Method and system for forecasting slamming load and structural response of composite sandwich structure

Abstract

The invention discloses a composite material sandwich structure slamming load and structure response forecasting method, which comprises the following steps: a data preprocessing stage: collecting a characteristic value and a target value of a composite material sandwich structure, performing normalization processing to obtain an original label data set, establishing a first part neural network and enhancing data; a physical information neural network learning stage: constructing a physical information neural network as a second part neural network, and training based on a preset network architecture; and a network prediction stage: predicting a central slamming pressure peak value of the composite material sandwich structure and a central tensile micro-strain peak value along the plate length direction by using the trained network. The beneficial effects of the invention are that the method comprises two parts of neural networks, can not only learn the characteristics of existing data, but also follows the physical principle for prediction, effectively improves the generalization ability of the neural networks, and improves the reasonability and accuracy of a prediction result.

Description

Technical Field

[0001] The present invention relates to the technical field of ship and ocean engineering, and particularly relates to a method and system for predicting slamming loads and structural responses of a composite sandwich structure. Background Art

[0002] Unmanned boats have excellent maneuverability, can quickly respond in complex and changeable sea areas, perform various high-difficulty and high-risk tasks, and significantly improve the success rate of military operations, civil rescue and other tasks. At the same time, due to the small scale of unmanned boats, the radar cross-section is relatively small and they can form cluster operations, so they gradually show great potential and broad prospects in the military field. With the rapid development of technology and the maturity of intelligent technology, some developed countries have officially incorporated unmanned boats into the modernization innovation projects of their military weaponry and civil application equipment.

[0003] The emergence and application of high-performance composite materials with multi-phase components and multi-level microstructures have significantly improved the maneuverability of unmanned boats when carrying various equipment to perform tasks in complex sea conditions. However, when a composite unmanned boat sails at high speed in severe sea conditions to perform tasks, there will be a violent relative motion between the hull and the waves, resulting in the bow of the unmanned boat being subjected to severe transient high-amplitude slamming loads. The transient high-amplitude slamming loads will significantly reduce the overall motion performance of the unmanned boat, and may even cause local structural strength damage, plastic deformation or even structural failure. Therefore, in order to ensure that the unmanned boat can successfully complete the specified tasks, its anti-slamming performance must be fully considered in the initial design stage.

[0004] Due to the characteristics of transient high amplitude and multi-field coupling of the slamming problem, with strong non-linearity, the existing slamming theoretical models have assumptions and simplifications in application, such as ignoring fluid viscosity, structural elastic deformation and fluid compressibility, and it is difficult to comprehensively reflect the real slamming phenomenon, resulting in large errors in the prediction results. At the same time, the cost of slamming tests is high, and the universality of test results is also limited. In addition, although numerical simulation methods have been widely used and developed in slamming research, their calculation time is long, especially when dealing with multi-physical field coupling problems, and there are still challenges in terms of calculation efficiency.

[0005] In recent years, intelligent machine learning methods based on data-driven have been increasingly widely applied in various fields. Machine learning learns patterns from data and makes predictions and decisions. Neural network is a special machine learning algorithm, mainly used to construct complex non-linear relationships, and has higher computational efficiency compared with traditional simulation methods. However, the application of neural network methods in the research of non-linear water entry slamming problems is limited and still in the development stage. At the same time, traditional neural network prediction only considers the prediction under data-driven, and the prediction results may violate basic physical laws, resulting in unreasonable prediction results. Traditional neural network prediction is highly dependent on the data set and requires a large amount of labeled data to cover the boundary. Small data is likely to lead to overfitting or underfitting. Summary of the Invention

[0006] The purpose of the present invention is to propose a method and system for predicting slamming loads and structural responses of a composite sandwich structure, aiming at the deficiencies of the existing technology, so as to reduce the error of the prediction results and improve the rationality of the predicted structure.

[0007] The technical solution adopted by the present invention is as follows: A method for predicting slamming loads and structural responses of a composite sandwich structure, the method comprising the following steps: Data preprocessing stage: Collect the characteristic values and target values of the slamming test of the composite sandwich structure and the fluid-structure interaction numerical simulation calculation as the original data set; perform normalization processing on the original data set to obtain the original labeled data set; build the first part of the neural network, and optimize the first part of the optimal neural network based on K-fold cross-validation, learn the non-linear mapping relationship between the characteristic values and target values in the original labeled data set of the composite sandwich structure, and enhance the data using the first part of the neural network. Physical information neural network learning stage: Based on the original labeled data set and the newly added data set, construct a physical information neural network, and perform training based on the preset network architecture to learn the non-linear relationship between all characteristic values and target values. Network prediction stage: Use the trained network to input new characteristic values to predict the peak slamming pressure at the center of the composite sandwich structure and the peak tensile micro-strain along the plate length direction at the center.

