A neural network-based method for predicting curing deformation of a typical structure of a composite material

Abstract

The application provides a composite material typical structure curing deformation prediction method based on a neural network and belongs to the field of composite material structures; specifically, for a typical structure part, the type and value range of parameters affecting curing quality are summarized; based on a Monte Carlo method, a plurality of groups of input parameters are randomly generated and introduced into a finite element analysis method to perform curing deformation calculation and generate a preliminary data set; then, a feature selection algorithm is used to classify the preliminary data to generate a new data set containing only important features; next, an ANN neural network architecture is established, and parameters with greater influence on curing deformation in less important features are divided; the important features are superimposed with different less important features as common inputs to train the network architecture; finally, after a new data set containing only important features is obtained for a new typical structure part, the trained ANN neural network is input to obtain a prediction result; and the application improves the calculation efficiency while ensuring the prediction accuracy and precision.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of composite material structure manufacturing, and relates to a curing deformation prediction method for typical composite material structures based on a neural network. BACKGROUND

[0002] Composite materials are increasingly widely applied in aviation, automobiles, ships and light industry, but the curing deformation problem caused by residual stress existing in the manufacturing process of thermosetting composite materials will cause the structure to be unable to be assembled or even scrapped. The curing deformation introduced in the manufacturing process is affected by various factors, such as curing temperature, mold action, material properties and curing pressure. At present, the research on the influencing factors of curing deformation still stays in qualitative or quantitative analysis by using traditional tests and finite element simulation methods. Whether the test or the finite element simulation will consume a large amount of cost and is very low in efficiency, and the fast and accurate prediction of the curing deformation cannot be realized.

[0003] With the rapid development of computer technology and the field of artificial intelligence, machine learning can replace many traditional solutions to scientific research and engineering problems in a simpler way. According to professional knowledge, the input variables are reasonably selected or established, and a suitable machine learning algorithm is selected to solve the problem of curing deformation prediction.

[0004] In the field of composite material research, many scholars have successfully used machine learning methods to solve problems. Some scholars use ANN to predict the fiber direction in long fiber compression molded composite materials, and obtain the fiber angle through information at different positions; such as reference 1: Application of Artificial Neural Networks to predict fibre orientation in long fibre compression moulded composite materials. Some scholars predict the curing springback deformation of asymmetrically laminated laminates based on ANN, and use FEM to calculate the data set for training, such as reference 2: Rapid prediction and inverse design of distortion behaviors of composite materials using artificial neural networks. Some scholars use decision tree regression model to predict the elastic modulus of single fiber based on macro modulus. Cross-scale FEM model is used to obtain the data set for training, such as reference 3: Prediction of mechanical properties of carbon fiber based on cross-scale FEM and machine learning.

[0005] The factors affecting the curing deformation are summarized, and the random forest method and lasso method are used to classify the related factors to determine the degree of influence. The constructed ANN artificial neural network is used to complete the final classification of features, and the method of artificial intelligence is used to realize the rapid and accurate prediction of curing deformation. SUMMARY

[0006] In the prior art, the prediction of curing deformation of composite materials requires complex modeling, and the calculation is time-consuming, which limits the application in the simulation of curing of composite material structure; and the traditional trial and error method of composite material structure manufacturing process has high cost, and cannot realize efficient prediction of curing deformation of composite materials, and cannot meet the requirements of composite material structure simulation. The present application proposes a high-efficiency prediction method for curing deformation of typical composite structure based on neural network, which replaces the complex modeling process and a large number of iterative calculations by using neural network, reduces the calculation time, ensures the prediction accuracy, realizes the classification of curing factors of composite materials, completes the rapid and efficient simulation of curing deformation of composite material structure, and reduces the production cost of composite material structure.

