Aero-engine uncertainty modeling method based on graph neural network

Abstract

The invention relates to the field of aero-engine design, and discloses an aero-engine uncertainty modeling method based on a graph neural network, and the method comprises the steps: generating a high-dimensional sample set containing geometric uncertainty features in all parts of an aero-engine through Latin hypercube sampling; meanwhile, considering the influence of the input uncertainty characteristics of each component on component indexes, the influence between different components and the influence of each component index on whole machine indexes, constructing a graph structure data set of the aero-engine components, taking geometric parameters as node characteristics, and learning the relationship between the components through a graph attention network; a graph neural network model is trained, single component, component coupling and complete machine level uncertainty quantitative analysis and modeling are realized, and multiple component indexes and complete machine indexes are predicted at the same time; according to the method, on one hand, the prediction precision and robustness are improved, on the other hand, the multi-level dependency relationship is effectively captured through the graph neural network, the limitation of a traditional method in processing complex uncertainty is solved, and aero-engine design optimization is promoted.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of aero-engine design, and discloses an aero-engine uncertainty modeling method based on a graph neural network. BACKGROUND

[0002] There are a large number of uncertainty factors in an aero-engine, including geometric uncertainty factors such as manufacturing deviations and use deformations, and working condition uncertainty factors such as inflow boundary conditions and matching working point changes. The number of these uncertainty factors can be as high as hundreds or even thousands, which can cause the engine performance to deviate from the expected design value, present a probability distribution, and bring problems such as low manufacturing qualification rate and difficult use and maintenance. Therefore, it is very important to construct a geometric-component-performance-engine performance uncertainty correlation model and to explore the action mechanism of the uncertainty factors.

[0003] Patent CN120068699A discloses a deep learning-based performance proxy modeling scheme for an aero-engine, which focuses on data-driven prediction and uncertainty evaluation at the component level such as the turbine, and improves the model interpretability by combining CFD batch data and feature importance analysis. However, this design has the following problems and shortcomings: it only covers the geometric-turbine component level, the input is mainly a plane feature vector, it does not extend to joint modeling and prediction at the engine level, it does not explicitly use the hierarchical topology prior and influence direction of geometry-component-engine, and it is difficult to reflect the cross-level coupling and mutual influence between geometry-component-engine.

[0004] Patent CN115688316A discloses an engine reliability evaluation method, which is mainly based on hierarchical resampling and Monte Carlo sequential statistics combination of component life or performance samples to obtain the estimation and confidence interval of system-level (engine) indicators. However, it does not establish a forward mechanism / data joint modeling of "geometry / working condition-component-engine", only performs hierarchical sampling and statistical combination, and lacks explicit modeling and learning of continuous coupling quantities. The expression of the interaction between components, the directionality and the variable topology is limited, and it is difficult to capture the multi-level cross-influence relationship and its transmission effect on engine performance.

[0005] In summary, the existing technology has the following deficiencies: lack of a unified framework that directly integrates structural prior (such as component connection relationship, coupling strength and directionality) into the learning process, making it difficult to model in small sample and strong coupling scenarios; lack of an end-to-end method that can simultaneously output component-level and engine-level predictions and quantify their uncertainties, making it difficult to carry out consistent and reliable evaluation and closed-loop checking at different levels; lack of unified processing capability for heterogeneous node types and multi-task output, making it difficult to cover different component categories and multi-index prediction requirements in one model, and making it difficult to meet the multi-level cross-influence modeling of geometry-component-engine. SUMMARY

[0006] The purpose of the present application is to provide an aero-engine uncertainty modeling method based on a graph neural network, which improves the prediction accuracy and robustness of the aero-engine simulation model, solves the limitations of traditional methods in dealing with complex uncertainty, and promotes the optimization of aero-engine design.

