Structural reliability updating method based on domain adaptation network integration

Abstract

The application discloses a structure reliability updating method based on domain self-adapting network integration, comprising the following steps: constructing a source domain and a target domain Monte Carlo sample pool; constructing a proxy model based on deep integration and active learning as a source domain proxy model; generating a migration training set sample point; constructing a domain self-adapting network integration proxy model as a target domain proxy model; selecting a fine-tuning training set sample point based on a learning function; calculating a sample point limit state function value to obtain a fine-tuning training set; partially fine-tuning the target domain proxy model; adopting a Monte Carlo method to estimate a failure probability and judge whether the failure probability meets a convergence criterion; and outputting a failure probability estimation value updated based on load variable distribution change. The method can accurately and efficiently complete failure probability updating and solving according to load variable distribution change, is helpful for quantifying the safety degree of structure continued service in real time, and thus can help in formulating more scientific and economic maintenance measures and overhaul plans.

Description

Technical Field

[0001] This invention belongs to the field of structural reliability analysis technology, specifically relating to a structural reliability update method based on domain adaptive network integration. Background Technology

[0002] To address the various uncertainties arising from load conditions, material properties, and manufacturing processes in complex engineering problems, probabilistic structural reliability methods have become essential for accurate structural safety assessment. These methods treat various uncertainties as random variables, quantifying them through probabilistic approaches to better balance safety and economy. The limit state functions of complex engineering structures often possess implicit, multidimensional, and highly nonlinear characteristics, requiring time-consuming finite element models for calculation. To improve the efficiency of structural reliability calculations, fitting complex implicit limit state functions using surrogate models to solve for failure probabilities has become a research hotspot.

[0003] However, during service, the performance of a structure does not remain static but is affected by various time-related and environmental factors, including but not limited to load variations, environmental corrosion, material aging, and accidental effects. These factors can lead to a decrease in structural strength and consequently affect overall reliability. Therefore, by comprehensively considering the various uncertainties of the structure and dynamically updating and adjusting the prediction model by incorporating real monitoring information, the deviation between the predicted results and the actual situation can be effectively reduced.

[0004] Considering that some structures are not convenient to directly monitor stress and strain information, updates can only be completed through indirect monitoring variables such as loads. This means that the distribution of input variables in structural reliability analysis will change. Changes in the distribution of input variables will lead to corresponding changes in the failure boundary, resulting in a decrease in the local fitting accuracy of the original surrogate model to the failure boundary. The process of reconstructing a new surrogate model is time-consuming, ignores the usability of the original surrogate model, and has low computational efficiency. Summary of the Invention

[0005] To address the shortcomings of the existing technologies, the present invention aims to provide a structural reliability update method based on domain adaptive network integration, thereby solving the problem of low efficiency in structural reliability updates in the existing technologies.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] The present invention provides a structural reliability update method based on domain adaptive network integration, comprising the following steps:

[0008] 1) Determine the distribution of the load variables before and after the change, and use them as the source domain distribution and target domain distribution, respectively;

[0009] 2) Construct source domain Monte Carlo sample pools and target domain Monte Carlo sample pools;

[0010] 3) Construct a proxy model based on deep ensemble and active learning according to the source domain distribution, as the source domain proxy model;

[0011] 4) The source domain transfer training set sample points and the target domain transfer training set sample points are generated using the uniform distribution corresponding Latin hypercube sampling method. The source domain transfer training set sample points are used to calculate the corresponding limit state function values ​​through the source domain proxy model.

[0012] 5) Construct a domain-adaptive network integrated proxy model based on the source domain proxy model and the total loss function, and use it as the target domain proxy model;

[0013] 6) Select fine-tuned training set sample points from the Monte Carlo sample pool of the target domain based on an active learning strategy;

[0014] 7) Calculate the limit state function values ​​of the sample points in the fine-tuning training set to obtain the fine-tuning training set;

[0015] 8) The target domain proxy model is partially fine-tuned, and the Monte Carlo method is used to estimate the structural failure probability;

[0016] 9) Determine whether the structural failure probability meets the convergence criterion. If it does, end the process and output the current failure probability to complete the failure probability calculation based on the load variable distribution change update. If it does not meet the criterion, proceed to step 10).

