Method and device for predicting stretching curve of metal plate strip

Abstract

The invention discloses a method and a device for predicting a tensile curve of a metal plate strip, relates to the technical field of metal mechanical property research, and aims to solve the problem of poor prediction precision of the tensile curve of the metal plate strip in the prior art. The method comprises the following steps: acquiring target process data, microscopic structure images and real tensile curves of a plurality of metal plate strips; performing data interception processing on the real stretching curve to obtain target stretching data; performing n-order polynomial fitting on the target stretching data to obtain a target polynomial coefficient; respectively constructing an association relationship between the target process data and the target polynomial coefficient as well as between the microscopic structure image and the target polynomial coefficient to obtain a target mixed data set; based on the target mixed data set, performing model training on the initial mixed prediction model until the prediction precision of the initial mixed prediction model meets a preset convergence condition, and obtaining a target mixed prediction model; and the stretching curve of the target metal plate strip is predicted by using the target mixed prediction model, so that the prediction precision is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of metal mechanical property research, and in particular to a metal plate strip tensile curve prediction method and device. BACKGROUND

[0002] With the transformation and upgrading of the national manufacturing industry and the high attention to the planning of the precision manufacturing industry field, metal plate strip materials with excellent performance have gradually become a focus. Metal plate strip materials have excellent surface precision, corrosion resistance, flexibility and high ductility, and have become important raw materials for micro-forming and precision manufacturing, and are widely used in the fields of aerospace, atomic energy industry, automobile manufacturing and electrical and electronic industries, and have become indispensable structural materials and functional materials in manufacturing industry. With the rapid development of related fields, the demand for precision forming raw material metal plate strip materials has rapidly increased, and the ultra-high precision of precision manufacturing has also increased the requirements for the performance, thickness precision and surface quality of metal plate strip materials. In recent years, artificial intelligence technology and application have become mature, and the self-learning model driven by data is introduced into the rolling process optimization and mechanical property prediction, which expands the research methods in the field of metal rolling.

[0003] At present, the intelligent prediction of metal plate strip tensile curve using artificial intelligence algorithm and data-driven modeling method can not only greatly improve the production efficiency, but also be beneficial to the optimization of rolling process and rapid prediction of performance. However, the data-driven algorithm used by artificial intelligence is still the traditional data-driven algorithm, which still only takes production data as input and ignores the fact that the mechanical property prediction of metal plate strip is directly related to the microstructure. The traditional data-driven algorithm method cannot meet the requirements of modern manufacturing technology for the prediction accuracy of the mechanical properties of metal plate strip, and the task of improving the model or controlling the accuracy becomes urgent.

[0004] Therefore, it is urgent to design a more advanced metal plate strip mechanical property prediction method to solve the problem of poor prediction accuracy of metal plate strip tensile curve in the prior art. SUMMARY

[0005] The present application aims to provide a metal plate strip tensile curve prediction method and device, which combines the influence of the production process parameters of the metal plate strip and the microstructure image on the mechanical properties of the plate strip to construct a metal plate strip tensile curve prediction hybrid model, and uses the model to obtain a metal plate strip tensile curve with higher prediction accuracy, thereby solving the problem of poor prediction accuracy of metal plate strip tensile curve in the prior art.

[0006] In order to achieve the above-mentioned purpose, the present application provides the following technical solutions: In a first aspect, the present application provides a metal plate strip tensile curve prediction method, which can include: obtaining target process data, microstructure images and real tensile curves of a plurality of metal strip materials; the target process data is process data obtained by performing multi-dimensional data fusion on an M-dimensional process data vector in initial process data and an N-dimensional one-hot encoding vector; performing data interception processing on the real tensile curves to obtain target tensile data; the target tensile data is tensile data points in the real tensile curves from a yield strength point to a tensile strength point; performing n-order polynomial fitting on the target tensile data to obtain target polynomial coefficients; respectively constructing a correlation between the target process data and the microstructure images and the target polynomial coefficients to obtain a target mixed data set; based on the target mixed data set, performing model training on an initial mixed prediction model until the prediction accuracy of the initial mixed prediction model meets a preset convergence condition to obtain a target mixed prediction model; the initial mixed prediction model is a network model for outputting mechanical properties by fusing a multilayer perceptron prediction model for process data training and a convolutional neural network prediction model for image data training; using the target mixed prediction model to predict a tensile curve of a target metal strip material.

[0007] Preferably, the using the target mixed prediction model to predict a tensile curve of a target metal strip material can include: obtaining target process data and microstructure images of a target metal strip material; based on the target process data and microstructure images of the target metal strip material, using the target mixed prediction model to predict a tensile curve to obtain polynomial coefficients for the target metal strip material; based on the polynomial coefficients, constructing a tensile curve for the target metal strip material.

[0008] Preferably, before the based on the target mixed data set, performing model training on an initial mixed prediction model, can include: constructing the initial mixed prediction model; the constructing the initial mixed prediction model includes: constructing a multi-input network model containing a ResNet network and an MLP network branch; a combination layer of the multi-input network model is a network layer combining output neurons of the ResNet network and the MLP network, and an Adam optimization algorithm of stochastic gradient descent is used to optimize network model parameters of the multi-input network model to obtain the initial mixed prediction model; the combination layer is used to combine the last layer of the regression output layer of the ResNet network and the MLP network into one layer.

[0009] Preferably, the ResNet network model can comprise a convolution block attention module, which is used to increase the weight of the combined feature map related to the prediction task.

[0010] Preferably, the ResNet network is used for data prediction of microscopic image data of metal strip; The MLP network is used to process numerical target process data of metal strip.

