A prediction method for dynamic response of plate-frame structure under underwater explosion load based on convolutional neural network

Abstract

The present invention discloses a method for predicting the dynamic response of a plate frame structure under underwater explosion load based on a convolutional neural network, which belongs to the technology of ship damage under underwater explosion load. The method first uses CEL to perform numerical simulation on the dynamic response of the plate frame structure under underwater explosion load to obtain data samples; generates different explosive equivalent-damage picture sets according to the damage caused to the plate frame by different equivalent explosives; performs dimensionality reduction processing on the aforementioned data samples according to the symmetrical structure of the plate frame; binds the reduced-dimensional data to the picture set; executes a convolutional neural network training program, uses the reduced-dimensional data and the picture set to train and obtain training results; executes a neural network training program, trains the input explosive equivalent-damage picture and the displacement data of each node of the plate frame, so that the explosive equivalent-plate frame damage-displacement of each node of the plate frame form a fitting relationship to achieve the purpose of prediction. The present invention uses a coupled Euler-Lagrangian method (CEL) to perform numerical simulation to obtain a database, trains the convolutional neural network based on the numerical calculation results, and realizes the rapid prediction of the dynamic response of the plate frame structure under underwater explosion load.

Description

Technical Field

[0001] The present invention relates to a technology for damaging a ship under underwater explosion loads, and in particular to a method for rapidly predicting the dynamic response of a plate frame structure under underwater explosion loads based on a convolutional neural network algorithm. Background Art

[0002] Surface vessels and submarines are essential components of naval equipment. Underwater blast loads pose a significant threat to the operational capabilities and safety of ships. Rapidly and accurately assessing damage characteristics of ships subjected to underwater blast loads plays a significant military role on the battlefield. Current research on underwater blasts focuses on simulation and simulation testing. Publicly available test data is limited to simplified ship structures and small-scale tests. Due to the limitations of scale, scaled-down model testing cannot accurately replicate structural welding and other connection methods, and boundary conditions cannot be fully consistent with the actual environment. Therefore, mapping the results from scaled-down models to the actual structure is extremely difficult and is only suitable for initial mechanistic studies. Large-scale scaled-down and full-ship testing, on the other hand, are expensive, and factors such as feasibility and safety can limit their implementation. Current research on the dynamic response of ship structures subjected to underwater blasts primarily relies on three-dimensional numerical methods, which generate large amounts of data. Analysis and mining of numerical simulation results can provide a reference for research on ship resistance to blast and shock.

[0003] With the development of deep learning algorithms, methods for building predictive models using data-driven approaches have achieved promising results. For example, invention patent CN 113312832 A proposes using a deep neural network to learn from a database of data samples of the response of hull plate structures to underwater explosion loads, obtained using an arbitrary Euler-Lagrangian method. The ant colony algorithm is then used to optimize the structure and hyperparameters of the deep neural network, enabling rapid prediction of the dynamic response of hull plate structures to underwater explosion loads. Convolutional neural networks, which have emerged in recent years, are capable of mining deep features. Compared to traditional artificial neural networks, convolutional neural networks can use fewer resources to solve more complex problems using complex functions. Using convolutional neural networks to mine deep data from underwater explosion numerical simulation results, rapid on-site prediction of plate structure damage to underwater explosion loads is achieved, providing a new approach to solving traditional underwater explosion problem solving. Summary of the Invention

[0004] The present invention uses the coupled Euler-Lagrange method (CEL) to perform numerical simulation to obtain a database, and trains a convolutional neural network based on the numerical calculation results to achieve rapid prediction of the dynamic response of the plate frame structure under underwater explosion loads.

[0005] In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is: a method for predicting the dynamic response of a plate frame structure under underwater explosion load based on a convolutional neural network, comprising the following steps:

[0006] Step 1: Use the coupled Euler-Lagrangian method (CEL) to numerically simulate the dynamic response of the plate frame structure under underwater explosion load to obtain data samples;

[0007] Step 2: Generate different explosive equivalent-damage image sets based on the damage caused to the frame by explosives of different equivalents;

[0008] Step 3: According to the symmetrical structure of the frame, the data samples are processed with dimension reduction;

[0009] Step 4: Bind the reduced-dimensional data to the image set;

[0010] Step 5: Execute the convolutional neural network training program, use the dimensionality reduced data and image set for training and obtain the training results.

[0011] Step 6: Use the trained convolutional neural network to predict the damage and node displacement of the dynamic response of the frame.

