Adapter strain field prediction method based on information fusion feature equalization operator

Abstract

The present invention proposes a strain field prediction method for an adapter based on an information fusion feature balance operator, which belongs to the field of adapters, including: S1, obtaining the geometric features, design variables, and design variable variation range of the adapter, and the strain field information is the value to be predicted; S2, establishing an adapter geometric model according to the design variables of the adapter, and constructing a metamaterial statics finite element model based on the adapter geometric model, using the Latin hypercube sampling method to obtain sample points, and using the metamaterial statics finite element model to generate a training data set according to simulation; S3, constructing a neural network model based on an information fusion feature balance operator, and using the training data set to train the neural network model based on the information fusion feature balance operator to obtain a trained neural network model based on the information fusion feature balance operator; S4, using the trained neural network model based on the information fusion feature balance operator to predict the strain field information to be predicted, and obtaining the strain field distribution of the adapter. The present application enhances the extraction and modeling capabilities of geometric features, so that the neural network model based on the information fusion feature balance operator can quickly predict the strain field distribution of adapters with different geometric configurations.

Description

Technical Field

[0001] The present invention relates to the technical field of adapters, and in particular to an adapter strain field prediction method based on an information fusion feature equalization operator. Background Art

[0002] Box launch is a common way of missile launch, in which the adapter is a key component connecting the missile and the launch box, which plays a role in fixing, shock absorption, radial support and guiding the missile launch. In order to improve the storage safety and launch performance of missiles, it is necessary to optimize the structure of the adapter. Since the strain field distribution of the adapter determines its support performance, accurate strain field distribution evaluation plays a vital role in calculating its support performance and guiding the optimal configuration design.

[0003] In the existing technology, the use of traditional high-precision finite element simulation models will result in time-consuming simulation and high computational costs; and the physical field prediction method based on deep learning still has insufficient prediction accuracy and computational efficiency when dealing with variable geometry structures and strain concentration areas.

[0004] Therefore, finding a method that can predict the strain of a variable set adapter and improve the prediction accuracy and speed is a technical problem that needs to be solved urgently by those skilled in the art. Summary of the invention

[0005] In view of this, the present invention proposes an adapter strain field prediction method based on information fusion feature equalization operator, which can enhance the extraction and modeling capabilities of geometric features, significantly improve the prediction accuracy, and enable the neural network model based on information fusion feature equalization operator to quickly predict the adapter strain field distribution of different geometric configurations.

[0006] The technical solution of the present invention is implemented as follows: The present invention provides an adapter strain field prediction method based on information fusion feature equalization operator, comprising the following steps:

[0007] S1, obtaining the geometric characteristics, design variables, and design variable variation range of the adapter, and the strain field information is the value to be predicted;

[0008] S2. Establish an adapter geometry model according to the design variables of the adapter, and construct a metamaterial statics finite element model based on the adapter geometry model, use the Latin hypercube sampling method to obtain sample points, and use the metamaterial statics finite element model to generate a training data set according to simulation;

[0009] S3, constructing a neural network model based on information fusion feature equalization operator, and using a training data set to train the neural network model based on information fusion feature equalization operator to obtain a trained neural network model based on information fusion feature equalization operator;

[0010] S4. Use the trained information fusion feature equalization operator-based neural network model to predict the strain field information to be predicted, and obtain the strain field distribution of the adapter.

[0011] On the basis of the above technical solutions, preferably, the information fusion feature equalization operator neural network model is based on a deep neural operator network, comprising an information fusion module, a feature equalization learning module and a Fourier feature embedding layer, and the information fusion module backbone network and branch network respectively include n hidden layers and one embedding layer.

[0012] On the basis of the above technical solution, preferably, the training process of the neural network model based on information fusion feature equalization operator is as follows:

[0013] S31, load and design parameters P are input into the two branch networks for feature extraction, and the characteristic vectors of load and design parameters are obtained respectively. and The two branch networks are the first branch network and the second branch network. Indicates the first branch network from the load The Pth feature extracted from represents the Pth feature extracted by the second branch network from the design parameter P;

[0014] S32, input the grid node coordinate y into each backbone network and obtain the corresponding output feature Sum the output features to get the fused output features in represents the Pth feature output by the kth backbone network;

[0015] S33, perform matrix multiplication calculation on the features of each layer of the corresponding backbone network and branch network, so that the features between each layer are integrated with each other, and the inner product of the three vectors is taken to obtain the final output result

[0016] S34, use the sine function as the activation function to construct the Fourier feature embedding layer, and use the Fourier feature embedding layer to the final output result After processing, high-frequency strain field characteristics are obtained.

