Design network training method, design method and system of thin plate type acoustic metamaterial

Abstract

The invention discloses a design network training method, a design method, a system, equipment, a medium and a product of a thin-plate type acoustic metamaterial, and relates to the technical field of acoustic metamaterial design, the method comprises the following steps: obtaining a plurality of thin-plate type acoustic metamaterial structures, mass blocks of each thin-plate type acoustic metamaterial structure being randomly distributed; preprocessing each sheet type acoustic metamaterial structure to obtain a sample data set; the sample data set comprises a mass distribution pixel image and a sound transmission loss curve corresponding to each thin plate type acoustic metamaterial structure; and training the bidirectional neural network model based on the sample data set to obtain a forward prediction network and a reverse design network. According to the method, the sound transmission loss curve can be quickly predicted, the matched mass block distribution structure can be efficiently generated, and the design efficiency is remarkably improved.

Description

Technical Field

[0001] This application relates to the field of acoustic metamaterial design technology, and in particular to a design network training method, design method, system, equipment, medium and product for thin-plate acoustic metamaterials. Background Technology

[0002] With the continuous acceleration of industrial modernization, noise pollution is causing increasingly serious damage to the ecological environment and harm to human health. Developing efficient and precise noise control technologies has become an urgent need for industry development. Acoustic metamaterials, with their unique physical sound insulation mechanisms, can overcome the limitations of traditional materials and achieve precise control of sound waves at specific frequencies, thus demonstrating significant research value and application prospects in the field of noise control.

[0003] Thin-plate acoustic metamaterials, due to their simple structure, mature manufacturing process, and wide applicability, have significant application potential in noise control engineering. The acoustic transmission loss performance of these materials is highly dependent on the distribution of the mass blocks, and the arrangement of the mass blocks directly determines the material's sound insulation effect. In practical applications, engineers need to quickly predict the acoustic transmission loss curve based on the mass block distribution (forward prediction), and simultaneously efficiently generate a matching mass block distribution structure based on the target acoustic transmission loss curve (reverse design). This is the core step in realizing the engineering application of thin-plate acoustic metamaterials.

[0004] However, in existing technologies, obtaining acoustic transmission loss curves mainly relies on time-consuming finite element simulations or physical experimental tests. The calculation process is complex and resource-intensive, failing to meet the needs of rapid design. Furthermore, the process of deriving mass block distribution from target performance lacks systematic tools, often requiring repeated trial and error and manual adjustments, resulting in low design efficiency and lengthy cycles. These problems severely hinder the promotion and application of thin-plate acoustic metamaterials in practical noise control scenarios. Summary of the Invention

[0005] The purpose of this application is to provide a design network training method, design method, system, device, medium and product for thin-plate acoustic metamaterials, which can quickly predict the acoustic transmission loss curve and efficiently generate a matching mass block distribution structure, avoiding time-consuming finite element simulation and repeated trial and error process, and significantly improving design efficiency.

[0006] To achieve the above objectives, this application provides the following solution: In a first aspect, this application provides a method for training a design network for thin-plate acoustic metamaterials, comprising: acquiring a plurality of thin-plate acoustic metamaterial structures, wherein the mass blocks of each thin-plate acoustic metamaterial structure are randomly distributed; preprocessing each thin-plate acoustic metamaterial structure to obtain a sample dataset; the sample dataset includes: a mass distribution pixel image and an acoustic transmission loss curve corresponding to each thin-plate acoustic metamaterial structure; training a bidirectional neural network model based on the sample dataset to obtain the forward prediction network and the reverse design network, wherein the bidirectional neural network model includes a first network model and a second network model; the forward prediction network is used to obtain a predicted acoustic transmission loss curve based on the mass distribution pixel image, and the reverse design network is used to obtain a mass distribution pixel design image based on the target acoustic transmission loss curve.

[0007] Optionally, obtaining a plurality of thin-plate acoustic metamaterial structures specifically includes: Obtain a single thin-plate acoustic metamaterial, comprising a thin plate and several mass blocks, wherein each mass block has the same size; Using simulation software, several mass blocks are randomly matched to the thin-plate acoustic metamaterial to generate several thin-plate acoustic metamaterial structures in batches, wherein the number and / or position of the mass blocks in each thin-plate acoustic metamaterial structure are different.

[0008] Optionally, the preprocessing of each of the thin-plate acoustic metamaterial structures to obtain a sample dataset specifically includes: Frequency domain response calculations were performed on each of the thin-plate acoustic metamaterial structures to obtain the acoustic transmission loss curves corresponding to each of the thin-plate acoustic metamaterial structures. Each of the thin-plate acoustic metamaterial structures is encoded to obtain a mass distribution pixel image, wherein the mass distribution pixel image is a binary pixelated grayscale image, the mass block region corresponds to the black pixels of the grayscale image, and the blank region of the thin plate corresponds to the white pixels of the grayscale image; The acoustic transmission loss curves are normalized, and the normalized acoustic transmission loss curves are associated with the quality distribution pixel images to form the sample dataset.

