Data-driven composite material strain field distribution prediction method

Abstract

The invention discloses a data-driven composite material strain field distribution prediction method, and belongs to the technical field of composite material mechanical property prediction. The method comprises the following steps: firstly, collecting a microstructure graph of the composite material and a corresponding actual strain field distribution graph, and constructing a data set after preprocessing; downsampling and feature extraction are carried out on the preprocessed image through an encoder, the encoder adopts a multi-stage residual structure, and each stage comprises a downsampling residual block and a standard residual block; obtaining potential variables from coding output sampling through a re-parameterization technique; performing up-sampling and feature reconstruction through a decoder with a symmetrical structure; and finally obtaining a normalized strain field prediction map through an output layer. And after the network training is completed, inputting the microstructure graph of the to-be-tested material to output the predicted strain field distribution. According to the method, accurate and efficient mapping from the microstructure of the material to the mechanical property is realized, and a reliable data driving solution is provided for strain analysis of the composite material.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of composite material mechanical property prediction, and particularly relates to a data-driven composite material strain field distribution prediction method. BACKGROUND

[0002] The strain field distribution of a composite material in a stress process is a key parameter for evaluating the mechanical properties and damage behavior of the composite material. In the prior art, strain field measurement relies on Digital Image Correlation (DIC) technology. As a non-contact full-field strain measurement method, DIC technology can obtain the full-field strain distribution on the surface of a material by analyzing the speckle pattern on the surface of the material. Although this method can provide accurate strain distribution data, it still has limitations: the experimental setup is complex, a random speckle pattern needs to be prepared, and image acquisition and processing are time-consuming; high-precision DIC analysis requires high computing resources, and it is difficult to achieve real-time or high-throughput analysis.

[0003] In recent years, deep learning methods, especially convolutional neural networks, have shown great capabilities in image recognition and regression prediction, can extract multi-level features from images, and establish a mapping relationship from input images to output images, and have achieved remarkable results in medical image analysis, autonomous driving and other fields. If a convolutional neural network is used to train the existing composite material micro-topography and the corresponding DIC measured strain field distribution map, it is expected to directly predict the strain field distribution map from the composite material micro-topography, and to predict the mechanical behavior with high precision and high efficiency, and to guide the development of new high-strength and lightweight composite materials in engineering applications. SUMMARY

[0004] The purpose of the present application is to overcome the limitations of the existing in-situ experiment combined with DIC testing technology in composite material strain field analysis, which has a long experimental process and is tedious to work, and to provide a data-driven composite material strain field distribution prediction method.

[0005] The specific technical solutions adopted by the present application are as follows:

[0006] The present application provides a data-driven composite material strain field distribution prediction method, and the specific steps are as follows:

[0007] S1: Collecting multiple groups of micro-topography and corresponding actual strain field distribution maps of composite materials under different strain levels; performing size unification, gray scale and normalization processing on the micro-topography and actual strain field distribution maps, and performing data enhancement to obtain a data set;

[0008] S2: resizing the data set in the data set to The gray-scale image of the pixel is input into a network input layer for processing to output a preliminary feature map; the network input layer comprises, in sequence, a convolutional layer, a batch normalization layer and a ReLU activation function layer;

[0009] S3: inputting the preliminary feature map into an encoding module for downsampling and feature extraction; the encoding module comprises three stages, each stage being composed of one down-sampling residual block and a plurality of standard residual blocks; the down-sampling residual block and the standard residual block are each composed of L convolutional layers, each convolutional layer being connected with a batch normalization layer and a ReLU activation function layer, and the input and the output being added through a shortcut connection;

[0010] S4: based on the mean and the variance output by the encoding module of step S3, the latent variable is obtained by sampling through a reparameterization technique;

[0011] S5: inputting the latent variable into a decoding module symmetrical to the encoding module of step S3, and performing upsampling and reconstruction based on the latent variable to obtain a high-level feature map; the decoding module comprises three stages, each stage being composed of one up-sampling residual block and a plurality of standard residual blocks;

[0012] S6: inputting the high-level feature map into a network output layer to output a predicted strain field distribution map normalized to the range of [0, 1]; the network output layer comprises, in sequence, a convolutional layer and a Sigmoid activation function layer;

[0013] S7: training the network model constructed in steps S2-S6 using the data set in step S1; inputting a micro-morphology map of a composite material to be predicted into the trained network model to obtain a normalized predicted strain field distribution map; and obtaining a real strain field distribution map of the composite material after denormalization.

