Multi-layer protection structure performance optimization method based on SHAP technology

Abstract

The invention relates to a multilayer protection structure performance optimization method based on an SHAP technology, belongs to the technical field of protection structure performance optimization, and solves the problem that efficient and reliable optimization of multilayer protection structure performance cannot be realized by a multilayer structure optimization method in the prior art. Comprising the steps of obtaining an original optimization condition of a to-be-optimized multi-layer protection structure; the method comprises the following steps: constructing a sample data set based on an original optimization condition of a to-be-optimized multi-layer protection structure, and training a constructed proxy model to obtain a trained proxy model; based on the sample data set, the proxy model and an SHAP technology, obtaining SHAP characteristics of the to-be-optimized multi-layer protection structure; and based on the SHAP characteristics of the to-be-optimized multi-layer protection structure, the trained agent model and the SHAP technology, obtaining the optimal value of each characteristic parameter of the to-be-optimized multi-layer protection structure, and completing the performance optimization of the multi-layer protection structure. And efficient, reliable and interpretable performance optimization of the multi-layer protection structure is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of optimizing the performance of protection structures, and particularly to a method for optimizing the performance of a multi-layer protection structure based on SHAP technology. Background Art

[0002] Currently, a large number of new materials and structures have been continuously proposed to absorb and dissipate the energy of external damage elements. Among them, polyurea and silicon carbide are representative new materials, while foamed aluminum, honeycomb, and grille with multi-cell characteristics are typical new structures. Compared with traditional solid protection structures, composite structures with a metal multi-cell structure as the core layer can not only dissipate energy through the bending, stretching, and fracture of the front and rear panels and the crushing of the sandwich layer, but also have many advantages such as low cost, light weight, and high specific energy absorption on the premise of having the high hardness and stiffness characteristics of metals. Therefore, they have been widely applied to many fields such as aerospace, tank armor, rail transit, automobiles, and ships.

[0003] Currently, the main research directions for optimizing protection structures include two aspects: structural topology and morphology, such as multi-layer gradient and negative Poisson design, and their improvement effects have been verified in field tests and numerical simulations. However, although these methods can improve the comprehensive performance of the structure, many new problems such as numerous parameters, difficult design, and complex cooperative interaction effects also arise. Traditional structural optimization design methods need to go through multiple tests to meet the established performance requirements, so the cost is extremely high during the optimization implementation process. In order to reduce the cost-effectiveness ratio of the structural optimization scheme, numerical simulation technology is used to reduce the number of tests and improve the understanding of the internal response mechanism of the structure. However, a series of problems such as difficult numerical modeling, long solution time, high computing power requirements, and complex software operations still result in a relatively long structural optimization design cycle.

[0004] In view of this, an optimization strategy method combining a surrogate model and an optimization algorithm is proposed. First, an effective database is obtained through numerical and field tests, and then a surrogate model that can reflect the internal mapping relationship of parameters is constructed on this basis. Finally, an optimization design scheme can be obtained by combining the optimization algorithm. However, since the optimization algorithm is extremely prone to falling into local optimal values during the optimization process, the optimization results sometimes cannot meet the established requirements. In addition, since the surrogate model is a typical "black box" model, that is, the model can only feedback the corresponding output information based on the input data, researchers often lack an understanding of the internal response mechanism of the model, thereby reducing the degree of trust in the optimization results.

[0005] Therefore, the current multi-layer structure optimization methods cannot achieve efficient, reliable, and interpretable optimization of the performance of multi-layer protection structures. Summary of the Invention

[0006] In view of the above analysis, the embodiments of the present invention aim to provide a method for optimizing the performance of a multi-layer protection structure based on SHAP technology, so as to solve the problem that the existing multi-layer structure optimization methods cannot achieve efficient, reliable, and interpretable performance of the multi-layer protection structure.

[0007] The embodiments of the present invention provide a method for optimizing the performance of a multi-layer protection structure based on SHAP technology, including the following steps:

[0008] Obtain the original optimization conditions of the multi-layer protection structure to be optimized; wherein, the original optimization conditions include an objective function and constraint conditions, and the constraint conditions are constructed by the characteristic parameters of the multi-layer protection structure to be optimized;

[0009] Based on the original optimization conditions of the multi-layer protection structure to be optimized, construct a sample data set, and then train the constructed surrogate model to obtain a trained surrogate model; wherein, each sample data includes the specific values of each characteristic parameter and the corresponding specific value of the objective function;

[0010] Based on the sample data set, the surrogate model, and SHAP technology, obtain the SHAP characteristics of the multi-layer protection structure to be optimized; wherein, the SHAP characteristics include the SHAP values of each characteristic parameter in each sample and the global importance of each characteristic parameter;

[0011] Based on the SHAP characteristics of the multi-layer protection structure to be optimized, the trained surrogate model, and SHAP technology, obtain the optimal values of each characteristic parameter of the multi-layer protection structure to be optimized, and complete the performance optimization of the multi-layer protection structure.

[0012] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects:

[0013] A method for optimizing the performance of a multi-layer protection structure based on SHAP technology provided by the present invention

[0014] 1. Use SHAP technology to obtain the SHAP values of each characteristic parameter to characterize the global and local effects of each characteristic parameter in the sample. The global importance ranking result formed based on this can not only make each characteristic parameter physically comparable, but also filter out the characteristic parameters with high cost and low impact, thereby improving the generalization performance of the surrogate model and the optimization efficiency and quality of the optimization algorithm while ensuring the scientific and reasonable selection of characteristics;

[0015] 2. Before the algorithm optimization, obtain the sensitive range of each characteristic parameter based on SHAP technology, so as to formulate reasonable and efficient algorithm constraint conditions, which can reduce the probability of the algorithm falling into local optimum, accelerate the algorithm optimization speed, and improve the optimization quality;

[0016] 3. The optimization method without using an optimization algorithm can obtain a clear, definite, and interpretable adjustment method in real time through SHAP technology, and complete the structural performance optimization task in an extremely short time. The optimization cycle is short, and the optimization scheme has clear interpretability, which can enhance the trust in the prediction results of the surrogate model.

