Structural reliability analysis method and device, equipment, storage medium and product

Abstract

The invention relates to a structural reliability analysis method and device, equipment, a storage medium and a product. The method comprises the following steps: determining an analysis sample pool of a to-be-analyzed equipment structure; inputting each analysis sample into a pre-trained agent model to obtain a prediction response value of the analysis sample; determining the failure probability of the to-be-analyzed equipment structure based on the predicted response value of the analysis sample; wherein the pre-trained agent model is obtained by performing iterative training on an initial agent model, and each iterative training comprises the following steps: determining optimal samples corresponding to a plurality of different learning functions from a candidate sample pool; based on the accumulated reward value of the last iteration of each learning function, determining the probability value of the current iteration of each learning function; determining an optimal learning function from the learning functions based on the probability value of the current iteration of the learning functions; and retraining the current agent model based on an optimal sample corresponding to the optimal learning function. By adopting the method, the structural reliability analysis efficiency can be improved.

Description

Technical Field

[0001] This application relates to the field of mechanical structure analysis technology, and in particular to a structural reliability analysis method, apparatus, equipment, storage medium, and product. Background Technology

[0002] Reliability is a crucial indicator for evaluating the structural quality of complex equipment. It permeates every stage of the entire lifecycle, from individual components to large, complex systems, including design, manufacturing, testing, use, and maintenance. Therefore, conducting reliability analysis on equipment structures is of significant guiding importance for improving the reliability design capabilities of the equipment manufacturing industry.

[0003] With the rise of machine learning technology, surrogate model-based structural reliability analysis methods are increasingly becoming the mainstream. In these methods, reliability analysis based on surrogate models requires pre-selected active learning strategies to guide the iterative training of the surrogate model. However, these methods are subjectively dependent, which can easily lead to low training efficiency and consequently affect the overall efficiency of reliability analysis. Summary of the Invention

[0004] Therefore, it is necessary to provide a structural reliability analysis method, apparatus, equipment, storage medium, and product that can improve the efficiency of structural reliability analysis in response to the above-mentioned technical problems.

[0005] Firstly, this application provides a structural reliability analysis method, including:

[0006] A sample pool for analyzing the structure of the device to be analyzed is determined; the sample pool includes analytical samples of the structure of the device to be analyzed, and the analytical samples include parameters that affect the reliability of the structure of the device to be analyzed.

[0007] Each analysis sample in the analysis sample pool is input into a pre-trained surrogate model to obtain the predicted response value of the analysis sample.

[0008] The failure probability of the device structure to be analyzed is determined based on the predicted response value of the analyzed sample.

[0009] The pre-trained agent model is obtained by iteratively training an initial agent model, and each iteration includes:

[0010] Determine the optimal sample corresponding to multiple different learning functions from the candidate sample pool;

[0011] Based on the cumulative reward value of the previous iteration of each learning function, the probability value of the current iteration of each learning function is determined;

[0012] The optimal learning function is determined from among all learning functions based on the probability value of the current iteration of the learning function.

[0013] The current agent model is retrained based on the optimal sample corresponding to the optimal learning function.

[0014] In one embodiment, each iteration of training further includes: for each learning function, determining the immediate reward value of the learning function in this iteration based on the optimal sample of the learning function; determining the cumulative reward value of the learning function in this iteration based on the immediate reward value of the current iteration and the cumulative reward value of the previous iteration; the cumulative reward value of the current iteration is used to calculate the probability value of the next iteration.

[0015] In one embodiment, determining the immediate reward value of each learning function in this iteration based on the optimal sample of each learning function includes: inputting the optimal sample corresponding to each learning function into the current surrogate model to obtain the predicted response value corresponding to each optimal sample; and determining the immediate reward value of the corresponding learning function in this iteration based on the negative of the absolute value of the predicted response value corresponding to the optimal sample.

[0016] In one embodiment, determining the cumulative reward value of the learning function for the current iteration based on the immediate reward value of the current iteration and the cumulative reward value of the previous iteration includes: determining the cumulative reward value of the learning function for the current iteration based on the immediate reward value of the current iteration, the cumulative reward value of the previous iteration, and the memory factor; wherein the memory factor is greater than 0 and less than 1.

[0017] In one embodiment, determining the probability value of each learning function in the current iteration based on the cumulative reward value of the previous iteration of each learning function includes: normalizing the cumulative reward value of each learning function in the previous iteration according to the maximum reward value and the minimum reward value to obtain each normalized cumulative reward value; wherein the maximum reward value and the minimum reward value are the maximum and minimum values ​​of the cumulative reward values ​​of each learning function in the previous iteration, respectively; for each learning function, determining the probability value of the learning function in the current iteration according to the balance factor and the normalized cumulative reward value; wherein the balance factor is a positive number less than or equal to 10.

[0018] In one embodiment, determining the failure probability of the structure to be analyzed based on the predicted response value of the analysis sample includes: determining the number of failure samples based on the predicted response value of the analysis sample; and determining the failure probability of the device structure to be analyzed based on the number of failure samples and the total number of samples in the analysis sample pool.

[0019] Secondly, this application also provides a structural reliability analysis device, comprising:

[0020] The first determining module is used to determine the analysis sample pool of the structure of the device to be analyzed; the analysis sample pool includes analysis samples of the structure of the device to be analyzed, and the analysis samples include parameters that affect the reliability of the structure of the device to be analyzed;

[0021] The input module is used to input each analysis sample in the analysis sample pool into the pre-trained surrogate model to obtain the predicted response value of the analysis sample.

[0022] The second determining module is used to determine the failure probability of the structure of the device to be analyzed based on the predicted response value of the analysis sample;

[0023] The pre-trained agent model is obtained by iteratively training an initial agent model, and during each iteration of training, the device further includes:

[0024] The third determination module is used to determine the optimal sample corresponding to multiple different learning functions from the candidate sample pool;

[0025] The fourth determining module is used to determine the probability value of each learning function in the current iteration based on the cumulative reward value of each learning function in the previous iteration;

[0026] The fifth determining module is used to determine the optimal learning function from among the learning functions based on the probability value of the learning function in this iteration;

[0027] The retraining module is used to retrain the current agent model based on the optimal samples corresponding to the optimal learning function.