[0007] In order to achieve the above technical effects, the technical scheme adopted by the present application is: An aero-engine uncertainty modeling method based on a graph neural network, comprising: S1, generating a sample set based on geometric uncertainty factors according to the geometric uncertainty factors in each component of the engine by using Monte Carlo sampling; obtaining the output indicators of each component and the overall indicators corresponding to each sample data in the sample set by using simulation or measurement means; the geometric uncertainty factors include the chord length of the blades of the compressor or turbine, the blade tip clearance, and the flow passage area of the corresponding component and the turbine cooling hole size; S2, constructing a graph structure dataset using the structure of the aero-engine and the sample set, regarding the geometric parameters of each component as a node feature vector, unifying the feature dimension by using a linear projection method, and constructing a graph representation containing multiple nodes, wherein each node corresponds to an engine component; S3, defining edge connections according to the physical coupling and functional interaction relationship between the aero-engine components, using full connection or sparse connection based on topological structure, forming an aero-engine graph structure to capture the dynamic interaction between multiple components; S4, designing a graph neural network model based on the multi-level association relationship in the aero-engine graph structure, forming a graph neural network model containing a projection layer, multiple GATConv layers and an activation function; S5, designing a component prediction head for each component node in the graph neural network model to output the predicted value of the component indicator, and designing a global prediction head to aggregate the predicted values of all component indicators to output the predicted value of the overall indicator; S6, taking all or part of the sample set as a training set, constructing an objective function for the training process according to the first weighted sum of the mask MSE loss of all component output indicators and the second weighted sum of the mask MSE loss of all overall indicators, and completing the training of the graph neural network model by using adaptive optimization and early stopping strategy.

[0008] Further, in step S4, the attention coefficient of the GATConv layer is determined according to is obtained by analysis, wherein is the attention coefficient of node and , the numerator is the attention weight of node to , and For activation function, Assign a scoring vector to the attention. Indicates transpose. Let be the linear projection matrix of the linear projection method. For components Node feature vectors For components The node feature vectors, This indicates that features with the same dimension will be used. and Vectors are concatenated into paired features. For nodes The set of neighboring nodes, For neighboring nodes The corresponding component node feature vector, This indicates that features with the same dimension will be used. and Vectors are concatenated into paired features, denominator Indicates to The attention weights of all neighboring nodes are summed.

[0009] Furthermore, after completing the training of the graph neural network model, the attention coefficients of any two nodes in the graph neural network model are extracted. Based on the order of the attention coefficients of each component and other components from large to small, the influence of the corresponding component on other components is sorted. The other components after the first q are identified as non-critical components. In the formed aero-engine graph structure, all connections between non-critical components and their corresponding components are removed. Steps S4-S6 are repeated to complete the graph neural network model update training.

[0010] Furthermore, in step S6, the component output indicator mask loss... Overall performance mask loss ,in The number of samples in the training set. For graph neural networks in the 1st... The first sample Predicted values ​​for individual component indicators For the first The first sample Test or simulation values ​​of individual component indicators; For graph neural networks in the 1st... The first sample Predicted values ​​for individual machine performance indicators For the first The first sample Test or simulated values ​​of the overall machine performance indicators.

[0011] Further, in step S6, the objective function ,in a number of component indicators, a number of whole-machine indicators, a loss weight of the first component indicator, a loss weight of the first whole-machine indicator.

[0012] Further, after completing the training of the graph neural network model, the GNNExplainer method is used to identify the edge connection and node feature that contributes most to the prediction of the component indicator and the whole-machine indicator, so as to control the dispersion of the edge connection and the node feature with the greatest contribution in the engine structure design and the part processing and manufacturing process; or through contribution degree sorting, the edge connection and the node feature with a contribution degree ranking less than a preset contribution degree threshold are deleted in the formed aero-engine graph structure, and then the graph neural network model is retrained.

[0013] Further, the method for identifying the edge connection and the node feature that contributes most to the prediction of the component indicator and the whole-machine indicator by using the GNNExplainer method comprises: The GNNExplainer maximizes the mutual information of the prediction by optimizing the mask The edge connection and the node feature corresponding to the maximum mutual information are identified as the edge connection and the node feature that contribute most to the prediction of the component indicator and the whole-machine indicator, wherein is the mask, is the subgraph obtained after the graph neural network is superimposed with the mask ; is the prediction of the component indicator or the whole-machine indicator by the graph neural network, is the mutual information, is the subgraph size, is the weight coefficient of the subgraph size.

[0014] Further, before step S2, the sample data in the sample set and the corresponding output indicators and whole-machine indicators are respectively subjected to missing value filling and standardization processing to obtain a normalized sample set and a normalized indicator data set; in step S2, the graph structure data set is constructed by using the aero-engine structure and the normalized sample set, and in step S6, all or part of the normalized sample set is used as a training set.