[0017] In the present invention, the above technical solutions can also be combined with each other to achieve more preferred combination schemes. Other features and advantages of the present invention will be described in the following specification. Moreover, some advantages can be made obvious from the specification or understood by implementing the present invention. The objectives and other advantages of the present invention can be achieved and obtained through the content specifically pointed out in the specification and the drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The drawings are only for the purpose of showing specific embodiments and are not considered to be a limitation of the present invention. Throughout the drawings, the same reference signs denote the same components.

[0019] Figure 1 It is a schematic flowchart of the multi-layer protection structure performance optimization method based on SHAP technology provided in Embodiment 1 of the present invention;

[0020] Figure 2 It is a schematic diagram of the three-layer aluminum foam sandwich structure provided in Embodiment 2 of the present invention;

[0021] Figure 3 It is a schematic diagram of the time history curve of the displacement of the rear panel under the action of an external load provided in Embodiment 2 of the present invention;

[0022] Figure 4(a) is a schematic diagram of the global importance ranking of each characteristic parameter provided in Embodiment 2 of the present invention;

[0023] Figure 4(b) is a schematic diagram of the global correlation of each characteristic parameter provided in Embodiment 2 of the present invention;

[0024] Figure 5(a) is a waterfall plot of the SHAP values of each characteristic parameter under the configuration of a uniform aluminum foam sandwich panel provided in Embodiment 2 of the present invention;

[0025] Figure 5(b) is a waterfall plot of the SHAP values of each characteristic parameter after the first update provided in Embodiment 2 of the present invention;

[0026] Figure 5(c) is a waterfall plot of the SHAP values of each characteristic parameter after the second update provided in Embodiment 2 of the present invention;

[0027] Figure 5(d) is a waterfall plot of the SHAP values of each characteristic parameter after the third update provided in Embodiment 2 of the present invention;

[0028] Figure 6(a) is a characteristic interaction diagram of the characteristic parameters ρ1 and ρ2 provided in Embodiment 2 of the present invention;

[0029] Figure 6(b) is the characteristic interaction diagram of characteristic parameters ρ1 and ρ3 provided in Embodiment 2 of the present invention;

[0030] Figure 7 is the characteristic parameter t provided in Embodiment 2 of the present invention L and t U of the characteristic interaction diagram;

[0031] Figure 8 is the characteristic interaction diagram of characteristic parameters ρ2 and ρ3 provided in Embodiment 2 of the present invention;

[0032] Figure 9(a) is the characteristic interaction diagram of characteristic parameters t1 and t2 provided in Embodiment 2 of the present invention;

[0033] Figure 9(b) is the characteristic interaction diagram of characteristic parameters t1 and t3 provided in Embodiment 2 of the present invention;

[0034] Figure 10 is the characteristic interaction diagram of characteristic parameters t2 and t3 provided in Embodiment 2 of the present invention;

[0035] Figure 11 is the comparison diagram of the optimization process provided in Embodiment 2 of the present invention. Detailed implementation manners

[0036] The following combines the accompanying drawings to specifically describe the preferred embodiments of the present invention. Among them, the accompanying drawings form a part of this application and are used together with the embodiments of the present invention to explain the principle of the present invention, and are not used to limit the scope of the present invention.

[0037] Embodiment 1

[0038] A specific embodiment of the present invention discloses a method for optimizing the performance of a multi-layer protection structure based on the SHAP technology, as Figure 1 shown, including the following steps:

[0039] S1. Obtain the original optimization conditions of the multi-layer protection structure to be optimized; wherein, the original optimization conditions include an objective function and constraint conditions, and the constraint conditions are constructed by the characteristic parameters of the multi-layer protection structure to be optimized.

[0040] Specifically, the constraint conditions include fixed constraint conditions and adjustable constraint conditions; wherein, the adjustable constraint conditions are the value range constraints of each characteristic parameter. It should be noted that the constraints other than the adjustable constraint conditions are classified as fixed constraint conditions.

[0041] Specifically, the objective function is the minimization of the target performance. It can be understood that the target performance may have maximization and minimization. In this embodiment, it is uniformly minimized. If it is maximization, a negative sign can be added in front of it to achieve unification.

[0042] Specifically, the original optimization conditions are obtained according to the actual optimization task. The optimization index is determined according to the optimization task of the multi-layer protection structure to be optimized. The objective function is constructed according to the performance index to be achieved by the optimization index. According to the influencing parameters of the performance in the objective function, each characteristic parameter is determined, and the constraint conditions are constructed.

[0043] Exemplarily, the target performance in the objective function can be panel deflection, total structure mass, and structure energy absorption. The characteristic parameters can be panel thickness, panel density, core layer thickness, core layer density, and external load.

[0044] It should be noted that the objective function in this embodiment is set for one kind of target performance. If there are multiple target performances, each step needs to be repeated.

[0045] S2. Based on the original optimization conditions of the multi-layer protection structure to be optimized, a sample data set is constructed, and then the constructed surrogate model is trained to obtain a trained surrogate model; wherein, each sample data includes the specific values of each characteristic parameter and the specific value of the corresponding objective function.

[0046] Specifically, a corresponding test plan is constructed through the original optimization conditions, and the specific values of the objective function corresponding to different values of each characteristic parameter are obtained through on-site or data tests, thereby generating a certain number of samples to obtain a sample data set. The number of samples is determined according to the optimization task. Preferably, the number of samples is not less than 50.

[0047] Specifically, the structure of the surrogate model in this embodiment is successively an input layer, a first hidden layer, a second hidden layer, and an output layer; the input of the surrogate model is the specific value of each characteristic parameter, and the output is the specific value of the objective function; wherein, the number of neurons in the input layer, the first hidden layer, the second hidden layer, and the output layer are 8, 7, 8, and 1 respectively, and the activation function is Relu.

[0048] It can be understood that in this embodiment, considering the interaction effects among the influencing characteristics and their obvious non-linear relationship with the structure performance characterization index, it can be more efficient and reliable.

