Method, device, medium and product for evaluating cross-condition reliability of high-temperature components for peak regulation and uncertainty propagation

Abstract

The peak-shaving high-temperature component cross-condition reliability evaluation and uncertainty propagation method, device, medium and product disclosed by the application relate to the field of high-temperature structural integrity and optimization design, and the method comprises the following steps: taking the design condition of a high-temperature component before grid peak shaving as a source domain and taking the design condition after grid peak shaving as a target domain; constructing an agent model based on relevant data of the source domain; determining inherent uncertainty and cognitive uncertainty of the source domain; constructing a graph neural network; performing migration training on the agent model based on the graph neural network, relevant data of the target domain and a loss function, obtaining an agent model of the target domain; determining a life distribution of the target domain based on the agent model of the target domain and relevant data of the target domain; determining a reliability value of the target domain based on the life distribution; and determining uncertainty of the target domain based on the graph neural network to realize uncertainty propagation. The application can predict the life distribution of the target domain and improve the credibility of the prediction.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of high-temperature structural integrity and optimal design, in particular to a cross-condition reliability evaluation and uncertainty propagation method for peak-shaving oriented high-temperature components, equipment, medium and product. BACKGROUND

[0002] With the demand for power grid peak shaving of high-temperature equipment such as gas turbines and steam turbines, accurate design life reliability evaluation is an important means to ensure the safe operation thereof. However, peak shaving brings complex new working conditions, and at the same time causes the need for optimization design of geometry and materials. Since it is difficult to obtain high-reliability component destruction test data, the traditional reliability evaluation often relies on high-fidelity finite element calculation, further replaces the time-consuming finite element calculation by means of a neural network or the like proxy model, and finally obtains the probability distribution of the life based on the Monte Carlo simulation of the proxy model to calculate the reliability. Compared with directly sampling based on the finite element calculation, this method has greatly improved the calculation efficiency, but the construction of the proxy model still needs hundreds of finite element calculations as a data set. When facing new design conditions, such as a series of components of the same type but with slightly different geometry sizes, adjustment of the component design working condition or replacement of the material during the design process, a single proxy model cannot adapt to the new design conditions, and the cost of retraining the proxy model under the new conditions is too high, which also causes waste of the original data.

[0003] The transfer learning method can make full use of the existing data and combine a small amount of new condition data to construct a proxy model under the new condition, and can play a great role in accelerating the calculation for the cross-condition reliability evaluation brought by peak shaving. At present, the transfer learning method is less applied in the field of reliability evaluation of high-temperature components, and the related method does not have the ability of uncertainty propagation. Specifically, the related technology has the following limitations: the inherent uncertainty and the cognitive uncertainty cannot be explicitly distinguished and propagated, resulting in the uncertainty of the reliability evaluation result being unclear; the transfer process is mostly a black box operation, lacking of mechanism guidance on "how to optimally transfer", and the transfer hyperparameters (such as the number of frozen layers, learning rate, etc.) depend on the manual setting of expert experience; in the case of scarce data in the target domain, the prediction uncertainty distribution of the proxy model in the input space is uneven, and the traditional experimental design method cannot efficiently guide the allocation of simulation calculation resources to reduce the global uncertainty.

[0004] Therefore, there is an urgent need in the art for a cross-condition reliability evaluation method that can systematically quantify and propagate uncertainty, adaptively optimize the transfer process, and intelligently guide data sampling, in order to cope with the challenges of high-temperature component design and safety evaluation under peak shaving conditions. SUMMARY

[0005] The application aims to provide a high-temperature component cross-condition reliability evaluation and uncertainty propagation method, device, medium and product for peak regulation, which can predict the life distribution of the target domain and improve the credibility of the prediction.

[0006] To achieve the above-mentioned purpose, the application provides the following solutions.

