A simulation-measured data fusion driven fault diagnosis method for aerospace structural components

Abstract

The application discloses a simulation-measured data fusion driven spaceflight structural part fault diagnosis method, and mainly solves the technical problems that the prior art relies on a large amount of measured data, but the simulation test of the spaceflight structural part is high in cost, and it is difficult to obtain a large amount of measured data; the fusion of simulation data and measured data cannot be effectively solved, low-cost simulation data cannot be fully utilized, and higher accuracy and reliability requirements cannot be met. The application effectively fuses and utilizes simulation data and measured data by constructing multi-fidelity graph structure data and designing a double-head graph convolutional neural network based on residual fusion, fully excavates the associated fault characterization features of the simulation data and the measured data, and significantly improves the fault recognition accuracy of the spaceflight structural part in the environment with scarce measured data.

Description

Technical Field

[0001] This invention relates to a fault diagnosis method for aerospace structural components, specifically a fault diagnosis method for aerospace structural components driven by simulation-measurement data fusion. Background Technology

[0002] Launch vehicles comprise multiple subsystems, among which structural components serve as the foundation connecting all instruments and equipment within each system. They bear various loads during attitude control, thrust adjustment, and trajectory correction, playing a crucial role in supporting, bearing, and protecting the various components within the system. Aerospace structural components typically operate under extreme conditions such as high temperatures and strong vibrations, which significantly increase the risk of failure. If structural component failures are not identified and repaired in a timely manner, they will directly lead to the failure of related systems, causing severe economic losses and even casualties.

[0003] Existing fault diagnosis technologies mainly include traditional methods based on signal analysis and data-driven methods. Traditional signal analysis methods rely on expert experience but suffer from significant drawbacks such as high misdiagnosis rates and low efficiency. In contrast, data-driven diagnostic methods do not rely on expert experience and can automatically extract features to achieve fault diagnosis, making them relatively more advanced.

[0004] However, in practical engineering applications, this technology faces two main challenges: First, existing technologies rely on a large amount of measured data, but simulation testing of aerospace structural components is costly and obtaining large amounts of measured data is difficult; second, existing methods have failed to effectively solve the problem of integrating simulation data with measured data, making it difficult to fully utilize low-cost simulation data. These limitations severely restrict the application effectiveness of existing data-driven intelligent diagnostic methods in practical engineering and cannot meet higher accuracy and reliability requirements. Therefore, a fault diagnosis method driven by the fusion of a large amount of simulation data and a small amount of measured data is needed to achieve efficient and reliable fault diagnosis of aerospace structural components. Summary of the Invention

[0005] The purpose of this invention is to address the technical problems of existing fault diagnosis technologies that rely on a large amount of measured data, but the cost of simulation testing of aerospace structural components is high, it is difficult to obtain a large amount of measured data, and the problem of fusion between simulation data and measured data is not effectively solved, making it difficult to make full use of low-cost simulation data and meet higher accuracy and reliability requirements. The invention provides a fault diagnosis method for aerospace structural components driven by simulation-measured data fusion.

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

[0007] A simulation-measurement data fusion-driven method for fault diagnosis of aerospace structural components, characterized by the following steps:

[0008] Step 1: Construct a finite element model of the aerospace structural component to be diagnosed, set the required working conditions and boundary conditions, and use simulation calculations to obtain low-fidelity simulation data;

[0009] Step 2: Conduct tests on the physical sample of the aerospace structural component to be diagnosed to obtain high-fidelity test data;

[0010] Step 3: Perform unified format conversion and standardization preprocessing on the acquired low-fidelity simulation data and high-fidelity real-measured data, and then perform time-frequency conversion on the standardized preprocessed low-fidelity simulation data and high-fidelity real-measured data to obtain the corresponding time-frequency domain low-fidelity simulation data and high-fidelity real-measured data.

[0011] Step 4: Based on the K-nearest neighbor strategy, establish the similarity relationship between low-fidelity simulation data and high-fidelity test data, and construct them into multi-fidelity graph structure data with node and edge features to achieve the fusion of multi-fidelity data; divide the constructed multi-fidelity graph structure data into training set and test set;

[0012] Step 5: Establish a dual-headed graph convolutional neural network based on residual fusion, and extract potential structural fault features from multi-fidelity graph structural data through the dual-headed graph convolutional neural network;

[0013] Step 6: Establish a lightweight feature classifier to perform nonlinear transformation and boundary discrimination on potential structural fault features, thereby classifying the fault categories of aerospace structural components;

[0014] Step 7: Combine a dual-head graph convolutional neural network based on residual fusion and a lightweight feature classifier based on convolutional neural network to establish a fault diagnosis model for aerospace structural components driven by simulation-measurement data fusion; complete the training of the fault diagnosis model using multi-fidelity graph structural data in the training set;

[0015] Step 8: Input the multi-fidelity graph structure data from the test set into the trained fault diagnosis model to perform fault diagnosis.

