Deep learning wave-absorbing structure rapid design prediction method based on imaging

Abstract

The invention discloses a deep learning wave-absorbing structure rapid design prediction method based on imagination. The method comprises the following steps: acquiring image data to be predicted; the to-be-predicted image data is input into a deep learning model, a prediction result is obtained, the deep learning model is constructed through a convolutional neural network CNN and is obtained through training of a training set, the training set is a two-dimensional image array constructed through a target response curve, and the two-dimensional image array is a two-dimensional image array constructed through a target response curve; the deep learning model is used for extracting spatial features of the image data, and performing prediction according to the spatial features. According to the method, rapid design of the metamaterial can be realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of the construction of optimization and prediction methods for specific structures, and particularly relates to a rapid design and prediction method for an absorbing structure based on deep learning with visualization. Background Art

[0002] Metamaterials are materials with artificially designed structures. Their properties originate from the microscopic structure or the arrangement of units inside the materials, rather than traditional materials that rely on their chemical composition and macroscopic morphology. By precisely designing and adjusting these structural units, metamaterials can exhibit unique properties that traditional materials do not possess under multiple physical fields (such as sound fields, elastic wave fields, electric fields, thermal fields, and mechanical fields). However, the design process of metamaterials is relatively complex, and a large number of different structural units are required to achieve wave field regulation. Traditional design methods usually rely on experience and manual adjustment, resulting in a long design cycle and being unable to accurately optimize the performance of metamaterials in a short time, which greatly limits the development of metamaterials. Summary of the Invention

[0003] To solve the above technical problems, the present invention proposes a rapid design and prediction method for an absorbing structure based on deep learning with visualization, which can achieve the rapid design of metamaterials.

[0004] The present invention provides a rapid design and prediction method for an absorbing structure based on deep learning with visualization, including:

[0005] Obtaining image data to be predicted;

[0006] Inputting the image data to be predicted into a deep learning model to obtain a prediction result, wherein the deep learning model is obtained by training with a training set, the training set is a two-dimensional image array constructed by a target response curve, and the deep learning model is used to extract the spatial features of the image data and make a prediction according to the spatial features.

[0007] Optionally, obtaining the training set includes:

[0008] Obtaining a target response curve;

[0009] Extracting the maximum and minimum values of the frequency points in the target response curve, normalizing the maximum and minimum values to obtain a normalized target value;

[0010] Based on the normalized target value, calculating pixel values to obtain pixel values corresponding to the frequency points;

[0011] Filling and arranging the pixel values corresponding to the frequency points in sequence to obtain a two-dimensional image array.

[0012] Optionally, the maximum and minimum values include: the maximum value and the minimum value.

[0013] Optionally, based on the normalized target value, pixel value calculation is performed to obtain the pixel value corresponding to the frequency point, including:

[0014] The normalized target value will be multiplied by the maximum pixel value 255 of the image to obtain the pixel value corresponding to the frequency point.

[0015] Optionally, the deep learning model is constructed by a convolutional neural network CNN.

[0016] Optionally, the deep learning model includes: a convolutional layer, a pooling layer, a fully connected layer, and an activation function;

[0017] The convolutional layer is used to extract local features in the input image and obtain a local feature map;

[0018] The pooling layer is used to reduce the dimension of the local feature map through a dimensionality reduction operation;

[0019] The fully connected layer is used to integrate and make decisions on the feature information extracted by the convolutional layer and the pooling layer, and output a prediction result;

[0020] The activation function is used to perform a non-linear transformation on the feature information extracted by the convolutional layer and the pooling layer.

[0021] Optionally, training the deep learning model includes:

[0022] It is carried out by the backpropagation algorithm, uses the gradient descent method to optimize the parameters of the network, gradually minimizes the loss function, and obtains the deep learning model.

[0023] Optionally, the loss function is:

[0024]

[0025] where MSE is the mean square error, is the predicted value of the model, y i is the true value, and n is the total number of samples.

[0026] Optionally, an early stopping mechanism is also incorporated during training. When the loss of the two-dimensional image array no longer decreases, the training process will stop prematurely.

[0027] Compared with the prior art, the present invention has the following advantages and technical effects:

[0028] High efficiency: After converting the response curve data into a two-dimensional image array, the deep learning model can complete multiple parameter predictions in a short time, improving the design efficiency.

