Armored vehicle floor-passenger-seat local model equivalent loading modeling method

Abstract

The invention relates to an equivalent loading modeling method for a floor-passenger-seat local model of an armored vehicle. The method comprises the following steps: firstly, constructing a whole vehicle finite element model, and extracting a passenger foot floor node group acceleration matrix; performing dimension reduction on the floor node group acceleration matrix to obtain a floor node group equivalent principal component acceleration matrix; then, a floor-passenger-seat local finite element model is extracted from the whole vehicle finite element model, and on the basis of the local finite element model, coordinates of an equivalent principal component node group in an equivalent principal component acceleration matrix of an equal floor node group are determined by adopting a time domain response error quantization-Bayesian optimization combined method; and finally, loading the equivalent principal component acceleration matrix of the floor node group to the passenger foot floor in the local finite element model to finish modeling. According to the method, the problems of large loading error and poor interpretability of a traditional local model equivalent loading modeling method are solved, and a new solution is provided for rapid and reliable development of a vehicle passenger protection system.

Description

Technical Field

[0001] The present invention belongs to the field of armored vehicle protection, and in particular relates to an equivalent loading modeling method for an armored vehicle floor-occupant-seat local model. Background Art

[0002] In modern warfare, military armored vehicles primarily face underbody explosion threats, such as improvised explosive devices, in asymmetric battlefields. The structural design and selection of occupant protection systems with optimal performance under blast impact often requires multiple rounds of optimization and verification, necessitating repeated armored vehicle explosion simulations or tests. Full-scale vehicle numerical simulation is an important tool for analyzing the performance of armored vehicle occupant protection systems under blast impact, offering advantages in cost and efficiency compared to blast testing. However, it should be noted that numerical simulation efficiency is a key factor influencing the performance development cycle of armored vehicle occupant protection systems under blast impact. In blast simulations requiring iterative optimization, the boundary conditions applied to the occupant and their protection system are essentially identical, resulting in a significant waste of computational resources and time. Solving the complete occupant response in a full-scale vehicle model on a typical local server typically takes several days, which is unacceptable for occupant response assessment problems requiring iterative optimization. Efficiently and cost-effectively evaluating and optimizing the physiological responses of occupants and the performance of occupant protection systems independently of the full-scale vehicle model is a critical step in the development of armored vehicle occupant protection performance.

[0003] The floor-occupant-seat local loading model offers an alternative to full-scale vehicle numerical simulation for rapidly evaluating the performance of occupant protection systems. By separating the full vehicle model from the local occupant-seat model, extracting the floor acceleration at coordinates such as the occupant's feet and seat mounting points from the full vehicle numerical simulation model and applying it to the local occupant-seat model, this approach significantly improves the performance evaluation of occupant protection systems such as protective foot mats and impact-blocking devices. However, the rationality and error of the loading of the thin-walled floor in the local model cannot be ignored.

[0004] There are currently three mainstream loading methods for local models: the first is to extract the acceleration of all nodes of the floor and average it, and use the synthesized average acceleration as the boundary condition for floor loading in the local model, and make certain corrections to the acceleration curve based on the test data. This is the most commonly used loading method, but this method ignores the motion heterogeneity of planar multi-degree-of-freedom particles. There is no unified and rigorous correction method for the average acceleration curve, and the resulting response error can never be solved. The second is to divide the floor into several relatively small blocks and calculate the average acceleration of each block for loading. This method is essentially an improvement on the first method, but it still has the problem of loading errors caused by the uncertain number of floor blocks and the non-unique size of the block areas. The third is to use triangular or sinusoidal waves for loading, and reversely infer the characteristic parameters of the triangular or sinusoidal waves from the occupant response in the full-scale vehicle numerical simulation. This method is only suitable for solving specific engineering problems and lacks feasibility in scientific research. Summary of the Invention

[0005] The purpose of the present invention is to provide a floor-passenger-seat local finite element model data dimensionality reduction equivalent loading modeling method to improve the development efficiency of the occupant protection system while ensuring the solution accuracy and interpretability.

