Full-field strain prediction method for aircraft structure static test based on operator neural network
By constructing residual fully connected and deep operator neural networks and combining them with finite element models, real-time and accurate prediction of the full-field strain of aircraft structures was achieved. This solved the problem that local measurements could not capture the full-field strain distribution, thus improving the accuracy and efficiency of the experiment.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINA AIRPLANT STRENGTH RES INST
- Filing Date
- 2026-03-12
- Publication Date
- 2026-06-19
AI Technical Summary
In static tests of aircraft structures, local point strain measurements are difficult to capture the strain distribution across the entire field, resulting in low simulation efficiency and poor real-time performance, which poses safety risks.
By employing an operator neural network-based approach, a residual fully connected neural network and a residual deep operator neural network are constructed, and combined with an aircraft structural finite element model, a combined residual deep operator neural network is trained to achieve real-time and accurate prediction of the full-field strain of the aircraft structure.
It significantly improves the prediction accuracy and computational efficiency of the full-field strain response in static tests of aircraft structures, and provides technical support for test monitoring and decision-making.
Smart Images
Figure CN122241864A_ABST
Abstract
Description
Technical Field
[0001] This application belongs to the field of aircraft structural static testing technology, specifically relating to a method for predicting full-field strain in aircraft structural static testing based on operator neural networks. Background Technology
[0002] Static tests on aircraft structures simulate the stress state of an aircraft in flight on the ground to verify the strength performance of the aircraft structure. During the test, the aircraft structure is subjected to severe loads, posing significant safety risks.
[0003] In static tests of aircraft structures, strain data of the aircraft structure are usually collected by devices such as resistance strain gauges and fiber optic sensors. The aircraft structure is then monitored in combination with the results of finite element pre-simulation. However, local point strain measurement is difficult to capture the details of the strain distribution in the entire physical field of the aircraft structure. Simulation calculation has problems such as difficulty in quantifying manufacturing and assembly uncertainties, low computational efficiency, and poor real-time performance.
[0004] In view of the aforementioned technical deficiencies, this application is hereby filed. Summary of the Invention
[0005] The purpose of this application is to provide a method for predicting the full-field strain of aircraft structures in static tests based on operator neural networks, so as to overcome or mitigate at least one of the known technical defects.
[0006] The technical solution of this application is:
[0007] A method for predicting full-field strain in static tests of aircraft structures based on operator neural networks includes:
[0008] Step 1: In the fully connected neural network, introduce residual connections to construct a residual fully connected neural network;
[0009] Step 2: In the deep operator neural network, a residual fully connected neural network is used to construct a residual deep operator neural network;
[0010] Step 3: Decompose the aircraft structure into multiple substructures, configure a corresponding residual deep operator neural network for each substructure, and combine the residual deep operator neural networks of each substructure into a combined residual deep operator neural network. The input of the combined residual deep operator neural network is the load of the aircraft structure, and the output is the strain of the aircraft structure.
[0011] Step 4: Train the combined residual deep operator neural network;
[0012] Step 5: Collect local load information of the aircraft structure, combine residual deep operator neural networks, and predict the full-field strain response of the aircraft structure.
[0013] Optionally, in the above-mentioned method for predicting the full-field strain of aircraft structure static test based on operator neural network, in step one, the number of neurons in each subsequent layer of the residual fully connected neural network is the same, except for the input layer.
[0014] Optionally, in the above-mentioned method for predicting the full-field strain of aircraft structure static test based on operator neural network, in step two, the branch network and trunk network of the residual deep operator neural network are both residual fully connected neural networks. The structures of the branch network and trunk network are consistent except for the input layer. The input of the branch network is the load vector, and the dimension is the load vector degree of freedom. The input of the trunk network is spatial coordinates.
[0015] Optionally, in the above-mentioned method for predicting the full-field strain of aircraft structure static test based on operator neural network, in step four, the combined residual deep operator neural network is trained using the finite element calculation data of the aircraft structure.
[0016] Optionally, in the above-mentioned method for predicting the full-field strain of aircraft structure static test based on operator neural network, in step four, the training of the combined residual deep operator neural network is a two-stage training. The first stage is to train the strain response prediction under a specific load scenario; the second stage is to train the strain response prediction under all load scenarios in the training set. Both stages of training use the Adam optimizer for iterative training.
[0017] This application has at least the following beneficial technical effects:
[0018] This paper presents a method for predicting the full-field strain of aircraft structures in static tests based on operator neural networks. Using finite element calculation data of aircraft structures as a basis, a surrogate model capable of predicting the full-field strain response of aircraft structures is obtained by training a combined residual deep operator neural network. This method can achieve real-time and accurate prediction of the full-field strain response of aircraft structures in static tests, significantly improving the prediction accuracy and computational efficiency of the test response, and providing technical support for test monitoring, early warning, and command decision-making. Attached Figure Description
[0019] Figure 1 This is a schematic diagram of the full-field strain prediction method for static test of aircraft structure based on operator neural network provided in the embodiments of this application.