[0017] On the basis of the above technical solution, preferably, step S32 specifically includes:

[0018] The backbone network is expanded into M sub-networks, each of which is responsible for learning a certain part of the features. The grid node coordinate y is input into each backbone network, and the corresponding output features are obtained. Sum the output features to get the fused output features

[0019] On the basis of the above technical solution, preferably, step S33 specifically includes:

[0020] The feature vector [t1,…,t P ] T and the eigenvector Perform element-by-element multiplication and sum to obtain the first eigenvector. The calculation formula of the first eigenvector is:

[0021]

[0022] Where y represents the coordinate of the adapter point, Represents the mth function vector The number of sampling points, b0 represents the bias term, represents the kth eigenvalue extracted from the mth branch network, P represents the design parameter, t k (y) represents the kth feature vector of adapter point y extracted by the backbone network.

[0023] On the basis of the above technical solution, preferably, the HuberLoss loss function is used in the training process of the information fusion feature equalization operator neural network model, and the calculation formula is:

[0024]

[0025] Among them, δ is a hyperparameter, represents the coordinates of the adapter point output based on the information fusion feature equalization operator neural network model, and y represents the coordinates of the adapter point.

[0026] On the basis of the above technical solution, preferably, the step of using the metamaterial statics finite element model to simulate and generate a training data set specifically includes:

[0027] The Latin hypercube sampling method is used to generate multiple adapter configurations as the first training set based on the variation range of the design variables of the adapter;

[0028] The first training set is batch simulated using a metamaterial statics finite element model to obtain the strain field of the adapter at each sample point;

[0029] The strain field at each sample point is preprocessed to obtain a training data set.

[0030] On the basis of the above technical solution, preferably, the pretreatment includes:

[0031] The intermediate nodes of the strain field of the adapter at each sample point are removed, and the strain field of the adapter at each sample point is downsampled to obtain a sparse strain field;

[0032] According to the symmetry of the adapter, half of the strain field after rarefaction is used as the training dataset.

[0033] More preferably, the adapter strain field prediction method based on information fusion feature equalization operator further includes:

[0034] Perform compression tests on adapters of different configurations and obtain test results;

[0035] The test results are compared with the strain field of the adapter at each sample point to verify the metamaterial static finite element model.

[0036] The adapter strain field prediction method of the present invention has the following beneficial effects compared with the prior art:

[0037] (1) By using finite element simulation to generate training data, the neural network model based on information fusion feature balance operator is trained to enhance the ability to extract and model geometric features, significantly improve the prediction accuracy, and enable the neural network model based on information fusion feature balance operator to quickly predict the strain field distribution of adapters with different geometric configurations, thereby improving the efficiency of adapter design;

[0038] (2) The different features of the input data are captured by different branch networks of the deep neural operator network, and the feature outputs of different branch networks are effectively integrated using the information fusion module. The feature weights of each branch network are adaptively optimized using the feature energy concentration learning module, and a Fourier feature embedding layer is added at the end of the neural network model based on the information fusion feature equalization operator to enhance the richness of feature expression and improve the ability of the neural network model based on the information fusion feature equalization operator to capture the strain field distribution. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0040] Figure 1 A schematic diagram of an adapter configuration of the adapter strain field prediction method of the present invention;

[0041] Figure 2 A schematic diagram of adapter parameterization dimensions of the adapter strain field prediction method of the present invention;

[0042] Figure 3 is a flow chart of the adapter strain field prediction method of the present invention;

[0043] Figure 4 The adapter automated finite element simulation process of the adapter strain field prediction method of the present invention, wherein (a) is the geometric configuration, (b) is the meshing, (c) is the load and boundary conditions, and (d) is the calculation solution;

[0044] Figure 5 It is a schematic diagram of the simulation data processing flow of the adapter strain field prediction method of the present invention;

[0045] Figure 6 It is a schematic diagram of the information fusion module structure of the adapter strain field prediction method of the present invention;

[0046] Figure 7 The calculation formula between each layer of the information fusion module of the adapter strain field prediction method of the present invention;

[0047] Figure 8 It is a structural schematic diagram of a feature balance learning module of the adapter strain field prediction method of the present invention;

[0048] Fig. 9 A schematic diagram of an adapter test compression device for the adapter strain field prediction method of the present invention;

