;

[0040] Determine the prior sensitivity coefficient associated with the combination of structural layer features;

[0041] The prior sensitivity coefficient is determined as the prior sensitivity coefficient of the structural layer;

[0042] The type of structural layer features is predetermined.

[0043] Furthermore, the process of constructing a relational database through structural layer sensitivity experiments includes,

[0044] Each structural layer with known structural layer feature combinations is simplified, and the change in runtime memory of the simplified sample simulation proxy model is recorded.

[0045] Record the average change in runtime memory corresponding to the feature combinations of each structural layer;

[0046] The ratio of the mean change in running memory to a predetermined threshold for the change in running memory is determined as the prior sensitivity coefficient for the association of structural layer feature combinations.

[0047] Store the prior sensitivity coefficients of the feature combinations associated with each structural layer to construct a relational database.

[0048] Furthermore, the process of calculating the sensitivity for each structural layer based on the prior sensitivity coefficient and the actual sensitivity coefficient includes,

[0049] The sensitivity is determined by the weighted sum of the prior sensitivity coefficient and the actual sensitivity coefficient.

[0050] Compared with existing technologies, this invention constructs a pre-proxy model, determines parameter distribution representation values ​​and feature correlation values, calculates the structural layer optimization contribution, adjusts the pre-proxy model, traverses several structural layers of the pre-proxy model, determines the dynamic redundancy of each structural layer, determines the actual sensitivity coefficient of each structural layer, determines the structural layer feature combination of each structural layer, determines the prior sensitivity coefficient of each structural layer based on the association database constructed through structural layer sensitivity experiments, calculates the sensitivity for each structural layer based on the prior sensitivity coefficient and the actual sensitivity coefficient, determines the sensitivity ranking based on the sensitivity ranking, and simplifies the structural layers based on the sensitivity ranking. This invention, by constructing a pre-proxy model and performing targeted processing, determines the industrial simulation proxy model, reduces the waste of computing power, and improves the computational efficiency of the model.

[0051] In particular, by determining the parameter distribution characterization value and feature correlation value, a basis is provided for calculating the contribution of the structural layer optimization. It can be understood that the parameter distribution characterization value represents the discrete difference of the structural layer of the constructed pre-proxy model. The larger the discrete difference, the higher the tendency of the pre-proxy model to adjust the structural layer direction. At the same time, the iterative complexity of the algorithm involved in the structural layer also affects the optimization of the proxy model. By characterizing the complexity of the structural layer of the constructed pre-proxy model participating in iterative calculations through feature correlation value, and combining the parameter distribution characterization value to calculate the structural layer optimization contribution, the necessity and effectiveness of optimizing the structural layer direction of the pre-proxy model can be comprehensively considered. In this way, the pre-proxy model can be optimized in a targeted manner, reducing the waste of computing power and improving the construction efficiency of the proxy model.

[0052] In particular, by calculating the contribution of structural layer optimization, the direction of adjustment for the pre-proxy model can be determined. In practice, the pre-proxy model may be a common model, such as the classic linear regression model or a simple decision tree model. These models have been widely used in similar scenarios in the past and have relatively mature architectures and parameter settings, and can be used directly. Alternatively, it may be a rare or even newly constructed model. In this case, the pre-proxy model may contain factors that lead to model complexity. Due to its high complexity, directly identifying it as the industrial simulation proxy model would lead to a waste of computing resources and may even affect the model's running efficiency. Based on this, this invention calculates the contribution of structural layer optimization after constructing the pre-proxy model, analyzes the probability of the structural layer's influence on the optimization of the pre-proxy model, and provides a basis for whether the model needs to be adjusted later, thereby reducing the waste of computing power and improving the model's computing efficiency.

