Reliability evaluation method combining adaptive correlation vector machine and Markov chain important sampling
By combining the adaptive correlation vector machine and the Markov chain important sampling method, the problems of low structural reliability evaluation efficiency and large result error in the prior art are solved, and efficient and accurate structural reliability evaluation is achieved, reducing the calculation amount And improve prediction accuracy.
Patent Information
- Application Number
- CN202411171511.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-26
- Publication Date
- 2024-11-29
AI Technical Summary
The prior art When evaluating structural reliability, especially engineering problems involving implicit performance functions, the calculation efficiency is low and the result error is large, especially when the performance function has multiple design points or is highly nonlinear, the existing methods are as follows. The calculation cost of order reliability method, Monte Carlo simulation method, etc. is high and difficult to meet actual needs.
Combining the adaptive correlation vector machine and Markov chain important sampling (RVM-MIS) method, important samples are generated through the improved Markov chain important sampling algorithm, initial RVM is constructed, and experimental design is updated using active learning functions until the stop condition is met. , realize high-precision structural reliability evaluation.
It significantly reduces the amount of calculation, improves evaluation efficiency and accuracy, and can accurately determine samples that affect RVM prediction accuracy, making RVM updates more efficient and predictive results more accurate, and does not require the identification of design points and the construction of important sampling density. .
Smart Images

Figure CN119047318A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of structural reliability assessment, and in particular to a reliability assessment method combining an adaptive relevance vector machine with Markov chain importance sampling. Background Art
[0002] In actual engineering systems, there are many uncertainties in geometric parameters, loads, and material properties. Reliability analysis of engineering problems aims to estimate the probability of failure under these uncertainties. In addition, many engineering problems involve implicit performance functions. Evaluating the failure probability of these problems usually requires a large number of time-consuming finite element simulations, and the computing time required to obtain accurate results is usually unbearable. In order to improve computational efficiency, it is necessary to reduce the number of expensive model simulations. In recent years, in order to effectively and accurately evaluate the failure probability of structures, various reliability assessment methods have been proposed, including approximation methods, sampling simulation methods, and surrogate model methods.
[0003] The first-order reliability method and the second-order reliability method are classic approximation methods. However, when the performance function has multiple design points or is highly nonlinear, the results obtained by these two methods have large errors. The Monte Carlo simulation method is widely used as a reference method. It is the most intuitive and accurate method, but the computational cost is prohibitive. In order to overcome the shortcomings of the Monte Carlo simulation method, many variance reduction methods have been developed to reduce the number of calls to the true performance function, such as line sampling, importance sampling, subset simulation, etc. However, the computational burden of these methods is still unacceptable in practical engineering problems.
[0004] In recent years, in order to further improve the computational efficiency of reliability analysis, surrogate model methods have received widespread attention. Commonly used surrogate models include neural networks, response surface methods, support vector machines, and Kriging models, which are used to approximate the real limit state. Among them, the most widely used is the Kriging model because it can provide prediction variance information. Based on this feature, various active learning functions have been proposed to reduce the number of actual performance function evaluations required to build a high-precision Kriging model.
[0005] Currently, the relevance vector machine (RVM), a machine learning method for classification and regression, provides a new tool for reliability assessment. RVM is based on a probabilistic Bayesian learning framework that overcomes the shortcomings of support vector machines. Compared with similar support vector machines, it can derive more accurate prediction models with fewer kernel functions, and also has many other advantages, such as automatic estimation of complex parameters, convenience of using arbitrary kernel functions, good generalization performance, and excellent sparsity. In addition, RVM can also achieve probabilistic prediction similar to the Kriging model, which provides the predicted value of the sample and the corresponding variance, which allows RVM to be incorporated into existing active learning functions to improve its prediction accuracy. Summary of the invention
[0006] In view of the above problems, the technical problem to be solved by the present invention is to provide a more accurate and efficient reliability assessment method to meet the needs of practical engineering problems.
[0007] In order to achieve the above-mentioned purpose, the present invention proposes a reliability assessment method combining adaptive relevance vector machine and Markov chain importance sampling (RVM-MIS). The implementation process of the method is as follows: based on the improved Markov chain importance sampling algorithm (MIS), important samples for reliability assessment are generated; the actual evaluation value of the important sampling center is used as the initial experimental design to establish the initial RVM; based on the constructed initial RVM, the predicted mean and variance of each important sample are obtained; based on the active learning function U, the experimental design is updated and the RVM is rebuilt until the stopping condition is met; and the structural reliability is evaluated using the high-precision RVM.
