A fast correction method for aircraft model wings considering random uncertainty

Abstract

This invention proposes a rapid correction method for aircraft model wings that takes random uncertainties into account. This method, which belongs to the field of aerospace technology, addresses the technical issues that traditional random model correction methods exhibit high computational complexity during model correction and may result in errors between the model and the actual structure when dealing with multi-source uncertainties. The invention simulates an aircraft model using finite element technology to generate a large number of random samples containing wing structural parameters. A CNN is used to construct a mapping relationship between the wing dynamic response and parameters, and this model is used for parameter correction. A CNN model is designed and trained, using five sets of wing natural frequencies as input to predict the mean and variance of the wing dimensions. A data generation and processing pipeline is established in the ANSYS and Python environments to automate sample generation. The generated data is used to train the CNN and perform parameter inversion to rapidly correct for model uncertainties.

Description

Technical Field

[0001] The invention relates to a method for quickly correcting an aircraft model wing taking random uncertainty into consideration, and belongs to the technical field of aerospace. Background Art

[0002] In the aerospace industry, finite element technology (FEM) is commonly used to simulate aircraft structures to improve efficiency and reduce costs. Finite element technology is a numerical analysis method used to predict the response behavior of complex structures. By dividing the structure into small finite element units (FEMs), engineers can accurately simulate the response of each component under different conditions. This technology has become an essential tool for studying aircraft structural performance. At the same time, the rise of deep learning technology has brought new changes to various fields. By constructing multi-layer neural networks, deep learning can effectively capture complex nonlinear relationships and has achieved remarkable results in pattern recognition, image processing, and other fields. Combining deep learning with traditional engineering techniques has become a hot topic of research.

[0003] Existing technologies suffer from the following drawbacks: 1. Traditional stochastic model correction methods rely on the construction of optimization algorithms, which typically consume significant computational resources and time. This approach exhibits high computational complexity during model correction, making real-time or rapid correction difficult to achieve. 2. Existing methods can introduce errors between the model and the actual structure when dealing with multiple sources of uncertainty, limiting their applicability and accuracy in dynamic environments. Summary of the Invention

[0004] The present invention aims to solve the technical problems that traditional random model correction methods exhibit high computational complexity in model correction and may cause errors between the model and the actual structure when dealing with multi-source uncertainties. A rapid correction method for aircraft model wings considering random uncertainties is proposed, which specifically includes:

[0005] Step 1: Perform finite element modeling on the aircraft wing model, perform dynamic simulation on the finite element model of the aircraft wing, and complete the parameterization of the aircraft wing model. In the parameterization, the geometric dimensions of the aircraft wing, wing length a and wing width b, are selected as uncertain parameters.

[0006] Step 2: Input the parameters of the wing model to be corrected into the finite element model for simulation calculation to obtain the corresponding structural characteristics;

[0007] Step 3: In the Python and Nastran environments, generate multiple random samples containing wing structural parameters, calculate the natural frequencies in batches, and complete the construction of the training data set;

[0008] Step 4: Build a convolutional neural network to learn the mapping relationship between wing dynamic response and parameters;

[0009] Step 5: Use the training data set to train the convolutional neural network, input the parameters of the wing model to be corrected into the trained convolutional neural network for parameter inversion, and obtain the model parameters generated by the inversion;

[0010] Step 6: Input the model parameters generated by inversion into the finite element model to calculate the corrected structural characteristics;

[0011] Step 7: Calculate the error between the corrected structural feature quantity and the actual structural feature quantity. If the error is greater than the preset value, process the random samples and the convolutional neural network until the error between the corrected structural feature quantity and the actual structural feature quantity is less than the preset value, completing the rapid correction of the aircraft model wing.

[0012] Preferably, the step of establishing a finite element model of the aircraft wing in step 1 includes:

[0013] Step 1.1.1: Select quadrilateral elements as mesh elements, use the mesh generator to perform global meshing of the wing model, and ensure that the mesh distribution is uniform.

[0014] Step 1.1.2: Densely mesh the stress concentration areas of the wing model and use finite element software to check the mesh quality to complete the meshing of the wing model.

[0015] Step 1.1.3: Select aluminum as the material for the wing model. Set the elastic modulus, Poisson's ratio, and density of the aluminum to complete the finite element model.