[0028] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the structural reliability analysis method provided in the first aspect of this application.

[0029] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the structural reliability analysis method provided in the first aspect of this application.

[0030] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, implements the structural reliability analysis method provided in the first aspect of this application.

[0031] The aforementioned structural reliability analysis method, apparatus, computer equipment, computer-readable storage medium, and computer program product determine an analysis sample pool for the structure of the device to be analyzed. Each analysis sample in the analysis sample pool is a multi-dimensional vector. Each analysis sample in the analysis sample pool is input into a pre-trained surrogate model to obtain the predicted response value of the analysis sample. The failure probability of the structure of the device to be analyzed is determined based on the predicted response value of the analysis sample. The pre-trained surrogate model is obtained by iteratively training an initial surrogate model. Each iteration includes: determining the optimal sample corresponding to multiple different learning functions from the candidate sample pool; determining the probability value of each learning function in this iteration based on the optimal sample of each learning function; determining the optimal learning function from all learning functions based on the probability value of the learning function in this iteration; and retraining the current surrogate model based on the optimal sample corresponding to the optimal learning function. This application employs multiple different learning functions and determines the optimal learning function in each iteration. The optimal learning function is then used as the learning function for the current iteration training, and the current surrogate model is retrained based on the optimal samples corresponding to the optimal learning function. This achieves active learning based on dynamically updated learning functions, rather than manually selecting a fixed learning function. This directly eliminates the process of manually selecting an active learning function, effectively avoids interference from human factors on the analysis results, avoids the trouble caused by manual selection, and improves training efficiency, thereby enhancing the overall efficiency of structural reliability analysis. Attached Figure Description

[0032] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0033] Figure 1 This is a diagram illustrating the application environment of the structural reliability analysis method in one embodiment.

[0034] Figure 2 This is a flowchart illustrating a structural reliability analysis method in one embodiment;

[0035] Figure 3 This is a schematic diagram of the training process for each iteration in one embodiment;

[0036] Figure 4 This is a schematic diagram of the training process for each iteration in another embodiment;

[0037] Figure 5 In one embodiment Figure 3 A flowchart illustrating step 302 in the middle section;

[0038] Figure 6 This is a schematic diagram illustrating the principle of combining and configuring active learning strategies in one embodiment;

[0039] Figure 7 This is a schematic diagram illustrating the determination of the optimal learning function based on probability values ​​during each iteration in one embodiment;

[0040] Figure 8 This is a flowchart illustrating the training process for each iteration in a specific example.

[0041] Figure 9 This is a structural diagram of a roof truss structure in a specific example.

[0042] Figure 10 This is a structural block diagram of a structural reliability analysis device in one embodiment;

[0043] Figure 11 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0044] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0045] It should be noted that the terms "first," "second," etc., used in this application can be used to describe various elements, but these elements are not limited by these terms. These terms are only used to distinguish the first element from the second element. The terms "comprising" and "having," and any variations thereof, used in this application, are intended to cover non-exclusive inclusion. The term "multiple" used in this application refers to two or more. The term "and / or" used in this application refers to one of the embodiments, or any combination of multiple embodiments.

[0046] In practical engineering problems involving structural reliability analysis, these problems often exhibit characteristics such as high dimensionality, computational complexity, and strong nonlinearity. For real-world engineering problems, when performing reliability analysis based on surrogate models, selecting an appropriate learning function can significantly reduce the computational iteration process, effectively decreasing the computational and time costs of structural reliability analysis and thus improving its engineering practicality. However, in related technologies, reliability analysis based on surrogate models requires pre-selected active learning strategies to guide the iterative training of the surrogate model, which is highly dependent on subjectivity and easily leads to low training efficiency, thereby affecting the overall efficiency of reliability analysis and reducing its engineering practicality. Therefore, this application proposes a structural reliability analysis method that improves the training efficiency of the surrogate model and enhances its engineering practicality.

[0047] The structural reliability analysis method provided in this application can be applied to, for example... Figure 1 In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be integrated onto server 104, or it can be located in the cloud or on other network servers. Terminal 102 can be, but is not limited to, various personal computers, laptops, smartphones, tablets, drones, low-altitude aircraft, IoT devices, and portable wearable devices. IoT devices can include smart speakers, smart TVs, smart air conditioners, smart vehicle devices, projection devices, etc. Server 104 can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services.

[0048] In one exemplary embodiment, such as Figure 2 As shown, a structural reliability analysis method is provided, which can be applied to... Figure 1 Taking the server in the example, the following steps 201 to 203 are used as an example:

[0049] Step 201: Determine the analysis sample pool for the structure of the device to be analyzed.

[0050] The analysis sample pool includes analysis samples of the equipment structure to be analyzed, which include parameters affecting the reliability of the equipment structure. Each analysis sample in the analysis sample pool is a multi-dimensional vector. The analysis samples of the equipment structure to be analyzed include both samples of the overall structure and samples of each component within the structure.

[0051] The analysis sample pool consists of samples drawn from the probability distribution of the input random variables of the equipment structure under analysis, used to evaluate the overall reliability of the equipment structure. Each analysis sample refers to a combination of specific parameter values ​​of the input random variables that affect the structural reliability, i.e., a multi-dimensional vector. For example, when there are 6 input random variables, each analysis sample is a 6-dimensional vector. The number of analysis samples is set according to actual needs.

[0052] For example, the server first obtains the input random variables (such as material properties, geometric parameters, etc.) and their probability distribution type (such as normal distribution) and distribution parameters (such as mean and standard deviation) of the equipment structure to be analyzed, uploaded by the operator, and loads them into memory. Then, the server calls a built-in or linked parameter generation algorithm through its processor, and by executing the algorithm, randomly generates an analysis sample pool based on the probability distribution of the input random variables. Optionally, a Monte Carlo method is used to randomly sample based on the joint probability distribution of the input random variables to generate an analysis sample pool, and the analysis samples are stored in memory for later use. The analysis sample pool includes multiple analysis samples, each of which is a multi-dimensional vector, with the dimension being the number of variables. For example, for three input random variables A, B, and C of a certain bridge structure, one analysis sample (three-dimensional vector) can be: {A1, B1, C1}, where A1, B1, and C1 are the specific parameter values ​​of the corresponding variables. A1 follows the probability distribution of the corresponding input random variable A, B1 follows the probability distribution of variable B, C1 follows the probability distribution of variable C, and {A1,B1,C1} follows the joint probability distribution of all variables.

