Method for efficiently predicting collision performance of reinforcing beam structure of power battery pack of electric vehicle

Abstract

The invention provides an efficient prediction method for collision performance of an electric vehicle power battery pack stiffening beam structure, which comprises the following steps of: establishing a basic closed beam structure model with four section configurations by adopting a typical working condition of quasi-static compression, carrying out preliminary sampling through a simulation analysis method, and collecting a collision performance model of the stiffening beam structure under the conditions of different sizes and wall thicknesses of different configurations; and collision safety key indexes such as peak cross-section force and specific energy absorption under a quasi-static compression working condition. Data obtained through simulation analysis is trained and optimized through an MR-LVGP model, and then indexes such as peak cross section force and specific energy absorption of the four configurations under different height-width ratios and thicknesses are obtained. Compared with a traditional simulation analysis method, the method for predicting the corresponding indexes through MR-LVGP model fitting has the advantages of being high in precision and short in consumed time, the efficiency of vehicle collision performance design can be improved, and repeated analysis work can be reduced.

Description

Technical Field

[0001] The present invention relates to the field of automotive structural design, and proposes an efficient prediction method for the collision performance of the reinforcing beam structure of an electric vehicle's power battery pack based on the MR-LVGP model. Background Art

[0002] With the rapid development of computer technology, computer simulation plays an important role in current scientific and engineering research. Due to limitations in cost and R & D cycle in the development work of actual vehicle models, it is often impossible to conduct a large number of real vehicle collision tests. With the help of computer simulation analysis, researchers can explore and analyze the process of vehicle collision at a lower cost. At the same time, with the development of statistics and neural networks, more high-precision and multi-parameter response surface fitting models, such as the multi-response Gaussian latent variable (MR-LVGP) model. Combining computer collision simulation with response surface fitting can further improve the efficiency of collision analysis, and through neural networks, the relationship between multiple inputs and outputs can be obtained by implicit solution methods, and it has a high degree of accuracy.

[0003] In the process of vehicle collision simulation, its dynamic response analysis and optimization involve strong non-linear conditions (material non-linearity, geometric non-linearity, contact non-linearity), a large number of design variable types (material selection, cross-section topology configuration and dimensions), and numerous design objectives (component level: energy absorption, cross-section force, mass, etc.; system level: maximum deformation of the cockpit, driver acceleration, mass, etc.). At the same time, in practical engineering applications, factors such as process and cost need to be considered, thus bringing great challenges to the crashworthiness design and engineering application of automotive structures. Summary of the Invention

[0004] The object proposed by the present invention is to provide an efficient prediction method for the collision performance of the reinforcing beam structure of an electric vehicle's power battery pack based on the MR-LVGP model. First, the simulation results of four commonly used closed energy-absorbing structures are statistically analyzed. Based on the existing dimensions of the cross beam installed on a certain vehicle model's power battery pack, a certain range is extended in both the increasing and decreasing directions. Then, according to the statistical data, the MR-LVGP model is used for fitting to predict the corresponding parameters of the non-sampled size configurations, so as to simplify the design work.

[0005] The above object of the present invention is achieved through the following scheme: An efficient prediction method for the collision performance of the reinforcing beam structure of an electric vehicle's power battery pack based on the MR-LVGP model, characterized by including the following steps:

[0006] Step 1: Initially design the dimensions such as the basic length, width, height, and thickness of the thin-walled closed structures of the four configurations of the square, rectangular, cross-shaped, and grid-shaped involved in the automotive power battery pack's reinforcing cross beam, and determine the sampling range;

[0007] Step 2: Based on the quasi-static compression condition, measure the collision energy absorption effect and peak force performance of the cross sections with different parameters;

[0008] Step 3: According to the created CAD model, perform corresponding pre-processing work such as mesh division and parameter setting, and perform corresponding simulation calculations, data collection, etc.;

[0009] Step 4: According to the data fitting principle of the MR-LVGP model, the corresponding covariance matrix calculation and hyperparameter solution are performed for 100 sets of sampled data, the data points to be predicted are input, the new covariance vector is calculated, and the predicted value is provided;

[0010] Furthermore, in step 1, the four thin-walled closed section structures constructed are based on engineering data, the width of the section is limited to a fixed value, and the size in the height direction is variable, and the value range is consistent with the size range of the body structure beam.