[0009] Optionally, training the bidirectional neural network model based on the training dataset to obtain a forward prediction network and a reverse design network specifically includes: Multiple quality distribution pixel images are randomly selected from the training dataset as first samples and input into the first network model. At the same time, the acoustic transmission loss curve corresponding to the first sample is extracted as the first label. The first network model is trained iteratively multiple times to obtain the positive prediction network. Wherein, any iteration of training of the first network model includes: The first feature map is obtained by convolving the first sample with the first convolution kernel; The second feature map is obtained by convolving the first feature map with the second convolution kernel; After normalizing the second feature map, it is input into the first fully connected layer, and the predicted acoustic transmission loss curve is output through the first fully connected layer. The predicted acoustic transmission loss curve and the first loss value of the first tag are calculated using the mean square error loss function. The weights of the first convolutional kernel, the second convolutional kernel, and the first fully connected layer are adjusted based on the first loss value.

[0010] Optionally, training the bidirectional neural network model based on the training dataset to obtain the forward prediction network and the reverse design network further includes: Multiple acoustic transmission loss curves are randomly selected from the training dataset as second samples and input into the second network model. At the same time, the quality distribution pixel images corresponding to the second samples are extracted as second labels to perform several iterations of training on the second network model to obtain the reverse design network. The training of the second network model in any iteration includes: The second sample is input into the second fully connected layer to obtain the first reconstructed feature map; The first reconstructed feature map is convolved by a convolutional layer to obtain the second reconstructed feature map; Global average pooling and global max pooling are performed on the second reconstructed feature map to obtain the average acoustic transmission loss and the peak acoustic transmission loss. The average sound transmission loss and the peak sound transmission loss are fused through a second fully connected layer to generate a spatial attention weight matrix. The spatial attention weight matrix is ​​processed by a decision function to generate a pixel image of predicted quality distribution. The second loss value between the predicted quality distribution pixel image and the second label is calculated using the mean square error loss function. The weights of the second fully connected layer and the convolutional layer are adjusted based on the second loss value, and the fusion coefficients of the global average pooling and global max pooling are also adjusted.

[0011] Secondly, this application provides a design method for a thin-plate acoustic metamaterial, based on a forward prediction network and an inverse design network obtained through the training method described in any of the first aspects above. The design method for the thin-plate acoustic metamaterial includes: The positive prediction network is used to obtain the predicted acoustic transmission loss curve based on the quality distribution pixel image; The reverse design network is used to obtain a quality distribution pixel design image based on the target sound transmission loss curve.

[0012] Thirdly, this application provides a design network training system for thin-plate acoustic metamaterials, including: The acquisition module is used to acquire several thin-plate acoustic metamaterial structures, wherein the mass blocks of each thin-plate acoustic metamaterial structure are randomly distributed. The processing module is used to preprocess each of the thin-plate acoustic metamaterial structures to obtain a sample dataset; the sample dataset includes: a mass distribution pixel image and a sound transmission loss curve corresponding to each thin-plate acoustic metamaterial structure; The training module trains the bidirectional neural network model based on the sample dataset to obtain the forward prediction network and the reverse design network. The bidirectional neural network model includes a first network model and a second network model. The forward prediction network is used to obtain the predicted acoustic transmission loss curve based on the quality distribution pixel image, and the reverse design network is used to obtain the quality distribution pixel design image based on the target acoustic transmission loss curve.

[0013] Fourthly, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the design network training method for the thin-plate acoustic metamaterial described in any one of the above.

[0014] Fifthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the design network training method for the thin-plate acoustic metamaterial described in any one of the above.

[0015] In a sixth aspect, this application provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the design network training method for the thin-plate acoustic metamaterial described above.