[0014] Preferably, 500-1000 groups of data of the composite material are collected in step S1, and the strain level is 0.5%-5%; the micro-morphology map is obtained by a scanning electron microscope; and the strain field distribution map is obtained by a digital image correlation technique; wherein 128, 256, 512 or 1024 are taken;

[0015] The data enhancement operation in step S1 includes random horizontal or vertical flipping, random rotation of -10°-+10°, and random adjustment of brightness and contrast of -10%-+10%; the data set is randomly divided into a training set, a validation set and a test set at a ratio of 7:2:1.

[0016] Preferably, the convolution kernel size of the convolution layer in the network input layer in step S2 is 7*7, the input channel number is 1, the output channel number is , and the convolution step is 2.

[0017] Preferably, in the encoding module in step S3: in the down-sampling residual block of the first-stage encoder, the input channel number is , the output channel number is , the main path is L 7*7 convolution, and one standard residual block is connected after the down-sampling residual block; in the down-sampling residual block of the second-stage encoder, the input channel number is , the output channel number is , the main path is L 5*5 convolution, and two standard residual blocks are connected after the down-sampling residual block; in the down-sampling residual block of the third-stage encoder, the input channel number is , the output channel number is 2 , the main path is L 3*3 convolution, and one standard residual block is connected after the down-sampling residual block.

[0018] The output feature map of the third-stage encoder is flattened into a one-dimensional vector at the tail of the encoding module, and then two independent fully connected layers are used to output the logarithm of the mean vector and the variance vector of the latent space, respectively.

[0019] Further, the number of layers of the convolution layer in step S3 satisfies the corresponding relationship with the pixel value : , wherein the convolution step of the first convolution layer is 2, and the convolution steps of the remaining layers are all 1; the convolution kernel size of the shortcut connection is 1*1, and the convolution step is 2.

[0020] Preferably, the expression of the latent variable in step S4 is ;

[0021] In the formula: is the mean vector output by the encoding module, is the square root vector of the variance, is a random noise vector sampled from a standard normal distribution, represents element-wise multiplication.

[0022] Preferably, the head of the decoding module in step S5 is provided with a fully connected layer for mapping the latent variable to a feature map ; the input channel number of the up-sampling residual block of the first-stage decoder is , and the output channel number is , the input channel number of the up-sampling residual block of the second stage decoder is , the output channel number is , the main path is L 5x5 convolution, the convolution kernel size of the shortcut connection is 5x5, the convolution step is 2, and then 2 standard residual blocks are connected; the input channel number of the up-sampling residual block of the third stage decoder is , the output channel number is , the main path is L 7x7 convolution, the convolution kernel size of the shortcut connection is 7x7, the convolution step is 2, and then 1 standard residual block is connected; the up-sampling residual block uses transpose convolution for up-sampling.

[0023] Preferably, in step S6, the convolution kernel size of the convolution layer in the network output layer is 4x4, the input channel number is , the output channel number is 1, and the convolution step is 2.

[0024] Preferably, in step S7, the network model is trained using the PyTorch deep learning framework based on the set loss function, optimizer and training hyperparameters.

[0025] Further, the optimizer uses the Adam optimizer; the initial learning rate is set to 0.00001-0.0001; the cosine annealing scheduling strategy is used; the training period is 500-800 training rounds; the batch size is selected as 8, 32 or 64 according to the graphics processing unit; and the validation set early stopping patience value is set to 10-20 training rounds.

[0026] Compared with the prior art, the present application has the following beneficial effects:

[0027] (1) The present application provides a series of scanning electron microscope (SEM) images and corresponding strain field distribution maps generated by DIC technology analysis as data, trains a convolutional neural network model, and realizes end-to-end prediction from composite material micro-morphology maps to strain field distribution maps. This method can directly generate corresponding strain field distribution maps from composite material micro-morphology maps without the need for traditional in-situ experiments combined with DIC test technology and the tedious characterization process, avoiding the complex image correlation calculation and parameter optimization process in traditional DIC technology, and significantly improving the analysis efficiency.

[0028] (2) The model building and training process of the present application is convenient and simple to operate, only needs to adjust the corresponding parameters according to the actual needs, and can complete the training of the model, which is convenient for users to operate.

[0029] (3) This invention has good scalability and is applicable to different types of composite materials and loading conditions. It can also be further extended to other materials, which is of great significance in academic research and can also provide effective guidance for the design and development of industrial materials. Attached Figure Description

[0030] Figure 1 The network model structure diagram provided by this invention;

[0031] Figure 2 The image shows the predicted strain field distribution of the SiC particle-reinforced aluminum alloy composite material under different strains in Example 1. Detailed Implementation

[0032] The present invention will be further described and illustrated below with reference to the accompanying drawings and specific embodiments. The technical features of each embodiment of the present invention can be combined accordingly, provided that there is no mutual conflict.