[0011] Furthermore, in step 4, the simulation data is imported into the MR-LVGP model for fitting, which specifically includes:

[0012] The covariance function is calculated using the following formula:

[0013] r(s,s ′ )=exp{-(xx ′ ) T Φ(xx ′ )-(z(t)-z(t ′ ) T (z(t)-z(t ′ ))}

[0014] Where x is a quantitative variable vector containing input variables; set the diagonal matrix Φ = diag(φ), φ = [φ1, φ2, …, φ p ] T , to avoid over-parameterization of qualitative and quantitative variables; z(t) is also a latent variable optimized by machine learning, which transforms the qualitative variable t and maps it to a latent vector z(t), which is convenient for converting the multidimensional model into a one-dimensional model by linearly combining z(t). The specific value of z(t) is obtained by machine learning optimization through the training data set; s = [x T ,z(t) T ] T That is, the combined input vector of x and z(t) is the input vector of the covariance model. The covariance model mainly calculates the covariance between two points in the input vector and generates a covariance matrix:

[0015]

[0016] In the above formula, the input parameters of the data set are used to determine the covariance matrix. The covariance matrix related hyperparameters, such as Φ and z(t), need to be trained in combination with the output parameters in the data set. The log-likelihood function is written as:

[0017]

[0018] Where R is N calculated according to the above formula data ×N data The covariance matrix contains the hyperparameters Z and Φ, where Z is the latent vector matrix composed of z(t); H is N resp ×N data 1 matrix, N data The number of sample input data, N resp is the number of output response sampling data; vec(.) is the column vectorization conversion of the matrix; is the Kronecker product; q is the number of qualitative variables t; D is the N of the data set resp ×N data dimensional output sampling data matrix; B and ∑ are the parameter matrices to be determined based on the training data. By setting the derivatives of the function represented by the above formula with respect to B and ∑ to 0, the maximum likelihood estimates of the two can be obtained:

[0019]

[0020] Obtain and Then, substitute it into the log-likelihood function and obtain the estimation of Z and Φ by minimizing the log-likelihood function.

[0021] When a new parameter point to be predicted is input, the covariance vector between the new predicted point and each point in the original data set is first calculated:

[0022] r=[r(s * ,s (1) ),…,r(s * ,s (n) )] T

[0023] In the formula, s * is the input parameter vector of the new input point, s (1) ,…,s (n) is the input parameter vector of the original training parameter set data points;

[0024] Then, based on the calculated covariance vector, the newly imported data point w can be obtained * The predicted value is as follows:

[0025]

[0026] In the formula, h(w * ) indicates that N resp The prior mean vector is composed of 1s.

[0027] The beneficial effects of the present invention include:

[0028] Combining the establishment of simulation models with the actual models under development is more engineering-oriented. Compared with the traditional empirical design method, simulation analysis can complete collision analysis more quickly and at a lower cost, and the fitting results of the MR-LVGP model have the advantages of multi-parameterization and high precision. The combination of the two can guide the parametric design process of closed-section beams more accurately, effectively and with high precision, minimize the intrusion of the power battery pack during the collision, increase energy absorption efficiency, and improve design efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] In order to more clearly illustrate the three-dimensional structure of the closed-section beam to which the present invention is applicable, the following briefly introduces the drawings required for use in the present invention. It should be understood that the following drawings only illustrate certain embodiments of the present invention and should not be considered as limiting the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative work.

[0030] Attached Figure 1 It is a schematic diagram of the structure of four different cross-section beams in the present invention.

[0031] Attached Figure 2 for Figure 1 Schematic diagram of the LS-DYNA simulation mesh model of the Chinese-shaped mouth among the four configurations.

[0032] Attached Figure 3 for Figure 2 The cross-sectional force curve of the U-shaped LS-DYNA simulation results.

[0033] Attached Figure 4 for Figure 2 Schematic diagram of LS-DYNA simulation results of the U-shaped section.

[0034] Attached Figure 5 Schematic diagram of the battery pack beam suitable for the fitting prediction of the present invention.