[0017] 10) Select new sample points based on the learning function value and add them to the fine-tuning training set sample points, then return to step 7).

[0018] Furthermore, in step 2), based on the probability density distributions of the source domain random variables and the target domain random variables, respectively, the inverse transformation method is used to obtain variable samples that follow the distribution function, and a source domain Monte Carlo sample pool S of the random variables is constructed. MC_s Target Domain Monte Carlo Sample Pool S MC_t .

[0019] Furthermore, step 3) specifically includes:

[0020] Based on the source domain distribution, an initial training set of sample points is generated using the Latin hypercube sampling method corresponding to a uniform distribution. The true limit state function values ​​of the training set sample points are calculated to obtain the training set. A weighted mean square error loss function is established based on the limit state function values. M artificial neural network sub-models are constructed based on the weighted mean square error loss function, and a surrogate model based on deep ensemble and active learning is constructed through uniform weighted mixing. The direct Monte Carlo method is used to estimate the structural failure probability. The learning function values ​​of all sample points in the source domain Monte Carlo sample pool are calculated, new sample points are selected and added to the training set sample points, and iterative training is performed until the failure probability convergence criterion is met.

[0021] Further, in step 4), based on the upper and lower bounds of the source domain Monte Carlo sample pool and the target domain Monte Carlo sample pool, a uniformly distributed Latin hypercube sampling method is used to sample within the interval, selecting n transfer training set sample points X respectively. s ={x s1 ,x s2 ,x s3 ,...,x sn}, X t ={x t1 ,x t2 ,x t3 ,…,x tn}, where X s The corresponding limit state function value is output through the source domain proxy model, x. sn X represents the source domain transfer of training set sample points. s For x sn The set of x tn X represents the target domain transfer training set sample points. t For x tn A set of.

[0022] Furthermore, step 5) specifically includes:

[0023] Using the constructed source domain proxy model, the network structures of the artificial neural network sub-models are read respectively to complete the construction of M domain adaptive network sub-models. Each domain adaptive network sub-model consists of three parts: a feature extraction module, an adaptive module, and a regression module. The feature extraction module is shared by the source and target domain networks and is used to extract deep features from the source and target domain data samples. The input layer and the first hidden layer of the artificial neural network sub-model are selected as the feature extractors. The adaptive module introduces the maximum mean difference to calculate the distribution difference between the source and target domain deep features. The calculation expression is as follows:

[0024]

[0025] In the formula, the subscript H denotes the regenerated kernel Hilbert space; n s ,nt φ(·) represents the number of samples in the source and target domains, respectively; φ(·) represents the nonlinear feature mapping function that maps the original data to the RHKS space; L MMD (D s D t () represents the maximum mean difference between the source domain distribution and the target domain distribution;

[0026] The regression module consists of a fully connected layer network, used to establish the mapping relationship between deep features and predicted response values;

[0027] The total loss function is a loss function that comprehensively considers regression prediction error and inter-domain differences, and its calculation expression is as follows:

[0028] L = L MSE (X L ,y)+λL MMD (D s D t )

[0029] In the formula, L MSE (X L (x, y) represents the regression loss of the domain adaptive network sub-model, and (x, y) represents the difference between the predicted values ​​and the label values ​​on the source domain training set, using the MSE loss function; L Let be the label value of the source domain sample, and y be the predicted value of the source domain sample after passing through the entire domain adaptive network sub-model; L MMD (D s D t ) represents the distributional difference between the source and target domain data; λ is a constant coefficient; L is the total loss function;

[0030] After constructing and training the domain adaptive network sub-models, the M domain adaptive network sub-models are uniformly weighted and mixed to obtain the domain adaptive network ensemble proxy model, which serves as the target domain proxy model; the domain adaptive network ensemble proxy model outputs sample points x. i The limit state function predicts the mean and standard deviation as follows:

[0031]

[0032] In the formula, μ(x) i ) represents the sample point x i The limit state function prediction mean represents the predicted output value of the domain adaptive network ensemble model for sample point xi; σ(x i ) represents the sample point x i The standard deviation of the limit state function prediction represents the domain adaptive network ensemble model's prediction of sample point x. i The uncertainty in output value prediction; M is the number of sub-models in the domain adaptive network; Represents the adaptive network sub-model of the j-th domain. For sample point x i The predicted value of the limit state function, whose model parameter is θ. j .