[0007] In a first aspect, the application provides a high-temperature component cross-condition reliability evaluation and uncertainty propagation method for peak regulation, comprising:

[0008] Obtaining relevant data of the design conditions of the high-temperature component before and after the grid peak regulation; taking the design conditions of the high-temperature component before the grid peak regulation as a source domain and the design conditions after the grid peak regulation as a target domain; the relevant data of the source domain includes input condition data and output data considering uncertainty; the relevant data of the target domain includes input condition data; the high-temperature component is a device component working in a set temperature range;

[0009] Building an agent model based on the relevant data of the source domain;

[0010] Determining the inherent uncertainty and cognitive uncertainty of the source domain;

[0011] Taking the relevant data of the source domain, the inherent uncertainty and the cognitive uncertainty as node attributes, and combining the relevant data of the target domain to build a graph neural network;

[0012] Migrating and training the agent model based on the graph neural network, the relevant data of the target domain and a loss function to obtain a target domain agent model;

[0013] Determining whether the uncertainty predicted by the target domain agent model is less than a set threshold;

[0014] If the uncertainty predicted by the target domain agent model is greater than or equal to the set threshold, performing active learning based on the uncertainty measure and coverage measure of the target domain to obtain a point set; using the point set to perform adaptive point addition in the relevant data of the target domain, initializing the target domain agent model and training the initialized target domain agent model until the uncertainty predicted by the target domain agent model is less than the set threshold;

[0015] If the uncertainty predicted by the target domain agent model is less than the set threshold, using the target domain agent model to determine the life distribution of the target domain based on the relevant data of the target domain; determining the reliability value of the target domain based on the life distribution of the target domain; determining the uncertainty of the target domain based on the graph neural network to realize uncertainty propagation.

[0016] In an embodiment, building an agent model based on the relevant data of the source domain comprises:

[0017] The proxy model is constructed based on the input condition data and the output data.

[0018] In one embodiment, determining the inherent uncertainty and cognitive uncertainty of the source domain includes:

[0019] The inherent uncertainty is determined based on the input condition data; the inherent uncertainty includes material properties, load conditions, and geometric dimensions.

[0020] The cognitive uncertainty is determined based on the input conditional data using a model committee and the Monte Carlo Dropout algorithm; the model committee consists of multiple neural network models with different structures.

[0021] In one embodiment, the proxy model is trained based on the graph neural network, relevant data of the target domain, and a loss function to obtain a proxy model for the target domain, including:

[0022] An optimization algorithm is used to select multiple layers in the surrogate model for freezing, and a new set layer of neural network is added to obtain an initial surrogate model; the optimization algorithm includes gradient-based optimization algorithm, heuristic modeling method and probabilistic modeling method;

[0023] Based on the relevant data of the target domain and the output of the graph neural network, the initial proxy model is transferred and trained using a loss function to obtain the proxy model of the target domain.

[0024] In one embodiment, the loss function includes a Bayesian negative log-likelihood term and a structural consistency term; the loss function is expressed as:

[0025] ;

[0026] In the formula, This represents the value of the loss function. This represents the value of the Bayesian negative log-likelihood term. This represents the value of the structural consistency term. The weighting coefficients for the structural consistency term are used to balance the importance of uncertainty modeling and structural preservation. This represents the variance of the prediction results from the initial surrogate model. The first data in the target domain represents the relevant data. The true value of each data point. This represents the predicted value output by the initial surrogate model during transfer training. The total amount of data in the relevant data of the target domain; This indicates that during the initial transfer training of the proxy model from the source domain to the target domain, the relevant data in the target domain is the [missing information]. The data and the first Feature similarity of data; This represents the first relevant data in the target domain of a graph neural network. The data and the first Feature similarity of data.

[0027] In one embodiment, the coverage metric is expressed as:

[0028] ;

[0029] In the formula, Indicates coverage metric, Indicates candidate samples The feature description vector, Indicates the sampled points The feature description vector, This indicates the number of candidate sampled points.

[0030] In one embodiment, a graph neural network is constructed by using relevant data from the source domain, the inherent uncertainty, and the cognitive uncertainty as node attributes, combined with relevant data from the target domain, including:

[0031] Using the relevant data from the source domain, the inherent uncertainty, and the cognitive uncertainty as node attributes, an initial graph neural network is constructed.

[0032] The adjacency matrix of the initial graph neural network is determined based on the node attributes;

[0033] The node attributes and adjacency matrix of the initial graph neural network are updated using an attention mechanism based on a set number of times and relevant data of the target domain, thus obtaining the graph neural network.

[0034] In a second aspect, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the method for cross-condition reliability assessment and uncertainty propagation of high-temperature components oriented to peak shaving as described above.

[0035] Thirdly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method for cross-condition reliability assessment and uncertainty propagation of high-temperature components for peak shaving as described above.

[0036] Fourthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the method for cross-condition reliability assessment and uncertainty propagation of high-temperature components oriented towards peak shaving, as described above.