[0016] Preferably, the step of performing dynamic simulation on the finite element model of the aircraft wing in step 1 includes:

[0017] Step 1.2.1: Perform finite element analysis using Nastrn software;

[0018] Step 1.2.2: Read the finite element analysis results and obtain the first six natural frequencies of the wing;

[0019] Step 1.2.3: Perform a static analysis on the finite element model of the aircraft wing to evaluate the strength and displacement of the wing under all loads and complete the parameterization of the aircraft wing model.

[0020] Preferably, step 1.2.1 specifically includes:

[0021] Step 1.2.1.1: Check the integrity of the wing model and the correctness of the finite element model parameter settings;

[0022] Step 1.2.1.2: Import the configured finite element model and set the calculation options and solution parameters;

[0023] Step 1.2.1.3: Start the solver to solve the problem and monitor the calculation status and convergence in real time during the finite element analysis process.

[0024] Preferably, step 3 specifically includes:

[0025] Step 3.1: Set the standard deviation of the two sets of random variables a and b and the range of the means of the two sets of random variables a and b;

[0026] Step 3.2: Use the rand function to generate m random number sample sets;

[0027] Step 3.3: Based on the random standard deviation and mean set in step 3.1, use Gaussian distribution to generate n, n∈m wing length and wing width data for the right wing. Then, by reflecting the wing length and wing width data for the right wing, obtain the wing length and wing width data for the left wing. Repeat the above steps to iteratively generate wing size data.

[0028] Step 3.4: Save the wing size data to an Excel file;

[0029] Step 3.5: Read the wing size data from the Excel file and call Nastran software to batch calculate the first six natural frequencies.

[0030] Preferably, step 3.4 specifically includes:

[0031] Step 3.4.1: After completing each iteration of wing dimension data generation, save the currently calculated wing length a and wing width b data to a local file;

[0032] Step 3.4.2: Use the pandas library to read and process the wing size data for each iteration, calculate the summary wing length a and wing width b, and save the summary data to an Excel file named according to the iteration number;

[0033] Step 3.4.3: Use the shutil module and os library to manage the generated Excel file, including copying and deleting operations.

[0034] Preferably, step 3.5 specifically includes:

[0035] Step 3.5.1: Read the wing size data from the Excel file, generate the mesh coordinates of the finite element model, create the aircraft wing model bdf file, and use the generated finite element model mesh coordinates to replace the corresponding mesh coordinates in the bdf file;

[0036] Step 3.5.2: Use a Python script to call Nastran software to perform finite element analysis of the wing using the aircraft wing model bdf file and calculate the first six natural frequencies;

[0037] Step 3.5.2: Read and process the output of the Nastran software, extract the natural frequency data, and store the natural frequency data and the generated wing geometry parameters in the corresponding Excel file to complete the construction of the training data set.

[0038] Preferably, the convolutional neural network in step 4 includes an input layer, a convolution layer, a pooling layer, a DroOut layer, a flattening layer and a fully connected layer, and the loss function of the convolutional neural network is a mean square error function;

[0039] The input layer is used to receive data of shape (n, 5), where n is the number of wing natural frequency data and 5 is the first five natural frequency data of the wing;

[0040] The convolution layer includes the first convolution layer and the second convolution layer, which are used to extract the features of the first five natural frequency data;

[0041] The first convolutional layer has a layer type of Conv1D, a number of neurons of 64, a convolution kernel size of 3, and an activation function of ReLU.

[0042] The second convolutional layer has a layer type of Conv1D, a number of neurons of 128, a convolution kernel size of 3, and an activation function of ReLU.

[0043] The flattening layer is used to flatten the convolutional data output and convert it into one-dimensional data;

[0044] The pooling layer is used to reduce the dimension of the first five natural frequency data. The layer type of the pooling layer is MaxPooling1D, and the pooling window size is 2;

[0045] The fully connected layer is used to process the first five natural frequency data after pooling to obtain the predicted value of the natural frequency data. The number of neurons in the fully connected layer is 128, and the activation function is the ReLU function.

[0046] The DroOut layer is set to a ratio of 0.5, which is used to discard the neuron output at a ratio of 0.5;

[0047] The output layer is used to output the predicted value of the natural frequency data. The number of neurons in the output layer is 4, and the activation function is linear. The 4 neurons correspond to the mean value of wing length a, the variance of wing length a, the mean value of wing width b, and the variance of wing width b.