[0049] Specifically, the surrogate model is trained in the following manner:

[0050] S21. The sample data set is divided into a training set and a test set, and the optimizer, learning rate, and loss function during the training process are set.

[0051] Exemplarily, the optimizer is Adam, and the learning rate is 1e-6; the training set is 80% of the total number of sample data, and the test set is 20% of the total number of sample data.

[0052] S22. Input each sample in the training set into the surrogate model, and use the backpropagation method to adjust the model weights and bias parameters until the error between the predicted value and the actual value of the model is within the set range, obtaining a preliminarily trained surrogate model.

[0053] S23. Based on the test set, test the preliminarily trained surrogate model, and obtain a trained surrogate model after passing the test.

[0054] Specifically, the test set is used to test the performance of the trained surrogate model. If the surrogate model performs well on the training set but poorly on the test set, it indicates that the surrogate model is overfitting. At this time, based on the sample data set, retrain the surrogate model, and monitor the overfitting situation of the model through the test set.

[0055] Specifically, the loss function RMSE is expressed as:

[0056]

[0057] In the formula, H is the number of samples in the training set, and δ h 、 respectively represent the actual specific value and the predicted specific value in the objective function.

[0058] Preferably, before training the surrogate model, in order to make the model converge faster, perform a normalization operation on the sample data set; normalize the elements in the sample data set through the following method:

[0059]

[0060] In the formula, represents the element in the sample data set after normalization, x represents the element in the sample data set before normalization, μ represents the average value of the sample data set before normalization, and σ represents the standard deviation of the sample data set before normalization.

[0061] It should be noted that if normalization is performed before training the surrogate model, when using the surrogate model in subsequent operations, its input needs to be normalized before input, where the corresponding average value and standard deviation are the average value and standard deviation of the sample data set.

[0062] S3. Based on the sample data set, the surrogate model and the SHAP technology, obtain the SHAP characteristics of the multi-layer protection structure to be optimized; wherein, the SHAP characteristics include the SHAP values of each feature parameter in each sample and the global importance of each feature parameter.

[0063] Preferably, the LHS sampling method can also be used to expand the number of samples in the sample data set, and better characterize the coupling mechanism of each parameter inside the model based on the SHAP technology.

[0064] Specifically, the SHAP value of the characteristic parameter in the sample is expressed as:

[0065]

[0066] In the formula, φ i′ represents the SHAP value of the i'-th characteristic parameter in the sample, z' represents the set of characteristic parameters except the i'-th characteristic parameter in the sample, x' represents the set of characteristic parameters in the sample, M represents the number of characteristic parameters, f(z') represents the output prediction value of the surrogate model when the input is z'; f(x') represents the output prediction value of the surrogate model when the input is x', and | | represents the number of characteristic parameters in the set.

[0067] Specifically, the global importance of the characteristic parameter is expressed as:

[0068]

[0069] In the formula, η i′ represents the global importance of the i'-th characteristic parameter, N represents the total number of samples, represents the SHAP value of the i'-th characteristic parameter in the n'-th sample.

[0070] S4. Based on the SHAP characteristics of the multi-layer protection structure to be optimized, the trained surrogate model, and the SHAP technology, obtain the optimal values of the characteristic parameters of the multi-layer protection structure to be optimized, and complete the performance optimization of the multi-layer protection structure.

[0071] Optionally, the optimal values of the characteristic parameters of the multi-layer protection structure to be optimized are obtained by the following method:

[0072] S41. Based on the constraint conditions, classify the characteristic parameters.

[0073] Specifically, classify the characteristic parameters based on the physical meaning of each characteristic parameter and the corresponding constraints; among them, the number of categories can be set according to the requirements of the optimization task. It should be noted that this classification can be achieved by those skilled in the art based on the knowledge in this field.

[0074] S42. Based on the categories of the characteristic parameters and the SHAP characteristics of the multi-layer protection structure to be optimized, update the original optimization conditions to obtain the updated optimization conditions.

[0075] During implementation, the original optimization conditions are updated by the following method:

[0076] Based on the sorting of the global importance of the characteristic parameters, select the characteristic parameters in descending order and execute:

[0077] If there are no other unupdated feature parameters in the category to which the current feature parameter belongs, the value range constraint of the current feature parameter is not updated;

[0078] Otherwise, based on the objective function, the specific values of the current feature parameter in each sample, and the corresponding SHAP values, determine the sensitive range of the current feature parameter, and then update the value range constraint of the current feature parameter.

[0079] Specifically, when implementing, the value range constraint of the current feature parameter is updated in the following way:

[0080] a1. Obtain a feature map based on the specific values of the current feature parameter in each sample and the corresponding SHAP values; among them, the feature map uses the current feature parameter as the horizontal axis and the SHAP value of the current feature parameter as the vertical axis.

[0081] Preferably, based on the feature map, feature interaction maps can also be obtained; among them, other feature parameters in the category to which the current feature parameter belongs are respectively added to each feature interaction map, and the specific values of the added other feature parameters are represented by the colors of the feature points. On the basis of the constraint conditions, combined with the variation relationship between the current feature parameter and each other feature parameter in the category in each feature interaction map, the reliability and accuracy of the database and the proxy model are respectively verified. It should be noted that those skilled in the art can combine the previously established constraint conditions with the SHAP feature interaction map to complete the judgment of the positive and negative correlations between features.

[0082] Exemplarily, as shown in Fig. 6(a), it represents a feature interaction map of a density (main feature) and another density (interaction feature) in the same category. Among them, the horizontal axis is the main feature density, the vertical axis is the SHAP value of the main feature density, and the color represents the interaction feature density. If no color is added in this feature interaction map, it is the feature map; it can be seen from this figure that as the value of the main feature density increases, the color of the interaction feature density changes from red to blue, showing a negative correlation, which conforms to the constraint conditions and the internal mechanism of the structure.

[0083] It can be understood that in this embodiment, the SHAP technology can characterize the accuracy of numerical simulations and sample data sets. For example, when there is a large gap or even contradiction between the SHAP values of the influencing factors in the optimization scheme and the theoretical basis, researchers need to check the accuracy of the original data and model configuration. On the premise of ensuring no errors, it is necessary to combine experimental and numerical simulation methods to verify the results of the optimization scheme. If the existing theory still cannot explain the cause mechanism after verification, it is necessary to re-examine the relevant theoretical framework to better understand the internal physical mechanism of this phenomenon and provide more options for the later structure performance improvement scheme.