[0012] Furthermore, as mentioned above, step 1 specifically includes: establishing a computational domain model and a plate frame structure model in the finite element software ABAQUS in sequence, determining the computational domain size according to the radius of the maximum bubble generated by an underwater explosion of a certain equivalent of explosives, and the size of the computational domain should be 2 to 4 times the maximum diameter of the bubble generated by a certain equivalent of explosives at the corresponding water depth; and assembling the computational domain model and the plate frame structure model; wherein, the computational domain model uses a three-dimensional Euler body, and the plate frame structure uses a three-dimensional deformable body; dividing the computational domain model into regions, dividing the water region and the explosive region; defining a predetermined number of material properties; meshing the computational domain model and the plate frame structure model; setting the contact properties, applying boundary conditions and load boundary conditions of the computational domain model and the plate frame structure model; and assigning the defined material properties to the pre-analyzed water region and the explosive region in the predefined field; and acquiring data samples through the Euler volume fraction in the visualization results in combination with the change of the Euler volume fraction.

[0013] Furthermore, the above-mentioned step 2 specifically includes: obtaining an explosive equivalent-damage picture set by changing the TNT equivalent to cause different degrees of damage to the frame.

[0014] Furthermore, the above step 3 specifically includes: reducing the dimension of the data obtained in step 1 by the symmetrical structure of the frame. The displacement matrix of each node of the frame structure is:

[0015]

[0016]

[0017] Among them, A is the displacement matrix of the upper left node of the plate frame, and B is the displacement matrix of the lower left node of the plate frame. Due to the symmetrical structure of the plate frame, we can get a mn =b mn , so the matrix B can be predicted to achieve the purpose of data dimensionality reduction.

[0018] Similarly, the displacement matrix of the entire frame node can be obtained as:

[0019]

[0020] Among them, A is the displacement matrix of the upper left node of the frame, B is the displacement matrix of the lower left node of the frame, C is the displacement matrix of the upper right node of the frame, and D is the displacement matrix of the lower right node of the frame. According to the symmetrical structure of the frame, we can get a mn =b mn , c mn =d mn , so the data of matrix M can be predicted to achieve the purpose of data dimensionality reduction.

[0021] Furthermore, as described above, step 4 specifically includes: binding the explosive equivalent-damage image set obtained in step 2 with the data after dimensionality reduction in step 3: each explosive equivalent-damage image corresponds to the displacement data of each node of a plate frame, and the matched images and displacement parameters are placed in the same cell.

[0022] Furthermore, step 5 specifically includes: using a convolutional neural network, the main structural layers are as follows:

[0023] The first layer consists of four parts: convolution, normalization, ReLU nonlinear activation, and pooling. Its main function is to extract features from the input image, and the result is used as the input of the second layer.

[0024] The second layer consists of four parts: convolution, normalization, ReLU nonlinear activation, and pooling. Its main function is to further extract features from the input image, and the results serve as the input of the third layer.

[0025] The third layer consists of two parts: convolution and ReLU nonlinear activation. Its main function is to further extract features from the input image, and the result is used as the input of the fourth layer.

[0026] The fourth layer consists of two parts: convolution and ReLU nonlinear activation. Its main function is to further extract features from the input image, and the result is used as the input of the fifth layer.

[0027] The fifth layer consists of three parts: convolution, ReLU nonlinear activation, and pooling. Its main function is to further extract features from the input image, and the result serves as the input of the sixth layer.

[0028] The sixth layer consists of three parts: full connection, ReLU nonlinear activation, and dropout. Its main function is to combine and classify the image features extracted by the previous five convolutional layers;

[0029] The seventh layer consists of three parts: full connection, ReLU nonlinear activation, and dropout. Its main function is to further combine and classify the features of the image set;

[0030] The eighth layer is the softmax output part, which outputs the final result.

[0031] Furthermore, step 6 as described above specifically includes executing a neural network training program to train the input explosive equivalent-damage image and the displacement data of each node of the frame, so that the explosive equivalent-frame damage-displacement of each node of the frame form a fitting relationship to achieve the purpose of prediction.