[0007] The high-efficiency prediction method for curing of typical composite structure based on neural network comprises the following steps:

[0008] Step one, for composite material typical structure, determine the parameter type and the value range of each parameter which affect the curing quality of the structure;

[0009] Step two, based on Monte Carlo method, generate several groups of parameter data sets randomly based on the value range of each parameter;

[0010] The parameter data set includes mold thermal expansion coefficient, mold and component interface friction coefficient, curing pressure, peak temperature, single layer thickness, number of layers, four layer proportion parameters, nine laminate parameters, corner angle, single layer chemical shrinkage, single layer mechanical parameters, single layer thermal expansion coefficient;

[0011] Where the stress and strain of the laminate are related by the laminate stiffness matrix:

[0012]

[0013] Where A is the in-plane stiffness matrix relating the in-plane forces to the mid-surface strain, B is the coupling stiffness matrix relating the in-plane strain to the bending-torsional internal forces, or the bending-torsional deformation to the in-plane forces, D is the bending stiffness matrix relating the bending curvature, torsional curvature and bending-torsional internal forces, ε is the mid-surface strain, and k is the mid-surface curvature;

[0014] The material stiffness invariant U and the stiffness attribute Q are calculated by the following formula:

[0015]

[0016] Where E 11 is the axial modulus of the single layer, E 22 is the transverse modulus of the single layer, G 12 is the shear modulus of the single layer, and v 12 is the main Poisson's ratio of the single layer;

[0017] The laminate parameters are calculated by the following formula:

[0018]

[0019] Where θ represents the layer direction at the thickness Z position, the origin of the Z direction coordinate is at the mid-surface; h is the laminate thickness, and Z k is the position of the kth layer;

[0020] Step three, call ABAQUS finite element analysis method, automatically input parameter data set, calculate the curing deformation corresponding to different parameters, and perform normalization processing to generate the preliminary data set of neural network;

[0021] Step four, using feature selection algorithm to classify the preliminary data, quantifying the influence degree of each factor, the data is divided into important features, secondary important features and unimportant features, and a new data set containing only important features is generated using the finite element method;

[0022] Step five, establish the ANN neural network architecture, use the control variable method to determine the number of neurons and layers, the selection of activation function, the selection of optimizer, and the setting of loss function;

[0023] Step six, using the established ANN neural network architecture, divide the parameters in the secondary important features that have greater impact on the curing deformation, and complete the final feature classification; using the important features superimposed with different secondary important features as common input, train the ANN neural network architecture;

[0024] Divide the secondary important features into important features, use the important features superimposed with different secondary important features as common input, train the ANN neural network architecture, and input the new data set into the ANN neural network architecture, get the prediction result as the benchmark; get the optimal parameters of the ANN neural network architecture;

[0025] Step seven, repeat steps two to four for a new typical structural part, directly input the new data set containing only important features of the structure into the trained ANN neural network to obtain the prediction result of the curing deformation of the structure;

[0026] The advantages and positive effects of the present application are:

[0027] (1) A composite material typical structure curing deformation prediction method based on neural network, compared with the traditional multi-physical field coupling based composite material structure curing deformation simulation calculation method, the material parameter type is complex, the finite element grid number is large, the calculation efficiency is low, and a large number of iterative calculations are needed, the present application replaces a large number of iterative calculations through neural network, which significantly improves the efficiency of composite material structure curing deformation simulation.

[0028] (2) A composite material typical structure curing deformation prediction method based on neural network, using feature selection method, classifying the factors affecting curing deformation, extracting the factors that have greater impact on curing deformation, and combining the optimized ANN neural network model to obtain accurate prediction of curing deformation, which improves the calculation efficiency while ensuring the accuracy of the prediction of the curing deformation of the composite material typical structure. BRIEF DESCRIPTION OF DRAWINGS

[0029] Figure 1 is a flowchart of a composite material typical structure curing deformation prediction method based on neural network of the present application;

[0030] Figure 2L type structure curing springback deformation influencing factor of the present application;

[0031] Figure 3 L type structure finite element model established in the present application;

[0032] Figure 4 Data set generation process chart simulated by finite element method in the present application;

[0033] Figure 5 Random forest regression feature selection process chart in the present application;

[0034] Figure 6 Lasso regression feature selection chart in the present application;

[0035] Figure 7 Artificial neural network structure schematic diagram adopted in the present application;

[0036] Figure 8 Prediction accuracy comparison of different data amounts and input features in the present application;

[0037] Figure 9 Accuracy result chart of each secondary important factor after superposition in the present application. DETAILED DESCRIPTION

[0038] In order to facilitate those skilled in the art to understand and implement the present application, the present application is further described in detail below in combination with the drawings and examples. Obviously, the described examples are only partial examples of the present application, but not all examples. Based on the examples in the present application, all other examples obtained by those skilled in the art without creative labor should belong to the protection scope of the present application.

