A C-Mn steel microstructure image generation method based on an improved autoencoder network

Abstract

The present invention discloses a C-Mn steel microstructure image generation method of an improved autoencoder network, which belongs to the technical field of steel microstructure image prediction, including obtaining a C-Mn data set; preprocessing and expanding the C-Mn data set; improving the autoencoder; constructing an encoder and decoder network model; applying the encoder model for feature extraction; constructing a PCA model; constructing an XGBoost model; inputting self-set component process data into the trained XGBoost model, PCA model and decoder model in sequence to obtain a predicted microstructure prediction image. The present invention introduces a dense convolutional block into a traditional autoencoder structure to form a DCB-AE network, solves the problem of traditional AE network fitting for large-scale images through a data dimensionality reduction and dimensionality increase method based on the DCB-AE network, and in addition, improves the generalization ability of the microstructure image generation model by combining a machine learning model.

Description

Technical Field

[0001] The present invention relates to the technical field of steel microstructure image prediction, and in particular to a C-Mn steel microstructure image generation method using an improved autoencoder network. Background Art

[0002] Microstructure image analysis plays a core role in the research of metal material composition, process, structure and performance, and has a profound impact on the development of materials science. The microstructure image of steel contains rich organizational morphology information, which plays a vital role in the mechanical properties of steel materials.

[0003] At present, in the design and optimization process of steel, microstructure is an important factor to be considered. Generally speaking, microstructural characteristics are mainly calculated through physical metallurgical models. For example, patent CN114896770A discloses a method for calculating the volume fraction of continuous cooling microstructure phase of low-alloy high-strength steel. This method calculates the phase volume fraction corresponding to the phase transformation type by establishing the two-phase process phase transformation kinetic equation and its differential equation. Patent CN102323981B discloses a method for predicting the static recrystallization microstructure evolution of austenite in hot-rolled steel. This method can accurately predict the evolution law of static recrystallization of austenite with fewer parameters, and optimize the hot rolling process parameters accordingly. However, the microstructural characteristics obtained by the current method are numerical information, which cannot provide image information of the specific morphological distribution of the microstructure, and cannot provide more accurate guidance for the design and optimization of steel. With the continuous development of AI technology, researchers have begun to use deep learning models to directly obtain microstructure images based on the chemical composition and production process of steel. For example, Lambard et al. used the StyleGAN2 model to generate various 512×512 pixel high-quality alloy microstructure images from limited data. Iyer et al. used the Wasserstein GAN network to synthesize microstructure images under given processing conditions. This method combines the cooling method with random latent space encoding as the input of the model to achieve the synthesis of high-quality multiphase structure under a given cooling method. However, the current microstructure image generation method is to conduct research on specific components and processes. The modeling process does not consider the influence of components and processes on the generated microstructure. At the same time, the random matrix is ​​used as the basic data for image generation, resulting in the generated images having no physical meaning and only similar appearance. Therefore, it is impossible to establish a relationship model between components and processes and microstructure images with physical metallurgical significance, and thus it is impossible to provide effective research guidance in the design and optimization process of steel. Summary of the invention

[0004] The purpose of the present invention is to provide a C-Mn steel microstructure image generation method with an improved autoencoder network. A dense convolution block DCB module is introduced into the traditional autoencoder AE structure to form a DCB-AE network. The problem of large-scale image fitting of the traditional AE network is solved by a data dimensionality reduction and dimensionality increase method based on the DCB-AE network. In addition, the generalization ability of the microstructure image generation model is improved by combining a machine learning model.

[0005] To achieve the above object, the present invention provides a C-Mn steel microstructure image generation method using an improved autoencoder network, the steps comprising:

[0006] S1. Collect samples from different positions of hot-rolled plates during steel production, and obtain a sample data set of C-Mn steel including composition, process and microstructure images through experiments. The microstructure images include ferrite and pearlite phase, etc.