[0049] Fig.10 It is the loss function variation curve of the training set and the validation set in the different model training processes of the adapter strain field prediction method of the present invention. DETAILED DESCRIPTION

[0050] The following will be combined with the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0051] like Figure 1 and Figure 2 As shown, in the embodiment provided by the present invention, the size of the adapter is 150mm×80mm×60mm, the material is polyurethane with superelastic mechanical properties, and the adapter is composed of a series of honeycomb unit cell structures. Therefore, the geometric shape of the adapter is controlled by the unit cell structure. The geometric shape of the unit cell structure of the adapter is controlled by 6 design parameters L1 to L6, so the design variables of the adapter are the design parameters L1 to L6 of the unit cell structure, and the number of design variables is 6.

[0052] like Figure 3 As shown, the present invention provides an adapter strain field prediction method based on information fusion feature equalization operator, comprising the following steps:

[0053] S1. Obtain the geometric characteristics, design variables, and design variable variation range of the adapter, and the strain field information is the value to be predicted.

[0054] S2. Establish an adapter geometric model according to the design variables of the adapter, and build a metamaterial statics finite element model based on the adapter geometric model. Use the Latin hypercube sampling method to obtain sample points, and use the metamaterial statics finite element model to generate a training data set according to simulation.

[0055] like Figure 4 As shown, specifically, the establishment of a metamaterial static finite element model is described with a specific embodiment:

[0056] Use SOLID186 elements to mesh the z = 40 mm end surface;

[0057] The mesh of the metamaterial static finite element model is obtained by sweeping along the z direction, and the lower end surface is set as a fixed support constraint;

[0058] All nodes of the upper end surface are coupled with the node P, and a longitudinal load is applied to the node P, so as to achieve the effect of applying a uniformly distributed load to the upper end surface.

[0059] By using SOLID186 units to mesh the z=40mm end face, the geometric details are fully captured, and the mesh is generated by sweeping along the z direction. The complete metamaterial static finite element model of the adapter structure is constructed quickly and efficiently. The lower end face is set as a fixed constraint and a longitudinal load is applied to the node P. The purpose of simulating the actual working load is achieved, and the mechanical response of the adapter under actual working conditions is effectively simulated.

[0060] In one embodiment of the present application, an independence analysis is performed on the metamaterial statics finite element model, and an optimal grid size is set in the metamaterial statics finite element model to eliminate the influence of the grid size of the metamaterial statics finite element model on the simulation results.

[0061] Preferably, the optimal mesh size is 0.75 mm.

[0062] Furthermore, the use of the metamaterial statics finite element model to simulate and generate a training data set specifically includes:

[0063] The Latin hypercube sampling method is used to generate multiple adapter configurations as the first training set based on the variation range of the design variables of the adapter;

[0064] The first training set is batch simulated using a metamaterial statics finite element model to obtain the strain field of the adapter at each sample point;

[0065] The strain field at each sample point is preprocessed to obtain a training data set.

[0066] By generating multiple adapter configurations as the first training set through the Latin hypercube sampling method, the variation range of the design variables can be covered more efficiently and evenly, reducing the number of samples required for training while ensuring the representativeness and richness of the training data. The first training set can be batch simulated using the metamaterial statics finite element model, which can greatly improve the efficiency of data generation. The quality of the training data can be enhanced by preprocessing the strain field data at each sample point, thereby improving the accuracy of the final strain field prediction.

[0067] The generation of the first training set is described with a specific embodiment:

[0068] The 183 adapter configurations generated by the Latin supersampling method with the design variable variation range of the adapter are used as the second training set, that is, the second training set includes 183 sample points;

[0069] A total of 20 loads of different sizes, {125, 250, …, 2500} N, were applied to each configuration to obtain the first training set, and the training set configuration, validation set configuration, and test set configuration were divided into 8:1:1 ratios, that is, the total number of samples in the first training set was 3660.

[0070] In one embodiment of the present application, MATLAB drives ANSYS Mechanical to automatically perform a series of operations such as meshing, setting boundary conditions and loads, and calculating and solving to obtain simulation results, namely, the strain field of the adapter at each sample point.

[0071] like Figure 5 As shown, further, the preprocessing includes:

[0072] The intermediate nodes of the strain field of the adapter at each sample point are removed, and the strain field of the adapter at each sample point is downsampled to obtain a sparse strain field;

[0073] According to the symmetry of the adapter, half of the strain field after rarefaction is used as the training dataset.