[0053] In particular, by quickly calculating the absolute value of the gradient, a mathematical basis is provided for determining the actual sensitivity coefficient of the structural layer. In the practical application of the surrogate model, the learnable parameters are the adjustable parameters in the structural layer. These parameters usually participate in function operations or constraints. The absolute value of the gradient of the learnable parameters characterizes the degree of influence of these parameters on the function. The larger the absolute value of the gradient, the greater the influence of the parameter on the model output. Based on this, the actual sensitivity of the corresponding structural layer to simplification is characterized, providing data support for subsequent adaptive simplification of the structural layer, thereby reducing the waste of computing power and improving the computational efficiency of the model.

[0054] In particular, by conducting structural layer sensitivity experiments to construct an association database, the pre-proxy model state corresponding to different combinations of structural layer features is analyzed. Then, by combining prior sensitivity coefficients and actual sensitivity coefficients, the sensitivity of each structural layer is determined, providing a theoretical basis for subsequently determining the optimal structural layer feature combinations. In practice, all structural layer features affect subsequent simplification, and the degree of influence varies with different combinations of structural layer features. Based on this, this invention considers constructing the sensitivity law of structural layers to subsequent simplification under different combinations of structural layer features through structural layer sensitivity experiments, thereby forming a correlation between structural layer feature combinations and prior sensitivity coefficients. This facilitates rapid subsequent sensitivity calculation. The sensitivity of structural layers to simplification is comprehensively considered from both practical and prior perspectives, providing support for simplification. Priority is given to simplifying structural layers with higher sensitivity. Simplifying fewer structural layers allows the proxy model to meet practical application standards, thereby reducing wasted computing power and improving the model's computational efficiency. Attached Figure Description

[0055] Figure 1 A schematic diagram illustrating the steps of a multi-algorithm collaborative industrial simulation proxy model construction method according to an embodiment of the invention;

[0056] Figure 2 This is a schematic diagram illustrating the steps of determining the parameter distribution characterization value in an embodiment of the invention;

[0057] Figure 3 This is a schematic diagram illustrating the steps of determining feature-related values ​​in an embodiment of the invention;

[0058] Figure 4 A schematic diagram illustrating the steps for calculating the contribution of the structural layer optimization in an embodiment of the invention;

[0059] Figure 5 This is a block diagram illustrating the logic for determining whether to adjust the pre-proxy model according to an embodiment of the invention.

[0060] Figure 6 A schematic diagram illustrating the steps of determining the dynamic redundancy of each structural layer based on the absolute value of the gradient in an embodiment of the invention;

[0061] Figure 7 This is a schematic diagram illustrating the steps for determining the actual sensitivity coefficient of a structural layer in an embodiment of the invention.

[0062] Figure 8 A schematic diagram illustrating the steps for determining the prior sensitivity coefficients of each structural layer in an embodiment of the invention;

[0063] Figure 9 This is a schematic diagram illustrating the steps of constructing an association database according to an embodiment of the invention. Detailed Implementation

[0064] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0065] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0066] Please see Figure 1 , Figure 1 This is a schematic diagram illustrating the method steps of a multi-algorithm collaborative industrial simulation proxy model construction method according to an embodiment of the invention. The multi-algorithm collaborative industrial simulation proxy model construction method of the present invention includes:

[0067] Step S1: Construct a pre-proxy model, determine the parameter distribution representation value based on the parameters of each structural layer in the pre-proxy model, and determine the feature correlation value based on the function of each structural layer;

[0068] Step S2: Calculate the structural layer optimization contribution based on the parameter distribution characterization value and feature correlation value to determine whether to adjust the pre-proxy model;

[0069] Step S3: Traverse several structural layers of the pre-proxy model, determine the dynamic redundancy of each structural layer based on the absolute value of the gradient, and determine the actual sensitivity coefficient of the structural layer by combining the feature correlation value of the corresponding structural layer.

[0070] Step S4: Determine the combination of structural features for each structural layer, and determine the prior sensitivity coefficient of each structural layer based on the association database constructed through the structural layer sensitivity experiment.

[0071] Step S5: Calculate the sensitivity for each structural layer based on the prior sensitivity coefficient and the actual sensitivity coefficient, determine the sensitivity ranking based on the sensitivity, and simplify the structural layers based on the sensitivity ranking.

