Multi-scale reliability analysis method for braided composite materials based on hybrid uncertainty

Abstract

The present invention discloses a multi-scale reliability analysis method for woven composite materials based on mixed uncertainty, belonging to the field of composite material technology. Compared with existing reliability analysis methods for woven composite materials, this patent comprehensively considers the cognitive uncertainty of the mechanical properties of the yarns inside the material and the objective uncertainty of the geometric dimensions. It uses the interval method and kernel density estimation method to comprehensively quantify the above two uncertainties, which can achieve accurate simulation of the mechanical response of the structure. It also adopts a sample space optimization method based on stratified sampling and an improved harmony algorithm to achieve uniform filling of the large parameter sample space. A two-level neural network method is used to establish a proxy model from microscopic properties to structural response, avoiding the microscopic modeling work of complex models and achieving accurate prediction of target loads under highly nonlinear conditions.

Description

Technical Field

[0001] The present invention belongs to the technical field of composite materials, and in particular relates to a multi-scale reliability analysis method for braided composite materials based on mixed uncertainty. Background Art

[0002] Woven composites are widely used in aerospace, energy, and military applications, offering low density, high strength, and design flexibility. However, due to the compression of fiber bundles during the molding process, defects such as deformation, bending, and buckling can easily occur. Furthermore, during densification, the matrix is ​​unevenly distributed within the material, resulting in localized uncertainty in the woven composite. As the fundamental component of woven composites, this uncertainty can be interpreted as uncertainty in yarn properties, significantly impacting the overall structural reliability of the woven composite.

[0003] Currently, reliability analysis of composite materials often begins with the material's macroscopic properties and dimensions. These macroscopic properties (such as modulus and strength) or geometric parameters (such as ply angle and thickness) are assumed to be random variables that follow a normal or Weibull distribution. Proxy models are then established to map random variables to structural responses. Publicly available literature examining the reliability of composite materials at the microscale, such as fibers, matrices, or yarns, is limited. Furthermore, research on the propagation of uncertainty from the microscale to the macroscale is inadequate, making it difficult to support improvements in composite material performance from a manufacturing perspective. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to address the deficiencies of the above-mentioned prior art and provide a multi-scale reliability analysis method for woven composite materials based on mixed uncertainty. Compared with the existing reliability analysis methods for woven composite materials, this patent comprehensively considers the cognitive uncertainty of the mechanical properties of the yarns inside the material and the objective uncertainty of the geometric dimensions, and adopts the interval method and the kernel density estimation method to comprehensively quantify the above two uncertainties, which can achieve accurate simulation of the mechanical response of the structure; a sample space optimization method based on stratified sampling and improved harmony algorithm is adopted to achieve uniform filling of large parameter sample space. A two-level neural network method is used to establish a proxy model from microscopic properties to structural responses, which avoids the microscopic modeling work of complex models and achieves accurate prediction of target loads under highly nonlinear conditions.

[0005] In order to achieve the above technical objectives, the technical solution adopted by the present invention is:

[0006] The multi-scale reliability analysis method of woven composite materials based on mixed uncertainty includes the following steps:

[0007] Step 1: Perform XCT scanning on the woven composite material structure to obtain a microscopic image inside the woven composite material structure.

[0008] Step 2: Measure the yarn micro-geometric parameters in the XCT image and obtain the geometric parameter samples.

[0009] Step 3: Partition the woven composite material structure according to the spatial coordinates of the statistical yarn micro-geometric parameters, and use the kernel density estimation method to perform non-parametric estimation on the yarn micro-geometric parameters in each region to obtain the probability density function of the geometric parameters based on the real yarn.

[0010] Step 4: Perform stratified random sampling of the yarn micro-geometric parameters according to the probability density function of the geometric parameters to obtain the first random sample, which is used to characterize the random distribution of the yarn micro-geometric parameters.

[0011] Step 5: Obtain the distribution range of the stress-strain constitutive response curve of the yarn in each direction through numerical simulation or public literature, obtain the boundary curve of its envelope surface based on the interval method, perform polynomial fitting on the boundary curve, use uniform distribution in the range of 0 to 1 for stratified random sampling, interpolate the boundary curve according to the sampling results, and obtain a second random sample to characterize the random properties of the yarn in each direction.

