Transient multi-physics field order reduction reconstruction method and device driven by time sequence working condition

Abstract

The invention discloses a transient multi-physics field reduced-order reconstruction method and device driven by a time sequence working condition, and relates to the technical field of digital twinning and calculation simulation acceleration. The method comprises the following steps: performing Latin hypercube sampling on monitoring working condition parameters to obtain a working condition parameter combination covering a sample space; establishing a high-fidelity finite element model to obtain geometric topology information; establishing a working condition parameter amplitude curve, performing batch finite element calculation in combination with the high-fidelity finite element model to obtain transient multi-physics field response data, and constructing a graph data structure; the training graph neural network auto-encoder comprises an encoder and a decoder; and training a time sequence neural network, establishing a mapping relation from time sequence working condition parameters to low-dimensional latent variables, and realizing rapid prediction and reconstruction from the time sequence working condition parameters to transient multi-physical field whole-field distribution in combination with a graph neural network decoder. According to the method, online rapid prediction of the transient physical field of the high-temperature turbine component can be realized, and the method has the advantages of high prediction precision, fast calculation response, strong cross-working-condition generalization capability and the like.

Description

Technical Field

[0001] This application relates to the fields of digital twin and computational simulation acceleration technology, and in particular to a time-series condition-driven transient multiphysics order reduction reconstruction method and apparatus. Background Technology

[0002] Gas turbines and other high-temperature rotating machinery operate under complex and variable conditions, often subjected to extreme loads. Real-time sensing and assessment of the structural response and temperature distribution of critical components are fundamental to ensuring the long lifespan, high reliability, and safe and stable operation of such equipment. However, due to the harsh operating conditions of gas turbines, such as high temperature and high pressure, the placement of sensors and the information they can acquire are very limited, making it impossible to directly obtain full-field physical quantities. Currently, engineering practice mainly relies on high-fidelity finite element models to analyze and predict temperature, stress, and strain fields. While finite element models offer high computational accuracy and good physical interpretability, their solution process is computationally expensive; when the model involves multi-physics coupling such as thermo-mechanical fields, the computation time often increases exponentially, making it difficult to meet the needs of online real-time calculation and rapid decision-making.

[0003] Some existing order reduction models compress the high-dimensional physical field across all nodes into a small number of latent variables or modal coefficients by learning or constructing low-dimensional representations. This can reduce the time required for a single transient solution from minutes / hours to seconds or even milliseconds, alleviating computational pressure. However, traditional linear order reduction methods struggle to handle stress concentrations caused by geometric discontinuities and strong nonlinear mappings resulting from thermo-mechanical coupling effects in complex components such as gas turbines. This leads to insufficient accuracy in "order reduction-reconstruction" and inadequate preservation of local peak values. Summary of the Invention

[0004] The purpose of this application is to provide a time-series operating condition-driven transient multiphysics order reduction reconstruction method and device, which can realize online and rapid prediction of transient physical fields of high-temperature turbine components, and has the advantages of high prediction accuracy, fast computation response and strong generalization ability across operating conditions.

[0005] To achieve the above objectives, this application provides the following solution.

[0006] In a first aspect, this application provides a time-series condition-driven transient multiphysics order reduction reconstruction method, including the following:

[0007] Latin hypercube sampling was performed on the monitoring operating parameters during the start-up and shutdown of high-temperature turbine components to obtain several sets of operating parameter combinations covering the sample space; the monitoring operating parameters include speed, gas temperature, initial temperature of turbine components, temperature rise rate, temperature fall rate, speed rise rate, speed fall rate, output power rise rate, and output power fall rate.

[0008] Based on each set of operating condition parameter combinations and physical constraints, a corresponding operating condition parameter amplitude curve is constructed; the physical constraints include: the time when the temperature rise begins is after the time when the speed rise begins, the time when the temperature falls begins is before the time when the speed falls, and the amplitude curve of the transient heat transfer coefficient satisfies the equivalent heat transfer boundary theory.

[0009] Based on each set of operating condition parameter combinations, a corresponding high-fidelity finite element model is established, a finite element mesh is generated, and geometric topology information is obtained.

[0010] Based on the high-fidelity finite element model, batch finite element calculations are performed on the amplitude curves of the operating conditions to obtain transient multiphysics response data.

[0011] A graph data structure is constructed based on the transient multiphysics response data and the geometric topology information.

[0012] A graph neural network autoencoder is trained based on the graph data structure to obtain the trained graph neural network autoencoder; the graph neural network autoencoder includes a graph neural network encoder and a graph neural network decoder.

[0013] Based on the graph neural network encoder, low-dimensional latent variables of the entire physical field data are obtained.

[0014] Based on a time-series neural network, the time-series operating condition parameter sequence obtained from the amplitude curve of the operating condition parameter is used as input, and the low-dimensional latent variable is used as output to obtain a trained time-series neural network and establish a mapping relationship from the time-series operating condition parameter to the low-dimensional latent variable.

[0015] Based on the trained temporal neural network and the graph neural network decoder, rapid prediction and reconstruction of the transient multiphysics field distribution from temporal operating parameters is achieved.

[0016] 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 time-series condition-driven transient multiphysics order reduction reconstruction method described above.

[0017] Thirdly, this application provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the time-series condition-driven transient multiphysics order reduction reconstruction method described above.

[0018] Fourthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the time-series condition-driven transient multiphysics order reduction reconstruction method described above.

[0019] According to the specific embodiments provided in this application, this application has the following technical effects.