[0053] Step 202: Input each analysis sample in the analysis sample pool into the pre-trained surrogate model to obtain the predicted response value of the analysis sample.

[0054] The pre-trained proxy model is obtained by iteratively training an initial proxy model. The initial proxy model is trained on an initial sample set based on the sample structure. Optionally, the training process of the initial proxy model includes:

[0055] First, determine the probability distribution type (e.g., normal distribution) and distribution parameters (e.g., mean and standard deviation) of the input random variables for the structural reliability analysis problem of the sample equipment. Then, establish the limit state function for the reliability analysis problem based on actual engineering requirements, as shown below:

[0056] (1)

[0057] Where x represents an N-dimensional vector consisting of independent input random variables, and N is the number of input random variables. This is the limit state function of the structure, used to determine whether the object being analyzed is safe. If the critical threshold of the limit state function is given, then the safe region and failure region of the sample structural response can be defined as follows:

[0058] (2)

[0059] Then, an initial sample pool is drawn according to the probability distribution (for example, containing 2N+2 initial samples, each of which is an N-dimensional vector, where N is the number of input variables).

[0060] Then, the original model or real simulation model of the sample structure is called, and each initial sample in the initial sample pool is substituted into the limit state function to calculate the real response value corresponding to each initial sample. This yields the initial data pair (initial sample and its corresponding real response value). Based on the initial data pair, an initial surrogate model (such as the Kriging surrogate model) is trained.

[0061] After obtaining the initial proxy model, a pre-trained proxy model is obtained by iteratively training the initial proxy model. For example... Figure 3 As shown, each iteration of training includes steps 301 to 304:

[0062] Step 301: Determine the optimal sample corresponding to multiple different learning functions from the candidate sample pool.

[0063] Prior to this step, a candidate sample pool for the sample structure and several different learning functions for structural reliability analysis can be predetermined. Optionally, the candidate sample pool (e.g., containing 10...) is generated by sampling based on the probability distribution of the input random variables of the sample structure. 5 (There are 10 candidate samples, each of which is an N-dimensional vector). Simultaneously, multiple different learning functions can be identified or selected for structural reliability analysis, such as the U-function, Expected Fesibility Function (EFF), Expected Risk Function (ERF), H-function, Reliability-based Lower Confidence Bounding Function (RLCB), and Reliability-based Expected Improvement Function (REIF).

[0064] The optimal sample corresponding to the learning function refers to the candidate sample that makes the learning function value optimal (minimum or maximum) in this iteration. For example, for the U function, if the U value of sample A1 is the minimum, then sample A1 is the optimal sample; for the H function, if the H value of sample A2 is the maximum, then sample A2 is the optimal sample.

[0065] For example, in each iteration, firstly, for each candidate sample in the candidate sample pool, the predicted response value is calculated using the current (in this iteration) surrogate model, and the predicted response value is substituted into each learning function to obtain multiple learning function values ​​for each candidate sample. Then, for each learning function, the optimal sample that makes the learning function reach its optimal value is selected from the candidate sample pool, thus obtaining the optimal sample corresponding to each learning function in this iteration.

[0066] Step 302: Determine the probability value of each learning function in the current iteration based on the cumulative reward value of each learning function in the previous iteration.

[0067] The cumulative reward value of the previous iteration refers to the total reward accumulated by the learning function throughout its history of iterations up to and including the previous iteration, but excluding the current iteration. This cumulative reward value is determined during the previous iteration. The cumulative reward value characterizes the performance of the learning function in each iteration, while the probability value represents the likelihood of it being selected in the current iteration. For each learning function in its history of iterations, better performance results in a higher cumulative reward value and a greater probability value; worse performance results in a lower cumulative reward value and a smaller probability value. The probability value for the current iteration can be understood as the probability of the learning function being selected in this iteration; a higher probability value means a greater chance of being selected in the current iteration.

[0068] For example, at each iteration, the cumulative reward value of each learning function in the previous iteration is obtained. Then, for each learning function, the cumulative reward value of its previous iteration is pre-calculated to obtain the probability value of the learning function in the current iteration. For example, in the t-th iteration, the probability value of the t-th iteration is determined based on the cumulative reward value G(t-1) of the (t-1)-th iteration.

[0069] Optionally, this step includes: selecting the maximum and minimum reward values ​​from the cumulative reward values ​​of the previous iteration of all learning functions; and for each learning function, determining the probability value of the learning function in the current iteration based on the maximum reward value, the minimum reward value, and the cumulative reward value of the learning function in the previous iteration.

[0070] It should be noted that before the first iteration, the initial probability values ​​for each learning function can be set, for example, to 1: ,in, Let represent the initial probability value of the i-th learning function. This represents the number of learning functions. Based on the initial cumulative reward value, the cumulative reward value of each learning function is updated in each iteration according to its performance. The updated reward value is then used to calculate the probability value for the next iteration.

[0071] Step 303: Determine the optimal learning function from among the learning functions based on the probability value of the learning function in this iteration.

[0072] For example, in each iteration, based on the probability value of each learning function in this iteration, a learning function is randomly selected and determined as the optimal learning function for this iteration, avoiding getting trapped in local optima and improving the model's generalization ability. Optionally, the probability value of each learning function in this iteration is mapped one-to-one to an interval between 0 and 1, so each interval corresponds to a learning function. A random number between 0 and 1 is generated, and the interval in which the random number falls is determined as the optimal learning function for this iteration. For example, there are three learning functions: function 1, function 2, and function 3, with probability values ​​of 0.5, 0.3, and 0.2, respectively, and corresponding intervals of [0.0.5), [0.5, 0.8), and [0.8, 1), respectively. If the random number is 0.45 and falls in [0.0.5), then function 1 is selected as the optimal learning function; if r is 0.75 and falls in [0.5, 0.8), then function 2 is selected as the optimal learning function for this iteration.