[0029] High precision: The convolutional neural network can automatically extract features, effectively improving the accuracy of the prediction results.

[0030] Strong adaptability: This method can be applied to the design of structural parameters of different types of metamaterials, including absorbing materials, stealth structures, etc. Description of the Drawings

[0031] The drawings constituting a part of this application are used to provide a further understanding of this application. The schematic embodiments of this application and their descriptions are used to explain this application and do not constitute an improper limitation of this application. In the drawings:

[0032] Figure 1 is a flowchart of a rapid design prediction method for an image-based deep learning absorbing structure according to an embodiment of the present invention;

[0033] Figure 2 is a schematic diagram of the size parameters of a vibration isolation metamaterial according to an embodiment of the present invention;

[0034] Figure 3 is a schematic diagram of the prediction result according to an embodiment of the present invention. Detailed Description of the Embodiment

[0035] It should be noted that, without conflict, the embodiments in this application and the features in the embodiments can be combined with each other. The following will refer to the drawings and combine the embodiments to detail this application.

[0036] It should be noted that the steps shown in the flowchart of the drawings can be executed in a computer system such as a set of computer-executable instructions, and although the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in a different order than here.

[0037] This embodiment proposes a rapid design prediction method for an image-based deep learning absorbing structure, as Figure 1 shown, specifically including the following steps:

[0038] Obtain the image data to be predicted;

[0039] Input the image data to be predicted into the deep learning model to obtain the prediction result, where the deep learning model is obtained by training with a training set, the training set is a two-dimensional image array constructed by the target response curve, and the deep learning model is used to extract the spatial features of the image data and make predictions based on the spatial features.

[0040] Furthermore, obtaining the training set includes:

[0041] Obtain the target response curve;

[0042] Extract the maximum and minimum values of the frequency points in the target response curve, and perform normalization processing on the maximum and minimum values to obtain the normalized target values;

[0043] Based on the normalized target values, pixel values are calculated to obtain the pixel values corresponding to the frequency points;

[0044] The pixel values corresponding to the frequency points are filled and arranged in order to obtain a two-dimensional image array.

[0045] Specifically, the target response curve is the reflection / transmission / absorption coefficient value corresponding to each frequency point. Through discretization and normalization, it is constructed into a two-dimensional image array.

[0046] The response data of each group of parameters can generate a single-channel, two-channel or three-channel image matrix:

[0047] Single-channel: When only a group of curve data is input (for example: S11 amplitude curve), the response value of each frequency point will be mapped to a pixel value in the image. The finally generated image is a single-channel two-dimensional matrix with certain gray-scale changes, reflecting the changes of the curve at each frequency point.

[0048] Two-channel: When two groups of curve data are input, the response values of each group of curves will be mapped to different channels respectively. In this way, the information of two groups of frequency response curves can be displayed in the same image, and the information of each group of curves will be reflected in two channels at the same time. Each pixel of the image is jointly composed of the mapped values of the two groups of response data.

[0049] Construction method:

[0050] Extraction of maximum and minimum values: Before constructing the image, it is necessary to first extract the maximum and minimum values of each frequency point from the input curve data. These values are used for subsequent normalization processing to ensure that the data is unified within a standard range. By extracting the maximum and minimum values, the influence of outliers in the data can be reduced, making the response values of each frequency point comparable.

[0051] The normalization formula is: The key to normalization is to compress the data into the range of 0 and 1 through the minimum and maximum values of the data, so that different response data can be under the same processing standard, thus avoiding calculation inconsistencies caused by data scale differences:

[0052]

[0053] where, x norm represents the normalized value, x is the original response value, x min and x max represent the maximum and minimum values in this dataset respectively.

[0054] Pixel value calculation: The normalized target value will be multiplied by the maximum pixel value of the image (usually 255) to obtain the pixel value corresponding to that point. For a two-dimensional image, the response value at each frequency point is converted into a pixel, forming a single luminance information in the image.

[0055] Image matrix construction: The pixel values corresponding to each frequency point are filled into a two-dimensional matrix in sequence, and finally an image is generated. These matrices usually adopt a square arrangement to match the image size with the number of frequency points. In this way, the response curve data of each frequency point is transformed into a pixel matrix of a two-dimensional image, making the original curve data visually presented effectively.