[0011] Preferably, the target process data of the plurality of metal strips can be obtained by: Obtaining initial process data of a plurality of metal strips; the initial process data at least includes rolling process data and heat treatment process data; One-hot encoding processing is performed on the cooling mode of the plurality of metal strips to obtain an N-dimensional one-hot encoding vector; the N-dimensional one-hot encoding vector at least includes a water cooling one-hot encoding vector and a furnace cooling one-hot encoding vector; Multi-dimensional data fusion is performed on the M-dimensional process data vector in the initial process data and the N-dimensional one-hot encoding vector to obtain the target process data; the M-dimensional process data vector at least includes rolling thickness process data vector, reduction ratio process data vector, annealing temperature process data vector and holding time process data vector; the target process data is M plus N-dimensional process data.

[0012] Preferably, the target process data of the plurality of metal strips, the microscopic structure image and the real tensile curve can be obtained by: Taking pictures of the microscopic structure images of the plurality of metal strips respectively to obtain a plurality of microstructure photos; Splicing a plurality of microstructure photos belonging to the same metal strip to obtain the microscopic structure image of the plurality of metal strips.

[0013] Preferably, the plurality of microstructure photos at least includes a pole figure, an Euler angle diagram, a diffraction quality diagram and a recrystallization fraction diagram; Splicing a plurality of microstructure photos belonging to the same metal strip to obtain the microscopic structure image of the plurality of metal strips can include: Converting the pole figure, the Euler angle diagram, the diffraction quality diagram and the recrystallization fraction diagram of the same metal strip to obtain four pictures with a pixel size of ; wherein, represents the length of the picture ; and Converting the pole figure, the Euler angle diagram, the diffraction quality diagram and the recrystallization fraction diagram of the same metal strip to obtain four pictures with a pixel size of Images of different sizes are combined and stitched together into a single image with a resolution of [number] pixels. The image was taken at a certain size to obtain a microstructure image of the metal sheet / strip.

[0014] Preferably, the step of constructing the correlation between the target process data and the microstructure image and the target polynomial coefficients to obtain the target hybrid dataset may include: The microscopic tissue images are preprocessed, and the image data after each preprocessing is retained to obtain a target microscopic tissue image with a greater number of images than the number of images corresponding to the microscopic tissue image; the image preprocessing includes at least one of translation, rotation, and cropping; The correlation between the target process data, the target microstructure image, and the target polynomial coefficients is constructed to obtain the target hybrid dataset.

[0015] In a second aspect, the present invention provides a device for predicting the tensile curve of metal sheet and strip, which may include: The acquisition module is used to acquire target process data, microstructure images and actual tensile curves of multiple metal strips; the target process data is the process data obtained by multidimensional data fusion of the M-dimensional process data vector and the N-dimensional unique thermal encoding vector in the initial process data. The data extraction module is used to extract data from the actual tensile curve to obtain target tensile data; the target tensile data is the tensile data points from the yield strength point to the tensile strength point in the actual tensile curve. The fitting module is used to perform nth-order polynomial fitting on the target stretching data to obtain the target polynomial coefficients. A correlation construction module is used to construct the correlation between the target process data and the microstructure image and the target polynomial coefficients, respectively, to obtain a target hybrid dataset; The model training module is used to train the initial hybrid prediction model based on the target hybrid dataset until the prediction accuracy of the initial hybrid prediction model meets the preset convergence condition, thereby obtaining the target hybrid prediction model. The initial hybrid prediction model is a network model that fuses a multilayer perceptron prediction model used for process data training and a convolutional neural network prediction model used for image data training, and outputs a mechanical performance network model. The prediction module is used to predict the tensile curve of the target metal sheet / strip using the target hybrid prediction model.

[0016] Compared with existing technologies, this invention provides a method for predicting the tensile curve of metal sheet and strip. This method acquires target process data, microstructure images, and actual tensile curves of multiple metal sheet and strip materials. The target process data is obtained by fusing M-dimensional process data vectors and N-dimensional uniquely encoded vectors from the initial process data. The actual tensile curves are truncated to obtain target tensile data, which consists of tensile data points from the yield strength point to the tensile strength point in the actual tensile curves. An nth-order polynomial is fitted to the target tensile data to obtain the target polynomial coefficients. Correlation relationships are constructed between the target process data, the microstructure images, and the target polynomial coefficients to obtain a target mixed dataset. Based on the target mixed dataset, an initial mixed prediction model is trained until the initial mixed prediction model is obtained. The prediction accuracy meets the preset convergence condition, resulting in a target hybrid prediction model. The initial hybrid prediction model is a network model that integrates a multilayer perceptron prediction model used for process data training and a convolutional neural network prediction model used for image data training, and outputs the mechanical properties. The target hybrid prediction model is used to predict the tensile curve of the target metal sheet / strip. Based on this, the present invention uses a target hybrid prediction model with multilayer perceptron and convolutional neural network functions, constructed based on process data, microstructure images, and tensile curve data of metal sheet / strip, to predict the tensile curve of the target metal sheet / strip. Compared with the prediction models in the prior art that are only based on process data or image-driven, this greatly improves the prediction accuracy of the tensile curve of metal sheet / strip and solves the problem of poor prediction accuracy of the tensile curve of metal sheet / strip in the prior art. Attached Figure Description