[0007] S2. Preprocess and expand the microstructure image dataset obtained in S1 to obtain dataset D ae ;

[0008] S3. Improve the autoencoder AE network and introduce dense convolution block DCB into the AE network;

[0009] S4, based on S2 dataset D ae Train the DCB-AE network and build encoder and decoder network models;

[0010] S5: Use the encoder network model to extract features from the microstructure image preprocessed in S2 to obtain a new data set D encoder ;

[0011] S6. Use PCA algorithm to analyze the data set D in S5 encoder Perform dimensionality reduction and build a PCA model to obtain a dimensionality reduction feature encoding dataset D of the microstructure image after encoding by the encoder and PCA dimensionality reduction ld ;

[0012] S7, reduce the dimension of the dataset D in S6 ld Partition, divide into training set and test set, train and test the XGBoost algorithm, build the XGBoost model, the input data of the XGBoost model is the component process, and the output data is the corresponding dimension reduction feature encoding data;

[0013] S8. Input the self-set component process data into the trained XGBoost model, PCA model and decoder model in sequence, and finally obtain the predicted microstructure prediction image.

[0014] Preferably, step S2 includes:

[0015] S21, cropping the microstructure image data in S1;

[0016] S22, expanding the cropped microstructure image by flipping and rotating;

[0017] S23. Define the size of the microstructure image as a three-dimensional matrix, expressed as I(C',H,W).

[0018] Preferably, in step S3, the dense convolution block includes a batch normalization layer, an activation function and a convolution layer.

[0019] Preferably, in step S4, the encoder includes a convolution layer and a dense convolution block, the decoder includes an upsampling layer, a dense convolution block and a convolution layer, and the training process specifically includes:

[0020] S41, extract features from I(C',H,W), and change the number of channels of I through the convolution layer to obtain a feature map A1(C,H,W);

[0021] S42, apply dense convolution block to extract features of A1, change the number of feature maps through two batch normalization, activation function, convolution layer and splicing operation, and then change the number and size of A1 feature map by convolution layer to obtain feature map

[0022] S43, repeat the above steps twice to obtain the feature map

[0023] S44, use the upsampling layer to change the size of the feature map A4, and then change the number of feature maps through the dense convolution block to obtain the feature map

[0024] S45. Repeat the above steps twice to obtain the feature map A7(C,H,W);

[0025] S46, applying a convolutional layer to change the number of feature maps A7, and obtaining feature maps A8 (C', H, W);

[0026] S47. Calculate the square of the difference between input I and output A8 as the loss of the network model, optimize the parameters of the autoencoder model, and obtain the trained encoder and decoder network models.

[0027] Preferably, step S8 includes:

[0028] S81, input the self-set component process data into the XGBoost model obtained in step S7, and obtain the predicted value V pca_pre .

[0029] S82, the predicted value V obtained in S81 pca_preInput the PCA model in S6 to perform data dimension increase to obtain microstructure feature data V decoder_pre .

[0030] S83, V obtained in S82 decoder_pre The decoder network model in S4 is input to obtain the C-Mn steel microstructure prediction image corresponding to the self-set composition and process data.

[0031] Therefore, the present invention adopts a C-Mn steel microstructure image generation method of an improved autoencoder network of the above structure, and the beneficial effects achieved are:

[0032] (1) The DCB module is introduced into the structure of the AE network, and the front and back microstructure features are spliced ​​using a connection structure. This connection method allows the features of the previous network layer to be directly applied to the subsequent network layer, which strengthens the propagation of features in the network, effectively alleviates the problem of gradient disappearance, and enables image features to be reused;

[0033] (2) Through machine learning, PCA and decoder models, the conversion from preset one-dimensional steel material composition and process data to three-dimensional microstructure images is realized. Compared with directly establishing a correspondence model between composition process and microstructure images, the nonlinear fitting ability and generalization performance of the model are improved, which provides effective production guidance for researchers and is conducive to improving the production efficiency of steel enterprises.