[0008] According to the above solution, in the data preprocessing stage, a slamming test is designed for the composite sandwich structure, and combined with fluid-structure interaction simulation, multiple groups of original data are obtained as the original data set. The original data includes characteristic value parameters and target value parameters; the original data set is standardized to obtain the original labeled data set.

[0009] According to the above solution, in the data preprocessing stage, perturbations are applied to the characteristic values in the original labeled data set to generate new characteristic values.

[0010] According to the above scheme, the specific method for enhancing data using the first part of the neural network is to use the new eigenvalue generated by perturbing the original dataset as the input value, and obtain the output value through the first part of the neural network, which is the target value corresponding to the newly added eigenvalue, forming an enhanced dataset, and merging it with the original label dataset for the subsequent training of the second part of the physics-informed neural network.

[0011] According to the above scheme, the specific method for optimizing the first part of the optimal neural network based on K-fold cross-validation is as follows: evenly and non-overlappingly divide the original label dataset into K subsets; in each round of training, select 1 subset from the K subsets as the validation set, and the remaining K - 1 subsets are merged into the training set, and train based on the preset network architecture to obtain a sub-model with specific weights and biases, and calculate the mean square error of the model on the corresponding validation set, and select the model with the smallest mean square error as the first part of the optimal neural network.

[0012] According to the above scheme, in the learning stage of the physics-informed neural network, the neural network is constrained by combining the data loss and the physical loss equations.

[0013] According to the above scheme, the total loss in the physics-informed neural network includes data loss, physical loss, and the L2 regularization term, and its complete expression is: (1); (2); (3); (4); Among them, L total The total loss function in the second part of the neural network; L date_loss is the data loss function, L physics_loss is the physical loss function; L reg is the L2 regularization term; λ d , λ p , λ r are the weight coefficients for controlling the data loss, physical loss, and regularization strength in sequence; N is the number of sample points, y true,i and y pred,i are the predicted value and the true data value of the neural network respectively; F inertia is the inertia term, F stiff is the stiffness term,F fluid is the fluid action term; f is the natural frequency under the equivalent stiffness of the composite sandwich structure, f pred,i is the natural frequency obtained from the predicted value of the second part of the neural network; ω k is the k layer weight parameter of the model; f is the natural frequency under the equivalent stiffness of the composite sandwich structure; f pred,i is the predicted natural frequency under the equivalent stiffness of the composite sandwich structure.

[0014] According to the above scheme, the relevant training parameters for training the physics-informed neural network are: using the Adam optimizer, the learning rate is 0.0001, and the maximum number of training epochs is 3000.

[0015] The present invention also discloses a slamming load and structural response prediction system for a composite sandwich structure, and the system includes: A data preprocessing module, which is used to obtain an original data set by using slamming tests and fluid-structure coupling, normalize the original data set to obtain an original label data set, and add new feature values; A first part of neural network construction and processing module, which is used to build a first part of neural network based on the original label data set, and optimize the first part of the optimal neural network based on K-fold cross-validation to learn the non-linear mapping relationship between the feature values and the target values in the original label data set of the composite sandwich structure; and use the new feature values added by the first-order part of the neural network input to obtain the corresponding target values; A second part of neural network construction and processing module, which is used to construct a second part of physics-informed neural network based on the original label data set and the newly added data set, combine data loss and physical equation constraint neural network, set training parameters, and train based on a preset network architecture to learn the non-linear relationship between all feature values and target values; A network prediction and evaluation module, which is used to use the trained network to predict the slamming load and structural response of the composite sandwich structure based on the input new feature values.

[0016] The beneficial effects of the present invention are: 1. The present invention includes two parts of neural networks. The first part of the neural network is located in the data preprocessing stage. The first part of the neural network after optimizing the configuration through cross-validation enhances the original data set, and normalizes the feature values and target values in the data set as the learning samples for the second part of the neural network. The second part of the neural network is a physics-informed neural network located in the learning stage of the physics-informed neural network, enabling the constructed neural network to not only learn the features of existing data but also make predictions following physical principles, effectively improving the generalization ability of the neural network, enhancing the rationality and accuracy of the prediction results. Compared with the traditional neural network prediction method, the method of the present invention reduces the demand for original label data, enabling it to still achieve fast and accurate prediction of the slamming load and structural response of the composite sandwich structure under the limitation of small sample data, expanding the application scope in data-scarce scenarios, and enhancing the practicality and applicability of the prediction method.