[0020] To better illustrate this embodiment, some content in the accompanying drawings may be omitted. They are for illustrative purposes only and should not be construed as limiting the scope of this application. Detailed Implementation
[0021] To make the technical solution and advantages of this application clearer, the technical solution of this application will be described in a clearer and more complete manner below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only some embodiments of this application, and are only used to explain this application, not to limit this application. It should be noted that, for ease of description, only the parts related to this application are shown in the accompanying drawings, and other related parts can be referred to the general design.
[0022] Furthermore, unless otherwise defined, the technical or scientific terms used in this application description shall have the ordinary meaning understood by one of ordinary skill in the art to which this application pertains. The word "comprising" as used in this application description indicates that the concept preceding the word encompasses the concepts listed following the word and their equivalents, without excluding other related concepts.
[0023] Operator neural networks, with their strong modeling ability for high-dimensional nonlinear mappings, can integrate analysis and experiment, learning the complex correlation between strain fields and loads from experimental data, thus significantly improving prediction accuracy and computational efficiency. Based on this, this application provides a method for full-field strain prediction in static tests of aircraft structures based on operator neural networks, such as... Figure 1 As shown, by constructing an operator neural network for static testing, based on the finite element calculation results, real-time and accurate prediction of the aircraft structure from local load input to full-field strain response can be achieved. It can make high-precision real-time prediction of the full-field response of the aircraft structure during the test loading process, providing support for test visualization and early warning decision-making.
[0024] Step 1: In the fully connected neural network, introduce residual connections to construct a residual fully connected neural network.
[0025] The mathematical expressions of two adjacent layers in a fully connected neural network are the same, the first... Layer output By the Layer output After linear transformation and activation function, its mathematical expression is as follows:
[0026] ;
[0027] in,
[0028] For fully connected neural networks, the activation function can be selected. , calculated as ,in, This is the input to the activation function;
[0029] For the fully connected neural network, the first Layer weight matrix;
[0030] For the fully connected neural network, the first Layer bias.
[0031] In residual connections, the residual function is used. For the fully connected neural network, the first Layer output Perform the transformation, and then combine it with the first... Layer output Add them together, and through the activation function, we obtain the first... Layer output Its mathematical expression is as follows:
[0032] .
[0033] residual function It contains two linear layers, namely:
[0034] ;
[0035] in,
[0036] This is the weight matrix of the first linear layer within the residual block;
[0037] This is the weight matrix of the second linear layer within the residual block;
[0038] The bias of the first linear layer within the residual block;
[0039] This is the bias of the second linear layer within the residual block.
[0040] In a residual fully connected neural network, all layers except the input layer have the same number of neurons to ensure that the residual connections are effective. Specifically, a residual fully connected neural network can be designed with 7 layers, 3 neurons in the input layer, taking the degrees of freedom of the load vector, and 50 neurons in each subsequent layer, as shown below: .
[0041] Step 2: In the deep operator neural network, a residual fully connected neural network is used to construct a residual deep operator neural network.
[0042] Deep operator neural networks Based on the universal approximation theorem of functionals, the system consists of two fully connected networks: a branch network and a trunk network. Its mathematical expression is as follows:
[0043] ;
[0044] in,
[0045] For the branch network In the component, corresponding to the first... The weights of each neuron;
[0046] is the activation function for the branch network;
[0047] For the branch network In the component, the first The input dimension to the first The connection weights of each neuron;
[0048] For the branch network The first component Bias of each neuron;
[0049] The first input to the branch network load vector One component;
[0050] is the activation function for the trunk network;
[0051] For trunk network The weight vector corresponding to each component;
[0052] These are the input spatial coordinates for the trunk network;
[0053] For the first trunk network The bias of each component.
[0054] Can be attached to , This eliminates the need for explicit expression in the network structure, thus... It will also not appear in the network structure;
[0055] , , This represents the number of network components, neurons, and the number of input components for the load vector.
[0056] Both the branch network and the trunk network use residual fully connected neural networks. The structures of the branch network and the trunk network are consistent except for the input layer. The input of the branch network is the load vector, and the dimension is the load vector degrees of freedom. The input of the trunk network is spatial coordinates.
[0057] Step 3: Decompose the aircraft structure into multiple substructures, configure a corresponding residual deep operator neural network for each substructure, and combine the residual deep operator neural networks of each substructure into a combined residual deep operator neural network. The input of the combined residual deep operator neural network is the load of the aircraft structure, and the output is the strain of the aircraft structure.
[0058] The residual deep operator neural network is designed for a specific substructure (such as skin, frame, beam, rib) in an aircraft structure. Based on the structure modeled in the finite element model of the aircraft structure, the aircraft structure can be decomposed into multiple substructures and combined. The residual deep operator neural network is composed of the residual deep operator neural networks of each structure, and each residual deep operator neural network corresponds to the learning task of the full-field strain response of a substructure.