[0016] According to the specific embodiments provided in this application, the following technical effects are disclosed: This application provides a training method, design method, system, device, medium, and product for a design network of thin-plate acoustic metamaterials. First, several thin-plate acoustic metamaterial structures are acquired, with the mass blocks of each structure randomly distributed, providing a diverse sample base for subsequent deep learning model training. These samples cover different distribution patterns of mass block quantity and location, ensuring that the neural network can fully learn the complex relationship between mass block randomness and acoustic performance. Further, each thin-plate acoustic metamaterial structure is preprocessed to obtain a sample dataset. This dataset includes: a mass distribution pixel image and an acoustic transmission loss curve corresponding to each thin-plate acoustic metamaterial structure. This step achieves digital representation of the physical structure and standardized processing of performance data, ensuring the reliability of model evaluation. Second, a bidirectional neural network model is trained based on the training dataset to obtain a forward prediction network and a reverse design network. The bidirectional neural network model includes a first network model and a second network model. The forward prediction network is used to obtain the predicted acoustic transmission loss curve based on the mass distribution pixel image, and the reverse design network is used to obtain the mass distribution pixel design image based on the target acoustic transmission loss curve. The forward prediction network utilizes a neural network to capture the nonlinear mapping relationship between pixel-level distribution features in a mass distribution pixel image and the frequency response characteristics of the acoustic transmission loss curve. Therefore, given any mass distribution, it can quickly generate the corresponding acoustic transmission loss prediction curve, avoiding the high computational cost of traditional finite element simulation methods. Simultaneously, the core of the reverse design network lies in learning the mapping law from the acoustic transmission loss curve to the mass distribution pixel image through a neural network. This reverse design mechanism allows for the direct derivation of a matching mass block distribution scheme based on specific acoustic performance requirements, overcoming the limitations of traditional trial-and-error methods in structural design. By constructing both the forward prediction network and the reverse design network, the rapid prediction of the acoustic transmission loss curve and the efficient generation of the mass block distribution are achieved. This approach enables rapid prediction of the acoustic transmission loss curve and efficient generation of a matching mass block distribution structure, avoiding time-consuming finite element simulation and repeated trial-and-error processes, significantly improving design efficiency. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is an application environment diagram of a design network training method for a thin-plate acoustic metamaterial according to an embodiment of this application; Figure 2A flowchart illustrating a design network training method for a thin-plate acoustic metamaterial provided in one embodiment of this application; Figure 3 for Figure 2 A detailed flowchart illustrating step S202; Figure 4 This is a schematic diagram illustrating the process of training a first network model according to an embodiment of this application. Figure 5 A schematic diagram illustrating the process of training a second network model according to an embodiment of this application; Figure 6 for Figure 2 A schematic diagram of the test results for detecting the positive prediction network in step S205; Figure 7 for Figure 2 A schematic diagram of the test results for detecting the reverse-designed network in step S205; Figure 8 A schematic diagram of the functional modules of a design network training system for thin-plate acoustic metamaterials provided in an embodiment of this application; Figure 9 A schematic diagram illustrating the process of forward prediction network and reverse design network provided in the embodiments of this application; Figure 10 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0019] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0020] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0021] The design network training method for thin-plate acoustic metamaterials provided in this application can be applied to, for example... Figure 1In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be set up independently, integrated into server 104, or placed in the cloud or on other servers. Terminal 102 can send the thin-plate acoustic metamaterial structure to be processed to server 104. After receiving the thin-plate acoustic metamaterial structure, server 104 preprocesses each structure to obtain a sample dataset; divides the sample dataset into a training dataset and a test dataset; trains a bidirectional neural network model based on the training dataset to obtain a forward prediction network and a reverse design network, wherein the bidirectional neural network model includes a first network model and a second network model; and tests the forward prediction network and the reverse design network based on the test dataset. Server 104 can feed back the test results of the forward prediction network and the reverse design network based on the test dataset to terminal 102. Furthermore, in some embodiments, the design network training method for thin-plate acoustic metamaterials can also be implemented separately by the server 104 or the terminal 102. For example, the terminal 102 can directly preprocess the thin-plate acoustic metamaterial structure to be processed, or the server 104 can obtain the thin-plate acoustic metamaterial structure to be processed from the data storage system, preprocess each thin-plate acoustic metamaterial structure to be processed to obtain a sample dataset, divide the sample dataset into a training dataset and a test dataset, and train the bidirectional neural network model based on the training dataset to obtain the forward prediction network and the reverse design network.

[0022] The terminal 102 can be, but is not limited to, various desktop computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices. IoT devices can include smart speakers, smart TVs, smart air conditioners, and smart in-vehicle devices. Portable wearable devices can include smartwatches, smart bracelets, and head-mounted devices. The server 104 can be implemented using a standalone server or a server cluster composed of multiple servers, or it can be a cloud server.

[0023] In one exemplary embodiment, such as Figure 2 As shown, a method for designing and training a network for thin-plate acoustic metamaterials is provided. This method is executed by a computer device, specifically by a terminal or server alone, or by both a terminal and a server. In this embodiment, the method is applied to... Figure 1 Taking server 104 as an example, the explanation includes the following steps 201 to 205. Wherein: S201. Obtain several thin-plate acoustic metamaterial structures, with the mass blocks of each thin-plate acoustic metamaterial structure randomly distributed. In some embodiments, the above step S201 (obtaining a plurality of thin-plate acoustic metamaterial structures) may be implemented in a manner that includes: obtaining a plurality of thin-plate acoustic metamaterial structures, wherein the mass blocks of each thin-plate acoustic metamaterial structure are randomly distributed.

[0024] That is, the number of mass blocks of each thin-plate acoustic metamaterial is different, or the position of the mass blocks of each thin-plate acoustic metamaterial is different, or both the number and position of the mass blocks of each thin-plate acoustic metamaterial are different.

[0025] Specifically, step S201 above, which involves obtaining several thin-plate acoustic metamaterial structures, may specifically include: Obtain a single thin-plate acoustic metamaterial, comprising a thin plate and several mass blocks, wherein each mass block has the same size; use simulation software to randomly match several mass blocks for the thin-plate acoustic metamaterial, and batch generate several thin-plate acoustic metamaterial structures, wherein the number and / or position of mass blocks in each thin-plate acoustic metamaterial structure are different.

[0026] For example, the dimensions of the thin plate can be 100×80×1 mm. 3 Its density can be 1170 kg / m³ 3 The Young's modulus can be 3.2 GPa, the Poisson's ratio can be 0.35, and the anisotropy loss factor can be 0.05.

[0027] As another example, the size of the mass block can be 10×10×3 mm. 3 Its density can be 7870 kg / m³ 3 The Young's modulus can be 200 GPa, and the Poisson's ratio can be 0.29.