[0033] Furthermore, in step 6), the training set sample points are adaptively selected and fine-tuned using an active learning strategy, and the target domain Monte Carlo sample pool S is calculated. MC_t The learning function values ​​U(x) of all sample points are used to select the k sample points T with the smallest U(x) values ​​as the fine-tuning training set. t The expression for calculating U(x) is:

[0034]

[0035] In the formula, μ(x) i ) and σ(x i ) represent sample points x respectively i The limit state function predicts the mean and standard deviation.

[0036] Furthermore, in step 7), the limit state function value G of the fine-tuning training set sample points is calculated by calling the limit state function or the finite element model. t The fine-tuned training set {T} is obtained. t G t}

[0037] Furthermore, step 8) specifically includes:

[0038] For each domain adaptive network sub-model in the target domain proxy model, the parameters of the input layer and the first hidden layer are first fixed, and the output layer parameters are retrained based on the fine-tuned training set; then the parameter freeze is lifted, and all network parameters are fine-tuned, at which point the model parameters only undergo minor adjustments; the structural failure probability estimate is calculated based on the Monte Carlo method. The expression is as follows:

[0039]

[0040] In the formula, N F N is the number of sample points where the limit state function value is less than 0, and S is the sample pool. MC_t Number of sample points in the middle.

[0041] Furthermore, the convergence criterion expression in step 9) is:

[0042]

[0043] in, This represents the estimated failure probability in the τth iteration. e represents the estimated failure probability in the (τ-1)th iteration;c Set the threshold and set e c =0.005; convergence condition is that the convergence criterion is met for three consecutive iterations.

[0044] Furthermore, step 10) specifically includes:

[0045] Calculate the target domain sample pool S MC_t The learning function values ​​U(x) for all sample points are calculated, and the minimum value U is selected. min The corresponding sample points are added to the fine-tuning training set as new sample points.

[0046] The beneficial effects of this invention are:

[0047] This invention introduces the maximum mean difference to calculate inter-domain differences and constructs a domain-adaptive network model to update the surrogate model. Network training is completed by comprehensively considering the regression loss and the total loss function of the maximum mean difference, thereby minimizing prediction error and inter-domain differences.

[0048] This invention combines an active learning strategy to adaptively sample the points where the model fine-tuning effect is optimal, thereby reducing the number of sample points required for model fine-tuning.

[0049] This invention can make full use of the original surrogate model to complete the efficient and accurate update of the structural failure probability when the distribution of input variables changes. Attached Figure Description

[0050] Figure 1 This is a flowchart of the method of the present invention.

[0051] Figure 2 This is a schematic diagram of the cantilevered cylindrical structure used in the embodiment.

[0052] Figure 3 This is a schematic diagram of a domain adaptive network sub-model. Detailed Implementation

[0053] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to embodiments and accompanying drawings. The content mentioned in the embodiments is not intended to limit the present invention.

[0054] This embodiment addresses the reliability analysis problem of a cantilever cylindrical structure with 9-dimensional random variables;

[0055] like Figure 2 As shown, considering the structural yield strength failure, its limit state function is as follows:

[0056] g(x) = S - σ max

[0057] Where S represents the structural yield strength; σ maxThis indicates the maximum Misses stress experienced by the structure;

[0058] Specifically:

[0059]

[0060] τ zx =(Td / 4I)

[0061] Among them, θ1=5°, θ2=10°; σ x and τ zx Let P be the normal stress and F1 be the torsional stress; F2 be the concentrated forces acting on the structure from the outside; T be the torque acting on the structure; M be the bending moment; A be the cross-sectional area of ​​the tube; and I be the moment of inertia of the cross section. The relevant parameters can be calculated using the following formula:

[0062]

[0063] In the structural reliability update problem, the distribution of the monitored variable before the change and the distribution of other variables are taken as the source domain data distribution, and the distribution of the monitored variable after the change and the distribution of other variables are taken as the target domain data distribution.