[0006] The technical solution to achieve the present invention is: an equivalent loading modeling method for an armored vehicle floor-occupant-seat local model, comprising:

[0007] S1: Data extraction, i.e. building the finite element model of the entire vehicle and the acceleration matrix A of the occupant foot floor node group.

[0008] S2: Data dimensionality reduction, that is, using the improved kernel principal component analysis method to reduce the data dimension of the floor node group acceleration matrix A and obtain the equivalent principal component acceleration matrix of the floor node group .

[0009] S3: Coordinate determination, that is, extracting the floor-passenger-seat local finite element model from the vehicle finite element model. Based on this local finite element model, the time domain response error quantification-Bayesian optimization method is used to determine the equivalent principal component acceleration matrix of the floor node group. The optimal equivalent principal component node group coordinate vector in the equivalent principal component node group .

[0010] S4: Load modeling, equivalent principal component acceleration matrix of floor node group The equivalent node group acceleration data in the optimal equivalent principal component node group coordinate vector The coordinates in are loaded into the occupant foot floor in the floor-occupant-seat local finite element model to complete the equivalent loading modeling.

[0011] S1 includes: constructing a finite element model of the entire vehicle under explosion impact, extracting the acceleration matrix of the occupant foot floor node group from the finite element model of the entire vehicle , The form is:

[0012]

[0013] Where, is the length of the time series, is the node dimension; is the time series index, ; is the node dimension index, ; for The node group acceleration vector at time instant.

[0014] S2 includes: S21: principal component analysis dimensionality reduction of the acceleration matrix A of the floor node group and S22: magnitude restoration of the dimensionality reduction result.

[0015] S21: Principal component analysis dimensionality reduction of floor node group acceleration matrix A includes:

[0016] The kernel principal component analysis method (KPCA) is used to introduce the kernel function To quantify Correlation between the accelerations of node groups at different times:

[0017]

[0018] Where, is a time series index different from i, ; for The node group acceleration row vector at time t; is the dynamically variable bandwidth of the Gaussian kernel function, which is calculated as:

[0019]

[0020] in is the correlation coefficient, which is calculated as:

[0021]

[0022] Where, for No. Value ; for No. Value ;

[0023] The kernel matrix containing the correlation between the accelerations of all node groups in A at different times is It can be expressed as:

[0024]

[0025] In order to remove the impact of data offset, Perform centralization processing, then the centralized kernel matrix It can be expressed as:

[0026]

[0027] in is an idempotent symmetric matrix, and the calculation method is:

[0028]

[0029] Where, is the m-order identity matrix; is an m-dimensional all-one vector.

[0030] Solution The eigenvalue of and eigenvectors :

[0031]

[0032] Get the eigenvalue matrix and the eigenvector matrix :

[0033]

[0034]

[0035] Where, For the eigenvalues The corresponding eigenvector, Sort by value in descending order.

[0036] Calculate the principal component acceleration matrix B of the floor node group:

[0037]

[0038] In the formula for Before The principal component eigenvalue matrix composed of eigenvalues is in the form of:

[0039]

[0040] for Before The principal component eigenvector matrix composed of eigenvectors is in the form of:

[0041]

[0042] in, is the dimensionality reduction order, which is determined by the Cumulative Variance Contribution Rate (CVCR):

[0043]

[0044] Where, is the coefficient of determination, which is generally taken as 0.9.

[0045] In formula (11), the form of B is:

[0046]

[0047] Where, for The principal component acceleration vector of the node group at the moment . At this point, the data dimension reduction process has been completed. midpoint group nodes, reduced to The principal component node group principal component nodes, but The acceleration data of the principal component node group has no actual physical meaning and needs to be restored to its magnitude.

[0048] S22: Dimensionality reduction results include:

[0049] By introducing the correction factor to restore the magnitude, the equivalent principal component acceleration matrix of the floor node group is solved :

[0050]

[0051] Where, For The correction matrix of the same order, that is for and The Hadamard is the correction factor.