[0048] The expression of the mean square error function is:

[0049]

[0050] In formula (1), n ​​is the number of wing natural frequency data, y i is the true value of the natural frequency data, ypred is the predicted value of the natural frequency data.

[0051] Preferably, the step of using the training data set to train the convolutional neural network in step 5 includes:

[0052] Step 5.1: Perform min-max normalization on the input training data to scale the training data to the range of (0, 1) to facilitate processing of data of different magnitudes and improve the accuracy of the algorithm;

[0053] Step 5.2: Train using early stopping.

[0054] Step 5.3: Use the LOSS function and mean absolute error function to evaluate the predicted value of the natural frequency data output by the training;

[0055] The expression for minimum-maximum normalization is:

[0056]

[0057] In formula (2), x′ is the normalized training data, x min is the minimum value of the training data, x max is the maximum value of the training data.

[0058] Preferably, the processing of the random samples and the convolutional neural network in step 7 includes:

[0059] The training data set of the convolutional neural network is expanded, the structure of the convolutional neural network is adjusted, and the convolutional neural network is retrained.

[0060] The beneficial effects of the present invention are:

[0061] 1. The present invention uses CNN for parameter inversion, which significantly reduces calculation and correction time, improves the speed of model correction, and enables it to be applied to design adjustments in real time.

[0062] 2. The present invention constructs a CNN network, which can extract and learn complex data features, improve the accuracy of prediction, make the correction of the mean and variance of model parameters more precise, and enhance the reliability of the design.

[0063] 3. The present invention constructs a Python and Nastran interface to achieve one-click data generation and analysis, realizes process automation, reduces manual intervention, and improves work efficiency.

[0064] 4. The present invention utilizes the structural adaptability of the CNN network to process diverse input parameters, adapt to different modeling requirements, and easily cope with various design changes.

[0065] 5. The present invention reduces sample requirements and improves data utilization through deep learning, effectively reducing experimental data requirements while controlling cost consumption.

[0066] 6. The present invention overcomes the high computational complexity of traditional methods by using a convolutional neural network to construct a mapping model between the random response of structural dynamics and the random characteristics of model parameters, thereby improving the correction accuracy and efficiency. In addition, by generating a large number of random samples and corresponding structural natural frequency data sets and using neural network training, rapid and accurate correction of the random characteristics of model parameters is achieved. The present invention significantly improves the shortcomings of traditional methods in dealing with random characteristics and supports efficient decision-making and improvements in practical engineering applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 A schematic flow chart of a method for rapidly correcting an aircraft model wing taking random uncertainty into account provided by the present invention;

[0068] Figure 2 A schematic diagram of the finite element model mesh and specific mesh parameters provided by the present invention;

[0069] Figure 3 A schematic diagram of the finite element model of an aircraft wing provided by the present invention;

[0070] Figure 4 The simulation result diagram of the natural frequency and corresponding modal vibration shape of the wing provided by the present invention;

[0071] Figure 5 The CNN network architecture diagram provided by the present invention;

[0072] Figure 6 A schematic diagram showing the comparison between the random parameter values ​​and the target values ​​of the wing model provided by the present invention;

[0073] Figure 7 A schematic diagram comparing the corrected wing length a provided by the present invention with the target value randomly distributed;

[0074] Figure 8 A schematic diagram comparing the corrected wing width b provided by the present invention with the target value randomly distributed;

[0075] Figure 9 This is a random distribution diagram of the first five corrected natural frequencies provided by the present invention. DETAILED DESCRIPTION

[0076] Combine Figure 1-3 This embodiment is described as follows. Figure 1 As shown, the steps of a method for quickly correcting an aircraft model wing considering random uncertainty described in this embodiment include:

[0077] S1: Finite element modeling of aircraft wing model;

[0078] S101: Select unit type: Common unit types include tetrahedral unit and hexahedral unit, etc. In this embodiment, quadrilateral unit (Quad) is used;

[0079] S102: Meshing step: Global mesh generation: Use the Paver mesh generator to perform global meshing on the wing model to ensure uniform mesh distribution;

[0080] S103: Local Mesh Optimization: Dense meshing is performed in stress concentration areas of the wing model to improve the calculation accuracy of these areas. For example, the mesh near the wing root and wingtip needs to be particularly refined.