[0064] Table 1 below shows the parameters of all random variables before and after the change in the distribution of the load variables:

[0065] Table 1

[0066]

[0067] Reference Figures 1 to 3 As shown, the present invention provides a structural reliability update method based on domain adaptive network integration, comprising the following steps:

[0068] 1) Determine the distribution of the load variables before and after the change, and use them as the source domain distribution and target domain distribution, as shown in Table 1.

[0069] 2) Construct source domain Monte Carlo sample pools and target domain Monte Carlo sample pools;

[0070] Based on the probability density distributions of the source domain random variables and the target domain random variables, respectively, the inverse transformation method is used to obtain variable samples that follow a distribution function, and a Monte Carlo sample pool S of the source domain random variables is constructed. MC_s Target Domain Monte Carlo Sample Pool S MC_t ;

[0071] 3) Construct a proxy model based on deep ensemble and active learning according to the source domain distribution, as the source domain proxy model;

[0072] Based on the source domain distribution, an initial training set of sample points is generated using the Latin hypercube sampling method corresponding to a uniform distribution. The true limit state function values ​​of the training set sample points are calculated to obtain the training set. A weighted mean square error loss function is established based on the limit state function values. M artificial neural network sub-models are constructed based on the weighted mean square error loss function, and a surrogate model based on deep ensemble and active learning is constructed through uniform weighted mixing. The direct Monte Carlo method is used to estimate the structural failure probability. The learning function values ​​of all sample points in the source domain Monte Carlo sample pool are calculated, new sample points are selected and added to the training set sample points, and iterative training is performed until the failure probability convergence criterion is met.

[0073] 4) The source domain transfer training set sample points and the target domain transfer training set sample points are generated using the uniform distribution corresponding Latin hypercube sampling method. The source domain transfer training set sample points are used to calculate the corresponding limit state function values ​​through the source domain proxy model.

[0074] In step 4), based on the upper and lower bounds of the source domain Monte Carlo sample pool and the target domain Monte Carlo sample pool, a uniformly distributed Latin hypercube sampling method is used to sample within the interval, selecting n transfer training set sample points X. s ={x s1 ,x s2 ,x s3 ,...,x sn}, X t ={x t1 ,x t2 ,x t3 ,...,x tn}, where X s The corresponding limit state function value is output through the source domain proxy model, x. sn X represents the source domain transfer of training set sample points. s For x sn The set of x tn X represents the target domain transfer training set sample points. t For x tn The set; in this embodiment, n is 1000.

[0075] 5) Construct a domain-adaptive network integrated proxy model based on the source domain proxy model and the total loss function, and use it as the target domain proxy model;

[0076] Using the constructed source domain proxy model, the network structures of the artificial neural network sub-models are read respectively to complete the construction of M domain adaptive network sub-models. Each domain adaptive network sub-model consists of three parts: a feature extraction module, an adaptive module, and a regression module. The feature extraction module is shared by the source and target domain networks and is used to extract deep features from the source and target domain data samples. The input layer and the first hidden layer of the artificial neural network sub-model are selected as the feature extractors. The adaptive module introduces the maximum mean difference to calculate the distribution difference between the source and target domain deep features. The calculation expression is as follows:

[0077]

[0078] In the formula, the subscript H denotes the regenerated kernel Hilbert space; n s ,n t φ(·) represents the number of samples in the source and target domains, respectively; φ(·) represents the nonlinear feature mapping function that maps the original data to the RHKS space; L MMD (D s D t () represents the maximum mean difference between the source domain distribution and the target domain distribution;

[0079] The regression module consists of a fully connected layer network, used to establish the mapping relationship between deep features and predicted response values;

[0080] The total loss function is a loss function that comprehensively considers regression prediction error and inter-domain differences, and its calculation expression is as follows:

[0081] L = L MSE (X L ,y)+λL MMD (D s D t )