[0008] Furthermore, the use of MIS to generate important samples can start important sampling by selecting an initial sample x1 as long as it is in the fault area, and use it as the first important sampling center.
[0009] Furthermore, based on the jth important sampling center, using the distribution f(τ|x j ) Generate K samples where f(τ|x j ) can be expressed as:
[0010]
[0011] Furthermore, randomly select from K samples And the ratio r is calculated by the following formula.
[0012]
[0013] Furthermore, if r>1, then accept As the j+1th important sampling center x j+1 Otherwise, The sampling center is regarded as rejected and randomly reselected from the remaining samples until the ratio r>1.
[0014] Furthermore, the above steps are repeated until M important sampling centers [x1, x2, ..., x M ].
[0015] Furthermore, the important sampling density function h is constructed by moving the sampling center from the mean to the i-th important sampling center. i (x);
[0016] Furthermore, through each important sampling density function h i (x) Generate H important samples.
[0017] Furthermore, the unbiased estimate of the failure probability can be calculated as follows:
[0018]
[0019] where β i is the weight coefficient whose value is 1 / M, It represents the importance sampling density function h i (x) The kth sample generated.
[0020] Furthermore, the initial RVM for reliability assessment is constructed by using actual assessment values of important sampling centers as an initial experimental design.
[0021] Furthermore, based on the constructed initial RVM, the predicted mean and variance of each important sample are obtained.
[0022] Furthermore, based on the active learning U function, we search for important samples with the smallest U value, use them as update samples, and add experimental design to improve the prediction accuracy of RVM until the stop condition is reached. The sample U value can be calculated by the following formula:
[0023]
[0024] Furthermore, when the U values of all important samples are greater than or equal to 2, it is considered that the update stop condition is reached and the update of RVM is stopped.
[0025] Finally, the structural reliability is evaluated using high-precision RVM, so that the failure probability of the structure can be obtained accurately and efficiently.
[0026] Compared with the prior art, the present invention has the following advantages and effects:
[0027] 1. The present invention inherits the advantages of RVM in accurately approaching the real limit state and MIS in effectively filling important fault areas, and significantly reduces the amount of calculation while ensuring the accuracy of reliability results.
[0028] 2. Using the U function in the active learning process can accurately determine the samples that have a significant impact on the RVM prediction accuracy, making the RVM update more efficient and the prediction results more accurate.
[0029] 3. Compared with previous important sampling methods, MIS does not require the identification of design points and the construction of important sampling density, making the generation of important samples more efficient. BRIEF DESCRIPTION OF THE DRAWINGS
[0030] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0031] Figure 1 The present invention is a flow chart of a reliability assessment method combining an adaptive relevance vector machine and Markov chain importance sampling.
[0032] Figure 2 The actual performance function value and predicted value graph of RVM-MIS and its comparison method.
[0033] Figure 3 This is a visualization result analysis diagram of the RVM-MIS method.
[0034] Figure 4 The results comparison chart of relevance vector machine and support vector machine. DETAILED DESCRIPTION
[0035] In order to make the objectives, technical solutions and advantages of the present invention more clear, the technical solutions of the present invention will be described in detail below with reference to the accompanying drawings.
[0036] like Figure 1 As shown, the present invention provides a reliability assessment method combining an adaptive relevance vector machine with Markov chain importance sampling, and its detailed process mainly consists of 8 steps:
[0037] 1. Generate M×H important samples using the improved Markov chain importance sampling algorithm;
[0038] 2. Construct the initial RVM using the actual evaluation values of M0 important sampling centers;
[0039] 3. Predict each important sample through the constructed RVM to obtain the predicted mean and variance of each important sample;
[0040] 4. Calculate the U value of each important sample, and regard the sample with the smallest U value as the updated sample. If the U value of the updated sample meets the stop condition, the program executes step 6, otherwise executes step 5;
[0041] 5. Use the true performance function to evaluate the updated samples, add them to the initial DOE, update the DOE and rebuild the RVM.
[0042] Then, the program returns to step 2;
[0043] 6. Estimate the failure probability and coefficient of variation using the constructed RVM;
[0044] 7. If the obtained coefficient of variation is higher than 0.05, it is considered that the number of current important samples is too small to produce inaccurate results. Another set of important samples continues to be generated through MIS. When the number of important samples increases, the program returns to step 3;
[0045] 8. If the coefficient of variation is less than 0.05, terminate the program and use the latest failure probability assessment value as the final result.
[0046] A classic numerical example is used to verify the efficiency and accuracy of RVM-MIS. The example also uses some other mainstream reliability assessment methods, including MCS, IS, AK-MCS, AK-IS and AK-MIS. The results of MCS can be used as a benchmark for comparing accuracy. In addition, RVM-MIS and AK-MIS use the same initial important samples, so the results obtained can reflect the performance of RVM.