[0032] The present invention achieves the following benefits: by using a coupled Euler-Lagrangian (CEL) method to perform numerical simulations to generate a database, and then training a convolutional neural network based on the numerical calculation results, it enables rapid prediction of the dynamic response of a panel structure under underwater explosion loads. Compared to existing technologies, this method can achieve rapid prediction of panel structure responses. Given the same computing resources, this method can produce more accurate results than traditional ABAQUS prediction methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 Flowchart for prediction of dynamic response of plate frame structure under underwater explosion load;

[0034] Figure 2 It is the assembly drawing of the calculation model and the plate frame structure;

[0035] Figure 3 Grid division diagram for calculation model and plate frame structure;

[0036] Figure 4 This is a comparison chart between the damage prediction results of the plate frame structure under underwater explosion load and the ABAQUS simulation results. DETAILED DESCRIPTION

[0037] To further understand the content, features, and effectiveness of this invention, the following examples provide further explanation. This invention utilizes a coupled Euler-Lagrangian (CEL) method for numerical simulation to generate a database. A convolutional neural network is trained based on the numerical results to achieve rapid prediction of the dynamic response of a panel structure under underwater explosive loads. The following describes the operational steps in detail.

[0038] Step 1: Use the coupled Euler-Lagrangian method (CEL) to numerically simulate the dynamic response of the plate frame structure under underwater explosion load to obtain data samples;

[0039] Specifically, the computational domain model and the plate frame structure model were established in ABAQUS. The computational domain size was determined based on the radius of the maximum bubble generated by an underwater explosion of a certain explosive equivalent. The computational domain size should be 2 to 4 times the maximum diameter of the bubble generated by a certain explosive equivalent at the corresponding water depth. The JWL state equation was used to describe the detonation process of TNT. Its expression is as follows:

[0040]

[0041] Where:

[0042] η—the ratio of product density to initial density, i.e. η=ρ / ρ0;

[0043] A, B, R1, R2, and ω are constants related to the state of the explosive;

[0044] e—internal energy per unit mass of explosive;

[0045] The parameters of TNT explosives are as follows

[0046] Table 1 TNT explosive parameters

[0047]

[0048]

[0049] Then we assemble the computational domain model and the plate frame structure model, and the resulting assembly diagram is as follows: Figure 2 As shown in the figure, the computational domain model uses a three-dimensional Euler body, and the plate frame structure uses a three-dimensional deformable body; the computational domain model is divided into regions, and the water area and the explosive area are divided; a predetermined number of material properties are defined; the computational domain model and the plate frame structure model are meshed, as shown in the schematic diagram. Figure 3 As shown in the figure; set the contact properties of the calculation domain model and the plate frame structure model, apply boundary conditions and load boundary conditions; assign the defined material properties to the pre-analyzed water area and explosive area in the predefined field; save the built model as a cae file to prepare for subsequent data acquisition

[0050] Step 2: Create an ODB job from the CAE file obtained in Step 1 and submit it for analysis. By visualizing the Euler volume fraction in the results, we can obtain data samples. Furthermore, by varying the radius of the spherical TNT to change the TNT equivalent, we can generate different sets of explosive equivalent-damage images.

[0051] Step 3: Reduce the dimension of the data obtained in step 1 by the symmetrical structure of the frame. The displacement matrix of each node of the frame structure is:

[0052]

[0053]

[0054] Among them, A is the displacement matrix of the upper left node of the plate frame, and B is the displacement matrix of the lower left node of the plate frame. Due to the symmetrical structure of the plate frame, we can get a mn =b mn , so the matrix B can be predicted to achieve the purpose of data dimensionality reduction.

[0055] Similarly, the displacement matrix of the entire frame node can be obtained as:

[0056]

[0057] Among them, A is the displacement matrix of the upper left node of the frame, B is the displacement matrix of the lower left node of the frame, C is the displacement matrix of the upper right node of the frame, and D is the displacement matrix of the lower right node of the frame. According to the symmetrical structure of the frame, we can get a mn =b mn , c mn =d mn , so the data of matrix M can be predicted to achieve the purpose of data dimensionality reduction.

[0058] Step 4: Bind the image set to the data rows reduced in step 3. Each explosive equivalent-damage image corresponds to the displacement data for each node of the frame. This time, only the explosive equivalent is varied to obtain different degrees of damage to the frame and generate different explosive equivalent-frame damage images. Each image corresponds to the displacement of each frame node, thus generating a damage and displacement dataset. Match the matched images and displacement parameters to the same cell.

[0059] Step 5: This time we use the AlexNet convolutional neural network, the main structure of which is as follows:

[0060] First layer: convolution + normalization + ReLU + pooling

[0061] The input image is a picture in the image collection, with a size of 227×227×3. After 96 convolution kernels of 11×11 and a stride of 4, the resulting size should be 55×55×96;

[0062] Perform normalization operation and use ReLU function for nonlinear activation. The ReLU function is

[0063]

[0064] Pooling is performed using a 3×3 pooling unit with a stride of 2 to obtain a feature map of size 27×27×96, which is output separately, with the upper and lower sides sized at 27×27×48 respectively.