[0073] Since the probability value is determined in each iteration, and thus the optimal learning function is determined, the optimal learning function is not fixed during the entire iterative training process, but changes dynamically and adaptively. The optimal learning function corresponding to two adjacent iterations may be the same or different.

[0074] Therefore, based on the cumulative reward value of each learning function in the previous iteration, the probability value of each learning function in the current iteration is determined, and the optimal learning function is determined based on this probability value. This allows the probability value to characterize the historical performance of the learning function, thus enabling reliable selection of the optimal learning function based on the probability value. This improves the accuracy of the surrogate model in the current iteration training, thereby shortening the overall iteration training time and reducing time costs.

[0075] Step 304: Retrain the current agent model based on the optimal sample corresponding to the optimal learning function.

[0076] For example, the steps also include: updating the initial sample pool based on the optimal sample corresponding to the optimal learning function to obtain an updated sample pool; using the initial sample pool to train the initial agent model; and retraining the current agent model using the updated sample pool.

[0077] In one possible implementation, the original model or a real simulation model is invoked. The optimal sample corresponding to the optimal learning function is substituted into the limiting state function to calculate the limiting state function value of the optimal sample (which can be called the global optimal sample) corresponding to the optimal learning function. This yields a new data pair (the global optimal sample and its corresponding limiting state function value). This new data pair is added to the initial sample pool to obtain an updated sample pool. Then, the updated sample pool is used to retrain the current surrogate model, realizing the current iteration training of the surrogate model. Each iteration improves the accuracy of the model until the end of the iteration, resulting in a pre-trained surrogate model.

[0078] Each iteration of training also includes: after retraining, based on the current surrogate model, determining whether the current iteration meets the convergence condition. If yes, the current surrogate model (i.e., the surrogate model retrained in this iteration) is determined as the pre-trained surrogate model, and the iteration ends; otherwise, return to step 301 and proceed to the next iteration.

[0079] By configuring the active learning strategy through steps 301 to 304, a pre-trained agent model is obtained through iterative training based on the initial agent model and multiple different learning functions. In each iteration, an optimal learning function and its corresponding globally optimal sample are determined, and retraining is performed. The optimal learning function for any two iterations may be the same or different. Throughout the iterative training process, at least two different optimal learning functions are determined and used. For example, the first iteration uses the U function as the optimal learning function; the second iteration uses the ERF function; the third iteration uses the H function, and so on. This embodiment selects the optimal learning function in each iteration, ensuring that the optimal learning function is used throughout the entire training process. Since a suitable learning function can significantly reduce the computational iteration process and improve training efficiency, this embodiment dynamically selects the optimal learning function to greatly reduce the computational iteration process and improve training efficiency.

[0080] A pre-trained surrogate model is used for reliability analysis of the device structure to be analyzed, wherein the probability distribution of each input random variable of the device structure to be analyzed is the same as the probability distribution of the sample structure.

[0081] For example, the server accesses the memory storing the analysis sample pool, inputs each analysis sample (multidimensional vector) in the analysis sample pool into the pre-trained proxy model (such as the Kriging proxy model) obtained through steps 301 to 304, and obtains the predicted response value (such as the predicted mean) of the analysis sample.

[0082] Step 203: Determine the failure probability of the equipment structure to be analyzed based on the predicted response values ​​of the analysis samples.

[0083] For example, the server divides all analysis samples into safe samples and failed samples based on the relationship between the predicted response value and 0, and determines the failure probability of the reliability analysis problem based on the number of safe samples. This is the reliability analysis result of the device structure to be analyzed. The higher the failure probability, the lower the reliability; the lower the failure probability, the higher the reliability.

[0084] For example, taking a roof truss structure as the device to be analyzed, a server deployed in an engineering computing environment obtains from the terminal multidimensional input random variables (uncertain variables) that affect the mechanical reliability of the structure, as determined by the design specifications and material standards of the structure. These variables include load variables acting on the overall structure, geometric dimension variables of each component, and material property variables. The server, through its processor, performs sampling calculations based on the joint probability distribution of these variables using Monte Carlo sampling, generating an analysis sample pool for reliability assessment and storing it in memory. Each analysis sample represents a possible physical state of the roof truss structure under uncertainty, containing a set of specific physical parameter values ​​for all variables. Then, the server accesses memory and inputs each analysis sample into a pre-trained proxy model called from the model library. This model simulates the physical response process of the roof truss structure and outputs the predicted response value for each analysis sample. Based on the predicted response value, the server calculates the failure probability of the roof truss structure under given uncertainty using a statistical algorithm and outputs it to the terminal, thus obtaining the reliability analysis result, which guides the engineering design and evaluation of the roof truss structure.

[0085] In the above structural reliability analysis method, an analysis sample pool is determined for the structure of the device to be analyzed. The analysis sample pool includes analysis samples of the structure of the device to be analyzed, and the analysis samples include parameters that affect the reliability of the structure of the device to be analyzed. Each analysis sample in the analysis sample pool is input into a pre-trained surrogate model to obtain the predicted response value of the analysis sample. The failure probability of the structure of the device to be analyzed is determined based on the predicted response value of the analysis sample. The pre-trained surrogate model is obtained by iteratively training an initial surrogate model. Each iteration includes: determining the optimal sample corresponding to multiple different learning functions from the candidate sample pool; determining the probability value of each learning function in this iteration based on the optimal sample of each learning function; determining the optimal learning function from each learning function based on the probability value of the learning function in this iteration; and retraining the current surrogate model based on the optimal sample corresponding to the optimal learning function. This application employs multiple different learning functions and determines the optimal learning function in each iteration. The optimal learning function is then used as the learning function for the current iteration training, and the current surrogate model is retrained based on the optimal samples corresponding to the optimal learning function. This achieves active learning based on dynamically updated learning functions, rather than manually selecting a fixed learning function. This directly eliminates the process of manually selecting an active learning function, effectively avoids interference from human factors on the analysis results, avoids the trouble caused by manual selection, and improves training efficiency, thereby enhancing the overall efficiency of structural reliability analysis.

[0086] In one exemplary embodiment, such as Figure 4 As shown, each training iteration also includes steps 401 and 402. Wherein:

[0087] Step 401: For each learning function, determine the instantaneous reward value for this iteration of the learning function based on the optimal sample of the learning function.