[0017] 2. Based on the physics-informed neural network, the present invention constructs a physical loss function by simultaneously considering the structural natural frequency under the equivalent stiffness of the composite sandwich structure and the fluid-structure interaction theory, and proposes a fast prediction method for slamming load and structural response with small samples in the case of limited test conditions or insufficient numerical simulation calculation resources. Compared with traditional neural network prediction, it considers physical prior knowledge and is more interpretable, improving the rationality and accuracy of the prediction results, and providing a more efficient evaluation means for the preliminary design of the anti-slamming performance of ship structures.

[0018] 3. The prediction method of the present invention can accurately evaluate the slamming load and structural response in the design stage of the composite sandwich structure unmanned boat, optimize the design scheme, improve the anti-slamming performance of the unmanned boat, shorten the R & D cycle, and reduce the R & D cost, which is of great significance for promoting the application of composite unmanned boats in practical engineering. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a flowchart of Example 1.

[0020] Figure 2 It is a schematic diagram of the slamming process of the composite sandwich structure in Example 1.

[0021] Figure 3 It is a schematic diagram of the characteristic parameters and target parameters of the composite sandwich structure in Example 1.

[0022] Figure 4 It is the architecture diagram of the first part of the neural network in Example 1 ( a is a neuron).

[0023] Figure 5 It is the structure diagram of the second part of the physics-informed neural network in Example 1 ( a is a neuron).

[0024] Figure 6 Schematic diagram of the regression curve between the predicted value and the true value of the peak midpoint slamming pressure in the composite sandwich structure in Embodiment 1.

[0025] Figure 7 Schematic diagram of the regression curve between the peak value of the microstrain and the true value along the plate length direction of the composite sandwich structure in Embodiment 1.

[0026] Figure 8 Schematic diagram of the predicted results of the slamming pressure and microstrain of the composite sandwich structure in Embodiment 2. Detailed implementation manners

[0027] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0028] Physics-Informed Neural Networks (hereinafter referred to as PINNs) is a neural network that incorporates physical laws described by differential equations into its loss function to guide the learning process to obtain solutions that better conform to basic physical laws. Based on PINNs, the present invention predicts the slamming load and structural response of the composite sandwich structure.

[0029] Embodiment 1 As Figure 3 The shown composite sandwich structure is a typical local structure of the bow of a composite ship, including two side panel layers and a middle core layer; the panel layers are made of glass fiber reinforced resin-based materials, and the middle core layer is PVC foam.

[0030] In this embodiment, a prediction method for the slamming load and structural response of the composite sandwich structure as described above is proposed. A prediction method for the slamming load and structural response of a composite sandwich structure, specifically a prediction method for the slamming load and structural response of a composite sandwich structure based on Physics-Informed Neural Networks (hereinafter referred to as PINNs), includes the following steps: S1. Data preprocessing stage: Collect the eigenvalues and target values of the slamming test and fluid-structure interaction numerical simulation calculation of the composite sandwich structure as the original data set; perform normalization processing on the original data set to obtain the original label data set; add new eigenvalues; build the first part of the neural network based on the original label data set, and optimize the optimal neural network of the first part based on K-fold cross-validation to learn the non-linear mapping relationship between the eigenvalues and target values in the original label data set of the composite sandwich structure, and use the first part of the neural network to enhance the data, and obtain the corresponding new target values by inputting the added new eigenvalues through the first part of the neural network. Specifically as follows: S11. Obtain the original data set through the slamming test and fluid-structure interaction simulation of the composite sandwich structure, and perform normalization processing on the original data set to obtain the original label data set.

[0031] In the present invention, for the composite sandwich structure, a slamming test is designed and combined with CFD-FEM fluid-structure interaction simulation to obtain multiple groups of original data (45 groups are obtained in this embodiment) as the original data set, and the original data includes eigenvalue parameters and target value parameters.