[0012] Step 6: Use the improved harmony algorithm to optimize the combination of the first random sample and the second random sample so that the shortest distance between the combined samples is maximized to obtain the optimized sample set.

[0013] Step 7: Establish a braided structure RVE model based on the optimized sample set, and use the finite element method to calculate the stress-strain constitutive response of the RVE model in various directions to obtain the "micro-parameter-macro-performance" data set.

[0014] Step 8: Establish a neural network framework, use the microscopic parameters as input layer parameters, and the macroscopic performance as output layer parameters to obtain the "microscopic parameter-macroscopic performance" neural network proxy model, completing the mapping from "microscopic parameters" to "macroscopic performance".

[0015] Step 9: Based on the “micro-parameter-macro-performance” neural network proxy model, sample the yarn constitutive parameters and RVE geometric parameters according to the spatial distribution to obtain the corresponding material macro-performance range.

[0016] Step 10: Sampling is performed within the range of the macroscopic properties of the material, assigning the macroscopic properties of the material to the corresponding areas in the structure, performing finite element simulation, obtaining the failure load value of the structure, and obtaining the "macroscopic properties-structural response" data set.

[0017] Step 11: Establish a neural network framework, take macro-performance as input layer parameters, and structural response as output layer parameters, to obtain the "macro-performance-structural response" neural network proxy model, and complete the mapping from "macro-performance" to "structural response".

[0018] Step 12: Based on the above-mentioned “micro-parameters-macro-performance” and “macro-performance-structural response” neural network proxy models, the Monte Carlo method is used to calculate the failure probability of each region in the structure.

[0019] To optimize the technical solution, the present invention further adopts the following measures:

[0020] In step 2, the yarn cross section is simplified into a quadrilateral, an ellipse or a convex lens, and the yarn microscopic geometric parameter is the side length or radius corresponding to the yarn cross section.

[0021] In step 3, the kernel density estimation method is used to obtain the probability density function of the geometric parameters based on the real yarn:

[0022]

[0023] Where N is the number of samples, h is the bandwidth selected for kernel density estimation, and x i is a microscopic geometric parameter of the i-th yarn sample, K(·) represents the kernel function, and the bandwidth h is obtained by cross-validation or iteration method to obtain the optimal value. is the probability density function of the geometric parameters.

[0024] In step 5, the boundary curves are the upper boundary curve and the lower boundary curve, and the expression for polynomial fitting of the boundary curves is: t is the horizontal coordinate value of a point on the boundary curve, f(t) is the vertical coordinate value of a point on the boundary curve, n is the number of fitting points, c j is the j-th coefficient of the boundary curve. The highest degree of the polynomial obtained by boundary curve fitting should be consistent. By interpolating the two curves, an arbitrary curve between the curves can be obtained. The expression of the arbitrary curve is: k is a random number ranging from 0 to 1, a j and b j are the j-th order coefficients corresponding to the upper boundary curve and the lower boundary curve respectively.

[0025] In step 6, the improved harmony algorithm is used to combine the first random sample and the second random sample to obtain a new sample, and the original sample number is used as the new sample parameter for optimization. The goal is to maximize the shortest distance between the new samples. A harmony memory library is set, which contains M initial sample number combinations with K parameters. The harmony memory library is then updated. The updating method is: loop the parameter m from 1 to K: calculate the shortest distance based on the sample number combination in the current harmony memory library, and give the parameters of the best and worst sample number combinations. and Get the number update range Such as x R More than x m The upper bound x mU or lower bound x mL , then assign the upper and lower bounds to x R In this process, a random number r1 between 0 and 1 is generated and compared with the preset mutation probability P m Compare, if the random number is less than the mutation probability P m , then generate new parameters x′ m =x mL +r1×(x mU -x mL ), otherwise follow Generate new parameters, r2 is a random number between 0 and 1. If x′ is used m Replace x worst in After that, the shortest distance between sample number combinations is greater than x worst The shortest spacing, then replace the harmony library During the optimization process of the sample number combination, the problem of generating decimal parameters will be encountered. The nearest integer method is used for rounding. When the maximum and shortest distances converge to no longer change, the iteration ends.