[0088] The immediate reward value for this iteration refers to the reward (performance score) obtained by the corresponding learning function in this iteration.

[0089] For example, after obtaining the optimal samples for each learning function, a reward function for the combined active learning strategy can be set for each different learning function. Based on the optimal samples and reward functions of each learning function, the instantaneous reward value of the learning function in this iteration is determined.

[0090] Step 402: Based on the immediate reward value of this iteration and the cumulative reward value of the previous iteration, determine the cumulative reward value of the learning function for this iteration.

[0091] The cumulative reward value of this iteration is used to calculate the probability value of the next iteration.

[0092] The cumulative reward value for this iteration refers to the total cumulative reward obtained by the learning function throughout its history of iterations up to and including this iteration. The cumulative reward value for the previous iteration refers to the total cumulative reward obtained by the learning function throughout its history of iterations up to and including this iteration, but excluding this iteration.

[0093] Example 1: For each learning function, calculate the sum of the current instant reward value and the cumulative reward value of the previous iteration, which is the cumulative reward value of the learning function in this iteration.

[0094] Example 2: For each learning function, after processing the cumulative reward value of the previous iteration according to a preset rule, the sum of the processed cumulative reward value of the previous iteration and the instantaneous reward value of the current iteration is used as the cumulative reward value of the learning function for the current iteration.

[0095] The cumulative reward value of each learning function in this iteration is used to calculate the probability value of each learning function in the next iteration, guiding the selection of the learning function for the next iteration.

[0096] For example, in the t-th iteration (this iteration), based on the optimal samples corresponding to the three different learning functions, the immediate reward value r(t) for the t-th iteration of each function is calculated. Then, for each learning function, based on the immediate reward value r(t) of the t-th iteration and the cumulative reward value G(t-1) of the (t-1)-th iteration, the cumulative reward value G(t) of the t-th iteration of that learning function is determined. In the next iteration, the probability value of that learning function is calculated based on G(t-1).

[0097] It should be noted that before iteration, the initial cumulative reward value for each learning function can be set, for example, it can be 0: ,in, Let represent the initial cumulative reward value of the i-th learning function. This indicates the number of learning functions. The cumulative reward value for each learning function is updated based on the initial cumulative reward value through steps 401 to 402.

[0098] Therefore, in this embodiment, the immediate reward of each learning function is determined at each iteration, and the cumulative reward value of each learning function is updated accordingly to guide the selection of the next learning function. By calculating the probability value of the next iteration using the reward value that characterizes the performance of the learning function, and selecting the optimal learning function based on the probability value, the reliability and accuracy of the selection can be guaranteed.

[0099] In an exemplary embodiment, step 401 includes: inputting the optimal sample corresponding to each learning function into the current surrogate model to obtain the predicted response value corresponding to each optimal sample; and determining the instantaneous reward value of the corresponding learning function for this iteration based on the negative number of the absolute value of the predicted response value corresponding to the optimal sample.

[0100] The smaller the absolute value, the larger the corresponding instant reward value; the larger the absolute value, the smaller the corresponding instant reward value.

[0101] For example, in each iteration, for each learning function, its corresponding optimal sample is input into the current surrogate model (e.g., the Kriging surrogate model) to obtain the predicted response value (e.g., the mean) corresponding to the optimal sample. Then, each predicted response value is input into a pre-set reward function, which represents the mapping relationship between the predicted response value and the immediate reward value of this iteration. The calculated reward function value is the immediate reward value of the learning function for this iteration. Alternatively, the immediate reward value of the corresponding learning function for this iteration can be determined based on the negative of the absolute value of the predicted response value corresponding to the optimal sample. Optionally, the negative of the absolute value of the predicted response value corresponding to the optimal sample can be used as the immediate reward value of the learning function for this iteration.

[0102] The absolute value of the predicted response directly reflects how close the corresponding optimal sample is to the predicted failure boundary; the smaller the absolute value, the closer the sample is to the actual failure boundary. Taking a negative absolute value results in a smaller absolute value (the closer the sample is) and a larger immediate reward value.

[0103] Optionally, when the surrogate model is the Kriging surrogate model, the predicted response value is the mean of the predicted response. Inputting this into the following reward function will yield the immediate reward value:

[0104] (3)

[0105] in, This represents the immediate reward value of the i-th learning function in the t-th iteration. This represents the optimal sample in the t-th iteration of the i-th learning function. This represents the mean predicted response (output by the current surrogate model) corresponding to the optimal sample in the t-th iteration of the i-th learning function. Taking six learning functions—U, EFF, ERF, H, RLCB, and REIF—as an example, the optimal samples are as follows:

[0106] (4)

[0107] Therefore, the instant reward value determined in this embodiment can quantify the potential value of the optimal sample of each learning function, reward the learning function that is closer to the real failure boundary, ensure the reliability of the instant reward value, provide effective support for the calculation of subsequent probability values, thereby driving the system to automatically tend to select the better strategy that is better at sampling in key areas (failure boundaries), and finally achieve efficient and accurate model training.

[0108] In an exemplary embodiment, step 402 above includes: determining the cumulative reward value of the learning function for the current iteration based on the immediate reward value of the current iteration, the cumulative reward value of the previous iteration, and the memory factor.

[0109] Among them, the memory factor is greater than 0 and less than 1, such as 0.7.

[0110] For example, for each learning function, the product of the memory factor and the cumulative reward value of the previous iteration is calculated, and the sum of this product and the immediate reward value of the current iteration is calculated as the cumulative reward value of the learning function for this iteration. The calculation formula is as follows:

[0111] (5)

[0112] in, This represents the cumulative reward value of the i-th learning function in the t-th iteration. This represents the cumulative reward value of the i-th learning function in the (t-1)-th iteration. This represents the immediate reward value of the i-th learning function in the t-th iteration. This represents the memory factor, which can be 0.7.

[0113] Therefore, this embodiment determines the cumulative reward value of the learning function for this iteration based on the immediate reward value of the current iteration, the cumulative reward value of the previous iteration, and the memory factor. This ensures the reliability and accuracy of the cumulative reward value, providing reliable support for the calculation of subsequent probability values.