[0032] In this embodiment, a drop test of the composite sandwich structure entering the water is carried out. Figure 1 It is a schematic diagram of the slamming of the composite sandwich structure. In this slamming test, as Figure 2 shown, the relevant parameters include: the total mass M of the composite sandwich structure entering the water is 68.24 kg, the length L of the composite sandwich structure is 450 mm, the width B is 400 mm, the ramp angle β is 13°, and the boundary condition C B =1 is fixed support on all four sides; the composite sandwich structure consists of two panel layers (hereinafter referred to as panel layers) made of glass fiber reinforced resin matrix material and an intermediate PVC foam core layer. The thickness of one-sided panel layer is 0.8 mm, and the total thickness t f of the two panel layers E f is 1.6 mm, the elastic modulus ρ f of the panel layer is 20.7 GPa, the density 2 of the panel layer is 1834 kg / m 2 , the Poisson's ratio ν f of the panel layer is 0.12, the thickness t c of the intermediate PVC foam core layer is 4 mm, the elastic modulus E c= 130 MPa, core layer density ρ c = 100 kg / m 2 , core layer Poisson's ratio ν c = 0.4, water density ρ w = 1000 kg / m 2 . When the test fixture releases the composite sandwich structure, the water entry velocity of the composite sandwich structure is obtained through a velocity sensor V , the pressure sensor and strain gauges respectively capture the peak midpoint slamming pressure P peak and the peak micro-strain along the plate length direction ε peak . Through the free-fall water entry slamming test, 30 groups of original label datasets are obtained. At the same time, CFD-FEM fluid-structure interaction numerical simulation calculations are carried out for this composite sandwich structure, and the STAR-CCM+ software and ABAQUS software are used for joint simulation to solve its slamming pressure and structural response, obtaining the peak midpoint slamming pressure P peak and the peak micro-strain along the plate length direction ε peak . Another 15 groups of original label datasets are obtained through numerical simulation. In this embodiment, a total of 45 groups of original datasets are obtained.

[0033] In the present invention, the original dataset contains eigenvalue and target value, as Figure 2 shown, where the target value includes the peak midpoint slamming pressure P peak and the peak micro-strain along the plate length direction ε peak , that is, the dependent variable; the eigenvalue is a variable that directly affects the target value but is not affected by it, that is, the independent variable, including slamming water entry parameters, dimension parameters and material parameters. Among them, the slamming water entry parameters include the total water entry mass M (unit: kg), water entry velocity V (unit: m / s), ramp angle β , boundary condition C B , water density ρ w (unit: kg / m³); the dimension parameters include length L (unit: m), width B (unit: m), panel thickness t f (unit: m), core thickness t c (unit: m); the material parameters include the elastic modulus of the panel layer Ef (unit: N / m²), the density of the panel layer ρ f (unit: kg / m³), the Poisson's ratio of the panel layer ν f , the elastic modulus of the sandwich layer E c (unit: N / m²), the density of the sandwich layer ρ c (unit: kg / m³), the Poisson's ratio of the sandwich layer ν c .

[0034] In the present invention, in order to accelerate the training speed of the neural network and avoid excessive influence of certain features on the model due to too large numerical ranges, the original data set is standardized to unify data with different ranges and scales into a relatively consistent interval, obtaining the original label data set, specifically as shown in the following formula: (5); Among them, X and X norm are the original value and the standardized value respectively, μ X is the X mean value of σ X is the X standard deviation of

[0035] Taking the water entry speed eigenvalue as an example, the water entry speed in the original label data X = [1, 2.3, 3.9…, 7] for a total of 45 groups, the mean value of the water entry speed eigenvalue μ X = 5.2, the standard deviation σ X = 1.7, then the standardized water entry speed X norm = [-2.4, -1.7, -0.8…, 1.1].

[0036] S12. Apply ± α perturbations to the eigenvalues in the original label data set to generate new eigenvalues.

[0037] Since the original label data set only contains 45 groups of samples and the data volume is small, directly training the physics-informed neural network is likely to lead to overfitting or non-convergent training. Therefore, it is necessary to enhance the original label data set to improve the prediction robustness and generalization ability. In this embodiment, ±5 perturbations are applied to the eigenvalues in the original label data set to generate 200 groups of new eigenvalues.

[0038] Taking the characteristic value of water entry speed as an example, data enhancement is achieved through the following steps: Assume that the original eigenvalue velocity is V 0, the disturbance amplitude is ±5%, then the new velocity characteristic value generated is V new As shown in the following formula, 200 new eigenvalues are generated using normal distribution within the interval.

[0039] , (6); in, is random noise that obeys the standard normal distribution, σ =0.005.