[0084] a2. Based on the objective function and the feature map, determine the sensitive range of the current feature parameter.

[0085] Specifically, the sensitive range of the current characteristic parameter is determined in the following manner:

[0086] Based on the objective function, select the region in the feature map where the SHAP value is less than or equal to 0, and then obtain the ratio of the number of feature points in the selected region to the number of feature points in the feature map; that is, the ratio of the number of all sample points in the selected region of the feature map to the number of all feature points in the entire feature map is the feature point ratio.

[0087] If the feature point ratio is less than the set first ratio threshold, then use the value range in the value range constraint of the current characteristic parameter as the sensitive range of the current characteristic parameter;

[0088] If the feature point ratio is greater than or equal to the set first ratio threshold, then determine the sensitive range of the current characteristic parameter based on the distribution of the feature points in the selected region.

[0089] More specifically, the determining the sensitive range of the current characteristic parameter based on the distribution of the feature points in the selected region includes:

[0090] a21. Cluster the feature points in the selected region to obtain each set of dense feature points. Among them, each set of dense feature points is obtained by the density clustering method based on the feature points in the selected region.

[0091] Exemplarily, the density clustering method is the DBSCAN algorithm, and the parameters in the DBSCAN algorithm are set according to specific requirements. The parameters include the neighborhood radius and the minimum number of points in the neighborhood.

[0092] a22. Based on each feature point in each set of dense feature points, obtain the value range of the current characteristic parameter of each set of dense feature points. Among them,

[0093] By obtaining the minimum value and the maximum value of the current characteristic parameter among each feature point in the set of dense feature points, the value range of the current characteristic parameter of the set of dense feature points is formed.

[0094] a23. Use the union of the value ranges of the current characteristic parameter of each set of dense feature points as the sensitive range of the current characteristic parameter.

[0095] a3. Take the intersection of the sensitive range of the current characteristic parameter and the value range in the value range constraint of the current characteristic parameter in the original optimization condition to obtain the updated value range constraint of the current characteristic parameter.

[0096] S43. Based on the updated optimization condition, the trained surrogate model, and the constructed optimization algorithm, obtain the optimal values of the characteristic parameters of the multi-layer protection structure to be optimized.

[0097] It should be noted that the above-mentioned method for updating the original optimization conditions is applicable to objective functions with multiple target performance requirements. The process from step S2 to this point is repeated according to different target performances to obtain the updated optimization conditions respectively, and the intersection of the updated optimization conditions is taken as the updated optimization condition.

[0098] More specifically, the optimization algorithm uses the swarm-based metaheuristic algorithm SSA.

[0099] Exemplarily, the population size N and the maximum number of iterations T of the SSA optimization algorithm are determined according to the computing resources, and the dimension D (the number of features to be optimized) and the fitness function f (the objective function) of the individuals in the SSA optimization algorithm population are set in combination with the optimization task. In addition, the producer ratio P, the proportion SD of the population perceiving danger, and the safety threshold ST involved in the algorithm all adopt the recommended values, which are 0.2, 0.1, and 0.8 respectively.

[0100] Initialize and assign values to the positions of the individuals in the population, that is, determine X i =(x i1 ,x i2 ,…,x iD ), and calculate the fitness values of each population based on the fitness function; where, X i represents the i-th population, and x i1 ,x i2 ,…,x iD represent the 1st to D-th individuals in the population respectively.

[0101] Update the position of the producer in the population based on the following formula:

[0102]

[0103] In the formula, represents the value of the j-th individual in the i-th population in the (t + 1)-th iteration; t is the current iteration number; α is a random number in the range of (0, 1]; R1 is a random number used to indicate whether a predator is found; Q is a random number following a Gaussian distribution; L is a matrix of dimension 1×D composed of elements 1.

[0104] Update the position of the beggar in the population based on the following formula:

[0105]

[0106] In the formula, represents the position of the individual with the lowest fitness value within the t-th iteration; represents the position of the individual with the highest fitness value within the (t + 1)-th iteration; A + =A T (AA T )-1 , where A is a 1×D matrix composed of elements 1 or -1.

[0107] Update the positions of some individuals in the population that can perceive external threats based on the following formula:

[0108]

[0109] In the formula, represents the position of the individual with the highest fitness during the t-th iteration; β is used to adjust the step size and follows a standard Gaussian distribution; K is a random number within the range [-1, 1]; ε is an infinitesimal; f i , f g and f w represent the fitness values of the i-th, optimal, and worst individuals respectively.

[0110] It can be understood that by obtaining the optimal values of each characteristic parameter in the above manner, when optimizing the performance of the multi-layer protection structure, it can meet the requirements of higher optimization performance improvement.

[0111] It can be understood that when optimizing the performance of multi-layer and multi-parameter structures such as gradients, honeycombs, and grids, it is generally necessary to first construct a sample data set of the surrogate model through on-site experiments and numerical simulations. However, due to the lack of a full understanding of the internal mechanism between features and optimization objectives at the initial stage of structural optimization, as many relevant influencing factors as possible will be considered, which will lead to problems such as long numerical simulation calculation time, large storage space for solution data, and high difficulty in convergence of the surrogate model. In addition, since traditional research methods can only give the qualitative importance comparison results of the same type of features, there is a lack of a full understanding of the optimization direction of structural performance. In view of this, the optimization method proposed in this embodiment can quickly obtain the quantitative ranking and positive and negative correlation information of physically comparable feature importance of different categories based on the SHAP technology, so as to filter out feature factors with high cost and low influence, ensure the scientific and reasonable selection of features, and achieve the purpose of accelerating the model convergence speed, improving the model robustness, and improving the cost performance of the optimization scheme. The above optimization method in this embodiment jointly uses the SHAP value and the optimization algorithm to specifically adjust the number and search range of influencing factors in the constraint conditions, that is, to achieve the purpose of accelerating the convergence speed, avoiding local optima, and improving the quality of the optimization results from two dimensions of feature quantity and range.