[0015] Compared with the prior art, the present application has the following beneficial effects: 1. The present application generates a high-dimensional sample set containing geometric uncertainty characteristics in each component of an aero-engine by Latin hypercube sampling, considers the influence of input uncertainty characteristics of each component on component indicators, the influence between different components and each component indicator on the overall machine indicator, constructs a graph structure data set of aero-engine components, takes geometric parameters as node features, and learns the relationship between components through a graph attention network; trains a graph neural network model to realize single-component, component coupling and overall machine level uncertainty quantification analysis and modeling, and can simultaneously predict multiple component indicators and overall machine indicators.

[0016] 2. On the one hand, the present application improves the prediction accuracy and robustness, and on the other hand, effectively captures multi-level dependency relationships through a graph neural network, solves the limitations of traditional methods in dealing with complex uncertainty, and promotes aero-engine design optimization. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 The flowchart of the aero-engine uncertainty modeling method in Example 1; Figure 2 The flowchart of the aero-engine uncertainty modeling method in Example 2. DETAILED DESCRIPTION

[0018] The present application will be further described in conjunction with the embodiments and the accompanying drawings. However, it should not be understood that the scope of the above-mentioned subject matter of the present application is limited to the following embodiments, and any technology realized based on the content of the present application falls within the scope of the present application.

[0019] Example 1 Referring to Figure 1 A graph neural network-based aero-engine uncertainty modeling method, comprising: S1. According to the geometric uncertainty factors in each component of the engine, a sample set based on the geometric uncertainty factors is generated by Monte Carlo sampling; the output indicators of each component and the overall machine indicators corresponding to each sample data in the sample set are obtained by simulation or measurement; the geometric uncertainty factors include the blade chord length, the blade tip clearance of the compressor or turbine, and the flow passage area and turbine cooling hole size of the corresponding component; S2. A graph structure data set is constructed using the structure of the aero-engine and the sample set, the geometric parameters of each component are regarded as node feature vectors, the feature dimension is unified by a linear projection method, and a graph representation containing multiple nodes is constructed, wherein each node corresponds to an engine component; S3. According to the physical coupling and functional interaction relationship between the aero-engine components, the edge connection is defined, the full connection or sparse connection based on the topological structure is adopted, and the aero-engine graph structure is formed to capture the dynamic interaction between multiple components; S4, based on the multi-level correlation relationship in the aero-engine graph structure, a graph neural network model is designed to form a graph neural network model including a projection layer, a plurality of GATConv layers and an activation function; S5, for each component node in the graph neural network model, a component prediction head is designed to output component index prediction value, and a global prediction head is designed to output the whole machine index prediction value by aggregating the component index prediction values of all components; S6, using all or part of the sample set as a training set, a target function of the training process is constructed according to the first weighted sum of all component output index mask MSE loss and the second weighted sum of all whole machine index mask MSE loss, and the graph neural network model is trained by using adaptive optimization and early stopping strategy.

[0020] In the embodiment, a high-dimensional sample set containing geometric uncertainty features of each component of the aero-engine is generated by Latin hypercube sampling, considering the influence of each component input uncertainty feature on the component index, the mutual influence between different components, and the influence of each component index on the whole machine index, a graph structure data set of the aero-engine component is constructed, the geometric parameters are taken as node features, and the relationship between components is learned through graph attention network; the graph neural network model is trained to realize single component, component coupling and whole machine level uncertainty quantification analysis and modeling, and multiple component indexes and whole machine indexes are predicted; the aero-engine uncertainty modeling method of the embodiment improves the prediction accuracy and robustness on the one hand, and effectively captures the multi-level dependency relationship through the graph neural network on the other hand, solves the limitations of traditional methods in dealing with complex uncertainty, and promotes the aero-engine design optimization.

[0021] Embodiment 2 Referring to Figure 2 A graph neural network-based aero-engine uncertainty modeling method, comprising: Step one, according to the geometric uncertainty factors in each component of the engine, a sample set based on geometric uncertainty factors is generated by Monte Carlo sampling; the output indexes of each sample data in the sample set and the whole machine indexes are obtained by simulation or measurement; the geometric uncertainty factors include blade chord length, blade tip clearance of the compressor or turbine, and flow passage area, turbine cooling hole size of the corresponding component; In this embodiment, two components, air system and turbine, are selected from an aero-engine to carry out research. The geometric uncertainty factors in the air system are sorted out, including 26 factors such as air system blade chord length, blade tip clearance, labyrinth clearance, cooling hole diameter, etc. The geometric uncertainty factors in the turbine are sorted out, including 53 factors such as turbine blade chord length of each row, blade tip clearance, blade geometry, etc. Latin hypercube sampling is used to generate a sample set, and the number of samples is 1000. For each sample, the output indicators of each component (including 7 indicators such as air system flow distribution, 6 indicators such as turbine expansion ratio, flow, efficiency, etc.) and the whole machine indicators (including 4 indicators such as thrust, specific fuel consumption, turbine inlet temperature and stability margin, etc.) are obtained by simulation.