[0020] First, this application solves the problem of traditional methods' difficulty in covering multiple operating conditions and sample spaces during the start-up and shutdown process of high-temperature turbine components by performing Latin hypercube sampling on the monitored operating condition parameters, thus achieving efficient construction of multi-operating condition samples. Second, based on each sampled operating condition parameter combination and physical constraints, an operating condition parameter amplitude curve is constructed, and a corresponding high-fidelity finite element model is established for batch finite element calculations to obtain transient multi-physics response data. Then, based on the transient multi-physics response data and geometric topology information, a graph data structure is constructed, and a graph neural network autoencoder is trained on the graph data structure. The encoder obtains low-dimensional latent variables of the entire physical field data, achieving order reduction of the physical field and solving the problems of high data storage overhead and high computational complexity of high-dimensional physical fields. Third, using the time-series operating condition parameter sequence obtained from the operating condition parameter amplitude curve as input and low-dimensional latent variables as output, a time-series neural network is trained to establish a mapping relationship from time-series operating condition parameters to low-dimensional latent variables. By capturing the temporal dependencies and nonlinear characteristics of operating condition parameter sequences, this method solves the problem that traditional order reduction methods struggle to establish nonlinear mappings between operating condition parameters and physical fields. Finally, based on a trained temporal neural network, low-dimensional latent variables are predicted in real time. Then, a graph neural network decoder rapidly reconstructs the low-dimensional latent variables into a full-field physical field, achieving rapid prediction and reconstruction from temporal operating condition parameters to the transient multiphysics field full-field distribution. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is an application environment diagram of a time-series condition-driven transient multiphysics order reduction and reconstruction method according to an embodiment of this application.

[0023] Figure 2 This is a flowchart illustrating a time-series condition-driven transient multiphysics order reduction reconstruction method provided in an embodiment of this application.

[0024] Figure 3 This is a schematic diagram illustrating the offline training and online prediction process of a time-series condition-driven transient multiphysics order reduction reconstruction method provided in an embodiment of this application.

[0025] Figure 4This is a schematic diagram of the geometric model and mesh generation in a finite element model of a gas turbine blade provided in an embodiment of this application.

[0026] Figure 5 This is a schematic diagram showing the setting of boundary conditions and loads in a finite element model of a gas turbine blade provided in an embodiment of this application.

[0027] Figure 6 This is a schematic diagram illustrating the timing changes of different operating parameters during the start-up and shutdown process of a gas turbine, as provided in an embodiment of this application.

[0028] Figure 7 The original cloud map of the stress field of a gas turbine blade at a certain moment is provided for an embodiment of this application.

[0029] Figure 8 This is a reconstructed cloud map of the stress field of a gas turbine blade at a certain moment, provided as an embodiment of this application.

[0030] Figure 9 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0031] First, the technical terms involved in the embodiments of this application will be introduced.

[0032] Graph Neural Networks (GNNs) are a class of deep learning models for graph-structured data. Based on nodes and their neighborhood connections, they extract structural and attribute features through message passing or aggregation mechanisms. An autoencoder, consisting of an encoder and a decoder, is used to compress high-dimensional data into low-dimensional latent variables and then reconstruct the original data. Specifically, a graph neural network autoencoder uses a GNN as its encoding and decoding core to perform reduced-order representation learning and full-field reconstruction on physical field graph data with grid nodes as nodes and topological adjacencies as edges.

[0033] Temporal neural networks are used to model time-varying sequential data, learning temporal correlations and dynamic evolution patterns. Typical models include Temporal Convolutional Networks (TCNs), Long Short-Term Memory Neural Networks (LSTMs), and Gated Recurrent Units (GRUs). Specifically: TCNs are based on one-dimensional convolution and dilated convolution to model long-term dependencies, featuring high parallel computing efficiency and good stability; LSTMs and GRUs are variants of recurrent neural networks that alleviate the gradient vanishing problem through gating mechanisms, making them suitable for nonlinear temporal mapping and state memory.

[0034] Latin hypercube sampling (LHS) is a space-filling experimental design method that divides the range of each input variable into several intervals and randomly samples within each interval, while ensuring that each variable is covered in each interval. This allows for a uniform exploration of the high-dimensional parameter space with fewer samples and is often used to build training data and improve the model's generalization ability across operating conditions.

[0035] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0036] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0037] This application provides a time-series condition-driven transient multiphysics order reduction reconstruction method, which can be applied to, for example... Figure 1In 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 set up independently, integrated into server 104, or placed in the cloud or on other servers. Terminal 102 can send monitoring operating parameters during the start-up and shutdown process of high-temperature turbine components to server 104. After receiving the monitoring operating parameters, server 104 performs Latin hypercube sampling on the monitoring operating parameters during the start-up and shutdown process of high-temperature turbine components to obtain several sets of operating parameter combinations covering the sample space; constructs corresponding operating parameter amplitude curves based on each set of operating parameter combinations and physical constraints; establishes corresponding high-fidelity finite element models based on each set of operating parameter combinations, divides finite element meshes, and obtains geometric topology information; performs batch finite element calculations on the operating parameter amplitude curves based on the high-fidelity finite element models to obtain transient multiphysics response data; and performs batch finite element calculations on the transient multiphysics response data and... The geometric topology information is used to construct a graph data structure; a graph neural network autoencoder is trained based on the graph data structure to obtain a trained graph neural network autoencoder; the graph neural network autoencoder includes a graph neural network encoder and a graph neural network decoder; based on the graph neural network encoder, low-dimensional latent variables of the entire physical field data are obtained; based on a temporal neural network, a trained temporal neural network is obtained with the temporal operating condition parameter sequence obtained according to the amplitude curve of the operating condition parameter as input and the low-dimensional latent variables as output, establishing a mapping relationship from the temporal operating condition parameters to the low-dimensional latent variables; based on the trained temporal neural network and the graph neural network decoder, rapid prediction and reconstruction from the temporal operating condition parameters to the transient multiphysics field distribution is achieved. The server 104 can feed back the reconstructed transient multiphysics field to the terminal 102. In addition, in some embodiments, the time-series operating condition driven transient multiphysics order reduction reconstruction method can also be implemented by the server 104 or the terminal 102 separately. For example, the terminal 102 can directly perform relevant operations on the monitoring operating condition parameters during the start-up and shutdown process of the high-temperature turbine component, or the server 104 can obtain the monitoring operating condition parameters during the start-up and shutdown process of the high-temperature turbine component from the data storage system and perform relevant operations on the monitoring operating condition parameters during the start-up and shutdown process of the high-temperature turbine component.

