AK-MCS reliability analysis method adopting improved learning function

The method addresses the technical problem of efficiently and accurately determining the probability of structural failure, which is addressed by using an improved learning function to select optimal sample points for the Kriging surrogate model, enhancing accuracy and reducing computational effort.

CN119066956APending Publication Date: 2024-12-03BEIJING UNIV OF TECH
View PDF 0 Cites 6 Cited by

Patent Information

Application Number
CN202411073517.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-06
Publication Date
2024-12-03

AI Technical Summary

Technical Problem

Existing structural reliability analysis methods face challenges in accurately and efficiently calculating the probability of failure, particularly for complex engineering problems involving implicit functions, due to high computational demands and inaccuracies in existing surrogate models.

Method used

An AK-MCS reliability analysis method using an improved learning function, which employs a WU learning function to select optimal sample points for updating a Kriging surrogate model, prioritizing points near the limit state surface and adjusting weights based on prediction variance to enhance accuracy and reduce computational effort.

Benefits of technology

The method achieves efficient and accurate estimation of structural failure probability with improved convergence and reduced computational cost, ensuring high accuracy and stability in failure probability estimation.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119066956A_ABST
    Figure CN119066956A_ABST
Patent Text Reader

Abstract

The invention discloses an AK-MCS reliability analysis method adopting an improved learning function, and the method comprises the following steps: extracting a sample set through Latin hypercube sampling according to the data characteristics of a structure input variable, carrying out the experimental design, and constructing an initial Kriging agent model according to an initial sample point and a response value thereof; the WU learning function provided by the invention is used for calculating an active learning function value and selecting an optimal sample point for updating the Kriging agent model, the WU learning function considers the influence of a sample point prediction value and a prediction variance on the failure probability estimation precision at the same time, the sample points closer to a limit state curved surface in a sample space are preferentially selected, and the Kriging agent model is updated. The prediction variance of the sample points close to the limit state curved surface is endowed with a large weight, so that the optimal sample points selected by adopting the WU learning function can fall on the limit state curved surface in the iterative optimization process of the Kriging agent model; according to the defined WU learning function stopping criterion, whether the WU learning function is converged and whether the variable coefficient meets the requirement are judged, a final Kriging agent model is obtained, and the failure probability of the structure is calculated. According to the method, for the reliability problem with complex performance functions, stable convergence of the failure probability can be ensured, and meanwhile the failure probability of the structure can be efficiently and accurately estimated.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to an AK-MCS (Active learning reliability method combining Kriging and Monte Carlo Simulation) reliability analysis method using an improved learning function, belonging to the field of structural reliability analysis. Background Art

[0002] For practical complex engineering problems, domestic and foreign scholars have proposed a series of methods to effectively calculate the probability of structural failure, mainly including: approximate analytical method, numerical simulation method and function substitution method. Approximate analytical methods include first-order reliability method, second-order reliability method, etc. Although such methods are simple in calculation, when dealing with reliability problems with complex implicit functions, the calculation results will have large errors. The numerical simulation method predicts the probability of structural failure by using Monte Carlo method, importance sampling method, subset simulation method, linear sampling method and other methods. Although the numerical simulation method can provide the highest accuracy solution for the probability of structural failure, the huge amount of calculation makes such methods less applicable. The function substitution method uses the characteristics of the proxy model to fit the black box function, improves the model accuracy through the adaptive point addition process, and after obtaining the proxy model with ideal accuracy, it combines the numerical simulation method to predict the probability of structural failure. Common proxy models include polynomial chaos expansion, support vector machine model, neural network model, Kriging model, etc.

[0003] The Kriging model is a surrogate model based on Gaussian process regression. Compared with other surrogate models, the Kriging model can not only provide prediction values, but also calculate the prediction variance of sample points. The size of the variance reflects the uncertainty of the prediction value. The basic idea of ​​the AK-MCS method is to use the characteristics of the Kriging model to effectively fit the black box function, and use the Kriging surrogate model to replace the real response function for solving reliability problems in engineering, so as to perform reliability analysis. In the adaptive structural reliability analysis based on the Kriging model, the active learning function is used to select the best sample points. The sample point selection strategy has a great influence on the accuracy and computational efficiency of the Kriging model. An active learning function suitable for structural reliability analysis can estimate the probability of structural failure efficiently, accurately and robustly. Therefore, it is of great significance to propose a method that uses an improved learning function combined with the Kriging model for structural reliability analysis. Summary of the invention

[0004] The present invention provides an AK-MCS reliability analysis method using an improved learning function, the purpose of which is to be able to evaluate the probability of structural failure efficiently and accurately.