[0026] In step 7, periodic boundary conditions are used to set the boundaries of the RVE model to ensure that the deformation of corresponding nodes on the symmetry plane is consistent.

[0027] In step 8 and step 11, the neural network is a feedforward neural network, a recurrent neural network, or a convolutional neural network.

[0028] The specific steps of step 8 are: establish a neural network framework, set several hidden layers between the input layer and the output layer, each hidden layer contains several neurons, and use the stochastic gradient descent method to adjust the weights and biases in the neural network to minimize the neural network loss function. The yarn axial constitutive parameters, yarn transverse constitutive parameters, yarn shear constitutive parameters and RVE geometric parameters are used as input layer parameters, and the axial constitutive parameters, transverse constitutive parameters and shear constitutive parameters of the woven material are used as output layer parameters to obtain the "micro-parameter-macro-performance" neural network agent model and complete the mapping from "micro-parameters" to "macro-performance".

[0029] The specific steps in step 11 are as follows: Establish a neural network framework, using the macroscopic material properties and load values ​​of regions 1 to N as input layer parameters, and the failure status of regions 1 to N as output layer parameters. Several hidden layers are set between the input and output layers, each containing several neurons. Stochastic gradient descent is used to adjust the weights and biases in the neural network to minimize the neural network loss function, resulting in a "macro-performance-structural response" neural network proxy model, completing the mapping from "macro-performance" to "structural response."

[0030] The specific method of step 12 is: according to the probability density function of the mesoscopic parameters, the yarn constitutive parameters and mesoscopic geometric parameters are randomly sampled on a large scale according to each region, and the improved harmony algorithm in step 6 is used to optimize the parameter combination to obtain the mesoscopic parameters, and the mesoscopic parameters are input into the "mesoscopic parameters-macroscopic performance" neural network proxy model to obtain the parameter distribution of "macroscopic performance", and according to the probability density function of the macroscopic performance, the macroscopic material properties of the woven material in each region are randomly sampled on a large scale, and input into the "macroscopic performance-structural response" neural network proxy model together with the given load to calculate the failure probability of each region in the structure under the given load.

[0031] Compared with the prior art, the present invention has the following beneficial effects:

[0032] 1. Taking into account the epistemic uncertainty of the mechanical properties of the yarn inside the material and the objective uncertainty of the geometric dimensions, the interval method and kernel density estimation method are used to comprehensively quantify the above two uncertainties, respectively, which can achieve accurate simulation of the structural mechanical response.

[0033] 2. A sample space optimization method based on stratified sampling and improved harmony algorithm is adopted. Compared with the existing Latin hypercube sampling method, it can achieve uniform filling of large parameter sample space, occupy less memory space and have faster calculation speed.

[0034] 3. A two-level neural network is used to establish a proxy model from mesoscopic properties to structural macroscopic responses, which takes into account the influence of mesoscopic properties while avoiding the mesoscopic modeling work of complex models, and realizes accurate prediction of target loads under highly nonlinear conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 It is a schematic diagram of the structural partition of the braided composite material;

[0036] Figure 2 The left figure is the sample space optimization result diagram of the improved harmony algorithm; the right figure is the sample space optimization iterative process diagram of the improved harmony algorithm;

[0037] Figure 3 It is a multi-scale RVE model for woven composite materials;

[0038] Figure 4 It is a schematic diagram of the “microscopic parameters-macroscopic performance” neural network agent model;

[0039] Figure 5 It is a schematic diagram of the “macro-performance-structural response” neural network agent model;

[0040] Figure 6 Schematic diagram of the reliability algorithm of the present invention. DETAILED DESCRIPTION

[0041] In order to make the purpose, technical solutions and advantages of this application more clearly understood, the present application is described and illustrated below in conjunction with the accompanying drawings and examples. It should be understood that the specific embodiments described herein are merely used to explain this application and are not intended to limit this application. Based on the embodiments provided in this application, all other embodiments obtained by those of ordinary skill in the art without making any creative efforts are within the scope of protection of this application.