[0033] This invention provides a data-driven method for predicting strain field distribution in composite materials. The overall network model structure is as follows: Figure 1 As shown. The specific steps are as follows:

[0034] Step 1: Data Collection and Preprocessing

[0035] (1) Collect 500 to 1000 sets of paired data from transverse compression experiments of composite materials. Each set of data includes scanning electron microscope images (SEM morphology maps) obtained at a specific strain level (0.5% to 5%), and the corresponding actual strain field distribution map calculated by digital image correlation (DIC) technology.

[0036] (2) Unify image size: Unify the size of all SEM topography images and actual strain field distribution images to the same size. pixels, of which Choose 128, 256, 512, or 1024 to ensure batch processing efficiency and meet network input requirements.

[0037] (3) Gray normalization: Convert the above image into a grayscale image and transform the pixel values ​​to [0,1].

[0038] (4) Data augmentation: Perform augmentation operations on the data, including random horizontal or vertical flipping, random rotation from -10° to +10°, and random brightness and contrast adjustment from -10% to +10%, to expand the dataset and improve the generalization ability and robustness of the model. The processed dataset is randomly divided into training set, validation set and test set in a ratio of 7:2:1.

[0039] Step 2: Network Input Layer Processing

[0040] The central size of the dataset obtained in step one is... A grayscale image of pixels is input into the network's input layer for processing, and a preliminary feature map is output.

[0041] The network input layer consists of a convolutional layer, a batch normalization layer, and a ReLU activation function layer connected in sequence. The convolutional layers have a 7×7 kernel size, 1 input channel, and [number missing] output channels. The convolution stride is 2. The final output preliminary feature map size is... .

[0042] Step 3: Encoding Processing

[0043] The encoder module is designed to extract multi-level features from the input SEM topography image and map them to the statistical distribution in the latent space. The encoding module consists of three stages, each comprising a downsampled residual block and several standard residual blocks. Both the downsampled and standard residual blocks are composed of L convolutional layers. Each convolutional layer is followed by a batch normalization layer and a ReLU activation function layer. The input and output are summed via a shortcut connection to address the vanishing gradient problem in deep networks.

[0044] Number of layers With pixel value The following correspondence is satisfied:

[0045]

[0046] That is when When, corresponding Furthermore, the stride of the first convolutional layer is 2, while the stride of the rest is 1. Specifically:

[0047] (1) The number of input channels in the downsampling residual block of the first-stage encoder is The number of output channels is The main path consists of L 7×7 convolutions, with 1×1 kernels in the shortcut connections and a stride of 2. This downsampling residual block is followed by a convolutional path with inputs and outputs of 1×1. The standard residual block of the channel. The output feature map size is... .

[0048] (2) The number of input channels in the downsampling residual block of the second-stage encoder is The number of output channels is The main path consists of L 5×5 convolutions, with 1×1 kernels in the shortcut connections and a stride of 2. This downsampling residual block is followed by two inputs and two outputs. The standard residual block. The output feature map size is... .

[0049] (3) The number of input channels in the downsampling residual block of the third-stage encoder is The number of output channels is 2 The main path consists of L 3×3 convolutions, with the shortcut connections using 1×1 kernels and a stride of 2. This downsampling residual block is followed by a convolutional path with inputs and outputs of 1×1. The standard residual block.

[0050] (4) The encoder tail flattens the output feature map obtained in the above steps into a one-dimensional vector, and then outputs the mean vector of the latent space through two independent fully connected layers. Sum of variance vectors The logarithm of .

[0051] Step 4: Constructing the potential space

[0052] The mean value output in step three and variance Latent variables are obtained through reparameterization techniques. latent variables The expression is ;

[0053] In the formula: It is the mean vector output by the encoding module. It is the square root vector of the variance. It is a random noise vector sampled from the standard normal distribution. This indicates element-wise multiplication. This technique allows the gradient to propagate back through the sampling operation.

[0054] Step 5: Decoding Process

[0055] The decoding module is responsible for decoding from latent variables. The corresponding high-level feature map (strain field distribution map) is reconstructed, which also includes three stages. Each stage consists of an upsampled deconvolution residual block and several standard residual blocks. The deconvolution residual block is upsampled using transposed convolution. The structure of the decoding module is symmetrical to that of the encoding module.

[0056] (1) The header of the decoding module is set to store latent variables Mapped to Fully connected layer of feature maps.

