An oriented electrical steel normalizing organization prediction and regulation method and system based on a multi-layer perception deep learning network and a cellular automaton, and a medium

Abstract

The application discloses an oriented electrical steel normalization structure prediction and regulation method and system based on a multi-layer perception deep learning network and a cellular automaton, and a medium, takes two-dimensional normalization process parameters as input, and takes an ODF slice map as output; a virtual labeling data of a "(t, T)-ODF" is generated in a preset process design space through a cellular automaton model driven by a physical mechanism, and a data set is constructed; then, a deep learning agent model is used to realize end-to-end nonlinear mapping from the two-dimensional normalization process parameters to the ODF slice map, so that fast prediction and visualization of the normalization structure are realized without calling the cellular automaton model one by one. The application solves the key problem of considering the prediction speed, structure accuracy and inversion design capability in an extremely short process window.

Description

Technical Field

[0001] This invention relates to metallic materials and intelligent manufacturing technology, and more specifically, to a method, system, and medium for predicting and controlling the normalization microstructure of oriented electrical steel based on multilayer perceptual deep learning networks and cellular automata. Background Technology

[0002] In the manufacturing process of grain-oriented electrical steel, short-time normalization after hot rolling is a crucial step, serving as a bridge between the preceding and following stages. Its main function is to rapidly reduce the dislocation density and texture inhomogeneity caused by the thickness gradient during a heating regime typically ranging from 780°C to 1100°C, lasting only a few minutes. The normalization result is not only affected by process parameters such as heating temperature, holding time, and heating rate, but also closely related to the initial microstructure after hot rolling. If the microstructure of the grain-oriented electrical steel is not properly shaped within this short time window, the texture evolution during subsequent cold rolling, decarburization, nitriding, and final annealing will be difficult to control, and key properties such as magnetic flux density and iron loss will fluctuate accordingly, thus profoundly determining the final product quality.

[0003] Currently, the prediction of tissue composition and optimization of the process mainly rely on the following methods, but they all have significant drawbacks:

[0004] (1) Experience-based trial and error method: This method relies heavily on engineers' experience and a large number of repeated laboratory test pieces for verification. The development cycle of this method is long (months or even years), the cost is high, and it is difficult to cover complex process spaces with high dimensions and nonlinearity, and it cannot guarantee the acquisition of the globally optimal process.

[0005] (2) Pure physical mechanism models (such as phase field method, cellular automata-CA): Although based on physical principles, the models have high credibility, but they consume huge computational resources. A single simulation often takes several hours or even days, which cannot meet the needs of industrial sites for rapid prediction and real-time feedback.

[0006] (3) Pure data-driven machine learning models: These require a large amount of high-quality labeled data (normalized EBSD data) for training. However, in industrial production, such data samples are scarce and extremely expensive to obtain. This results in poor generalization ability of the model under small sample conditions, a sharp drop in prediction accuracy when faced with different grades and batches of raw materials, and insufficient interpretability of the model, making it difficult to provide physical insights.

[0007] Although existing methods for tissue prediction and process design have achieved certain results in experimental verification and theoretical modeling, they still cannot simultaneously meet the industry's core requirements for "fast prediction speed, accurate prediction results, and strong reverse design capability" in practical engineering applications, demonstrating significant technical bottlenecks.

[0008] In conclusion, the current prediction and control of normalization microstructure faces the core contradiction of being unable to achieve both speed and accuracy. This limitation is particularly prominent in oriented electrical steel, because its product performance (such as magnetic induction and iron loss) is extremely sensitive to the evolution of microstructure, and the microstructure evolution is highly dependent on the coupling results of multiple factors such as the initial state of hot rolling, the trajectory of normalization heat treatment, and the evolution of the inhibiting phase. Summary of the Invention

[0009] To address the shortcomings of existing technologies, the present invention aims to provide a method, system, and medium for predicting and controlling the normalized microstructure of oriented electrical steel based on multilayer perceptual deep learning networks and cellular automata, which solves the key problem of balancing prediction speed, structural accuracy, and inversion design capability within an extremely short process window.