[0065] Second layer: convolution + normalization + ReLU + pooling

[0066] The input is the matrix output by the first layer and the corresponding damage and displacement values. The two sets of data are 27×27×48 in size. They are collected using a 128×5×5 convolution kernel with a convolution stride of 1. The output is two sets of 128×27×27 pixel layers.

[0067] The ReLU function activates nonlinearity;

[0068] Pooling is performed using a 3×3 pooling unit with a stride of 2, resulting in two sets of pixel layers of size 13×13×128.

[0069] The third layer: convolution + ReLU

[0070] The input is the matrix output by the second layer and the corresponding damage and displacement values. The two sets of data are 13×13×128 in size. They are collected using a 3×3×192 convolution kernel with edge padding of 1 and a convolution stride of 1. The output is two sets of pixel layers of 13×13×192 in size.

[0071] The ReLU function activates nonlinearity.

[0072] Fourth layer: convolution + ReLU

[0073] The input is the matrix output by the third layer and the corresponding damage and displacement values. The size of the two data sets is 13×13×192. A convolution kernel of 3×3×192 is used with edge padding = 1 and a convolution stride of 1. The output is two pixel layers of size 13×13×192.

[0074] The ReLU function activates nonlinearity.

[0075] Fifth layer: convolution + ReLU + pooling

[0076] The input is the matrix output by the fourth layer and the corresponding damage and displacement values. The two sets of data are 13×13×192 in size. A convolution kernel of 3×3×128 is used with edge padding of 1 and a convolution stride of 1. The output is two sets of data of 13×13×128 in size.

[0077] The ReLU function activates nonlinearity;

[0078] For pooling, input two groups of 13×13×128, use 3×3 pooling units with a step size of 2, and obtain two groups of 6×6×128 data.

[0079] Sixth layer: full connection + ReLU + dropout

[0080] The input is the matrix output by the fifth layer and the corresponding damage and displacement values. The size of the two sets of data is 6×6×128. This layer has 4096 convolution kernels, and the size of each convolution kernel is 6×6×256. Since the size of the convolution kernel is exactly the same as the size of the feature map to be processed, that is, each coefficient in the convolution kernel is multiplied by only one pixel value of the feature map size. Therefore, this layer is called a fully connected layer. Since the convolution kernel is the same size as the feature map, there is only one value after the convolution operation. The pixel layer size after convolution is 4096×1×1, that is, there are 4096 neurons;

[0081] The ReLU function activates nonlinearity;

[0082] The dropout layer suppresses overfitting by randomly disconnecting some neurons or not activating some neurons.

[0083] Seventh layer: full connection + ReLU + dropout

[0084] Fully connected, the input is the matrix output by the sixth layer and the corresponding damage and displacement values, the data size is 4096×1×1;

[0085] The ReLU function activates nonlinearity;

[0086] The dropout layer suppresses overfitting by randomly disconnecting certain neurons or not activating certain neurons.

[0087] Layer 8: Output

[0088] The input is the matrix output by the seventh layer and the corresponding damage and displacement values, which are combined to obtain a final result matrix dimension of 2, which is the two parameters required to predict damage and plate displacement.

[0089] Step 6: Execute the neural network training program to train the input explosive equivalent-damage image and the displacement data of each node of the frame, so that the explosive equivalent-frame damage-frame displacement of each node form a fitting relationship to achieve the purpose of prediction, such as Figure 4 The comparison chart of the plate frame damage predicted by ABAQUS and convolutional neural network for 0.4kg equivalent TNT. The maximum displacement error is 0.213, and the calculated error rate is 7.9%. Therefore, it can be concluded that the neural network can predict the dynamic response of the plate frame structure under underwater explosion load.

Claims

1. A method for predicting the dynamic response of a plate frame structure under underwater explosion load based on a convolutional neural network, characterized in that: The following steps are involved: Step 1: Use the coupled Euler-Lagrangian method to numerically simulate the dynamic response of the plate frame structure under underwater explosion load to obtain data samples; Step 2: Generate different explosive equivalent-damage image sets based on the damage caused to the frame by explosives of different equivalents; Step 3: According to the symmetrical structure of the frame, perform dimensionality reduction processing on the data samples obtained in step 1; Step 4: Bind the reduced-dimensional data to the image set; Step 5: Execute the convolutional neural network training program, use the reduced-dimensional data and the image set to train and obtain the training results; Step 6: Use the trained convolutional neural network to predict the damage and node displacement of the dynamic response of the frame.