[0016] Furthermore, in step 3, the standardization preprocessing is performed using Z-score standardization, calculated as follows:

[0017] , , ;

[0018] In the formula: For low-fidelity simulation data and high-fidelity real-world test data, the first One value, ; The average of low-fidelity simulation data and high-fidelity real-world measurement data. The standard deviations of low-fidelity simulation data and high-fidelity measured data are given. For the standardized low-fidelity simulation data and high-fidelity real-world measurement data, the first... Values.

[0019] Furthermore, step 4 specifically involves:

[0020] Step 4.1: Determine the nodes

[0021] One data point is taken from the high-fidelity measured data as a high-fidelity sample, based on Euclidean distance. The K-nearest neighbor strategy obtains the k nearest neighbor samples of the high-fidelity sample from the low-fidelity simulation data, thereby determining the k nodes of the high-fidelity sample; Euclidean distance. The calculation formula is:

[0022] ;

[0023] In the formula: To provide feature vectors for high-fidelity real-world test data, The feature vector of low-fidelity simulation data; For feature dimensions; The first in high-fidelity real measurement data One value; The first in low-fidelity simulation data One value;

[0024] Step 4.2: Using the following formula, based on k nodes of the high-fidelity sample... Determine the edge weights between nodes :

[0025] ;

[0026] In the formula: To ensure the accuracy of the test data, This is low-fidelity simulation data; for k neighbors, For k nearest neighbors and The average Euclidean distance between them;

[0027] Step 4.3: Following the same method as steps 4.1 and 4.2, sequentially obtain the k nodes corresponding to all high-fidelity samples. and edge weights Through the k nodes corresponding to all high-fidelity samples and edge weights Establish similarity relationships between low-fidelity simulation data and high-fidelity measured data, and construct them into multi-fidelity graph structure data with node and edge features;

[0028] Step 4.4: Divide the obtained multi-fidelity graph structure data into training set and test set according to the set ratio.

[0029] Furthermore, in step 5, the dual-head graph convolutional neural network includes two input branches and a first adder;

[0030] Both input branches include a first Mean module, and a first ChebConv module, a first EdgePool module, a second ChebConv module, a second EdgePool module, a second Mean module, and a second adder connected in sequence;

[0031] The input terminal of the first Mean module is connected to the second output terminal of the first EdgePool module, and the output terminal is connected to the second input terminal of the second adder;

[0032] The output of the second adder in the two input branches is connected to the two inputs of the first adder, and the output of the first adder is connected to the input of the lightweight feature classifier.

[0033] Furthermore, in step 6, the lightweight feature classifier includes two fully connected layers and a Dropout regularization layer connected between the two fully connected layers;

[0034] The input of the first fully connected layer is connected to the output of the first adder;

[0035] The two fully connected layers are used to perform nonlinear transformations on the input structural fault features to determine the category boundaries and achieve efficient classification of different fault types.

[0036] The Dropout regularization layer is used to ignore the connections of some neurons in the fully connected layer during training, preventing the classifier from overfitting.

[0037] Furthermore, in step 7, the following cross-entropy loss function is used during the training process. Perform model optimization:

[0038] ;

[0039] In the formula: Number of categories; For the first The true label of the class; For the fault diagnosis model to the first The predicted probability of a class.

[0040] Furthermore, step 2 specifically involves conducting tests on the physical samples of the aerospace structural components to be diagnosed under various operating conditions to obtain high-fidelity test data under these conditions.

[0041] Compared with the prior art, the present invention has the following beneficial technical effects:

[0042] This breakthrough overcomes the dependence of existing aerospace structural component data-driven fault diagnosis technologies on large-scale measured data. By constructing multi-fidelity graph structural data and designing a dual-head graph convolutional neural network based on residual fusion and a lightweight feature classifier based on convolutional neural networks, it fully explores the associated fault characterization features of simulation and measured data using simulation and measured data, significantly improving the fault identification accuracy of aerospace structural components in environments where measured data is scarce. Attached Figure Description

[0043] Figure 1 This is a flowchart of the intelligent fault diagnosis method for aerospace structural components based on digital-analog linkage of the present invention.