[0040] S13. Based on the original labeled data set, build the first part of the neural network, train and evaluate the original labeled data set based on K-fold cross validation, and obtain the first part of the optimal neural network to learn the nonlinear mapping relationship between the eigenvalues and target values in the 45 groups of original labeled data sets of the composite sandwich structure, and use the first part of the neural network to enhance the data and obtain the corresponding target values using the newly added eigenvalues.

[0041] In this embodiment, the first part of the feedforward neural network architecture is built to learn the mapping rules between eigenvalues and target values in 45 groups of original label data sets as much as possible without using physical priors, and select the sub-model with the smallest verification error as the optimal neural network of the first part through K-fold cross-validation; at the same time, the optimal neural network of the first part is used to generate target values corresponding to new eigenvalues, laying the data set foundation for high-precision learning of the physical information neural network in the second part.

[0042] The specific method is as follows: First, build the first part of the neural network based on the original label data set, such as Figure 4 As shown, the first part of the neural network contains 1 input layer, 3 hidden layers (containing 64, 32, and 16 neurons respectively a(0) and 1 output layer, using the ELU activation function to enhance the non-linear expression ability and alleviate the vanishing gradient problem. The input layer is used to receive the input feature values and transfer them to the hidden layer. The hidden layer is used to perform a series of non-linear transformations and feature extractions on the input feature values. The first hidden layer has 64 neurons, which receive the features from the input layer and process them, and then transfer the results to the second hidden layer. The second hidden layer has 32 neurons, which further perform feature extraction and transformation on the data, and then transfer it to the third hidden layer. The third hidden layer has 16 neurons, which continue to process the data to prepare for the output layer. The hidden layer introduces non-linearity through various activation functions (such as the ELU activation function, etc.), so that the neural network can learn complex functional relationships. The output layer is used to output the target value according to the information transferred from the hidden layer.

[0043] Secondly, in order to enable the first part of the neural network to be effectively trained, relevant training parameters are determined: the Adam optimizer is selected, the learning rate is set to 0.001, and the maximum number of training epochs is set to 500 (which means that during the entire training process, the data will be traversed at most 500 times); the mean squared error (MSE) is selected as the loss function.

[0044] Thirdly, after determining the network architecture and training parameters, the K-fold (specifically 10-fold) cross-validation method is used to evaluate the model performance. The specific operation is as follows: the 45 groups of original label datasets are evenly and non-overlappingly divided into K subsets; in each round of training, 1 subset is selected from these K subsets as the validation set, whose role is to test the performance of the model on the data that has not participated in the training; the remaining K - 1 subsets are combined into the training set and trained based on the preset network architecture, and finally a sub-model with specific weights and biases is obtained; after each round of training, the mean squared error of the model is calculated on the validation set corresponding to this fold, and the whole process will be repeated K times to ensure that each subset has the opportunity to participate in the model evaluation as the validation set.

[0045] Finally, the model with the smallest mean squared error is selected as the optimal neural network for the first part.

[0046] In this embodiment, through comparison, it is found that the mean squared error of the model obtained by the 5th fold training is the smallest, so it is determined as the optimal neural network for the first part.

[0047] S14. Enhance the data using the first part of the neural network: Use the new feature values generated by perturbing the original dataset as the input values, and obtain the output values through the first part of the neural network, which are the target values corresponding to the newly added feature values, to form an enhanced dataset, and merge it with the original label dataset for the training of the subsequent second part of the physics-informed neural network.

[0048] In the present invention, the specific method for enhancing data is as follows: Based on the optimal neural network of the first part, the 200 groups of new eigenvalue generated in S12 are predicted to obtain the corresponding 200 groups of new target values. The 45 groups of original label datasets and the 200 groups of new label data after data enhancement are merged for the subsequent learning stage of the physics-informed neural network.

[0049] S2. Learning stage of the physics-informed neural network: Based on the original label dataset and the newly added dataset, a physics-informed neural network is constructed, that is, the neural network of the second part. Considering the data loss and the physical equation to constrain the neural network, the training parameters are set, and training is carried out based on the preset network architecture to learn the non-linear relationship between all eigenvalues and target values.

[0050] In the present invention, the neural network of the first part trains a basic neural network through data-driven to generate 200 groups of new enhanced label datasets after physical perturbation; based on the 45 groups of original label data and the 200 groups of enhanced data, a physics-informed neural network is constructed, that is, the neural network of the second part. Considering the structural natural frequency under the equivalent stiffness of the composite sandwich structure and the fluid-structure coupling theory to construct a physical loss function to improve the generalization and physical rationality of the model; then the training parameters are set, and training is carried out based on the preset network architecture to learn the non-linear relationship between all eigenvalues and target values. The two stages form a progression from "data augmentation" to "physical correction". The neural network of the first part solves the problem of insufficient samples, and the latter ensures that the learning prediction results conform to physical laws.