[0028] More specifically, the surface of the thin plate is divided into a 10×8 pixel grid, and each pixel grid can independently choose to embed a mass block or remain blank. Using a computer random algorithm, the existence state of the mass block in each pixel grid is randomly assigned (for example, using a 0-1 random number generator, where 1 represents that the pixel grid is embedded with a mass block and 0 represents that the pixel grid is blank). Each random assignment corresponds to a unique mass distribution structure. The above random assignment process is repeated to generate a sufficient number (for example, 14336) of thin plate-type acoustic metamaterial structures with different mass block distributions, so as to provide a source for subsequent sample data acquisition.

[0029] S202. Preprocess each thin-plate acoustic metamaterial structure to obtain a sample dataset.

[0030] Optionally, the sample dataset includes: pixel images of the mass distribution and acoustic transmission loss curves corresponding to each thin-plate acoustic metamaterial structure; In an exemplary embodiment, such as Figure 3 As shown, step S202 (preprocessing each thin-plate acoustic metamaterial structure to obtain a sample dataset) can specifically include: S2021. Frequency domain response calculations are performed on each thin-plate acoustic metamaterial structure to obtain the acoustic transmission loss curves corresponding to each thin-plate acoustic metamaterial structure.

[0031] Specifically, to ensure the consistency and accuracy of the acoustic transmission loss curve calculation, a sound-structure coupled simulation environment is built in finite element simulation software for various thin-plate acoustic metamaterial structures. This environment can include: A thin-plate acoustic metamaterial structure is embedded in a rectangular waveguide, and a single-sided plane wave reflection is applied inside the rectangular waveguide to simulate the propagation scenario of actual noise. Perfect matching layers are set at both ends of the rectangular waveguide to avoid the boundary reflection from interfering with the calculation of the sound transmission loss curve, ensuring that the simulation data truly reflects the sound insulation performance of the thin-plate acoustic metamaterial structure.

[0032] Furthermore, the frequency domain response of each thin-plate acoustic metamaterial structure was calculated using the aforementioned acoustic-structure coupling simulation environment, and the acoustic transmission loss curve of the thin-plate acoustic metamaterial structure in the 200-1500Hz frequency band was output.

[0033] S2022. Encode each thin-plate acoustic metamaterial structure to obtain a mass distribution pixel image. The mass distribution pixel image is a binary pixelated grayscale image. The mass block area corresponds to the black pixels of the grayscale image, and the blank area of ​​the thin plate corresponds to the white pixels of the grayscale image. Specifically, a pixelation encoding rule is formulated: for thin-plate acoustic metamaterial structures with different mass distributions, a 10×8 binary pixelated grayscale image encoding standard is defined, the mass block region in the thin-plate acoustic metamaterial structure is corresponding to the black pixels of the grayscale image, the blank region of the thin plate is corresponding to the white pixels of the grayscale image, and the grayscale value of the binary pixelated grayscale image is linearly scaled to the range of [-1,1] (where -1 is assigned to black pixels and 1 is assigned to white pixels), providing pixel conversion rules for the pixel conversion steps required in the embodiments of this application.

[0034] For example, if a pixel grid is embedded with a quality block, the corresponding pixel is encoded as a black pixel and the grayscale value is scaled to -1; if a pixel grid remains blank, the corresponding pixel is encoded as a white pixel and the grayscale value is scaled to 1.

[0035] S2023. Normalize each sound transmission loss curve, and associate the normalized sound transmission loss curve with each quality distribution pixel image to form a sample dataset.

[0036] Specifically, the acoustic transmission loss curves obtained in step S2021 are normalized, and the normalized acoustic transmission loss curves are paired one by one with their corresponding quality distribution pixel images to obtain a sample dataset. The sample dataset may include 14,336 sets of acoustic transmission loss curves and their corresponding quality distribution pixel images.

[0037] S203. Train the bidirectional neural network model based on the training dataset.

[0038] In some embodiments, the above step S203 (training the bidirectional neural network model based on the training dataset) may include: training the bidirectional neural network model based on the sample dataset to obtain a forward prediction network and a reverse design network, wherein the bidirectional neural network model includes a first network model and a second network model; the forward prediction network is used to obtain the predicted acoustic transmission loss curve based on the quality distribution pixel image, and the reverse design network is used to obtain the quality distribution pixel design image based on the target acoustic transmission loss curve.

[0039] It should be noted that before training the bidirectional neural network model based on the training dataset, the sample dataset needs to be divided into a training dataset, a validation dataset, and a test dataset.

[0040] Specifically, the sample dataset, containing acoustic transmission loss curves and corresponding quality distribution pixel images, is rationally distributed into different subsets to facilitate subsequent model training and evaluation. Typically, the data in the sample dataset is divided into three main parts: the training dataset, the test dataset, and the validation dataset. The training dataset (TrainingSet) is primarily used to train the model, and its data volume usually accounts for approximately 60%-80% of the total data. The test dataset (TestSet) is used to adjust model parameters and prevent overfitting, accounting for approximately 10%-20%. The validation dataset (ValidationSet) is used to evaluate the final performance of the model, accounting for approximately 10%-20%. For example, the `train_test_split` function provided by the `model_selection` module in the sklearn library of the Python programming language can be used to randomly distribute the dataset into the training, test, and validation datasets.