[0044] Figure 2 This is a schematic diagram of the dual-headed graph convolutional neural network and lightweight feature classifier in an embodiment of the aerospace structural component fault diagnosis method driven by simulation-measurement data fusion of the present invention.

[0045] Figure 3 This is the finite element model of the compartment structure used in the embodiments of the present invention.

[0046] Figure 4 This refers to the predictive accuracy of the fault diagnosis test results in this embodiment of the invention. Detailed Implementation

[0047] To make the objectives, advantages, and features of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Those skilled in the art should understand that these embodiments are merely used to explain the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0048] like Figure 1 As shown in the figure, this embodiment presents a simulation-measurement data fusion-driven method for fault diagnosis of aerospace structural components, which specifically includes the following steps:

[0049] Step 1: Based on the relevant design documents, construct a finite element model of the cabin structure, clarify the required operating conditions and boundary conditions, and obtain the required low-fidelity simulation data through finite element calculations. Specifically:

[0050] like Figure 3As shown, a finite element model of the compartment structure was established based on the structural design documents and combined with design parameters such as material parameters and boundary conditions. The constructed finite element model can simulate the dynamic response of the compartment structure under different excitations, obtain the required low-fidelity simulation data, and provide large-scale simulation data for the subsequent training of the intelligent diagnostic model. During the construction of the finite element model, corresponding nodes were set at the actual locations of the accelerometer sensors in the experiment to ensure the consistency of the spatial distribution between the measured data and the simulation data.

[0051] Step 2: Conduct mechanical tests on physical samples of aerospace structural components to obtain sparse, high-fidelity test data. Specifically:

[0052] like Figure 4 As shown, vibration tests were conducted on the cabin structure, and data were collected under five different operating conditions: (1) State 1: normal state; (2) State 2: additional small counterweight; (3) State 3: additional large counterweight; (4) State 4: bolt loosening; (5) State 5: small cylinder falling off.

[0053] Step 3: Perform unified format conversion and standardization preprocessing on the acquired low-fidelity virtual space simulation data and high-fidelity physical space measurement data. Then, perform time-frequency conversion on the processed low-fidelity simulation data and high-fidelity physical space measurement data to obtain corresponding time-frequency domain low-fidelity simulation data and high-fidelity physical space measurement data. Specifically:

[0054] The data obtained in steps 1 and 2 are standardized using Z-score, and the calculation formula is as follows:

[0055] , , ;

[0056] In the formula: For low-fidelity simulation data and high-fidelity real-world test data, the first One value, ; The average of low-fidelity simulation data and high-fidelity real-world measurement data. The standard deviations of low-fidelity simulation data and high-fidelity measured data are given. For the standardized low-fidelity simulation data and high-fidelity real-world measurement data, the first... Each value represents the position of the original value in a distribution with a mean of 0 and a standard deviation of 1.

[0057] Step 4: Based on the K-nearest neighbor strategy, establish the similarity relationship between low-fidelity simulation data and high-fidelity physical space measured data. Construct the standardized low-fidelity simulation data and high-fidelity physical space measured data from Step 3 into a multi-fidelity graph structure with node and edge features, achieving multi-fidelity data fusion. Specifically:

[0058] Step 4.1: Determine the nodes

[0059] Take a high-fidelity sample from the high-fidelity test data, and obtain the k nearest neighbor samples of the high-fidelity sample from the low-fidelity simulation dataset based on the K nearest neighbor strategy.

[0060] The K-nearest neighbor strategy uses Euclidean distance, and its calculation formula is as follows:

[0061] ;

[0062] In the formula: To provide feature vectors for high-fidelity real-world test data, The feature vector of low-fidelity simulation data; For feature dimensions; The first in high-fidelity real measurement data One value; The first in low-fidelity simulation data Values.

[0063] Step 4.2: Determine the edge weights between nodes.

[0064] The formula for calculating the edge weights of nodes in multi-fidelity graph structure data is as follows:

[0065] ;

[0066] In the formula: To ensure the accuracy of the test data, This is low-fidelity simulation data; for k neighbors, For k nearest neighbors and The average Euclidean distance between them.

[0067] Step 4.3: Following the same method as steps 4.1 and 4.2, sequentially obtain the k nodes corresponding to all high-fidelity samples. and edge weights Through the k nodes corresponding to all high-fidelity samples and edge weights Establish similarity relationships between low-fidelity simulation data and high-fidelity measured data, and construct them into multi-fidelity graph structure data with node and edge features to achieve the fusion of multi-fidelity data.