[0081] S104: Mesh quality check: Use the quality check function in the finite element software to check the mesh quality, such as the unit distortion rate, aspect ratio, mesh angle, etc., to ensure that the mesh quality meets the requirements;

[0082] Finite element model and mesh specific parameters such as Figure 2 As shown in the figure, the specific parameters of the mesh are as follows: (1) Number of nodes: 1979; (2) Number of elements: 1337; (3) Element shape: Quad; (4) Mesher type: Paver; (5) Topology type: Quad4; (6) Global edge length: Automatically calculated, value 0.01; (7) Surface list: Surface 1:14.

[0083] S105: Material property setting: Material properties are an important component of the finite element analysis model. In this embodiment, aluminum is used as the material. Aluminum is widely used in the aerospace field and has excellent specific strength and specific stiffness. The setting of material properties directly affects the accuracy of the model and the reliability of the analysis results. According to the characteristics of aluminum, relevant material parameters are set to accurately simulate its mechanical behavior. The following are the setting steps:

[0084] S10501: Elastic modulus: The elastic modulus of aluminum is set at 70 GPa, which is a measure of its stiffness;

[0085] S10502: Poisson's ratio: The Poisson's ratio of aluminum is usually taken as 0.33, which represents the ratio of the material's lateral deformation to its longitudinal deformation;

[0086] S10503: Density: Set the density of aluminum to 2.7g / cm 3 , which is used to calculate the model mass and inertial properties.

[0087] The finite element model of the wing of the aircraft model established later in this embodiment is as follows Figure 3As shown, for parameterization, this embodiment selects the geometric dimensions of the aircraft wing, wing length a and wing width b, as uncertain parameters.

[0088] S2: Perform dynamic simulation on the finite element model of the aircraft wing;

[0089] To study the dynamics of the aircraft model structure, this embodiment first performs a dynamic finite element simulation of the aircraft model based on Patran and Nastran software. The specific steps are as follows:

[0090] S201: Analysis Type: Select the appropriate analysis type based on the research purpose. This paper mainly conducts modal analysis to obtain the natural frequencies and vibration shapes of the wing, and performs static analysis to evaluate the strength and displacement of the wing under various loads.

[0091] S202: Simulation software: Use Nastran software for finite element analysis, which has powerful solving capabilities and rich analysis functions.

[0092] S203: Calculation process:

[0093] S20301: Initial setup: Check the integrity of the model and the correctness of parameter settings.

[0094] S20302: Load simulation file: Import the set finite element model file, set calculation options and solution parameters.

[0095] S20303: Run simulation: Start the solver and monitor the calculation status and convergence in real time during the calculation process.

[0096] S204: Result processing: importing the output result file into Patran software to read and obtain the natural frequency and corresponding vibration mode of the wing;

[0097] like Figure 4 As shown, this embodiment calculates and analyzes the first six natural frequencies of the wing, and obtains the first six natural frequencies of the wing and their corresponding modal vibration shape results through simulation.

[0098] S3: Input the parameters of the wing model to be corrected into the finite element model for simulation calculation to obtain the corresponding structural characteristics;

[0099] S4: Generate multiple random samples of wing structural parameters in Python and Nastran environments, calculate natural frequencies in batches, and complete the construction of the training dataset.

[0100] S401: Build Python and Nastran interface;

[0101] The main goal of building the Python and Nastran interface is to automate the process of performing dynamic finite element simulations on aircraft models. The implementation steps are as follows:

[0102] S40101: Prepare and parameterize the aircraft wing model: First, create an initial finite element model of the aircraft wing.

[0103] The geometric parameters of the model (wing length and wing width) are set as variable parameters to facilitate subsequent adjustment through the program.

[0104] S40102: Python Program Development: Write Python programs to implement the following functions:

[0105] S40103: Parameter generation: Generate different combinations of wing geometry parameters based on research requirements.

[0106] S40104: Model File Modification: The program automatically modifies the Nastran input model file based on generated parameters to reflect different geometric size changes.

[0107] S40105: Batch Nastran: Automatically start Nastran for batch analysis by calling system commands or using the appropriate API interface.

