Protective material performance analysis method based on double-branch time sequence model

Abstract

The invention provides a protective material performance analysis method based on a double-branch time sequence model, and relates to the technical field of material performance analys.The method comprises the steps that a standard penetration model is built in finite element simulation software, and a penetration time sequence data set is made by training the standard penetration model; according to the penetration time sequence data set, a double-branch time sequence model is trained; and utilizing the trained double-branch time sequence model to obtain an analysis result of the performance of the protective material. According to the method, the problems of less quantitative analysis, low efficiency and insufficient accuracy when complex time sequence data is processed in a traditional material performance analysis method are solved. According to the method, a double-branch time sequence model based on two-stage training and a loss function suitable for time sequence data are introduced, a research method combining deep learning and finite element simulation analysis is explored, the analysis efficiency is remarkably improved, meanwhile, more comprehensive quantitative analysis is provided, and powerful technical support is provided for design, optimization and evaluation of protective materials.

Description

Technical Field

[0001] The present invention belongs to the technical field of material property analysis, and particularly relates to a method for analyzing the performance of protective materials based on a dual-branch time series model. Background Art

[0002] When exploring complex engineering problems, the penetration process and its analysis play a crucial role. Penetration, in short, refers to the process in which a rigid object (such as a projectile) penetrates another object (such as a target) at a certain speed. This process not only involves the dynamic interaction between the projectile and the target, but also covers multiple aspects such as material deformation, energy dissipation, and possible structural damage. Penetration analysis is to quantitatively describe and predict the physical phenomena involved in the penetration process through mathematical models and simulation techniques, in order to evaluate the penetration ability of the projectile, the protection performance of the target, and the possible consequences.

[0003] However, traditional finite element simulation analysis (FEA) often seems inadequate when dealing with highly nonlinear and complex problems such as penetration. Although finite element simulation can meticulously simulate the stress and deformation process of objects, its calculation process is complex and time-consuming. Especially when dealing with large-scale data sets or high-complexity models, the low calculation efficiency becomes a key factor restricting its wide application. In addition, for the massive time series data generated during the penetration process, traditional finite element simulation is difficult to efficiently capture and utilize the implicit information in these data, such as penetration speed, energy changes, and the dynamic relationship between material parameters.

[0004] In addition, although deep learning technology has made remarkable progress in recent years, especially showing strong capabilities in dealing with large-scale data sets and complex pattern recognition, existing deep learning methods still have deficiencies in combining the time series data of the penetration process with the material parameters of the penetrator and the penetrated object. Although deep learning models can learn the time series characteristics during the penetration process, when dealing with a dynamic process such as penetration that highly depends on physical laws, they often ignore the important influence of material parameters on the penetration behavior.

[0005] Existing research rarely conducts quantitative analysis on the protection performance of protective materials. It often only conducts qualitative analysis, without giving the degree of influence of a certain parameter of the protective material on the protection performance or without giving the comprehensive degree of influence among the various parameters of the protective material on the protection performance. Summary of the Invention

[0006] Aiming at the above deficiencies in the prior art, the present invention provides a method for analyzing the performance of protective materials based on a dual-branch time series model, which solves the problems of less quantitative analysis, low efficiency, and insufficient accuracy in traditional material performance analysis methods when dealing with complex time series data.

[0007] To achieve the above object, the technical solution adopted by the present invention is: a method for analyzing the performance of a protective material based on a dual-branch time series model, comprising the following steps:

[0008] S1. Construct a standard penetration model in finite element simulation software, and through training it, produce a penetration time series data set;

[0009] S2. Train a dual-branch time series model according to the penetration time series data set;

[0010] S3. Use the trained dual-branch time series model to obtain the analysis result of the performance of the protective material.