[0034] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 It is a schematic diagram of the process of the present invention;

[0036] Figure 2 The schematic diagram of the improved autoencoder network of the present invention;

[0037] Figure 3 It is a schematic diagram of the data transmission process of the dense convolution block unit of the present invention;

[0038] Figure 4 is an example of test results in the test set of the present invention;

[0039] Figure 5 This is an example diagram of a microstructure image generated by the present invention based on self-set composition and process data. DETAILED DESCRIPTION

[0040] The technical solution of the present invention is further described below through the accompanying drawings and embodiments.

[0041] Unless otherwise defined, the technical terms or scientific terms used in the present invention shall have the common meanings understood by people with ordinary skills in the field to which the present invention belongs. The words "include" or "comprise" and the like used in the present invention mean that the elements or objects appearing before the word include the elements or objects listed after the word and their equivalents, but do not exclude other elements or objects.

[0042] Example

[0043] like Figure 1-5 As shown, the present invention provides a C-Mn steel microstructure image generation method using an improved autoencoder network, the steps comprising:

[0044] S1. Collect samples from different positions of hot-rolled plates during steel production, and obtain a sample data set of C-Mn steel containing m components, processes and microstructure images through experiments. The microstructure images include ferrite and pearlite, etc.

[0045] Specifically, m = 1000 composition processes and steel samples were collected from a hot rolling production line of a steel plant, where the composition process data included C, Si, Mn, P, S, Al, thickness (H), rough rolling outlet temperature (RDT), finishing rolling outlet temperature (FDT), and coiling temperature (CT). The microstructure images of the C-Mn steel samples were obtained by experimental methods, and finally an initial data set D containing the composition process and the corresponding microstructure images was obtained.

[0046] S2. Crop the microstructure image data in S1 to obtain a data set D containing the composition process and the corresponding microstructure image. crop ; The cropped microstructure images were expanded by using data enhancement methods such as upside down flipping, left-right flipping, and rotating 90°, 180°, and 270° to obtain a microstructure image dataset D containing 6000 ae ; The microstructure image size is defined as a three-dimensional matrix, represented as I(C',H,W).

[0047] Specifically, the image is cropped to obtain an image size I(3,1024,1024) that meets the input of the improved autoencoder model.

[0048] S3. Improve the autoencoder AE network by introducing a dense convolution block DCB into the AE network. The dense convolution block includes a batch normalization layer, an activation function, and a convolution layer. Figure 2 good Figure 3 , Figure 2 Among them, A1-A4 is the encoder part, and A5-A8 is the decoder part.

[0049] S4, based on S2 dataset Da eThe DCB-AE network is trained to construct encoder and decoder network models, where the encoder includes convolutional layers and dense convolutional blocks, and the decoder includes upsampling layers, dense convolutional blocks, and convolutional layers. The training steps include:

[0050] S41. Extract features from I(C',H,W), and obtain feature map A1(C,H,W) by changing the number of channels of I through the convolution layer.

[0051] Specifically, the number of convolution kernels of the convolution layer is 16, the size is 3*3, the step size is 1, and the padding is 1. The convolution layer is applied to extract the features of I to obtain the feature map A1 (16, 1024, 1024).

[0052] S42, apply dense convolutional blocks to extract features from A1, refer to Figure 3 After two batch normalizations, activation functions, convolutional layers, and concatenation operations, the number of feature maps is changed to achieve feature extraction and reuse. The number and size of the A1 feature map are then changed by the convolutional layer to obtain the feature map

[0053] The batch normalization function in the dense convolution block is the BatchNorm2d function, which is expressed as:

[0054]

[0055] Among them, E(x) is the mean of the input data, Var(x) is the variance of the input data, ∈ is the stability coefficient, and its default value is 1e-5. γ and β are affine operations on the input values, that is, linear transformations. The default values ​​of γ and β are 1 and 0, respectively.

[0056] The activation function is the Leaky ReLU function, and its expression is:

[0057] y(x)=max(0,r)+0.01*min(0,z).

[0058] Among them, max(0, x) represents the maximum value between the output 0 and x, and min(0, x) represents the minimum value between the output 0 and x.