[0074] It can be understood that by batch simulation of the first training set through the metamaterial statics finite element model, the strain field of the adapter at each sample point is obtained, that is, a large number of nodes of the adapter configuration are obtained using the metamaterial statics finite element model, and each node has a corresponding strain value. The strain field is a set formed by the node coordinates and the strain value, that is, the strain field is data in the form of a point cloud.

[0075] After batch simulation of the first training set using the metamaterial statics finite element model, the strain field under the sample points is obtained, in which the strain field contains dense grid nodes and intermediate nodes. Therefore, it is necessary to remove the intermediate nodes of the strain field and downsample the strain field to remove redundant information. According to the symmetry of the adapter during compression, half of the strain field is taken as the training data set.

[0076] By removing the intermediate nodes of the strain field at each sample point, the number of data points is greatly reduced, thereby reducing the computational burden of subsequent training. At the same time, the strain field of the adapter at each sample point is downsampled to further sparse the data, retaining the strain values ​​of key feature points and optimizing training efficiency. The symmetry of the adaptation and structure is used to use half of the sparse strain field as the training data set, thereby reducing the number of training samples and further improving the training efficiency without losing accuracy.

[0077] Specifically, the preprocessing is described with a specific embodiment:

[0078] The first training set is batch simulated using the metamaterial static finite element model. The simulation result of the adapter with configuration parameters of [7.92, 4.64, 3.16, 17.91, 9.24, 3.11] is taken as an example. The number of grids is 23305 and the number of nodes is 119917.

[0079] Since the SOLID186 unit type is a hexahedral structure with 20 nodes, there are "middle points" at the center of the face and the center of the edge, and the middle points are only used in the calculation process, and their values ​​are empty. Therefore, the middle nodes of the strain field under the sample point are first removed, and the number of nodes after removal is 80532;

[0080] Considering that the strain change in the z direction is extremely small, only the points on the front face are selected;

[0081] Since the nodes are denser when the number of grids is 6, there is redundant information, which will greatly increase the training time of the model. Therefore, the strain field under the sample point is downsampled to obtain a sparse strain field. The number of vertices in the sparse strain field is 3599.

[0082] According to the symmetry of the adapter during compression, only half of the sparse strain field is used for prediction. During the model reasoning process, the prediction results of the other half of the strain field can be obtained according to the symmetry, and finally the strain fields in the x and y directions are obtained.

[0083] S3. Construct a neural network model based on information fusion feature equalization operator, and use a training data set to train the neural network model based on information fusion feature equalization operator to obtain a trained neural network model based on information fusion feature equalization operator.

[0084] Specifically, the information fusion feature equalization operator neural network model is based on a deep neural operator network, and includes an information fusion module, a feature equalization learning module and a Fourier feature embedding layer. The information fusion module backbone network and branch network respectively include n hidden layers and one embedding layer.

[0085] Different branch networks of the deep neural operator network are used to capture different features of the input data, and the information fusion module is used to effectively integrate the feature outputs of different branch networks to enhance the richness of feature expression. The feature energy concentration learning module is used to adaptively optimize the feature weights of each branch network to improve the effect of feature fusion. A Fourier feature embedding layer is added to the end of the neural network model based on information fusion feature equalization operator to map the data to the frequency domain space, and the frequency domain features are input together with the spatial domain features into the neural network model based on information fusion feature equalization operator to improve the ability of the neural network model based on information fusion feature equalization operator to capture the strain field distribution.

[0086] Furthermore, the training process of the information fusion feature equalization operator neural network model is as follows:

[0087] S31, load and design parameters P are input into the two branch networks for feature extraction, and the characteristic vectors of load and design parameters are obtained respectively. and The two branch networks are the first branch network and the second branch network. Indicates the first branch network from the load The Pth feature extracted from represents the Pth feature extracted by the second branch network from the design parameter P;

[0088] S32, input the grid node coordinate y into each backbone network and obtain the corresponding output feature Sum the output features to get the fused output features in represents the Pth feature output by the kth backbone network;

[0089] S33, perform matrix multiplication calculation on the features of each layer of the corresponding backbone network and branch network, so that the features between each layer are integrated with each other, and the inner product of the three vectors is taken to obtain the final output result

[0090] S34, use the sine function as the activation function to construct the Fourier feature embedding layer, and use the Fourier feature embedding layer to the final output result After processing, high-frequency strain field characteristics are obtained.