[0108] S40106: Data Collection and Processing: Extract key analysis results such as natural frequencies from Nastran output files, and perform preliminary data cleaning and formatting to prepare data for neural network training.

[0109] S402: Data preparation and storage of wing geometric random parameters;

[0110] The random parameters of the wing length and wing width b of the aircraft model are generated as follows:

[0111] S40201: Define parameter ranges: First, set the ranges for the standard deviations of the two sets of random variables a and b (standard deviation A and standard deviation B) and the two sets of means (mean A and mean B). These parameters are designed to reflect physical properties related to wing size and are derived from a uniform distribution.

[0112] S40202: Generate random numbers: Use the rand function in numpy to generate 1000 random number samples. These values ​​serve as the basic parameters for generating wing dimensions, ensuring the randomness and diversity of the dimension data a and b.

[0113] S40203: Wing size data generation:

[0114] S4020301: Wing size calculation: In the program loop, based on the random mean and standard deviation generated previously,

[0115] Gaussian distribution (normal distribution) is used to generate specific wing length and wing width data of 60 wings, and each set of data specifically reflects the mean and variance of the wing length and wing width.

[0116] S4020302: Data reflection processing: To achieve symmetry of the wing, the dimensional data of the left wing part is obtained by reflecting the data of the right part.

[0117] S40204: Data storage and processing:

[0118] S4040401: Save calculation results: After completing each data generation iteration, save the wing length and wing width data to a text file to preserve the details of the data generation and facilitate subsequent analysis.

[0119] S4020402: Export to Excel: Use the pandas library to read and process the previous wing size data, calculate and summarize the wing length and width, and save this data in an Excel file named with the number of iterations to ensure that each data is recorded independently to avoid data overwriting.

[0120] S40205: File management: Use shutil and os libraries to manage generated Excel files, including copying

[0121] Copy and delete operations to ensure the organization of data files and effective use of storage space.

[0122] S40206: Subsequent applications of random data:

[0123] S4020601: Data reading and processing: Read wing size data from Excel files, generate mesh coordinates for the finite element model, create an aircraft wing model bdf file, and replace the corresponding mesh coordinates in the bdf file for subsequent Nastran calculations.

[0124] S4020602: Natural frequency calculation: Use a Python script to call the aircraft wing model bdf file into Nastran software to perform finite element analysis of the wing and calculate its natural frequency.

[0125] S4020603: Results Summary: Read and process the Nastran output (.f06 file), extract the natural frequency data, and store it and the generated wing geometry parameters in the corresponding Excel table to complete the data set.

[0126] Through the above steps, not only the randomness and systematicity of data generation are guaranteed, but also the generated dataset can play a key role in uncertainty assessment of aircraft design and analysis. The format of the generated dataset is shown in Table 1;

[0127] Table 1

[0128]

[0129]

[0130] S5: Construct a convolutional neural network to learn the mapping relationship between wing dynamic response and parameters;

[0131] The convolutional neural network constructed in this embodiment is as follows Figure 5 As shown in the figure, it includes input layer, convolution layer, pooling layer, DroOut layer, flattening layer and fully connected layer. The loss function of the convolutional neural network is the mean square error function.

[0132] The expression of the mean square error function is:

[0133]

[0134] In formula (1), n ​​is the number of wing natural frequency data, y i is the true value of the natural frequency data, y pred is the predicted value of the natural frequency data;

[0135] The input layer is used to receive data of shape (n, 5), where n is the number of wing natural frequency data and 5 is the first five natural frequency data of the wing;

[0136] The convolution layer includes the first convolution layer and the second convolution layer, which are used to extract the features of the first five natural frequency data;

[0137] The first convolutional layer has a layer type of Conv1D, a number of neurons of 64, a convolution kernel size of 3, and an activation function of ReLU.

[0138] The second convolutional layer has a layer type of Conv1D, a number of neurons of 128, a convolution kernel size of 3, and an activation function of ReLU.

[0139] The flattening layer is used to flatten the convolutional data output and convert it into one-dimensional data;

[0140] The pooling layer is used to reduce the dimension of the first five natural frequency data. The layer type of the pooling layer is MaxPooling1D, and the pooling window size is 2;

[0141] The fully connected layer is used to process the first five natural frequency data after pooling to obtain the predicted value of the natural frequency data. The number of neurons in the fully connected layer is 128, and the activation function is the ReLU function.