[0072] Step S6: Fix the pre-proxygen model to complete the construction;

[0073] The structural layer sensitivity experiment includes simplifying each structural layer of a known structural layer feature combination, recording the change in runtime memory of the simplified sample simulation proxy model, and storing the correlation between several structural layer feature combinations and prior sensitivity coefficients. The structural layer feature combination includes the numerical range to which several structural layer features belong.

[0074] Specifically, the structural layers in the implementation are the sub-layers of the intermediate layer of the proxy model. The proxy model mainly includes an input layer, an intermediate layer, and an output layer. The input layer is used to receive input data and perform preprocessing. The intermediate layer is the core of the proxy model and usually contains several sub-layers. Each sub-layer is responsible for handling specific computation or data transformation tasks. It can be composed of models or combined operation functions. Through the intermediate layer, the input data can be processed to obtain the output results, which are then output through the output layer. Usually, the construction of the intermediate layer is determined based on the actual engineering problem and application scenario, which will not be elaborated here.

[0075] Specifically, the parameters of the structural layer are the parameters that participate in the calculation within the structural layer.

[0076] Specifically, there is no limitation on the specific traversal method. For example, it can be a breadth-first traversal, visiting the structure layer by layer. Of course, other traversal methods can also be used, as long as it can ensure that the structure layer of the pre-proxy model can be fully accessed. Those skilled in the art can choose according to the actual situation, which will not be elaborated here.

[0077] Specifically, structural layer characteristics include the number of functions, the number of constraints, the coupling strength, and the average order of the polynomials.

[0078] The number of functions refers to the number of functions contained in the structure layer. The number of functions represents the computational complexity of the structure layer.

[0079] The number of constraints refers to the number of constraints contained in the structural layer. The number of constraints represents the complexity of the structural layer in terms of constraints.

[0080] The coupling strength is the average number of functions associated with each parameter in the structural layer, which characterizes the degree of association between the functions. A parameter is considered to participate in the function operation if it is associated with that function.

[0081] The average order of polynomials is the average of the highest order of the polynomials of each function in the structure layer, which characterizes the computational complexity of the functions in the structure layer.

[0082] Please see Figure 2 , Figure 2 This is a schematic diagram illustrating the steps of determining parameter distribution representation values ​​according to an embodiment of the invention. Specifically, the process of determining parameter distribution representation values ​​based on the parameters of each structural layer in the pre-proxy model includes,

[0083] Step S21: Determine several structural layers of the pre-proxy model;

[0084] Step S22: Determine the number of parameters corresponding to the plurality of structural layers;

[0085] Step S23: Determine the quantity variance of the parameters as the parameter distribution characterization value.

[0086] Please see Figure 3 , Figure 3 This is a schematic diagram illustrating the steps of determining feature correlation values ​​according to an embodiment of the invention. Specifically, the process of determining feature correlation values ​​based on the functions of each structural layer includes,

[0087] Step S31: Record the number of local minima and saddle points corresponding to the functions in several structural layers of the pre-proxy model;

[0088] Step S31: Calculate the feature correlation value based on the number of local minimum points and the number of saddle points.

[0089] It is understandable that local minima and saddle points are relative to the function. After iterative processing, the number of local minima and the number of saddle points are recorded.

[0090] Specifically, the steps for calculating the feature correlation value are as follows.

[0091] The average number of local minimum points in several structural layers is recorded as the first average number.

[0092] The average number of saddle points in several structural layers is recorded as the second average number.

[0093] The average of the first quantity average and the second quantity average is determined as the characteristic correlation value.

[0094] Specifically, a local minimum point refers to a point within a local region of the function's domain where the function value is less than or equal to the function values ​​of all other points in its neighborhood.

[0095] There are no restrictions on the method for determining local minima. For example, local minima can be determined by gradient descent, which involves calculating the gradient of the function and performing multiple iterations. Of course, those skilled in the art can also choose a method for determining local minima based on the actual situation, which will not be elaborated here.