[0091] Furthermore, step S33 specifically includes:

[0092] The feature vector [t1,…,t P ] T and the eigenvector Perform element-by-element multiplication and sum to obtain the first eigenvector. The calculation formula of the first eigenvector is:

[0093]

[0094] Where y represents the coordinate of the adapter point, Represents the mth function vector The number of sampling points, b0 represents the bias term, represents the kth eigenvalue extracted from the mth branch network, P represents the design parameter, t k (y) represents the kth feature vector of adapter point y extracted by the backbone network.

[0095] like Figure 6 As shown, in one embodiment of the present application, the information fusion module includes a backbone network and two branch networks.

[0096] Specifically, the two branch networks include branch network 1 and branch network 2, and the two branch networks are used to load the adapter respectively. And design parameters p for feature extraction, that is, load Input to branch network 1, branch network 2 and main network, the hidden layer 1 and embedding layer in branch network 1 respectively load Perform feature extraction. The hidden layer 1 and the embedding layer in the branch network 2 respectively extract features of the design parameter p, and the hidden layer 1 and the embedding layer in the main network respectively extract features of the coordinate y of the adapter point;

[0097] The features extracted from the embedded layer in the branch network 1, the features extracted from the embedded layer in the main network, and the features extracted from the hidden layer 1 in the branch network 1 are fused to obtain feature F1, and feature F1 is input into the hidden layer 2 of the branch network 1, and the features extracted from the embedded layer in the branch network 1, the features extracted from the embedded layer in the main network, and the features extracted from the hidden layer 2 are fused to obtain feature F2, and so on, until all hidden layers n in the branch network 1 are feature extracted, and the features extracted from the embedded layer in the branch network 1, the features extracted from the embedded layer in the main network, and the features extracted from the hidden layer n are fused to obtain the load feature Fn;

[0098] The features extracted from the embedded layer in the branch network 2, the features extracted from the embedded layer in the main network, and the features extracted from the hidden layer 1 in the branch network 2 are fused to obtain feature P1, and feature P1 is input into the hidden layer 2 of the branch network 2, and the features extracted from the embedded layer in the branch network 2, the features extracted from the embedded layer in the main network, and the features extracted from the hidden layer 2 are fused to obtain feature P2, and so on, until all hidden layers n in the branch network 2 are feature extracted, and the features extracted from the embedded layer in the branch network 2, the features extracted from the embedded layer in the main network, and the features extracted from the hidden layer n are fused to obtain the design variable Pn;

[0099] The features extracted from the embedding layer in the main network, the features extracted from the embedding layer in the branch network 1, the features extracted from the embedding layer in the branch network 2, and the features extracted from the hidden layer 1 in the main network are fused to obtain feature T1; feature T1 is input into the hidden layer 2 of the main network, and the features extracted from the embedding layer in the main network, the features extracted from the embedding layer in the branch network 1, the features extracted from the embedding layer in the branch network 2, and the features extracted from the hidden layer 2 are fused to obtain feature T2, and so on, until all hidden layers n in the main network are feature extracted, and the features extracted from the embedding layer in the main network, the features extracted from the embedding layer in the branch network 1, the features extracted from the embedding layer in the branch network 2, and the features extracted from the hidden layer n are fused to obtain the third feature vector Tn;

[0100] The load characteristic Fn, design variable Pn and the third characteristic vector Tn are combined to obtain the final output result

[0101] like Figure 7 As shown, specifically, the calculation formula of the embedding layer is:

[0102]

[0103] E p =σW p l+b p

[0104] E y =σW y l+b y

[0105] in, E p and E y Respectively represent strain load The feature embedding vector of the strain design parameter p and the coordinate y of the strain adapter, σ represents the nonlinear activation function, W p and W y Respectively represent strain load The weight matrix of the feature embedding vector of the strain design parameter p and the coordinate y of the strain adapter, b p and b y Respectively represent strain load The strain design parameter p and the offset term of the coordinate y of the strain adapter.

[0106] The calculation formula for the first hidden layer (i.e. hidden layer 1) is:

[0107]

[0108] and Respectively represent strain load The strain design parameter p and the coordinate y of the strain adapter are the activation outputs of the first hidden layer, and Respectively represent strain load The strain design parameters p and the coordinates y of the strain adapter are mapped to the weight matrix of the first hidden layer, and Respectively represent strain load The strain design parameters p and the coordinates y of the strain adapter are the bias terms in the first hidden layer.

[0109] The calculation formula for the first output layer (i.e. the output of hidden layer 1) is:

[0110]

[0111] and Respectively represent strain load The strain design parameters p and the coordinates y of the strain adapter are the outputs of the first hidden layer.

