A neural network-based constrained damped panel shell two-stage topology configuration prediction method and system

Abstract

This invention discloses a two-stage topology prediction method and system for constrained damped plate and shell structures based on neural networks. The method includes the following steps: constructing a Res-CB-Unet neural network model; obtaining the strain energy of the substrate element of the constrained damped structure, obtaining the optimal topology of the constrained damped structure using a SIMP model, and constructing a training dataset based on the substrate element strain energy and the optimal topology; performing two-stage training on the Res-CB-Unet neural network model based on the training dataset to obtain a configuration prediction model; and predicting the topology of the constrained damped structure to be predicted based on the configuration prediction model. This invention can achieve efficient prediction of the topology of constrained damped plate and shell structures without performing comprehensive optimization calculations, thereby accelerating the optimization process and improving prediction accuracy. Compared with traditional topology optimization methods, it can significantly improve computational efficiency while improving prediction accuracy, exhibiting a clear time advantage.

Description

Technical Field

[0001] This invention belongs to the field of constrained damping structure design technology, specifically relating to a two-stage topology prediction method and system for constrained damping plate and shell based on neural networks. Background Technology

[0002] Constrained layer damping structures primarily rely on the shear deformation of viscoelastic damping materials to dissipate vibration energy, thereby controlling the vibration and noise of the entire shell. They offer advantages such as simple structure, convenient engineering applications, significant vibration and noise reduction effects, and high reliability, and have been widely used in the vibration and noise reduction design of load-bearing shells in large ships, aerospace, and automotive industries. While current methods for optimizing constrained layer damping structures meet the lightweight design requirements of engineering projects, they suffer from excessively long optimization times and the need to rebuild topology optimization models, which is clearly unfavorable for long-term engineering applications. Therefore, developing a rapid prediction method for the optimal topology configuration of constrained layer damping structures is of significant engineering application importance.

[0003] Significant progress has been made in the optimization design of constrained damped plate and shell structures. Most objective functions are set as peak response and loss factor, optimizing the geometry and material properties of the constrained damping material to achieve lightweight and high-efficiency structures. Ansari et al. optimized the distribution and geometry of damping materials in constrained damped thin plates based on the level set method, and experimental results further verified the effectiveness of their numerical simulation. Wang Mingxu et al. used the solid isotropic microstructure penalty (SIMP) model to perform topology optimization on constrained damped plate and shell structures to maximize the modal loss factor. Dong Li et al. proposed a dual-material hybrid topology optimization method, introducing the material mixing ratio as a design variable to achieve multi-material and structural integrated topology optimization based on the SIMP method, with the optimization results considering both material distribution and mixing ratio. Wu Yonghui et al., based on the finite element model and parametric level set method (PLSM), achieved topology optimization of constrained damped (CLD) structures by maximizing the modal loss factor, and explored the influence of material thickness variation on the modal loss factor and optimal configuration. The aforementioned literature indicates that topology optimization of constrained damping structures can effectively achieve structural lightweighting and significantly suppress vibration response. However, when the dimensions of the substrate and the dimensions of the constrained damping material in a constrained damping plate and shell structure change, a new numerical model needs to be established for optimization design, resulting in low computational efficiency and making it difficult to apply to the rapid topology configuration design of generalized constrained damping plate and shell structures. Therefore, it is necessary to propose a rapid prediction method and system for topology configurations of similar constrained damping plate and shell structures. Summary of the Invention

[0004] This invention aims to address the shortcomings of existing technologies and provides the following solutions:

[0005] A two-stage topology prediction method for constrained damped plate and shell based on neural networks includes the following steps:

[0006] Construct the Res-CB-Unet neural network model;

[0007] The strain energy of the base plate element of the constrained damped plate shell structure is obtained, the optimal topology of the constrained damped structure is obtained using the SIMP model, and a training dataset is constructed based on the strain energy of the base plate element and the optimal topology.

[0008] The Res-CB-Unet neural network model is trained in two stages based on the training dataset to obtain a configuration prediction model.

[0009] Based on the configuration prediction model, the topological configuration of the constrained damping structure to be predicted is predicted.

[0010] Preferably, the Res-CB-Unet neural network model includes: an encoding layer, a convolutional attention layer, a decoding layer, a concatenation layer, and a residual skip layer;

[0011] The encoding layer is used to convert the mechanical information of the constrained damping structure into a tensor and perform feature extraction to obtain a feature tensor.