[0038] The terminal 102 can be, but is not limited to, various desktop computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices. IoT devices can include smart speakers, smart TVs, smart air conditioners, and smart in-vehicle devices. Portable wearable devices can include smartwatches, smart bracelets, and head-mounted devices. The server 104 can be implemented using a standalone server or a server cluster composed of multiple servers, or it can be a cloud server.

[0039] In one exemplary embodiment, such as Figure 2 As shown, a time-series condition-driven transient multiphysics order reduction reconstruction method is provided. This method is executed by a computer device, specifically by a terminal or server alone, or by both a terminal and a server. In this embodiment, the method is applied to... Figure 1 Taking server 104 as an example, the explanation includes the following steps 201 to 209.

[0040] Step 201: Perform Latin hypercube sampling on the monitoring operating parameters during the start-up and shutdown process of the high-temperature turbine components to obtain several sets of operating parameter combinations covering the sample space; the monitoring operating parameters include, but are not limited to, speed, gas temperature, initial temperature of the turbine components, temperature rise rate, temperature fall rate, speed rise rate, speed fall rate, output power rise rate, and output power fall rate.

[0041] Specifically, in order to construct normalized operating condition parameter amplitude curves, it is necessary to monitor peak speed, peak gas temperature, and peak output power; the number of Latin hypercube samples must be sufficient to ensure coverage of conventional start-stop strategies, generally no less than 50 sets.

[0042] Step 202: Construct corresponding operating condition parameter amplitude curves based on each set of operating condition parameter combinations and physical constraints; the physical constraints include: the time when the temperature rise begins is after the time when the speed rise begins, the time when the temperature falls begins is before the time when the speed falls, and the amplitude curve of the transient heat transfer coefficient satisfies the equivalent heat transfer boundary theory.

[0043] Specifically, the amplitude curve of the operating condition parameters is calculated and generated based on the aforementioned combination of monitoring parameters and necessary physical constraints. The physical constraints include, but are not limited to: the time when the temperature rise begins is always after the time when the speed rise begins, the time when the temperature falls begins is always before the time when the speed falls, and the amplitude curve of the transient heat transfer coefficient must satisfy the equivalent heat transfer boundary theory, etc.

[0044] Step 203: Establish a corresponding high-fidelity finite element model based on each set of working condition parameter combinations, divide the finite element mesh, and obtain geometric topology information.

[0045] Step 204: Perform batch finite element calculations on the amplitude curves of the operating conditions parameters based on the high-fidelity finite element model to obtain transient multiphysics response data.

[0046] Specifically, the batch finite element calculation method includes: manually inputting the constructed operating condition parameter amplitude curves into a benchmark high-fidelity finite element model for calculation, or automatically inputting them into the high-fidelity finite element model for calculation using a batch processing script developed in Python. The multiphysics transient response data is spliced ​​physical field data for a set of operating condition parameter combinations. Each row of the data represents the physical quantities of all nodes in the field at a certain time point, and each column represents the process of physical quantity change of a certain node over a time scale. By sampling a certain number of operating condition parameter combinations in the aforementioned steps, a corresponding number of operating condition parameter amplitude curve combinations will be generated, which in turn correspond to a corresponding number of physical field transient response data.

[0047] Step 205: Construct a graph data structure based on the transient multiphysics response data and the geometric topology information.

[0048] Specifically, firstly, the graph data structure compatible with the graph learning framework refers to the data structure compatible with the PyTorch Geometric deep learning library, specifically including: data describing the geometric topology information of finite element model elements and nodes, time axis data, and physical field data. Secondly, further, for a finite element model, its nodes can be constructed as nodes in the graph data structure. Furthermore, because the nodes of shared elements are located in the same discrete matrix during finite element calculations, it is assumed that all nodes of shared elements have edges in the graph data structure.

[0049] Step 206: Train a graph neural network autoencoder based on the graph data structure to obtain the trained graph neural network autoencoder; the graph neural network autoencoder includes a graph neural network encoder and a graph neural network decoder.

[0050] Specifically, the graph neural network encoder extracts node and neighborhood features based on graph convolution / graph aggregation layers (such as GCN, GraphSAGE, GAT, etc.) to compress the full-field physical quantities at a single or multiple time step into low-dimensional latent variables (low-dimensional representation); the graph neural network decoder reconstructs the full-field physical field based on the low-dimensional latent variables. During training, the reconstruction error can be used as the objective function, and strategies such as weighting of key regions and peak preservation can be combined to reduce the reconstruction deviation in stress concentration areas.

[0051] Step 207: Based on the graph neural network encoder, obtain the low-dimensional latent variables of the entire physical field data.

[0052] Step 208: Based on the time-series neural network, using the time-series operating condition parameter sequence obtained from the amplitude curve of the operating condition parameter as input and the low-dimensional latent variable as output, a trained time-series neural network is obtained, and a mapping relationship from the time-series operating condition parameters to the low-dimensional latent variable is established.

[0053] Specifically, firstly, the time-series operating parameters refer to the operating parameters at each time point sampled from the aforementioned operating parameter amplitude curves based on the time axis. Secondly, temporal neural networks are a general term for neural network models that process temporal inputs, including but not limited to: temporal convolutional networks, long short-term memory neural networks, or gated recurrent units. Thirdly, the time-series operating parameters and the low-dimensional latent variables (or low-dimensional representations) generated during the aforementioned graph neural network autoencoder order reduction process have a one-to-one temporal correspondence. The number of time points in the time axis data during the aforementioned graph data construction process determines the number of time-series operating parameters and low-dimensional latent variables.

[0054] Step 209: Based on the trained temporal neural network and the graph neural network decoder, realize the rapid prediction and reconstruction of the transient multiphysics field distribution from the temporal operating parameters.

[0055] Specifically, firstly, the online inference and prediction process involves deploying the optimal model parameters of the temporal neural network and graph neural network decoder obtained during offline training to an online monitoring platform. Based on the input temporal operating parameters, the physical field state is calculated in real time and visualized in three dimensions. Secondly, the operating parameters are data such as engine speed, gas temperature, and output power obtained from monitoring signals from the gas turbine control system. Thirdly, the transient multiphysics fields include, but are not limited to, stress fields, strain fields, temperature fields, and damage fields.