[0114] In an exemplary embodiment, step 302 above includes: determining the probability value of each learning function in the current iteration based on the balance factor and the cumulative reward value of each learning function in the previous iteration. Figure 5 As shown, this may specifically include steps 501 and 502. Wherein:

[0115] Step 501: Normalize the cumulative reward value of the previous iteration of each learning function based on the maximum and minimum reward values ​​to obtain each normalized cumulative reward value.

[0116] The maximum reward value and the minimum reward value are the maximum and minimum values ​​of the cumulative reward value in the previous iteration of each learning function, respectively.

[0117] For example, the maximum and minimum reward values ​​are selected from the cumulative reward values ​​of all learning functions in the previous iteration, and the cumulative reward value of each learning function in the previous iteration is normalized using the following formula:

[0118] (6)

[0119] (7)

[0120] in, This represents the cumulative reward value after normalization by the i-th learning function. This represents the cumulative reward value of the i-th learning function in the (t-1)-th iteration. Indicates the maximum reward value. This represents the minimum reward value.

[0121] Step 502: For each learning function, determine the probability value of the learning function in this iteration based on the balance factor and the normalized cumulative reward value.

[0122] The balance factor is a positive number less than or equal to 10, such as 2.

[0123] For example, after obtaining the normalized cumulative reward value for each learning function, the first product of each normalized cumulative reward value and the balance factor is calculated. For each learning function, the second product of the balance factor and its normalized cumulative reward value is selected from the first product, and the probability value of the learning function for this iteration is determined based on the second product and the first product. Optionally, the calculation formula is:

[0124] (8)

[0125] in, Indicates the number of learning functions. This represents the balance factor, which can be 7. This represents the cumulative reward value after normalization of the j-th learning function. Indicates the second product. This represents the first product.

[0126] In this embodiment, after normalizing the cumulative reward value of the previous iteration, the probability value of the learning function for the current iteration is determined based on the balance factor and the normalized cumulative reward value. This enables adaptive dynamic updating of the probability value. By quantifying the historical performance (reward) of the learning function into the probability value of selection in the current iteration, and using the balance factor to regulate the intensity of exploration and utilization, the algorithm can automatically and dynamically identify the most effective sampling strategy for the specific reliability analysis problem, thereby ensuring high efficiency and strong robustness throughout the training process.

[0127] After obtaining the probability values, the optimal learning function for this iteration is randomly selected based on the probability, and the current surrogate model is retrained based on the optimal samples corresponding to the optimal learning function (step 304). The cumulative reward value for each learning function in this iteration is then determined (steps 401 and 402). It can then be determined whether the entire active learning process has converged. If yes, the iteration is terminated; otherwise, step 301 is returned for the next iteration. The convergence condition may include that the maximum possible relative error of the failure probability of the current surrogate model is less than or equal to a preset value. The following describes the convergence determination process for the combined active learning strategy.

[0128] In one possible implementation, the output response value of the Kriging model is known. Follows a standard normal distribution Therefore, at the unknown point x, its response value The probability that the symbol is mispredicted is:

[0129] (9)

[0130] That is, if ( The sign of the response value at the unknown point x is incorrectly predicted. The sample is divided into two parts, one of which satisfies , The other part satisfies , ,and Based on the current Kriging surrogate model and using the Monte Carlo method, the failure probability can be calculated as follows:

[0131] (10)

[0132] in, and These represent the number of failed samples in the first and second groups of samples, respectively.

[0133] Because the second group of samples satisfies This indicates that the second group of samples has an uncertainty regarding the incorrect prediction of the output response sign. First, we assume that the actual number of failed samples is... Therefore, the true value of the failure probability is:

[0134] (11)

[0135] Therefore, the relative error between the actual value and the calculated value of the failure probability is expressed as:

[0136] (12)

[0137] In the formula, This represents the number of truly invalid samples in the second group of samples, i.e. Therefore, relative error The maximum value is expressed as:

[0138] (13)

[0139] In the embodiments of this application, as long as This indicates that the active learning process of the combined configuration has converged iteratively. Therefore, when the maximum possible relative error of the failure probability of the current surrogate model is less than or equal to a preset value, such as 0.02, the active learning is considered to have converged, ensuring the reliability of iterative training.

[0140] like Figure 6 and Figure 7 As shown, the combined active learning strategy in this embodiment is as follows: taking multiple different learning functions as U, EFF, ERF, H, RLCB, and REIF functions as examples, in each iteration, the optimal sample corresponding to each function is selected from the candidate sample pool, respectively. , , , , and The combined configuration strategy selects a second time based on the probability value. The best sample After selecting the optimal sample, the cumulative reward value of each learning function in this iteration is updated based on the optimal sample, and used to calculate the probability value of the next iteration.

[0141] In practical engineering problems, due to insufficient experience in judging the problem itself, it is often difficult to select a suitable active learning function. The embodiments of this application ingeniously adopt an active learning solution strategy based on combined configuration in structural reliability analysis problems, which directly eliminates the process of manually selecting active learning functions. This can effectively avoid the interference of human factors on the analysis results, while ensuring that the overall strategy's analysis efficiency approaches the analysis efficiency of the optimal active learning strategy.

[0142] In an exemplary embodiment, step 203 includes: determining the number of failed samples based on the predicted response values ​​of the analyzed samples; and determining the failure probability of the device structure to be analyzed based on the number of failed samples and the total number of samples in the analyzed sample pool.

[0143] For example, the number of predicted response values ​​less than 0 is determined as the number of failure samples, and the ratio between the number of failure samples and the total number of samples is determined as the failure probability of the equipment structure to be analyzed.

[0144] Optionally, the failure probability of the structure can be calculated using the Monte Carlo method, as follows:

[0145] (14)

[0146] (15)

[0147] In the formula, and These are the estimated values ​​of the failure probability and the coefficient of variation, respectively. It is the number of samples in the Monte Carlo sampling. It represents the number of failed samples.

[0148] The following example illustrates the training process of each iteration of this application.