[0039] The present application proposes a composite material typical structure curing deformation prediction method based on neural network, which quickly and accurately predicts composite materials through neural network; as shown in the figure, including the following steps: Figure 1

[0040] Step one, for the composite material typical structure, determine the parameter type affecting the curing quality of the structure and the value range of each parameter;

[0041] The parameter type includes mold, process system, typical structure geometric factor and related factors of material properties; the value range of each parameter is determined according to the summary literature and the structure process actually adopted in the composite material forming process.

[0042] Step two, based on the Monte Carlo method, randomly generate several groups of parameter data sets based on the value range of each parameter;

[0043] ​The parameter data set contains material parameters, structure parameters and process parameters, and specifically includes a mold thermal expansion coefficient, a mold and component interface friction coefficient, a curing pressure, a peak temperature, a single layer thickness, a number of layups, four layup proportion parameters, nine laminate parameters, a corner angle, a single layer chemical shrinkage rate, single layer mechanical parameters, and a single layer thermal expansion coefficient.

[0044] Step three, calling the ABAQUS finite element analysis method, automatically inputting the parameter data set, calculating the curing deformation corresponding to different parameters, and performing normalization processing to generate a preliminary data set of the neural network;

[0045] The ABAQUS finite element analysis software is used for curing deformation simulation, a PYTHON script is written to call the ABAQUS program, material parameters, structure parameters and process parameters are automatically inputted, and the curing deformation corresponding to different parameters is calculated, and then the maximum and minimum value normalization method is used to process the results to generate a preliminary data set of the neural network;

[0046] The normalization processing formula is:

[0047]

[0048] Where x * is the data after transformation of a certain feature; x is the original data before transformation of the feature; x max and x min are the maximum and minimum values of the original data of the feature, respectively.

[0049] Step four, using a feature selection algorithm to classify the preliminary data, quantifying the influence degree of each factor to divide the data into important features, secondary important features and unimportant features, only considering the influence of important feature values on the curing deformation, and using a finite element method to generate a new data set containing only important features;

[0050] Step five, establishing an ANN neural network architecture, using a control variable method to determine the number of neurons and layers, an activation function selection, an optimizer selection, and a loss function setting;

[0051] The ANN architecture constructed in the application is a one-layer hidden layer with 10 neurons, the activation function is ReLU (Rectified Linear Unit), the output layer activation function is Linear, and the optimizer is Adam.

[0052] Step six, using the constructed ANN neural network architecture, dividing the parameters in the secondary important features that have a greater influence on the curing deformation, completing the final feature classification, and using important features superimposed with different secondary important features as common inputs to train the ANN neural network architecture;

[0053] The data set obtained by using only important features is input into the ANN neural network architecture for training and prediction. The training and prediction of different important features are calculated in turn, and the further screened more accurate feature classification is obtained.

[0054] The data set of all important features superimposed on the secondary important features can improve the accuracy of the ANN neural network architecture. Therefore, the secondary important features are divided into important features, and the solidification deformation data set corresponding to the new important features is calculated using the finite element method as the final data set. The ANN neural network model is trained to obtain the final ANN model parameters.

[0055] The trained ANN neural network architecture is used for further feature classification to continue screening out parameters that have a greater impact on solidification deformation from important features.

[0056] Step seven, repeat steps two to four for a new typical structural part. The new data set containing only important features of the structure is directly input into the trained ANN neural network to obtain the prediction result of the solidification deformation of the structure.

[0057] The trained ANN neural network architecture is directly input into other parts containing corner features to obtain the prediction result of the solidification corner springback deformation of the structure.

[0058] Embodiment:

[0059] Taking the solidification deformation of an L-shaped typical structure as an example;

[0060] Step one, according to the existing literature and the structure process parameters used in actual production, determine the parameter types affecting the solidification quality of the composite structure and the selection range of each parameter.