2. A method for predicting the dynamic response of a plate frame structure under underwater explosion load based on a convolutional neural network as claimed in claim 1, characterized in that: The specific process of step 1 is as follows: In the finite element software ABAQUS, a computational domain model and a plate frame structure model are successively established. The computational domain size is determined according to the radius of the maximum bubble generated by the underwater explosion of a certain equivalent of explosives. The size of the computational domain should be 2 to 4 times the maximum diameter of the bubble generated by a certain equivalent of explosives at the corresponding water depth. The computational domain model and the plate frame structure model are assembled. Among them, a three-dimensional Euler body is selected as the computational domain model, and a three-dimensional deformable body is selected as the plate frame structure. The computational domain model is divided into regions, and the water region and the explosive region are divided. A predetermined number of material properties are defined. The computational domain model and the plate frame structure model are meshed. The contact properties, applied boundary conditions and load boundary conditions of the computational domain model and the plate frame structure model are set. The defined material properties are respectively assigned to the pre-analyzed water region and the explosive region in the predefined field. Combined with the change of the Euler volume fraction, the data samples are acquired through the Euler volume fraction in the visualization results.

3. A method for predicting the dynamic response of a plate frame structure under underwater explosion load based on a convolutional neural network as claimed in claim 1, characterized in that: The specific process of reducing the dimension of the sample data obtained in step 1 by using the symmetrical structure of the plate frame in step 3 is as follows: The displacement matrix of each node of the plate frame structure is: (1) (2) in, is the displacement matrix of the upper left node of the frame, is the displacement matrix of the node at the lower left of the frame. Due to the symmetrical structure of the frame, we can get , so we can do the matrix Make predictions to achieve data dimensionality reduction; Similarly, the displacement matrix of the entire frame node can be obtained as: (3) in, is the displacement matrix of the upper left node of the frame, is the displacement matrix of the lower left node of the frame, is the displacement matrix of the upper right node of the frame, is the displacement matrix of the node at the lower right of the frame; according to the symmetrical structure of the frame, we can get , , so the matrix The data is used for prediction to achieve the purpose of data dimensionality reduction.

4. A method for predicting the dynamic response of a plate frame structure under underwater explosion load based on a convolutional neural network as claimed in claim 1, characterized in that: The specific process of step 4 is as follows: Bind the explosive equivalent-damage image set obtained in step 2 to the data obtained after dimensionality reduction in step 3: each explosive equivalent-damage image corresponds to the displacement data of each node of a frame, and place the matched images and displacement parameters into the same cell.

5. A method for predicting the dynamic response of a plate frame structure under underwater explosion load based on a convolutional neural network as claimed in claim 1, characterized in that: The convolutional neural network used in step 5 has the following main structural layers: The first layer consists of four parts: convolution, normalization, ReLU nonlinear activation, and pooling. Its main function is to extract features from the input image, and the result is used as the input of the second layer. The second layer consists of four parts: convolution, normalization, ReLU nonlinear activation, and pooling. Its main function is to further extract features from the input image, and the results serve as the input of the third layer. The third layer consists of two parts: convolution and ReLU nonlinear activation. Its main function is to further extract features from the input image, and the result is used as the input of the fourth layer. The fourth layer consists of two parts: convolution and ReLU nonlinear activation. Its main function is to further extract features from the input image, and the result is used as the input of the fifth layer. The fifth layer consists of three parts: convolution, ReLU nonlinear activation, and pooling. Its main function is to further extract features from the input image, and the result serves as the input of the sixth layer. The sixth layer consists of three parts: full connection, ReLU nonlinear activation, and dropout. Its main function is to combine and classify the image features extracted by the previous five convolutional layers; The seventh layer consists of three parts: full connection, ReLU nonlinear activation, and dropout. Its main function is to further combine and classify the features of the image set; The eighth layer is the softmax output part, which outputs the final result.

6. A method for predicting the dynamic response of a plate frame structure under underwater explosion load based on a convolutional neural network as claimed in claim 1, characterized in that: The specific process of step 6 is as follows: execute the neural network training program, train the input explosive equivalent-damage picture and the displacement data of each node of the frame, so that the explosive equivalent-frame damage-displacement of each node of the frame form a fitting relationship to achieve the purpose of prediction.

Images