[0051] S21. Build the neural network of the second part, that is, the physics-informed neural network.

[0052] In this embodiment, the structure of the physics-informed neural network is as Figure 5 shown. The specific architecture is: Based on the 45 groups of original datasets and the 200 groups of enhanced datasets generated by the first part, a deep neural network including an input layer, 4 hidden layers, and an output layer is constructed; among them, the hidden layers sequentially include 128, 64, 32, and 16 neurons.

[0053] S22. The second part of the neural network constrains the neural network training through the total loss function composed of the data loss, physical loss, and L2 regularization term. Among them, the data loss calculates the mean square error between the predicted value of the second part of the neural network and the original value (using the eigenvalue obtained in the first stage as the input value of the second part of the neural network to obtain the corresponding output value, that is, the predicted value, and then calculating the mean square error between this predicted value and the original target value corresponding to the eigenvalue), and backpropagates to continuously adjust the parameters (the weights and biases of the network) to ensure that the prediction result is close to the original value. The physical loss forces the prediction result of the neural network to not only be close to the original value but also satisfy the physical constraints defined by the control equation through the embedded control equation (such as Equation 4). The L2 regularization is to add the sum of squares penalty term of the network weights to the loss function to limit the complexity of the second part of the neural network and prevent overfitting in prediction. During the training process, the optimizer dynamically balances the gradient updates of the three losses to minimize the total loss function.

[0054] In the present invention, the total loss in the physics-informed neural network includes the data loss, physical loss, and L2 regularization term, and its complete expression is: (7); (8); (9); (10); Among them, L total is the total loss function in the second part of the neural network; L date_loss is the data loss function, L physics_loss is the physical loss function; L reg is the L2 regularization term; λ d , λ p , λ r are the weight coefficients for controlling the data loss, physical loss, and regularization strength in sequence. The data fitting accuracy is preferably guaranteed, that is, λ d > λ p , to avoid potential biases introduced by the simplification of the physical equation, where the weight coefficient λ d = 0.6, λ p = 0.4, λ r is set to 0.005; N ​y true,i and y pred,i are the neural network predicted value and the true data value respectively; F inertia is the inertia term, F stiff is the stiffness term, F fluid is the fluid action term; ω k is the weight parameter of the k th layer of the model; f is the natural frequency under the equivalent stiffness of the composite sandwich structure, Hz; f pred,i is the predicted natural frequency under the equivalent stiffness of the composite sandwich structure, Hz.

[0055] In S22, the physical loss function considers the natural frequency under the equivalent stiffness of the composite sandwich structure and the fluid-structure interaction, aiming to ensure that the neural network prediction result has reasonable mechanical behavior.

[0056] The natural frequency under the equivalent stiffness of the composite sandwich structure f is shown as follows: (11); In formula (6), K eff is the equivalent stiffness of the composite sandwich structure, N / m, shown as follows: (12); (13); In formulas (7) and (8), D f and D c are the bending stiffnesses of the face sheet layer and the core layer respectively, with the unit of N・m²; t f and t c are the thicknesses of the face sheet layer and the core layer respectively, with the unit of m; ν f and ν c are the Poisson's ratios of the face sheet layer and the core layer respectively.

[0057] In the present invention, the fluid-structure interaction describes the interaction between the slamming load and the composite sandwich structure by considering the inertia term, the stiffness term and the fluid action term. The inertia term F inertia characterizes the acceleration effect of the composite sandwich structure after being subjected to the slamming load, and the stiffness termF stiff Reflects the elastic response of the composite sandwich structure, and the fluid action term F fluid Then the influence of the slamming load is considered, and the inertia term F inertia , the stiffness term F stiff and the fluid action term F fluid Are respectively shown as follows: (14); (15); (16); In the formula, m mid Is the equivalent mass at the midpoint of the composite sandwich structure, ε peak_pred Is the peak value of the central strain prediction of the composite sandwich structure; diff ( ε peak_pred ) is ε peak_pred The difference of the position.

[0058] In the present invention, the calculations of the total loss, data loss, physical loss and L2 regularization term in the physics-informed neural network are all prior arts in the industry, which can be obtained through the prior arts and are not the focus of the present invention, so they will not be elaborated here.

[0059] S23. Design the training parameters and train the physics-informed neural network.