[0017] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings: Figure 1 A schematic diagram of the main process of a method for predicting the tensile curve of metal sheet and strip provided by the present invention; Figure 2a The inverse pole figure of the rolled surface grains of a certain Invar alloy thin strip is provided in the prediction method of the tensile curve of metal sheet and strip provided by the present invention. Figure 2b Euler angle diagram of the grain size of a certain Invar alloy thin strip in the rolling surface of a method for predicting the tensile curve of metal sheet and strip provided by the present invention; Figure 2c The grain diffraction quality diagram of the rolled surface of a certain Invar alloy thin strip in the method for predicting the tensile curve of metal sheet and strip provided by the present invention; Figure 2dA grain recrystallization microstructure of a certain Invar alloy thin strip on the rolling surface, which is used in the prediction method of the tensile curve of metal sheet and strip provided by the present invention. Figure 3 The image provided by this invention is a composite image of four microscopic images of Invar alloy thin strip after processing and reassembly. Figure 4 A schematic diagram of the target hybrid prediction model in the method for predicting the tensile curve of metal sheet and strip provided by the present invention; Figure 5 A schematic diagram illustrating the change in loss with iteration number after using the Adam optimization algorithm in the prediction method for the tensile curve of metal sheet and strip provided by the present invention. Figure 6 A schematic diagram showing the fitting effect of the average stress of the true value and the predicted value of the target hybrid prediction model in the prediction method of the tensile curve of metal sheet and strip provided by the present invention. Figure 7 A schematic diagram showing the fitting effect of the experimental curve and the predicted curve of the target hybrid prediction model in the prediction method for the tensile curve of metal sheet and strip provided by the present invention. Figure 8 This is a schematic diagram of the main structure of a device for predicting the tensile curve of metal sheet and strip provided by the present invention. Detailed Implementation

[0018] To facilitate a clear description of the technical solutions in the embodiments of the present invention, the terms "first" and "second" are used to distinguish identical or similar items with essentially the same function and effect. For example, the first threshold and the second threshold are merely used to distinguish different thresholds and do not limit their order. Those skilled in the art will understand that the terms "first" and "second" do not limit the quantity or execution order, and the terms "first" and "second" are not necessarily different.

[0019] It should be noted that in this invention, the terms "exemplary" or "for example" are used to indicate examples, illustrations, or descriptions. Any embodiment or design described as "exemplary" or "for example" in this invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0020] In this invention, "at least one" refers to one or more, and "more than one" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone, where A and B can be singular or plural. The character " / " generally indicates that the preceding related objects have an "or" relationship. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one of a, b, or c can represent: a, b, c, a combination of a and b, a combination of a and c, a combination of b and c, or a, b, and c, where a, b, and c can be single or multiple.

[0021] As metal sheets and strips become thinner, their internal microstructure and structure undergo significant changes, resulting in macroscopic alterations to their mechanical properties. While numerous scholars both domestically and internationally have developed comprehensive theoretical systems regarding metal deformation mechanisms, mechanical properties, and microstructural evolution in traditional metallurgical thick plate rolling processes, traditional research methods still require substantial investment of manpower and resources, involving numerous repeated experiments. This not only involves a large workload but also results in significant errors in experimental results due to the thinness of the samples, making it difficult to achieve the high-quality, low-cost production requirements.

[0022] While using artificial intelligence algorithms and data-driven modeling methods to intelligently predict the stretching curves of metal sheets and strips can greatly improve production efficiency and facilitate the optimization of rolling processes and rapid performance prediction, it is still impossible to meet the accuracy requirements of modern manufacturing technology for predicting the mechanical properties of metal sheets and strips without changing the algorithm-driven data.

[0023] In view of this, the present invention provides a method and apparatus for predicting the tensile curve of metal sheet and strip. By combining the influence of the production process parameters and microstructure images of metal sheet and strip on the mechanical properties of the sheet and strip, and fusing the multilayer perceptron prediction model used for process data training and the convolutional neural network prediction model used for image data training, a hybrid model for predicting the tensile curve of metal sheet and strip is constructed. Using this model, a metal sheet and strip tensile curve with higher prediction accuracy can be obtained, thereby solving the problem of poor prediction accuracy of the tensile curve of metal sheet and strip in the prior art.

[0024] The technical solution of the present invention will now be described in detail with reference to the accompanying drawings: In a first aspect, the present invention provides a method for predicting the tensile curve of metal sheet and strip. Please refer to [link to relevant documentation]. Figure 1 , Figure 1The main flowchart of the method for predicting the tensile curve of metal sheet and strip provided by the present invention is shown in the figure. The main body executing the method is a server or service terminal equipped with the method for predicting the tensile curve of metal sheet and strip provided by the present invention, such as a test platform or a handheld portable prediction device.

[0025] exist Figure 1 In this context, the method may include: Step 110: Obtain target process data, microstructure images, and actual tensile curves for multiple metal strips; the target process data is the process data obtained by multidimensional data fusion of the M-dimensional process data vector and the N-dimensional unique thermal encoding vector in the initial process data.

[0026] In step 110, the basic data acquired for model training in this invention is no longer limited to production process data as in the prior art. Instead, it acquires target process data, microstructure images, and actual tensile curves for multiple metal strips. The basic data source is no longer singular, and the amount of data is relatively richer. Specifically, the target process data is multidimensional process data obtained by fusing M-dimensional process data vectors with N-dimensional one-hot encoded vectors. Compared to the single-source process data used in the prior art, the process data of this invention covers a wider range and is more detailed. Furthermore, microstructure images are also acquired. This data, combined with the target process data, is more conducive to improving the reliability of the basic data and the prediction accuracy of the prediction model.

[0027] Step 120: Perform data extraction processing on the actual tensile curve to obtain target tensile data; the target tensile data is the tensile data points from the yield strength point to the tensile strength point in the actual tensile curve.