[0042] Obviously, the drawings described below are merely examples or embodiments of the present application. Those skilled in the art can, without inventive effort, apply the present application to other similar scenarios based on these drawings. Furthermore, it is also understood that, although the effort involved in such a development process may be complex and lengthy, for those skilled in the art related to the content disclosed in this application, changes in design, manufacturing, or production based on the technical content disclosed in this application are merely conventional technical means and should not be construed as an insufficiency of the content disclosed in this application.

[0043] References to "embodiments" in this application mean that a particular feature, structure, or characteristic described in connection with the embodiment may be included in at least one embodiment of the application. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor does it refer to independent or alternative embodiments that are mutually exclusive of other embodiments. It is understood, both explicitly and implicitly, by those skilled in the art that the embodiments described in this application may be combined with other embodiments unless there is a conflict.

[0044] This example uses the multi-scale reliability analysis method of woven composite materials based on mixed uncertainty described in the invention to perform reliability analysis on the woven composite material adjustment sheet structure.

[0045] Step 1: Perform XCT scanning on the regulating plate structure to obtain a microscopic image of its interior. The resolution is set to 10 microns, the X-ray voltage is 80 kV, the tube current is 80 microamperes, and the integration time is 2000 milliseconds. Ensure that the image can display basic information such as the yarn profile and contain a large number of yarn cross-section samples. When reconstructing, the image coordinate axis should be aligned with the main axis of the structure to ensure that the microscopic structure does not suffer from image distortion. Divide the structure into 7 regions, such as Figure 1 shown.

[0046] Step 2: Measure the yarn micro-geometric parameters from the XCT image to obtain a geometric parameter sample. The yarn cross section is simplified into an ellipse and a quadrilateral. The micro-geometric parameters are the major and minor axis radii of the warp ellipse, the major and minor axis radii of the weft ellipse, the weft spacing, and the side length of the weft quadrilateral, a total of six micro-geometric parameters.

[0047] Step 3: Partition according to the spatial coordinates of the statistical microscopic parameters, such as Figure 1 The yarn geometric parameters in each region are estimated using kernel density estimation method to obtain the probability density function of the geometric parameters based on the real yarn. The probability distribution function obtained by the KDE method is given by Definition, where N is the number of samples, h is the bandwidth selected for kernel density estimation, and x i is a microscopic geometric parameter of the i-th yarn sample. The bandwidth h can be optimized by cross-validation or iteration.

[0048] Step 4: Perform stratified random sampling according to the probability density function of the yarn micro-geometric parameters. The number of layers is set to 100 and the number of samples is 100 to obtain the first random sample, which is used to characterize the random distribution of the yarn micro-geometric parameters.

[0049] Step 5: Obtain the distribution range of the stress-strain constitutive response curve of the yarn in each direction through numerical simulation or public literature, obtain the boundary curve of its envelope surface based on the interval method, perform polynomial fitting on the boundary curve, use uniform distribution to perform stratified random sampling in the range of 0 to 1, interpolate the boundary curve according to the sampling results, and obtain a second random sample to characterize the random properties of the yarn in each direction. In this example, four parameters are set, namely the yarn axial random factor, the transverse random factor, the 12 / 13 shear random factor, and the 23 shear random factor. Use uniform distribution to perform stratified random sampling in the range of 0 to 1 for each of the four parameters, with the number of layers set to 100 and the number of samples being 100, to obtain a second random sample to characterize the random properties of the yarn in each direction. The boundary curves are the upper boundary curve and the lower boundary curve, and the polynomial expression is: t is the horizontal coordinate value of a point on the boundary curve, f(t) is the vertical coordinate value of a point on the boundary curve, n is the number of fitting points, c j is the j-th order coefficient of the boundary curve. The highest order of the polynomial obtained by fitting the boundary curve should be consistent. By interpolating the two curves, an arbitrary curve between the curves can be obtained. The expression of the arbitrary curve is: k is a random number ranging from 0 to 1, a j and b j are the j-th order coefficients corresponding to the upper boundary curve and the lower boundary curve respectively.