[0056] By implementing steps 201 to 209 above, this application can better characterize the strong nonlinear relationship caused by thermal-mechanical coupling and stress concentration through the combined modeling method of "graph autoencoder order reduction + temporal mapping prediction", reduce the reconstruction error in the peak region, thereby significantly improving the computational efficiency while ensuring accuracy, and has good cross-condition generalization ability.

[0057] In another exemplary embodiment of this application, in order to accurately construct the graph data structure based on the transient multiphysics response data and the geometric topology information, the above step 205 is replaced by the following steps 2051 to 2052.

[0058] Step 2051: Input the transient multiphysics response data and the geometric topology information into the graph data builder.

[0059] Step 2052: The graph data builder reads the geometric topology information and constructs a graph data structure based on the inclusion relationship between cells and nodes and the geometric coordinates of the nodes.

[0060] The graph data structure includes nodes and node features, edges and edge features; wherein, the nodes in the graph data structure are constructed based on the nodes in the finite element model; the node features in the graph data structure are the physical quantity values ​​of the corresponding nodes in the physical field data; the edges in the graph data structure are constructed based on the pairs of all nodes in each element of the finite element model, and the edge features are the weights of the edges, with a default edge feature of 1 (all edges are consistent), and duplicate edges shared between different elements should be removed.

[0061] In another exemplary embodiment of this application, in order to accurately train the graph neural network autoencoder based on the graph data structure and obtain the trained graph neural network autoencoder, the above step 206 is replaced by the following steps 2061 to 2063.

[0062] Step 2061: Define the structure of the graph neural network autoencoder and obtain the network parameters of the graph neural network autoencoder.

[0063] Step 2062: Define the loss function of the graph neural network autoencoder, and determine the loss value of the graph neural network autoencoder based on the reconstruction error using the loss function of the graph neural network autoencoder.

[0064] Step 2063: Input the graph data structure into the graph neural network autoencoder, and optimize the network parameters of the graph neural network autoencoder through iterative training based on the loss value of the graph neural network autoencoder to obtain the trained graph neural network autoencoder.

[0065] In another exemplary embodiment of this application, in order to accurately train the graph neural network autoencoder based on the graph data structure and obtain the trained graph neural network autoencoder, the above step 208 is replaced by the following steps 2081 to 2084.

[0066] Step 2081: Based on the amplitude curve of the operating condition parameters, a time-series operating condition parameter sequence is obtained by sampling; each time-series operating condition parameter sequence corresponds to a low-dimensional latent variable.

[0067] Step 2082: Input the time-series operating condition parameter sequence into the time-series neural network to obtain the output of the time-series neural network.

[0068] Step 2083: Based on the output of the time-series neural network, the low-dimensional latent variables corresponding to the time-series operating condition parameter sequence, and the determined loss function of the time-series neural network, determine the loss value of the time-series neural network.

[0069] Step 2084: Based on the loss value of the time-series neural network, optimize the network parameters of the time-series neural network to obtain the trained time-series neural network, and establish a mapping relationship from time-series operating parameters to low-dimensional latent variables.

[0070] In another exemplary embodiment of this application, in order to accurately predict and reconstruct the rapid distribution of transient multiphysics field from time-series operating parameters based on the trained temporal neural network and the graph neural network decoder, the above step 209 is replaced by the following steps 2091 to 2093.

[0071] Step 2091: Obtain the real-time time-series operating condition parameter sequence online.

[0072] Step 2092: Input the real-time time series operating condition parameter sequence into the trained time series neural network to obtain the predicted low-dimensional latent variables.

[0073] Step 2093: Input the predicted low-dimensional latent variables into the decoder to reconstruct the transient multiphysics field distribution.

[0074] In summary, this application, based on Latin hypercube sampling and combined with a high-fidelity finite element model, constructs and accumulates multi-condition transient data capable of characterizing the start-up and shutdown processes of real gas turbine high-temperature turbine components. Through offline training of a graph neural network autoencoder and a temporal neural network, a rapid prediction and reconstruction process is established from temporal operating condition parameters to the full-field distribution of multi-physics fields, thereby enabling online inference of the real-time evolution of the physical field. This method can provide technical support for rapid condition assessment and online operation and maintenance management of gas turbine high-temperature turbine components, possessing advantages such as high prediction accuracy, fast computational response, and strong cross-condition generalization ability.

[0075] This application provides an offline training and online prediction process for a time-series condition-driven transient multiphysics order reduction reconstruction method, such as... Figure 3 As shown, the specific process is as follows.

[0076] The offline training phase includes the following steps.

[0077] 1. Data Preparation: Export the transient response data of the physical field and the geometric topology information of the model mesh obtained from the aforementioned calculations from the finite element method (FEM) software; sample the time-series load parameter data from the previously constructed load parameter amplitude curves. Specifically, the FEM software generally refers to ABAQUS or ANSYS, etc.

[0078] 2. Graph Data Construction: Input the prepared physics data and geometric topology information into the graph data builder to construct data compatible with the PyTorch Geometric deep learning library.

[0079] 3. Training of graph neural network autoencoder: Input the constructed graph data into the graph neural network autoencoder to achieve low-dimensional representation learning and reconstruction of the entire physical field. The specific operation process is as follows.

[0080] Specifically, there is a graph structure corresponding to the finite element mesh, and the graph structure expression is as follows.

[0081] .

[0082] in, For graph structure; A set of nodes; Let it be the set of edges.

[0083] For any point in time The physical quantities of the entire field at that moment are written as nodal characteristic matrices, and the expression of the characteristic matrix is ​​as follows.

[0084] .

[0085] in, Indicates a point in time The node characteristics, i.e., time points. The total physical quantities of the entire field; T This represents the number of sampling points on the time axis. For a point in time; For the number of nodes, the number of nodes ; In this embodiment, the physical quantity dimension for each node is... .

[0086] The encoder updates node representations through graph message passing, as shown in the following formula.

[0087] .