[0056] Furthermore, the extreme values include: the maximum value and the minimum value.

[0057] Furthermore, based on the normalized target value, pixel value calculation is performed to obtain the pixel value corresponding to the frequency point, including:

[0058] The normalized target value is multiplied by the maximum pixel value of the image to obtain the pixel value corresponding to the frequency point.

[0059] Furthermore, the deep learning model is constructed by a convolutional neural network CNN;

[0060] The deep learning model includes: a convolutional layer, a pooling layer, a fully connected layer, and an activation function;

[0061] Convolutional layer, the convolutional layer is used to extract local features in the input image, especially the local patterns and important information that may exist in the frequency response curve. The convolutional layer extracts features from the input image through a convolution operation (i.e., sliding the convolutional kernel), and is usually used for the recognition of basic features such as edges, textures, and shapes in the image.

[0062] Pooling layer, the pooling layer reduces the dimension of the feature map through a dimensionality reduction operation (such as max pooling or average pooling), reducing the computational amount and memory occupancy. The role of the pooling layer is not only dimensionality reduction, but also to enhance the robustness of the network and reduce the risk of overfitting.

[0063] Fully connected layer, after several convolutional and pooling layers, the network will transfer the extracted feature information to the fully connected layer. The fully connected layer is responsible for integrating and making decisions on the abstract features, and finally outputs the prediction result of the model.

[0064] Activation function, the output of each layer usually undergoes a non-linear transformation through an activation function. Commonly used activation functions include ReLU (Rectified Linear Unit) and Sigmoid function, etc.

[0065] Specifically, the present invention constructs a deep learning model based on the Convolutional Neural Network (CNN) architecture. CNN performs excellently in fields such as image recognition and speech processing, and can effectively extract spatial features from two-dimensional images. The input of this model is the two-dimensional image matrix generated through the above steps, and the network learns the feature information of these images to make effective predictions and classifications.

[0066] Further, training the deep learning model includes:

[0067] It is carried out through the backpropagation algorithm, uses the gradient descent method to optimize the parameters of the network, gradually minimizes the loss function, and obtains the deep learning model.

[0068] Further, the loss function is:

[0069]

[0070] where MSE is the mean squared error, is the predicted value of the model, y i is the true value, and n is the total number of samples.

[0071] Further, an early stopping mechanism is also incorporated during the training process. When the loss of the two-dimensional image array no longer decreases, the training process will stop prematurely.

[0072] Specifically, to prevent the model from overfitting during training, this embodiment introduces an early stopping mechanism. When the loss of the validation set no longer decreases within several rounds, the training process will stop prematurely. This can not only effectively avoid overfitting, but also save computing resources and reduce unnecessary training time. The early stopping mechanism dynamically adjusts the training process based on the change of the validation set loss, so as to ensure that the model can stop at the optimal training rounds.

[0073] More specifically, to further improve the performance of the model, an automated hyperparameter optimization technique is adopted. Tools such as Keras Tuner are used in this article for hyperparameter tuning, automatically adjusting hyperparameters such as the number of convolutional layers, the size of convolutional kernels, the learning rate, and the batch size. Hyperparameter optimization can significantly improve the accuracy of the model and reduce the complexity of manual adjustment.

[0074] (1) Number of convolutional layers and size of convolutional kernels: The number of convolutional layers and the size of convolutional kernels are key factors affecting the performance of the convolutional neural network. A reasonable number of layers and size of convolutional kernels can extract more and more effective feature information.

[0075] (2) Learning rate: An appropriate learning rate can effectively control the convergence speed of the model. Too high a learning rate may lead to unstable training, while too low a learning rate may lead to an overly slow training process.

[0076] (3) Batch Size: The batch size has an important impact on the training speed and convergence. Selecting an appropriate batch size can accelerate the training process and avoid overfitting.

[0077] After the model training is completed, it needs to be verified through the test set to evaluate its performance. The specific performance metrics include:

[0078] Mean Squared Error (MSE): MSE is an important metric for evaluating the prediction error of a regression model. A smaller MSE value indicates that the model's prediction is more accurate. Through MSE, the size of the model's prediction error can be quantified, providing a direction for model improvement.