[0059] In a specific example, the wing is divided into 19 substructures, named ELE_HL, ELE_LE_0_H, ELE_LE_0_Q, ELE_LE_0_Z, ELE_LE_1_H, ELE_LE_1_Q, ELE_LE_1_Z, ELE_LE_2_H, ELE_LE_2_Q, ELE_LE_2_Z, ELE_LE_3_H, ELE_LE_3_Q, ELE_LE_3_Z, ELE_LE_4_H, ELE_LE_4_Q, ELE_LE_4_Z, ELE_QL, ELE_SMP, and ELE_XMP. The combined residual deep operator neural network consists of 19 residual deep operator neural networks with the same structure, and each residual deep operator neural network corresponds to the learning task of the full-field strain response of one substructure.
[0060] Step 4: Train the combined residual deep operator neural network.
[0061] Using finite element analysis data of aircraft structures, a combined residual deep operator neural network is trained. This allows for the calculation of the full-field strain response of aircraft structures under 1000 load conditions using finite element analysis software. The ratio of the data sets is divided into training, validation, and test sets to train the combined residual deep operator neural network.
[0062] The training of the combined residual deep operator neural network is a two-stage training process. The first stage trains the strain response prediction under a specific load scenario (i.e., arbitrarily selecting a set of data from the training set for training). The second stage trains the strain response prediction under all load scenarios in the training set (using all sets of data in the training set for training). Both stages of training use the Adam optimizer for iterative training.
[0063] Now assume that the combined residual deep operator neural network is ,in, As the input to the neural network that combines residual deep operators, To combine the undetermined weights and biases in the residual deep operator neural network, the first-stage training loss function is... Represented as:
[0064] ;
[0065] Second-stage training loss function Represented as:
[0066] ;
[0067] in,
[0068] This represents the total number of training sample points;
[0069] for Adaptive weighting coefficients for each sample point;
[0070] To combine residual deep operator neural networks for the first The predicted output for each sample;
[0071] For the first The true response of each sample point;
[0072] The coefficient of the regularization term can be taken as... ;
[0073] This represents the total number of parameters in the combined residual deep operator neural network.
[0074] For the first in the combined residual deep operator neural network One parameter.
[0075] Second-stage training loss function , compared with the first stage training loss function In contrast, it does not include adaptive weights, which are iterated continuously during training, in the following manner:
[0076] ;
[0077] in,
[0078] The weighting coefficients for the update step can be taken as follows: .
[0079] Step 5: Collect local load information of the aircraft structure, combine residual deep operator neural networks, and predict the full-field strain response of the aircraft structure.
[0080] The average relative error can be used. To measure the accuracy of the prediction of the full-field strain response of the aircraft structure:
[0081] .
[0082] Verification has shown that using a combined residual deep operator neural network to predict the full-field strain response of an aircraft structure yields a mean relative error. It is within 5% and has high efficiency.
[0083] The above-described embodiment discloses a method for predicting the full-field strain of aircraft structures in static tests based on operator neural networks. By training a combined residual deep operator neural network, a surrogate model capable of predicting the full-field strain response of the aircraft structure is obtained. While meeting the accuracy requirements, it can significantly improve the calculation speed of the full-field strain response, providing important support for command and decision-making in static tests of aircraft structures.
[0084] The technical solution of this application has been described in conjunction with the preferred embodiments shown in the accompanying drawings. Those skilled in the art should understand that the scope of protection of this application is obviously not limited to these specific embodiments. Without departing from the principles of this application, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of this application.
Claims
1. A method for predicting full-field strain in static tests of aircraft structures based on operator neural networks, characterized in that, include: Step 1: In the fully connected neural network, introduce residual connections to construct a residual fully connected neural network; Step 2: In the deep operator neural network, a residual fully connected neural network is used to construct a residual deep operator neural network; Step 3: Decompose the aircraft structure into multiple substructures, configure a corresponding residual deep operator neural network for each substructure, and combine the residual deep operator neural networks of each substructure into a combined residual deep operator neural network. The input of the combined residual deep operator neural network is the load of the aircraft structure, and the output is the strain of the aircraft structure. Step 4: Train the combined residual deep operator neural network; Step 5: Collect local load information of the aircraft structure, combine residual deep operator neural networks, and predict the full-field strain response of the aircraft structure.
2. The method for predicting full-field strain in static tests of aircraft structures based on operator neural networks according to claim 1, characterized in that, In step one, in the residual fully connected neural network, except for the input layer, the number of neurons in each subsequent layer is the same.
3. The method for predicting full-field strain in static tests of aircraft structures based on operator neural networks according to claim 2, characterized in that, In step two, in the residual deep operator neural network, both the branch network and the trunk network are residual fully connected neural networks. The structures of the branch network and the trunk network are consistent except for the input layer. The input of the branch network is the load vector, and the dimension is the load vector degree of freedom. The input of the trunk network is spatial coordinates.
4. The method for predicting full-field strain in static tests of aircraft structures based on operator neural networks according to claim 3, characterized in that, In step four, the combined residual deep operator neural network is trained using finite element calculation data of the aircraft structure.
5. The method for predicting full-field strain in static tests of aircraft structures based on operator neural networks according to claim 4, characterized in that, In step four, the training of the combined residual deep operator neural network is a two-stage training. The first stage is to train the strain response prediction under a specific load scenario; the second stage is to train the strain response prediction under all load scenarios in the training set. Both stages of training use the Adam optimizer for iterative training.