[0112] The calculation formula for the kth hidden layer is:

[0113]

[0114]

[0115] and Respectively represent strain load The strain design parameters p and the coordinates y of the strain adapter are the activation outputs of the kth hidden layer, and Respectively represent strain load The strain design parameters p and the coordinates y of the strain adapter are bias terms in the kth hidden layer.

[0116] The calculation formula of the k-th output layer is:

[0117]

[0118] and Respectively represent strain load The strain design parameters p and the coordinates y of the strain adapter are the outputs of the kth hidden layer.

[0119] The calculation formula of the output layer (the output of the information fusion module output layer) is:

[0120]

[0121] and Respectively represent strain load The strain design parameter p and the coordinate y of the strain adapter are output at the activation layer, and Represents the strain load of n-1 layer The strain design parameters p and the coordinates y of the strain adapter are mapped to the weight matrix of the output layer, and Respectively represent strain load The strain design parameters p and the coordinates y of the strain adapter are output at the n-1 output layer, and Respectively represent strain load The strain design parameter p and the coordinate y of the strain adapter are biased in the output layer.

[0122] like Figure 8 As shown in the figure, the feature balance learning module includes M backbone networks, 2 branch networks, 2 embedding layers and a shared embedding layer. The two branch networks are branch network 1 and branch network 2. The corresponding fusion features are input into the first embedding layer and the branch network 1 respectively. The n hidden layers in the branch network 1 perform load Perform feature extraction to obtain feature u;

[0123] The fusion features corresponding to the design variable p are input into the second embedding layer and the branch network 2 respectively. The n hidden layers in the branch network 2 extract features of the variable p in turn to obtain the feature v;

[0124] The coordinate y of the adapter point is input into the shared embedding layer and M backbone networks for feature extraction, where each backbone network obtains the feature The M features Sum the output features The output features Multiply and sum the corresponding elements with feature u and feature v to get the second eigenvector

[0125] Preferably, the number of hidden layers of the branch network is 4, the number of neurons in the hidden layer is 128, 6 trunk network branches are used, the scaling factors are [1, 2, 4, 8, 16, 32] (the scaling factor is a model hyperparameter, but the results of different settings in preliminary experiments are slightly different), and the number of model parameters is 498818.

[0126] Furthermore, the HuberLoss loss function is used in the training process of the neural network model based on the information fusion feature equalization operator, and the calculation formula is:

[0127]

[0128] Among them, δ is a hyperparameter, represents the coordinates of the adapter point output based on the information fusion feature equalization operator neural network model, and y represents the coordinates of the adapter point.

[0129] By adopting HuberLoss as the loss function, the anti-interference ability and stability of the neural network model based on information fusion feature equalization operator for abnormal data are improved, and a hyperparameter δ is introduced into the HuberLoss loss function to control the transition point of L1Loss and L2Loss. By adjusting δ, the robustness of training to outliers and the accuracy of training to normal data are weighed.

[0130] In one embodiment of the present application, evaluation indicators are used to evaluate the training results of the neural network model based on the information fusion feature equalization operator, wherein the evaluation indicators include mean absolute error (MAE), maximum absolute error (MMAE), root mean square error (RMSE) and mean relative error (MRE).

[0131] A variety of evaluation indicators are used to comprehensively evaluate the prediction performance of the neural network model based on the information fusion feature equalization operator from different angles. By comparing the evaluation results of different indicators, the network structure, hyperparameters, etc. can be adjusted in a targeted manner to optimize the neural network model based on the information fusion feature equalization operator.

[0132] Specifically, the calculation formula of the mean absolute error is:

[0133]

[0134] represents the jth output value of the i-th sample predicted by the model, represents the jth true output value of the ith sample, N i Represents the number of output values ​​in the i-th sample

[0135] The maximum absolute error is calculated as:

[0136]

[0137] The formula for calculating the root mean square error is:

[0138]

[0139] The calculation formula for the average relative error is:

[0140]

[0141] It can be understood that when the error between the predicted value and the true value is small and close to 0, the loss function is The MAE loss function can be smoothed. When the error is large, the loss function is Approximately close to MAE; the experiment determines the best training parameters for each model based on the MLE indicator, uses the AdamW optimizer, and adopts the cosine annealing learning rate decay curve. To prevent overfitting, the early stopping mechanism is uniformly adopted in the experiment to monitor the validation set loss of each iteration step in the training. When the loss decreases, the current model is saved; when the loss does not decrease for 100 consecutive iterations, the training is terminated. In the experiment, each group of tests is executed 10 times, and the average value is taken as the final evaluation result.