[0112] Optionally, the optimal values of each characteristic parameter of the multi-layer protection structure to be optimized can also be obtained through the following method:

[0113] SP1. Based on the constraint conditions, classify each characteristic parameter; and obtain the global correlation between each characteristic parameter and the SHAP value based on the SHAP value of each characteristic parameter in each sample.

[0114] Specifically, the global correlation includes positive correlation and negative correlation; the global correlation is obtained by analyzing the relationship between the transformation of the specific values of each characteristic parameter and the transformation of the corresponding SHAP value.

[0115] SP2. Based on the constraint conditions, obtain the specific values of each characteristic parameter of the multi-layer protection structure to be optimized under the initial setting.

[0116] Specifically, the initial setting of each characteristic parameter is the setting of any set of specific values of each characteristic parameter under the constraint conditions.

[0117] SP3. Input the specific values of the current characteristic parameters into the trained surrogate model to obtain the specific value of the objective function, and then obtain the current SHAP values of each characteristic parameter based on the SHAP technology.

[0118] Specifically, the current SHAP values of each characteristic parameter can be visualized through a waterfall plot, which is clearer and more understandable.

[0119] SP4. Based on the current SHAP values, global importance, global correlation, and constraint conditions of each adjustable characteristic parameter, adjust the specific values of each adjustable characteristic parameter; among them, the states of each characteristic parameter are adjustable characteristic parameters and non-adjustable characteristic parameters.

[0120] Specifically, all characteristic parameters are initially initialized as adjustable characteristic parameters.

[0121] Specifically, the specific values of each adjustable characteristic parameter are adjusted in the following way:

[0122] SP41. Based on the current SHAP values and global importance of each adjustable characteristic parameter, determine the characteristic parameter to be adjusted.

[0123] More specifically, select the characteristic parameter corresponding to the maximum value of the current SHAP value as the characteristic parameter to be adjusted; if the characteristic parameters corresponding to the maximum value of the current SHAP value are not unique, then select the characteristic parameter with the greatest global importance among them as the characteristic parameter to be adjusted.

[0124] SP42. Determine whether there are adjustable characteristic parameters in the category to which the characteristic parameter to be adjusted belongs whose global importance is lower than that of the characteristic parameter to be adjusted; if so, adjust the characteristic parameter to be adjusted based on the current SHAP value, global correlation, and constraint conditions of the characteristic parameter to be adjusted; if not, the characteristic parameter to be adjusted is not adjusted, and the characteristic parameter to be adjusted is used as a non-adjustable characteristic parameter.

[0125] More specifically, the adjustment of the characteristic parameter to be adjusted based on the current SHAP value, global correlation, and constraint conditions of the characteristic parameter to be adjusted includes:

[0126] If the global correlation of the feature parameter to be adjusted is negatively correlated, increase the specific value of the feature parameter to be adjusted based on the set change amount, and adjust the specific value of the adjustable feature parameter with the smallest global importance in the category to which the feature parameter to be adjusted belongs based on the adjusted specific value of the feature parameter to be adjusted and the constraint conditions; wherein,

[0127] If the current specific value of the parameter to be adjusted or the corresponding adjustable feature parameter exceeds the corresponding value range, the specific values of the feature parameter to be adjusted and the corresponding adjustable feature parameter are not adjusted, and the feature parameter to be adjusted is used as a non-adjustable feature parameter.

[0128] More specifically, the adjustment of the feature parameter to be adjusted based on the current SHAP value, global correlation and constraint conditions of the feature parameter to be adjusted includes:

[0129] If the global correlation of the feature parameter to be adjusted is positively correlated, reduce the specific value of the feature parameter to be adjusted based on the set change amount, and adjust the specific value of the adjustable feature parameter with the smallest global importance in the category to which the feature parameter to be adjusted belongs based on the adjusted specific value of the feature parameter to be adjusted and the constraint conditions; wherein,

[0130] If the current specific value of the parameter to be adjusted or the corresponding adjustable feature parameter exceeds the corresponding value range, the specific values of the feature parameter to be adjusted and the corresponding adjustable feature parameter are not adjusted, and the feature parameter to be adjusted is used as a non-adjustable feature parameter.

[0131] It can be understood that when increasing or decreasing the specific value of the feature parameter, it is necessary to correspondingly decrease or increase the feature parameters in the relevant constraints, that is, other feature parameters in the same category, according to the constraint conditions to ensure that the adjusted value still satisfies the constraint conditions.

[0132] SP5. Determine whether the optimization termination condition is reached. If so, use the current specific values of each feature parameter as the optimal values of each feature parameter. If not, repeat steps SP3 to SP5.

[0133] Specifically, the optimization termination condition includes that the current SHAP values of each feature parameter are all less than or equal to 0; or the number of adjustable feature parameters is 0.

[0134] It can be understood that in this embodiment, considering that the optimization algorithm has a long iterative optimization time and the optimization direction has randomness and non-explainability, it is proposed that only based on the surrogate model and SHAP technology, a clear and definite personalized structural performance optimization scheme can be quickly obtained, thereby accelerating the structural performance optimization cycle while improving the interpretability of the surrogate model and the optimization scheme. In the optimization process, a deeper understanding of the interaction mechanism between the various features and performance indicators of the structure can be obtained, and the requirements for the timeliness of the optimization cycle can be met.

[0135] Compared with the prior art, an optimization method for the performance of a multi-layer protection structure based on the SHAP technology provided in this embodiment uses the SHAP technology to obtain the SHAP values of each characteristic parameter, which are used to characterize the global and local effects of each characteristic parameter in the sample. The global importance ranking result formed based on this can not only make each characteristic parameter physically comparable, but also filter out the characteristic parameters with high cost and low impact, thereby improving the generalization performance of the surrogate model and the optimization efficiency and quality of the optimization algorithm while ensuring the scientific and reasonable selection of characteristics; before the algorithm optimization, the sensitive range of each characteristic parameter is obtained based on the SHAP technology, so as to formulate reasonable and efficient algorithm constraint conditions, which can reduce the probability of the algorithm falling into the local optimum, accelerate the algorithm optimization speed and improve the optimization quality; the provided optimization method without using the optimization algorithm can obtain a clear and interpretable adjustment method in real time through the SHAP technology, and complete the structural performance optimization task in an extremely short time. The optimization cycle is short and the optimization scheme has clear interpretability, which can improve the trust in the prediction results of the surrogate model.