[0079] Coefficient of Determination (R 2 ):The coefficient of determination (R 2 ) is used to measure the degree of fit of the model to data changes, and its value range is [0, 1]. A higher R 2 value indicates that the model can better fit the data and capture the trend of the data.

[0080] When validating on the test set, the present invention needs to comprehensively consider these two metrics to ensure that the model not only performs well on the training set but also has good generalization ability on unseen data.

[0081] Fast Prediction and Model Deployment:

[0082] Fast Prediction:

[0083] The trained deep learning model can be quickly put into practical applications for real-time prediction. In the application environment, users only need to input the target curve data, and the model can quickly output the prediction results. This real-time prediction function is of great significance for engineering designs that require quick responses.

[0084] Batch Prediction: In addition to single predictions, the model also supports batch data input, enabling the processing of the design requirements of multiple parameters at once. This allows users to improve the prediction efficiency and optimize the design cycle when conducting large-scale designs.

[0085] Model Deployment:

[0086] To achieve wide applications, the trained deep learning model will be deployed to the actual production environment. Through methods such as API interfaces, the model can be integrated with other systems to support online prediction and real-time parameter optimization. In addition, the model also supports the processing of batch input data to meet the requirements of multi-parameter designs.

[0087] The following will elaborate on this embodiment in conjunction with the accompanying drawings:

[0088] This embodiment introduces an application of predicting the performance of vibration isolation structures through a Convolutional Neural Network (CNN) model, mainly for modeling the mechanical properties of materials. Vibration isolation superstructures are usually used to reduce vibration transmission to improve the stability and working efficiency of equipment. To achieve rapid design and optimization of materials, this method trains a deep learning model to predict their vibration isolation performance by inputting the basic physical parameters of the materials (such as elastic modulus, density, hardness, etc.) and structural feature data (such as thickness, porosity, etc.), and then provides optimization suggestions for material design.

[0089] Step 1: Data Preparation:

[0090] Original Data: Collect the physical parameters and geometric features of different types of vibration isolation structures. The data includes:

[0091] Elastic modulus, Poisson's ratio, density, material thickness, unit size, internal parameter size, etc.

[0092] Response Data: By changing the structural parameters, measure the vibration isolation performance of structures with different parameters.

[0093] These performances include: vibration isolation frequency response, vibration attenuation coefficient, etc.

[0094] Image Construction: Preprocess the performance data of each material (such as normalization, standardization, etc.) and construct it into a two-dimensional image array. Each image represents a structural instance, where each pixel value of the image represents a certain physical property (such as elastic modulus, density, etc.). This step converts traditional numerical data into an image format that can be processed by a convolutional neural network.

[0095] Step 2: Data Partitioning:

[0096] Randomly partition the processed image data into a training set, a validation set, and a test set. Usually, the data is partitioned in a ratio of 8:2 or 7:3 to ensure that the training set has enough data for model training, and at the same time, the test set can effectively evaluate the generalization ability of the model.

[0097] Introduce a validation set for hyperparameter tuning and preventing overfitting.

[0098] Step 3: Model Construction:

[0099] Convolutional Neural Network (CNN) Architecture: This embodiment uses a deep convolutional neural network (CNN) to process the image data. The network structure includes:

[0100] Convolutional Layer: Extract local features in the image through convolution operations to identify the basic properties and structural features of the material.

[0101] Pooling Layer: Reduces the size of the image through pooling operations, while preserving important feature information and reducing computational complexity.

[0102] Fully Connected Layer: Integrates the feature information extracted by the convolutional layer and finally outputs the predicted values of the vibration isolation performance of the material (such as vibration isolation coefficient, vibration attenuation amount, etc.).

[0103] Step 4: Loss Function and Optimizer:

[0104] Loss Function: In this embodiment, the mean squared error (MSE) is used as the loss function to measure the difference between the predicted value and the true value. The calculation formula of the mean squared error is as follows:

[0105]

[0106] where, is the predicted value of the model, y i is the true value, and n is the total number of samples.

[0107] Optimizer: The Adam optimizer is used for model training. The Adam optimizer adjusts the learning rate adaptively to improve the convergence speed and reduce the training time.