[0037] According to the specific embodiments provided in the application, the application has the following technical effects:

[0038] The application provides a peak-regulation-oriented high-temperature component cross-condition reliability evaluation and uncertainty propagation method, device, medium and product. An agent model is constructed according to related data of a source domain. A graph neural network is constructed according to the related data of the source domain and inherent uncertainty and cognitive uncertainty of the source domain, in combination with related data of a target domain, so that the uncertainty of the target domain is uniformly distributed. The system quantifies and propagates the uncertainty of the target domain, and then efficiently guides the allocation of computing resources in subsequent agent model construction of the target domain to reduce global uncertainty. The constructed agent model is migrated and trained according to the graph neural network, the related data of the target domain and a loss function, the migration process is adaptively optimized, and then the agent model of the target domain is obtained, and it is determined whether the uncertainty predicted by the agent model of the target domain is less than a set threshold. If not, active learning is performed based on uncertainty measurement and coverage measurement, intelligent data sampling is guided, and a point set is obtained. The point set is adaptively added to the related data of the target domain, which can intelligently guide data sampling, initialize the agent model of the target domain and train the initialized agent model of the target domain, until the uncertainty predicted by the agent model of the target domain is less than the set threshold, and the related data of the source domain can be efficiently and accurately used to assist in constructing the agent model of the target domain. The agent model of the target domain is used to determine the life distribution of the target domain based on the related data of the target domain, and the reliability value of the target domain is determined according to the life distribution of the target domain, so that the prediction result is accompanied by a reliability index, and the credibility of the prediction is improved. The uncertainty of the target domain is determined based on the graph neural network to realize the uncertainty propagation of the target domain. BRIEF DESCRIPTION OF DRAWINGS

[0039] In order to more clearly illustrate the technical solutions in the embodiments of the application or the prior art, the following will briefly introduce the drawings needed in the embodiments. Obviously, the drawings in the following description are only some embodiments of the application, and other drawings can be obtained by those skilled in the art without creative labor.

[0040] Figure 1 A flowchart of a peak-regulation-oriented high-temperature component cross-condition reliability evaluation and uncertainty propagation method in an embodiment of the application;

[0041] Figure 2 A sample number independence verification schematic diagram provided by an embodiment of the application;

[0042] Figure 3 A precision comparison schematic diagram of the migration result after active learning and point iteration provided by an embodiment of the application;

[0043] Figure 4 A rotor structure stress cloud map provided for another embodiment of the application;

[0044] Figure 5 A source domain reliability value schematic diagram provided for another embodiment of the application;

[0045] Figure 6 A structural schematic diagram of a computer device provided for an embodiment of the application. DETAILED DESCRIPTION

[0046] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by a person of ordinary skill in the art without creative work are within the protection scope of the application.

[0047] In view of the deficiencies in the foregoing background art, the application provides a high-temperature component cross-condition reliability evaluation and uncertainty propagation method for peak regulation, which can efficiently and accurately use source domain (design conditions with sufficient data) knowledge to assist in building a proxy model of the target domain (new design conditions), and clearly quantifies various uncertainties in the evaluation process.

[0048] The above purposes, features and advantages of the application will be more apparent and easy to understand. The application will be further described in detail below with reference to the drawings and specific embodiments.

[0049] In the process of grid peak regulation, the steam turbine or gas turbine of the power plant is frequently started and stopped. For the high-temperature components (such as turbine blades) therein, frequent fluctuations in the load spectrum will cause greater damage (i.e., shorter life). The gas turbine does not consider the grid peak regulation during use at the beginning of design, and the designed load spectrum is quite different from the load spectrum during use. Therefore, the load working condition after peak regulation is one of the design conditions. Secondly, to cope with the high damage caused by grid peak regulation, the gas turbine manufacturer needs to optimize the structure of components such as turbine blades and select better materials, so the newly produced gas turbine and the active gas turbine are different in structure and material. If the data of the old gas turbine before grid peak regulation is to be applied to the new gas turbine, the technical solution provided by the application needs to be applied (according to the mapping relationship from the design condition before grid peak regulation to the life, combined with a small amount of data of the design condition after grid peak regulation, the mapping relationship from the design condition after grid peak regulation to the life is calculated, that is, transfer learning). Based on this, as shown in the figure, the high-temperature component cross-condition reliability evaluation and uncertainty propagation method for peak regulation provided by the application includes: Figure 1

[0050] ​In step 100, related data of design conditions of the high-temperature component before and after the grid peak shaving is obtained. The design condition of the high-temperature component before the grid peak shaving is taken as a source domain, and the design condition after the grid peak shaving is taken as a target domain. The related data of the source domain includes input condition data and output data considering uncertainty. The related data of the target domain includes input condition data. The high-temperature component is a device component working in a set temperature range.