[0096] Specifically, a saddle point represents a point where the gradient of a function is zero, but that point is not a local minimum.

[0097] There is no limitation on the method for determining saddle points. For example, saddle points can be determined by the Hessian matrix. If the Hessian matrix at a certain point is neither a positive definite matrix nor a negative definite matrix, then that point is determined to be a saddle point. Of course, those skilled in the art can also choose a method for determining saddle points according to the actual situation, which will not be elaborated here.

[0098] Specifically, by determining the parameter distribution representation value and feature correlation value, a calculation basis is provided for calculating the contribution of the structural layer optimization. It can be understood that the parameter distribution representation value represents the discrete difference of the structural layer of the constructed pre-proxy model. The larger the discrete difference, the higher the tendency of the pre-proxy model to adjust the structural layer direction. At the same time, the iterative complexity of the algorithm involved in the structural layer also affects the optimization of the proxy model. By representing the complexity of the structural layer of the constructed pre-proxy model participating in iterative calculations through feature correlation value, and combining the parameter distribution representation value to calculate the structural layer optimization contribution, the necessity and effectiveness of optimizing the structural layer direction of the pre-proxy model are comprehensively considered. In this way, the pre-proxy model is optimized in a targeted manner, reducing the waste of computing power and improving the construction efficiency of the proxy model.

[0099] Please see Figure 4 , Figure 4 This is a schematic diagram illustrating the steps of calculating the structural layer optimization contribution in an embodiment of the invention. Specifically, the process of calculating the structural layer optimization contribution includes:

[0100] Step S41: Determine the ratio of the parameter distribution characterization value to the benchmark parameter distribution characterization value as the first influencing factor;

[0101] Step S42: Determine the ratio of the feature correlation value to the benchmark feature correlation value as the second influencing factor;

[0102] Step S43: Determine the weighted sum of the first influence factor and the second influence factor as the structural layer optimization contribution.

[0103] Specifically, the baseline parameter distribution characterization value is calculated in advance. The parameter distribution characterization values ​​of several optimized proxy model structural layers are obtained in advance, the average value of several parameter distribution characterization values ​​is determined, and the baseline parameter distribution characterization value is set as the product of the average value of the parameter distribution characterization value and the offset coefficient.

[0104] The baseline feature correlation value is calculated in advance. The feature correlation values ​​of several simplified and completed proxy model structure layers are obtained in advance. The average value of several feature correlation values ​​is determined. The baseline feature correlation value is set as the product of the average value of the feature correlation values ​​and the offset coefficient. The offset coefficient is in the interval [1.35, 1.55].

[0105] Specifically, the sum of the weighting coefficients of the first impact factor and the second impact factor is 1, the weighting coefficient of the first impact factor is 0.42, and the weighting coefficient of the second impact factor is 0.58.

[0106] Specifically, by calculating the contribution of structural layer optimization, the direction of adjustment for the pre-proxy model is determined. In practice, the pre-proxy model may be a common model, such as a classic linear regression model or a simple decision tree model. These models have been widely used in similar scenarios in the past and have relatively mature architectures and parameter settings, making them ready for direct use. Alternatively, it may be a rare or even newly constructed model. In this case, the pre-proxy model may contain factors that lead to model complexity. Due to its potentially high complexity, directly identifying it as the industrial simulation proxy model would lead to a waste of computing resources and may even affect the model's operating efficiency. Based on this, this invention calculates the contribution of structural layer optimization after constructing the pre-proxy model, analyzes the probability of the structural layer's influence on the optimization of the pre-proxy model, and provides a basis for whether the model needs to be adjusted later, thereby reducing the waste of computing power and improving the model's computational efficiency.

[0107] Please see Figure 5 , Figure 5 This is a logic block diagram illustrating how to determine whether to adjust the pre-proxy model according to an embodiment of the invention. Specifically, determining whether to adjust the pre-proxy model involves...