[0088] in, For the first +1 level node The hidden feature vectors; It is a non-linear activation function; All of these are training parameter matrices; For the first Layer Time Node The hidden feature vectors; For the first Layer Time Node j The hidden feature vectors; ,in for The Rows represent nodes At any moment The physical quantity; Represents a node The set of neighboring nodes; This represents the neighborhood aggregation weight, which is specifically determined by the chosen pooling mechanism.

[0089] To obtain a full low-dimensional representation, we need to read out or pool the node embeddings, as shown in the following expression.

[0090] .

[0091] in, , For time points The corresponding low-dimensional latent variables (low-dimensional representation); To read out the operator, operations such as average pooling, max pooling, and attention pooling can be performed; For the process After layer encoding, a node-level embedding matrix is ​​obtained. ,in It represents the dimension of the low-dimensional latent space.

[0092] The decoder then reconstructs the entire physical field based on the latent variables, as shown in the following expression.

[0093] .

[0094] in, For the time points obtained from reconstruction The total physical quantities of the entire field; For parameters decoder function; This represents a graph structure, consistent with the graph structure described above.

[0095] The reconstruction error is used to train the loss function, and an optional form of mean squared error is given, as shown in the following expression.

[0096] .

[0097] in, This is the global reconstruction loss function based on the reconstruction error; It is a norm 2. Furthermore, in some embodiments, if the set of dangerous point nodes and their neighborhood are defined as , n To reduce the problem of peak data in stress concentration areas being homogenized or smoothed by global numerical values, the number of critical points is taken as... The neighborhood set consists of 1 neighboring nodes. And define the local subgraph node set, as shown in the following expression.

[0098] .

[0099] in, To focus on dangerous nodes The constructed local node set has the following number of nodes. ; This is a dangerous node; The number of neighboring nodes; for Each neighboring node constitutes a neighborhood set.

[0100] The expression for extracting a submatrix from a local set of nodes is as follows.

[0101] .

[0102] in, Indicates from Extracting nodes The corresponding local physical quantity matrix; Similarly.

[0103] Therefore, the local reconstruction loss is defined as follows.

[0104] .

[0105] Finally, the overall loss function (i.e. the loss function of the graph neural network autoencoder) is expressed as follows.

[0106] .

[0107] in, For local reconstruction loss function; This is a weighting factor used to balance the global reconstruction accuracy and the reconstruction accuracy of stress concentration areas.

[0108] 4. Temporal Neural Network Training: Using the temporal operating parameters obtained from the amplitude curve of the operating parameters as input features, and the low-dimensional latent variables (low-dimensional representation) obtained by the graph neural network autoencoder as output features, the temporal neural network is trained to establish a mapping relationship.

[0109] Specifically, this includes setting the sequence of operating parameters obtained from sampling the amplitude curve, as shown in the following expression.

[0110] .

[0111] in, For time points The working condition parameter vector; The dimension of the operating condition parameters, This is the vector of operating parameters at time point 1.

[0112] The low-dimensional latent variable sequence expression obtained from the aforementioned trained graph neural network autoencoder is as follows.

[0113] .

[0114] in, For time points The low-dimensional representation of .

[0115] The next training objective is to learn the mapping function. The expression is as follows.

[0116] .

[0117] in, It is a temporal neural network. These are trainable parameters; For the predicted low-dimensional latent variable sequence, .

[0118] The training process can use mean squared error as the loss function. The expression for the loss function (i.e., the loss function of sequential neural networks) is as follows.

[0119] .

[0120] in, For latent variable prediction error; For time points The corresponding low-dimensional latent variables; These are the low-dimensional latent variables obtained from the prediction.

[0121] In some embodiments, Temporal convolutional neural networks, long short-term memory neural networks, and gated recurrent units can be selected, specifically including the following:

[0122] (1) When using a temporal convolutional neural network, its first... The expression for the causal dilated convolution of the layer is as follows.

[0123] .

[0124] in, For the first Layer at a point in time Intermediate features; ; The kernel length is [length]. Void ratio; For the first Layer Each convolutional kernel weight; For bias; It is a non-linear activation function. The final result is obtained through the output layer. .

[0125] (2) When using a long short-term memory neural network, its long short-term memory units at time points The overall mapping expression is as follows.

[0126] .

[0127] in, For time points The timing condition parameter vector; It is in a hidden state; In cellular state; These are trainable parameters; The dimension of the operating condition parameters; Let be the dimension of the hidden state.

[0128] The gating parameters of the Long Short-Term Memory (LSTM) units are generated by the gating mapping function, and the state is updated accordingly. The relevant expressions are as follows.

[0129] .

[0130] .

[0131] .

[0132] in, These are the input gate, the forget gate, and the output gate; Candidate memories; This involves concatenating vectors. This is a gated mapping function implemented by a fully connected layer, where the gated component in the output is activated by Sigmoid and the candidate memory is activated by Tanh. It is the Hadamard product.

[0133] The output of a long short-term memory neural network can be used to obtain low-dimensional latent variable predictions through hidden state mapping, as shown in the following expression.

[0134] .

[0135] in, For time points The low-dimensional representation of the predicted value; and These are trainable parameters; Dimensionality is represented by a low-dimensional representation.

[0136] (3) When a gated loop unit is used, the gated loop unit at time point Receive input operating condition parameter vector And based on the hidden state of the previous moment Update to get the current hidden state The expression for its update method is as follows.

[0137] .

[0138] in, For time points Timing parameters; For time points The hidden state; Indicates parameters Gated loop unit mapping; The dimension of the operating condition parameters; Let be the dimension of the hidden state.

[0139] In some embodiments, the gated loop unit employs update gates and reset gates to control the fusion of historical information and current input, and its gated vectors can be merged, as shown in the following expression.

[0140] .

[0141] in, These are the update door and the reset door, respectively. It is a combination of trainable affine transformation and Sigmoid nonlinear mapping.

[0142] Based on the gating variable, the candidate hidden state and hidden state update of the gated loop unit can be simplified, as shown in the following expression.

[0143] .

[0144] .

[0145] in, This is a hidden candidate state. For trainable affine transformations; It is the Hadamard product.

[0146] The final output of the low-dimensional latent variable prediction can be obtained from the hidden state through a linear mapping, as shown in the following expression.

[0147] .