[0061] According to the actual manufacturing parameters and value range of the composite material, all the design parameters of the composite L-shaped structure are summarized, as shown in Figure 2 The key parameters mainly include mold, process system, L-shaped structure geometric factors and material attribute related factors. According to the actual forming process of the composite material and the research on the solidification process in the related research literature, the value range of each factor is determined.

[0062] Step two, use the Monte Carlo method to generate input parameters, write scripts to import ABAQUS finite element software for solidification deformation calculation, and obtain the springback angle of the L-shaped structure as the preliminary data set.

[0063] The Monte-Carlo method is used to generate the data set for finite element calculation based on the value range of each solidification parameter; as shown in Figure 4 The specific process is as follows:

[0064] The random numbers of the independent variables of the solidification factors are generated based on the Monte-Carlo method using a python script, and a total of 29 input parameters are included;

[0065] The value range of the parameters includes the mold thermal expansion coefficient (1.5E-6-24E-6), the mold and component interface friction coefficient (0-0.6), the solidification pressure (0.1-0.7 MPa), the peak temperature (350 K-460 K), the single layer thickness (0.1-0.2 mm), the number of layers (8-100), four layer proportion parameters, nine laminate parameters, the corner angle (10-170 degrees), the single layer chemical shrinkage rate (β1=-800--100, β2=-20000--5000), the single layer mechanical parameter (E1=10-260 GPa, E2=1-20 GPa, G 12 =1-10 GPa, G 23 =1-10 GPa, v 12 =0.3-0.4), the single layer thermal expansion coefficient (α1=-1e-6-1e-5, α2=1e-6-1e-4).

[0066] The layer proportion parameters are the proportions of each layer, i.e., four parameters, although there are only three independent parameters, but in the analysis, four parameters are still retained to facilitate the determination of the influence of each layer. The layer sequence is not a continuous variable, therefore, the layer sequence information needs to be converted into a continuous variable, and this conversion needs to adequately reflect the layer sequence information. In the classical laminate theory, the stress and strain of the laminate are related through the laminate stiffness matrix:

[0067]

[0068] where A is the in-plane stiffness matrix relating the in-plane forces to the mid-plane strain, B is the coupling stiffness matrix relating the in-plane strain to the bending-torsional internal forces, or the bending-torsional deformation to the in-plane forces, D is the bending stiffness matrix relating the bending curvature, torsional curvature, and bending-torsional internal forces, ε is the mid-plane strain, and k is the mid-plane curvature.

[0069] The three matrices are symmetric matrices, and the 6X6 total matrix formed by them is also a symmetric matrix, which can be calculated from the material stiffness invariant U and the 12 laminate parameters:

[0070]

[0071] The material stiffness invariant U and the stiffness property Q are calculated by the following formula:

[0072]

[0073] where E 11 is the axial modulus of the single layer, E 22 is the transverse modulus of the single layer, and G12 is the single ply shear modulus, v 12 is the single ply major Poisson's ratio.

[0074] 12 laminates parameters are calculated by the following formula:

[0075]

[0076] where θ represents the ply orientation at the thickness Z location (the Z coordinate origin is at the midplane). h is the laminate thickness, Z k is the kth ply location; when the ply angle is only 0, 90, 45 degrees, the fourth component is all 0.

[0077] When there is an asymmetric and non-equilibrium ply case, there are actually only 9 independent laminate parameters. The 9 laminate parameters are regarded as the quantitative representation of the ply sequence, and can be calculated by the formula when the known ply is input. The ply ratio and the ply sequence can be automatically calculated by the known ply, so the input variables corresponding to them are the specific ply settings.

[0078] Step three, the feature selection algorithm is used to classify the preliminary data set, input the ANN neural network model, while ensuring the prediction accuracy and precision, also improve the calculation efficiency, can realize the efficient simulation of the curing deformation of the composite L-shaped structure based on multi-physical field coupling.

[0079] In order to eliminate the dimensional influence between the features, the input data set needs to be normalized.

[0080] Since the selection of feature data in the embodiment is based on the Monte Carlo simulation random acquisition based on uniform distribution, first, the maximum and minimum value method is selected to transform each feature data;

[0081] The embodiment of the application adopts a python script, and a total of 8000 groups of data curing parameter variable data are generated.