[0052] The correction factor Solve by minimizing the following:

[0053]

[0054] at this time, The equivalent principal component node group contains Equivalent principal component nodes, which have actual physical meaning, but middle The coordinates of the equivalent principal component node groups need to be further determined.

[0055] S3 includes: S31: quantification of equivalent loading error and S32: determination of equivalent principal component node group coordinates.

[0056] S31: Equivalent loading error quantification includes:

[0057] Extracting full vehicle occupant response data from the full vehicle finite element model .

[0058] Extract the floor-passenger-seat local finite element model from the whole vehicle finite element model.

[0059] Using the Design of Experiment (DOE) method, The coordinates of all equivalent principal component nodes in the equivalent principal component node group are combined and sampled to obtain the coordinate vectors of multiple equivalent principal component node groups. .

[0060] Each equivalent principal component node group coordinate vector Load it into the floor-passenger-seat local finite element model to obtain the The corresponding local finite element model is loaded with occupant response data .

[0061] Calculating Phase Error :calculate and The mutual correlation coefficient :

[0062]

[0063] Where, To calculate the corresponding Phase shift when .

[0064] Calculated by the following formula and Phase error between :

[0065]

[0066] Calculated amplitude error : Define the distance cost matrix , the distance cost element is calculated using the Euclidean distance in the following formula :

[0067]

[0068] Define the distance cost matrix Regular path W on :

[0069]

[0070] Where, is the p-th point on the regular path W. Represents middle and middle Euclidean cost distance between two points .

[0071] calculate and The amplitude error between :

[0072]

[0073] Calculate the combined loading error of the local finite element model :

[0074]

[0075] Where, and are the phase error weight factor and the amplitude error weight factor respectively.

[0076] S32: Determining the coordinates of the equivalent principal component node group includes:

[0077] Multiple equivalent principal component node group coordinate vectors Coordinate vector set of the principal component node group of the combined effect , multiple with each Corresponding Combined into a local finite element model loading comprehensive error set F , together constitute the training data set .

[0078] Using the training dataset , build about Gaussian Process (GP) prior distribution approximation surrogate model , which has the form:

[0079]

[0080] Where, for Different from Another equivalent principal component node group coordinate vector; ,for The mean distribution of is the covariance kernel function.

[0081] Determine the current prior distribution approximate surrogate model middle, The maximum value Is it more than 80%? , the new equivalent principal component node group coordinate vector is determined by the expected improvement (EI) acquisition function :

[0082]

[0083] The new equivalent principal component node group coordinate vector is solved by the following formula The posterior distribution of :

[0084]

[0085] Where, The prior distribution defined for the Gaussian process is solved by equation (24); is the observation likelihood, usually assumed to be Gaussian noise .

[0086] Further solve the posterior distribution The predicted mean and the predicted covariance ,in .

[0087] In getting The posterior distribution of After the predicted value is obtained, the iterative optimization process is performed, and the Bayesian optimization steps described in S32 are repeated until , at this time is the optimal equivalent principal component node group coordinate vector .

[0088] S4 includes: Equivalent principal component acceleration matrix of floor node group The equivalent node group acceleration data in The coordinates in are loaded into the occupant foot floor in the floor-occupant-seat local finite element model to complete the equivalent loading modeling.

[0089] Compared with the prior art, the present invention has the following significant advantages:

[0090] (1) The present invention adopts a kernel principal component analysis method that takes into account dynamically variable bandwidth, which solves the problem of ignoring the anisotropy of the motion of the multi-degree-of-freedom node group of the floor in the existing averaging method. The nonlinear acceleration correlation of the floor node group can be accurately captured through the dynamically variable bandwidth, and the dimension of the node group can be significantly reduced while retaining the information characteristics of the original node group acceleration matrix.

[0091] (2) The present invention innovatively introduces a dimension reduction result magnitude restoration process to ensure that the node group acceleration data after dimension reduction is strictly equivalent to the original data in terms of physical magnitude and dynamic characteristics.