[0035] Attached Figure 6 This is an example of the prediction results after training the MR-LVGP model. DETAILED DESCRIPTION

[0036] To make the implementation method, purpose, technical solution and advantages of the present invention clearer and more complete. The part described and shown in the attached drawings presents the basic research method of the present invention, and the rest is repetitive data collection work, so it is not all listed. The specific measures described here are only used to explain the present invention and are not used to limit the present invention.

[0037] The present invention is an efficient prediction method for the collision performance of the reinforcing beam structure of the power battery pack of an electric vehicle based on the MR-LVGP model, so as to reduce repetitive calculation work and optimize the design and development efficiency.

[0038] Step 1: Preliminarily design the basic dimensions such as the basic length, width, height, and thickness of the thin-walled closed structures of the four configurations of the mouth shape, day shape, eye shape, and field shape involved in the cross beam of the automotive power battery pack, and determine the sampling range;

[0039] In this embodiment, due to the dimensional space limitation of the actual vehicle model, such as Figure 5 shown, the width of the energy-absorbing structure reinforcing cross beam is limited to 70 mm, the initial design height is 132 mm, and the specific structures of the four configurations are as Figure 1 shown. Considering the aspect ratio commonly used in engineering practice at present, five groups of sampling heights are set as 52.5 mm, 70 mm, 105 mm, 132 mm, and 175 mm.

[0040] Step 2: Measure the collision energy absorption effect and peak force performance of cross sections with different parameters;

[0041] Considering the design requirements of the energy-absorbing collapse area of the power battery pack reinforcing cross beam, the length of the cross sections of the four sizes is set to 200 mm. Since the cross section structures of the eye shape and field shape in the four structures are slightly more complex and are more likely to show instability and bending during compression, a shorter length is adopted, so as to more specifically analyze the energy-absorbing collapse effect of the corresponding structures.

[0042] In addition to the design guidance for the energy-absorbing collapse length of the cross beam, the simulation peak force results of the four configurations at this length have qualitative guiding significance for the buckling resistance performance of the cross beam of the same structure with a larger length. In engineering practice, the energy-absorbing capacity can be increased by applying a collapse guiding groove to avoid instability.

[0043] Step 3: According to the created CAD model, perform corresponding preprocessing work such as mesh generation and parameter setting, and perform corresponding simulation calculations, data collection, etc.;

[0044] In this example, in the developed vehicle model, the designed thickness of the corresponding thin-walled closed structure is 3 mm, within the common thickness range of aluminum alloy stamping, die-casting and other processes. Five groups of thicknesses of 2 mm, 2.5 mm, 3 mm, 3.5 mm, and 4 mm are set for sampling.

[0045] Based on the above input variable settings, the sampling data has a total of 100 data points, 25 data points for each configuration. The corresponding CAD model established by CATIA software is imported into ANSA software for pre-processing, and the resulting grid analysis model is as follows: Figure 2 shown.

[0046] The collision analysis model adopts the form of shell mesh, and the calculation is performed by applying thickness information on the mesh. SPC fixed constraints are applied to the rear end of the model, and a rigid wall is set at the front end to extrude the model at a constant speed, which is quasi-static compression. The material is set to 6082T6 aluminum alloy, and the material parameters are shown in Table 1. The corresponding contact and control parameters are set to complete the simulation settings.

[0047] Table 1 6082-T6 aluminum alloy related parameters

[0048]

[0049] Analyze the corresponding results output by LS-DYNA, such as Figure 3 and Figure 4 As shown, the peak force of this structure is 171.408 kN. During the entire collision process, the structure is evenly compressed and wrinkled, and the energy absorption form is good.

[0050] Figure 3 The figure shows the cross-sectional force curve during the entire collision process. Since the rigid wall moves in a uniform motion, the time on the horizontal axis of the curve can be equivalent to the deformation length, which is the cross-sectional force-deformation length curve.

[0051] The specific energy absorption can be obtained by calculating the structural mass obtained in ANSA and the change in internal energy output in LS-DYNA.

[0052] According to the above simulation method, the specific energy absorption and peak force data of the quasi-static compression test with 4 configurations, 5 lengths, and 5 thicknesses with a total of 100 data sampling points were collected and prepared to be entered into the MR-LVGP model for regression analysis.