[0010] To achieve the above objectives, the present invention adopts the following technical solution:

[0011] The first aspect of this invention provides a method for predicting and controlling the normalization microstructure of oriented electrical steel based on multilayer perceptual deep learning networks and cellular automata.

[0012] The two-dimensional normalization process parameters are used as input, and the ODF slice diagram is used as output.

[0013] Virtual labeled data of “(t,T)-ODF” is generated in batches within a preset process design space using a cellular automata model driven by physical mechanisms, thus constructing a dataset.

[0014] Then, a deep learning proxy model is used to realize the end-to-end nonlinear mapping from the two-dimensional normalization process parameters to the ODF slice diagram, thereby achieving rapid prediction and visualization of normalized tissues without having to call the cellular automata model one by one.

[0015] Preferably, the method for predicting and controlling the normalizing microstructure of grain-oriented electrical steel includes the following steps:

[0016] S1, Multi-source data acquisition and preprocessing;

[0017] S2, Dataset generation based on the cellular automata model;

[0018] S3, Construction and training of the deep learning agent model;

[0019] S4, under the condition of the two-dimensional normalized process parameters, rapid prediction and visualization of the ODF slice diagram.

[0020] Preferably, step S1 specifically includes:

[0021] First, electron backscatter diffraction data of hot-rolled raw materials were collected, covering the surface, intermediate layer and core, with a scanning step size of 0.5~2.0μm;

[0022] Furthermore, the initial grain topology, crystallographic orientation information, grain boundary information, and micro-region texture features are extracted from the electron backscatter diffraction data.

[0023] Subsequently, the two-dimensional normalization process parameters are expressed as a two-dimensional vector u=(t,T), including the holding temperature T and the holding time t;

[0024] Finally, all collected data were standardized, normalized, and missing values ​​were removed to obtain a normalized process vector. .

[0025] Preferably, the cellular automata model in step S2 is used to simulate recovery, recrystallization, grain growth, and texture evolution during the normalization process.

[0026] Preferably, the cellular automata model integrates the following physical fields:

[0027] Grain boundary mobility: expressed as Where θ is the grain boundary orientation difference, T is the temperature, M0 is the pre-exponential factor, Q is the activation energy, and R is the gas constant;

[0028] Nucleation rate I and growth rate G: are related to local energy storage distribution and real-time temperature history;

[0029] By running the cellular automata model in batches within a preset process design space, large-scale "virtual-labeled" data pairs are generated, from the initial state to the normalized orientation distribution function (ODF) features.

[0030] Preferably, measured electron backscatter diffraction data is introduced to calibrate the systematic bias of the cellular automata model, thereby improving the physical fidelity and reliability of the generated data.

[0031] Preferably, in step S3, the deep learning proxy model employs a deep neural network in the form of a multilayer perceptron deep learning network to directly learn the mapping relationship between the two-dimensional normalized process parameters and the ODF slice map, thereby achieving the learning from... arrive End-to-end nonlinear regression.

[0032] Preferably, the deep neural network in the form of a multilayer perceptual deep learning network includes:

[0033] Input layer: The input dimension is 2, corresponding to the normalized heat preservation time t and heat preservation temperature T;

[0034] Several hidden layers: consisting of multiple fully connected layers and nonlinear activation functions stacked alternately, used to extract high-dimensional features of the process-texture mapping;

[0035] Output layer: It is a fully connected layer with an output dimension of 4096, corresponding to the pixel-by-pixel intensity prediction value of the ODF slice image.

[0036] Preferably, the output layer is followed by a Sigmoid or Tanh activation function to limit the predicted values ​​to a preset numerical range, so as to be consistent with the normalization method of the target data in step S1.

[0037] Preferably, in step S3, the deep learning proxy model uses mean squared error as the basic loss function to measure the vector of the predicted ODF slice map. Vectors of ODF slices in the real dataset The difference between them can be expressed in the following form:

[0038] .

[0039] Preferably, L1 loss and structural similarity index are introduced as auxiliary terms to impose additional constraints on the local contrast and texture preservation of the predicted ODF slice image.