[0149] In a specific example, such as Figure 8 As shown, each iteration of the training process includes the following steps:

[0150] Step 801: Initialize parameters and establish the limit state function; the parameters include the balance factor, memory factor, initial cumulative reward value, initial probability value, and limit state function;

[0151] Step 802: Generate the initial sample pool and call the limit state function to calculate the true response value;

[0152] Step 803: Construct the Kriging proxy model based on the initial sample pool;

[0153] Step 804: Generate a candidate sample pool;

[0154] Step 805: Determine the optimal sample corresponding to multiple different learning functions from the candidate sample pool;

[0155] For detailed implementation instructions, please refer to step 301. Several different learning functions are U, EFF, ERF, H, RLCB, and REIF functions.

[0156] Step 806: Determine the instantaneous reward value of the learning function for this iteration based on the optimal sample of the learning function; see step 401 for details.

[0157] Step 807: Determine the cumulative reward value of the learning function for this iteration based on the immediate reward value of this iteration, the cumulative reward value of the previous iteration, and the memory factor; see the above description for specific implementation details.

[0158] Step 808: Determine the probability value of each learning function in this iteration based on the balance factor and the cumulative reward value of each learning function in the previous iteration; see steps 501 and 502 for specific implementation details.

[0159] Step 809: Based on the probability value of this iteration, randomly select the optimal learning function for this iteration according to probability.

[0160] Step 810: Retrain the current Kriging proxy model based on the optimal sample corresponding to the optimal learning function; see step 304 for details.

[0161] Step 811: Determine whether the maximum relative error of the current Kriging proxy model is less than 0.02. If yes, proceed to step 812; otherwise, return to step 805.

[0162] Step 812: Use the current Kriging proxy model as the pre-built Kriging proxy model.

[0163] To verify the combined configuration solution strategy of this application embodiment, a 6-dimensional application example is used. In this example, the structure of the device to be analyzed is a roof truss structure, specifically as follows: Figure 9 As shown. The roof's top and compression members are reinforced with concrete, while the bottom and tie members are made of steel. A uniformly distributed load q is applied to the roof, transforming the uniform load into a concentrated load at the nodes. The random variables are uniform load q, rod length l, and cross-sectional area of ​​the cement rod. cross-sectional area of ​​steel pole Elastic modulus of cement pole , elastic modulus of steel rod The limit state function in this example is that the maximum vertical position of the roof truss structure does not exceed 0.03m, that is:

[0164]

[0165] The distribution types and distribution parameters of all input random variables for this structure are shown in Table 1.

[0166] Table 1. Input Random Variable Distribution Types and Parameters

[0167]

[0168] In this calculation example, the U function, EFF function, ERF function, H function, RLCB function, and REIF function were used. The solution results of the combined configuration solution strategy of this application embodiment were compared with the solution results of the above different functions. All calculation results are shown in Table 2. Taking the calculation results obtained by Monte Carlo Simulation (MCS) as a reference, it can be seen that the calculation accuracy of different calculation methods all meet the requirements of engineering applications. However, their call counts (the number of times the real simulation model is called) are quite different. The more calls, the lower the calculation efficiency of the method. It can be seen that the calculation efficiency of the combined configuration solution strategy of this application embodiment is higher than that of most methods. In practical engineering applications, when it is difficult to choose an active learning strategy due to insufficient engineering information, the method proposed in this application embodiment can effectively avoid selecting an overlapping learning strategy, while its calculation efficiency tends to be closer to the better learning strategy.

[0169] Table 2 Comparison of calculation results from different methods

[0170]

[0171] In summary, the structural reliability analysis method of this application combines a combined configuration strategy and an active learning method. This avoids the problems caused by manually selecting active learning strategies in engineering applications, avoids subjective biases from manually selecting active functions, and ensures both computational efficiency and accuracy in structural reliability analysis. It effectively improves the efficiency and accuracy of structural reliability analysis and better meets practical engineering needs. It has advantages such as simple operation, accurate calculation, and high analysis efficiency, and has high practical and promotional value in the field of structural reliability analysis and evaluation technology.

[0172] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages in other steps. It is understood that the steps in different embodiments can be freely combined as needed, and all non-contradictory solutions formed by such combinations are within the scope of protection of this application.

[0173] Based on the same inventive concept, this application also provides a structural reliability analysis apparatus for implementing the structural reliability analysis method described above. The solution provided by this apparatus is similar to the implementation scheme described in the above method; therefore, the specific limitations in one or more embodiments of the structural reliability analysis apparatus provided below can be found in the limitations of the structural reliability analysis method described above, and will not be repeated here.

[0174] In one exemplary embodiment, such as Figure 10 As shown, a structural reliability analysis device is provided, comprising: a first determining module 1001, an input module 1002, and a second determining module 1003, wherein: the first determining module 1001 is used to determine an analysis sample pool of the structure of the device to be analyzed; the analysis sample pool includes analysis samples of the structure of the device to be analyzed, and the analysis samples include parameters affecting the reliability of the structure of the device to be analyzed; the input module 1002 is used to input each analysis sample in the analysis sample pool into a pre-trained surrogate model to obtain the predicted response value of the analysis sample; the second determining module 1003 is used to determine the structure of the device to be analyzed based on the predicted response value of the analysis sample. The device includes: a failure probability of the backup structure; wherein the pre-trained proxy model is obtained by iteratively training an initial proxy model, and during each iteration, the device further includes: a third determining module, used to determine the optimal sample corresponding to multiple different learning functions from the candidate sample pool; a fourth determining module, used to determine the probability value of each learning function in the current iteration based on the cumulative reward value of each learning function in the previous iteration; a fifth determining module, used to determine the optimal learning function from each learning function based on the probability value of the learning function in the current iteration; and a retraining module, used to retrain the current proxy model based on the optimal sample corresponding to the optimal learning function.

[0175] In one embodiment, during each iteration of training, the apparatus further includes: a sixth determining module, configured to determine, for each learning function, the immediate reward value of the learning function in this iteration based on the optimal sample of the learning function; and a seventh determining module, configured to determine the cumulative reward value of the learning function in this iteration based on the immediate reward value of the current iteration and the cumulative reward value of the previous iteration; the cumulative reward value of the current iteration is used to calculate the probability value of the next iteration.