[0041] The training dataset is used to train the model. Through this dataset, the model learns the mapping from input to output, enabling it to predict new input data. The training dataset contains a large amount of sample data, which guides the model in learning and recognizing different features and patterns. Through repeated iterations and optimization on the training dataset, the model gradually grasps the regularities in the data and establishes preliminary predictive capabilities. Since the model may overfit to the data in the training dataset, the test dataset is used to evaluate the model's generalization ability, i.e., its predictive ability on new data. After model training, testing the model on the test dataset verifies its generalization ability on unseen data. The data in the test dataset is not encountered by the model during training; therefore, it allows for model tuning. The validation dataset is used for performance monitoring and parameter tuning during model training. The data in the validation dataset also comes from the original labeled corpus, but it is neither used in model training nor for final performance testing; instead, it serves as a reference standard for model tuning. Evaluation through the validation dataset can objectively reflect the model's true performance, thus avoiding overfitting or underfitting by selecting appropriate model parameters.

[0042] In an exemplary embodiment, training a bidirectional neural network model based on a training dataset to obtain a forward prediction network and a reverse design network may specifically include: Multiple quality distribution pixel images are randomly selected from the training dataset as first samples and input into the first network model. At the same time, the acoustic transmission loss curve corresponding to the first sample is extracted as the first label. The first network model is then trained iteratively multiple times to obtain a positive prediction network.

[0043] For example, the number of quality distribution pixel images used as the first sample can be 32.

[0044] Specifically, 32 quality distribution pixel images are randomly selected from the training dataset as the first sample input to the first network model, and the acoustic transmission loss curves in the 200-1500 Hz frequency band corresponding to the 32 quality distribution pixel images are extracted as the first label.

[0045] Among them, such as Figure 4 As shown, any iteration of training for the first network model includes: S411. Convolve the first sample using the first convolution kernel to obtain the first feature map.

[0046] For example, the first convolutional kernel can be a 5×5 convolutional kernel with a stride of 2.

[0047] Specifically, the quality distribution pixel image is input into the first network model (neural network model), and the spatial resolution of the quality distribution pixel image is reduced to 5×4 and the number of channels is doubled through a 5×5 convolution kernel (stride-2 downsampling) to achieve dimensionality reduction feature extraction and obtain the first feature map.

[0048] S412. Convolve the first feature map using the second convolution kernel to obtain the second feature map.

[0049] For example, the second convolution kernel can be a 3×3 convolution kernel.

[0050] Specifically, the first feature map (5×4 feature map) is convolved twice using a 3×3 convolution kernel to enhance local feature interactions (e.g., the association between adjacent pixels in the quality block distribution) and retain the consistency of the 5×4 dimension and channels to obtain the second feature map.

[0051] S413. After normalizing the second feature map, input it into the first fully connected layer, and output the predicted acoustic transmission loss curve through the first fully connected layer.

[0052] Specifically, the second feature map output by the 3×3 convolution kernel is batch normalized to optimize the feature distribution and avoid training oscillations caused by differences in data size. The normalized second feature map (5×4 feature map) is unfolded into a 640-dimensional vector and input into the first fully connected layer (hidden layer and output layer), and finally outputs a predicted acoustic transmission loss curve with 131 parameters.

[0053] S414. Calculate the predicted acoustic transmission loss curve and the first loss value of the first tag using the mean square error loss function.

[0054] The smaller the first loss value, the higher the prediction accuracy of the first network model.

[0055] S415. Adjust the weights of the first convolutional kernel, the second convolutional kernel, and the first fully connected layer based on the first loss value.

[0056] Repeat steps S411-S415 above. After 600 training iterations, the first network model gradually converges, indicating that the training of the first network model is complete, and a positive prediction network is obtained.

[0057] In an exemplary embodiment, training the bidirectional neural network model based on a training dataset to obtain a forward prediction network and a reverse design network may further include: Multiple acoustic transmission loss curves are randomly selected from the training dataset as second samples and input into the second network model. At the same time, the quality distribution pixel images corresponding to the second samples are extracted as second labels. The second network model is then trained iteratively several times to obtain the reverse-designed network.

[0058] Specifically, 32 acoustic transmission loss curves in the 200-1500 frequency band are randomly selected from the training dataset as second samples and input into the second network model. At the same time, the quality distribution pixel images corresponding to the 32 acoustic transmission loss curves in the 200-1500 frequency band are extracted as second labels.

[0059] Among them, such as Figure 5 As shown, any iteration of training the second network model includes: S421. Input the second sample into the second fully connected layer to obtain the first reconstructed feature map.

[0060] The acoustic transmission loss curve in the 200-1500 frequency band is input into the second fully connected layer. Through linear transformation and activation function (e.g., ReLU), the acoustic transmission loss curve is reconstructed into a 10×8 dimension feature map (the first reconstructed feature map) to prepare for subsequent convolution.

[0061] S422. Convolve the first reconstructed feature map through a convolutional layer to obtain the second reconstructed feature map.

[0062] For example, the convolutional layer can be a 3×3 convolution.

[0063] Perform a 3×3 convolution on the 10×8 dimensional feature map (first reconstructed feature map) to restore the spatial resolution details of the 10×8 dimensional feature map (first reconstructed feature map) (e.g., edge features of quality block distribution), and the output is still a 10×8 dimensional feature map (second reconstructed feature map).