[0011] The beneficial effects of the present invention are as follows: By constructing a dual-branch time series model applied to penetration process fitting and parameter analysis, the present invention solves the following problems: First, the finite element simulation is slow and the efficiency in processing a large number of data sets is low; Second, the current deep learning time series model fails to effectively combine the time series data of the penetration process with the material parameters of the penetrator and the penetrated object; Third, qualitative and quantitative analyses are performed on the protection performance of the protective material, and the deep learning model is used to have the ability to learn from a large amount of data and infer the optimal or better parameter combination.

[0012] Furthermore, the step S1 includes the following steps:

[0013] S101. Construct a standard penetration model in finite element simulation software, and adopt the method of automatically batch modifying the parameters of the finite element simulation model by using a python script to generate a total of X finite element simulation models with different parameters;

[0014] S102. Calculate the penetration results of each finite element simulation model. Among them, the penetration results include M time points, and the data of each time point include time, the current speed of the bullet, the current displacement of the bullet, and the total energy of the system. And among the M time points, starting from the time when the bullet contacts the target plate, the time points after the bullet contacts the target plate are selected, and a total of M * time points are used as the effective experimental data for each simulation;

[0015] S103. According to the penetration results, select a reference time series with a length of lookback and a prediction time series with a length of pred as training data and labels respectively. Among them, for the effective experimental data M of a single finite element simulation model * , divide the time series data into N strips, and obtain X×N strips of penetration time series data, where X represents the number of finite element simulation models;

[0016] S104. Randomly sample the training set, validation set, and test set. According to the sampling results, for non-temporal data, it is sequentially corresponding to X×N pieces of penetration temporal data one by one, and the complete penetration temporal data is saved to complete the production of the penetration temporal data set, where the non-temporal data is the target plate material parameters.

[0017] The beneficial effect of the above further solution is that: through the post-processing results obtained by finite element simulation, the present invention constructs the temporal data set used for subsequent deep learning training and protection performance analysis, making preparations for subsequent analysis.

[0018] Furthermore, the step S2 includes the following steps:

[0019] S201. Input the penetration temporal data set into the LSTM network for training to obtain the first output variable y lstm , and calculate the first output variable y lstm and the label value in step S103 according to the loss function;

[0020] S202. Determine whether the number of update times reaches the threshold. If so, enter the second stage of training the dual-branch temporal model; otherwise, return to step S201;

[0021] S203. Freeze the LSTM network, input the penetration temporal data set into the LSTM network to obtain the first output variable y lstm , input the variable p into the KAN network, output the second variable y kan , connect the first variable y lstm and the second variable y kan through residual connection, and calculate the variable y co and the label value in step S103 according to the loss function to complete the training of the dual-branch temporal model.

[0022] The beneficial effect of the above further solution: The phased training of the present invention enables the dual-branch temporal model to fully learn the characteristics of the corresponding data. After the LSTM is trained in the first stage, this part of the network learns the characteristics of the temporal data set; in the second stage, the KAN network learns the influence of different material parameters on the bullet deceleration rate, and this influence is applied to the output of the LSTM network according to the residual connection to obtain the corrected prediction output.

[0023] Furthermore, the expression of the loss function in step S201 is as follows:

[0024]

[0025] Among them, loss represents the loss function, y i and respectively represent the true value at the i-th time point and the predicted value of the dual-branch time series model, n represents the length of the predicted sequence vector, and T(n,i) represents the weight function for the i-th time point in the predicted sequence vector of length n.

[0026] The beneficial effects of the above further scheme are as follows: In the multi-time point prediction task, by assigning different weights to the predicted values at different time points, the prediction accuracy of the dual-branch time series model can be effectively improved. Specifically, the weights of the earlier time points are higher, enabling the dual-branch time series model to pay more attention to the influence of early data, thereby improving the overall performance and stability when processing time series data. This improvement can better reflect the requirements for time sensitivity in practical applications and optimize the loss calculation process.

[0027] Furthermore, the weight function T(n,i) includes a harmonic weight and a normal distribution weight;

[0028] The expression of the harmonic weight is as follows:

[0029]

[0030] The expression of the normal distribution weight is as follows:

[0031]

[0032] Among them, σ represents the variance of the normal distribution, μ represents the mean of the normal distribution, and x represents the index of the i-th time point mapped into the standard normal distribution (0, 3σ).