[0022] Step two, the sample data in the sample set and the corresponding output indicators and whole machine indicators are respectively filled with missing values and standardized to obtain a normalized sample set and a normalized indicator data set. The graph structure data set is constructed by using the structure of the aero-engine and the normalized sample set. The geometric parameters of each component are regarded as node feature vectors. The feature dimensions are unified by linear projection method. A graph representation containing multiple nodes is constructed, wherein each node corresponds to an engine component. Specifically, it includes: S2.1: Standardize the sample data according to the mean and standard deviation of each variable; S2.2: According to the structure of the aero-engine and the data characteristics, a graph structure data set is constructed. The geometric parameters of each component are regarded as node feature vectors. In order to deal with the heterogeneity of different component feature dimensions, in this embodiment, the 26 dimensions of the air system are zero-padded to 53 dimensions, aligned with the 53 dimensions of the turbine, and a double-node graph (node 0 corresponds to the air system, and node 1 corresponds to the turbine) is constructed. Each node feature is a [2, 53] matrix; S2.3: Define edge connection according to the physical coupling and functional interaction relationship between aero-engine components. In this embodiment, the edge index is defined as [[0], [1]], indicating a one-way connection from the air system to the turbine, reflecting the influence of air system cold gas flow change on turbine aerodynamic performance; S2.4: Pack the graph data using the dataset packaging class, and package it into the SysGraphDataset class, including node-level labels (node 0, 7 dimensions) and (node 1, 6 dimensions), and global labels (4 dimensions), which are convenient for subsequent model input and batch processing.

[0023] Step three, define edge connection according to the physical coupling and functional interaction relationship between aero-engine components, adopt full connection or sparse connection based on topological structure, form aero-engine graph structure to capture dynamic interaction between multiple components; Step 4: Based on the multi-level relationships in the graph structure of the aero-engine, design a graph neural network model to form a graph neural network model containing a projection layer, two GATConv layers and a ReLU activation function; In this embodiment, the projection layers process the air system and turbine features respectively; the attention coefficient of the GATConv layer is based on... Analysis yielded, among which For nodes and Attention coefficient, molecule Pointer node right Attention weights For activation function, Assign a scoring vector to the attention. Indicates transpose. Let be the linear projection matrix of the linear projection method. For components Node feature vectors For components The node feature vectors, This indicates that features with the same dimension will be used. and Vectors are concatenated into paired features. For nodes The set of neighboring nodes, For neighboring nodes The corresponding component node feature vector, This indicates that features with the same dimension will be used. and Vectors are concatenated into paired features, denominator Indicates to The attention weights of all neighboring nodes are summed.

[0024] Step 5: Design a component prediction head for each component node in the graph neural network model to output the component index prediction value, and design a global prediction head to summarize the component index prediction values ​​of all components to output the overall machine index prediction value. In this embodiment, a dedicated prediction head is designed for each component node to output component indicators. The air system prediction head outputs 7-dimensional Y1, the turbine prediction head outputs 6-dimensional Y2, and the global prediction head splices Y1 and Y2 (13 dimensions) and maps them to the 4-dimensional whole machine indicator Z.

[0025] Step 6: Using all or part of the normalized sample set as the training set, construct the objective function for the training process based on the first weighted sum of the MSE loss of all component output index masks and the second weighted sum of the MSE loss of all whole machine index masks, and complete the training of the graph neural network model by adopting adaptive optimization and early stopping strategies. This embodiment uses masked MSE loss in the training of the graph neural network model. , ,in For component output specification mask loss, For overall system performance mask loss, The number of samples in the training set. For graph neural networks in the 1st... The first sample Predicted values ​​for individual component indicators For the first The first sample Test or simulation values ​​of individual component indicators; For graph neural networks in the 1st... The first sample Predicted values ​​for individual machine performance indicators For the first The first sample Test or simulation values ​​of the overall machine performance; The objective function ,in The number of samples in the training set. The number of component indicators, The number of indicators for the whole machine. For the first The loss weight of each component indicator, For the first The loss weight of each overall system indicator For graph neural networks in the 1st... The first sample Predicted values ​​for individual component indicators For the first The first sample Test or simulation values ​​of individual component indicators; For graph neural networks in the 1st... The first sample Predicted values ​​for individual machine performance indicators For the first The first sample Test or simulation values ​​of the overall machine performance; Using the Adam optimizer (lr=1e) -3 The model was trained using an early-stop, patience-based training strategy with 200 epochs and a 50% epoch limit. The overall loss and component losses for both the training and validation sets were recorded to complete the model training. The trained model was then evaluated on the validation set. The MAE and RMSE indices are used to verify the accuracy of the model's predictions of component and overall machine indices.