[0064] S423. Perform global average pooling and global max pooling on the second reconstructed feature map to obtain the average sound transmission loss and the peak sound transmission loss; and generate a weighted feature map by fusing the average sound transmission loss and the peak sound transmission loss.

[0065] Specifically, firstly, global average pooling (capturing the average acoustic transmission loss between channels) and global max pooling (identifying the peak acoustic transmission loss between channels) are performed in parallel on the 10×8 dimension feature map after convolution (the second reconstructed feature map). Specifically: Global average pooling calculates the average feature value for each channel, capturing the "average acoustic transmission loss" between channels (reflecting the overall sound insulation level). Global max pooling calculates the maximum feature value for each channel, identifying the "peak acoustic transmission loss" between channels (reflecting the sound insulation capability at key frequencies).

[0066] Secondly, the two pooling results (i.e., average acoustic transmission loss and peak acoustic transmission loss) are fused and the activation function outputs a 10×8 weighted feature map with the same spatial size as the 10×8-dimensional feature map after convolution (the second reconstructed feature map). This enhances the region corresponding to the peak frequency and weakens irrelevant regions.

[0067] S424. The weighted feature map is processed by the decision function to generate a pixel image of the predicted quality distribution.

[0068] Specifically, a decision function is applied to the weighted feature map (the output value is mapped to [0,1]), and then a threshold is used to determine (e.g., if the output value is >0.5, it is determined as 1, corresponding to white pixels or blank areas; if the output value is ≤0.5, it is determined as -1, corresponding to black pixels / quality block areas), generating a predicted quality distribution pixel image, which is a 10×8 binary pixelated grayscale image.

[0069] S425. Calculate the second loss value between the predicted quality distribution pixel image and the second label using the mean square error loss function.

[0070] The smaller the second loss value, the more accurate the predicted quality block distribution of the pixel image.

[0071] S426. Adjust the weights of the second fully connected layer and the convolutional layer based on the second loss value, and adjust the fusion coefficients of global average pooling and global max pooling.

[0072] Repeat steps S421-S426 above. After 450 training iterations, the second network model gradually converges, indicating that the training of the second network model is complete, and the reverse-designed network is obtained.

[0073] In an exemplary embodiment, testing the forward prediction network and the reverse design network based on a test dataset may specifically include: The positive prediction network was tested based on the test dataset, including: Multiple quality distribution pixel images are extracted from the test dataset as first test samples and input into the forward prediction network to obtain the first test acoustic transmission loss curve; the first average error between the first test acoustic transmission loss curve and the acoustic transmission loss curve corresponding to the first test sample in the test dataset is calculated; the forward prediction network is optimized based on the first average error.

[0074] To verify the prediction accuracy of the forward prediction network, this embodiment uses data from the test dataset that was not used in training to predict the acoustic transmission loss of a thin-plate acoustic metamaterial structure. These prediction results are then compared with simulation results. The test results and simulation results show excellent consistency, confirming the feasibility of the forward network. The test results of this forward prediction network are as follows: Figure 6 As shown, the computation time required for network prediction and simulation is compared. Although the simulation takes 210 seconds, the prediction process of the forward prediction network only takes 0.4 seconds, which greatly improves the computation speed by an order of magnitude.

[0075] The reverse-engineered network was tested using a test dataset, including: Multiple acoustic transmission loss curves are extracted from the test dataset and input into the reverse design network as second test samples to obtain a first test quality distribution pixel image. The first test quality distribution pixel image is then simulated using finite element software to calculate the test acoustic transmission loss curve corresponding to the second test result. The test acoustic transmission loss curve is compared with the second test sample to obtain the comparison result. The reverse design network is then optimized based on the comparison result.

[0076] To verify the model's predictive ability, this embodiment uses randomly selected untrained acoustic transmission loss curves from the test dataset as input. The quality block distribution of the test quality distribution pixel images shows high consistency, with minor local design deviations due to differences in grayscale values. Nevertheless, through approximation, the designed structure completely corresponds to the actual structure. The test results of this reverse-designed network are as follows: Figure 7 As shown, based on the input sound transmission loss curve and the obtained results, it can be clearly seen that the reverse design network provided in this application embodiment can effectively control the sound insulation spectrum by adjusting the number and distribution pattern of mass blocks.