[0060] In the present invention, the Adam optimizer is adopted, the learning rate is 0.0001, and the maximum number of training rounds is 3000 rounds.

[0061] S3. Network prediction stage: Use the trained network in S2 to input new feature values and predict the peak value of the central slamming pressure and the peak value of the tensile micro-strain along the plate length direction of the composite sandwich structure.

[0062] In this embodiment, the second part of the physics-informed neural network uses the slamming entry parameters in S11 (total entry mass M , entry velocity V , ramp angle β , boundary conditions C B , water density ρ w ), dimension parameters (length L , width B , panel layer thickness tf , sandwich thickness t c ), material parameters (elastic modulus of the panel layer E f , density of the panel layer ρ f , Poisson's ratio of the panel layer ν f , elastic modulus of the sandwich layer Ec , density of the sandwich layer ρ c , Poisson's ratio of the sandwich layer ν c ) as the input, and taking the peak midpoint slamming pressure P peak and the peak microstrain along the plate length direction ε peak as the output, to obtain the slamming pressure of the composite sandwich structure as shown in Figure 6 and Figure 7 (i.e., the peak midpoint slamming pressure P peak ) and the strain (i.e., the peak microstrain along the plate length direction ε peak ), where the coefficient of determination R 2 of the regression curve between the predicted slamming pressure value and the true value is 0.97, and the coefficient of determination R 2 of the regression curve between the predicted strain value and the true value is 0.96, indicating that the prediction result of this method is good. In the present invention, the true value regression curves of the peak midpoint slamming pressure P peak and the peak microstrain along the plate length direction ε peak can be obtained from the target values collected in Step 1.

[0063] Example 2 A system for predicting slamming loads and structural responses of a composite sandwich structure, the system comprising: A data preprocessing module, configured to obtain an original data set by using slamming tests and fluid-structure coupling, perform normalization processing on the original data set to obtain an original labeled data set, and add new feature values; A first part neural network construction and processing module, configured to construct a first part neural network based on the original labeled data set, and optimize the first part optimal neural network based on K-fold cross-validation to learn the non-linear mapping relationship between the feature values and the target values in the original labeled data set of the composite sandwich structure; and use the first part neural network to input the added new feature values to obtain the corresponding target values; The second part is the neural network construction and processing module, which is used to construct the second part of the physics-informed neural network based on the original label dataset and the newly added dataset, combine the data loss and the physical equation constraint neural network, set the training parameters, and perform training based on the preset network architecture to learn the non-linear relationship between all feature values and the target values; The network prediction and evaluation module is used to utilize the trained network to predict the slamming load and structural response of the composite sandwich structure based on the input new feature values. Specifically, for the peak midpoint slamming pressure P peak and the peak micro-strain along the plate length direction ε peak for prediction.

[0064] In this embodiment, as Figure 8 shown, by using the prediction system, inputting the new feature values, the predicted value of the slamming pressure of the composite sandwich structure (i.e., the peak midpoint slamming pressure P peak ) is 49744 Pa, and the predicted value of the strain (i.e., the peak micro-strain along the plate length direction ε peak ) is 652.

[0065] In the present invention, the slamming load structural response prediction model as a whole consists of two stages: dataset preprocessing and physics-informed neural network prediction. In the first stage of small sample dataset preprocessing, the original data is first normalized, and a neural network with three hidden layers having the best configuration is obtained through 10-fold cross-validation of the original data. Subsequently, new feature value samples generated by applying α perturbation to the original data are used, and the new slamming pressure and structural response corresponding to the new feature values are predicted through the best neural network as the new feature values. The original small sample data and the augmented data are combined. The second part is the physics-informed neural network prediction stage. The augmented data and the original data in the first stage are used as the learning data of the PINN. By simultaneously considering the data loss and the physical loss constraints in the loss function, the mapping law between the feature values and the target values is obtained, and by inputting the feature values, the slamming load and structural response of the composite sandwich structure can be quickly predicted.