[0040] Preferably, during the training process, the deep learning agent model employs a gradient descent-based optimization algorithm, iteratively optimizing the model parameters on the training set and monitoring the loss changes on the validation set. When the loss on the validation set converges or no longer decreases significantly, the training ends and the final model weights are saved.

[0041] Preferably, in step S4, after the deep learning agent model is trained, for any given combination of the two-dimensional constant process parameters, the ODF slice image can be quickly predicted through the following steps:

[0042] S41, apply the same standardization or normalization method as step S1 to the holding time t and holding temperature T to obtain the normalized process vector. ;

[0043] S42, the normalized process vector The intensity vector of the predicted ODF slice map is obtained by inputting it into the pre-trained deep neural network. ;

[0044] S43, the intensity vector of the ODF slice image Perform denormalization to restore the original ODF intensity dimensions;

[0045] S44, the vector of the ODF slice image Reconstructing the texture into ODF slices using a 64*64 spatial layout enables the visualization and prediction of texture evolution results under specific two-dimensional normalized process parameters.

[0046] Preferably, based on the prediction results of the ODF slice diagram, the operator can screen and optimize the normalizing process parameters to achieve control over the normalizing microstructure of the oriented electrical steel.

[0047] The second aspect of this invention provides a system for predicting and controlling the normalizing microstructure of oriented electrical steel, for implementing the method for predicting and controlling the normalizing microstructure of oriented electrical steel based on multilayer perceptron deep learning networks and cellular automata as described in the first aspect of this invention, comprising:

[0048] The data acquisition and virtual experiment module is used to acquire or simulate the initial microstructure of hot rolling, two-dimensional normalization process parameters and corresponding ODF slice diagrams, and to complete the data generation of the cellular automata model and the calibration with the measured data.

[0049] The data preprocessing and encoding module is used to normalize, flatten, and store the two-dimensional normalized process parameters and the ODF slice image to form a dataset that can be used for deep learning training.

[0050] The surrogate model training and prediction module is used to build, train and deploy a feedforward neural network in the form of a multilayer perceptual deep learning network to achieve fast forward prediction from the two-dimensional normalized process parameters to the ODF slice map.

[0051] The process scanning and visualization module is used to call proxy models in batches within a given process range, generate ODF prediction results and present them in the form of images or statistical features to assist engineers in process window evaluation and preliminary screening.

[0052] Preferably, the normalizing microstructure prediction and control system for grain-oriented electrical steel further includes:

[0053] An integrated module that interfaces with production line control or process design software to embed prediction results into existing process development workflows.

[0054] The present invention also discloses a computer-readable storage medium having a computer program stored thereon, which, when executed on a processor, implements the steps of the method for predicting and controlling the normalization microstructure of oriented electrical steel based on a multilayer perceptual deep learning network and cellular automata provided in the first aspect of the present invention.

[0055] This invention provides a method, system, and medium for predicting and controlling the normalization microstructure of oriented electrical steel based on multilayer perceptron deep learning networks and cellular automata. It relates to modeling and microstructure control of the heat treatment process of oriented electrical steel, designing and rapidly optimizing normalization processes, and digitally studying and verifying the evolution mechanism of microstructure. By using deep learning tools to accelerate materials science research and engineering applications, this invention is the first to apply deep learning methods to the field of electrical steel, and has the following beneficial effects:

[0056] (1) Significantly reduces the dependence on large-scale measured EBSD data and alleviates the problem of data scarcity. Compared with traditional pure data-driven machine learning methods that require a large amount of normalized EBSD data, this invention effectively alleviates the problem of high-quality EBSD data scarcity and high collection costs by using a "virtual experiment as the main method and measured data as a supplement" approach, and significantly lowers the threshold for model building;

[0057] (2) The complex physical forward modeling process is compressed into a lightweight MLP surrogate model, which greatly improves the prediction speed. This invention "distills" the complex micro-organism evolution process into a lightweight deep learning surrogate model, which significantly improves the prediction efficiency and makes it possible to quickly evaluate a large number of candidate process points within the production cycle.