[0077] The design network training method for thin-plate acoustic metamaterials provided in this application firstly acquires several thin-plate acoustic metamaterial structures. The mass blocks of each structure are randomly distributed, providing a diverse sample base for subsequent deep learning model training. These samples cover distribution patterns of different mass block quantities and positions, ensuring that the neural network can fully learn the complex relationship between mass block randomness and acoustic performance. Further, each thin-plate acoustic metamaterial structure is preprocessed to obtain a sample dataset. This dataset includes: a mass distribution pixel image and an acoustic transmission loss curve corresponding to each thin-plate acoustic metamaterial structure. This step achieves digital representation of the physical structure and standardized processing of performance data, ensuring the reliability of model evaluation. Secondly, a bidirectional neural network model is trained based on the training dataset to obtain a forward prediction network and a reverse design network. The bidirectional neural network model includes a first network model and a second network model. The forward prediction network is used to obtain the predicted acoustic transmission loss curve based on the mass distribution pixel image, and the reverse design network is used to obtain the mass distribution pixel design image based on the target acoustic transmission loss curve. The forward prediction network utilizes a neural network to capture the nonlinear mapping relationship between pixel-level distribution features in a mass distribution pixel image and the frequency response characteristics of the acoustic transmission loss curve. Therefore, given any mass distribution, it can quickly generate the corresponding acoustic transmission loss prediction curve, avoiding the high computational cost of traditional finite element simulation methods. Simultaneously, the core of the reverse design network lies in learning the mapping law from the acoustic transmission loss curve to the mass distribution pixel image through a neural network. This reverse design mechanism allows for the direct derivation of a matching mass block distribution scheme based on specific acoustic performance requirements, overcoming the limitations of traditional trial-and-error methods in structural design. By constructing both the forward prediction network and the reverse design network, the rapid prediction of the acoustic transmission loss curve and the efficient generation of the mass block distribution are achieved. This approach enables rapid prediction of the acoustic transmission loss curve and efficient generation of a matching mass block distribution structure, avoiding time-consuming finite element simulation and repeated trial-and-error processes, significantly improving design efficiency.

[0078] Based on the same inventive concept, this application also provides a design method for thin-plate acoustic metamaterials. Based on the forward prediction network and inverse design network obtained through the design network training method for thin-plate acoustic metamaterials provided in any of the above embodiments, the design method for thin-plate acoustic metamaterials includes: The forward prediction network is used to obtain the predicted acoustic transmission loss curve based on the quality distribution pixel image; the reverse design network is used to obtain the quality distribution pixel design image based on the target acoustic transmission loss curve.

[0079] Based on the same inventive concept, this application also provides a design network training system for thin-plate acoustic metamaterials for implementing the design network training method for the aforementioned thin-plate acoustic metamaterials. The solution provided by this system is similar to the solution described in the above method. Therefore, the specific limitations of one or more thin-plate acoustic metamaterial design system embodiments provided below can be found in the limitations of the design method for thin-plate acoustic metamaterials described above, and will not be repeated here.

[0080] Please see Figure 8 , Figure 8 This application provides a functional module diagram of a design network training system for thin-plate acoustic metamaterials according to an embodiment of the present application. The system specifically includes: The acquisition module 801 is used to acquire several thin-plate acoustic metamaterial structures, and the mass blocks of each thin-plate acoustic metamaterial structure are randomly distributed. The processing module 802 is used to preprocess each thin-plate acoustic metamaterial structure to obtain a sample dataset; the sample dataset includes: the mass distribution pixel image and the acoustic transmission loss curve corresponding to each thin-plate acoustic metamaterial structure; Training module 803 trains the bidirectional neural network model based on the sample dataset to obtain a forward prediction network and a reverse design network. The bidirectional neural network model includes a first network model and a second network model. The forward prediction network is used to obtain the predicted acoustic transmission loss curve based on the quality distribution pixel image, and the reverse design network is used to obtain the quality distribution pixel design image based on the target acoustic transmission loss curve.

[0081] This application also provides an application scenario in which the design method of the aforementioned thin-plate acoustic metamaterial is applied. Specifically: like Figure 9 As shown, Figure 9 This application provides a schematic diagram illustrating the process of forward prediction network and reverse design network in its embodiments.

[0082] In specific application scenarios, the design method of thin-plate acoustic metamaterials mentioned in any of the above embodiments can be used to generate a predicted curve of the target acoustic transmission loss curve corresponding to any thin-plate acoustic metamaterial structure based on the mass distribution pixel image corresponding to the structure through a forward prediction network; or, the reverse design network can generate a mass distribution pixel image corresponding to any thin-plate acoustic metamaterial structure based on the target curve of the target acoustic transmission loss curve corresponding to the structure, thereby completing the reverse design of the thin-plate acoustic metamaterial structure.

[0083] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 10 As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores video tag processing data. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a network training method for designing thin-plate acoustic metamaterials.

[0084] Those skilled in the art will understand that Figure 10 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0085] In one exemplary embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0086] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0087] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0088] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0089] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0090] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0091] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0092] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for designing and training a network for thin-plate acoustic metamaterials, characterized in that, The design network training method for the thin-plate acoustic metamaterial includes: Several thin-plate acoustic metamaterial structures are obtained, and the mass blocks of each thin-plate acoustic metamaterial structure are randomly distributed. Each of the thin-plate acoustic metamaterial structures is preprocessed to obtain a sample dataset; the sample dataset includes: a mass distribution pixel image and an acoustic transmission loss curve corresponding to each thin-plate acoustic metamaterial structure; The bidirectional neural network model is trained based on the sample dataset to obtain a thin-plate acoustic metamaterial design network. The thin-plate acoustic metamaterial design network includes a forward prediction network and a reverse design network. The bidirectional neural network model includes a first network model and a second network model. The forward prediction network is used to obtain the predicted acoustic transmission loss curve based on the mass distribution pixel image, and the reverse design network is used to obtain the mass distribution pixel design image based on the target acoustic transmission loss curve.