[0082] In the formula, L MSE (X L (x, y) represents the regression loss of the domain adaptive network sub-model, and (x, y) represents the difference between the predicted values ​​and the label values ​​on the source domain training set, using the MSE loss function; L Let be the label value of the source domain sample, and y be the predicted value of the source domain sample after passing through the entire domain adaptive network sub-model; L MMD (D s D t ) represents the distributional difference between the source and target domain data; λ is a constant coefficient; L is the total loss function;

[0083] After constructing and training the domain adaptive network sub-models, the M domain adaptive network sub-models are uniformly weighted and mixed to obtain the domain adaptive network ensemble proxy model, which serves as the target domain proxy model; the domain adaptive network ensemble proxy model outputs sample points x. iThe limit state function predicts the mean and standard deviation as follows:

[0084]

[0085]

[0086] In the formula, μ(x) i ) represents the sample point x i The limit state function prediction mean represents the predicted output value of the domain adaptive network ensemble model for sample point xi; σ(x i ) represents the sample point x i The standard deviation of the limit state function prediction represents the domain adaptive network ensemble model's prediction of sample point x. i The uncertainty in output value prediction; M is the number of sub-models in the domain adaptive network; Represents the adaptive network sub-model of the j-th domain. For sample point x i The predicted value of the limit state function, whose model parameter is θ. j .

[0087] 6) Select fine-tuned training set sample points from the Monte Carlo sample pool of the target domain based on an active learning strategy;

[0088] In step 6), the training set sample points are adaptively selected and fine-tuned using an active learning strategy, and the Monte Carlo sample pool S of the target domain is calculated. MC_t The learning function values ​​U(x) of all sample points are used to select the k sample points T with the smallest U(x) values ​​as the fine-tuning training set. t In this embodiment, k is 10; the expression for calculating U(x) is:

[0089]

[0090] In the formula, μ(x) i ) and σ(x i ) represent sample points x respectively i The limit state function predicts the mean and standard deviation.

[0091] 7) Calculate the limit state function values ​​of the sample points in the fine-tuning training set to obtain the fine-tuning training set;

[0092] The limit state function value G is calculated by calling the limit state function or the finite element model to fine-tune the training set sample points. t The fine-tuned training set {T} is obtained. t G t}

[0093] 8) The target domain proxy model is partially fine-tuned, and the Monte Carlo method is used to estimate the structural failure probability;

[0094] For each domain adaptive network sub-model in the target domain proxy model, the parameters of the input layer and the first hidden layer are first fixed, and the output layer parameters are retrained based on the fine-tuned training set; then the parameter freeze is lifted, and all network parameters are fine-tuned, at which point the model parameters only undergo minor adjustments; the structural failure probability estimate is calculated based on the Monte Carlo method. The expression is as follows:

[0095]

[0096] In the formula, N F N is the number of sample points where the limit state function value is less than 0, and S is the sample pool. MC_t Number of sample points in the middle.

[0097] 9) Determine whether the structural failure probability meets the convergence criterion. If it does, end the process and output the current failure probability to complete the failure probability calculation based on the load variable distribution change update. If it does not meet the criterion, proceed to step 10).

[0098] The convergence criterion expression in step 9) is:

[0099]

[0100] in, This represents the estimated failure probability in the τth iteration. e represents the estimated failure probability in the (τ-1)th iteration; c Set the threshold and set e c =0.005; convergence condition is that the convergence criterion is met for three consecutive iterations.

[0101] 10) Select new sample points based on the learning function value and add them to the fine-tuning training set sample points, then return to step 7).

[0102] The calculation results of reconstructing the target domain surrogate model using the Direct Monte Carlo (MCS) method, the Active Learning Kriging (AK-MCS+U) method, and the surrogate model method based on deep ensemble and active learning (DE-AL), and updating the target domain surrogate model using the method of this invention, are shown in Table 2. Using the calculation results obtained by the Direct Monte Carlo method as a reference for the true value of this structural reliability problem, the failure probability estimate obtained by the method of this invention is closest to the reference value, with a relative error of only -0.72%, which is smaller than the other three methods. At the same time, the number of calls to the limit state function is only 7, significantly lower than other calculation methods. This result shows that the method of this invention can ensure high accuracy of the failure probability estimate after changes in input variables with a small number of limit state function calls; Table 2 shows a comparison of the calculation results as follows:

[0103] Table 2

[0104]

[0105] This invention has many specific applications. The above description is only a preferred embodiment of this invention. It should be noted that for those skilled in the art, several improvements can be made without departing from the principle of this invention, and these improvements should also be considered within the scope of protection of this invention.