[0028] Step 130: Perform nth-order polynomial fitting on the target stretching data to obtain the target polynomial coefficients.

[0029] Step 140: Construct the correlation between the target process data and the microstructure image and the target polynomial coefficients respectively to obtain the target hybrid dataset.

[0030] In steps 120 to 140, data points from the yield strength point to the tensile strength point in the actual tensile curve are first extracted to obtain target tensile data. Then, an nth-order polynomial is fitted to the target tensile data to obtain target polynomial coefficients; for example, if a 3rd-order polynomial is fitted, four polynomial coefficients can be obtained. Finally, the correlation between the target process data and the target polynomial coefficients, as well as the correlation between the microstructure image and the target polynomial coefficients, are established to obtain a target mixed dataset. This target mixed dataset can then be used to train the tensile curve prediction model.

[0031] Step 150: Based on the target mixed dataset, train the initial mixed prediction model until the prediction accuracy of the initial mixed prediction model meets the preset convergence condition to obtain the target mixed prediction model; the initial mixed prediction model is a network model that fuses the multilayer perceptron prediction model used for process data training and the convolutional neural network prediction model used for image data training and outputs the mechanical performance.

[0032] Step 160: Predict the tensile curve of the target metal sheet / strip using the target hybrid prediction model.

[0033] In steps 150 to 160, based on the target hybrid dataset, using target process data and microscopic images as model inputs and target polynomial coefficients as model outputs, an initial hybrid model is obtained by fusing the multilayer perceptron prediction model used for process data training and the convolutional neural network prediction model used for image data training. This process continues until the prediction accuracy of the initial hybrid prediction model meets preset convergence conditions, such as when the absolute error (MAE) and root mean square error (RMSE) of the model both meet preset convergence conditions. The target hybrid prediction model is then obtained. This model can be used to predict the tensile curve of the target metal sheet / strip, resulting in a more accurate tensile curve and thus more precisely controlling the mechanical properties of the metal sheet / strip, meeting the mechanical property requirements of precision manufacturing for metal sheet / strip.

[0034] Based on this, the present invention provides a method for predicting the stretching curve of metal sheet and strip. By using process data, microstructure images, and stretching curve data of metal sheet and strip, a target hybrid prediction model with multilayer perceptron and convolutional neural network functions is constructed to predict the stretching curve of the target metal sheet and strip. This method can obtain a more accurate stretching curve of metal sheet and strip, solving the problem of poor prediction accuracy of stretching curve of metal sheet and strip in the prior art, and promoting the rapid development of metal sheet and strip in precision manufacturing.

[0035] As an optional embodiment, before step 110, i.e., before acquiring the target process data for multiple metal strips, the following steps may be included: acquiring initial process data for multiple metal strips; the initial process data includes at least rolling process data and heat treatment process data; performing unique thermal encoding processing on the cooling methods of the multiple metal strips to obtain an N-dimensional unique thermal encoding vector; the N-dimensional unique thermal encoding vector includes at least a water-cooled unique thermal encoding vector and a furnace-cooled unique thermal encoding vector; fusing the M-dimensional process data vector in the initial process data with the N-dimensional unique thermal encoding vector to obtain the target process data; the M-dimensional process data vector includes at least a rolling thickness process data vector, a reduction rate process data vector, an annealing temperature process data vector, and a holding time process data vector; the target process data is M plus N-dimensional process data.

[0036] As a specific embodiment, this invention uses 170 metal strips, 4 types of process data vectors, 2 types of cooling methods, and 3rd-order polynomial fitting as examples. In practical applications, the number of metal strips, more types of process data and cooling methods, and higher-order polynomial fitting techniques can be added or reduced according to requirements. For example, rolling process data includes all or some of the following: initial thickness, reduction rate, number of rolling passes, strip width, final rolling thickness, front and rear tension, rolling speed, etc.; heat treatment process data includes: heat treatment temperature, heat treatment holding time, and cooling method, etc., where the cooling method can also include air cooling, furnace cooling, and water cooling, etc.

[0037] Specifically, assuming there are 170 metal strips (a certain Invar alloy thin strip), obtaining the initial process data for these 170 metal strips involves acquiring initial process data for each strip. This process data can include at least two categories: rolling process data vectors and heat treatment process data vectors. Rolling process data can include rolling thickness and reduction rate, where the unit for rolling thickness is h / mm; the unit for reduction rate is %; and the heat treatment process data can include at least annealing temperature process data vectors, holding time process data vectors, and cooling methods, where the unit for annealing temperature is T / ℃, the unit for holding time is t / min, and the cooling methods include furnace cooling and water cooling.

[0038] Furthermore, the furnace cooling and water cooling methods can be processed by unique thermal encoding to obtain a 2D unique thermal encoding vector; the unique thermal encoding results of the cooling methods are detailed in Table 1.

[0039] Table 1 Results of Independent Heat Coding for Cooling Methods Furthermore, the process data vector can be encoded to obtain a 4-dimensional process data vector, namely the rolling thickness process data vector, the pressure reduction process data vector, the annealing temperature process data vector, and the holding time process data vector; then the process data vector is fused with the unique heat encoding vector to obtain 6-dimensional process data; here M is 4 and N is 2.

[0040] Furthermore, statistical methods can be employed. In principle, noise reduction processing was performed on the production process data of 170 ultra-thin metal strips. In this embodiment, there is no noise data, so the number of samples after noise reduction processing is still 170 rolling process data of Invar alloy strips. The partial data after their respective mechanical property tests are shown in Table 2.