[0108] If the optimization contribution of the structural layer is greater than the optimization contribution threshold of the structural layer, it is determined that the pre-proxy model should be adjusted. The pre-proxy model is traversed through several structural layers, and the dynamic redundancy of each structural layer is determined based on the absolute value of the gradient. The actual sensitivity coefficient of the structural layer is determined by combining the feature correlation value of the corresponding structural layer.

[0109] If the contribution of the structural layer optimization is less than or equal to the structural layer optimization contribution threshold, it is determined that the pre-proxy model will not be adjusted.

[0110] Specifically, the structural layer optimization contribution threshold represents a boundary that the pre-proxy model needs to be adjusted. It is calculated in advance by obtaining the historical structural layer optimization contribution of several adjusted pre-proxy models in advance, solving the average value of each historical structural layer optimization contribution, and determining the product of the average value and the influence coefficient as the structural layer optimization contribution threshold. The influence coefficient is selected in the interval [0.8, 0.95]. In order to improve the construction accuracy of the pre-proxy model, the influence coefficient is determined to be 0.87 in the implementation.

[0111] Please see Figure 6 , Figure 6 This is a schematic diagram illustrating the steps of determining the dynamic redundancy of each structural layer based on the absolute value of the gradient, as per an embodiment of the invention. Specifically, the process of determining the dynamic redundancy of each structural layer based on the absolute value of the gradient includes:

[0112] Step S61: Calculate the absolute values ​​of the gradients of several learnable parameters of the pre-proxy model;

[0113] Step S62: Determine the number of learnable parameters corresponding to gradient absolute values ​​being less than the gradient absolute value threshold;

[0114] Step S63: The ratio of the quantity to the total number of learnable parameters is determined as the dynamic redundancy.

[0115] Specifically, learnable parameters are adjustable parameters in the structural layer, which typically participate in function operations or constraints. Their specific form depends on the type of surrogate model, which will not be elaborated here.

[0116] Specifically, the absolute value of the gradient represents the degree of influence of the learnable parameters on the function. There are no restrictions on how the absolute value of the gradient is calculated. For example, the partial derivatives of the function with respect to the model parameters can be obtained according to the chain rule. The absolute value of the gradient can be obtained by taking the absolute value of each element of the gradient. Of course, those skilled in the art can also use other methods, as long as the absolute value of the gradient can be obtained. This will not be elaborated further.

[0117] Specifically, the gradient absolute value threshold is pre-calculated data. Several simplified pre-proxy models are obtained in advance, and several corresponding gradient absolute values ​​are determined. The average value of these gradient absolute values ​​is then used as the gradient absolute value threshold.

[0118] Specifically, by quickly calculating the absolute value of the gradient, a mathematical basis is provided for determining the actual sensitivity coefficient of the structural layer. In the practical application of the surrogate model, the learnable parameters are the adjustable parameters in the structural layer. These parameters usually participate in function operations or constraints. The absolute value of the gradient of the learnable parameters characterizes the degree of influence of these parameters on the function. The larger the absolute value of the gradient, the greater the influence of the parameter on the model output. Based on this, the actual sensitivity of the corresponding structural layer to simplification is characterized, providing data support for subsequent adaptive simplification of the structural layer, thereby reducing the waste of computing power and improving the computational efficiency of the model.

[0119] Please see Figure 7 , Figure 7 This is a schematic diagram illustrating the steps for determining the actual sensitivity coefficient of a structural layer according to an embodiment of the invention. Specifically, the process of determining the actual sensitivity coefficient of a structural layer includes:

[0120] Step S71: Determine the ratio of the dynamic redundancy to the baseline dynamic redundancy as the dynamic redundancy influence factor.

[0121] Step S72: Determine the ratio of the feature correlation value of the corresponding structural layer to the baseline feature correlation value as the correlation value influence factor;

[0122] Step S73: Determine the weighted sum of the dynamic redundancy influence factor and the correlation value influence factor as the actual sensitivity coefficient of the structural layer.