[0033] The beneficial effects of the above further scheme are as follows: By analyzing the characteristics of different functions, targeted method selection is provided. The stable gradient of the harmonic weight enables it to effectively learn the information of all time points and is suitable for time series data that requires global information; the normal distribution weight has a large gradient change within the range of (0, 3σ), can better focus on the information of the earlier time points, and is suitable for scenarios that require local information. This flexibility enables the model to more effectively adapt to the data characteristics when processing different types of time series data and improve the prediction performance.

[0034] Furthermore, the expression of the residual connection is as follows:

[0035] y co =[tanh(y kan ) + 1] × y lstm

[0036] Among them, y co represents the analysis result of the performance of the protective material.

[0037] The beneficial effects of the above further scheme are as follows: The second variable y kanThe output of the KAN network, representing the influence degree of material parameters on the change of bullet velocity, is scaled to the interval (-1, 1) using the tanh(·) function, so that the coefficient of the first output variable y lstm is within the interval (0, 2), meeting the requirements of residual connection.

[0038] Furthermore, the step S3 includes the following steps:

[0039] S301. Based on the trained dual-branch time series model, save the weights of the KAN network to obtain a weight map;

[0040] S302. Construct a new standard penetration model in the finite element simulation software, select some target plate material parameters to make a new penetration time series dataset for parameter qualitative and quantitative analysis, and randomly assign values to the selected target plate material parameters to obtain multiple new finite element simulation models;

[0041] S303. According to the random assignment results, calculate the penetration results of each new finite element simulation model to obtain a new penetration time series dataset;

[0042] S304. Use the trained dual-branch time series model to predict the new penetration time series dataset and perform regression analysis on the prediction results;

[0043] S305. According to the weight map and the prediction results after regression analysis, draw a parameter heat map to obtain the analysis results of the performance of the protective material.

[0044] The beneficial effects of the above further solution are as follows: The analysis process of the performance of the protective material is optimized through systematic steps. Specifically, the implementation of step S3 can accurately save the weights of the KAN network and generate a weight map based on the trained dual-branch time series model, providing a strong basis for subsequent analysis. The process of constructing a standard penetration model and randomly assigning values to the target plate material parameters enhances the data diversity and ensures that the obtained penetration time series dataset is more representative. By using the trained dual-branch time series model to predict these data and performing regression analysis, the material performance can be evaluated more precisely. Finally, by drawing a parameter heat map based on the weight map and the regression analysis results, the performance distribution of the protective material is visually presented, helping researchers deeply understand the material characteristics and improving the visualization and practicality of the analysis results. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is a flowchart of the method of the present invention.

[0046] Figure 2 It is a schematic diagram of the KAN network structure adopted by the present invention.

[0047] Figure 3Schematic diagram of the LSTM-KAN Embed network model designed for the present invention. Specific implementation manners

[0048] The following describes the specific implementation manners of the present invention to facilitate those skilled in the art of this technology to understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation manners. For those of ordinary skill in the art of this technology, as long as various changes are within the spirit and scope of the present invention defined and determined by the appended claims, these changes are obvious, and all inventions and creations using the concept of the present invention are within the scope of protection.

[0049] Embodiment

[0050] Before explaining the present invention, the following terms are first explained:

[0051] KAN (Kolmogorov-Arnold Networks): Kolmogorov-Arnold network, that is Figure 3 the material information embedding network.

[0052] LSTM-KAN Embed: Long short-term memory-KAN embedding network.