[0082] Step four, using ABAQUS finite element analysis software, the data after normalization of the L-shaped structure is simulated and calculated for curing deformation.

[0083] For the curing simulation of the composite laminated plate, three-dimensional solid element modeling calculation is used, and the use of C3D20R second-order reduced integration three-dimensional solid element to simulate the composite structure can avoid the generation of shear self-locking, so that the calculation is more stable.

[0084] The mold unit type is C3D8R first-order linear reduced integration three-dimensional solid element. The interaction between the mold and the structure is set as friction, and the friction coefficient is corrected through the test result. The analysis step is a general quasi-static analysis step, and the mold is removed after curing is completed by using the Model change function, and demolding is completed. The finite element model is as followsFigure 3 As shown, the material properties are implemented through user-defined subroutines UMAT and UEXPAN.

[0085] Step five, using random forest regression and Lasso regression algorithm, the key parameters are selected into important features, secondary important features and unimportant features.

[0086] The present application divides a large number of key parameters into important features, secondary important features and unimportant features through feature selection. Important features refer to decisive role in the accuracy of the prediction result, and unimportant features refer to little effect on the accuracy of the prediction result; the secondary important features are between the two.

[0087] The feature selection process of random forest regression is as shown in Figure 5 First, input the data, complete data standardization, find the optimal model hyperparameter through cross-validation, randomly divide 70% training data set and 30% test data set for training, return feature importance data, and this process is completed 5 times to take the average value to reduce the influence of randomness.

[0088] The feature selection process of Lasso regression is as shown in Figure 6 Similar to random forest regression, after standardizing the data set, randomly divide the training set and the test set, select the lambda parameter, train, get the feature coefficient of each lambda parameter, and each lambda parameter is randomly divided into 5 data sets and trained, and the average value is taken as the feature coefficient result corresponding to the lambda parameter.

[0089] First, select a small lambda parameter, and the present application takes 0.0001, at this time the coefficient of the feature that becomes 0 is considered to be an unimportant feature; then take a larger lambda parameter, and the present application takes 0.002, at this time the coefficient of the feature that remains unchanged to 0 is considered to be an important feature; and the remaining is a secondary important feature.

[0090] Combining theoretical analysis, the results of random forest regression algorithm and Lasso regression algorithm are summarized to finally determine a reasonable classification.

[0091] Step six, construct an artificial neural network ANN model, and train using the selected features.

[0092] Feature selection helps to simplify the ANN model, reduce the calculation cost, and reduce the risk of overfitting.

[0093] This embodiment is aimed at the curing deformation of the L-shaped structure of the composite material, selects the open source artificial neural network library Keras for modeling and training, and the ANN architecture established by the present application mainly reflects several key information, including the number of neurons and the number of layers, the selection of activation function, the selection of optimizer and the like. The present application determines the key information of the ANN architecture through control variable method.

[0094] In the embodiment of the present application, the damage initiation model is a three-layer network, including an input layer, a single hidden layer and an output layer, having 29, 10 and 1 neurons respectively, the hidden layer activation function is "ReLU", and the output layer activation function is "Linear". The neural network diagram of the solidification deformation model is shown as Figure 7 .

[0095] The Adam algorithm is used as the optimizer for the established neural network in the embodiment of the present application, and the learning rate is set to 0.05. The data set obtained in step three is used for training. The prediction accuracy of different data amounts and input features is shown as Figure 8 . The prediction accuracy of the trained model in the present example is above 94.9%.

[0096] The accuracy results after superposition of each secondary important factor are shown as Figure 9 . The data amount is selected as 5000. The prediction accuracy using only the important features in the result is 0.968. When another secondary important feature is additionally superimposed, the introduction of the 90° laying ratio, the laminate parameter C2 and the material property parameter E2 can obviously improve the result accuracy and reduce the random fluctuation of the result, so it can be considered that these three features are useful features. The introduction of other features will reduce the result accuracy, which can be considered as noise terms, and is a useless feature.

[0097] A is an important feature, Mould CTE is the mold thermal expansion coefficient, and CTE is the thermal expansion coefficient.

[0098] The above is only the preferred specific implementation of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art within the technical scope of the present application conclusion, according to the technical scheme and the inventive concept of the present application for equivalent replacement or change, should be covered in the protection scope of the present application.