[0108] Step 5: Hyperparameter Tuning and Early Stopping Mechanism:

[0109] Hyperparameter Tuning: Use an automated tool (such as Keras Tuner) for hyperparameter optimization. By adjusting hyperparameters such as the number of convolutional layers, the size of the convolutional kernel, the learning rate, and the batch size, the performance of the model is optimized.

[0110] Early Stopping Mechanism: To prevent overfitting, an early stopping mechanism is introduced. When the loss on the validation set no longer decreases within a certain number of rounds, the training is stopped in advance to improve the generalization ability of the model.

[0111] Step 6: Model Evaluation and Validation:

[0112] Evaluation Metrics: Use the mean squared error (MSE) and the coefficient of determination (R 2 ) to evaluate the model. MSE is used to quantify the prediction error, and the R 2 value evaluates the fitting effect of the model.

[0113] Test Set Validation: Perform validation on the test set, calculate the prediction effect of the model on new data, and ensure the accuracy and reliability of the model.

[0114] Step 7: Fast Prediction and Model Deployment:

[0115] Real-time Prediction: Deploy the trained CNN model to the actual application environment, such asFigure 2 Input the parameters of the newly designed vibration isolation material (such as the physical properties and structure of the material) as shown, and obtain the prediction results, such as Figure 3 shown, and predict its vibration isolation performance in real time.

[0116] Batch prediction: Support batch prediction of multiple vibration isolation materials, which is convenient for material designers to screen out the optimal material combination in a short time.

[0117] The above is only a preferred specific implementation mode of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed in the present application should be covered by the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A rapid design prediction method for an image-based deep learning absorbing structure, characterized in that, Including: Obtain the image data to be predicted; Input the image data to be predicted into the deep learning model to obtain a prediction result, where the deep learning model is obtained by training with a training set, the training set is a two-dimensional image array constructed by a target response curve, and the deep learning model is used to extract the spatial features of the image data and make a prediction based on the spatial features.

2. A rapid design prediction method for an image-based deep learning absorbing structure according to claim 1, characterized in that Obtaining the training set includes: Obtain the target response curve; Extract the maximum and minimum values of the frequency points in the target response curve, perform normalization processing on the maximum and minimum values to obtain the normalized target values; Based on the normalized target values, perform pixel value calculation to obtain the pixel values corresponding to the frequency points; Fill and arrange the pixel values corresponding to the frequency points in sequence to obtain a two-dimensional image array.

3. A rapid design prediction method for an image-based deep learning absorbing structure according to claim 2, characterized in that The maximum and minimum values include: the maximum value and the minimum value.

4. A rapid design prediction method for an image-based deep learning absorbing structure according to claim 2, characterized in that Based on the normalized target values, performing pixel value calculation to obtain the pixel values corresponding to the frequency points includes: The normalized target value will be multiplied by the maximum pixel value 255 of the image to obtain the pixel value corresponding to the frequency point.

5. A rapid design prediction method for an image-based deep learning absorbing structure according to claim 1, characterized in that, The deep learning model is constructed by a convolutional neural network CNN.

6. A method for rapid design prediction of an image-based deep learning absorbing structure according to claim 5, characterized in that The deep learning model includes: a convolutional layer, a pooling layer, a fully connected layer, and an activation function; The convolutional layer is used to extract local features in the input image to obtain a local feature map; The pooling layer is used to reduce the dimension of the local feature map through a dimensionality reduction operation; The fully connected layer is used to integrate and make decisions on the feature information extracted by the convolutional layer and the pooling layer, and output a prediction result; The activation function is used to perform a non-linear transformation on the feature information extracted by the convolutional layer and the pooling layer.

7. A rapid design prediction method for an image-based deep learning absorbing structure according to claim 6, characterized in that Training the deep learning model includes: It is carried out by the backpropagation algorithm, uses the gradient descent method to optimize the parameters of the network, gradually minimizes the loss function, and obtains the deep learning model.

8. A rapid design prediction method for an image-based deep learning absorbing structure according to claim 7, characterized in that The loss function is: Among them, MSE is the mean square error, is the predicted value of the model, y i is the true value, and n is the total number of samples.

9. A rapid design prediction method for an image-based deep learning absorbing structure according to claim 7, characterized in that An early stopping mechanism is also incorporated during training. When the loss of the two-dimensional image array no longer decreases, the training process will stop prematurely.

Images