[0142] S4. Use the trained information fusion feature balance operator-based neural network model to predict the strain field information to be predicted and obtain the real strain field distribution of the adapter.

[0143] This application uses finite element simulation to generate training data, trains the neural network model based on information fusion feature balance operator, enhances the ability to extract and model geometric features, and significantly improves the prediction accuracy. The neural network model based on information fusion feature balance operator can quickly predict the strain field distribution of adapters with different geometric configurations, thereby improving the efficiency of adapter design.

[0144] like Fig. 9 As shown, in one embodiment of the present application, the adapter strain field prediction method based on the information fusion feature equalization operator further includes:

[0145] Perform compression tests on adapters of different configurations and obtain test results;

[0146] The test results are compared with the strain field of the adapter at each sample point to verify the metamaterial static finite element model.

[0147] The strain field prediction method is described with a specific embodiment:

[0148] Two 3 × 3 configurations were used, whose unit cell parameters were [6.30, 2.97, 3.99, 20.00, 14.78, 4.00] and [4.00, 3.09, 5.00, 20.00, 14.37, 4.00] respectively;

[0149] Use a vernier caliper to test the size of the sample, then wipe the surface of the sample clean and place it on the testing machine. In order to ensure uniform force, add a rectangular plate under the disc of the testing machine to completely cover the upper end surface of the adapter. During the compression process, the testing machine maintains a uniform descending speed of 5mm / min, and the maximum pressure is set to 750N. At the same time, in order to reduce the test error, four consecutive loading-unloading processes are used, and the test result of the last loading is taken as the final test result. After the test is completed, the displacement-force change curves of the simulation and the test are compared, as shown in the figure. Figure 5 As shown, the change trends of simulation and test are similar. For sample 1, the maximum displacement measured in the test is 7.71mm, and the simulation displacement is 7.76mm, with a difference of 0.65%. For sample 2, the maximum displacement measured in the test is 11.89mm, and the simulation displacement is 10.97mm, with a difference of 7.74%, which is within the engineering error range. Sample 2 has significantly worse support performance than sample 1. Therefore, when the compression force of sample 2 is less than 400N, it is basically the same as the simulation. After it is greater than 400N, due to the large deformation, uneven force occurs in the test, resulting in increased error.

[0150] Therefore, simulation can be used as an approximate substitute for experiments. By establishing a strain field prediction model for the adapter, the full-domain strain field information of the adapter under different configurations and loads can be provided quickly and accurately, thus shortening the design time of the adapter.

[0151] In one embodiment of the present application, a training data set is used to compare the prediction results of the information fusion feature equalization operator neural network model (IFE-ONet model) and other models, and the comparison results are shown in Table 1. Other models include geometry-aware Fourier neural operator (Geo-FNO), graph neural operator (GNO), multiple-input deep neural operator (MIONet), operator self-attention model (Operator transformer, OFormer), point cloud neural network (Point-cloud neural network, PointNet).

[0152] Table 1

[0153]

[0154]

[0155] As shown in Table 1, the proposed method is used to predict the adapter stress field. The IFE-ONet model proposed in this application is superior to the other six comparison methods in terms of four error evaluation indicators: MAE, MMAE, RMSE and MRE. It also has the following advantages: (1) The performance of the IFE-ONet model is the most stable, and the MRE index of the model is optimal at different training sample points and different resolutions; (2) The IFE-ONet model has strong scalability. With the increase of training sample points or resolution, the performance of the model shows a steady growth trend without fluctuation; (3) IFE-ONet has low requirements on the number of sample points and resolution. Even with a small number of sample points or resolution, the model can achieve an error of about 20%, which is significantly better than other comparison methods. At the same time, the ablation experiment was used to verify the effectiveness of the three modules proposed.

[0156] Depend on Fig.10 It can be seen that the convergence speeds of IFE-ONet, Geo-FNO and OFormer are relatively close, and significantly faster than GNO, MIONet and PointNet. From the perspective of convergence, on the training set, OFormer has the lowest loss, followed by IFE-ONet, but very close to OFormer, with a loss of , while MIONet has the worst convergence, significantly higher than the other five methods. On the validation set, IFE-ONet has the lowest loss value, and MIONet still has the largest.