[0041] Table 2. Process data and mechanical property data of some Invar alloy strips Furthermore, deviation standardization is performed on the data in Table 2: this involves a linear transformation of the original rolling process data for ultra-thin metal strips, mapping the transformed numerical results of each process data point to the range [0-1]. This can be achieved using the following formula: (1) Determine the transformation function; where, Here, is the transformation function, max is the maximum value of the sample data, and min is the minimum value of the sample data. The production parameters after deviation standardization are shown in Table 3.

[0042] Table 3. Results of deviation standardization for some Invar alloy rolling process data and mechanical property test data. Table 3 (continued) In this embodiment, the non-numerical variable of cooling method is processed by one-hot encoding, which converts it into a computable vector space. The converted one-hot vectors are shown in Table 1.

[0043] Furthermore, prior to step 110, i.e., before acquiring the target process data, microstructure images, and actual tensile curves of multiple metal strips, the process may further include: taking photos of the microstructure images of multiple metal strips layer by layer to obtain multiple microstructure photographs; stitching together the multiple microstructure photographs belonging to the same metal strip to obtain microstructure images for multiple metal strips. The layers are divided according to the final rolled thickness of the metal strip; the multiple microstructure photographs include at least an inverse pole figure, Euler angle diagram, diffraction quality diagram, and recrystallization fraction diagram. Finally, the multiple microstructure photographs belonging to the same metal strip are stitched together to obtain microstructure images for multiple metal strips; that is, the inverse pole figure, Euler angle diagram, diffraction quality diagram, and recrystallization fraction diagram of the same metal strip can be image converted to obtain four images with a pixel count of [missing information]. Images of various sizes; among them, Indicates the length of the image Width; four pixels as Images of different sizes are combined and stitched together into a single image with a resolution of [number] pixels. Images of various sizes are obtained to capture the microstructure of multiple metal strips.

[0044] Continuing with the example of 170 metal strips (a certain Invar alloy thin strip), the microstructure images of the 170 metal strips were tested and photographed layer by layer. The resulting microstructure images of the extremely thin metal strip included four types of images. Please refer to the relevant documentation. Figures 2a to 2d , Figure 2a The inverse pole figure of the rolled surface grains of a certain Invar alloy thin strip is provided in the prediction method of the tensile curve of metal sheet and strip provided by the present invention. Figure 2b Euler angle diagram of the grain size of a certain Invar alloy thin strip in the rolling surface of a method for predicting the tensile curve of metal sheet and strip provided by the present invention; Figure 2c The grain diffraction quality diagram of the rolled surface of a certain Invar alloy thin strip in the method for predicting the tensile curve of metal sheet and strip provided by the present invention; Figure 2d This invention provides a method for predicting the tensile curve of metal sheet and strip, which includes a grain recrystallization microstructure diagram of the rolled surface of an Invar alloy thin strip. Based on this, inverse pole figure (IPF), Euler angle figure (EA), diffraction quality figure (DQ), and recrystallization fraction figure (RF) corresponding to the microstructure image of the 170 metal sheet and strip can be obtained respectively; among which, the length of the four types of microstructure images is... Width is Image of a certain size.

[0045] Furthermore, four types of images of the microstructure of extremely thin metal strips can be stitched together to obtain, as shown below. Figure 3 The images shown are microstructure images of multiple metal strips. Figure 3This invention provides a method for predicting the tensile curve of metal sheet and strip, which uses four microscopic images of Invar alloy thin strip that have been processed and re-stitched together to form a complete image. Specifically, it uses four images with a pixel count of... Images of different sizes are combined and stitched together into a single image with a resolution of [number] pixels. Images of various sizes are obtained to capture the microstructure of multiple metal strips.

[0046] Furthermore, the mechanical properties of 170 metal sheet and strip samples were tested, resulting in 170 tensile curves. Specific mechanical property testing methods can be employed using existing technologies. For example, an axial tensile force can be applied using a universal testing machine, the load-displacement curve can be recorded, and the curve characteristics can be analyzed to obtain mechanical property parameters, such as the yield load corresponding to the yield strength and the maximum load corresponding to the tensile strength, thus obtaining the true tensile curves of the 170 metal sheet and strip samples.

[0047] As an optional embodiment, in steps 120 to 130, the 170 tensile curves obtained can be processed, and the tensile data points from the yield point to the tensile point of the tensile curves can be extracted as target tensile data. Then, a third-order polynomial fitting is performed on the target tensile data to obtain the four polynomial coefficients after fitting.

[0048] For example: the fitting polynomial form using a third-order polynomial is as follows: To illustrate, we will take one of the stretching sample data as an example; the stretching sample data is shown in Table 4.

[0049] Table 4 Data for a certain stretching sample After fitting the above stretching curve, the following formula can be used: (2) This represents the explicit expression for the stretch curve; where, , , , The values ​​are 173.59, 8.23, -0.18, and 0.00092, respectively.

[0050] Furthermore, in step 140, the polynomial coefficients corresponding to the 170 process data, 170 combined microstructure images, and 170 stretching curves can be matched one-to-one according to the correspondence relationship; the matched mixed dataset is divided into two sets, a training set and a test set, according to a certain ratio, and the set division should maintain the consistency of data distribution.

[0051] Preferably, to expand the dataset, the correlation between the target process data and the microstructure images and the target polynomial coefficients is constructed respectively to obtain a target hybrid dataset. This may further include: performing image preprocessing on the microstructure images and retaining the image data after each preprocessing to obtain a target microstructure image with a greater number of images than the corresponding number of microstructure images; the image preprocessing includes at least one of translation, rotation, and cropping; and constructing the correlation between the target process data and the target microstructure images and the target polynomial coefficients respectively to obtain a target hybrid dataset.