[0012] The convolutional attention layer includes channel attention and spatial attention, which are used to improve the ability to distinguish the feature tensors.

[0013] The decoding layer is used to reassemble the feature tensor to obtain the output image;

[0014] The splicing layer is used to connect the encoding layer and the decoding layer;

[0015] The residual jump is used for the internal connection between the encoding layer and the decoding layer, so as to preserve shallow low-level features in deeper layers.

[0016] Preferably, the strain energy of the substrate unit is:

[0017] MSE = U T ×K p ×U

[0018] Where MSE represents the element strain energy of the substrate in a constrained damping shell with constrained damping material, U represents the substrate displacement vector, and K p Let T represent the stiffness matrix of the substrate unit, and let T represent the transpose of the matrix.

[0019] Preferably, the method for training the Res-CB-Unet neural network model includes:

[0020] The planar diagram of the strain energy of the substrate unit is mapped proportionally onto a blank background image to obtain a mapped image;

[0021] Using the mapped image as input and the optimal topology as output, the Res-CB-Unet neural network model is trained to obtain an intermediate configuration prediction model;

[0022] The intermediate configuration prediction model is used to predict the constrained damped plate and shell structure corresponding to the mapped image to obtain the suboptimal topological configuration.

[0023] Using the suboptimal topology as input and the optimal topology as output, the intermediate configuration prediction model is trained to obtain the configuration prediction model.

[0024] The present invention also provides a two-stage topology prediction system for constrained damped plate and shell based on neural networks. The prediction system applies the prediction method described above and includes: a model building module, a dataset building module, a model training module, and a configuration prediction module.

[0025] The model building module is used to build the Res-CB-Unet neural network model;

[0026] The dataset construction module is used to obtain the strain energy of the base plate element of the constrained damping structure, obtain the optimal topology of the constrained damping structure using the SIMP model, and construct a training dataset based on the strain energy of the base plate element and the optimal topology.

[0027] The model training module performs two-stage training on the Res-CB-Unet neural network model based on the training dataset to obtain the configuration prediction model.

[0028] The configuration prediction module performs topological configuration prediction on the constrained damping structure to be predicted based on the configuration prediction model.

[0029] Preferably, the Res-CB-Unet neural network model includes: an encoding layer, a convolutional attention layer, a decoding layer, a concatenation layer, and a residual skip layer;

[0030] The encoding layer is used to convert the mechanical information of the constrained damping structure into a tensor and perform feature extraction to obtain a feature tensor.

[0031] The convolutional attention layer includes channel attention and spatial attention, which are used to improve the ability to distinguish the feature tensors.

[0032] The decoding layer is used to reassemble the feature tensor to obtain the output image;

[0033] The splicing layer is used to connect the encoding layer and the decoding layer;

[0034] The residual jump is used for the internal connection between the encoding layer and the decoding layer, so as to preserve shallow low-level features in deeper layers.

[0035] Preferably, the strain energy of the substrate unit is:

[0036] MSE = U T ×K p ×U

[0037] Where MSE represents the element strain energy of the substrate in a constrained damping shell with constrained damping material, U represents the displacement vector of the substrate, and K p Let T represent the element stiffness matrix of the substrate, and let T represent the transpose of the matrix.

[0038] Preferably, the model training module includes: a mapping unit, a first-stage training unit, a suboptimal configuration prediction unit, and a second-stage training unit;

[0039] The mapping unit maps the planar diagram of the strain energy of the substrate unit of the existing constrained damping structure onto the blank background image on a proportional scale to obtain a mapped image;

[0040] The first-stage training unit uses the mapped image as input and the optimal topology configuration as output to train the Res-CB-Unet neural network model to obtain an intermediate configuration prediction model.

[0041] The suboptimal configuration prediction unit uses the intermediate configuration prediction model to predict the existing constrained damping structure corresponding to the mapping image to obtain the suboptimal topological configuration.

[0042] The second-stage training unit uses the suboptimal topology as input and the optimal topology as output to train the intermediate configuration prediction model, thereby obtaining the configuration prediction model.