[0176] In one embodiment, the sixth determining module is specifically used to: input the optimal sample corresponding to each learning function into the current surrogate model to obtain the predicted response value corresponding to each optimal sample; and determine the instantaneous reward value of the corresponding learning function in this iteration based on the negative number of the absolute value of the predicted response value corresponding to the optimal sample.

[0177] In one embodiment, the seventh determining module is specifically used to: determine the cumulative reward value of the learning function in the current iteration based on the immediate reward value of the current iteration, the cumulative reward value of the previous iteration, and the memory factor; wherein the memory factor is greater than 0 and less than 1.

[0178] In one embodiment, the fourth determining module is specifically used to: normalize the cumulative reward value of each learning function in the previous iteration based on the maximum reward value and the minimum reward value to obtain each normalized cumulative reward value; wherein the maximum reward value and the minimum reward value are the maximum and minimum values ​​of the cumulative reward value in the previous iteration of each learning function, respectively; for each learning function, determine the probability value of the learning function in the current iteration based on the balance factor and the normalized cumulative reward value; wherein the balance factor is a positive number less than or equal to 10.

[0179] In one embodiment, the second determining module 1003 is specifically used to: determine the number of failed samples based on the predicted response values ​​of the analyzed samples; and determine the failure probability of the device structure to be analyzed based on the number of failed samples and the total number of samples in the analyzed sample pool.

[0180] Each module in the aforementioned structural reliability analysis device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0181] In one exemplary embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 11 As shown, this computer device includes a processor, memory, input / output interfaces (I / O), and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operating system and computer programs stored in the non-volatile storage media. The database stores structural reliability analysis data. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a structural reliability analysis method.

[0182] Those skilled in the art will understand that Figure 11 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0183] In one exemplary embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the structural reliability analysis method of the above embodiment.

[0184] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a processor, implements the structural reliability analysis method of the above embodiment.

[0185] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the structural reliability analysis method of the above embodiments.

[0186] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. The databases involved in the embodiments provided in this application can include at least one of relational databases and non-relational databases. Non-relational databases can include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application can be general-purpose processors, central processing units, graphics processors, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0187] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0188] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A structural reliability analysis method, characterized in that, The method includes: A sample pool for analyzing the structure of the device to be analyzed is determined; the sample pool includes analytical samples of the structure of the device to be analyzed, and the analytical samples include parameters that affect the reliability of the structure of the device to be analyzed. Each analysis sample in the analysis sample pool is input into a pre-trained surrogate model to obtain the predicted response value of the analysis sample. The failure probability of the device structure to be analyzed is determined based on the predicted response value of the analyzed sample. The pre-trained agent model is obtained by iteratively training an initial agent model, and each iteration includes: Determine the optimal sample corresponding to multiple different learning functions from the candidate sample pool; Based on the cumulative reward value of the previous iteration of each learning function, the probability value of the current iteration of each learning function is determined; The optimal learning function is determined from among all learning functions based on the probability value of the current iteration of the learning function. The current agent model is retrained based on the optimal sample corresponding to the optimal learning function.

2. The method according to claim 1, characterized in that, Each iteration of training also includes: For each learning function, the instantaneous reward value of the learning function in this iteration is determined based on the optimal sample of the learning function; Based on the immediate reward value of the current iteration and the cumulative reward value of the previous iteration, the cumulative reward value of the learning function for the current iteration is determined; the cumulative reward value of the current iteration is used to calculate the probability value of the next iteration.

3. The method according to claim 2, characterized in that, The step of determining the instantaneous reward value for each learning function in this iteration based on the optimal sample of each learning function includes: Input the optimal sample corresponding to each learning function into the current agent model to obtain the predicted response value corresponding to each optimal sample; The instantaneous reward value for this iteration of the corresponding learning function is determined based on the negative of the absolute value of the predicted response value corresponding to the optimal sample.

4. The method according to claim 2, characterized in that, Determining the cumulative reward value of the learning function for the current iteration based on the immediate reward value of the current iteration and the cumulative reward value of the previous iteration includes: The cumulative reward value of the learning function in this iteration is determined based on the immediate reward value of the current iteration, the cumulative reward value of the previous iteration, and the memory factor; wherein the memory factor is greater than 0 and less than 1.

5. The method according to claim 1, characterized in that, Determining the probability value of each learning function in the current iteration based on the cumulative reward value of the previous iteration of each learning function includes: The cumulative reward value of each learning function in the previous iteration is normalized based on the maximum and minimum reward values ​​to obtain each normalized cumulative reward value; where the maximum and minimum reward values ​​are the maximum and minimum values ​​of the cumulative reward value in the previous iteration of each learning function, respectively. For each learning function, the probability value of the learning function in this iteration is determined based on the balance factor and the normalized cumulative reward value; wherein the balance factor is a positive number less than or equal to 10.

6. The method according to any one of claims 1-5, characterized in that, Determining the failure probability of the structure to be analyzed based on the predicted response value of the analyzed sample includes: The number of failed samples is determined based on the predicted response values ​​of the analyzed samples; The failure probability of the device structure to be analyzed is determined based on the number of failed samples and the total number of samples in the analysis sample pool.

7. A structural reliability analysis device, characterized in that, The device includes: The first determining module is used to determine the analysis sample pool of the structure of the device to be analyzed; the analysis sample pool includes analysis samples of the structure of the device to be analyzed, and the analysis samples include parameters that affect the reliability of the structure of the device to be analyzed; The input module is used to input each analysis sample in the analysis sample pool into the pre-trained surrogate model to obtain the predicted response value of the analysis sample. The second determining module is used to determine the failure probability of the structure of the device to be analyzed based on the predicted response value of the analysis sample; The pre-trained agent model is obtained by iteratively training an initial agent model, and during each iteration of training, the device further includes: The third determination module is used to determine the optimal sample corresponding to multiple different learning functions from the candidate sample pool; The fourth determining module is used to determine the probability value of each learning function in the current iteration based on the cumulative reward value of each learning function in the previous iteration; The fifth determining module is used to determine the optimal learning function from among the learning functions based on the probability value of the learning function in this iteration; The retraining module is used to retrain the current agent model based on the optimal samples corresponding to the optimal learning function.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images