[0057] (2) The number of input channels of the upsampled residual block of the first-stage decoder is The number of output channels is The main path consists of L 3×3 convolutions, with the shortcut connections having 3×3 kernels and a stride of 2, followed by one input and output convolution. The standard residual block of the channel. The output feature map size is... .

[0058] (3) The number of input channels of the upsampled residual block of the second-stage decoder is The number of output channels is The main path consists of L 5×5 convolutions, with the shortcut connections having 5×5 kernels and a stride of 2, followed by two inputs and outputs that are both 5×5. The standard residual block of the channel. The output feature map size is... .

[0059] (4) The number of input channels of the upsampled residual block of the third-stage decoder is The number of output channels is The main path consists of L 7×7 convolutions, with the shortcut connections having 7×7 kernels and a stride of 2, followed by one input and output convolution. The standard residual block for the channel. The output feature map (high-level feature map) size is... .

[0060] Step Six: Network Output Layer Processing

[0061] The aforementioned high-level feature maps are input into the network output layer, which is a single output layer. Convolution, with the number of input channels being The output has 1 channel, a convolution stride of 2, and is followed by a sigmoid activation function to normalize the output values ​​to the range [0,1], resulting in the predicted strain field distribution map. The output image size is [size missing]. .

[0062] Step 7: Train the model

[0063] The constructed network model is trained using the dataset from step one, as follows:

[0064] (1) Setting the loss function: A combined loss function is adopted, which combines the reconstruction loss and KL divergence loss with certain weights to simultaneously optimize numerical accuracy and structural fidelity. Among them, the reconstruction loss uses mean square error to measure the pixel difference between the strain field distribution map output by the network and the strain field distribution map obtained by the actual DIC analysis. The KL divergence loss is used to normalize the distribution of the latent space to approximate the standard normal distribution, ensuring the continuity and smoothness of the latent space.

[0065] (2) Set up the optimizer: Use the Adam optimizer, set the initial learning rate to 0.0001, and adopt the cosine annealing scheduling strategy.

[0066] (3) Training parameters. The training cycle is 500 to 800 training epochs, and the batch size is adjusted to 8, 32, or 64 depending on the graphics processing unit (GPU); the validation set is used to monitor the training process, and the early stopping patience value is set to 10 to 20 epochs.

[0067] (4) Hardware and software. Training was performed on a workstation equipped with an AMD 9950X 5090 GPU, based on the PyTorch deep learning framework.

[0068] Step 8: Predict the strain field distribution map

[0069] The preprocessed SEM topography image of the composite material is input into the trained model. The network propagates forward, and the decoder outputs the corresponding predicted strain field distribution map. The output normalized image is denormalized and converted into strain values ​​with actual physical meaning based on the statistics of the training set labels. It can also be further converted into a color image to directly observe the strain concentration area.

[0070] Example 1

[0071] This embodiment uses the above method to predict the strain field distribution of SiC particle-reinforced aluminum alloy composites under different strains. The results are as follows: Figure 2 As shown, (a) to (d) are strain field distribution maps predicted by the model, using SEM morphology images of SiC particle-reinforced aluminum alloy composites at strains of 0.6%, 0.8%, 1.0%, and 1.2% as input. The changes in the strain field can be intuitively observed from the prediction maps, proving that the method provided by this invention can accurately and efficiently predict the strain field based on microscopic morphology images.

[0072] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the invention. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the invention. Therefore, all technical solutions obtained through equivalent substitution or transformation fall within the protection scope of the present invention.

Claims

1. A data-driven method for predicting strain field distribution in composite materials, characterized in that, The specific steps are as follows: S1: Collect multiple sets of microstructure images of composite materials at different strain levels and corresponding actual strain field distribution maps; perform size unification, grayscale conversion and normalization on the microstructure images and actual strain field distribution maps, and perform data augmentation to obtain the dataset; S2: The dataset with dimensions of... A grayscale image of pixels is input into the network input layer for processing, and a preliminary feature map is output; the network input layer includes a convolutional layer, a batch normalization layer and a ReLU activation function layer connected in sequence; S3: The preliminary feature map is input into the encoding module for downsampling and feature extraction; the encoding module contains three stages, each stage consisting of a downsampled residual block and several standard residual blocks; both the downsampled residual block and the standard residual block are composed of L convolutional layers, each convolutional layer is followed by a batch normalization layer and a ReLU activation function layer, and the input and output are added through a shortcut connection; S4: The mean value based on the output of the encoding module in step S3 and variance Latent variables are obtained through reparameterization techniques. ; S5: The latent variables The input is fed into the decoding module, which is symmetrical to the encoding module in step S3, based on the latent variables. Upsampling and reconstruction are performed to obtain a high-level feature map; the decoding module consists of three stages, each stage consisting of an upsampled residual block and several standard residual blocks; S6: Input the high-level feature map into the network output layer and output a predicted strain field distribution map normalized to the range [0,1]; the network output layer includes a convolutional layer and a sigmoid activation function layer connected in sequence; S7: Use the dataset from step S1 to train the network model constructed in steps S2 to S6; input the microstructure map of the composite material to be predicted into the trained network model to obtain the normalized predicted strain field distribution map. After inverse normalization, the true strain field distribution of the composite material is obtained.