[0136] Embodiment 2

[0137] To verify the correctness and effectiveness of the method in Embodiment 1, a specific Embodiment 2 is provided. Taking the widely used aluminum foam sandwich panel as an example, the structural performance optimization task under explosion is carried out. On the premise that the mass of the multi-layer gradient design is equal to that of the homogeneous sandwich structure, the overall performance of the structure can be further improved, so a three-layer aluminum foam sandwich structure is selected as the optimization target. By combining the existing research results and the specific characteristics of this optimization task, the characteristic parameters are finally selected as the TNT equivalent, explosion distance, thickness of the front and rear panels, and thickness and density of the three-layer sandwich structure, as Figure 2 shown.

[0138] To simplify the optimization task, in this embodiment, the values of the TNT equivalent and the explosion distance are limited to 60 g and 100 mm respectively. In addition, it is stipulated that the side lengths of the front and rear panels and the three-layer sandwich structure are all 300 mm, and the thickness value ranges are 0.5 mm to 1.5 mm and 6 mm to 12 mm respectively. Limited by the manufacturing process of the aluminum foam sandwich panel, the density value of the three-layer sandwich structure is 0.20 g / cm 3 ~0.70 g / cm 3After determining the values of each characteristic parameter, the requirements for this optimization task are formulated as follows: on the premise that the total thickness of the front and rear panels of the structure, the total thickness of the internal sandwich structure, and the total mass are 2 mm, 30 mm, and 1215 g ± 10 g respectively, the optimized configuration data of the sandwich structure are given to minimize the damage of the external load to the protected target. It should be noted that since the protected target is often connected to the rear panel of the sandwich structure, the maximum deflection δ at the center of the rear panel is used as the characterization index of the structural performance in this optimization task, that is, the maximum position point of the displacement time history curve of the rear panel under the action of the external load, as shown in Figure 3 shown. Thus, the original optimization conditions are obtained, expressed as:

[0139]

[0140] After determining the optimization goal, 56 groups of effective samples are collected in this embodiment through on-site tests and numerical simulation techniques, and they are divided into a training set and a test set in a ratio of 8:2. The training set is only used to adjust the parameters of the BPNN model, and the test set is only used to test the performance of the model. During the training process of the model, the hyperparameters of the model are debugged through the grid search technique in this example. The hyperparameters of the surrogate model are debugged through the above method, and the BPNN model with the best performance is finally selected.

[0141] The characteristic parameters are classified. Since the total thickness of the front and rear is limited to 2 mm in the constraint conditions, the thickness t of the front and rear panels U and t L are classified as one category. Similarly, the three-layer sandwich panel can be classified as the second category under the constraint of a total thickness of 30 mm. The physical meaning of the density of the three-layer sandwich panel is significantly different from the thickness, so ρ1, ρ2, and ρ3 are used as the third category.

[0142] In order to better characterize the coupling mechanism of each parameter inside the model based on the SHAP technology, 100 groups of samples are generated using the LHS sampling method on the basis of the existing data. Calculate the SHAP values of each feature in all samples, and then obtain the global importance of each feature, as shown in Figure 4(a). It can be seen from the figure that among the 8 selected influencing factors, the density ρ1 of the first-layer sandwich structure has the most significant influence on the deflection of the rear panel, far greater than the similar features ρ2 and ρ3. The influence of the thickness t of the rear panel L is second only to ρ1, but is greater than the thickness t of the front panel of the similar feature Uhas a greater influence. For the thickness of the three-layer sandwich structure, the importance ranking is: t1>t2>t3. In addition, it is not difficult to find from Figure 4(b) that all influencing factors as a whole show a negative correlation with the rear panel deflection, that is, the protective performance of the structure will decrease with the increase of each characteristic value, which is also more in line with the existing relevant theories. At the same time, it also proves that the database collected in this embodiment is accurate, and the surrogate model trained based on it can be used for subsequent structural performance optimization tasks.

[0143] Optionally, the optimal values of each characteristic parameter are determined by the following method:

[0144] Since the constraint conditions limit the total thickness of the front and rear panels, the total thickness of the internal sandwich structure, the density of the sandwich panel, and the total mass, the corresponding uniform aluminum foam sandwich panel is selected as the initial setting: the thickness of the upper and lower panels is 1 mm each, and the thickness and density of the internal sandwich structure are 10 mm and 0.45 g / cm respectively 3 .

[0145] In the initial stage of optimization, first, based on the BPNN model, predict the rear panel deflection δ of the uniform sandwich structure under the action of the explosion load, and then use the SHAP technique to calculate the SHAP values of each characteristic parameter under this working condition and visualize them through a waterfall plot, as shown in Figure 5(a). It is not difficult to find that among the 8 selected characteristic parameters, the density ρ1 of the first-layer sandwich structure is the main reason for the increase in the rear panel deflection δ. According to the global correlation analysis, ρ1 and δ are negatively correlated. Therefore, if δ is to be reduced, ρ1 should be increased. Among them, the increase amount of the density ρ1 of the first-layer sandwich structure is determined by the following values of the density ρ1: 0.20, 0.29, 0.37, 0.45, 0.53, 0.61, 0.70. It should be noted that in order to meet the requirement of the total mass being 1205 g to 1225 g in the constraint conditions, when ρ1 is increased to 0.53, ρ3, which is the smallest in global importance among the same category of characteristics, should be reduced to 0.37.