Claims

1. A structural reliability updating method based on domain adaptation network integration, characterized in that, The steps are as follows: 1) determine the distribution before and after the change of the load variable, respectively as the source domain distribution and the target domain distribution; 2) construct a source domain Monte Carlo sample pool and a target domain Monte Carlo sample pool; 3) construct a proxy model based on deep integration and active learning according to the source domain distribution, as a source domain proxy model; 4) generate source domain transfer training set sample points and target domain transfer training set sample points by using the Latin hypercube sampling method corresponding to the uniform distribution, and calculate the corresponding limit state function values of the source domain transfer training set sample points by the source domain proxy model; 5) construct a domain adaptive network integrated proxy model based on the source domain proxy model and the total loss function, as a target domain proxy model; 6) select fine-tuning training set sample points in the target domain Monte Carlo sample pool based on an active learning strategy; 7) calculate the limit state function values of the fine-tuning training set sample points to obtain a fine-tuning training set; 8) partially fine-tune the target domain proxy model and estimate the structural failure probability by using the Monte Carlo method; 9) judge whether the structural failure probability meets the convergence criterion, if yes, end and output the current failure probability, and complete the failure probability calculation based on the change of the load variable distribution update; if not, go to step 10); 10) select new sample points based on the learning function value and add them to the fine-tuning training set sample points, and return to step 7).

2. The structural reliability updating method based on domain adaptation network integration according to claim 1, characterized in that, In step 2), based on the probability density distribution of the source domain random variable and the probability density distribution of the target domain random variable respectively, inverse transformation method is used to obtain variable samples subject to distribution function, and a source domain Monte Carlo sample pool S of random variables is constructed MC_s , and a target domain Monte Carlo sample pool S MC_t . 3.The structural reliability updating method based on domain adaptation network integration according to claim 1, characterized in that, The step 3) specifically comprises: generate initial training set sample points by using the Latin hypercube sampling method corresponding to the uniform distribution according to the source domain distribution; calculate the true limit state function values of the training set sample points to obtain a training set; establish a weighted mean square error loss function based on the limit state function values; construct M artificial neural network submodels according to the weighted mean square error loss function, and construct a proxy model based on deep integration and active learning by uniform weighted mixing; estimate the structural failure probability by using the direct Monte Carlo method; calculate the learning function values of all sample points in the source domain Monte Carlo sample pool, select new sample points to add to the training set sample points, and iterate training until the failure probability convergence criterion is met.

4. The structural reliability updating method based on domain adaptation network integration according to claim 1, characterized in that, The step 4) respectively based on the upper and lower bounds of the source domain Monte Carlo sample pool and the target domain Monte Carlo sample pool, adopts the Latin hypercube sampling method corresponding to the uniform distribution to sample in the interval, respectively selects n transfer training set sample points X s ={x s1 ,x s2 ,x s3 ,...,x sn} t =X t1 ={x t2 ,x t3 ,x tn ,...,x s} , wherein X sn outputs the corresponding limit state function value through the source domain proxy model, x s represents the source domain transfer training set sample point, X sn is the set of x tn , x t represents the target domain transfer training set sample point, X tn is the set of x .