[0092] (3) The present invention proposes a time domain response error quantification-Bayesian optimization joint method to achieve the determination of the optimal equivalent principal component node group coordinates based on the objective error evaluation.

[0093] (4) The local model equivalent loading modeling method proposed in this invention greatly shortens the evaluation cycle of the occupant protection system performance while ensuring the solution progress, and breaks through the engineering adaptability limitations of traditional average loading and triangular wave loading methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0094] Figure 1 is the overall framework diagram of the equivalent loading modeling method for the local finite element model.

[0095] Figure 2 is a flow chart of the acceleration matrix dimensionality reduction method.

[0096] Figure 3 is a schematic diagram of the time domain response error quantization-Bayesian optimization joint method.

[0097] Figure 4 is a schematic diagram of equivalent loading modeling. DETAILED DESCRIPTION

[0098] This example is described in further detail below with reference to the accompanying drawings.

[0099] See also Figure 1-4 The present invention describes an armored vehicle floor-occupant-seat local finite element model data dimensionality reduction equivalent loading modeling method.

[0100] Please see the attached Figure 1 The equivalent loading modeling method for the armored vehicle floor-occupant-seat local model of the present invention includes:

[0101] S1: Data extraction, i.e. building the finite element model of the entire vehicle and the acceleration matrix A of the occupant foot floor node group.

[0102] S2: Data dimensionality reduction, that is, using the improved kernel principal component analysis method to reduce the data dimension of the floor node group acceleration matrix A and obtain the equivalent principal component acceleration matrix of the floor node group .

[0103] S3: Coordinate determination, that is, extracting the floor-passenger-seat local finite element model from the vehicle finite element model. Based on this local finite element model, the time domain response error quantification-Bayesian optimization method is used to determine the equivalent principal component acceleration matrix of the floor node group. The optimal equivalent principal component node group coordinate vector in the equivalent principal component node group .

[0104] S4: Load modeling, equivalent principal component acceleration matrix of floor node group The equivalent node group acceleration data in The coordinates in are loaded into the occupant foot floor in the floor-occupant-seat local finite element model to complete the equivalent loading modeling.

[0105] Please see the attached Figure 1 , S1 includes: constructing a finite element model of the entire vehicle under explosion impact, extracting the acceleration matrix of the occupant foot floor node group from the finite element model of the entire vehicle , The form is:

[0106]

[0107] Where, is the length of the time series, is the node dimension; is the time series index, ; is the node dimension index, ; for The node group acceleration vector at time instant.

[0108] Please see the attached Figure 1 and attached Figure 2 , S2 includes: S21: principal component analysis dimensionality reduction of the acceleration matrix A of the floor node group and S22: magnitude restoration of the dimensionality reduction result.

[0109] S21: Principal component analysis dimensionality reduction of floor node group acceleration matrix A includes:

[0110] Since the node group acceleration is affected by the floor impact dynamic response and presents a Gaussian distribution-like attenuation characteristic in space, the kernel principal component analysis (KPCA) method is used to introduce the Gaussian kernel function To quantify Correlation between the accelerations of node groups at different times:

[0111]

[0112] Where, It is also a time series index. ; for The node group acceleration row vector at time t; is the dynamically variable bandwidth of the Gaussian kernel function, which is calculated as:

[0113]

[0114] in is the correlation coefficient, which is calculated as:

[0115]

[0116] Where, for No. Value ; for No. Value ;

[0117] The kernel matrix containing the correlation between the accelerations of all node groups in A at different times is It can be expressed as:

[0118]

[0119] Kernel Matrix is an m-order positive semidefinite symmetric matrix, which contains All feature information in , that is, the similarity between the acceleration vectors of node groups at different times.

[0120] In order to avoid the deviation of acceleration matrix data from the principal component calculation in high-dimensional space, Perform centralization processing, then the centralized kernel matrix It can be expressed as:

[0121]

[0122] in is an idempotent symmetric matrix, and the calculation method is:

[0123]

[0124] Where, is the m-order identity matrix; is an m-dimensional all-one vector; It is still an m-order positive semidefinite symmetric matrix.