[0050] Step 6: Use the improved harmony algorithm to optimize the combination of the first random sample and the second random sample so that the shortest distance between the combined samples is maximized. The optimization results are as follows: Figure 2 As shown in the left figure, the iterative convergence process is as follows Figure 2 As shown in the right figure. The first random sample and the second random sample are combined to obtain a new sample, and the original sample number is used as the new sample parameter for optimization, with the goal of maximizing the shortest distance between the new samples. Set a harmony memory library, which contains M initial sample number combinations with K parameters, and then update the harmony memory library. The update method is: loop the parameter m from 1 to K: calculate the shortest distance based on the sample number combination in the current harmony memory library, and give the parameters of the best and worst sample number combinations. and Get the number update range Such as x R More than x m The upper bound x mU or lower bound x mL , then assign the upper and lower bounds to x R In this process, a random number r1 between 0 and 1 is generated and compared with the preset mutation probability P m Compare, if the random number is less than the mutation probability P m , then generate new parameters x′m =x mL +r1×(x mU -x mL ), otherwise follow Generate new parameters, r2 is a random number between 0 and 1. If x′ is used m Replace x worst in After that, the shortest distance between sample number combinations is greater than x worst The shortest spacing, then replace the harmony library When optimizing the sample number combination, we will encounter the problem of generating decimal parameters, and we will use the nearest integer method to round them up.

[0051] Step 7: Establish a braided structure RVE model based on the optimized sample set, such as Figure 3 As shown in Figure 2, periodic boundary conditions were used to set the boundaries of the RVE finite element model to ensure consistent deformation of corresponding nodes on the symmetry plane. The finite element method was used to calculate the stress-strain constitutive response of the RVE model in various directions, obtaining a "micro-parameter-macro-performance" data set.

[0052] Step 8: Establish a neural network framework, take microscopic parameters such as yarn axial constitutive parameters, yarn transverse constitutive parameters, yarn shear constitutive parameters, and RVE geometric parameters as input layer parameters, and take macroscopic properties such as woven material axial constitutive parameters, woven material transverse constitutive parameters, and woven material shear constitutive parameters as output layer parameters. Set up several hidden layers between the input layer and the output layer, and each hidden layer contains several neurons. Use methods such as random gradient descent to adjust the weights and biases in the neural network to minimize the neural network loss function and complete the mapping from "microscopic parameters" to "macroscopic properties", such as Figure 4 shown.

[0053] Step 9: Based on the “micro-parameter-macro-performance” neural network proxy model, the yarn constitutive parameters and RVE geometric parameters are sampled according to the spatial distribution to obtain the corresponding material macro-performance range.

[0054] Step 10: Sampling is performed within the range of the macroscopic properties of the material, assigning the macroscopic properties of the material to the corresponding areas in the structure, performing finite element simulation, obtaining the failure load value of the structure, and obtaining the "macroscopic properties-structural response" data set.

[0055] Step 11: Establish a neural network framework, taking macroscopic properties such as macroscopic material properties and load values ​​of regions 1 to N as input layer parameters, and structural responses such as whether failure occurs in regions 1 to N as output layer parameters. Set up several hidden layers between the input layer and the output layer, and each hidden layer contains several neurons. Use methods such as stochastic gradient descent to adjust the weights and biases in the neural network to minimize the neural network loss function and complete the mapping from "macroscopic properties" to "structural response", such as Figure 5 shown.

[0056] Step 12: Based on the above-mentioned "micro-parameters-macro-performance" and "macro-performance-structural response" neural network proxy models, the Monte Carlo method is used to calculate the failure probability of each region in the structure. The specific method is: according to the probability density function of the micro-parameters, the yarn constitutive parameters and micro-geometric parameters are randomly sampled on a large scale according to each region, and the improved harmony algorithm in step 6 is used to optimize the parameter combination to obtain the micro-parameters. The micro-parameters are input into the "micro-parameters-macro-performance" neural network proxy model to obtain the parameter distribution of the "macro-performance". According to the probability density function of the macro-performance, the macro-material properties of the woven material in each region are randomly sampled on a large scale, and input into the "macro-performance-structural response" neural network proxy model together with the given load to calculate the failure probability of each region in the structure under the given load. The reliability analysis flow chart is as follows Figure 6 shown.

[0057] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions based on the principles of the present invention are within the scope of protection of the present invention. It should be noted that for those skilled in the art, various improvements and modifications that do not depart from the principles of the present invention should be considered within the scope of protection of the present invention.