[0052] Specifically, the 170 denoised process data points, the redefined and stitched microstructure images, and the mechanical properties of each Invar alloy strip can be mapped one-to-one. To expand the dataset, the images are translated, rotated, and cropped, resulting in a complete dataset of 1020 data points, which is the target hybrid dataset. The mapped hybrid dataset is then divided into training and testing sets according to a certain ratio, maintaining consistency in data distribution. Specifically, 80% of the dataset can be selected as the training set, and 20% as the testing set. To further verify model performance, 20% of the training set is selected as the validation set. Thus, the training set data = Test set data = This allows for model training of hybrid prediction models based on the target hybrid dataset.

[0053] As an optional embodiment, before step 150, i.e., before training the initial mixture prediction model based on the target mixture dataset, the process may include: constructing the initial mixture prediction model. Please refer to [reference needed]. Figure 4 and Figure 5 , Figure 4 A schematic diagram of the target hybrid prediction model in the method for predicting the tensile curve of metal sheet and strip provided by the present invention; Figure 5 This diagram illustrates the change in loss with the number of iterations after using the Adam optimization algorithm in the prediction method for the tensile curve of metal sheet and strip provided by the present invention.

[0054] exist Figure 4The initial hybrid prediction model can be constructed by: building a multi-input network model containing branches of ResNet and MLP networks; the merging layer of the multi-input network model is a network layer that combines the output neurons of ResNet and MLP networks; and optimizing the network model parameters of the multi-input network model using the Adam optimization algorithm with stochastic gradient descent to obtain the initial hybrid prediction model; the merging layer is used to merge the layer above the regression output layer of ResNet and MLP networks into one layer. The ResNet network model includes a convolutional block attention module, which increases the weights of the combined feature maps relevant to the prediction task, allowing the model to focus more on regions relevant to the training process rather than other noise. The ResNet network is used for data prediction based on microscopic image data of metal sheets and strips; the MLP network is used to process numerical data for data prediction based on target process data of metal sheets and strips.

[0055] It should be noted that the Convolutional Block Attention Module (CBAM) is a lightweight attention mechanism. Through two independent sub-modules, spatial attention and channel attention, it enhances the extraction of key information from feature maps. The model can learn more about which feature maps are more important and which regions on the feature maps are more important, thereby improving the quality of the extracted microstructure features and increasing the accuracy of the model in the task of predicting the tensile curves of metal sheets and strips.

[0056] Specifically, a multilayer perceptron prediction model trained on process data and a convolutional neural network prediction model trained on image data can be constructed, and the two models can be combined to determine a common mechanical performance output. In this invention, polynomial coefficients a0, a1, a2, and a3 are selected as the common output of the model.

[0057] For the model combination part, a combination layer is defined to combine the output neurons of MLP and CNN models. This layer merges the previous layer of the optional regression output layer of the two models into one layer, a fully connected hidden layer with ReLU activation function, and a fully connected regression output with Linear activation function. Figure 4 The hidden layer of the model shown can have 16 neurons.

[0058] Furthermore, the denoised process data (target process data) and image data (microscopic tissue images) are jointly combined to form a hybrid modeling dataset, which serves as the joint input to a multilayer perceptron (MLP) and a convolutional neural network (CNN). The Adam optimization algorithm based on stochastic gradient descent is used to optimize the parameters of this model. Figure 5 shows the change in loss with the number of iterations after applying the Adam optimization algorithm. Based on this, this invention constructs a mechanical property prediction model for ultra-thin metal strips by combining a multilayer perceptron and a convolutional neural network using the optimal parameter combination obtained through optimization, thus obtaining an initial hybrid prediction model.

[0059] In step 150, a mechanical property prediction model for ultrathin metal strips combining multilayer perceptron and convolutional neural network (initial hybrid prediction model) can be trained using the training set of the target hybrid dataset. The generalization performance of the mechanical property prediction model is tested using the test set of the target hybrid dataset. The target hybrid prediction model is obtained when the prediction accuracy of the initial hybrid prediction model meets the preset convergence condition.

[0060] Specifically, after mixed data modeling, the test set The prediction results are shown in Table 5.

[0061] Table 5 Model Prediction Results Furthermore, the coefficient of determination R is used. 2 The overall predictive performance of the metal ultrathin strip mechanical property prediction model combining multilayer perceptron and convolutional neural network was evaluated using mean absolute error (MAE) and root mean square error (RMSE). The prediction results of the model are shown in Table 6.

[0062] Table 6. Calculation results of model error Substitute the predicted values ​​a1, a2, a3, and a4 into... The predicted stretching curve can then be obtained.

[0063] The fitting effect of the prediction model on the mean stress of the true and predicted values ​​on the test set is as follows: Figure 6 As shown, Figure 6 This invention provides a method for predicting the tensile curve of metal sheet and strip. The schematic diagram illustrates the fitting effect of the target hybrid prediction model on the average stress of the true and predicted values ​​on the test set; the coefficient of determination R... 2 =0.96, the fitting effect of the actual curve and the predicted curve on the test set is as follows: Figure 7 As shown, Figure 7 The fitting effect diagram of the target hybrid prediction model on the test set of the actual curve and the predicted curve in the prediction method of metal sheet and strip provided by the present invention; from Figure 6 andFigure 7 The curves shown clearly demonstrate that the target hybrid prediction model of this invention can achieve good prediction results, illustrating the feasibility of stitching together process data extracted by linear layers and image data extracted by convolutional layers during machine learning. This provides a new basis for extending to other materials and extracting data from other modalities.