[0068] Step 5, see Figure 2A dual-headed graph convolutional neural network based on residual fusion is established to perform feature propagation and aggregation on multi-fidelity graph structure data, extracting potential structural fault features from the multi-fidelity graph structure data, and mitigating information loss through residual connections. Specifically:

[0069] A dual-head graph convolutional neural network is applied using a multi-layer architecture based on a residual fusion mechanism. The network input employs a dual-head mode with joint time-frequency domain input to enhance the richness of information in multi-fidelity graph structure data samples. In the feature extraction stage, the time-frequency domain multi-fidelity graph structure data samples use a separate graph neural network channel. Convolutional operations on the graph are implemented through the ChebConv normalization module, while the EdgePool module is introduced to achieve graph structure compression. By using multi-fidelity graph structure data as input, the designed network performs message passing and feature aggregation between graph nodes and edges to uncover potential fault features.

[0070] Specifically, the aforementioned dual-head graph convolutional neural network includes two input branches and a first adder; each of the two input branches includes a first Mean module, and a first ChebConv module, a first EdgePool module, a second ChebConv module, a second EdgePool module, a second Mean module, and a second adder connected in sequence; the input end of the first Mean module is connected to the second output end of the first EdgePool module, and the output end is connected to the second output end of the second adder; the output end of the second adder in the input branches is respectively connected to the two input ends of the first adder, and the output end of the first adder is connected to the input end of the lightweight feature classifier.

[0071] Step 6: Establish a lightweight feature classifier based on a convolutional neural network to perform nonlinear transformation and boundary discrimination on the potential fault features obtained in Step 5, thereby achieving efficient classification of fault categories. Specifically:

[0072] A lightweight convolutional neural network classifier is used to identify and classify the potential fault features extracted in step 5. The designed classifier has a simple and lightweight structure, consisting of only two fully connected layers and a Dropout regularization layer connected between the two fully connected layers. The Dropout layer is used to ignore some neuron connections during training to prevent the classifier from overfitting.

[0073] Step 7: Combining the dual-headed graph convolutional neural network based on residual fusion established in Step 5 and the lightweight feature classifier based on convolutional neural networks established in Step 6, a fault diagnosis model is constructed; model training is completed using the constructed multi-fidelity graph structure data. Specifically:

[0074] The fault diagnosis model takes multi-fidelity graph structure data as input and corresponding fault category labels as output. For each of the five structural states (one normal state and four damaged states), 20 sets of data samples are selected, for a total of 100 sets of data samples, to train the model. The cross-entropy loss function is used for model optimization during training, and its mathematical expression is:

[0075] ;

[0076] In the formula: Number of categories; For the first The true label of the class; For the fault diagnosis model to the first The predicted probability of the class. Furthermore, the initial learning rate is set to 0.01, and the optimizer is set to Adam.

[0077] Step 8: Use the trained fault diagnosis model to diagnose faults in the tested structural components. Simulate the tested structural components' measured data using untrained measured data. Select 100 sets of measured data for each of the five structural states, for a total of 500 sets of measured data. Calculate the corresponding diagrammatic structural data samples and input them into the fault diagnosis model for fault diagnosis. Test the fault identification accuracy of this invention. The test results are as follows: Figure 4 As shown, the prediction accuracies for states 1 to 5 are 97%, 96%, 94%, 97%, and 99%, respectively, and the average prediction accuracy for the five states is 96.6%. The test results verify the prediction accuracy of the present invention.

[0078] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the present invention.

Claims

1. A method for fault diagnosis of aerospace structural components driven by simulation-measurement data fusion, characterized in that, Includes the following steps: Step 1: Construct a finite element model of the aerospace structural component to be diagnosed, set the required working conditions and boundary conditions, and use simulation calculations to obtain low-fidelity simulation data; Step 2: Conduct tests on the physical sample of the aerospace structural component to be diagnosed to obtain high-fidelity test data; Step 3: Perform unified format conversion and standardization preprocessing on the acquired low-fidelity simulation data and high-fidelity real-measured data, and then perform time-frequency conversion on the standardized preprocessed low-fidelity simulation data and high-fidelity real-measured data to obtain the corresponding time-frequency domain low-fidelity simulation data and high-fidelity real-measured data. Step 4: Based on the K-nearest neighbor strategy, establish the similarity relationship between low-fidelity simulation data and high-fidelity measured data, and construct the two into multi-fidelity graph structure data with node and edge features; The constructed multi-fidelity graph structure data is divided into training and testing sets; Step 5: Establish a dual-headed graph convolutional neural network based on residual fusion, and extract potential structural fault features from multi-fidelity graph structural data through the dual-headed graph convolutional neural network; Step 6: Establish a lightweight feature classifier to perform nonlinear transformation and boundary discrimination on potential structural fault features, thereby classifying the fault categories of aerospace structural components; Step 7: Combine a dual-head graph convolutional neural network based on residual fusion and a lightweight feature classifier based on convolutional neural network to establish a fault diagnosis model for aerospace structural components driven by simulation-measurement data fusion; complete the training of the fault diagnosis model using multi-fidelity graph structural data in the training set; Step 8: Input the multi-fidelity graph structure data from the test set into the trained fault diagnosis model to perform fault diagnosis.