[0125] Solution The eigenvalue of and eigenvectors :

[0126]

[0127] Where, represents the main direction of data projection, Indicates the data dispersion variance in the main direction. The larger it is, the more information features the data projection retains in this direction. Get the eigenvalue matrix and the eigenvector matrix :

[0128]

[0129]

[0130] Where, For the eigenvalues The corresponding eigenvector, Arrange in descending order of value, then The corresponding eigenvector is considered the first main direction, and so on.

[0131] Calculate the principal component acceleration matrix B of the floor node group:

[0132]

[0133] In the formula for Before The principal component eigenvalue matrix composed of eigenvalues is in the form of:

[0134]

[0135] for Before The principal component eigenvector matrix composed of eigenvectors is in the form of:

[0136]

[0137] in, is the dimensionality reduction order, which is determined by the Cumulative Variance Contribution Rate (CVCR):

[0138]

[0139] Where, is the coefficient of determination, which is generally taken as 0.9. That is, it is generally believed that when the cumulative variance contribution exceeds 90%, Can carry out research instead of A.

[0140] In formula (11), the form of B is:

[0141]

[0142] Where, for The node group principal component acceleration vector at the moment, the node group principal component acceleration matrix B is retained The main acceleration information is Lower node dimension. At this point, the data dimensionality reduction process has been completed. midpoint group nodes, reduced to The principal component node group principal component nodes, but The acceleration data of the principal component node group has no actual physical meaning and needs to be restored to its magnitude.

[0143] S22: Dimensionality reduction results include:

[0144] By introducing the correction factor to restore the magnitude, the equivalent principal component acceleration matrix of the floor node group is solved :

[0145]

[0146] Where, For The correction matrix of the same order, that is for and The Hadamard is the correction factor.

[0147] The correction factor Solve by minimizing the following:

[0148]

[0149] So far we have completed S2: Dimensionality reduction of the floor acceleration matrix and determined the acceleration value of each principal component node. The equivalent principal component node group contains Equivalent principal component nodes, which have actual physical meaning, but middle The coordinates of the equivalent principal component node groups need to be further determined.

[0150] Please see the attached Figure 1 and attached Figure 3 , S3 includes: S31: equivalent loading error quantification and S32: equivalent principal component node group coordinate determination.

[0151] S31: Equivalent loading error quantification includes:

[0152] Extracting full vehicle occupant response data from the full vehicle finite element model .

[0153] Extract the floor-passenger-seat local finite element model from the whole vehicle finite element model.

[0154] Using the Design of Experiment (DOE) method, The coordinates of all equivalent principal component nodes in the equivalent principal component node group are combined and sampled to obtain the coordinate vectors of multiple equivalent principal component node groups. .

[0155] Each equivalent principal component node group coordinate vector Load it into the floor-passenger-seat local finite element model to obtain the The corresponding local finite element model is loaded with occupant response data .

[0156] Calculating Phase Error :calculate and The mutual correlation coefficient :

[0157]

[0158] Where, To calculate the corresponding The phase shift when . Able to characterize and The time domain phase correlation between them.

[0159] Calculated by the following formula and Phase error between :

[0160]

[0161] Right now When the value is maximum, and The correlation between The largest Phase error .

[0162] Through the phase error and After phase calibration, the amplitude error can be accurately quantified .

[0163] Calculated amplitude error : Define the distance cost matrix , the distance cost element is calculated using the Euclidean distance in the following formula :

[0164]

[0165] Define the distance cost matrix Regular path W on :

[0166]

[0167] Where, is the p-th point on the regular path W. Represents middle and middle Euclidean cost distance between two points The starting point of the regular path W is specified as , the end point is , with monotonicity and continuity that cannot cross nodes.

[0168] When the total cost on the regular path, that is, the Euclidean cost distance corresponding to each point When the sum is minimized, the square root of the total cost is considered and The amplitude error between :

[0169]

[0170] Calculate the combined loading error of the local finite element model :

[0171]

[0172] Where, and are the phase error weight factor and the amplitude error weight factor respectively. Generally, the weighted average value of 1 / 2 is taken.