[0058] (3) While maintaining the integrity of the texture information, the end-to-end mapping from "two-dimensional process to full-field ODF image" was realized. This invention realizes the end-to-end prediction of the entire ODF field driven by a very small number of process features (two parameters, t and T), which not only ensures the richness of the texture representation, but also maintains the simplicity of the model structure, which is conducive to subsequent coupling with the magnetic property model to carry out integrated analysis of "process-texture-performance". Attached Figure Description

[0059] Figure 1 This is a flowchart illustrating the method for predicting and controlling the normalization microstructure of grain-oriented electrical steel according to the present invention.

[0060] Figure 2 This is a schematic diagram of the cellular automata (CA) "virtual experiment" module structure and DoE sampling in the method for predicting and controlling the normalization microstructure of oriented electrical steel of the present invention;

[0061] Figure 3 This is a schematic diagram of measured normalized ODF data used for optimization in the method for predicting and controlling the normalized microstructure of oriented electrical steel in this invention;

[0062] Figure 4 This is a schematic diagram of the MLP model architecture used in Example 3 of the method for predicting and controlling the normalized microstructure of oriented electrical steel in this invention;

[0063] Figure 5 This is a schematic diagram comparing the ODF diagram predicted by the MLP model and the ODF effect obtained by actual CA simulation in Example 3 of the method for predicting and controlling the normalized microstructure of oriented electrical steel of the present invention. Detailed Implementation

[0064] To better understand the above-mentioned technical solutions of the present invention, the technical solutions of the present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0065] Combination Figure 1As shown, this invention provides a method for predicting and controlling the normalization microstructure of oriented electrical steel based on multilayer perceptual deep learning networks and cellular automata.

[0066] The two-dimensional normalization process parameters (holding temperature T and holding time t) are used as input, and an ODF slice image with a pixel size of 64*64 is used as output.

[0067] Virtual labeled data of “(t,T)-ODF(64*64)” can be generated in batches within a preset process design space using a cellular automaton (CA) model driven by physical mechanisms. A small amount of measured EBSD data can be used to calibrate the CA model and construct a training set.

[0068] Then, a deep learning surrogate model is used to realize the end-to-end nonlinear mapping from two-dimensional normalized process parameters to ODF slice diagrams, thereby enabling rapid prediction and visualization of normalized tissues without having to call the cellular automata model one by one.

[0069] The method for predicting and controlling the normalized microstructure of grain-oriented electrical steel according to the present invention specifically includes the following steps:

[0070] S1, Multi-source data acquisition and preprocessing;

[0071] First, electron backscatter diffraction (EBSD) data of hot-rolled raw materials were collected, covering the surface, intermediate layer and core, with a scanning step size of 0.5~2.0μm;

[0072] Furthermore, the initial grain topology, crystallographic orientation information, grain boundary information, and micro-region texture features are extracted from the electron backscatter diffraction data.

[0073] Subsequently, the two-dimensional normalization process parameters are expressed as a two-dimensional vector u=(t,T), including the holding temperature T and the holding time t;

[0074] Finally, all collected data were standardized, normalized, and missing values ​​were removed to obtain a normalized process vector. .

[0075] S2, a dataset for generating "virtual experiments" based on a cellular automata model, such as... Figure 2 As shown;

[0076] A cellular automata (CA) model based on physical mechanisms was constructed to simulate recovery, recrystallization, grain growth, and texture evolution during the normalization process.

[0077] The cellular automata model integrates the following key physics fields:

[0078] Grain boundary mobility: expressed as Where θ is the grain boundary orientation difference, T is the temperature, M0 is the pre-exponential factor, Q is the activation energy, and R is the gas constant;

[0079] Nucleation rate I and growth rate G: are related to local energy storage distribution and real-time temperature history;

[0080] By running cellular automata models in batches within a pre-defined process design space (DoE), large-scale “virtual-annotated” data pairs are generated, from the initial state (hot-rolled EBSD and process u) to the normalized orientation distribution function ODF features (pixel size 64*64).