2. The data-driven method for predicting strain field distribution in composite materials according to claim 1, characterized in that, In step S1, 500-1000 sets of composite material data are collected, with strain levels ranging from 0.5% to 5%; the microstructure images are obtained using scanning electron microscopy; the strain field distribution maps are obtained using digital image correlation technology; wherein... Choose 128, 256, 512, or 1024; The data augmentation operations described in step S1 include random horizontal or vertical flipping, random rotation from -10° to +10°, and random adjustment of brightness and contrast from -10% to +10%; the dataset is randomly divided into training set, validation set, and test set in a ratio of 7:2:

1.

3. The data-driven method for predicting strain field distribution in composite materials according to claim 1, characterized in that, In step S2, the convolutional kernel size of the input layer in the network is 7×7, the number of input channels is 1, and the number of output channels is [missing information]. The convolution stride is 2.

4. The data-driven method for predicting strain field distribution in composite materials according to claim 1, characterized in that, In step S3, within the encoding module: the number of input channels in the downsampling residual block of the first-stage encoder is... The number of output channels is The main path consists of L 7×7 convolutions, followed by one standard residual block; the number of input channels in the downsampling residual block of the second-stage encoder is... The number of output channels is The main path consists of L 5×5 convolutions, followed by two standard residual blocks after the downsampled residual block; the number of input channels in the downsampled residual block of the third-stage encoder is... The number of output channels is 2 The main path consists of L 3×3 convolutions, and this downsampled residual block is followed by one standard residual block. At the end of the encoding module, the output feature map of the third-stage encoder is flattened into a one-dimensional vector, and then passed through two independent fully connected layers to output the mean vector of the latent space. Sum of variance vectors The logarithm of .

5. The data-driven method for predicting strain field distribution of composite materials according to claim 4, characterized in that, The number of convolutional layers in step S3 With pixel value Satisfy the correspondence: The first convolutional layer has a stride of 2, and the strides of the others are all 1; the convolutional kernel size of the shortcut connection is 1×1, and the stride is 2.

6. The data-driven method for predicting strain field distribution in composite materials according to claim 1, characterized in that, Latent variables in step S4 The expression is ; In the formula: It is the mean vector output by the encoding module. It is the square root vector of the variance. It is a random noise vector sampled from the standard normal distribution. This indicates element-wise multiplication.

7. The data-driven method for predicting strain field distribution in composite materials according to claim 1, characterized in that, The header of the decoding module described in step S5 is set to include a latent variable. Mapped to The fully connected layer of the feature map; the number of input channels of the upsampled residual block of the first-stage decoder is The number of output channels is The main path consists of L 3×3 convolutions, with the shortcut connections using 3×3 kernels and a stride of 2, followed by one standard residual block; the second-stage decoder's upsampled residual block has the following input channels: The number of output channels is The main path consists of L 5×5 convolutions, with the shortcut connections using 5×5 kernels and a stride of 2, followed by two standard residual blocks; the third-stage decoder's upsampled residual blocks have the following input channels: The number of output channels is The main path consists of L 7×7 convolutions, the convolution kernel size of the shortcut connection is 7×7, the convolution stride is 2, followed by 1 standard residual block; the upsampling residual block uses transposed convolution for upsampling.

8. The data-driven method for predicting strain field distribution in composite materials according to claim 1, characterized in that, In step S6, the convolutional kernel size of the output layer of the network is 4×4, and the number of input channels is... The number of output channels is 1, and the convolution stride is 2.

9. The data-driven method for predicting strain field distribution in composite materials according to claim 1, characterized in that, In step S7, the network model is trained using the PyTorch deep learning framework based on the set loss function, optimizer, and training hyperparameters.

10. The data-driven method for predicting strain field distribution in composite materials according to claim 9, characterized in that, The optimizer used is the Adam optimizer; the initial learning rate is set to 0.00001~0.0001; a cosine annealing scheduling strategy is adopted; the training cycle is 500~800 training rounds; the batch size is selected as 8, 32 or 64 depending on the graphics processor; the validation set early stopping patience value is set to 10~20 training rounds.

Images