[0043] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0044] This invention enables efficient prediction of the topology of constrained damped plate and shell without performing comprehensive optimization calculations, thereby accelerating the optimization process and improving prediction accuracy. Compared with traditional topology optimization methods (such as the SIMP method), it can significantly improve computational efficiency while improving prediction accuracy, and has obvious time advantages. Attached Figure Description

[0045] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0046] Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention;

[0047] Figure 2 This is a schematic diagram of the Res-CB-Unet neural network structure according to an embodiment of the present invention;

[0048] Figure 3 This is a schematic diagram of the CBAM layer structure according to an embodiment of the present invention;

[0049] Figure 4 This is a schematic diagram of the finite element dynamic model of the constrained damping plate and shell structure according to an embodiment of the present invention. Detailed Implementation

[0050] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0051] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0052] Example 1

[0053] In this embodiment, as Figure 1 As shown, a two-stage topology prediction method for constrained damped plate and shell based on neural networks includes the following steps:

[0054] S1. Construct the Res-CB-Unet neural network model.

[0055] The Res-CB-Unet neural network model includes: an encoding layer, a convolutional attention layer, a decoding layer, a concatenation layer, and a residual jump layer. The encoding layer converts the mechanical information of the constrained damping structure into tensors and extracts features to obtain feature tensors. The convolutional attention layer includes channel attention and spatial attention to improve the discriminative ability of the feature tensors. The decoding layer reassembles the feature tensors to obtain the output image. The concatenation layer connects the encoding and decoding layers. The residual jump layer is used for internal connections between the encoding and decoding layers to preserve shallow, low-level features in deeper layers.

[0056] In this embodiment, the Res-CB-UNet neural network consists of an Encoder, a CBAM layer (Convolutional Block Attention Module), a Decoder, and a Crop layer, with an additional residual skip layer. The residual skip layer allows low-level features from shallow layers (such as edges and textures) to be preserved in deeper layers, thus helping the model better understand image details. The constructed neural network structure is as follows: Figure 2 As shown.

[0057] The encoding layer converts the appropriate constraint damping plate structural mechanical information into a tensor, which is then processed by the input layer, convolutional layer, and downsampling layer for feature extraction. The tensor undergoes several padding, convolution, batch normalization, activation function, and downsampling (the tensor size is further reduced after pooling) in the encoding layer to extract features at different levels. The size of the convolutional kernel in the encoding layer is set to 3×3, and the downsampling stride is set to 2, so that the size of the output tensor is reduced to 1 / 64 of the input tensor.

[0058] The decoding layer is responsible for reconstructing the feature tensor output by the encoding layer. It includes convolutional layers, upsampling, concatenation, and an output layer. In the decoding layer, multiple padding and interpolation deconvolution operations are performed. After concatenation, upsampling restores the image to a larger size, where the size of the convolutional kernel is the same as that of the encoding layer. The concatenation mechanism fuses some downsampled information with the upsampled features, retaining some features of the original image and further enhancing the network's learning ability. Finally, the output image generated by the decoding layer maintains the same size as the original topology.

[0059] To enhance information transmission efficiency during topology optimization, more input channels were added and information was distributed; to improve information correlation between channels, a CBAM layer was added between the encoding and decoding layers, such as... Figure 3 As shown, this module includes a SAM (channel attention) module and a CAM (spatial attention) module, aiming to improve the feature discrimination ability of the multi-channel input model. The SAM module includes operations such as adaptive pooling (adaptive max pooling, adaptive average pooling), convolution, and activation functions (ReLU) to capture the weight differences between channels. The CAM module uses the sigmoid activation function to smoothly complete the activation process and non-linear mapping of the neural network. In addition, channel stacking is used instead of element-wise addition to better explore the weight differences between pixels.

[0060] S2. Obtain the strain energy of the base plate element of the constrained damped plate and shell structure, use the SIMP model to obtain the optimal topology of the constrained damped structure, and construct a training dataset based on the base plate element strain energy and the optimal topology.