[0148] in, For time points The low-dimensional representation of the predicted value; and These are trainable parameters; The dimension of the low-dimensional representation.

[0149] Online prediction phase: After obtaining the optimal model parameters through offline training, the trained temporal neural network is... Graph Neural Network Decoder Deployed to an online monitoring platform or digital twin system to achieve rapid prediction and reconstruction of the entire field distribution of transient multiphysics, specifically including the following steps.

[0150] 1. Online data acquisition and preprocessing: During online operation, operating condition parameter sequences are collected in real time from the gas turbine control system or monitoring system. ,in, For time points The operating parameter vector includes, but is not limited to, engine speed, gas temperature, output power, heat transfer coefficient, or other effective characterizing parameters. Online input data undergoes scaling and alignment sampling according to the normalization method determined in the offline phase, including but not limited to: amplitude normalization, time resampling, missing value imputation, and outlier filtering, to ensure consistency between the online input and training distributions.

[0151] 2. Online prediction of latent variables: The preprocessed time series operating parameters are input into the deployed neural network to obtain the predicted low-dimensional latent variables at the corresponding time points, as shown in the following expression.

[0152] .

[0153] in, For time points The low-dimensional representation of the predicted value; To represent the dimension in a low-dimensional way; These are the optimal model parameters obtained through offline training.

[0154] In some embodiments, the temporal neural network may perform inference using a sliding window approach, i.e., the most recent... Operating parameters at each time point As input, to reduce latency and improve stability in online computing.

[0155] 3. Fast full-field reconstruction of the physical field: The low-dimensional latent variables are input into the graph neural network decoder, combined with the pre-stored grid graph structure. Output time point The distribution of physical quantities across the entire field is expressed as follows.

[0156] .

[0157] in, The reconstructed physical field matrix; This represents the number of nodes in the finite element mesh (equivalent to the number of nodes in the graph data). The dimension of the physical field (default) ); These are the optimal decoder parameters obtained through offline training. The total physical quantities include, but are not limited to, the temperature field, stress field, strain field, and the damage field derived from them.

[0158] In some embodiments, to achieve continuous online output, the following can be configured: The above prediction and reconstruction process is repeated every time step or every sampling period to obtain a continuous sequence of the physical field's evolution over time. .

[0159] 4. Online visualization and status assessment: In some embodiments, the online platform can be based on... It provides 3D cloud map visualization and can further output time history curves of physical quantities for key components (hazardous nodes or hazardous areas). The hazardous area can be a predefined set of nodes. and its neighborhood set This is to support key monitoring and alarm threshold determination in hotspot areas.

[0160] 5. Online Result Output and Interface: Online prediction results can be output to the operation and maintenance management system via a data interface for applications such as condition diagnosis, remaining life prediction, maintenance decision support, or control strategy optimization. Output formats include, but are not limited to: physical quantity files of all nodes, key node curves, statistical characteristics (peak value, mean, gradient, etc.), and risk indicators.

[0161] In this embodiment, a method for reducing the order of transient multiphysics field driven by time-series operating conditions is described using a heavy-duty gas turbine blade as an example. The specific implementation steps are as follows.

[0162] The blade material is selected from UGTC48 high-temperature alloy or equivalent high-temperature alloy, with an operating temperature of 900℃. Specific material parameters are set as follows: the material creep power law coefficient is set to... The stress value was set to 39.11, the time order to -0.311, and the material density to [value missing]. Young's modulus is set to Poisson's ratio was set to 0.3; the linear expansion coefficient was set to... Thermal conductivity is set to Specific heat set to The thermo-mechanical coupling response of the blades under the combined effects of high-temperature combustion gas scouring and high-speed rotation is considered. For ease of explanation, this embodiment uses the temperature field, equivalent stress field, and equivalent strain field as the main output physical quantities, and calculates fatigue damage and creep damage based on these.

[0163] like Figure 4 As shown, Figure 4 (a) is a schematic diagram of the geometric model in the finite element model of a gas turbine blade; Figure 4(b) is a schematic diagram of the geometric model mesh generation in the finite element model of a gas turbine blade.

[0164] A three-dimensional geometric model of the blade was established and a finite element mesh was generated. Constraint boundaries were set at the connection between the blade root and the rotor disk. The mesh type was set to C3D4, generating a total of 33,023 elements and 8,682 nodes. Figure 5 As shown, time-varying thermal and load boundaries are applied to the blade surface, including but not limited to: changes in combustion gas temperature over time, changes in transient heat transfer coefficient over time, and centrifugal loads caused by changes in rotational speed over time. The transient solution can employ a thermo-mechanical coupling analysis step, with equal-interval sampling along the time axis to obtain the output of physical quantities at multiple time points across the entire field, thus yielding transient multiphysics response data. In this embodiment, a set of full-field data nodes corresponding to a set of operating conditions is sequentially concatenated to obtain a matrix-like data representation: each row corresponds to the distribution of physical quantities at a certain time point across the entire field, and each column corresponds to the evolution of physical quantities at a certain node in the time dimension.

[0165] Latin hypercube sampling is performed on the sample space of monitoring parameters during start-up and shutdown. The sampled parameters include, but are not limited to: peak speed and its acceleration and deceleration, turbine blade initial temperature, peak output power and its acceleration and deceleration, and equivalent characterization parameters of transient heat transfer boundaries. The number of samples is no less than 50 sets to cover common start-up and shutdown strategies and enhance cross-condition generalization capability. Based on each set of sampled parameters, an amplitude curve of the operating condition parameters is constructed. A schematic diagram of the time-series changes of different operating condition parameters during the start-up and shutdown process of a gas turbine is shown below. Figure 6 As shown. The generation of the amplitude curve needs to meet certain physical constraints. Generally, the starting time of temperature rise is set after the starting time of rotational speed rise, the starting time of temperature fall is set after the starting time of rotational speed fall, and the amplitude curve of the heat transfer coefficient satisfies the consistency of the equivalent heat transfer boundary, etc. The amplitude curve is input into the finite element model for batch transient solution to obtain a multi-condition transient multiphysics response dataset covering the sample space.