[0026] Step seven, after the training of the graph neural network model is completed, the attention coefficients of any two nodes of the graph neural network model are extracted, the influence of the corresponding components on other components is sorted according to the order of the attention coefficients of each component and other components from large to small, the other components after the first q are identified as non-key components, and the connections between all non-key components and corresponding components in the formed aero-engine graph structure are cancelled, and steps four to six are repeated to complete the update training of the graph neural network model.

[0027] Step eight, after the training of the graph neural network model is completed, the GNNExplainer method is used to identify the edge connection and node feature that contribute most to the prediction of the component index and the whole machine index, wherein the GNNExplainer maximizes the mutual information of the prediction by optimizing the mask The edge connection and node feature corresponding to the maximum mutual information are identified as the edge connection and node feature that contribute most to the prediction of the component index and the whole machine index. is the mask. is the graph neural network superimposed with the mask The subgraph obtained after the mask is removed. is the prediction of the component index or the whole machine index by the graph neural network. is the mutual information, which measures the importance of the subgraph to the prediction. is the size of the subgraph, which is used as a sparsity constraint. is the weight coefficient of the size of the subgraph.

[0028] In this embodiment, by identifying the contribution rates of a total of 79 geometric features in the air system and the turbine to the whole machine index, 10 key geometric uncertainty factors that have important influence on the thrust and specific fuel consumption are identified, including certain labyrinth clearance in the air system, turbine blade trailing edge angle, etc.

[0029] The embodiment finds a compact subgraph by the GNNExplainer, which contains the edge connection and node feature with the largest contribution to the component index and the whole machine index prediction. On the one hand, the dispersion of the edge connection and node feature with the largest contribution is controlled in the engine structure design and the part processing and manufacturing process, and the dispersion of the whole machine index is reduced; or the edge connection and node feature with the contribution less than the preset contribution threshold are deleted in the formed aero-engine graph structure through the contribution ranking, and then the graph neural network model is retrained to realize the lightweight of the graph neural network, thereby reducing the simulation or measured sample amount required for training the high-precision graph neural network, or realizing the construction of a higher-precision graph neural network model with the same sample amount; on the other hand, by analyzing the optimized subgraph, the uncertainty influence chain of geometric feature-component index-whole machine index is revealed, the influence of geometric uncertainty on the whole machine robustness through component interaction is quantified, the parameter sensitivity analysis can be carried out based on the interpretable machine learning algorithm, the key influencing factors and their influence law are identified, the uncertainty optimization decision is supported, the dispersion of the whole machine index is reduced, and the guidance for engine design improvement is provided.

[0030] Step nine, generate an explanation report to reveal the basis of model decision, based on 10 key uncertainty factors and their influence law on performance, accurately tighten the tolerance, control the dispersion of the edge connection and node feature with the largest contribution in the engine structure design and the part processing and manufacturing process, and realize the improvement of the dispersion of the thrust and specific fuel consumption and the like within the cost acceptable in engineering.