[0081] In step S2, measured electron backscatter diffraction data can be further introduced to calibrate the systematic bias of the cellular automata model, so as to improve the physical fidelity and reliability of the generated data.

[0082] S3, Construction and training of deep learning agent models, such as Figure 3 As shown;

[0083] The deep learning proxy model employs a deep neural network in the form of a multi-layer perceptron (MLP) to directly learn the mapping relationship between two-dimensional normalized process parameters and ODF slice diagrams, thereby achieving... arrive End-to-end nonlinear regression.

[0084] Among them, deep neural networks in the form of MLP include:

[0085] Input layer: The input dimension is 2, corresponding to the normalized heat preservation time t and heat preservation temperature T;

[0086] Several hidden layers: consisting of multiple fully connected layers (the number of fully connected neurons can be 64~512) stacked alternately with nonlinear activation functions (preferably ReLU), used to extract high-dimensional features of the process-texture mapping;

[0087] Output layer: It is a fully connected layer with an output dimension of 4096, which corresponds to the pixel-by-pixel intensity prediction value of the 64*64 ODF slice image.

[0088] After the output layer, a Sigmoid or Tanh activation function can be added to limit the predicted values ​​to a preset range, so as to maintain consistency with the normalization method of the target data in step S1.

[0089] The deep learning proxy model uses mean squared error (MSE) as the basic loss function to measure the vector of the predicted ODF slice map. Vectors of ODF slices in the real dataset The difference between them can be expressed in the following form:

[0090] .

[0091] In step S3, L1 loss, structural similarity (SSIM) index, etc. can also be introduced as auxiliary terms to impose additional constraints on the local contrast and texture preservation of the predicted ODF slice image, thereby improving the visual and physical rationality of the predicted ODF image.

[0092] During the training process of the deep learning surrogate model, gradient descent-based optimization algorithms are employed, such as Adam or stochastic gradient descent (SGD) with momentum, or other gradient descent-based optimization algorithms such as RMSprop and AdamW. Appropriate learning rates, batch sizes, and training epochs are set. The model parameters are iteratively optimized on the training set, and the loss is monitored on the validation set. Training ends and the final model weights are saved when the validation set loss converges or no longer decreases significantly.

[0093] S4, Rapid prediction and visualization of ODF slice diagrams under two-dimensional normalized process parameters.

[0094] After the deep learning proxy model is trained, for any given combination of two-dimensional constant process parameters, the ODF slice map can be quickly predicted through the following steps:

[0095] S41, apply the same standardization or normalization method as in step S1 to the holding time t and holding temperature T to obtain the normalized process vector. ;

[0096] S42, normalize the process vector The intensity vector of the predicted 4096-dimensional ODF slice map is obtained by inputting into a pre-trained deep neural network. ;

[0097] S43, Intensity vector of ODF slice image The process involves inverse normalization to restore the original ODF intensity dimensions. Specifically, this means reconstructing the intensity values ​​of the ODF slice image before normalization by applying an inverse normalization mapping consistent with step S1 to the normalized prediction results output by the model. The original ODF intensity values ​​are derived from the orientation distribution function ODF slice calculated from crystal orientation data measured by EBSD or XRD.

[0098] ODF slice images are generated from raw orientation data acquired by EBSD and / or XRD. Specifically, the raw orientation data is first represented by Euler angles and orientation statistics are performed to obtain the orientation distribution function; then a preset ϕ2 section is selected, and the ODF intensity on this section is discretized and interpolated using a regular grid; finally, the discretized ODF intensity matrix is ​​converted into a two-dimensional image representation, which serves as the ODF slice image for model input or output.

[0099] S44, vector of ODF slice image Reconstructing the texture into ODF slices using a 64*64 spatial layout enables the visualization and prediction of texture evolution results under specific two-dimensional normalized process parameters.

[0100] Through the above steps, this invention can quickly obtain ODF slices under different time-temperature combinations without repeating complex simulations or time-consuming experiments, providing process engineers with an efficient tool for process window evaluation, texture optimization, and performance prediction.