[0061] In this embodiment, a finite element dynamic model of the constrained damping structure is first constructed, such as... Figure 4 As shown, it is divided into a constraint layer, a damping layer, and a substrate layer. Each element has four nodes, and each node has seven degrees of freedom, which are the displacements u along the x and y directions within the neutral plane of the constraint layer. c Displacement u along the x and y directions within the neutral plane of the substrate layer b The lateral displacement *w* of the composite unit and the deflections along the x and y axes. Based on vibration dynamics and plate and shell theory, assuming the shear modulus of the damping layer material is a complex constant, the finite element dynamic model of the constrained damping structure is:

[0062]

[0063] Where X represents the generalized displacement vector of all nodes, M represents the global mass matrix, and K... R K represents the real part of the global stiffness matrix. I Represents the imaginary part of the global stiffness matrix. Let represent the generalized acceleration vector of all nodes, and j represent the imaginary unit. Based on the above equation, the characteristic equation for the free vibration of a constrained damped plate-shell structure can be obtained:

[0064]

[0065] Wherein: Represents the r-th complex eigenvalue, Let represent the r-th order complex eigenvector. Based on the modal loss factor characterizing the vibration suppression capability of constrained damped structures, the r-th order modal loss factor is defined as:

[0066]

[0067] Where, ψ r Let represent the r-th real mode shape vector corresponding to the complex eigenvector, and T represent the transpose of the matrix. The constrained damping material topology optimization configuration with modal loss factor as the objective function has some similarity to the substrate MSE distribution; moreover, the strain energy distribution of the base structure is easier to obtain, and the input information is convenient and simple to construct. Therefore, the strain energy of the substrate element is:

[0068] MSE = U T ×K p ×U

[0069] Where MSE represents the strain energy of the substrate element in a constrained damping plate shell with constrained damping material, U represents the substrate displacement vector, and K p Let T represent the stiffness matrix of the substrate unit, and let T represent the transpose of the matrix.

[0070] Next, the optimal topology of the constrained damping structure is obtained using the SIMP model: Based on the SIMP interpolation model, the mass matrix and stiffness matrix of the constrained damping structure can be written as:

[0071]

[0072] Where M represents the mass matrix of the constrained damping structure, and K represents the stiffness matrix of the constrained damping structure. b K b , M represents the global mass matrix, global stiffness matrix, element mass matrix, and element stiffness matrix of the substrate layer, respectively; v K R K I , M represents the global mass matrix of the damping layer, the real part of the global stiffness matrix of the constrained damping structure, the imaginary part of the global stiffness matrix of the constrained damping structure, the real part of the shear stiffness matrix of the constrained damping structure element, and the imaginary part of the shear stiffness matrix of the constrained damping structure element, respectively; c K c , These represent the global mass matrix, global stiffness matrix, element mass matrix, and element stiffness matrix of the constraint layer, respectively. K represents the unit mass matrix of the damping layer. v-R The real part of the global stiffness matrix of the damping layer is represented by n; n represents the number of elements; x i The state of the i-th constrained damping structural element is represented by its element density; p represents the penalty factor, which is chosen to be 3 in this embodiment. Using the element density of the constrained damping structure as the design variable, maximizing the modal loss factor as the design objective, and material consumption as the constraint, a mathematical model for topology optimization of the constrained damping structure is constructed as follows:

[0073] find:x i (i = 1, 2, ..., n)

[0074] Min:

[0075]

[0076] st

[0077]

[0078] 0 <x min ≤x i ≤1

[0079] Where ξ(x) is the objective function; η r Here, x is the modal loss factor, and σ is the modal order. i The state of existence of the i-th constrained damping material, here referring to the design variable, where x = 1 indicates the complete presence of the damping composite material, and x = 0 indicates the presence of only the substrate unit; V 0 This represents the volume of the constraint damping material that fully covers the surface of the substrate layer; x min denoted by , where represents the minimum unit density; and f represents the volume fraction of the constrained damping material. Finally, the constrained damping structure is topologically optimized using a mathematical model for topology optimization to obtain the optimal topology.

[0080] In this embodiment, the input data in the training dataset consists of two-dimensional images of the strain energy of the substrate elements. When constructing the input dataset images, the parameters of the constrained damping structure are set as follows: the constraint method is four-sided fixed support; the substrate dimensions of the constrained damping plate are: width 0.4m-0.3m, length 0.5m-0.3m, damping layer thickness 0.3mm-0.9mm, and constraint layer thickness 0.1mm-0.7mm. A uniform sampling method is used to uniformly sample the substrate width and length, damping layer thickness, and constraint layer thickness, and then randomly combine them to obtain 300 constrained damping plates of different sizes and specifications. Finite element analysis is performed on these 300 constrained damping plates to obtain the modal strain energy of the substrate, which is then converted into two-dimensional images. When constructing the output dataset images, the SIMP model is used to perform topology optimization on these 300 constrained damping plates to obtain the optimal topology configuration, which is then converted into two-dimensional images as the output dataset.