[0142] The DroOut layer is set to a ratio of 0.5, which is used to discard the neuron output at a ratio of 0.5;

[0143] The output layer is used to output the predicted value of the natural frequency data. The number of neurons in the output layer is 4, and the activation function is linear. The 4 neurons correspond to the mean value of wing length a, the variance of wing length a, the mean value of wing width b, and the variance of wing width b.

[0144] S6: Using the training data set to train the convolutional neural network, inputting the parameters of the wing model to be corrected into the trained convolutional neural network to perform parameter inversion, and obtaining the model parameters generated by the inversion;

[0145] S601: Performing minimum-maximum normalization processing on the input training data to scale the training data to the range of (0, 1) to facilitate processing of data of different magnitudes and improve the accuracy of the algorithm;

[0146] The expression for minimum-maximum normalization is:

[0147]

[0148] In formula (2), x′ is the normalized training data, x m i n is the minimum value of the training data, x max is the maximum value of the training data.

[0149] S602: Use early stopping for training. This stops training when the validation loss stops decreasing significantly, thus preventing the model from overfitting on the training set. Specifically, the validation loss function is monitored and training is terminated early when it stops decreasing over several consecutive training rounds.

[0150] S603: Use the LOSS function and the mean absolute error function to evaluate the predicted value of the natural frequency data output by the training. The evaluation results are shown in Table 2.

[0151] Table 2

[0152]

[0153]

[0154] As shown in Table 2, the convolutional neural network constructed in this embodiment has high accuracy and efficiency when processing test data. The low value of the mean error rate (MERR) indicates that the average deviation between the predicted parameters and the target parameters is small, indicating that the convolutional neural network has high reliability and generalization ability.

[0155] S7: inputting the model parameters generated by the inversion into the finite element model to calculate the corrected structural characteristic quantities;

[0156] S8: Calculate the error between the corrected structural feature quantity and the actual structural feature quantity. If the error is greater than a preset value, expand the training data set of the convolutional neural network, adjust the structure of the convolutional neural network, and retrain the convolutional neural network until the error between the corrected structural feature quantity and the actual structural feature quantity is less than the preset value, thereby completing the rapid correction of the aircraft model wing.

[0157] The comparison between the corrected random parameter values ​​of the aircraft model wing and the target values ​​is shown in Table 3.

[0158] Table 3

[0159] parameter Target value Correction value Error rate <![CDATA[μ a ]]> 0.298078297 0.298486398 0.1369％ <![CDATA[σ a ]]> 0.003927089 0.003787407 3.5576％ <![CDATA[μ b ]]> 0.025039665 0.025018017 0.0864％ <![CDATA[σ b ]]> 0.001696540 0.001735816 2.3146％

[0160] As shown in Table 3, in this embodiment, the average error of the corrected wing geometric dimension mean is 0.1117%, and the average error of the wing geometric dimension variance is 2.9361%. The fitting results of the distribution of the corrected value and the target value are shown in Table 3. Figure 6 As shown, it can be seen that the correction value of this embodiment is relatively well-matched with the target value distribution, and the corrected model has achieved a good correction effect.

[0161] In this embodiment, the corrected mean and variance of the wing length and wing width are used to regenerate the wing length (a) and wing width (b) values ​​and import them into Nastran software to calculate their corresponding fifth-order natural frequencies. The corrected wing length (a) and wing width (b) values ​​and natural frequencies are compared with the target values. The comparison results are as follows: Figure 7 、 Figure 8 , as shown in Table 4 and Table 5.

[0162] Table 4

[0163]

[0164]

[0165] Table 5

[0166] Mean natural frequency Target value Correction value Error rate <![CDATA[f1]]> 0.151648 0.146359 3.489％ <![CDATA[f2]]> 0.406761 0.389584 4.218％ <![CDATA[f3]]> 3.990503 4.100371 2.752％ <![CDATA[f4]]> 3.778299 3.985891 5.497％ <![CDATA[f5]]> 4.161709 4.308347 3.526％

[0167] like Figure 9 As shown in Table 4, the distribution of the first five natural frequencies after correction in this embodiment is highly fitted with the first five natural frequencies of the target structure. As shown in Tables 4 and 5, the mean average error between the corrected natural frequencies and the target values ​​is 0.450%, and the variance average error is 3.896%, indicating that the constructed convolutional neural network has a good correction effect.