[0101] The specific architecture of the deep learning agent model in this invention adopts a feedforward neural network in the form of an MLP, using the normalized process vector. The input is processed through several fully connected layers and a non-linear activation function (preferably ReLU) for feature extraction. The final output is a 4096-dimensional vector, corresponding to the pixel-wise intensity of a 64*64 ODF slice. After the output layer, a Sigmoid or Tanh activation function is applied to limit the prediction results to a numerical range consistent with the normalization method of the training data. The 4096-dimensional vector is then restored to a 64*64 ODF image through a reshape operation.

[0102] In this invention, the data representation and loss function design uniformly encode the normalized process parameters (holding time, holding temperature) into two-dimensional continuous features and perform standardization or interval normalization; the ODF cross-section is discretized into 64*64 pixels with a fixed spatial layout and flattened into a target vector of length 4096. Optionally, the ODF intensity is logarithmically transformed or normalized to the maximum value before training to improve numerical stability; mean square error (MSE) is used as the basic loss function to measure the difference between the predicted ODF vector and the true ODF vector. Preferably, auxiliary terms such as L1 loss and structural similarity (SSIM) can be superimposed to improve the performance of the predicted image in terms of local contrast and texture preservation.

[0103] In this invention, process window scanning and quasi-inverse screening based on a surrogate model utilize a pre-trained MLP model to perform high-density sampling and batch inference within a continuous time-temperature process space, quickly obtaining ODF slice prediction maps corresponding to a large number of candidate process points. By filtering and sorting specific features of the predicted ODF (such as the intensity of a certain orientation interval, texture uniformity, etc.), process parameter regions that meet preset texture indices are identified, providing candidate solution sets for subsequent refined process optimization or coupling with performance models, thereby achieving process window search that supports approximate "reverse design" with low computational cost.

[0104] The present invention also provides a system for predicting and controlling the normalizing microstructure of grain-oriented electrical steel for implementing the method of predicting and controlling the normalizing microstructure of grain-oriented electrical steel of the present invention, comprising:

[0105] The data acquisition and virtual experiment module is used to acquire or simulate the initial microstructure of hot rolling, two-dimensional normalization process parameters and corresponding ODF slice diagrams, and to complete the data generation of the cellular automata model and the calibration with the measured data.

[0106] The data preprocessing and encoding module is used to normalize, flatten, and store the two-dimensional normalized process parameters and ODF slice diagrams to form a dataset that can be used for deep learning training.

[0107] The surrogate model training and prediction module is used to build, train and deploy a feedforward neural network in the form of an MLP to achieve fast forward prediction from two-dimensional normalized process parameters to ODF slice diagrams.

[0108] The process scanning and visualization module is used to call proxy models in batches within a given process range, generate ODF prediction results and present them in the form of images or statistical features to assist engineers in process window evaluation and preliminary screening.

[0109] The present invention's system for predicting and controlling the normalized microstructure of grain-oriented electrical steel also includes:

[0110] The integrated module interfaces with production line control or process design software to embed the prediction results into the existing process development process, thereby realizing the engineering application of the method for predicting and controlling the normalization microstructure of oriented electrical steel in the industrial field.

[0111] Example 1

[0112] In this embodiment 1, oriented electrical steel hot-rolled plates from the same heat number were selected, and EBSD scanning was performed on the thickness direction of the plates with a step size of 2μm to obtain representative initial texture information.

[0113] The design space for the two-dimensional normalization process parameters is set as follows:

[0114] Temperature range: 800~1100℃

[0115] Heat preservation time range: 60s~240s

[0116] The holding temperature and holding time are used to construct a two-dimensional process vector u=(t,T), which is then mapped to the interval [0,1] using the min-max normalization method to obtain the normalized process vector. .

[0117] Within the design space of the normalization process, 400 sets of process points (ti,Ti) are generated by combining regular grids and random sampling. The normalization process is simulated one by one in the CA model, and the normalized ODF slice is output.