[0031] The above is only a preferred embodiment of the present application, and is not used to limit the present application, any modification, equivalent replacement and improvement made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for modeling uncertainties in aero-engines based on graph neural networks, characterized in that, include: S1. Based on the geometric uncertainty factors in each component of the engine, Monte Carlo sampling is used to generate a sample set based on the geometric uncertainty factors; The output indicators of each component and the overall machine indicators corresponding to each sample data in the sample set are obtained by simulation or actual measurement. The geometric uncertainty factors include the blade chord length and tip clearance of the compressor or turbine, as well as the flow channel area and turbine cooling hole size of the corresponding components. S2. Construct a graph structure dataset using the aero-engine structure and the sample set, treat the geometric parameters of each component as node feature vectors, unify the feature dimensions through linear projection, and construct a graph representation containing multiple nodes, where each node corresponds to an engine component. S3. Define edge connections based on the physical coupling and functional interaction relationships between aero-engine components, and use fully connected or topology-based sparse connections to form an aero-engine graph structure to capture the dynamic interaction between multiple components. S4. Based on the multi-level relationships in the graph structure of aero-engines, a graph neural network model is designed to form a graph neural network model containing a projection layer, multiple GATConv layers and activation functions. S5. Design a component prediction head for each component node in the graph neural network model to output the component index prediction value, and design a global prediction head to summarize the component index prediction values ​​of all components to output the overall machine index prediction value. S6. Using all or part of the sample set as the training set, construct the objective function for the training process based on the first weighted sum of the MSE loss of all component output index masks and the second weighted sum of the MSE loss of all whole machine index masks, and complete the training of the graph neural network model by adopting adaptive optimization and early stopping strategies.

2. The uncertainty modeling method for aero-engines according to claim 1, characterized in that, In step S4, the attention coefficients of the GATConv layer are determined according to... Analysis yielded, among which For nodes and Attention coefficient, molecule Pointer node right Attention weights For activation function, Assign a scoring vector to the attention. Indicates transpose. Let be the linear projection matrix of the linear projection method. For components Node feature vectors For components The node feature vectors, This indicates that features with the same dimension will be used. and Vectors are concatenated into paired features. For nodes The set of neighboring nodes, For neighboring nodes The corresponding component node feature vector, This indicates that features with the same dimension will be used. and Vectors are concatenated into paired features, denominator Indicates to The attention weights of all neighboring nodes are summed.

3. The aero-engine uncertainty modeling method according to claim 2, characterized in that, After completing the training of the graph neural network model, the attention coefficients of any two nodes in the graph neural network model are extracted. Based on the order of the attention coefficients of each component and other components from large to small, the influence of the corresponding component on other components is sorted. The other components after the first q are identified as non-critical components. In the formed aero-engine graph structure, all connections between non-critical components and their corresponding components are removed. Steps S4-S6 are repeated to complete the graph neural network model update training.

4. The uncertainty modeling method for aero-engines according to claim 1, characterized in that, In step S6, the component output index mask loss Overall performance mask loss ,in The number of samples in the training set. For graph neural networks in the 1st... The first sample Predicted values ​​for individual component indicators For the first The first sample Test or simulation values ​​of individual component indicators; For graph neural networks in the 1st... The first sample Predicted values ​​for individual machine performance indicators For the first The first sample Test or simulated values ​​of the overall machine performance indicators.

5. The aero-engine uncertainty modeling method according to claim 4, characterized in that, In step S6, the objective function ,in The number of component indicators, The number of indicators for the whole machine. For the first The loss weight of each component indicator, For the first The loss weight of each overall system indicator.

6. The aero-engine uncertainty modeling method according to claim 4, characterized in that, After training the graph neural network model, the GNNExplainer method is used to identify the edge connections and node features that contribute the most to the prediction of component indicators and overall engine indicators, so as to control the dispersion of the edge connections and node features that contribute the most during engine structure design and component manufacturing; or the edge connections and node features whose contribution ranking is less than the preset contribution threshold are deleted from the formed aero-engine graph structure by ranking the contribution, and then the graph neural network model is retrained.

7. The uncertainty modeling method for aero-engines according to claim 6, characterized in that, Methods for identifying edge connections and node features that contribute most to the prediction of component and overall system performance using the GNNExplainer approach include: GNNExplainer maximizes the mutual information of predictions by optimizing the mask. The edge connections and node features corresponding to the maximum mutual information values ​​are identified as the edge connections and node features that contribute the most to the prediction of component indicators and overall machine indicators. For the mask, For graph neural network overlay mask The resulting subgraph; For graph neural networks to predict component or overall machine performance; For mutual information, For the size of the subgraph, This is the weighting coefficient for the size of the subgraph.

8. The method for modeling uncertainties in aero-engines according to any one of claims 1-7, characterized in that, Before constructing the graph structure dataset using the aero-engine structure and the sample set in step S2, missing value imputation and standardization are performed on the sample data and corresponding output indicators and overall indicators in the sample set to obtain a normalized sample set and a normalized indicator dataset. In step S2, the graph structure dataset is constructed using the aero-engine structure and the normalized sample set. In step S6, all or part of the normalized sample set is used as the training set.

Images