[0051] In step 200, an agent model is constructed based on the related data of the source domain. This includes constructing an agent model based on the input condition data and the output data.

[0052] In step 300, the inherent uncertainty and the cognitive uncertainty of the source domain are determined.

[0053] In step 400, the related data, the inherent uncertainty and the cognitive uncertainty of the source domain are taken as node attributes, and the graph neural network is constructed in combination with the related data of the target domain.

[0054] In step 500, the agent model is migrated and trained based on the graph neural network, the related data of the target domain and a loss function, to obtain an agent model of the target domain.

[0055] In step 600, it is determined whether the uncertainty predicted by the agent model of the target domain is less than a set threshold.

[0056] If the uncertainty predicted by the agent model of the target domain is greater than or equal to the set threshold, active learning is performed based on an uncertainty measure and a coverage measure of the target domain, to obtain a point set. The point set is used for adaptive point adding in the related data of the target domain, the agent model of the target domain is initialized and the initialized agent model of the target domain is trained, until the uncertainty predicted by the agent model of the target domain is less than the set threshold.

[0057] If the uncertainty predicted by the agent model of the target domain is less than the set threshold, the agent model of the target domain is used to determine a life distribution of the target domain based on the related data of the target domain. The reliability value of the target domain is determined based on the life distribution of the target domain. The uncertainty of the target domain is determined based on the graph neural network, to realize uncertainty propagation.

[0058] The coverage measure is represented as: In the formula, the coverage measure is represented as: The feature description vector of the candidate sample The feature description vector of the sampled point The feature description vector of the sampled point ​​The candidate sample point number represents all data in the relevant data of the target domain; the sampled point represents the data point that has been sampled in the relevant data of the target domain, and the number of all the sampled points is the candidate sample point number.

[0059] As an optional implementation, the implementation process of step 300 can be described as follows: determining the inherent uncertainty according to the input condition data. The inherent uncertainty includes material properties, load conditions and geometric dimensions. The cognitive uncertainty is determined based on the input condition data through the model committee and the Monte-Carlo Dropout algorithm. The model committee is composed of multiple neural network models with different structures.

[0060] For example, the relevant data of the source domain includes the input condition data (i.e. the probability distribution data of the design condition of the source domain) considering the uncertainty and the output data (i.e. the life obtained through the experiment or the finite element simulation according to the input condition data). The inherent uncertainty of the source domain is determined according to the probability distribution of the input condition data. And the cognitive uncertainty caused by human knowledge is calculated according to the input condition data, and the difference between models and the randomness within the model are considered at the same time through the model committee and the Monte-Carlo Dropout algorithm, which can be expressed as:

[0061] .

[0062] In the formula, represents the output result of the i-th neural network model in the model committee, represents the number of neural network models in the model committee, represents the cognitive uncertainty, represents the variance of the output result of the i-th neural network model, represents the expectation of the i-th neural network model calculated by the Monte-Carlo Dropout algorithm, represents the variance of the i-th neural network model calculated by the Monte-Carlo Dropout algorithm, represents the expectation of the M neural network model output results. represents the expectation of the M neural network model output results.

[0063] ​​​The model committee consists of multiple independently trained neural network models (typically with different initializations, training data subsets, or model structures). The training dataset is constructed based on relevant data from the source domain (including input conditional data and output data). Inputting the input conditional data from the source domain into the neural network models within the model committee yields the output results (i.e., the lifetime predicted by the neural network models). The distribution of the outputs from all the neural network models in the model committee determines the cognitive uncertainty of the source domain. The Monte Carlo Dropout algorithm simulates Bayesian neural networks by enabling dropout sampling multiple times when making predictions based on the input conditional data, thus generating an uncertainty estimate.

[0064] The implementation process of step 400 includes: constructing an initial graph neural network by using relevant data from the source domain, inherent uncertainty, and cognitive uncertainty as node attributes; determining the adjacency matrix of the initial graph neural network based on the node attributes; and updating the node attributes and adjacency matrix of the initial graph neural network using an attention mechanism based on a set number of iterations and relevant data from the target domain to obtain the graph neural network.