[0168] The above description is merely a preferred embodiment of the present invention and does not constitute any form of limitation to the present invention. Although the present invention has been disclosed as a preferred embodiment as above, it is not intended to limit the present invention. Any technician familiar with the present profession can make some changes or modifications to equivalent embodiments of equivalent changes using the technical content disclosed above without departing from the scope of the technical solution of the present invention. However, any simple modification, equivalent replacement and improvement of the above embodiments made according to the technical essence of the present invention, within the spirit and principles of the present invention, without departing from the content of the technical solution of the present invention, shall still fall within the scope of protection of the technical solution of the present invention.

Claims

1. A method for rapid correction of aircraft model wings considering random uncertainty, characterized in that: The steps of the described method for quickly correcting the wing of an aircraft model considering random uncertainty include: Step 1: Perform finite element modeling on the aircraft wing model, perform dynamic simulation on the finite element model of the aircraft wing, and complete the parameterization of the aircraft wing model. In the parameterization, the geometric dimensions of the aircraft wing, wing length a and wing width b, are selected as uncertain parameters. Step 2: Input the parameters of the wing model to be corrected into the finite element model for simulation calculation to obtain the corresponding structural characteristics; Step 3: In the Python and Nastran environments, generate multiple random samples containing wing structural parameters, calculate the natural frequencies in batches, and complete the construction of the training data set; Step 4: Build a convolutional neural network to learn the mapping relationship between wing dynamic response and parameters; Step 5: Use the training data set to train the convolutional neural network, input the parameters of the wing model to be corrected into the trained convolutional neural network for parameter inversion, and obtain the model parameters generated by the inversion; Step 6: Input the model parameters generated by inversion into the finite element model to calculate the corrected structural characteristics; Step 7: Calculate the error between the corrected structural feature quantity and the actual structural feature quantity. If the error is greater than the preset value, process the random samples and the convolutional neural network until the error between the corrected structural feature quantity and the actual structural feature quantity is less than the preset value, completing the rapid correction of the aircraft model wing.

2. A method for rapidly correcting an aircraft model wing taking random uncertainty into account according to claim 1, characterized in that: The steps in Step 1 to create the finite element model of the aircraft wing include: Step 1.1.1: Select quadrilateral elements as mesh elements, use the mesh generator to perform global meshing of the wing model, and ensure that the mesh distribution is uniform. Step 1.1.2: Densely mesh the stress concentration areas of the wing model and use finite element software to check the mesh quality to complete the meshing of the wing model. Step 1.1.3: Select aluminum as the material for the wing model. Set the elastic modulus, Poisson's ratio, and density of the aluminum to complete the finite element model.

3. A method for rapidly correcting an aircraft model wing taking random uncertainty into account according to claim 1, characterized in that: The steps for performing dynamic simulation on the finite element model of the aircraft wing in step 1 include: Step 1.2.1: Perform finite element analysis using Nastrn software; Step 1.2.2: Read the finite element analysis results and obtain the first six natural frequencies of the wing; Step 1.2.3: Perform a static analysis on the finite element model of the aircraft wing to evaluate the strength and displacement of the wing under all loads and complete the parameterization of the aircraft wing model.

4. A method for rapidly correcting an aircraft model wing taking random uncertainty into account according to claim 3, characterized in that: Step 1.2.1 specifically includes: Step 1.2.1.1: Check the integrity of the wing model and the correctness of the finite element model parameter settings; Step 1.2.1.2: Import the configured finite element model and set the calculation options and solution parameters; Step 1.2.1.3: Start the solver to solve the problem and monitor the calculation status and convergence in real time during the finite element analysis process.

5. A method for rapidly correcting an aircraft model wing taking random uncertainty into account according to claim 1, characterized in that: Step 3 specifically includes: Step 3.1: Set the standard deviation of the two sets of random variables a and b and the range of the means of the two sets of random variables a and b; Step 3.2: Use the rand function to generate m random number sample sets; Step 3.3: Based on the random standard deviation and mean set in step 3.1, use Gaussian distribution to generate n, n∈m wing length and wing width data for the right wing. Then, by reflecting the wing length and wing width data for the right wing, obtain the wing length and wing width data for the left wing. Repeat the above steps to iteratively generate wing size data. Step 3.4: Save the wing size data to an Excel file; Step 3.5: Read the wing size data from the Excel file and call Nastran software to batch calculate the first six natural frequencies.