[0053] Step 4: According to the data fitting principle of the MR-LVGP model, the corresponding covariance matrix calculation and hyperparameter solution are performed for 100 sets of sampled data, the data points to be predicted are input, the new covariance vector is calculated, and the predicted value is provided;

[0054] The MR-LVGP model is a non-parametric model. Each fitting requires calling all data sampling points to participate in the calculation, which makes the increase and decrease of data points more flexible. And through machine learning methods, according to the existing data points, the hyperparameters of the relevant Gaussian kernel in the model kernel function are hyper-optimized to reduce the error of model prediction. In the classic GP process, the value of the hyperparameter determines the shape of the probability distribution of the function fitted by the GP process, that is, the size of the prediction error, and does not affect whether the fitting can be performed.

[0055] The basic steps of regression through the MR-LVGP model are as follows:

[0056] Select a suitable Gaussian kernel function, i.e., covariance function: In the present invention, the covariance function used is as follows:

[0057] r(s,s ′ )=exp{-(xx ′ ) T Φ(xx ′ )

[0058] -(z(t)-z(t ′ ) T (z(t)-z(t ′ ))}

[0059] Where x is a quantitative variable vector containing input variables; set the diagonal matrix Φ = diag(φ), φ = [φ1, φ2, …, φ p ] T , to avoid over-parameterization of qualitative and quantitative variables; z(t) is also a latent variable optimized by machine learning, which transforms the qualitative variable t and maps it to a latent vector z(t), which is convenient for converting the multidimensional model into a one-dimensional model by linearly combining z(t). The specific value of z(t) is obtained by machine learning optimization through the training data set; s = [x T ,z(t) T ] T That is, the combined input vector of x and z(t) is the input vector of the covariance model. The covariance model mainly calculates the covariance between two points in the input vector and generates a covariance matrix:

[0060]

[0061] In the above formula, the input parameters of the data set are mainly used to determine the covariance matrix. The covariance matrix related hyperparameters, such as Φ and z(t), need to be trained in combination with the output parameters in the data set. The log-likelihood function is:

[0062]

[0063] Where R is N calculated according to the above formula data ×N data The covariance matrix contains the hyperparameters Z and Φ, where Z is the latent vector matrix composed of z(t); H is N resp ×N data 1 matrix, N data The number of sample input data, N resp is the number of output response sampling data; vec(.) is the column vectorization conversion of the matrix; is the Kronecker product; q is the number of qualitative variables t; D is the N of the data set resp ×N data dimensional output sampling data matrix; B and ∑ are the parameter matrices to be determined based on the training data. By setting the derivatives of the function represented by the above formula with respect to B and ∑ to 0, the maximum likelihood estimates of the two can be obtained:

[0064]

[0065] Obtain and Then, it is substituted into the log-likelihood function, and the estimation of Z and Φ is obtained by minimizing the log-likelihood function. In the present invention, the parameter values ​​obtained by optimization are: Z = [-0.778774464982514, -0.432391962384804, -0.591677515402194, -0.0498745849512014, 0.539903650411312];

[0066] Φ=[1.57316865977923,0.526365817479556].

[0067] Based on the hyperparameters calculated above and the covariance matrix of the training samples, the relevant parameter values ​​of the new input points can be predicted.

[0068] When a new parameter point to be predicted is input, the covariance vector between the new predicted point and each point in the original data set is first calculated:

[0069] r=[r(s * ,s (1) ),…,r(s * ,s (n) )] T

[0070] In the formula, s * is the input parameter vector of the new input point, s (1) ,…,s (n) is the input parameter vector of the original training parameter set data points.

[0071] Then, based on the calculated covariance vector, the newly imported data point w can be obtained * The predicted value is as follows:

[0072]

[0073] In the formula, h(w * ) indicates that N resp The prior mean vector composed of 1

[0074] According to the fitting process described in the MR-LVGP model described above, the simulated data points are imported into the model, and the predicted values ​​can be obtained through fitting calculation, such as Figure 6 shown.

[0075] The shapes of the various components in the drawings are schematic, and certain differences from their actual shapes are not excluded. The drawings are only used to illustrate the principles of the present application and are not intended to limit the present application.