[0047] The example is a high-order problem whose performance function is:
[0048]
[0049] where d1 and d2 are independent standard normal variables. The results of these methods are shown in the following table, with the total number of samples N s , the number of times the performance function is called N call , failure probability P f , Coefficient of variation of failure probability COVP f and the failure probability error percentage ε compared to the benchmark. ε and N call They are two important indicators for evaluating the accuracy and efficiency of a method respectively.
[0050]
[0051] For AK-MCS N call , “10+24” means 10 initial samples and 24 samples added in the active learning process. call, “21+13” means 21 approximate samples and 13 added samples. For RVM-MIS, “10+16+7” means 10 important sampling centers, 16 rejected sampling centers, and 7 added samples.
[0052] As can be seen from Table 1, RVM-MIS has made considerable improvements in computational efficiency while also providing extremely accurate failure probabilities. s and N call The number is much lower than that of MCS, which only calls the true performance function of 33 samples. In addition, RVM-MIS does not require the determination of the design point. Therefore, the computational complexity of RVM-MIS is lower than that of IS and AK-IS. Compared with AK-MCS, the error is slightly larger. However, due to N s The number of N is large, and the calculation time of AK-MCS to select the updated sample is too long. AK-MCS needs 2716 seconds to solve this problem, while RVM-MIS only needs 47 seconds. Therefore, compared with AK-MCS, RVM-MIS still shows its advantage. Since RVM-MIS and AK-MIS use the same important samples and initial experimental design, N s The number of samples and the number of sample centers are the same. The active learning process of RVM-MIS only adds 7 samples, while AK-MIS adds 10 samples. The failure probability estimated by RVM-MIS is slightly higher than that of AK-MIS, indicating that the performance of the RVM model is better than that of the Kriging model. The actual performance function values and predicted values of RVM-MIS and AK-MIS for 25 important samples are shown in Figure 2. Figure 2 As shown in the figure, it can be seen that in most cases, the predicted value of RVM-MIS is closer to the actual value. The predicted sign of RVM-MIS for each sample is consistent with the actual sign, but there is a prediction sign error for AK-MIS, which verifies that the failure probability estimated by RVM-MIS is more accurate than that of AK-MIS.
[0053] Important sampling centers, rejected sampling centers, important samples, added samples and predicted limit states of RVM-MIS are as follows: Figure 3 As shown in the figure, it can be seen that the 10 important sampling centers are all located in the failure area. The sample locations added during the active learning process are close to the actual limit state. In addition, the important samples generated by MIS cover the failure area well. It can be seen from the figure that the limit state predicted by RVM-MIS is very accurate and basically consistent with the actual limit state.
[0054] Sparsity is one of the outstanding features of RVM. RVM is considered to be more sparse than SVM because it can obtain more accurate prediction results with fewer kernel functions. To verify this conclusion, RVM and SVM were constructed based on the same training samples used by RVM-MIS. The results are shown in Figure 2. Figure 4As shown in the figure, it can be seen that the limit states predicted by RVM and SVM are almost the same and very accurate. The number of relevant vectors of RVM is 7, and the number of support vectors of SVM is 16, which verifies that the sparsity of RVM is better.
[0055] The above are only exemplary embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.
Claims
1. A reliability assessment method combining adaptive relevance vector machine and Markov chain importance sampling, characterized in that: Based on the improved Markov chain importance sampling algorithm, it can effectively generate important samples to fill important fault areas; Based on the active learning function, the initial experimental design is updated and the relevance vector machine is reconstructed. The high-precision relevance vector machine is further used to evaluate the structural reliability efficiently and accurately.
2. The reliability assessment method combining adaptive relevance vector machine and Markov chain importance sampling according to claim 1, characterized in that: The improved Markov chain importance sampling algorithm generates a series of important sampling centers through the Markov chain algorithm, moves the initial sampling center from the mean to each important sampling center, constructs their respective important sampling density functions, and generates a series of important samples at each important sampling center based on the important sampling density function, so that the important samples can effectively fill the fault area.
3. The reliability assessment method combining adaptive relevance vector machine and Markov chain importance sampling according to claim 1, characterized in that: The related vector machine first constructs an initial related vector machine through the actual evaluation value of the important sampling center as the initial experimental design, uses the active learning U function to find the important samples with the smallest U value as the update samples, and adds the update samples to the experimental design to improve the prediction accuracy of the related vector machine until the stop condition of the related vector machine update is reached.