[0005] To achieve the above technical objectives, the present invention is implemented through the following technical solutions:

[0006] An AK-MCS reliability analysis method using an improved learning function, comprising:

[0007] Step 1: Determine the failure mode of the structure and the corresponding functional function, and obtain the input variables and distribution information that affect the structural functional function. According to the characteristics of random input variable data, Latin hypercube sampling is used to extract m groups of input samples to establish the MCS sample set N MCS . Perform experimental design (DoE) and select the sample set N MCS Extract n initial sample points S from the kernel, calculate the true response of the sample points, record it as G, S and G form the initial training sample point set T, and construct the initial Kriging model based on T;

[0008] The specific expression of the Kriging model is:

[0009] G(x)=f T (x)β+z(x) (1)

[0010] Where: x is any input variable, G is the corresponding predicted response value, f(x) = [f1(x), f2(x), …, f m (x)] T is the basis function vector of the polynomial, m is the number of polynomials, β=[β1,β2,…,β m ] T is the regression coefficient vector.

[0011] Step 2: Calculate N according to the Kriging model MCS The predicted value and predicted variance of the sample point in the example are calculated by using the WU learning function proposed in the present invention to calculate the active learning function value of the sample point and select the best sample point X new , used to update the Kriging surrogate model.

[0012] The characteristic of the WU function is that it considers the influence of the predicted value and prediction variance of the sample point on the accuracy of failure probability estimation, gives priority to selecting sample points closer to the limit state surface in the sample space, and assigns a larger weight to the prediction variance of the sample points close to the limit state surface. Its mathematical expression is:

[0013]

[0014] Where: and are the sample prediction mean and prediction standard deviation of the Kriging model, is the adaptive orientation function.

[0015] The data characteristics of the adaptive orientation function are as follows Figure 1 As shown. Figure 1 It is known that when When the value of approaches 0, the adaptive orientation function value is 2. This function can ensure that when the predicted value of a sample point is close to the limit state surface, a larger weight is given to the sample point, thereby ensuring that the best sample point selected by the WU learning function falls on the limit state surface as much as possible, improving the accuracy of the Kriging proxy model construction and reducing the number of calls to the performance function.

[0016] Select the sample point corresponding to the minimum value of WU(x) as the optimal sample point of the current Kriging model, that is,

[0017] x new =argmin[WU(x)] (3)

[0018] Step 3: Determine whether the WU learning function converges and the coefficient of variation Cov(P f ) meets the requirements. The present invention defines the convergence criterion of the WU learning function as min{WU(x)≥1}, requiring the coefficient of variation to be less than 0.05. When min{WU(x)≥1}, calculate the coefficient of variation. If Cov(P f )<0.05, that is, the estimation of failure probability is considered acceptable, and the final Kriging model is obtained, and the probability estimation value obtained is That is, as the failure probability P f , if Cov(P f )≥0.05, generate a new MCS sample set, expand the number of initial sample points, and re-execute step 1. When min{WU(x)<1}, calculate X new The response value G new , update the training sample point set T, update the Kriging model, and execute step 2.

[0019] The present invention defines the stopping criterion of the WU learning function as

[0020] min{WU(x)≥1} (4)

[0021] When the WU learning function reaches the convergence criterion, the adaptive orientation function have

[0022]

[0023] Right now At this time, the probability that the sign of the sample point in S is predicted correctly is at least Φ(2)=97.7%, and the Kriging model is accurate enough.

[0024] The failure probability calculation formula is as follows:

[0025]

[0026] Where N G≤0 It is the number of sample points in the MCS sample set whose predicted response value of the Kriging model is less than or equal to 0.

[0027] The coefficient of variation is calculated as follows:

[0028] BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 Adaptive orientation function.

[0030] Figure 2 This disclosure discloses a technical roadmap.

[0031] Figure 3 Schematic diagram of a nonlinear oscillating system.

[0032] Figure 4 Failure probability convergence curve DETAILED DESCRIPTION

[0033] The analysis method of the present invention is further explained below in conjunction with specific examples.

[0034] Figure 2 : is a flow chart of the disclosed method. In this example, an AK-MCS reliability analysis method using an improved learning function is provided, comprising:

[0035] Step 1: Determine the failure mode of the structure and the corresponding functional function, and obtain the input variables and distribution information that affect the structural functional function. According to the data characteristics of the structural input variables, Latin hypercube sampling is used to extract the initial sample points S, calculate the true response G of the sample points, and construct the initial Kriging model;

[0036] Figure 3 The nonlinear oscillator shown is widely used to verify the effectiveness of reliability analysis methods. Its performance function is

[0037]

[0038] Where:

[0039] The parameters of each random variable are shown in Table 1. The random variables c1, c2, r, m, t and F(t) are independent of each other and obey the normal distribution. Where: c1, c2 are spring stiffness, m is the mass of the car, and r is the given threshold.