[0166] The blade's mesh topology and transient physics output are organized into a graph data structure: finite element nodes are used as graph nodes, with node characteristics representing the physical quantity (e.g., temperature, stress, or strain) at a specific time point; edges are constructed by connecting nodes sharing the same finite element element pairwise, and duplicate edges are removed. The edge characteristic can be set to a default constant of 1. This results in a data structure compatible with the PyTorch Geometric graph learning framework, including geometric topology information, time axis information, and physics data. In this embodiment, the finite element model has 33,023 elements and 8,682 nodes. After the graph data constructor, a total of 8,682 nodes and 95,668 edges are constructed, with 421 frames sampled for each operating condition.

[0167] Subsequently, a graph neural network autoencoder was trained. The encoder compressed the entire physical field into low-dimensional latent variables, and the decoder reconstructed the entire physical field based on these low-dimensional latent variables. The GNN encoder consisted of 3 GraphSAGE convolutional layers with a hidden layer width of 64 and attention pooling as the pooling method. The GNN decoder also consisted of 3 GraphSAGE convolutional layers with a hidden layer width of 64 and a latent space dimension of 16. The training objective function was global reconstruction mean squared error; the learning rate was set to 0.0005; the training epochs were 250, with an early stopping tolerance threshold of 30 epochs; the optimizer was Adam, with an optimizer weight of 0.000001; and the batch size was 20. Furthermore, to better incorporate geometric topological information, normalized node geometric coordinates were concatenated to node features; to prevent gradient explosion during training and subsequent failure, gradient clipping was enabled with a threshold of 5. The training device used was an NVIDIA GeForce RTX 5080 with CUDA version 13.1.

[0168] After training the graph neural network autoencoder, for each set of operating condition samples, the time-series operating condition parameter sequence is sampled from the amplitude curve according to the time axis, and aligned with the low-dimensional latent variable sequence obtained by the graph autoencoder in the time dimension to form a supervised learning training sample pair. The time-series neural network (including but not limited to temporal convolutional neural networks, long short-term memory neural networks, gated recurrent units, etc.) is trained to establish the mapping relationship from the time-series operating condition parameters to the low-dimensional latent variables.

[0169] During online inference (the sequence of operating parameters collected online and normalized in an offline manner is input into a time-series neural network to obtain the predicted value of the low-dimensional latent variable, and then the latent variable is input into a graph neural network decoder to quickly obtain the distribution of the physical quantities of the entire field at the current moment), real-time prediction and reconstruction of the distribution of the transient multiphysics field from the monitored operating parameters is realized.

[0170] Figure 7 This is an original cloud map of the stress field of a gas turbine blade at a certain moment, provided in an embodiment of this application. Figure 8 This is a reconstructed stress field cloud diagram of a gas turbine blade at a certain moment, provided in an embodiment of this application. A comparison of the two diagrams shows the original finite element cloud diagram and the cloud diagram reconstructed using the reduced-order method. It can be seen that this method possesses excellent reduction-order and reconstruction accuracy. In the online inference stage, under the same computer equipment, this method takes approximately 20 seconds to perform a full physical prediction for a single operating condition, while the finite element analysis takes approximately 30 minutes to calculate the full physical field for a single operating condition. The reduced-order model reduces the time consumption by approximately 98.9% compared to the full-order model.

[0171] In summary, this embodiment takes heavy-duty gas turbine blades as the object, constructs transient thermo-mechanical coupling high-fidelity data covering the parameter space of multiple operating conditions during start-up and shutdown, and realizes rapid prediction and reconstruction from time-series operating conditions parameters to the full-field distribution of transient multiphysics fields based on a modeling framework combining graph neural network autoencoders and temporal neural networks.

[0172] By introducing a graphical representation of finite element mesh topological constraints, this embodiment can effectively characterize the complex geometry of the blade and the characteristics of stress concentration regions. By using low-dimensional latent variables as intermediate representations, the difficulty of temporal mapping and the computational complexity of the model are significantly reduced. Furthermore, by combining a temporal neural network to learn the dynamic evolution of the start-up and shutdown process, online inference of the continuous evolution of the transient physical field is achieved. Implementation results show that, while maintaining high reconstruction accuracy, this method offers an order-of-magnitude improvement in computational efficiency compared to traditional finite element transient solutions, meeting the application requirements for online monitoring and rapid evaluation.

[0173] Therefore, this embodiment verifies the feasibility and effectiveness of the method described in this application in transient multiphysics modeling of high-temperature turbine components, demonstrating that this method can provide reliable technical support for online status perception, performance evaluation, and operation and maintenance decision-making of key gas turbine components, and has good engineering application prospects. This application also provides an application scenario in which the aforementioned time-series operating condition-driven transient multiphysics field order reduction and reconstruction method is applied. Specifically, the time-series operating condition-driven transient multiphysics field order reduction and reconstruction method provided in this embodiment can be applied in the digital twin and intelligent operation and maintenance scenario of gas turbines. This application scenario includes a data acquisition stage, a data processing stage, and a state assessment stage. The data acquisition stage is used to acquire the time-series operating condition parameters during the operation of the gas turbine in real time; the time-series operating condition parameters enter the data processing stage from the data acquisition stage, and are rapidly inferred through the order reduction and reconstruction model constructed by the method of this application to obtain the corresponding transient multiphysics field distribution. In the data processing stage, the physical field reconstruction of the real-time operating condition parameters can be performed based on the collaborative method of "offline training-online inference", that is, the corresponding transient multiphysics field distributions such as temperature field and stress field are generated in real time for high-temperature turbine components.

[0174] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 9As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, 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 in the non-volatile storage media to run. The database stores monitoring parameters during the start-up and shutdown of high-temperature turbine components. The I / O interfaces are used for information exchange between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When executed by the processor, the computer program implements a time-series-driven transient multiphysics order reduction reconstruction method.

[0175] Those skilled in the art will understand that Figure 9 The structures shown are merely block diagrams of some structures related to the present application and do 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 shown in the figures, or combine certain components, or have different component arrangements. In an exemplary embodiment, a computer device is provided, including 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-described method embodiments.