[0118] The CA output is used directly as the "truth value" to train the surrogate model.

[0119] The ODF slices are also standardized to 64*64 pixels and expanded into a 4096-dimensional vector. To alleviate the problem of large dynamic range of ODF, in this embodiment 1, the ODF intensity is first logarithmically transformed (log(ODF+0.01)) and then min-max normalized to improve numerical stability.

[0120] This embodiment 1 adopts a more compact MLP structure to reduce deployment difficulty and parameter scale.

[0121] In the network architecture, the input layer has a dimension of 2, corresponding to the normalized process vector; the hidden layer consists of two fully connected layers with 64 and 128 neurons respectively, each followed by a ReLU activation function; a dropout of 0.1 is added after the second hidden layer to suppress overfitting; the output layer has a dimension of 4096, and instead of additional non-linear activation, it directly outputs the data and matches the normalized log-ODF intensity in the training objective.

[0122] The training configuration involved dividing 400 sets of dummy data into training, validation, and test sets in a 7:1.5:1.5 ratio; the loss function was MSE, the optimizer was Adam, and the learning rate was set to 5×10⁻⁶. -4 The batch size is 16, the maximum number of training rounds is 300, and early stopping is used on the validation set.

[0123] After training, the average pixel-level MSE of the test set remained within the range of 0.001 to 0.003 across the entire design space of 800℃ to 920℃ and 60s to 420s, meeting the requirements for process pre-research and preliminary screening. This surrogate model takes ≤10ms for a single ODF prediction, while traditional cellular automata take ≥4 hours for a single simulation, representing a prediction speed improvement of over 1.4 million times, which can meet the real-time prediction needs of industrial sites.

[0124] For the i-th sample in the test set, let the true ODF slice image be... Predicted ODF slice image as The single-sample pixel-level mean square error is defined as:

[0125]

[0126] The average pixel-level MSE of the test set is defined as follows:

[0127]

[0128] Where M is the number of samples in the test set. The test set is an independent set of samples that did not participate in model training and hyperparameter selection. It is obtained by dividing the overall dataset according to a preset ratio and is independent of the training set and validation set. The MSE statistic is calculated based on the prediction results obtained after reasoning through all test samples one by one.

[0129] Example 2

[0130] This second embodiment is more refined than the first embodiment, especially in scenarios with abundant data and model computing resources.

[0131] In this embodiment 2, oriented electrical steel hot-rolled plates from the same heat number were selected, and EBSD scanning was performed on the thickness direction of the plates with a step size of 0.5 μm to obtain representative initial texture information with higher microstructure characterization accuracy.

[0132] The design space for the two-dimensional normalization process parameters is set as follows:

[0133] Temperature range: 780~1150℃

[0134] Heat preservation time range: 30s~480s

[0135] The holding temperature and holding time are used to construct a two-dimensional process vector u=(t,T), and the normalized process vector is obtained using the z-score standard method. .

[0136] Based on the initial microstructure of each hot-rolled component, a corresponding CA model is constructed to simulate the normalization process within the aforementioned design space. 800 process points (ti, Ti) are generated using Latin hypercube sampling, and the normalization process is simulated one by one in the CA model, outputting the normalized ODF slice diagram.

[0137] Meanwhile, within the preferred process sub-range of 900℃~1050℃ and 180s~360s, 20 representative process combinations were selected for actual normalization tests and EBSODF data were measured to form 20 sets of "process-measured ODF" data pairs.

[0138] This embodiment 2 employs a multi-source fusion strategy:

[0139] Use 800 sets of CA virtual data as the main dataset;

[0140] Twenty sets of actual test data were used as a high-fidelity calibration set and given higher weight in the training loss to guide the surrogate model to align with the actual test data.

[0141] All ODF slices are uniformly interpolated to 64*64 pixels and then processed by log transformation and normalization to further improve training stability.