[0064] Furthermore, in step 160, the stretching curve of the target metal sheet / strip is predicted using a target hybrid prediction model, including: acquiring target process data and microstructure images of the target metal sheet / strip; predicting the stretching curve using the target hybrid prediction model based on the target process data and microstructure images of the target metal sheet / strip to obtain polynomial coefficients for the target metal sheet / strip; and constructing the stretching curve for the target metal sheet / strip based on the polynomial coefficients. This results in a more accurate stretching curve, further enabling optimization of the rolling process and rapid performance prediction.

[0065] It should be noted that the target process data and microstructure images of the target metal sheet and strip are obtained using the same methods described above, and will not be repeated here.

[0066] Secondly, the present invention provides a device for predicting the tensile curve of metal sheet and strip, please refer to [reference needed]. Figure 8 , Figure 8 This is a schematic diagram of the main structure of a device for predicting the tensile curve of metal sheet and strip provided by the present invention.

[0067] exist Figure 8 In this context, the device may include: The acquisition module 810 is used to acquire target process data, microstructure images and actual tensile curves of multiple metal strips; the target process data is the process data obtained by multidimensional data fusion of the M-dimensional process data vector and the N-dimensional unique thermal encoding vector in the initial process data.

[0068] The data extraction module 820 is used to perform data extraction processing on the real tensile curve to obtain target tensile data; the target tensile data is the tensile data points from the yield strength point to the tensile strength point in the real tensile curve.

[0069] The fitting module 830 is used to perform nth-order polynomial fitting on the target stretching data to obtain the target polynomial coefficients.

[0070] The association construction module 840 is used to construct the association between the target process data and the microstructure image and the target polynomial coefficients, respectively, to obtain the target hybrid dataset.

[0071] The model training module 850 is used to train the initial hybrid prediction model based on the target hybrid dataset until the prediction accuracy of the initial hybrid prediction model meets the preset convergence condition, thereby obtaining the target hybrid prediction model; the initial hybrid prediction model is a network model that fuses the multilayer perceptron prediction model used for process data training and the convolutional neural network prediction model used for image data training, and outputs the mechanical performance. The prediction module 860 is used to predict the stretch curve of the target metal sheet and strip using the target hybrid prediction model.

[0072] Based on this, the present invention provides a device for predicting the tensile curve of metal sheet and strip. The device can acquire target process data, microstructure images, and actual tensile curves of multiple metal sheet and strip materials using an acquisition module 810. The target process data is obtained by fusing M-dimensional process data vectors and N-dimensional uniquely encoded vectors from the initial process data. Then, a data extraction module 820 performs data extraction processing on the actual tensile curves to obtain target tensile data. The target tensile data consists of tensile data points from the yield strength point to the tensile strength point in the actual tensile curve. Next, a fitting module 830 performs n-order polynomial fitting on the target tensile data to obtain target polynomial coefficients. Further, an association construction module 840 is used to construct associations between the target process data, microstructure images, and the target polynomial coefficients to obtain a target mixed dataset. Finally, a model training module 850 uses the target mixed dataset to train the initial mixed data. The initial hybrid prediction model is trained until its prediction accuracy meets a preset convergence condition, thus obtaining the target hybrid prediction model. The initial hybrid prediction model is a network model that integrates a multilayer perceptron prediction model used for process data training and a convolutional neural network prediction model used for image data training, and outputs a mechanical performance network model. Finally, the prediction module 860 uses the target hybrid prediction model to predict the tensile curve of the target metal sheet / strip, thereby improving the prediction accuracy of the tensile curve of the metal sheet / strip and solving the problem of poor prediction accuracy of the tensile curve of metal sheet / strip in the prior art.

[0073] Although the invention has been described herein in conjunction with various embodiments, those skilled in the art will understand and implement other variations of the disclosed embodiments by reviewing the accompanying drawings, the disclosure, and the appended claims in carrying out the claimed invention. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude a plurality. A single processor or other unit can implement several functions listed in the claims. While different dependent claims may recite certain measures, this does not mean that these measures cannot be combined to produce good results.

[0074] Although the invention has been described in conjunction with specific features and embodiments, it is obvious that various modifications and combinations can be made therein without departing from the spirit and scope of the invention. Accordingly, this specification and drawings are merely exemplary descriptions of the invention as defined by the appended claims, and are considered to cover any and all modifications, variations, combinations, or equivalents within the scope of the invention. Clearly, those skilled in the art can make various alterations and modifications to the invention without departing from its spirit and scope. Thus, if such modifications and modifications of the invention fall within the scope of the claims and their equivalents, the invention is also intended to include such modifications and modifications.

Claims

1. A method for predicting the tensile curve of metal sheet and strip, characterized in that, include: Acquire target process data, microstructure images, and actual tensile curves for multiple metal sheet and strip materials; The target process data is the process data obtained by multi-dimensional data fusion of the M-dimensional process data vector and the N-dimensional one-hot encoded vector in the initial process data. The actual tensile curve is processed by data extraction to obtain target tensile data; the target tensile data is the tensile data points from the yield strength point to the tensile strength point in the actual tensile curve. The target stretching data are fitted with an nth-order polynomial to obtain the target polynomial coefficients. The correlation between the target process data and the microstructure image and the target polynomial coefficients are constructed respectively to obtain the target hybrid dataset; Based on the target mixed dataset, the initial mixed prediction model is trained until the prediction accuracy of the initial mixed prediction model meets the preset convergence condition, thus obtaining the target mixed prediction model; the initial mixed prediction model is a network model that fuses the multilayer perceptron prediction model used for process data training and the convolutional neural network prediction model used for image data training, and outputs the mechanical performance. The target hybrid prediction model is used to predict the tensile curve of the target metal sheet and strip.