[0066] Finally, it should be noted that the above are only the preferred embodiments of the present invention and are not used to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, for those skilled in the art, they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements for some of the technical features. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A prediction method for slamming load and structural response of a composite sandwich structure, characterized in that The method comprises the following steps: Data preprocessing: The eigenvalues and target values from the composite sandwich structure slamming test and fluid-structure interaction numerical simulation are collected as the original data set. The original data set is normalized to obtain the original labeled data set. The first part of the neural network is constructed, and the optimal neural network of the first part is selected based on K-fold cross-validation. The nonlinear mapping relationship between the eigenvalues and target values in the original labeled data set of the composite sandwich structure is learned, and the data is enhanced using the first part of the neural network. Physical information neural network learning phase: Based on the original labeled dataset and the newly added dataset, a physical information neural network is constructed, trained based on the preset network architecture, and learns the nonlinear relationship between all feature values and target values; Network prediction stage: Using the trained network and inputting new eigenvalues, the central impact pressure peak and the central tensile microstrain peak along the length of the plate of the composite sandwich structure are predicted.

2. The method for predicting slamming load and structural response of the composite sandwich structure according to claim 1, characterized in that, In the data preprocessing stage, a slamming test is designed for a composite sandwich structure and combined with fluid-structure coupling simulation to obtain multiple sets of raw data as the original data set. The raw data includes characteristic value parameters and target value parameters. The original dataset is standardized to obtain the original label dataset.

3. The method for predicting slamming load and structural response of the composite sandwich structure according to claim 1, wherein In the data preprocessing stage, perturbations are applied to the eigenvalues in the original labeled data set to generate new eigenvalues.

4. The method for predicting slamming load and structural response of the composite sandwich structure according to claim 3, characterized in that, The specific method of using the first part of the neural network to enhance data is: taking the new eigenvalues generated by perturbing the original data set as input values, obtaining the output values through the first part of the neural network, that is, the target values corresponding to the newly added eigenvalues, forming an enhanced data set, and merging the enhanced data set with the original label data set for subsequent training of the second part of the physical information neural network.

5. The method for predicting slamming load and structural response of the composite sandwich structure according to claim 1, characterized in that, The specific method for selecting the optimal neural network in the first part based on K-fold cross-validation is as follows: the original labeled data set is divided into K subsets evenly and without overlap; in each round of training, one subset is selected from the K subsets as the validation set, and the remaining K-1 subsets are merged into the training set. Training is performed based on the preset network architecture to obtain a sub-model with specific weights and biases, and the mean square error of the model is calculated on the corresponding validation set. The model with the smallest mean square error is selected as the optimal neural network in the first part.

6. The method for predicting slamming load and structural response of the composite sandwich structure according to claim 1, characterized in that, During the physical information neural network learning phase, the data loss and physical loss equations are combined to constrain the neural network.

7. The prediction method for slamming load and structural response of the composite sandwich structure according to claim 1, characterized in that The total loss in the physical information neural network includes data loss, physical loss and L2 regularization term, and its complete expression is: （1）； （2）； （3）； （4）； Among them, L total The total loss function in the second part of the neural network; L date_loss Is the data loss function, L physics_loss Is the physical loss function; L reg Is the L2 regularization term; λ d , λ p , λ r Are the weight coefficients for controlling the data loss, physical loss, and regularization strength in sequence; N Is the number of sample points, y true,i And y pred,i Are the predicted value of the neural network and the true data value respectively; F inertia Is the inertia term, F stiff Is the stiffness term, F fluid Is the fluid action term; ω k Is the k layer weight parameter of the model; f Is the natural frequency under the equivalent stiffness of the composite sandwich structure; f pred,i Is the predicted natural frequency under the equivalent stiffness of the composite sandwich structure.

8. The method for predicting slamming load and structural response of the composite sandwich structure according to claim 1, characterized in that, The relevant training parameters for training the physical information neural network are: Adam optimizer, learning rate of 0.0001, and maximum training rounds of 3000 rounds.

9. A slamming load and structural response prediction system for a composite sandwich structure, characterized in that The system includes: The data preprocessing module is used to obtain the original data set using the slamming test and fluid-structure interaction, normalize the original data set to obtain the original label data set, and add new eigenvalues; The first - part neural network construction and processing module is used to construct the first - part neural network based on the original label dataset, and optimize the first - part optimal neural network based on K - fold cross - validation to learn the non - linear mapping relationship between the eigenvalue and the target value in the original label dataset of the composite sandwich structure; and use the first - part neural network to input the newly added eigenvalue to obtain the corresponding target value. The second - part neural network construction and processing module is used to construct the second - part physics - informed neural network based on the original label dataset and the newly added dataset, combine the data loss and the physical equation to constrain the neural network, set the training parameters, and train based on the preset network architecture to learn the non - linear relationship between all eigenvalues and the target value. The network prediction and evaluation module is used to utilize the trained network to predict the slamming load and structural response of the composite sandwich structure based on the input new eigenvalue.

Images