[0081] S3. The Res-CB-Unet neural network model is trained in two stages based on the training dataset to obtain the configuration prediction model.

[0082] The method for training the Res-CB-Unet neural network model includes: mapping the planar diagram of the strain energy of the substrate element onto a blank background image on a proportional scale to obtain a mapped image; training the Res-CB-Unet neural network model with the mapped image as input and the optimal topological configuration as output to obtain an intermediate configuration prediction model; using the intermediate configuration prediction model to predict the constraint damping structure corresponding to the mapped image to obtain a suboptimal topological configuration; and training the intermediate configuration prediction model with the suboptimal topological configuration as input and the optimal topological configuration as output to obtain a configuration prediction model.

[0083] To address the issue of varying image pixel counts for substrate unit strain energy and optimal topology configurations formed by constrained damping plates of different sizes and specifications, a proportional mapping method is employed to map the obtained two-dimensional image onto a 384×512 blank background. This ensures that the pixel ratio of the input and output information images remains consistent regardless of the damping plate size, and both input and output information are converted into coded tensors.

[0084] On a computer equipped with an Intel(R) Core(TM) i5-12500H CPU and an Nvidia 3060 GPU, using the PyTorch framework, the Adam algorithm was employed to train the Res-CB-UNet neural network, with substrate unit strain energy as input and optimal topology as output. The initial learning rate was set to 0.001. If the loss value did not show a decreasing trend after 20 consecutive training epochs, the learning rate was reduced to 0.1 times the original value. To evaluate the predictive performance of the neural network model, the cross-entropy loss function was used to characterize the prediction accuracy. As the number of training epochs increased, the loss value of the training dataset gradually decreased; after more than 207 training epochs, the loss value of the test set tended to stabilize, converging to around 0.1036, while the loss value of the training set continued to decrease slowly, reaching 0.0894 at the end of training. The neural network model with 205 training epochs was selected as the intermediate configuration prediction model. In the 205th training epoch, the strain energy cloud map of the substrate unit in the validation set was used as input to the intermediate configuration prediction model to obtain the suboptimal topological configuration. For constrained damping structures, to improve the significant limitations of the Res-CB-UNet neural network with single input information in handling edge positions and details, a second stage of neural network training was performed. In this stage, the intermediate configuration prediction model continued to learn the detailed information of the optimized configuration of the constrained damping structure. The suboptimal topological configuration obtained in the first stage of training was used as input, and the optimal topological configuration was used as output to establish the dataset for the second training. Then, the Adam algorithm was used to train the intermediate configuration prediction model in the second stage. The initial learning rate was set to 0.001. If the loss value did not show a decreasing trend after 20 consecutive training epochs, the learning rate was reduced to 0.1 times the original value to ensure that the model maintained a stable optimization direction during training. To evaluate the prediction effect of the neural network model, the cross-entropy loss function value was still used to characterize the prediction accuracy. After training, the configuration prediction model was obtained.

[0085] S4. Based on the configuration prediction model, perform topological configuration prediction on the constrained damping structure to be predicted.

[0086] Example 2

[0087] In this embodiment, the accuracy of the constructed model is evaluated.

[0088] Before evaluating the model's prediction accuracy, its stability needs to be examined. Input information from 300 samples was fed into the trained neural network to obtain the corresponding predicted images. The cross-entropy loss between the predicted results and the output information was calculated, and a loss value distribution map was constructed based on this. The loss value distribution map shows that the loss values ​​of the vast majority of samples are concentrated in the region of small values, while a small number of samples have relatively large loss values, indicating that there are instances of poor prediction performance, but this does not affect the overall prediction quality of the model. The neural network model trained in two stages exhibits a more concentrated loss value distribution when predicting the optimal configuration, demonstrating good stability and significantly outperforming the neural network model trained in only one stage.