5. The structural reliability updating method based on domain adaptation network integration according to claim 1, characterized in that, The step 5) specifically comprises: read the network structure of the artificial neural network submodel through the constructed source domain proxy model, and complete the construction of M domain adaptive network submodels; the domain adaptive network submodel comprises a feature extraction module, an adaptive module and a regression module: the feature extraction module is shared by the source domain and the target domain network, and is used for extracting deep features of the source domain and the target domain data samples, and the input layer and the first hidden layer of the artificial neural network submodel are selected as the feature extractor; the adaptive module introduces the maximum mean difference, calculates the distribution difference between the source domain deep features and the target domain deep features, and the calculation expression is as follows: where subscript H denotes the reproducing kernel Hilbert space; n s ,n t are the number of source and target domain samples, respectively; φ(·) denotes a non-linear feature mapping function that maps the original data into the RHKS space; L MMD (D s ,D t ) denotes the maximum mean discrepancy of source and target domain distributions. the regression module is composed of a fully connected layer network, and is used for establishing the mapping relationship between the deep features and the predicted response values; the total loss function is a loss function considering the regression prediction error and the domain difference, and the calculation expression is as follows: L = L MSE (X L ,y) + λL MMD (D s ,D t ) In the formula, L MSE (X L , y) represents the regression loss of the domain adaptive network submodel, represents the difference between the predicted value and the label value of the source domain training set, and adopts the MSE loss function; X L is the label value of the source domain sample, and y is the predicted value output after the source domain sample passes through the entire domain adaptive network submodel; L MMD (D s , D t ) represents the distribution difference between the source domain and the target domain data; λ is a constant coefficient; and L is a total loss function; After the domain adaptation network sub-model is constructed and trained, M domain adaptation network sub-models are uniformly weighted and mixed to obtain a domain adaptation network integrated proxy model as a target domain proxy model; the domain adaptation network integrated proxy model outputs a limit state function prediction mean and standard deviation of a sample point x i as follows: wherein μ(x i ) is the limit state function prediction mean value of the sample point x i , representing the predicted output value of the sample point x i by the domain adaptive network ensemble model; σ(x i ) is the limit state function prediction standard deviation of the sample point x i , representing the prediction uncertainty of the output value of the sample point x i by the domain adaptive network ensemble model; M is the number of domain adaptive network sub-models; μj(x ) represents the limit state function prediction value of the sample point x i by the jth domain adaptive network sub-model, and the model parameters thereof are θ j .

6. The structural reliability updating method based on domain adaptation network integration according to claim 1, characterized in that, The step 6) adaptively selects the fine-tuning training set sample points by the active learning strategy, selects k sample points with the minimum U(x) value as the fine-tuning training set sample points T by calculating the learning function value U(x) of all sample points in the target domain Monte Carlo sample pool S MC_t t The calculation expression of U(x) is:​ where μ(x i ) and σ(x i ) represent the predicted mean and predicted standard deviation of the limit state function at sample point x i , respectively.

7. The structural reliability updating method based on domain adaptation network integration according to claim 1, characterized in that, The step 7) calculates the limit state function value G of the fine-tuning training set sample points by calling the limit state function or the finite element model t , and obtains the fine-tuning training set {T t , G t}.

8. The structural reliability updating method based on domain adaptation network integration according to claim 1, characterized in that, The step 8) specifically comprises: For each domain adaptation network sub-model in the target domain agent model, the model input layer and the first hidden layer network parameters are fixed, and the output layer parameters are retrained based on the fine-tuning training set; then the parameter freezing is released, and all network parameters are fine-tuned, at this time the model parameters only undergo small adjustments; the structural failure probability estimate value is calculated based on the Monte Carlo method The expression is as follows: In the formula, N F is the number of sample points with limit state function values less than 0, and N is the number of sample points in the sample pool S MC_t .

9. The structural reliability updating method based on domain adaptation network integration according to claim 1, characterized in that, The convergence criterion expression in the step 9) is: wherein, denotes the failure probability estimate of the τth iteration, denotes the failure probability estimate of the (τ-1)th iteration; e c is a threshold value; the convergence criterion is met when three consecutive iterations satisfy the convergence criterion.

10. The structural reliability updating method based on domain adaptation network integration according to claim 1, characterized in that, The step 10) specifically comprises: Compute the target domain sample pool S MC_t The learning function value U(x) of all sample points in the pool S is calculated, and the minimum value U min The sample point corresponding to the minimum value is added to the fine-tuning training set as a new sample point.

Images