[0076] Although the present application is disclosed in detail with reference to the accompanying drawings, it should be understood that these descriptions are merely exemplary and are not intended to limit the application of the present application. The scope of protection of the present application is defined by the appended claims and may include various modifications, alterations and equivalents made to the invention without departing from the scope and spirit of the present application.

Claims

1. An efficient prediction method for the collision performance of the reinforcing beam structure of the power battery pack of an electric vehicle based on the MR-LVGP model, characterized by: It includes the following steps: Step 1: Initially design the dimensions such as the basic length, width, height, and thickness of the thin-walled closed structures of the four configurations of the mouth-shaped, day-shaped, eye-shaped, and field-shaped for the crossbeam of the automotive battery pack, and determine the sampling range; Step 2: Based on the quasi-static compression condition, measure the collision energy absorption effect and peak force performance of cross-sections with different parameters; Step 3: According to the created CAD model, perform corresponding pre-processing work such as mesh generation and parameter setting, and carry out corresponding simulation calculations, data collection, etc.; Step 4: According to the data fitting principle of the MR-LVGP model, calculate the corresponding covariance matrix and solve the hyperparameters for 100 groups of sampling data, input the data points to be predicted, calculate the new covariance vector, and provide the predicted value.

2. The prediction method for the collision energy absorption index of the crossbeam of the automotive power battery pack based on the MR-LVGP model according to claim 1, wherein: In step 1, the four thin-walled closed cross-section structures are constructed based on engineering data. The width of the cross-section is restricted and is a fixed value, and the dimension in the height direction is variable, and the value range conforms to the dimension range of the crossbeam of the body power battery pack structure.

3. The prediction method for the collision energy absorption index of the automotive crossbeam based on the MR-LVGP model according to claim 1, wherein: In step 4, import the simulation data into the MR-LVGP model for fitting, specifically including: Use the following formula to calculate the covariance function: r(s,s ′ )=exp{-(x-x ′ ) T Φ(x-x ′ )-(z(t)-z(t ′ ) T (z(t)-z(t ′ ))} Where x is a quantitative variable vector containing input variables; set the diagonal matrix Φ = diag(φ), φ = [φ1, φ2, …, φ p ] T , to avoid over-parameterization of qualitative and quantitative variables; z(t) is also a latent variable optimized by machine learning, which transforms the qualitative variable t and maps it to a latent vector z(t), which is convenient for converting the multidimensional model into a one-dimensional model by linearly combining z(t). The specific value of z(t) is obtained by machine learning optimization through the training data set; s = [x T ,z(t) T ] T That is, the combined input vector of x and z(t) is the input vector of the covariance model. The covariance model mainly calculates the covariance between two points in the input vector and generates a covariance matrix: In the above formula, the input parameters of the data set are used to determine the covariance matrix. The hyperparameters related to the covariance matrix, such as Φ and z(t), need to participate in the training in combination with the output parameters in the data set. Write the log-likelihood function as: Where R is N calculated according to the above formula data ×N data The covariance matrix contains the hyperparameters Z and Φ, where Z is the latent vector matrix composed of z(t); H is N resp ×N data 1 matrix, N data The number of sample input data, N resp is the number of output response sampling data; vec(.) is the column vectorization conversion of the matrix; is the Kronecker product; q is the number of qualitative variables t; D is the N of the data set resp ×N data dimensional output sampling data matrix; B and ∑ are the parameter matrices to be determined based on the training data. By setting the derivatives of the function represented by the above formula with respect to B and ∑ to 0, the maximum likelihood estimates of the two can be obtained: Obtain and Then, substitute it into the log-likelihood function and obtain the estimation of Z and Φ by minimizing the log-likelihood function. When inputting new parameter points to be predicted, first calculate the covariance vector between the new prediction points and each point in the original data set: r=[r(s * ,s (1) ),…,r(s * ,S (n) )] T In the formula, s * is the input parameter vector of the new input point, s (1) ,…,s (n) is the input parameter vector of the original training parameter set data points; Then, based on the calculated covariance vector, the newly imported data point w can be obtained * The predicted value is as follows: In the formula, h(w * ) indicates that N resp The prior mean vector is composed of 1s.

Images