[0065] The implementation process of step 500 includes: selecting multiple layers in the surrogate model for freezing using an optimization algorithm, and adding several new neural network layers to obtain an initial surrogate model. The optimization algorithms include gradient-based optimization algorithms, heuristic modeling methods, and probabilistic modeling methods. Based on relevant data from the target domain and the output of the graph neural network, a loss function is used to perform transfer training on the initial surrogate model to obtain a surrogate model for the target domain.

[0066] The loss function includes a Bayesian negative log-likelihood term and a structural consistency term. The loss function is expressed as:

[0067] .

[0068] In the formula, This represents the value of the loss function. This represents the value of the Bayesian negative log-likelihood term. This represents the value of the structural consistency term. The weighting coefficients for the structural consistency term are used to balance the importance of uncertainty modeling and structural preservation. This represents the variance of the prediction results from the initial surrogate model. The first data in the target domain represents the relevant data. The true value of each data point. This represents the predicted value output by the initial surrogate model during transfer training. The total amount of data in the relevant data of the target domain; This indicates that during the initial transfer training of the proxy model from the source domain to the target domain, the relevant data in the target domain is the [missing information]. a feature similarity between the first data and the second data in the related data of the target domain in the graph neural network. a feature similarity between the first data and the second data in the related data of the target domain in the graph neural network. a feature similarity between the first data and the second data in the related data of the target domain in the graph neural network. a feature similarity between the first data and the second data in the related data of the target domain in the graph neural network. a feature similarity between the first data and the second data in the related data of the target domain in the graph neural network.

[0069] Based on the above description, by implementing the above steps 100~600, the present application realizes the collaborative improvement of the precision, efficiency and reliability of the reliability evaluation of high-temperature components under cross-design conditions based on the systematic fusion of transfer learning, graph neural networks and active learning. It has the following advantages:

[0070] 1) Systematic: using graph neural networks to explicitly and interpretively propagate inherent and cognitive uncertainty, so that the evaluation results are accompanied by a reliability index.

[0071] 2) Automation and self-adaptation: the optimal transfer strategy is determined by an optimization algorithm, and the data sampling is intelligently guided by active learning, which maximally reduces the dependence on expert experience and expensive simulation data.

[0072] 3) High efficiency and high precision: in the case of extremely scarce target domain data, a high-precision proxy model of the target domain can be quickly constructed, significantly reducing the evaluation cost, and providing a powerful quantitative tool for the rapid design and safety decision of high-temperature components under peak shaving conditions.

[0073] In an exemplary embodiment, a numerical case is used to illustrate the feasibility and accuracy of the cross-condition reliability evaluation and uncertainty propagation method for high-temperature components under peak shaving. The numerical case can be used to compare the deviation between the predicted value and the true value. Based on this, the implementation process of the cross-condition reliability evaluation and uncertainty propagation method for high-temperature components under peak shaving in this embodiment includes:

[0074] S1, given four polynomials and the probability distribution of each variable, which can be represented as different design conditions, as the source domain and the target domain, for verifying that the adaptive transfer can be performed under the conditions of variable dimensionality, increased variables, and changed probability distribution. Adaptive transfer can obtain better prediction results with fewer points than ordinary transfer. The results obtained by transfer learning are close to the results of direct training. The design conditions are shown in formulas (1)~(4).

[0075] (1)

[0076] (2)

[0077] (3)

[0078] (4)

[0079] wherein, Output data (known) as source domain, , and Output data (unknown, i.e. data to be solved in this embodiment) as different target domains. The variable distribution of the input condition data of formulas (1)~(3) is , , , And The working condition expression is the same as , but the variable distribution of the input condition data becomes , , .

[0080] According to the input condition data and output data of the source domain, the proxy model is constructed, and the input and output points are taken from formulas (1)~(4) respectively to train the feedforward neural network to obtain the proxy model. In order to exclude the influence of the number of samples on the prediction ability, first verify the sample independence, as shown in Figure 2 , the model average absolute percentage error (MAPE) accuracy tends to converge after the number of training samples exceeds 700, so 1000 samples are selected as the number of subsequent research samples.

[0081] The ability of the proxy model to fit different polynomials is shown in Table 1. Using 1000 training samples directly in the source domain and target domain can obtain better results, but if only a small number of samples are used, the prediction effect is poor in the case of insufficient training set. The purpose of transfer learning is to use a small number of samples in the target domain combined with the source domain data to obtain better results.