[0053] The basic idea of the present invention is to use the post-processing data of finite element simulation to complete the training of the deep learning model by making a time series dataset with a lookback length, that is, a penetration time series dataset of the lookback type. First, the LSTM network can effectively solve the problems of gradient disappearance and long-term dependence in the RNN, and learn the features in the penetration time series data with higher accuracy; second, the KAN network is good at solving mathematical equations to learn the influence of target plate material parameters on the penetration process; third, after the double-branch network model is trained, the weight visualization ability of the KAN network is used to quantitatively and qualitatively analyze the influence of target plate material parameters on the protection performance. The present invention includes two parts, namely network training and analysis of protection performance parameters. The deep learning networks used are LSTM (long short-term memory network) and KAN network. Among them, LSTM is more common and will not be elaborated here; the KAN network structure is as Figure 2 shown. The KA theorem proposed by the KAN network explains that any set of input variables can obtain the desired output through the nesting of non-linear functions, which is similar to the universal approximation theorem proposed by the MLP network. Different from the MLP network, the KAN network no longer trains the weights of the edges between neurons, but fixes the weights of each edge and trains the non-linear activation functions on each neuron node. The comparison of its model formulas is shown in Formulas (1) and (2):

[0054]

[0055]

[0056] Among them, Φ L represents the non-linear activation function matrix of the L-th layer of the KAN network, and W L represents the connection weight matrix of each neuron in the L-th layer of the MLP network, σ represents the non-linear activation function of the neuron connection, and X all represent the feature matrix.

[0057] As Figure 1 shown, the present invention provides a method for analyzing the performance of a protective material based on a dual-branch time series model, and its implementation method is as follows:

[0058] S1. Build a standard penetration model in the finite element simulation software, and through training it, produce a penetration time series dataset, and its implementation method is as follows:

[0059] S101. Build a standard penetration model in the finite element simulation software, and adopt the method of automatically batch modifying the parameters of the finite element simulation model with a python script to generate a total of X finite element simulation models with different parameters;

[0060] S102. Calculate the penetration results of each finite element simulation model. Among them, the penetration results include M time points, and the data of each time point include time, the current speed of the bullet, the current displacement of the bullet, and the total energy of the system. And among the M time points, starting from the time when the bullet touches the target plate, select the time points after the bullet touches the target plate, a total of M * time points as the effective experimental data for each simulation;

[0061] S103. According to the penetration results, select a reference time series with a length of lookback and a prediction time series with a length of pred as the training data and labels respectively. Among them, for the effective experimental data M of a single finite element simulation model * , divide the time series data into N strips, and obtain X×N strips of penetration time series data, where X represents the number of finite element simulation models;

[0062] S104. Randomly sample the training set, validation set, and test set, and according to the sampling results, for the non-time series data, correspond them to the X×N strips of penetration time series data one by one in turn, and save the complete penetration time series data to complete the production of the penetration time series dataset, where the non-time series data is the target plate material parameters.

[0063] In this embodiment, a standard penetration model is established in the finite element simulation software ANSYS, and the method of automatically batch modifying the parameters of the finite element simulation model with a python script is used to generate a total of X finite element simulation models with different parameters, and the penetration results of each model are calculated by the software to produce a lookback type dataset.

[0064] After the finite element simulation software finishes the calculation, the time series data of the penetration results exported by each finite element simulation model contains M time points. The data at each time point consists of time, the current velocity of the bullet, the current displacement of the bullet, and the total energy of the system. Among these M time points, starting from the moment the bullet touches the target plate, the time points after the bullet touches the target plate are selected, totaling M * time points as the valid data for each simulation. Subsequently, according to the lookback method, a reference time series with a length of lookback and a prediction time series with a length of pred are selected as the training data and labels respectively. For the valid experimental data M of a single finite element simulation model * , the time series data can be divided into N strips, as shown in formula (3):

[0065] N = M * -lookback - pred (3)

[0066] The above X strips of penetration time series data can be divided into a total of X×N strips of lookback data, and are randomly sampled according to the training set, validation set, and test set, with a ratio of 8:1:1. For the non-time series data (target plate material parameters), they correspond one by one with the above X×N strips of penetration time series data. Finally, the complete data set is saved in csv format.