[0059] Specifically, the size of the convolution kernel in the dense convolution block is 3*3, the step size is 1, the padding is 1, the number of convolution kernels is 28, 40, and 32 respectively, and the obtained feature map is A2(32,512,512).

[0060] S33, apply dense convolution block to extract features from A2, change the number of feature maps through two batch normalization, activation function, convolution layer and splicing operation, and then change the number and size of A2 feature maps by convolution layer, and finally get feature map

[0061] Specifically, the size of the convolution kernel in the dense convolution block is 3*3, the step size is 1, the padding is 1, the number of convolution kernels is 44, 56, and 64 respectively, and the obtained feature map is A3(64,256,256).

[0062] S34, apply dense convolution block to extract features of A3, change the number of feature maps through two batch normalization, activation function, convolution layer and splicing operation, and then change the number and size of A3 feature map by convolution layer, and finally get feature map

[0063] Specifically, the size of the convolution kernel in the dense convolution block is 3*3, the step size is 2, the padding is 1, and the number of convolution kernels is 76, 88, and 128. The resulting feature map is A4 (128, 128, 128).

[0064] S35, apply the upsampling layer to change the size of the feature map A4, and obtain the feature map Then the number of feature maps is changed through dense convolution blocks to obtain feature maps

[0065] Specifically, all upsampling operations have a kernel size of 2*2 and a stride of 2.

[0066] In this step, A4 is obtained by upsampling layer After passing through dense convolution blocks with 140, 152, and 64 convolution kernels respectively, feature map A5 (64, 256, 256) is obtained.

[0067] S36, apply the upsampling layer to change the size of the feature map A5, and obtain the feature map Then the number of feature maps is changed through dense convolution blocks to obtain feature maps

[0068] Specifically, A5 is obtained by upsampling layer Then, after passing through dense convolution blocks with convolution kernels of 76, 88, and 32 respectively, the feature map A6 (32, 512, 512) is obtained.

[0069] S37, apply the upsampling layer to change the size of the feature map A6, and obtain the feature map Then, the number of feature maps is changed through dense convolution blocks to obtain feature map A7 (C, H, W).

[0070] Specifically, A6 is obtained through the upsampling layer Then, after passing through dense convolution blocks with convolution kernels of 44, 56, and 16 respectively, the feature map A7 (16, 1024, 1024) is obtained.

[0071] S38: Apply the convolution layer to change the number of feature maps A7 to obtain feature maps A8 (C', H, W), then calculate the square of the difference between the input I and the output A8 as the loss of the network model, optimize the parameters of the autoencoder model, and obtain the trained encoder and decoder models.

[0072] Specifically, in this step, the number of convolutional layers is 3, the size is 1*1, the step length is 1, and the padding is 1. The convolutional layer is applied to change the number of channels of A7 to obtain the feature map A8 (3, 1024, 1024). Then the mean square loss function MSELoss is applied to calculate the loss value between the input I and the output A8, thereby optimizing the parameters of the autoencoder model. After 50,000 iterations, the trained encoder and decoder models are obtained.

[0073] The expression of the loss function is loss(x i ,y i )=(x i -y i ) 2 .

[0074] Among them, x i is the predicted value of the i-th sample data, y i is the true value of the i-th sample data.

[0075] S5. Use the encoder network model to extract features from the preprocessed microstructure image to obtain a new data set D encoder .

[0076] Specifically, the trained encoder model is applied to the dataset D in step S2 crop The microstructure image in is feature encoded to obtain the encoded data set D encoder , where each encoded matrix size is (128,128,128).

[0077] S6. Use the principal component analysis (PCA) algorithm to analyze the data set D in S5. encoder Perform dimensionality reduction and build a PCA model to obtain a dimensionality reduction feature encoding dataset D of the microstructure image after encoding by the encoder and PCA dimensionality reduction ld .

[0078] S7, reduce the dimension of the dataset D in S6 ld Partition, divide into training set and test set, train and test the XGBoost algorithm, the input data is the component process, the output data is the corresponding dimensionality reduction feature encoding data, and the XGBoost model is constructed.