[0082] Table 1 Ability of neural network to fit different polynomials

[0083]

[0084] S2, the inherent uncertainty has been embodied in the probability distribution of the input variable data, and the cognitive uncertainty of each point (i.e. input and output points) in formula (1) is calculated according to step 300 in the above embodiment.

[0085] S3, adopt steps 400 and 500 in the above embodiment, construct a graph neural network, and use a particle swarm optimization algorithm to perform adaptive transfer learning on the initial proxy model to obtain the proxy model of the target domain.

[0086] S4, determine whether the uncertainty of the proxy model prediction of the target domain is less than the set threshold. If the uncertainty of the proxy model prediction of the target domain is greater than or equal to the set threshold, find points with large uncertainty and far from the current points by active learning (i.e. active learning based on uncertainty measurement and coverage measurement of the target domain to obtain a point set), and perform adaptive point addition. After each point addition, the proxy model of the target domain is reinitialized. The migration results of formulas (1) to (4) are verified, as shown in FIG. 6 (part a of FIG. 6 is the migration result of formulas (1) to (2), part b of FIG. 6 is the migration result of formulas (2) to (3), and part c of FIG. 6 is the migration result of formulas (3) to (4)). 10 / 30 / 50 / 100 different numbers of samples are used as initial samples, and the final result accuracy is basically consistent. After supplementing to 150 samples, they have all converged, which illustrates the feasibility of active learning. Figure 3 Figure 3 Figure 3 Figure 3

[0087]

[0088] Table 2 Comparison of adaptive adjustment and direct point addition strategy results

[0089]

[0090] S5, verify the ability of the proxy model of the target domain obtained by formula (1) as the source domain final migration learning to predict the target domain data. As shown in Table 3, this method can adaptively migrate in the case of dimensionality increase, variable increase, and distribution change. The migration learning has a result close to that of direct large sample training. After introducing uncertainty enhancement in the loss function, the accuracy can be further improved.

[0091] Table 3 Migration learning prediction result ability (R2 / MAPE) table

[0092]

[0093] ​​​​​​​​​​​​​​​​​In another exemplary embodiment, the application of the rotor real structure to the peak-shaving high-temperature component cross-condition reliability evaluation and uncertainty propagation method is illustrated, and the stress cloud of the rotor three-dimensional model is as shown in Figure 4 The conventional working condition is creep fatigue loading.

[0094] According to steps 100-300, the design conditions of the source domain and the target domain are determined. The cognitive uncertainty is calculated by the neural network model combined with the model committee and the Monte-Carlo Dropout algorithm, and the inherent uncertainty related parameters of the finite element model are shown in Table 4.

[0095] Table 4: Rotor inherent uncertainty variable probability distribution table

[0096]

[0097] The new conditions of the target domain caused by peak shaving include variable load conditions (design condition 1-1: the rotating speed is increased from 1904 rad / s to 2284 rad / s, design condition 1-2: the load holding time is extended from 1440 h to 2880 h); material replacement (design condition 2: the material is replaced from CrMoV steel to P92 steel); geometry size optimization (the contour 2 size length is optimized from 76.74 mm to 78.0 mm).

[0098] After the sample number independence verification, the precision obtained by using 100 samples to train the neural network has basically converged. Based on the Monte Carlo simulation of the neural network, the reliability on the source domain can be obtained, as shown in Figure 5 The (a) part of Figure 5 is the error band of the life distribution predicted by the proxy model of the target domain and the true value of the life, and all points fall within ±1.2 times the error band; Figure 5 The (b) part of is the failure probability-design life curve, and the design life under 90% reliability is 687 cycles.

[0099] According to steps 400 and 500, the graph neural network and the particle swarm optimization algorithm adaptive transfer learning network are constructed.

[0100] According to step 600, adaptive active learning is performed, and finally the neural network suitable for the target domain is obtained. The comparison results of the transfer learning method and the direct large sample training method under different target domains are shown in Table 5. Among them, the reliability obtained by directly based on a large number of finite element calculations (105 times) is considered as the accurate reliability result, and the deviation of the transfer learning method from it is calculated. The effect of transfer learning is better than direct training, because the regularity under other working conditions is fully utilized.