[0089] The evaluation metrics for topology optimization of constrained damped structures include pixel value error, modal loss factor error, and computational efficiency. Specific evaluation results are shown in Table 1. Compared to the SIMP method, the single-stage neural network shows an average reduction of 11.6% in pixel value error, -3.79% in quality score error, and 96.78% in computation time for any 100 samples; while the two-stage neural network shows reductions of 5.71%, -2.84%, and 94.53% in these three aspects compared to the SIMP method.

[0090] Table 1

[0091]

[0092] In summary, based on the SIMP calculation results, the optimal topology formed by the two-stage Res-CB-UNet neural network performs better in terms of pixel value error, especially at structural edges, where the two-stage neural network prediction results show clearer details and are more consistent with the SIMP calculation results. However, the quality score error decreases, mainly due to fewer interruptions and increased detail in the prediction results. The quality score errors of both single-stage and two-stage prediction results are lower than the baseline values ​​of the SIMP method. The two-stage prediction method has a smaller prediction model error, thus having a smaller impact on the modal loss factor. While computational efficiency decreases slightly due to the secondary import of the weight file, considering the significant improvement in accuracy, the overall performance of the two-stage prediction method is still superior to that of the single-stage prediction method.

[0093] Example 3

[0094] In this embodiment, a two-stage topology prediction system for constrained damped plate and shell based on neural networks includes: a model building module, a dataset building module, a model training module, and a configuration prediction module.

[0095] The model building module is used to build Res-CB-Unet neural network models.

[0096] The Res-CB-Unet neural network model includes: an encoding layer, a convolutional attention layer, a decoding layer, a concatenation layer, and a residual jump layer. The encoding layer converts the mechanical information of the constrained damping structure into tensors and extracts features to obtain feature tensors. The convolutional attention layer includes channel attention and spatial attention to improve the discriminative ability of the feature tensors. The decoding layer reassembles the feature tensors to obtain the output image. The concatenation layer connects the encoding and decoding layers. The residual jump layer is used for internal connections between the encoding and decoding layers to preserve shallow, low-level features in deeper layers.

[0097] The dataset construction module is used to obtain the strain energy of the substrate element of the constrained damping structure, obtain the optimal topology of the constrained damping structure using the SIMP model, and construct a training dataset based on the strain energy of the substrate element and the optimal topology.

[0098] The strain energy of the substrate unit is:

[0099] MSE = U T ×K p ×U

[0100] Where MSE represents the strain energy of the substrate unit, U represents the displacement vector, and K p Let T denote the element stiffness matrix, and T denote the transpose of the matrix.

[0101] The model training module performs two-stage training on the Res-CB-Unet neural network model based on the training dataset to obtain the configuration prediction model.

[0102] The model training module includes: a mapping unit, a first-stage training unit, a suboptimal configuration prediction unit, and a second-stage training unit. The mapping unit maps the planar diagram of the strain energy of the substrate element of the existing constrained damping structure onto a blank background image on a proportional scale to obtain a mapped image. The first-stage training unit uses the mapped image as input and the optimal topological configuration as output to train the Res-CB-Unet neural network model to obtain an intermediate configuration prediction model. The suboptimal configuration prediction unit uses the intermediate configuration prediction model to predict the existing constrained damping structure corresponding to the mapped image to obtain the suboptimal topological configuration. The second-stage training unit uses the suboptimal topological configuration as input and the optimal topological configuration as output to train the intermediate configuration prediction model to obtain the configuration prediction model.

[0103] The configuration prediction module performs topological configuration prediction on the constrained damping structure to be predicted, based on the configuration prediction model.

[0104] The above embodiments are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A neural network-based constrained damped panel shell two-stage topological configuration prediction method, characterized in that, The method comprises the following steps: constructing a Res-CB-Unet neural network model; obtaining a base plate unit strain energy of a constrained damping structure, obtaining an optimal topology configuration of the constrained damping plate-shell structure by using a SIMP model, obtaining an optimal topology configuration, converting the optimal topology configuration into a two-dimensional image as output data of a data set, and constructing a training data set based on the base plate unit strain energy and the optimal topology configuration; training the Res-CB-Unet neural network model in two stages based on the training data set to obtain a configuration prediction model; predicting a topology configuration of a to-be-predicted constrained damping plate-shell structure based on the configuration prediction model; The method for training the Res-CB-Unet neural network model comprises: mapping a planar graph of the base plate unit strain energy to a blank background graph in equal proportions to obtain a mapping image; training the Res-CB-Unet neural network model by taking the mapping image as input and the optimal topology configuration as output to obtain an intermediate configuration prediction model; predicting a constrained damping plate-shell structure corresponding to the mapping image by using the intermediate configuration prediction model to obtain a suboptimal topology configuration; training the intermediate configuration prediction model by taking the suboptimal topology configuration as input and the optimal topology configuration as output to obtain the configuration prediction model.