[0157] The convergence speed and convergence of IFE-ONet and OFormer are close, and IFE-ONet is slightly better. Due to the characteristics of its shallow network, MIONet has insufficient fitting ability, and its convergence and convergence speed are the worst. The convergence speed and convergence of GNO and PointNet are close.

[0158] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. An adapter strain field prediction method based on information fusion feature equalization operator, characterized in that: The following steps are involved: S1, obtaining the geometric characteristics, design variables, and design variable variation range of the adapter, and the strain field information is the value to be predicted; S2. Establish an adapter geometry model according to the design variables of the adapter, and construct a metamaterial statics finite element model based on the adapter geometry model, use the Latin hypercube sampling method to obtain sample points, and use the metamaterial statics finite element model to generate a training data set according to simulation; S3, constructing a neural network model based on information fusion feature equalization operator, and using a training data set to train the neural network model based on information fusion feature equalization operator to obtain a trained neural network model based on information fusion feature equalization operator; S4, using the trained information fusion feature balance operator-based neural network model to predict the strain field information to be predicted, and obtaining the strain field distribution of the adapter; The information fusion feature equalization operator neural network model is based on a deep neural operator network, and includes an information fusion module, a feature equalization learning module, and a Fourier feature embedding layer. The information fusion module backbone network and branch network include n hidden layers and one embedding layer respectively. The training process of the information fusion feature equalization operator neural network model is as follows: S31, load and design parameters Input two branch networks for feature extraction respectively, and obtain the feature vectors of load and design parameters respectively. and , where the two branch networks are the first branch network and the second branch network. Indicates the first branch network from the load The Pth feature extracted from Indicates the design parameters of the second branch network The Pth feature extracted from S32, input the grid node coordinate y into each backbone network and obtain the corresponding output feature , sum the output features to get the fused output features ,in represents the Pth feature output by the kth backbone network; S33, perform matrix multiplication calculation on the features of each layer of the corresponding backbone network and branch network, so that the features between each layer are integrated with each other, and the inner product of the three vectors is taken to obtain the final output result ; S34, use the sine function as the activation function to construct the Fourier feature embedding layer, and use the Fourier feature embedding layer to the final output result After processing, high-frequency strain field characteristics are obtained.

2. The adapter strain field prediction method based on information fusion feature equalization operator according to claim 1, characterized in that: Step S32 specifically includes: The backbone network is expanded into M sub-networks, each of which is responsible for learning a certain part of the features. The grid node coordinate y is input into each backbone network, and the corresponding output features are obtained. , sum the output features to get the fused output features .

3. The adapter strain field prediction method based on information fusion feature equalization operator as claimed in claim 2, characterized in that: Step S33 specifically includes: The feature vector and the eigenvector Perform element-by-element multiplication and sum to obtain the first eigenvector. The calculation formula of the first eigenvector is: ， in, represents the coordinates of the adapter point, Represents the mth function vector The number of sampling points, represents the bias term, represents the kth eigenvalue extracted by the mth branch network, represents the design parameters, Represents the k-th feature vector of adapter point y extracted by the backbone network.

4. The adapter strain field prediction method based on information fusion feature equalization operator as claimed in claim 3, characterized in that: The HuberLoss loss function is used in the training process of the neural network model based on the information fusion feature equalization operator. The calculation formula is: ； in, is a hyperparameter, represents the coordinates of the adapter point output based on the information fusion feature equalization operator neural network model, Indicates the coordinates of the adapter point.

5. The adapter strain field prediction method based on information fusion feature equalization operator according to claim 1, characterized in that: The use of the metamaterial statics finite element model to simulate and generate a training data set specifically includes: The Latin hypercube sampling method is used to generate multiple adapter configurations as the first training set based on the variation range of the design variables of the adapter; The first training set is batch simulated using a metamaterial statics finite element model to obtain the strain field of the adapter at each sample point; The strain field at each sample point is preprocessed to obtain a training data set.

6. The adapter strain field prediction method based on information fusion feature equalization operator according to claim 5, characterized in that: The pre-processing comprises: The intermediate nodes of the strain field of the adapter at each sample point are removed, and the strain field of the adapter at each sample point is downsampled to obtain a sparse strain field; According to the symmetry of the adapter, half of the strain field after rarefaction is used as the training dataset.

7. The adapter strain field prediction method based on information fusion feature equalization operator according to claim 1, characterized in that: The adapter strain field prediction method based on information fusion feature equalization operator also includes: Perform compression tests on adapters of different configurations and obtain test results; The test results are compared with the strain field of the adapter at each sample point to verify the metamaterial static finite element model.

Images