[0173] S32: Determining the coordinates of the equivalent principal component node group includes:

[0174] Multiple equivalent principal component node group coordinate vectors Coordinate vector set of the principal component node group of the combined effect , multiple with each Corresponding Combined into a local finite element model loading comprehensive error set F , together constitute the training data set .

[0175] Using the training dataset , build about Gaussian Process (GP) prior distribution approximation surrogate model , which has the form:

[0176]

[0177] Where, for Different from Another equivalent principal component node group coordinate vector; ,for The mean distribution of best estimate of; is the covariance kernel function, which defines the similarity between data points in the Gaussian process and is used to construct The key.

[0178] Determine the current prior distribution approximate surrogate model middle, The maximum value Is it more than 80%? , the new equivalent principal component node group coordinate vector is determined by the expected improvement (EI) acquisition function :

[0179]

[0180] Right now should be The value of is the largest.

[0181] The new equivalent principal component node group coordinate vector is solved by the following formula The posterior distribution of :

[0182]

[0183] Where, The prior distribution defined for the Gaussian process is solved by equation (24); is the observation likelihood, usually assumed to be Gaussian noise .

[0184] Further solve the posterior distribution The predicted mean and the predicted covariance :

[0185]

[0186]

[0187] Where, for The matrix form of .

[0188] In getting The posterior distribution of After the predicted value is obtained, the iterative optimization process is performed, and the Bayesian optimization steps described in S32 are repeated until , at this time is the optimal equivalent principal component node group coordinate vector .

[0189] Please see the attached Figure 4 , S4 includes: the floor node group equivalent principal component acceleration matrix The equivalent node group acceleration data in The coordinates in are loaded into the occupant foot floor in the floor-occupant-seat local finite element model to complete the equivalent loading modeling.

Claims

1. The equivalent loading modeling method for the armored vehicle floor-occupant-seat local model is characterized by: The steps include: S1: Data extraction: Construct a finite element model of the vehicle under explosion impact and extract the acceleration matrix A of the occupant foot floor node group; S2: Data dimensionality reduction: Using the improved kernel principal component analysis method, the floor node group acceleration matrix A is reduced to obtain the floor node group principal component acceleration matrix B. The floor node group principal component acceleration matrix B is restored to obtain the equivalent principal component acceleration matrix of the floor node group. ; S3: Determine coordinates: Extract the floor-passenger-seat local finite element model from the vehicle finite element model. Based on the local finite element model, use the time domain response error quantification-Bayesian optimization method to determine the equivalent principal component acceleration matrix of the floor node group. The optimal equivalent principal component node group coordinate vector in the equivalent principal component node group ; S4: Load modeling: Equivalent principal component acceleration matrix of floor node group The equivalent node group acceleration data in the optimal equivalent principal component node group coordinate vector The coordinates in are loaded into the occupant foot floor in the floor-occupant-seat local finite element model to complete the equivalent loading modeling.

2. The method according to claim 1, characterized in that Floor node group acceleration matrix in step S1 The form is: （1） Where, is the length of the time series, is the node dimension; is the time series index, ; is the node dimension index, ; for The node group acceleration vector at time instant.