[0146] After the first optimization is completed, calculate the SHAP values of each adjusted characteristic parameter again based on the SHAP technique and visualize them using a waterfall plot, as shown in Figure 5(b). Under the same external load, compared with the uniform sandwich structure, the rear panel deflection δ of the optimized structure decreases from 20.35 mm to 18.12 mm. In addition, the main characteristic that causes a large rear panel deflection δ under the current working condition becomes the density ρ2 of the second-layer sandwich structure. Since ρ2 and δ are also negatively correlated, ρ2 needs to be increased to 0.53 g / cm 3 . At the same time, continue to reduce ρ3, which is the smallest in global importance among the same category of characteristics, to 0.29 g / cm 3 to meet the constraint conditions.

[0147] Figure 5(c) shows the waterfall plot after the second parameter adjustment. It can be found that the deflection δ of the rear panel of the structure under the current working condition has been reduced to 17.49 mm. At this time, the characteristic that has a greater impact on δ is still the density ρ2 of the second-layer sandwich structure. Therefore, ρ2 is continued to be increased to 0.61 g / cm 3 , and accordingly ρ3 is adjusted to 0.20 g / cm 3 . The optimized structure waterfall plot is shown in Figure 5(d). It can be found that the SHAP values of each characteristic under the current working condition are all less than or equal to 0, indicating that no parameter promotes δ. Therefore, the final optimized parameter values of this method are: t L = t U = 1 mm, t1 = t2 = t3 = 10 mm, ρ1 = 0.53 g / cm 3 , ρ2 = 0.61 g / cm 3 , ρ3 = 0.20 g / cm 3 .

[0148] Optionally, the optimal values of each characteristic parameter are obtained by the following method:

[0149] Based on the specific values and SHAP values of each characteristic parameter in each sample in the sample dataset, a characteristic interaction diagram of each characteristic parameter is drawn. Considering that the global importance of ρ1 is the greatest, the characteristic interaction diagram of this characteristic parameter and other characteristic parameters of the same type is first characterized, as shown in Figure 6. It can be found through analysis that because the total mass of the sandwich layer is restricted when setting the optimization goal, there is a strong interaction between ρ1 and ρ2 and ρ3, that is, as the value of ρ1 becomes larger and larger, the characteristic values corresponding to ρ2 and ρ3 gradually fade, proving that this method can accurately characterize the interaction mechanism inside the model. In addition, from Figure 6(a) or Figure 6(b), the sensitive range of the first-layer density ρ1 is: 0.46 g / cm 3 ~ 0.70 g / cm 3 .

[0150] Similarly, Figure 7 shows the characteristic interaction diagram of the rear panel thickness t L and the front panel thickness t U of the same type of characteristics. The sensitive range of t L can be obtained as: 1 mm to 1.5 mm.

[0151] Since this optimization task involves the density of three layers of sandwich in total, the value range of ρ2 needs to be adjusted accordingly. Figure 8 shows the characteristic interaction diagram between this characteristic parameter and the characteristic parameter ρ3 of the same category. It is not difficult to find that under the restriction of the constraint conditions, there is also a negative correlation between ρ2 and ρ3. It can be obtained from the figure that the sensitive range of ρ2 is: 0.20 - 0.39 g / cm 3 or 0.51 g / cm 3~0.70 g / cm 3 。

[0152] Among the three-layer thicknesses of the sandwich structure, t1 and t2 play a more important role in predicting δ. Therefore, Figures 9 and Figure 10 respectively give the characteristic interaction diagrams of t1 and t2 with other characteristic parameters of the same type. It can be obtained from Figure 9 that the sensitive range of t1 is: 7.78 mm to 12 mm. It can Figure 10 also be easily obtained that the sensitive range of t2 is: 8.77 mm to 12 mm.

[0153] In summary, the updated optimization conditions are obtained, expressed as:

[0154]

[0155] Before using the meta-heuristic algorithm SSA for optimization, the following settings are made for the optimization algorithm: To ensure the diversity of the population and reduce the probability of the traditional optimization algorithm falling into a local optimum, the population size is set to 100; the maximum number of iterations is set to 1000 to ensure that the algorithm can fully converge; the producer ratio, the population ratio of perceiving danger, and the safety threshold are 0.2, 0.1, and 0.8 respectively; the individual dimension is 8.

[0156] Figure 11 shows the iterative convergence process during the optimization process of the SSA optimization algorithm under two optimization conditions. It can be seen from Figure 11 that the traditional optimization algorithm has an extremely slow convergence speed at the beginning of iteration. The deflection value only starts to decrease rapidly after 352 iterations and gradually stabilizes after 475 iterations. The optimal solution finally given is 13.72 mm. Compared with the traditional optimization algorithm, the optimization strategy proposed in the present invention can enable the optimization algorithm to identify solutions with lower deflections at the beginning of iteration and search for an optimization solution with a deflection value of 13.22 mm at the 209th iteration. After that, the deflection value remains unchanged, proving that it can fully converge after 209 iterations.

[0157] Table 1 gives the data information of the optimal solutions of the above two optimization methods. Compared with the traditional algorithm, the optimization solution obtained by the optimization method in Example 1 not only has a smaller deflection value of the rear panel, but also shows a trend of "dense on the top and sparse on the bottom" in density distribution, which is also consistent with the law obtained from the existing field test results.

[0158] In summary, the optimization method proposed in this embodiment can not only help the optimization algorithm converge quickly, but also provide high-quality optimization solutions that conform to the physical laws of the experiment with fewer iteration steps.

[0159] Table 1 Comparison of Optimization Solutions

[0160]

[0161] Those skilled in the art can understand that all or part of the processes of implementing the methods of the above embodiments can be completed by instructing relevant hardware through a computer program, and the program can be stored in a computer-readable storage medium. Among them, the computer-readable storage medium is a magnetic disk, an optical disk, a read-only memory or a random access memory, etc.

[0162] As mentioned above, only the preferred specific embodiments of the present invention are described, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed by the present invention should be covered by the protection scope of the present invention.