Claims

1. A neural network-based method for predicting curing distortion of a typical structure of a composite material, characterized by, Comprising the following steps: Step one, for the typical structure of composite materials, determine the parameter type that affects the curing quality of the structure and the value range of each parameter; Step two, based on the Monte Carlo method, generate a number of parameter data sets randomly based on the value range of each parameter; The parameter data set includes the mold thermal expansion coefficient, the mold and component interface friction coefficient, the curing pressure, the peak temperature, the single layer thickness, the number of layers, the four layer proportion parameters, the nine laminate parameters, the corner angle, the single layer chemical shrinkage rate, the single layer mechanical parameter, and the single layer thermal expansion coefficient; Wherein the stress and strain of the laminate are related by the laminate stiffness matrix: Wherein A is the in-plane stiffness matrix relating the in-plane force and the mid-plane strain, B is the coupling stiffness matrix relating the in-plane strain and the bending and torsional internal force, or the bending and torsional deformation and the in-plane force, D is the bending stiffness matrix relating the bending rate, the torsion rate and the bending and torsional internal force, ε is the mid-plane strain, and k is the mid-plane curvature; The material stiffness invariant U and the stiffness attribute Q are calculated by the following formula: where E 11 is the single ply axial modulus, E 22 is the single ply transverse modulus, G 12 is the single ply shear modulus, v 12 is the single ply major Poisson's ratio; The laminate parameters are calculated by the following formula: where θ represents the ply orientation at the thickness Z location, with the Z direction coordinate origin at the midplane; h is the laminate thickness, Z k is the position of the kth ply. Step three, call the ABAQUS finite element analysis method, automatically input the parameter data set, calculate the curing deformation corresponding to different parameters, and perform normalization processing to generate a preliminary data set for the neural network; Step four, classify the preliminary data using a feature selection algorithm, quantify the influence of each factor, divide the data into important features, secondary important features and unimportant features, and use the finite element method to generate a new data set containing only important features; Step five, establish an ANN neural network architecture, use the control variable method to determine the number of neurons and layers, the selection of activation function, the selection of optimizer, and the setting of loss function; Step six, use the established ANN neural network architecture to divide the parameters in the secondary important features that have a greater impact on the curing deformation, complete the final feature classification; use the important features and different secondary important features as common inputs to train the ANN neural network architecture; Divide the secondary important features into important features, use the important features and different secondary important features as common inputs to train the ANN neural network architecture, input the new data set into the ANN neural network architecture, and obtain the prediction result as the benchmark; obtain the optimal parameters of the ANN neural network architecture; Step seven, repeat steps two to four for a new typical structure, and directly input the important feature parameters of the structure into the trained ANN neural network to obtain the prediction result of the curing deformation of the structure; The trained ANN neural network architecture is directly input into the parameters of other parts containing corner features to obtain the deformation.

2. The neural network-based composite typical structure curing distortion prediction method of claim 1, wherein, In step one, the parameter types include mold, process system, typical structure geometric factors and material attribute related factors; the value range of each parameter is determined artificially according to the actual molding process of composite materials.

3. The neural network-based composite typical structure curing distortion prediction method of claim 1, wherein, In step three, the normalization processing formula is: where x * is the transformed data for a feature; x is the original data for the feature before transformation; x max and x min are the maximum and minimum values of the original data for the feature, respectively.

4. The neural network-based composite typical structure curing distortion prediction method of claim 1, wherein, In step four, the feature selection algorithm selects the Lasso regression algorithm, selects the λ parameter in the algorithm, trains to obtain the feature coefficients, and takes the average value as the feature coefficient result corresponding to this λ parameter; First, a λ parameter is randomly selected in the range of 0-0.001, and the features whose coefficients become 0 are considered as unimportant features; Then, the λ parameter is taken as 0.002, at this time, the features whose coefficients remain 0 are considered as important features, and the remaining features are considered as less important features.

5. The neural network-based composite typical structure curing distortion prediction method of claim 1, wherein, In step five, the constructed ANN architecture is 1 layer of hidden layer, 10 neurons, the activation function is ReLU, the output layer activation function is Linear, and the optimizer is Adam.

Images