[0067] S2. According to the penetration time series data set, train a dual-branch time series model, and its implementation method is as follows:

[0068] S201. Input the penetration time series data set into the LSTM network for training to obtain the first output variable y lstm , and calculate the first output variable y lstm and the label value in step S103 according to the loss function;

[0069] S202. Determine whether the number of update times has reached the threshold. If so, enter the second stage of training the dual-branch time series model; otherwise, return to step S201;

[0070] S203. Freeze the LSTM network, input the penetration time series data set into the LSTM network to obtain the first output variable y lstm , input the variable p into the KAN network, output the second variable y kan , connect the first variable y lstm and the second variable y kan through a residual connection to obtain the third output variable y co , and calculate the third output variable y co and the label value in step S103 according to the loss function to complete the training of the dual-branch time series model.

[0071] In this embodiment, the dual-branch time series model (LSTM-KAN Embed) adopted by the present invention is trained in two stages. The model framework is shown in Figure 3 . Among them, in the first stage, only the LSTM network is trained, and the output result of this part is the final result of the current stage; in the second stage, the parameters of the LSTM network in the model are frozen, the KAN network is trained, and then the output results of these two networks are subjected to residual connection to obtain the final prediction result.

[0072] In this embodiment, the present invention also designs a more reasonable weighted loss function for time series data, as shown in formula (4):

[0073]

[0074] Among them, loss represents the loss function, y i and respectively represent the true value and the predicted value of the dual-branch time series model at the i-th time point, n represents the length of the predicted sequence vector, and T(n,i) represents the weight function at the i-th time point for the predicted sequence vector of length n.

[0075] In this embodiment, the present invention designs two weight functions, namely harmonic weight and normal distribution weight. The harmonic weight uses the first n terms of the harmonic series as the weight coefficients at different time points, and the gradient of this function is relatively stable, which is more conducive to learning the information of the entire prediction length. The expression is shown in formula (5). The normal distribution weight selects the first n terms within the first 3σ of the normal distribution function as the weight coefficients, and its gradient is larger within the first 3σ, which is more inclined to learn the time series information closer to the front. In this study, σ = 1, μ = 0, and the indices of the first n time points are scaled to the interval (0, 3σ). The expressions are shown in formulas (6) and (7).

[0076]

[0077]

[0078]

[0079] Among them, σ represents the variance of the normal distribution, μ represents the mean of the normal distribution, and x represents the index of the i-th time point mapped to the standard normal distribution (0, 3σ).

[0080] When the weights at each time point are the same, that is, the loss function degenerates to MSE.

[0081] In this embodiment, in the first stage, the time series data set divided in step S1 is input into the LSTM network in batches for training. The input data of this LSTM network is as Figure 3 the variable X in (with the shape of RBatchSize×3×30 ), the output result is as Figure 3 the first output variable y in lstm (with the shape of R BatchSize×1×10 ). Calculate the first output variable and the label value according to the loss function. When the number of training epochs in this stage reaches the predetermined value of 30, end this stage; in the second stage, freeze the LSTM network, and all parameters of this network do not participate in the training in the second stage. Consistent with the first stage, also input Figure 3 the variable X in BatchSize×5 (with the shape of R BatchSize×1 ) into the LSTM network, input the variable p (with the shape of R Figure 3 ) into the KAN network, and the output result has the shape of R Figure 3 the second output variable y in co (with the shape of R BatchSize×10 ). The residual connection (

[0082] y co = [tanh(y kan ) + 1] × y lstm (8)

[0083] where y co represents the prediction result of the performance of the protective material, y kan represents the output of the KAN network, representing the influence degree of the material parameters on the change of the bullet speed. Use the tanh(·) function to scale it to the interval (-1, 1), so that the coefficient of y lstm belongs to the interval (0, 2), meeting the requirements of the residual connection.

[0084] In this embodiment, input the penetration time series data and non-time series data such as material parameters (target plate material parameters) into the trained LSTM network and KAN network respectively, and then make a residual connection between the outputs of the two networks to obtain the final prediction result, completing the test of the trained dual-branch time series model.