Claims

1. A multi-scale reliability analysis method for braided composite materials based on mixed uncertainty, characterized by: The following steps are involved: Step 1: Perform XCT scanning on the woven composite material structure to obtain a microscopic image inside the woven composite material structure. Step 2: Measure the yarn micro-geometric parameters in the XCT image and obtain the geometric parameter samples. Step 3: Partition the woven composite material structure according to the spatial coordinates of the statistical yarn micro-geometric parameters, and use the kernel density estimation method to perform non-parametric estimation on the yarn micro-geometric parameters in each region to obtain the probability density function of the geometric parameters based on the real yarn. Step 4: Perform stratified random sampling of the yarn micro-geometric parameters according to the geometric parameter probability density function to obtain the first random sample, which is used to characterize the random distribution of the yarn micro-geometric parameters. Step 5: Obtain the distribution range of the stress-strain constitutive response curve of the yarn in each direction through numerical simulation or public literature, obtain the boundary curve of its envelope surface based on the interval method, perform polynomial fitting on the boundary curve, use uniform distribution in the range of 0 to 1 for stratified random sampling, interpolate the boundary curve according to the sampling results, and obtain a second random sample to characterize the random properties of the yarn in each direction. Step 6: Use the improved harmony algorithm to optimize the combination of the first random sample and the second random sample so that the shortest distance between the combined samples is maximized to obtain the optimized sample set. Step 7: Based on the optimized sample set, a braided structure RVE model is established. The stress-strain constitutive response of the RVE model in each direction is calculated using the finite element method to obtain a "micro-parameter-macro-performance" data set. Step 8: Establish a neural network framework, use the microscopic parameters as input layer parameters, and the macroscopic performance as output layer parameters to obtain the "microscopic parameter-macroscopic performance" neural network proxy model, completing the mapping from "microscopic parameters" to "macroscopic performance". Step 9: Based on the "micro-parameter-macro-performance" neural network proxy model, sample the yarn constitutive parameters and RVE geometric parameters according to the spatial distribution to obtain the corresponding material macro-performance range. Step 10: Sampling is performed within the range of the material's macroscopic properties, assigning the material's macroscopic properties to the corresponding areas in the structure, performing finite element simulation, obtaining the failure load value of the structure, and obtaining the "macroscopic properties-structural response" data set. Step 11: Establish a neural network framework, take macro-performance as input layer parameters, and structural response as output layer parameters, to obtain the "macro-performance-structural response" neural network proxy model, and complete the mapping from "macro-performance" to "structural response". Step 12: Based on the aforementioned "micro-parameters-macro-performance" and "macro-performance-structural response" neural network proxy models, the Monte Carlo method is used to calculate the failure probability of each region in the structure.

2. The multi-scale reliability analysis method for braided composite materials based on hybrid uncertainty according to claim 1 is characterized in that: In step 2, the yarn cross section is simplified into a quadrilateral, an ellipse or a convex lens, and the yarn microscopic geometric parameter is the side length or radius corresponding to the yarn cross section.

3. The multi-scale reliability analysis method for braided composite materials based on hybrid uncertainty according to claim 2 is characterized in that: In step 3, the kernel density estimation method is used to obtain the probability density function of the geometric parameters based on the real yarn: Where N is the number of samples, h is the bandwidth selected for kernel density estimation, and x i is a microscopic geometric parameter of the i-th yarn sample, K(·) represents the kernel function, and the bandwidth h is obtained by cross-validation or iteration method to obtain the optimal value. is the probability density function of the geometric parameters.

4. The multi-scale reliability analysis method of woven composite materials based on hybrid uncertainty according to claim 3 is characterized in that: In step 5, the boundary curves are the upper boundary curve and the lower boundary curve, and the expression for polynomial fitting of the boundary curves is: t is the horizontal coordinate value of a point on the boundary curve, f(t) is the vertical coordinate value of a point on the boundary curve, n is the number of fitting points, c j is the j-th coefficient of the boundary curve. The highest degree of the polynomial obtained by boundary curve fitting should be consistent. By interpolating the two curves, an arbitrary curve between the curves can be obtained. The expression of the arbitrary curve is: k is a random number ranging from 0 to 1, a j and b j are the j-th order coefficients corresponding to the upper boundary curve and the lower boundary curve respectively.