[0123] Specifically, the baseline dynamic redundancy is calculated in advance. Several simplified pre-proxy models are obtained in advance, and several corresponding dynamic redundancies are determined. The average value of these dynamic redundancies is determined as the baseline dynamic redundancy.

[0124] Specifically, the sum of the weighting coefficients of the dynamic redundancy influence factor and the correlation value influence factor is 1, the weighting coefficient of the dynamic redundancy influence factor is 0.53, and the weighting coefficient of the correlation value influence factor is 0.47.

[0125] Please see Figure 8 , Figure 8 This diagram illustrates the steps for determining the prior sensitivity coefficients of each structural layer in an embodiment of the invention. Specifically, the process of determining the prior sensitivity coefficients of each structural layer based on an association database constructed through structural layer sensitivity experiments includes:

[0126] Step S81: Determine the range of values ​​for each structural layer feature and construct the structural layer feature combination.

[0127] Step S82: Determine the prior sensitivity coefficient associated with the combination of structural layer features;

[0128] Step S83: Determine the prior sensitivity coefficient as the prior sensitivity coefficient of the structural layer;

[0129] The type of structural layer features is predetermined.

[0130] Specifically, this invention constructs a relational database by conducting structural layer sensitivity experiments to analyze the pre-proxy model state under different combinations of structural layer features. Then, by combining prior sensitivity coefficients with actual sensitivity coefficients, the sensitivity of each structural layer is determined, providing a theoretical basis for subsequently determining the optimal structural layer feature combinations. In practice, all structural layer features affect subsequent simplification, and the degree of influence varies with different combinations of structural layer features. Based on this, this invention considers constructing the sensitivity law of structural layers to subsequent simplification under different combinations of structural layer features through structural layer sensitivity experiments, thereby forming a correlation between structural layer feature combinations and prior sensitivity coefficients. This facilitates rapid subsequent sensitivity calculation. By comprehensively considering both practical and prior dimensions of structural layer sensitivity to simplification, support is provided for simplification. Simplification is prioritized for structural layers with higher sensitivity, and simplification of fewer structural layers ensures that the proxy model meets practical application standards, thereby reducing computational waste and improving the model's computational efficiency.

[0131] Please see Figure 9 , Figure 9 This is a schematic diagram illustrating the steps of constructing an association database according to an embodiment of the invention. Specifically, the process of constructing the association database through structural layer sensitivity experiments includes:

[0132] Step S91: Simplify each structural layer of the known structural layer feature combination, and record the change in runtime memory of the simplified sample simulation proxy model.

[0133] Step S92: Record the average change in runtime memory corresponding to each structural layer feature combination;

[0134] Step S93: The ratio of the average change in running memory to a predetermined threshold for the change in running memory is determined as the prior sensitivity coefficient for the association of structural layer feature combinations.

[0135] Step S94: Store the prior sensitivity coefficients of the feature combination associations of each structural layer to construct the association database.

[0136] It is understandable that the change in running memory is the difference between the average running memory during the operation of the unsimplified sample simulation proxy model and the average running memory during the operation of the simplified sample simulation proxy model.

[0137] It is understandable that structural layer features are represented in numerical form, and the combination of structural layer features includes the numerical range to which the structural layer features belong. The numerical range of each type of structural layer feature is predefined, and there are no restrictions on the method of defining the numerical range.

[0138] In practice, the maximum and minimum values ​​of various structural layer features corresponding to several proxy models can be pre-calculated to obtain the maximum range interval. The maximum range interval is then divided into four numerical ranges for the corresponding type of structural layer feature.

[0139] Specifically, the threshold for changes in running memory is determined based on the sum of changes in running memory across all structural layers in the pre-proxy model. Several simplified historical pre-proxy models are obtained in advance, the sum of historical changes in running memory across all structural layers is determined, and the average of the historical sums of running memory changes is determined as the threshold for changes in running memory.

[0140] Specifically, the process of calculating the sensitivity for each structural layer based on the prior sensitivity coefficient and the actual sensitivity coefficient includes,

[0141] The sensitivity is determined by the weighted sum of the prior sensitivity coefficient and the actual sensitivity coefficient.