[0176] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0177] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0178] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0179] 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. When executed, the computer program 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 and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0180] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0181] 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 specification.

[0182] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A time-series condition-driven transient multiphysics order reduction and reconstruction method, characterized in that, The time-series-driven transient multiphysics order reduction and reconstruction method includes: Latin hypercube sampling was performed on the monitoring operating parameters during the start-up and shutdown process of the high-temperature turbine components to obtain several sets of operating parameter combinations covering the sample space; the monitoring operating parameters include speed, gas temperature, initial temperature of the turbine components, temperature rise rate, temperature fall rate, speed rise rate, speed fall rate, output power rise rate, and output power fall rate. Based on each set of operating condition parameter combinations and physical constraints, a corresponding operating condition parameter amplitude curve is constructed; the physical constraints include: the time when the temperature rise begins is after the time when the speed rise begins, the time when the temperature falls begins is before the time when the speed falls, and the amplitude curve of the transient heat transfer coefficient satisfies the equivalent heat transfer boundary theory. Based on each set of working condition parameter combinations, a corresponding high-fidelity finite element model is established, a finite element mesh is generated, and geometric topology information is obtained. Based on the high-fidelity finite element model, batch finite element calculations are performed on the amplitude curves of the operating conditions to obtain transient multiphysics response data. A graph data structure is constructed based on the transient multiphysics response data and the geometric topology information; A graph neural network autoencoder is trained based on the graph data structure to obtain a trained graph neural network autoencoder; the graph neural network autoencoder includes: a graph neural network encoder and a graph neural network decoder; Based on the graph neural network encoder, low-dimensional latent variables of the entire physical field data are obtained; Based on the time-series neural network, the time-series operating condition parameter sequence obtained from the amplitude curve of the operating condition parameter is used as input and the low-dimensional latent variable is used as output to obtain the trained time-series neural network and establish the mapping relationship from the time-series operating condition parameter to the low-dimensional latent variable. Based on the trained temporal neural network and the graph neural network decoder, rapid prediction and reconstruction of the transient multiphysics field distribution from temporal operating parameters is achieved.

2. The transient multiphysics order reduction and reconstruction method driven by time-series operating conditions according to claim 1, characterized in that, Based on the transient multiphysics response data and the geometric topology information, a graph data structure is constructed, specifically including: Input the transient multiphysics response data and the geometric topology information into the graph data builder; The graph data builder reads the geometric topology information and constructs a graph data structure based on the inclusion relationship between elements and nodes and the geometric coordinates of the nodes. The graph data structure includes nodes and node features, edges and edge features. The nodes in the graph data structure are constructed based on the nodes in the finite element model. The node features in the graph data structure are the physical quantity values ​​of the corresponding nodes in the transient multiphysics response data. The edges in the graph data structure are constructed based on the pairwise connections between all nodes in each element of the finite element model. The edge features are the weights of the edges.

3. The transient multiphysics order reduction and reconstruction method driven by time-series operating conditions according to claim 2, characterized in that, It also includes removing duplicate edges shared between different units in the graph data structure.

4. The transient multiphysics order reduction and reconstruction method driven by time-series operating conditions according to claim 1, characterized in that, Based on the aforementioned graph data structure, a graph neural network autoencoder is trained to obtain the trained graph neural network autoencoder, specifically including: Define the structure of the graph neural network autoencoder and obtain the network parameters of the graph neural network autoencoder. A loss function for a graph neural network autoencoder is defined. Based on this loss function, the reconstruction error is used to determine the loss value of the graph neural network autoencoder. The expression for the loss function of the graph neural network autoencoder is as follows: ； in, The global reconstruction loss function; For local reconstruction loss function; , λ For the weighting factor; The graph data structure is input into the graph neural network autoencoder. Based on the loss value of the graph neural network autoencoder, the network parameters of the graph neural network autoencoder are iteratively trained and optimized to obtain the trained graph neural network autoencoder.

5. The transient multiphysics order reduction and reconstruction method driven by time-series operating conditions according to claim 1, characterized in that, Based on a time-series neural network, using the time-series operating condition parameter sequence obtained from the amplitude curve of the operating condition parameters as input and the low-dimensional latent variable as output, a trained time-series neural network is obtained, establishing a mapping relationship from the time-series operating condition parameters to the low-dimensional latent variable, specifically including: Based on the amplitude curve of the operating condition parameters, a time-series operating condition parameter sequence is obtained by sampling; each time-series operating condition parameter sequence corresponds to a low-dimensional latent variable. The time-series operating condition parameter sequence is input into a time-series neural network to obtain the output of the time-series neural network; Based on the output of the time-series neural network, the low-dimensional latent variables corresponding to the time-series operating condition parameter sequence, and the determined loss function of the time-series neural network, the loss value of the time-series neural network is determined. Based on the loss value of the time-series neural network, the network parameters of the time-series neural network are optimized to obtain the trained time-series neural network, and a mapping relationship from time-series operating parameters to low-dimensional latent variables is established.

6. The transient multiphysics order reduction and reconstruction method driven by time-series operating conditions according to claim 5, characterized in that, The expression for the loss function of the temporal neural network is: ； in, For latent variable prediction error; It is a 2-norm; For a point in time; For time points The corresponding low-dimensional latent variables; These are the low-dimensional latent variables obtained from the prediction.

7. The transient multiphysics order reduction and reconstruction method driven by time-series operating conditions according to claim 1, characterized in that, Based on the trained temporal neural network and the graph neural network decoder, rapid prediction and reconstruction from temporal operating parameters to the transient multiphysics field distribution are achieved, specifically including: Online acquisition of real-time time-series operating condition parameter sequences; The real-time operating condition parameter sequence is input into the trained time-series neural network to obtain the predicted low-dimensional latent variables; The predicted low-dimensional latent variables are input into the decoder to reconstruct the transient multiphysics field distribution.

8. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that the processor executes the computer program to implement the time-series condition-driven transient multiphysics order reduction reconstruction method according to 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 time-series condition-driven transient multiphysics order reduction reconstruction method 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 time-series condition-driven transient multiphysics order reduction reconstruction method as described in any one of claims 1-7.

Images