2. The method for fault diagnosis of aerospace structural components driven by simulation-measurement data fusion according to claim 1, characterized in that, In step 3, the standardization preprocessing uses Z-score standardization, and the calculation formula is as follows: ， ， ； In the formula: For low-fidelity simulation data and high-fidelity real-world test data, the first One value, ; The average of low-fidelity simulation data and high-fidelity real-world measurement data. The standard deviations of low-fidelity simulation data and high-fidelity measured data are given. For the standardized low-fidelity simulation data and high-fidelity real-world measurement data, the first... Values.

3. The method for fault diagnosis of aerospace structural components driven by simulation-measurement data fusion according to claim 1, characterized in that, Step 4 specifically involves: Step 4.1: Determine the nodes; One data point is taken from the high-fidelity measured data as a high-fidelity sample, based on Euclidean distance. The K-nearest neighbor strategy obtains the k nearest neighbor samples of the high-fidelity sample from the low-fidelity simulation data, thereby determining the k nodes of the high-fidelity sample; Euclidean distance. The calculation formula is: ； In the formula: To provide feature vectors for high-fidelity real-world test data, The feature vector of low-fidelity simulation data; For feature dimensions; The first in high-fidelity real measurement data One value; The first in low-fidelity simulation data One value; Step 4.2: Using the following formula, based on k nodes of the high-fidelity sample... Determine the edge weights between nodes : ； In the formula: To ensure the accuracy of the test data, This is low-fidelity simulation data; for k neighbors, For k nearest neighbors and The average Euclidean distance between them; Step 4.3: Following the same method as steps 4.1 and 4.2, sequentially obtain the k nodes corresponding to all high-fidelity samples. and edge weights Through the k nodes corresponding to all high-fidelity samples and edge weights Establish similarity relationships between low-fidelity simulation data and high-fidelity measured data, and construct them into multi-fidelity graph structure data with node and edge features; Step 4.4: Divide the obtained multi-fidelity graph structure data into training set and test set according to the set ratio.

4. The method for fault diagnosis of aerospace structural components driven by simulation-measurement data fusion according to claim 1, characterized in that: In step 5, the dual-head graph convolutional neural network includes two input branches and a first adder; Both input branches include a first Mean module, and a first ChebConv module, a first EdgePool module, a second ChebConv module, a second EdgePool module, a second Mean module, and a second adder connected in sequence; The input terminal of the first Mean module is connected to the second output terminal of the first EdgePool module, and the output terminal is connected to the second input terminal of the second adder; The output of the second adder in the two input branches is connected to the two inputs of the first adder, and the output of the first adder is connected to the input of the lightweight feature classifier.

5. The method for fault diagnosis of aerospace structural components driven by simulation-measurement data fusion according to claim 1, characterized in that: In step 6, the lightweight feature classifier includes two fully connected layers and a Dropout regularization layer connected between the two fully connected layers; The input of the first fully connected layer is connected to the output of the first adder; The two fully connected layers are used to perform nonlinear transformations on the input structural fault features to determine the category boundaries and achieve efficient classification of different fault types. The Dropout regularization layer is used to ignore the connections of some neurons in the fully connected layer during training, preventing the classifier from overfitting.

6. The method for fault diagnosis of aerospace structural components driven by simulation-measurement data fusion according to claim 1, characterized in that: In step 7, the following cross-entropy loss function is used during training. Perform model optimization: ； In the formula: Number of categories; For the first The true label of the class; For the fault diagnosis model to the first The predicted probability of a class.

7. The method for fault diagnosis of aerospace structural components driven by simulation-measurement data fusion according to claim 1, characterized in that: Step 2 specifically involves conducting tests on physical samples of the aerospace structural components to be diagnosed under various operating conditions to obtain high-fidelity test data under these conditions.

Images