[0142] Specifically, the sum of the weighting coefficients of the prior sensitivity coefficient and the actual sensitivity coefficient is 1, the weighting coefficient of the prior sensitivity coefficient is 0.42, and the weighting coefficient of the actual sensitivity coefficient is 0.58.

[0143] Specifically, the sensitivity is sorted in descending order, and the combination of structural layer features is simplified according to the sensitivity order to improve the computational efficiency of the model.

[0144] Specifically, there are no restrictions on the simplification methods. For example, redundant neurons, connections, or layers in the industrial deep learning model can be removed through model architecture pruning (such as channel pruning and weight sparsification). Alternatively, the structural layers can be simplified by replacing the combined parameters of the complex structure with equivalent parameters (such as equivalent stiffness and equivalent thermal resistance). Of course, those skilled in the art can also use other methods to simplify, which will not be elaborated here.

[0145] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0146] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for constructing a multi-algorithm collaborative industrial simulation agent model, characterized in that, The method comprises the following steps: constructing a pre-agent model, determining parameter distribution representation values according to parameters of each structural layer in the pre-agent model, and determining feature correlation values according to functions of each structural layer; calculating structural layer optimization contribution degrees based on the parameter distribution representation values and the feature correlation values to determine whether to adjust the pre-agent model; if it is determined to adjust the pre-agent model, traversing a plurality of structural layers of the pre-agent model, determining dynamic redundancy degrees of each structural layer according to gradient absolute values, and determining structural layer actual sensitive coefficients in combination with the feature correlation values of the corresponding structural layers; determining structural layer feature combinations of each structural layer, determining prior sensitive coefficients of each structural layer according to a correlation relationship library constructed through structural layer sensitive experiments; calculating sensitivities of each structural layer according to the prior sensitive coefficients and the actual sensitive coefficients, determining a sensitivity ranking according to the sensitivities, and simplifying the structural layers according to the sensitivity ranking; fixing the pre-agent model and completing the construction; wherein the structural layer sensitive experiment comprises simplifying each structural layer of a known structural layer feature combination respectively, recording a running memory change amount of a simplified sample simulation agent model, the correlation relationship library stores the correlation relationship between a plurality of structural layer feature combinations and prior sensitive coefficients, and the structural layer feature combination contains a numerical range to which a plurality of structural layer features belong; the process of constructing the correlation relationship library through the structural layer sensitive experiment comprises, simplifying each structural layer of a known structural layer feature combination respectively, and recording a running memory change amount of a simplified sample simulation agent model; recording a mean value of the running memory change amount corresponding to each structural layer feature combination; determining the ratio of the mean value of the running memory change amount to a predetermined running memory change amount threshold as the prior sensitive coefficient associated with the structural layer feature combination; storing the prior sensitive coefficients associated with each structural layer feature combination to construct the correlation relationship library.

2. The multi-algorithm coordinated industrial simulation agent model construction method according to claim 1, characterized in that, The process of determining the parameter distribution representation values according to the parameters of each structural layer in the pre-agent model comprises, determining a plurality of structural layers of the pre-agent model; determining the number of parameters corresponding to the plurality of structural layers; determining the variance of the number of parameters as the parameter distribution representation value.

3. The multi-algorithm coordinated industrial simulation agent model construction method according to claim 2, characterized in that, The process of determining the feature correlation values according to the functions of each structural layer comprises, recording the number of local minimum points and the number of saddle points corresponding to the functions in the plurality of structural layers of the pre-agent model; calculating the feature correlation values based on the number of local minimum points and the number of saddle points.

4. The multi-algorithm coordinated industrial simulation agent model construction method according to claim 3, characterized in that, The process of calculating the structural layer optimization contribution degrees comprises, determining the ratio of the parameter distribution representation value to a benchmark parameter distribution representation value as a first influence factor; determining the ratio of the feature correlation value to a benchmark feature correlation value as a second influence factor; determining the weighted sum of the first influence factor and the second influence factor as the structural layer optimization contribution degree.