[0040] Table 1 Random variable distribution

[0041] random variable Distribution Type Mean Standard Deviation <![CDATA[c1]]> normal distribution 1 0.1 <![CDATA[c2]]> normal distribution 0.1 0.01 r normal distribution 0.5 0.05 <![CDATA[F1]]> normal distribution 1 0.2 m normal distribution 1 0.05 <![CDATA[t1]]> normal distribution 1 0.2

[0042] According to the characteristics of random input variable data, the Kriging model is constructed in the sample space [-2, +2], and Latin hypercube sampling is used to extract the initial sample points S = 12 for DoE and N for MCS. MCS For 10 6 , calculate the response value G of the initial sample point, S and G form the initial training sample point set T, build the initial Kriging model based on T, select the basis function and co-correlation function of the Kriging proxy model, and the Kriging proxy model is:

[0043] g k (x) = f(x) T β+z(x) (9)

[0044] In this embodiment, the Kriging model basis function is selected as a zero-order polynomial, and the co-correlation function is a Gaussian correlation function.

[0045] Step 2: Calculate the sample set N according to the Kriging model MCS 10 of 10 6 The predicted value and predicted variance of the sample points are calculated by using the WU learning function proposed in the present invention to calculate the active learning function value of the sample points and select the best sample point X new , used to update the Kriging surrogate model.

[0046] Step 3: Determine whether the WU learning function converges and the coefficient of variation Cov(P f ) meets the requirements. The present invention defines the convergence criterion of the WU learning function as min{WU(x)≥1}, and requires the coefficient of variation to be less than 0.05. The Kriging proxy model is continuously updated iteratively to obtain the final Kriging proxy model, and the probability estimate obtained by the Kriging proxy model is calculated. As the failure probability P f ;

[0047] The calculation results of this example are shown in Table 3. The calculation results are the average values ​​of 10 random tests. The failure probability convergence curve is shown in Figure 4 As shown. call Indicates the number of calls to the function, P f represents the failure probability, represents the coefficient of variation, ε error is the relative error between the estimated failure probability and the standard value, ε error The calculation formula is

[0048]

[0049] Where: is the failure probability calculated by the MCS method; is the failure probability predicted by the Kriging model.

[0050] Table 3 Calculation results

[0051]

[0052] Among them, the typical method is the U learning function method, which is a general technology and will not be described in detail here.

[0053] The coefficient of variation Cov (P f )=0.0058<0.05, that is, the final Kriging proxy model is obtained, and the failure probability P of the nonlinear oscillation system is calculated f =0.028, relative error ε error =0.31%, the number of calls to the functional function is 12+105.7 (12 is the number of initial sample points, 105.7 is the number of times the performance function is called during the active learning process).

[0054] According to the results in Table 3, by comparing the calculation results of the Monte Carlo simulation and the U learning function method, the AK-MCS reliability analysis method using the improved learning function proposed in the present invention can achieve the estimation of the structural failure probability with high efficiency and accuracy. Figure 4 It can be seen that the method proposed in the present invention can ensure the stable convergence of the failure probability and improve the convergence efficiency.

Claims

1. An AK-MCS reliability analysis method using an improved learning function, characterized in that: The following steps are involved: Step 1: According to the data characteristics of the structural input variables, Latin hypercube sampling is used to extract the sample set and conduct experimental design, and the initial Kriging proxy model is constructed based on the initial sample points and their response values; Step 2: Calculate N according to the Kriging model MCS The predicted value and predicted variance of the sample point in the WU learning function are used to calculate the active learning function value of the sample point, and according to x new =argmin[WU(x)] selects the best sample point X new , used to update the Kriging surrogate model; Step 3: According to the defined WU learning function stopping criterion, determine whether the WU learning function converges and whether the coefficient of variation meets the requirements, obtain the final Kriging proxy model, and calculate the failure probability of the structure.

2. The AK-MCS reliability analysis method using an improved learning function according to claim 1, characterized in that: In step 2, considering the influence of the predicted value and prediction variance of the sample points on the accuracy of failure probability estimation, the sample points closer to the limit state surface in the sample space are preferentially selected, and a larger weight is given to the prediction variance of the sample points close to the limit state surface. The WU learning function is proposed, and the mathematical expression is: Where: and are the sample prediction mean and prediction standard deviation of the Kriging model, is the adaptive orientation function.

3. The AK-MCS reliability analysis method using an improved learning function according to claim 1, characterized in that: In step 3, the convergence condition of the WU learning function is expressed as: min{WU(x)≥1} When the WU learning function reaches the convergence criterion, the adaptive orientation function have Right now At this time, the probability that the signs of the sample points in S are predicted correctly is at least Φ(2)=97.7%.

Citation Information

Cited By

  • Mechanism reliability calculation method based on multi-fidelity active learning agent model

    CN119918422A

  • Active learning reliability analysis method for reflector support system

    CN120509311A

  • Active learning reliability analysis method for mirror support system

    CN120509311B

  • Truss structure reliability analysis method and device, computer equipment and storage medium

    CN121188940A

  • High-efficiency and high-precision prediction method for minimum failure probability of aviation structure system based on exponential penalty learning mechanism

    CN121766163A