3. The method according to claim 2, characterized in that In step S2, "using the improved kernel principal component analysis method to perform data dimensionality reduction on the floor node group acceleration matrix A to obtain the floor node group principal component acceleration matrix B" is specifically as follows: The kernel principal component analysis method KPCA is used to introduce the kernel function Quantized floor node group acceleration matrix Correlation between the accelerations of node groups at different times: （2） Where, is a time series index different from i, ; for The node group acceleration row vector at time t; is the dynamically variable bandwidth of the Gaussian kernel function, and the calculation formula is: （3） in is the correlation coefficient, and the calculation formula is: （4） Where, for No. Value ; for No. Value ; The kernel matrix containing the correlation between the accelerations of all node groups at different times in the floor node group acceleration matrix A is Expressed as: （5） Nucleus Matrix Perform centralization processing to obtain the centralized kernel matrix : （6） in is an idempotent symmetric matrix, and the calculation method is: （7） Where, is the m-order identity matrix; is an m-dimensional all-one vector; Solving the centralized kernel matrix The eigenvalue of and eigenvectors : （8） Get the eigenvalue matrix and the eigenvector matrix : （9） （10） Where, For the eigenvalues The corresponding eigenvector, Sort in descending order of value; Calculate the principal component acceleration matrix B of the floor node group: （11） In the formula is the eigenvalue matrix Before The principal component eigenvalue matrix composed of eigenvalues is in the form of: （12） is the eigenvector matrix Before The principal component eigenvector matrix composed of eigenvectors is in the form of: （13） in, is the dimensionality reduction order, which is determined by the cumulative variance contribution rate: （14） Where, is the coefficient of determination; In formula (11), the form of the principal component acceleration matrix B of the floor node group is: （15） Where, for The principal component acceleration vector of the node group at time t.

4. The method according to claim 3, characterized in that In step S2, the principal component acceleration matrix B of the floor node group is restored to obtain the equivalent principal component acceleration matrix of the floor node group. Specifically: Introduce the correction factor to perform magnitude restoration processing and solve the equivalent principal component acceleration matrix of the floor node group : （16） Where, is the principal component acceleration matrix of the floor node group The correction matrix of the same order, that is for and The Hadamard is the correction factor; The correction factor Solve by minimizing the following: （17） 5. The method according to claim 4, characterized in that Step S3 is specifically as follows: S31: Quantification of equivalent loading error: Extracting full vehicle occupant response data from the full vehicle finite element model ; Extract the floor-passenger-seat local finite element model from the whole vehicle finite element model; Using the experimental design method DOE, the equivalent principal component acceleration matrix of the floor node group is calculated. The coordinates of all equivalent principal component nodes in the equivalent principal component node group are combined and sampled to obtain the coordinate vectors of multiple equivalent principal component node groups. ; Each equivalent principal component node group coordinate vector Load it into the floor-passenger-seat local finite element model to obtain the The corresponding local finite element model is loaded with occupant response data ; Calculating Phase Error :calculate and The mutual correlation coefficient : （18） Where, To calculate the corresponding Phase shift when Calculated by the following formula and Phase error between : （19） Calculated amplitude error : Define the distance cost matrix , the distance cost element is calculated using the Euclidean distance in the following formula : （20） Define the distance cost matrix Regular path W on : （21） Where, is the p-th point on the regular path W, Represents middle and middle Euclidean cost distance between two points ; calculate and The amplitude error between : （22） Calculate the combined loading error of the local finite element model : （23） Where, and are the phase error weight factor and the amplitude error weight factor respectively; S32: Determine the equivalent principal component node group coordinates: Multiple equivalent principal component node group coordinate vectors Coordinate vector set of the principal component node group of the combined effect , multiple with each Corresponding Combined into a local finite element model loading comprehensive error set F , together constitute the training data set ; Using the training dataset , build about Gaussian process GP prior distribution approximation surrogate model , which has the form: （24） Where, for Different from Another equivalent principal component node group coordinate vector; ,for The mean distribution of is the covariance kernel function; Determine the current prior distribution approximate surrogate model middle, The maximum value Is it more than 80%? , by expecting to improve the EI acquisition function to determine the new equivalent principal component node group coordinate vector : （25） The new equivalent principal component node group coordinate vector is solved by the following formula The posterior distribution of : （26） Where, The prior distribution defined for the Gaussian process is solved by equation (24); is the observation likelihood, usually assumed to be Gaussian noise ; Solving for the posterior distribution The predicted mean and the predicted covariance ,in ; In getting The posterior distribution of After the predicted value of , the iterative optimization process is performed, repeating the Bayesian optimization steps until , at this time is the optimal equivalent principal component node group coordinate vector .

Images