6. A method for rapidly correcting an aircraft model wing taking random uncertainty into account according to claim 5, characterized in that: Step 3.4 specifically includes: Step 3.4.1: After completing each iteration of wing dimension data generation, save the currently calculated wing length a and wing width b data to a local file; Step 3.4.2: Use the pandas library to read and process the wing size data for each iteration, calculate the summary wing length a and wing width b, and save the summary data to an Excel file named according to the iteration number; Step 3.4.3: Use the shutil module and os library to manage the generated Excel file, including copying and deleting operations.

7. The method for rapidly correcting an aircraft model wing taking random uncertainty into consideration according to claim 5, wherein: Step 3.5 specifically includes: Step 3.5.1: Read the wing size data from the Excel file, generate the mesh coordinates of the finite element model, create the aircraft wing model bdf file, and use the generated finite element model mesh coordinates to replace the corresponding mesh coordinates in the bdf file; Step 3.5.2: Use a Python script to call Nastran software to perform finite element analysis of the wing using the aircraft wing model bdf file and calculate the first six natural frequencies; Step 3.5.2: Read and process the output of the Nastran software, extract the natural frequency data, and store the natural frequency data and the generated wing geometry parameters in the corresponding Excel file to complete the construction of the training data set.

8. The method for rapidly correcting an aircraft model wing taking random uncertainty into consideration according to claim 1, wherein: In step 4, the convolutional neural network includes input layer, convolution layer, pooling layer, DroOut layer, flattening layer and fully connected layer. The loss function of the convolutional neural network is the mean square error function. The input layer is used to receive data of shape (n, 5), where n is the number of wing natural frequency data and 5 is the first five natural frequency data of the wing; The convolution layer includes the first convolution layer and the second convolution layer, which are used to extract the features of the first five natural frequency data; The first convolutional layer has a layer type of Conv1D, a number of neurons of 64, a convolution kernel size of 3, and an activation function of ReLU. The second convolutional layer has a layer type of Conv1D, a number of neurons of 128, a convolution kernel size of 3, and an activation function of ReLU. The flattening layer is used to flatten the convolutional data output and convert it into one-dimensional data; The pooling layer is used to reduce the dimension of the first five natural frequency data. The layer type of the pooling layer is MaxPooling1D, and the pooling window size is 2; The fully connected layer is used to process the first five natural frequency data after pooling to obtain the predicted value of the natural frequency data. The number of neurons in the fully connected layer is 128, and the activation function is the ReLU function. The DroOut layer is set to a ratio of 0.5, which is used to discard the neuron output at a ratio of 0.5; The output layer is used to output the predicted value of the natural frequency data. The number of neurons in the output layer is 4, and the activation function is linear. The 4 neurons correspond to the mean value of wing length a, the variance of wing length a, the mean value of wing width b, and the variance of wing width b. The expression of the mean square error function is: In formula (1), n ​​is the number of wing natural frequency data, y i is the true value of the natural frequency data, y pred is the predicted value of the natural frequency data.

9. The method for rapidly correcting an aircraft model wing taking random uncertainty into consideration according to claim 1, wherein: The steps in step 5 to train the convolutional neural network using the training dataset include: Step 5.1: Perform minimum-maximum normalization on the input training data and scale the training data to the range of (0, 1); Step 5.2: Train using early stopping. Step 5.3: Use the LOSS function and mean absolute error function to evaluate the predicted value of the natural frequency data output by the training; The expression for minimum-maximum normalization is: In formula (2), x′ is the normalized training data, x min is the minimum value of the training data, x max is the maximum value of the training data.

10. The method for rapidly correcting an aircraft model wing considering random uncertainty according to claim 1, characterized in that: The processing of random samples and convolutional neural networks in step 7 includes: The training data set of the convolutional neural network is expanded, the structure of the convolutional neural network is adjusted, and the convolutional neural network is retrained.

Images