[0101] Table 5 migration learning reliability calculation result table of different target domain conditions of rotor

[0102]

[0103] In an example embodiment, a computer device is provided, which can be a server or a terminal, and an internal structure diagram thereof can be as shown in Figure 6 The computer device includes a processor, a memory, an input / output interface (I / O) and a communication interface. The processor, the memory and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. The processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operating system and the computer program in the non-volatile storage medium to run. The database of the computer device is configured to store data related to the cross-condition reliability evaluation and uncertainty propagation method for peak-shaving high-temperature components. The input / output interface of the computer device is configured to exchange information between the processor and external devices. The communication interface of the computer device is configured to communicate with external terminals through network connection. The computer program is executed by the processor to implement a cross-condition reliability evaluation and uncertainty propagation method for peak-shaving high-temperature components.

[0104] Those skilled in the art can understand that Figure 6 the structure shown in the above description is only a block diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation on the computer device to which the scheme of the present application is applied. The specific computer device can include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components. In an example embodiment, a computer device is provided, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0105] In an example embodiment, a computer readable storage medium is provided, which stores a computer program. The computer program is executed by a processor to implement the steps in the above method embodiments.

[0106] In an example embodiment, a computer program product is provided, which includes a computer program. The computer program is executed by a processor to implement the steps in the above method embodiments.

[0107] It should be noted that the user information (including but not limited to user equipment information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in the present application are all information and data authorized by the user or authorized by all parties, and the collection, use and processing of related data need to comply with relevant regulations.

[0108] It can be understood by those skilled in the art that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing related hardware through a computer program, and the computer program can be stored in a non-volatile computer readable storage medium. When the computer program is executed, it can include the processes of the above-mentioned embodiments of each method. Among them, any reference to memory, database or other medium used in each embodiment provided by the present application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (Read-Only Memory, ROM), magnetic tape, floppy disk, flash memory, optical storage, high-density embedded non-volatile memory, resistive memory (ReRAM), magnetoresistive random access memory (Magnetoresistive Random Access Memory, MRAM), ferroelectric memory (Ferroelectric Random Access Memory, FRAM), phase change memory (Phase Change Memory, PCM), graphene memory, etc. Volatile memory can include random access memory (Random Access Memory, RAM) or external cache memory, etc. As an illustration but not limitation, RAM can be in various forms, such as static random access memory (Static Random Access Memory, SRAM) or dynamic random access memory (Dynamic Random Access Memory, DRAM), etc.

[0109] The database involved in each embodiment provided by the present application can include at least one of a relational database and a non-relational database. The non-relational database can include a distributed database based on a block chain, etc., without being limited thereto. The processor involved in each embodiment provided by the present application can be a general-purpose processor, a central processing unit, a graphics processing unit, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., without being limited thereto.

[0110] Each technical feature of the above embodiments can be combined arbitrarily. In order to make the description simple, not all possible combinations of each technical feature in the above embodiments are described, but as long as the combination of these technical features does not exist contradictory, it should be considered as the scope of the present application.

[0111] The principles and implementations of the present application are described in detail herein with specific examples. The above examples are only used to help understand the method of the present application and its core idea; meanwhile, for those skilled in the art, according to the idea of the present application, the specific implementation and application range will be changed. In summary, the content of the specification should not be understood as a limitation of the present application.

Claims

1. A method for cross-condition reliability assessment and uncertainty propagation of high-temperature components for peak shaving, characterized in that, include: Acquire relevant data on the design conditions of high-temperature components before and after power grid peak shaving; take the design conditions of high-temperature components before power grid peak shaving as the source domain and the design conditions after power grid peak shaving as the target domain; the relevant data of the source domain includes input condition data and output data considering uncertainties; the relevant data of the target domain includes input condition data. High-temperature components are equipment parts that operate within a set temperature range. A proxy model is constructed based on the relevant data from the source domain; Determine the inherent uncertainty and cognitive uncertainty of the source domain; The relevant data of the source domain, the inherent uncertainty, and the cognitive uncertainty are used as node attributes, and a graph neural network is constructed by combining the relevant data of the target domain. Based on the graph neural network, relevant data of the target domain, and loss function, the proxy model is transferred to train, and a proxy model of the target domain is obtained. Determine whether the uncertainty of the surrogate model prediction in the target domain is less than a set threshold; If the uncertainty predicted by the surrogate model in the target domain is greater than or equal to a set threshold, then active learning is performed based on the uncertainty metric and coverage metric of the target domain to obtain a point set; the point set is used to adaptively add points to the relevant data in the target domain, initialize the surrogate model of the target domain, and train the initialized surrogate model of the target domain until the uncertainty predicted by the surrogate model of the target domain is less than the set threshold. If the uncertainty predicted by the surrogate model of the target domain is less than a set threshold, the surrogate model of the target domain is used to determine the lifetime distribution of the target domain based on the relevant data of the target domain; the reliability value of the target domain is determined based on the lifetime distribution of the target domain; and the uncertainty of the target domain is determined based on the graph neural network to achieve uncertainty propagation.