Claims

1. A multi-layer protection structure performance optimization method based on SHAP technology, characterized in that: The following steps are involved: Obtaining original optimization conditions of the multi-layer protective structure to be optimized; wherein the original optimization conditions include an objective function and constraint conditions, and the constraint conditions are constructed by characteristic parameters of the multi-layer protective structure to be optimized; Based on the original optimization conditions of the multi-layer protection structure to be optimized, a sample data set is constructed, and then the constructed proxy model is trained to obtain a trained proxy model; wherein each sample data includes specific values ​​of each characteristic parameter and a corresponding specific value of the objective function; Based on the sample data set, the proxy model and the SHAP technology, the SHAP characteristics of the multi-layer protection structure to be optimized are obtained; wherein the SHAP characteristics include the SHAP value of each characteristic parameter in each sample and the global importance of each characteristic parameter; Based on the SHAP characteristics of the multi-layer protection structure to be optimized, the trained proxy model and the SHAP technology, the optimal values ​​of the characteristic parameters of the multi-layer protection structure to be optimized are obtained, and the performance optimization of the multi-layer protection structure is completed.

2. The multi-layer protection structure performance optimization method based on SHAP technology according to claim 1 is characterized in that: The optimal values ​​of the characteristic parameters of the multi-layer protective structure to be optimized are obtained by: Based on the constraints, each characteristic parameter is classified into categories; Based on the categories of each characteristic parameter and the SHAP characteristics of the multi-layer protection structure to be optimized, the original optimization conditions are updated to obtain updated optimization conditions; Based on the updated optimization conditions, the trained proxy model and the constructed optimization algorithm, the optimal values ​​of the characteristic parameters of the multi-layer protection structure to be optimized are obtained.

3. The multi-layer protection structure performance optimization method based on SHAP technology according to claim 2 is characterized in that: The constraint conditions include fixed constraint conditions and adjustable constraint conditions; wherein the adjustable constraint conditions are the value range constraints of each characteristic parameter; the original optimization conditions are updated in the following manner: Based on the global importance ranking of each feature parameter, select the feature parameters from large to small and execute: If there are no other unupdated feature parameters in the category to which the current feature parameter belongs, the value range constraint of the current feature parameter will not be updated; Otherwise, based on the objective function, the specific value of the current feature parameter in each sample and the corresponding SHAP value, the sensitive range of the current feature parameter is determined, and then the value range constraint of the current feature parameter is updated.

4. The multi-layer protection structure performance optimization method based on SHAP technology according to claim 3 is characterized in that: Update the value range constraints of the current feature parameters in the following ways: According to the specific value of the current feature parameter in each sample and the corresponding SHAP value, a feature map is obtained; wherein the feature map takes the current feature parameter as the horizontal axis and the SHAP value of the current feature parameter as the vertical axis; Determine the sensitive range of the current feature parameter based on the objective function and the feature map; The sensitive range of the current feature parameter and the value range in the current feature parameter value range constraint in the original optimization condition are intersected to obtain the updated current feature parameter value range constraint.

5. The multi-layer protection structure performance optimization method based on SHAP technology according to claim 4 is characterized in that: The objective function is minimization of target performance; The sensitivity range of the current characteristic parameter is determined by: Based on the objective function, a region in the feature map where the SHAP value is less than or equal to 0 is selected, thereby obtaining a ratio of the number of feature points in the selected region to that in the feature map; If the feature point quantity ratio is less than the set first ratio threshold, the value range in the current feature parameter value range constraint is used as the sensitive range of the current feature parameter; If the feature point quantity ratio is greater than or equal to the set first ratio threshold, the sensitive range of the current feature parameter is determined based on the distribution of the feature points in the selected area.

6. The multi-layer protection structure performance optimization method based on SHAP technology according to claim 5 is characterized in that: The step of determining the sensitive range of the current feature parameter based on the distribution of feature points in the selected area includes: Clustering is performed based on the feature points in the selected area to obtain a dense set of feature points; Based on each feature point in each feature point dense set, a value range of a current feature parameter of each feature point dense set is obtained; The union of the value ranges of the current feature parameters of the dense sets of feature points is taken as the sensitive range of the current feature parameters.

7. The multi-layer protection structure performance optimization method based on SHAP technology according to claim 6 is characterized in that: Based on the feature points in the selected area, a dense set of feature points is obtained through density clustering.

8. The multi-layer protection structure performance optimization method based on SHAP technology according to claim 7 is characterized in that: The minimum value and the maximum value of the current feature parameter in each feature point in the feature point dense set are obtained to form the value range of the current feature parameter of the feature point dense set.

9. The multi-layer protection structure performance optimization method based on SHAP technology according to claim 1 is characterized in that: The optimal values ​​of the characteristic parameters of the multi-layer protective structure to be optimized are also obtained in the following way: SP1. Based on the constraints, each feature parameter is classified into categories; and based on the SHAP value of each feature parameter in each sample, the global correlation between each feature parameter and the SHAP value is obtained; SP2. Based on the constraint conditions, obtain the specific values ​​of the characteristic parameters of the multi-layer protection structure to be optimized under the initial settings; SP3, input the specific values ​​of the current feature parameters into the trained proxy model to obtain the specific value of the objective function, and then obtain the current SHAP value of each feature parameter based on the SHAP technology; SP4. Based on the current SHAP value, global importance, global correlation and constraint conditions of each adjustable feature parameter, the specific value of each adjustable feature parameter is adjusted; wherein the state of each feature parameter is an adjustable feature parameter and a non-adjustable feature parameter. SP5. Determine whether the optimization termination condition is reached. If so, take the current specific value of each feature parameter as the optimal value of each feature parameter. If not, repeat steps SP3 to SP5.

10. According to the multi-layer protection structure performance optimization method based on SHAP technology as claimed in claim 9, the specific value of each adjustable characteristic parameter is adjusted by the following method: Determine the feature parameters to be adjusted based on the current SHAP value and global importance of each adjustable feature parameter; Determine whether there is an adjustable feature parameter whose global importance is lower than the global importance of the feature parameter to be adjusted in the category to which the feature parameter to be adjusted belongs; if so, adjust the feature parameter to be adjusted based on the current SHAP value, global correlation and constraint conditions of the feature parameter to be adjusted; if not, do not adjust the feature parameter to be adjusted, and treat the feature parameter to be adjusted as a non-adjustable feature parameter.

Images