5. The multi-scale reliability analysis method for braided composite materials based on hybrid uncertainty according to claim 4 is characterized in that: In step 6, the improved harmony algorithm is used to combine the first random sample and the second random sample to obtain a new sample, and the original sample number is used as the new sample parameter for optimization. The goal is to maximize the shortest distance between the new samples. A harmony memory library is set, which contains M initial sample number combinations with K parameters. The harmony memory library is then updated. The updating method is: loop the parameter m from 1 to K: calculate the shortest distance based on the sample number combination in the current harmony memory library, and give the parameters of the best and worst sample number combinations. and Get the number update range Such as x R More than x m The upper bound x mU or lower bound x mL , then assign the upper and lower bounds to x R In this process, a random number r1 between 0 and 1 is generated and compared with the preset mutation probability P m Compare, if the random number is less than the mutation probability P m , then generate new parameters x′ m =x mL +r1×(x mU -x mL ), otherwise follow Generate new parameters, r2 is a random number between 0 and 1, such as x′ m Replace x worst in After that, the shortest distance between sample number combinations is greater than x worst The shortest spacing, then replace the harmony library During the optimization process of the sample number combination, the problem of generating decimal parameters will be encountered. The nearest integer method is used for rounding. When the maximum and shortest distances converge to no longer change, the iteration ends.

6. The multi-scale reliability analysis method of woven composite materials based on hybrid uncertainty according to claim 5 is characterized in that: In step 7, periodic boundary conditions are used to set the boundaries of the RVE model to ensure that the deformation of corresponding nodes on the symmetry plane is consistent.

7. The multi-scale reliability analysis method of woven composite materials based on hybrid uncertainty according to claim 1 is characterized in that: In step 8 and step 11, the neural network is a feedforward neural network, a recurrent neural network, or a convolutional neural network.

8. The multi-scale reliability analysis method for braided composite materials based on hybrid uncertainty according to claim 6, characterized in that: The specific steps of step 8 are as follows: establish a neural network framework, set up several hidden layers between the input layer and the output layer, each hidden layer contains several neurons, use the stochastic gradient descent method to adjust the weights and biases in the neural network to minimize the neural network loss function, use the yarn axial constitutive parameters, yarn transverse constitutive parameters, yarn shear constitutive parameters and RVE geometric parameters as input layer parameters, and use the axial constitutive parameters, transverse constitutive parameters and shear constitutive parameters of the woven material as output layer parameters to obtain the "micro-parameter-macro-performance" neural network proxy model, completing the mapping from "micro-parameters" to "macro-performance".

9. The multi-scale reliability analysis method of woven composite materials based on hybrid uncertainty according to claim 1, characterized in that: The specific steps of step 11 are: establish a neural network framework, use the macro material properties and load values ​​of regions 1 to N as input layer parameters, use whether failure occurs in regions 1 to N as output layer parameters, set several hidden layers between the input layer and the output layer, each hidden layer contains several neurons, and use the stochastic gradient descent method to adjust the weights and biases in the neural network to minimize the neural network loss function, thereby obtaining a "macro performance-structural response" neural network proxy model and completing the mapping from "macro performance" to "structural response".

10. The multi-scale reliability analysis method of braided composite materials based on hybrid uncertainty according to claim 1, characterized in that: The specific method of step 12 is as follows: according to the probability density function of the mesoscopic parameters, large-scale random sampling of the yarn constitutive parameters and mesoscopic geometric parameters are performed in each region, and the improved harmony algorithm in step 6 is used to optimize the parameter combination to obtain the mesoscopic parameters. The mesoscopic parameters are input into the "mesoscopic parameters-macroscopic performance" neural network proxy model to obtain the parameter distribution of "macroscopic performance". According to the probability density function of the macroscopic performance, large-scale random sampling of the macroscopic material properties of the woven material in each region is performed, and the macroscopic material properties are input into the "macroscopic performance-structural response" neural network proxy model together with the given load to calculate the failure probability of each region in the structure under the given load.

Images