[0085] S3. Use the trained dual-branch time series model to obtain the analysis result of the performance of the protective material. The implementation method is as follows:

[0086] S301. Based on the trained dual-branch time series model, save the weights of the KAN network to obtain a weight map;

[0087] S302. Construct a new standard penetration model in the finite element simulation software, select some target plate material parameters to produce a new penetration time series dataset for qualitative and quantitative analysis of parameters, and randomly assign values to the selected target plate material parameters to obtain multiple new finite element simulation models;

[0088] S303. According to the random assignment results, calculate the penetration results of each new finite element simulation model to obtain a new penetration time series dataset;

[0089] S304. Use the trained dual-branch time series model to predict the new penetration time series dataset and perform regression analysis on the prediction results;

[0090] S305. According to the weight map and the prediction results after regression analysis, draw a parameter heat map to obtain the analysis results of the performance of the protective material.

[0091] In this embodiment, first, after completing the training of the dual-branch time series network model, save the weights of the KAN network. The weights of the KAN network reflect the different influences of each material parameter on the deceleration of the bullet. The greater the weight, the more the bullet decelerates during the penetration process, indicating better protective performance. The weights of the KAN network are determined by the activation function between each neuron. The more important the activation function, the greater its coefficient. Then, the coefficients of the activation functions of each neuron can be summed to obtain the weight of the neuron (similar to calculating the weights of the edges between neurons in the MLP).

[0092] Secondly, according to the obtained weight map, select some relatively important parameters to produce a lookback dataset for qualitative and quantitative analysis of parameters, use the trained dual-branch time series network model to make predictions, and then perform regression analysis.

[0093] Construct a standard bullet penetration model used in network training in the finite element simulation software ANSYS, fix and select some material parameters (for example, there are 5 target plate material parameters in the standard simulation model, fix 4 of them, and randomly assign values to the remaining 1 material parameter), randomly assign values to the selected parameters within a reasonable range (usually ±10%), calculate the penetration results of each model through the software, and produce a lookback type dataset. Use the network model to complete the prediction task of the new dataset, and use the prediction results as the data source for regression analysis.

[0094] In this embodiment, in step S101, "automatically modifying the parameters of the finite element simulation model in batches using a Python script" is to modify all the target plate material parameters. For example, if there are 5 material parameters in the standard finite element simulation model, then these 5 material parameters are randomized to generate "a total of X finite element simulation models with different parameters" mentioned in step S101; while in step S302, "selecting some target plate material parameters" means randomly selecting 1 of the 5 material parameters in the standard finite element simulation model (the remaining 4 material parameters are the same as those in the standard finite element simulation model) to generate several new finite element simulation models.

[0095] Regression analysis uses a regression function with interaction terms as shown in formula (9). And by taking the partial derivative method to obtain the influence of one variable on Z when a certain variable is fixed, as shown in formula (10):

[0096] Z = aX + bY + cXY (9)

[0097]

[0098] Among them, X and Y represent the material parameters selected in the pairwise regression analysis, Z represents the bullet deceleration situation, a and b both represent the influence bases, and c represents the influence weight.

[0099] Finally, according to the obtained weight map and prediction results, a parameter heat map is drawn to complete the analysis of the protection performance parameters.

[0100] In summary, the present invention simplifies the complex calculation process and improves the processing efficiency by using deep learning, providing a potential alternative to the traditional finite element simulation analysis. At the same time, the present invention explores the influence of the target plate material parameters on the bullet deceleration effect (such as steps S304 and S305), verifies the effectiveness of the double-branch time series model through quantitative analysis and qualitative analysis, and provides a data grid of key material parameters, providing reference information for the design of bulletproof materials, and is expected to be applied in industrial scenarios to accelerate the research and development of new materials.

Claims

1. A method for analyzing the performance of a protective material based on a dual-branch time series model, characterized in that, It includes the following steps: S1. Construct a standard penetration model in finite element simulation software, and produce a penetration time series dataset by training it; S2. Train a dual-branch time series model according to the penetration time series dataset; S3. Use the trained dual-branch time series model to obtain the analysis results of the performance of the protective material.