2. The method for cross-condition reliability assessment and uncertainty propagation of high-temperature components for peak shaving according to claim 1, characterized in that, Constructing a proxy model based on relevant data from the source domain includes: The proxy model is constructed based on the input condition data and the output data.

3. The method for cross-condition reliability assessment and uncertainty propagation of high-temperature components for peak shaving according to claim 2, characterized in that, Determining the inherent uncertainty and cognitive uncertainty of the source domain includes: The inherent uncertainty is determined based on the input condition data; the inherent uncertainty includes material properties, load conditions, and geometric dimensions. The cognitive uncertainty is determined based on the input conditional data using a model committee and the Monte Carlo Dropout algorithm; the model committee consists of multiple neural network models with different structures.

4. The method for cross-condition reliability assessment and uncertainty propagation of high-temperature components for peak shaving according to claim 3, characterized in that, The proxy model is trained based on the graph neural network, relevant data of the target domain, and a loss function to obtain a proxy model for the target domain, including: An optimization algorithm is used to select multiple layers in the surrogate model for freezing, and a new set layer of neural network is added to obtain an initial surrogate model; the optimization algorithm includes gradient-based optimization algorithm, heuristic modeling method and probabilistic modeling method; Based on the relevant data of the target domain and the output of the graph neural network, the initial proxy model is transferred and trained using a loss function to obtain the proxy model of the target domain.

5. The method for cross-condition reliability assessment and uncertainty propagation of high-temperature components for peak shaving according to claim 4, characterized in that, The loss function includes a Bayesian negative log-likelihood term and a structural consistency term; the loss function is expressed as follows: ； In the formula, This represents the value of the loss function. This represents the value of the Bayesian negative log-likelihood term. This represents the value of the structural consistency term. The weighting coefficients for the structural consistency term are used to balance the importance of uncertainty modeling and structural preservation. This represents the variance of the prediction results from the initial surrogate model. The first data in the target domain represents the relevant data. The true value of each data point. This represents the predicted value output by the initial surrogate model during transfer training. The total amount of data in the relevant data of the target domain; This indicates that during the initial transfer training of the proxy model from the source domain to the target domain, the relevant data in the target domain is the [missing information]. The data and the first Feature similarity of data; This represents the first relevant data in the target domain of a graph neural network. The data and the first Feature similarity of data.

6. The method for cross-condition reliability assessment and uncertainty propagation of high-temperature components for peak shaving according to claim 1, characterized in that, The coverage metric is expressed as: ； In the formula, Indicates coverage metric, Indicates candidate samples The feature description vector, Indicates the sampled points The feature description vector, This indicates the number of candidate sampled points.

7. The method for cross-condition reliability assessment and uncertainty propagation of high-temperature components for peak shaving according to claim 1, characterized in that, Using the relevant data from the source domain, the inherent uncertainty, and the cognitive uncertainty as node attributes, and combining them with the relevant data from the target domain, a graph neural network is constructed, including: Using the relevant data from the source domain, the inherent uncertainty, and the cognitive uncertainty as node attributes, an initial graph neural network is constructed. The adjacency matrix of the initial graph neural network is determined based on the node attributes; The node attributes and adjacency matrix of the initial graph neural network are updated using an attention mechanism based on a set number of times and relevant data of the target domain, thus obtaining the graph neural network.

8. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the method for cross-condition reliability assessment and uncertainty propagation of high-temperature components oriented towards peak shaving, as described in any one of claims 1-7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the cross-condition reliability assessment and uncertainty propagation method for high-temperature components oriented towards peak shaving, as described in any one of claims 1-7.

10. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the cross-condition reliability assessment and uncertainty propagation method for high-temperature components oriented towards peak shaving, as described in any one of claims 1-7.

Images