2. The method for predicting the tensile curve of metal sheet and strip as described in claim 1, characterized in that, The prediction of the tensile curve of the target metal sheet / strip using the target hybrid prediction model includes: Acquire target process data and microstructure images of the target metal sheet / strip; Based on the target process data and microstructure image of the target metal sheet and strip, the tensile curve is predicted using the target hybrid prediction model to obtain the polynomial coefficients for the target metal sheet and strip. Based on the polynomial coefficients, a stretching curve for the target metal sheet / strip is constructed.

3. The method for predicting the tensile curve of metal sheet and strip as described in claim 1, characterized in that, The step of training the initial mixture prediction model based on the target mixture dataset includes, prior to: Construct the initial hybrid prediction model; The construction of the initial hybrid prediction model includes: A multi-input network model containing branches of ResNet and MLP networks is constructed. The merging layer of the multi-input network model is a network layer that combines the output neurons of ResNet and MLP networks. The network model parameters of the multi-input network model are optimized using the Adam optimization algorithm with stochastic gradient descent to obtain the initial hybrid prediction model. The merging layer is used to merge the layer above the regression output layer of ResNet and MLP networks into one layer.

4. The method for predicting the tensile curve of metal sheet and strip as described in claim 3, characterized in that, The ResNet network model includes a convolutional block attention module, which is used to increase the weights of the combined feature maps related to the prediction task.

5. The method for predicting the tensile curve of metal sheet and strip as described in claim 3, characterized in that, The ResNet network is used for data prediction of microscopic image data of metal sheets and strips; The MLP network is used to process numerical data to predict target process data for metal sheets and strips.

6. The method for predicting the tensile curve of metal sheet and strip as described in claim 1, characterized in that, Before obtaining the target process data for multiple metal sheet and strip materials, the following steps are included: Obtain initial process data for multiple metal sheet and strip materials; the initial process data includes at least rolling process data and heat treatment process data; The cooling methods of multiple metal strips are individually encoded to obtain an N-dimensional individually encoded vector; the N-dimensional individually encoded vector includes at least a water-cooled individually encoded vector and a furnace-cooled individually encoded vector. The target process data is obtained by multi-dimensional data fusion of the M-dimensional process data vector in the initial process data and the N-dimensional unique thermal encoding vector; the M-dimensional process data vector includes at least the rolling thickness process data vector, the reduction rate process data vector, the annealing temperature process data vector, and the holding time process data vector; the target process data is M plus N-dimensional process data.

7. The method for predicting the tensile curve of metal sheet and strip as described in claim 1, characterized in that, The acquisition of target process data, microstructure images, and actual tensile curves for multiple metal strips includes, prior to: The microstructure images of multiple metal strips were photographed layer by layer to obtain various microstructure photographs; Multiple microstructure photographs belonging to the same metal strip are stitched together to obtain microstructure images of multiple metal strips.

8. The method for predicting the tensile curve of metal sheet and strip as described in claim 7, characterized in that, The various microstructure photographs include at least inverse pole figures, Euler angle figures, diffraction quality figures, and recrystallization fraction figures; The step of stitching together multiple microstructure photographs belonging to the same metal strip to obtain microstructure images of multiple metal strips includes: Image conversion was performed on the inverse pole figure, Euler angle figure, diffraction quality figure, and recrystallization fraction figure of the same metal strip to obtain four images with a resolution of 1000 pixels. Images of various sizes; among them, Indicates the length of the image Width; The four pixels are Images of different sizes are combined and stitched together into a single image with a resolution of [number] pixels. The image was taken at a certain size to obtain a microstructure image of the metal sheet / strip.

9. The method for predicting the tensile curve of metal sheet and strip as described in claim 1, characterized in that, The process involves constructing the correlation between the target process data and the microscopic tissue images and the target polynomial coefficients, respectively, to obtain a target hybrid dataset, including: The microscopic tissue images are preprocessed, and the image data after each preprocessing is retained to obtain a target microscopic tissue image with a greater number of images than the number of images corresponding to the microscopic tissue image; the image preprocessing includes at least one of translation, rotation, and cropping; The correlation between the target process data, the target microstructure image, and the target polynomial coefficients is constructed to obtain the target hybrid dataset.

10. A device for predicting the tensile curve of metal sheet and strip, characterized in that, include: The acquisition module is used to acquire target process data, microstructure images, and actual tensile curves of multiple metal strips; The target process data is the process data obtained by multi-dimensional data fusion of the M-dimensional process data vector and the N-dimensional one-hot encoded vector in the initial process data. The data extraction module is used to extract data from the actual tensile curve to obtain target tensile data; the target tensile data is the tensile data points from the yield strength point to the tensile strength point in the actual tensile curve. The fitting module is used to perform nth-order polynomial fitting on the target stretching data to obtain the target polynomial coefficients. A correlation construction module is used to construct the correlation between the target process data and the microstructure image and the target polynomial coefficients, respectively, to obtain a target hybrid dataset; The model training module is used to train the initial hybrid prediction model based on the target hybrid dataset until the prediction accuracy of the initial hybrid prediction model meets the preset convergence condition, thereby obtaining the target hybrid prediction model. The initial hybrid prediction model is a network model that fuses a multilayer perceptron prediction model used for process data training and a convolutional neural network prediction model used for image data training, and outputs a mechanical performance network model. The prediction module is used to predict the tensile curve of the target metal sheet / strip using the target hybrid prediction model.

Images