2. The method for analyzing the performance of a protective material based on a dual-branch time series model according to claim 1, wherein The step S1 includes the following steps: S101. Construct a standard penetration model in finite element simulation software, and adopt the method of automatically modifying the parameters of the finite element simulation model in batches by using a python script to generate a total of X finite element simulation models with different parameters; S102. Calculate the penetration results of each finite element simulation model. The penetration results include M time points. The data of each time point include time, the current velocity of the bullet, the current displacement of the bullet, and the total energy of the system. Among the M time points, starting from the time when the bullet contacts the target plate, select the time points after the bullet contacts the target plate, with a total of M * time points as the effective experimental data for each simulation; S103. According to the penetration results, select a reference time series with a length of lookback and a prediction time series with a length of pred as training data and labels respectively, where for the effective experimental data M of a single finite element simulation model * divide the time series data into N pieces to obtain X×N pieces of penetration time series data, where X represents the number of finite element simulation models; S104. Randomly sample the training set, validation set, and test set, and according to the sampling results, for non-time series data, correspond them to X×N pieces of penetration time series data one by one in sequence, and save the complete penetration time series data to complete the production of the penetration time series dataset, where the non-time series data is the target plate material parameters.

3. The method for analyzing the performance of the protective material based on the dual-branch time series model according to claim 2, wherein The step S2 includes the following steps: S201. Input the penetration time series data set into the LSTM network for training to obtain the first output variable y lstm , and calculate the first output variable y lstm and the label value in step S103 according to the loss function to complete the first-stage training of the dual-branch time series model; S202. Judge whether the number of update times reaches the threshold. If so, enter the second stage of training the dual-branch time series model; otherwise, return to step S201; S203. Freeze the LSTM network, input the penetration time series dataset into the LSTM network, and obtain the first output variable y lstm , input the variable p into the KAN network, and output the second variable y kan , for the first variable y lstm and the second variable y kan , through residual connection, obtain the third output variable y co , and calculate the third output variable y co and the label value in step S103 according to the loss function to complete the training of the dual-branch time series model.

4. The method for analyzing the performance of a protective material based on a dual-branch time series model according to claim 3, wherein, The expression of the loss function in the step S201 is as follows: Among them, loss represents the loss function, and y i and represent the true value and the predicted value of the dual-branch time series model at the i-th time point respectively, n represents the length of the predicted sequence vector, and T(n,i) represents the weight function at the i-th time point in the predicted sequence vector of length n.

5. The method for analyzing the performance of a protective material based on a dual-branch time series model according to claim 4, characterized in that The weight function T(n,i) includes a harmonic weight and a normal distribution weight; The expression of the harmonic weight is as follows: The expression of the normal distribution weight is as follows: Where, σ represents the variance of the normal distribution, μ represents the mean of the normal distribution, and x represents the index of the i-th time point mapped into the standard normal distribution (0, 3σ).

6. The method for analyzing the performance of a protective material based on a dual-branch time series model according to claim 3, wherein The expression of the residual connection is as follows: y co = [tanh(y kan ) + 1] × y lstm Among them, y co represents the prediction result of the performance of the protective material.

7. The method for analyzing the performance of a protective material based on a dual-branch time series model according to claim 2, wherein The step S3 includes the following steps: S301. Based on the trained dual-branch time series model, save the weights of the KAN network to obtain a weight map; S302. Construct a new standard penetration model in finite element simulation software, select some target plate material parameters to produce a new penetration time series dataset for qualitative and quantitative analysis of the parameters, and randomly take values for the selected target plate material parameters to obtain multiple new finite element simulation models; S303. According to the random value results, calculate the penetration results of each new finite element simulation model to obtain a new penetration time series dataset; S304. Use the trained dual-branch time series model to predict the new penetration time series dataset, and perform regression analysis on the prediction results; S305. According to the weight map and the prediction results after regression analysis, draw a parameter heat map to obtain the analysis results of the performance of the protective material.

Images