[0079] Reference Figure 4 , a is the real microstructure image, and b is the microstructure image predicted by the XGBoost model.

[0080] Specifically, dataset D ld Divide in a ratio of 4:1.

[0081] S8, inputting the self-set component process data into the trained XGBoost model, PCA model and decoder model in sequence, and finally obtaining the predicted microstructure prediction image, specifically including:

[0082] S81, input the self-set component process data into the XGBoost model obtained in step S7, and obtain the predicted value V pca_pre .

[0083] S82, the predicted value V obtained in S81 pca_pre Input the PCA model in S6 to perform data dimension increase to obtain microstructure feature data V decoder_pre .

[0084] S83, V obtained in S82 decoder_pre The decoder network model in S4 is input to obtain the C-Mn steel microstructure prediction image corresponding to the self-set composition and process data.

[0085] Specifically, the set C-Mn steel composition process values ​​are shown in Table 1. The three important parameters in the microstructure image, ferrite and pearlite phase fractions and average grain size, and the statistical results are shown in Table 2. The corresponding C-Mn steel microstructure prediction image and organizational characteristics are shown in Table Figure 5 As shown in Table 1, Table 2 and Figure 5 In the above figure, a and b correspond to each other.

[0086] Table 1

[0087]

[0088] Table 2

[0089]

[0090] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A method for generating a C-Mn steel microstructure image by improving an autoencoder network, characterized in that the steps include: S1. Obtain sample data set of C-Mn steel composition, process and microstructure images; S2, preprocessing and expanding the microstructure image data set obtained in S1; S3. Improve the autoencoder AE network and introduce a dense convolution block DCB into the AE network. The dense convolution block includes a batch normalization layer, an activation function and a convolution layer. S4. Based on the S2 data set, the DCB-AE network is trained to construct an encoder and decoder network model. The encoder includes a convolution layer and a dense convolution block, and the decoder includes an upsampling layer, a dense convolution block and a convolution layer. The training process specifically includes: S41, extract features from I(C',H,W), and change the number of channels of I through the convolution layer to obtain a feature map A1(C,H,W); S42, apply dense convolution block to extract features of A1, change the number of feature maps through two batch normalization, activation function, convolution layer and splicing operation, and then change the number and size of A1 feature map by convolution layer to obtain feature map S43, repeat the above steps twice to obtain the feature map S44, use the upsampling layer to change the size of the feature map A4, and then change the number of feature maps through the dense convolution block to obtain the feature map S45. Repeat the above steps twice to obtain the feature map A7(C,H,W); S46, applying a convolutional layer to change the number of feature maps A7, and obtaining feature maps A8 (C', H, W); S47, calculating the square of the difference between the input I and the output A8 as the loss of the network model, optimizing the parameters of the autoencoder model, and obtaining the trained encoder and decoder network models; S5, using the encoder network model to extract features from the microstructure image preprocessed in S2 to obtain a new data set; S6, use the PCA algorithm to reduce the dimension of the data set in S5 and build a PCA model; S7, divide the dimension reduction data set in S6, train and test the XGBoost algorithm, and build an XGBoost model; S8, inputting the self-set component process data into the trained XGBoost model, PCA model and decoder model in sequence to obtain the predicted microstructure prediction image, specifically including: S81, input the self-set component process data into the XGBoost model obtained in step S7, and obtain the predicted value V pca_pre ; S82, the predicted value V obtained in S81 pca_pre Input the PCA model in S6 to perform data dimension increase to obtain microstructure feature data V decoder_pre ; S83, V obtained in S82 decoder_pre The decoder network model in S4 is input to obtain the predicted image of C-Mn steel microstructure corresponding to the self-set composition and process data.

2. The C-Mn steel microstructure image generation method of the improved autoencoder network according to claim 1 is characterized in that: Step S2 includes: S21, cropping the microstructure image data in S1; S22, expanding the cropped microstructure image by flipping and rotating; S23. Define the size of the microstructure image as a three-dimensional matrix, expressed as I(C',H,W).

Images