2. The design network training method for thin-plate acoustic metamaterials according to claim 1, characterized in that, The acquisition of several thin-plate acoustic metamaterial structures specifically includes: Obtain a single thin-plate acoustic metamaterial, comprising a thin plate and several mass blocks, wherein each mass block has the same size; Using simulation software, several mass blocks are randomly matched to the thin-plate acoustic metamaterial to generate several thin-plate acoustic metamaterial structures in batches, wherein the number and / or position of the mass blocks in each thin-plate acoustic metamaterial structure are different.

3. The design network training method for thin-plate acoustic metamaterials according to claim 2, characterized in that, The thin-plate acoustic metamaterial structures described above are preprocessed to obtain a sample dataset, specifically including: Frequency domain response calculations were performed on each of the thin-plate acoustic metamaterial structures to obtain the acoustic transmission loss curves corresponding to each of the thin-plate acoustic metamaterial structures. Each of the thin-plate acoustic metamaterial structures is encoded to obtain a mass distribution pixel image, wherein the mass distribution pixel image is a binary pixelated grayscale image, the mass block region corresponds to the black pixels of the grayscale image, and the blank region of the thin plate corresponds to the white pixels of the grayscale image; The acoustic transmission loss curves are normalized, and the normalized acoustic transmission loss curves are associated with the quality distribution pixel images to form the sample dataset.

4. The design network training method for thin-plate acoustic metamaterials according to claim 3, characterized in that, The step of training the bidirectional neural network model based on the training dataset to obtain the forward prediction network and the reverse design network specifically includes: Multiple quality distribution pixel images are randomly selected from the training dataset as first samples and input into the first network model. At the same time, the acoustic transmission loss curve corresponding to the first sample is extracted as the first label. The first network model is trained iteratively multiple times to obtain the positive prediction network. Wherein, any iteration of training of the first network model includes: The first feature map is obtained by convolving the first sample with the first convolution kernel; The second feature map is obtained by convolving the first feature map with the second convolution kernel; After normalizing the second feature map, it is input into the first fully connected layer, and the predicted acoustic transmission loss curve is output through the first fully connected layer. The predicted acoustic transmission loss curve and the first loss value of the first tag are calculated using the mean square error loss function. The weights of the first convolutional kernel, the second convolutional kernel, and the first fully connected layer are adjusted based on the first loss value.

5. The design network training method for thin-plate acoustic metamaterials according to claim 4, characterized in that, The step of training the bidirectional neural network model based on the training dataset to obtain the forward prediction network and the reverse design network further includes: Multiple acoustic transmission loss curves are randomly selected from the training dataset as second samples and input into the second network model. At the same time, the quality distribution pixel images corresponding to the second samples are extracted as second labels to perform several iterations of training on the second network model to obtain the reverse design network. The training of the second network model in any iteration includes: The second sample is input into the second fully connected layer to obtain the first reconstructed feature map; The first reconstructed feature map is convolved by a convolutional layer to obtain the second reconstructed feature map; Global average pooling and global max pooling are performed on the second reconstructed feature map to obtain the average acoustic transmission loss and the peak acoustic transmission loss; and a weighted feature map is generated by fusing the average acoustic transmission loss and the peak acoustic transmission loss. The weighted feature map is processed by a decision function to generate a pixel image of predicted quality distribution. The second loss value between the predicted quality distribution pixel image and the second label is calculated using the mean square error loss function. The weights of the second fully connected layer and the convolutional layer are adjusted based on the second loss value, and the fusion coefficients of the global average pooling and global max pooling are also adjusted.

6. A design method for a thin-plate acoustic metamaterial, characterized in that, The design method for the thin-plate acoustic metamaterial, based on the forward prediction network and the inverse design network obtained by the training method described in any one of claims 1-5, includes: The positive prediction network is used to obtain the predicted acoustic transmission loss curve based on the quality distribution pixel image; The reverse design network is used to obtain a quality distribution pixel design image based on the target sound transmission loss curve.

7. A design network training system for thin-plate acoustic metamaterials, characterized in that, Based on the forward prediction network and the inverse design network obtained by the training method described in any one of claims 1-5, the design network training system for the thin-plate acoustic metamaterial includes: The acquisition module is used to acquire several thin-plate acoustic metamaterial structures, wherein the mass blocks of each thin-plate acoustic metamaterial structure are randomly distributed. The processing module is used to preprocess each of the thin-plate acoustic metamaterial structures to obtain a sample dataset; the sample dataset includes: a mass distribution pixel image and a sound transmission loss curve corresponding to each thin-plate acoustic metamaterial structure; The training module trains the bidirectional neural network model based on the sample dataset to obtain the forward prediction network and the reverse design network. The bidirectional neural network model includes a first network model and a second network model. The forward prediction network is used to obtain the predicted acoustic transmission loss curve based on the quality distribution pixel image, and the reverse design network is used to obtain the quality distribution pixel design image based on the target acoustic transmission loss curve.

8. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the steps of the design network training method for the thin-plate acoustic metamaterial according to any one of claims 1-5.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the design network training method for thin-plate acoustic metamaterials as described in any one of claims 1-5.

10. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the steps of the design network training method for thin-plate acoustic metamaterials as described in any one of claims 1-5.

Images