5. The multi-algorithm coordinated industrial simulation agent model construction method according to claim 1, characterized in that, The process of determining whether to adjust the pre-agent model comprises, If the structural layer optimization contribution degree is greater than a structural layer optimization contribution degree threshold value, it is determined that the pre-agent model is adjusted, a plurality of structural layers of the pre-agent model are traversed, a dynamic redundancy degree of each structural layer is determined according to a gradient absolute value, and a structural layer actual sensitive coefficient is determined in combination with the feature correlation value of the corresponding structural layer; If the structural layer optimization contribution degree is less than or equal to the structural layer optimization contribution degree threshold value, it is determined that the pre-agent model is not adjusted.

6. The multi-algorithm coordinated industrial simulation agent model construction method according to claim 3, characterized in that, The process of determining the dynamic redundancy degree of each structural layer according to the gradient absolute value includes, calculating the gradient absolute value of a plurality of learnable parameters of the pre-agent model; determining the number of learnable parameters corresponding to the gradient absolute value less than a gradient absolute value threshold value; determining the ratio of the number to the total number of learnable parameters as the dynamic redundancy degree.

7. The multi-algorithm coordinated industrial simulation agent model construction method according to claim 6, characterized in that, The process of determining the structural layer actual sensitive coefficient includes, determining the ratio of the dynamic redundancy degree to a reference dynamic redundancy degree as a dynamic redundancy degree influence factor; determining the ratio of the feature correlation value of the corresponding structural layer to a reference feature correlation value as a correlation value influence factor; determining the weighted sum value of the dynamic redundancy degree influence factor and the correlation value influence factor as the structural layer actual sensitive coefficient.

8. The multi-algorithm coordinated industrial simulation agent model construction method according to claim 1, characterized in that, The process of determining the prior sensitive coefficient of each structural layer according to the correlation relationship library constructed through the structural layer sensitive experiment includes, determining the range in which the numerical value of each structural layer feature of the structural layer is located, and constructing the structural layer feature combination; determining the prior sensitive coefficient associated with the structural layer feature combination; determining the prior sensitive coefficient as the prior sensitive coefficient of the structural layer; wherein the type of the structural layer feature is predetermined.

9. The multi-algorithm coordinated industrial simulation agent model construction method according to claim 1, characterized in that, The process of calculating the sensitivity of each structural layer according to the prior sensitive coefficient and the actual sensitive coefficient includes, determining the weighted sum value of the prior sensitive coefficient and the actual sensitive coefficient as the sensitivity. The process of determining the structural layer optimization contribution degree includes, calculating the gradient absolute value of a plurality of learnable parameters of the pre-agent model; determining the number of learnable parameters corresponding to the gradient absolute value less than a gradient absolute value threshold value; determining the ratio of the number to the total number of learnable parameters as the structural layer optimization contribution degree. The process of determining the structural layer actual sensitive coefficient includes, determining the ratio of the dynamic redundancy degree to a reference dynamic redundancy degree as a dynamic redundancy degree influence factor; determining the ratio of the feature correlation value of the corresponding structural layer to a reference feature correlation value as a correlation value influence factor; determining the weighted sum value of the dynamic redundancy degree influence factor and the correlation value influence factor as the structural layer actual sensitive coefficient. The process of determining the prior sensitive coefficient of each structural layer according to the correlation relationship library constructed through the structural layer sensitive experiment includes, determining the range in which the numerical value of each structural layer feature of the structural layer is located, and constructing the structural layer feature combination; determining the prior sensitive coefficient associated with the structural layer feature combination; determining the prior sensitive coefficient as the prior sensitive coefficient of the structural layer; wherein the type of the structural layer feature is predetermined. The process of calculating the sensitivity of each structural layer according to the prior sensitive coefficient and the actual sensitive coefficient includes, determining the weighted sum value of the prior sensitive coefficient and the actual sensitive coefficient as the sensitivity.

Images