2. The neural network-based constrained panel shell two-stage topology configuration prediction method according to claim 1, wherein, The Res-CB-Unet neural network model comprises an encoding layer, a convolutional attention layer, a decoding layer, a splicing layer, and a residual skip layer; The encoding layer is configured to convert mechanical information of a constrained damping plate-shell structure into a tensor and perform feature extraction to obtain a feature tensor; The convolutional attention layer comprises channel attention and spatial attention, and is configured to improve the distinguishing ability of the feature tensor; The decoding layer is configured to recombine the feature tensor to obtain an output graph; The splicing layer is configured to connect the encoding layer and the decoding layer; The residual skip is configured to internally connect the encoding layer and the decoding layer, and is configured to enable shallow low-level features to be retained in a deep level.

3. The neural network-based constrained panel shell two-stage topology prediction method of claim 1, wherein, The base plate unit strain energy is: wherein, MSE represents the unit strain energy of the base plate in the constrained damping panel shell on which the constrained damping material is applied, U represents the displacement vector of the base plate, K p represents the unit stiffness matrix of the base plate, T represents the transpose of a matrix.

4. A neural network-based constrained panel shell two-stage topology configuration prediction system, the prediction system applies the prediction method of any one of claims 1-3, characterized in that, comprises: a model construction module, a data set construction module, a model training module, and a configuration prediction module; The model construction module is configured to construct a Res-CB-Unet neural network model; The data set construction module is configured to obtain a base plate unit strain energy of a constrained damping structure, obtain an optimal topology configuration of the constrained damping structure by using a SIMP model, and construct a training data set based on the base plate unit strain energy and the optimal topology configuration; The model training module trains the Res-CB-Unet neural network model in two stages based on the training data set to obtain a configuration prediction model; The configuration prediction module predicts a topology configuration of a to-be-predicted constrained damping structure based on the configuration prediction model.

5. The neural network-based constrained panel shell two-stage topology prediction system of claim 4, wherein, The Res-CB-Unet neural network model comprises an encoding layer, a convolutional attention layer, a decoding layer, a splicing layer, and a residual skip layer; The encoding layer is configured to convert mechanical information of a constrained damping structure into a tensor and perform feature extraction to obtain a feature tensor; The convolution attention layer comprises channel attention and spatial attention, and is used to improve the distinguishing ability of the feature tensor; The decoding layer is used to reorganize the feature tensor to obtain an output graph; The concatenation layer is used to connect the encoding layer and the decoding layer; The residual jump is used for internal connection of the encoding layer and the decoding layer, and is used to make shallow low-level features remain in a deep level.

6. The neural network-based constrained panel shell two-stage topology prediction system of claim 4, wherein, The substrate unit strain energy is: wherein, MSE represents the unit strain energy of the base plate in the constrained damping panel shell on which the constrained damping material is applied, U represents a displacement vector of the base plate, K p represents a unit stiffness matrix of the base plate, T represents a transpose of a matrix.

7. The neural network-based constrained panel shell two-stage topology prediction system of claim 4, wherein, The model training module comprises a mapping unit, a first stage training unit, a suboptimal configuration prediction unit and a second stage training unit; The mapping unit maps a planar graph of the substrate unit strain energy of an existing constrained damping structure to a blank background graph in a same proportion to obtain a mapping image; The first stage training unit trains the Res-CB-Unet neural network model with the mapping image as input and the optimal topological configuration as output to obtain an intermediate configuration prediction model; The suboptimal configuration prediction unit uses the intermediate configuration prediction model to predict the existing constrained damping structure corresponding to the mapping image to obtain a suboptimal topological configuration; The second stage training unit trains the intermediate configuration prediction model with the suboptimal topological configuration as input and the optimal topological configuration as output to obtain the configuration prediction model.

Images