[0142] This embodiment 2 employs a high-capacity architecture: Input layer: 2-dimensional; First hidden layer: 256 neurons + ReLU; Second hidden layer: 512 neurons + ReLU; Third hidden layer: 256 neurons + ReLU; Dropout of 0.2 is added after the second hidden layer; 4096 dimensions are used to regress the normalized log-ODF pixel intensity; Tanh activation is applied after the output layer, and the data is linearly transformed to the normalized interval of the training data.

[0143] Subsequently, 800 sets of virtual data were mixed with 20 sets of real-world data, with 85% used as the training set, 10% as the validation set, and 5% as the test set. The loss function was primarily MSE, with different weights applied to the virtual and real-world data. The weighting coefficient for the real-world data samples was 3, and the weighting coefficient for the virtual data was 1, to highlight the importance of the real-world data. Simultaneously, a small weighted term of SSIM (structural similarity) loss was superimposed on the MSE to improve the texture consistency of the ODF images. The optimizer was AdamW (Adam with weight decay), with a learning rate of 1×10⁻⁶. -3 Weight decay coefficient 1×10 -4 The batch size is 32, and the maximum number of training rounds is 300. Early stopping on the validation set is adopted, and the model parameters with the minimum validation loss are saved during training.

[0144] After training, the pixel-level MSE on the entire test set was significantly lower than that in Example 1, and the position deviation of the main texture peak was controlled within 2°~3°. (Goss, {111}) <112> The intensity prediction error of key texture components is usually less than 5%. This invention only requires 20 sets of measured EBSD data to complete the model calibration, while traditional pure data-driven models have a prediction error of ≥30% with the same sample size, which greatly reduces the dependence on high-cost measured data.

[0145] Example 3

[0146] In this embodiment 3, oriented electrical steel hot-rolled plates of the same heat number were selected, and EBSD scanning with a step size of 1μm was used in the thickness direction to obtain the ODF orientation distribution of the initial microstructure of the hot-rolled plate.

[0147] The design space for the two-dimensional normalization process parameters is set as follows:

[0148] Temperature range: 800~1150℃

[0149] Heat preservation time range: 30s~360s

[0150] In this embodiment 3, the sub-intervals of 800℃~1100℃ and 60s~240s are focused on during the model training and process scanning stages to cover the short-term windows commonly used in actual industry.

[0151] To construct the surrogate model, a two-dimensional process vector u is formed by taking the constant holding temperature T and the holding time t. The mean and standard deviation of t and T for all samples are calculated, and a normalized process vector is obtained by standardization. .

[0152] Based on the initial microstructure of the hot-rolled EBSD, a cellular automaton (CA) model was constructed to simulate the recovery, recrystallization and grain growth during the normalization process, and the orientation distribution function (ODF) slice diagram after normalization was output as a characterization of the target microstructure.

[0153] Within the design space, 600 sets of process points (t) were generated using Latin hypercube sampling. i ,T i CA simulations were run one by one to obtain 600 sets of virtual "(t,T)–ODF slice" data pairs. The ODF slices were uniformly interpolated to a 64*64 discrete grid and expanded into a 4096-dimensional vector.

[0154] To improve the physical reliability of the CA virtual data, 12 additional process combinations were selected within process sub-intervals of 1 min, 2 min, 3 min, and 4 min at 800℃, 900℃, and 1000℃. Actual normalization experiments were conducted, and normalized EBSD data were obtained. These measured ODF values ​​were compared with the CA output, and a simple linear deviation correction method was used to correct the overall CA output. The corrected 600 sets of virtual data were used as the labeled dataset for subsequent deep learning training. Figure 3 This is the measured process ODF diagram used for calibration.

[0155] The MLP model architecture used in this embodiment 3 is as follows:

[0156] Input layer: 2-dimensional, corresponding to the normalized process vector; Hidden layer: three fully connected layers with 128, 256, and 512 neurons respectively, each followed by a ReLU activation function; Output layer: a fully connected layer with an output dimension of 4096, corresponding to the pixel-by-pixel intensity prediction value of a 64*64 slice image. A Sigmoid activation is applied after the output layer to restrict the output to the [0,1] interval, matching the min-max normalization operation of the ODF intensity in the preprocessing stage. The model architecture of this embodiment 3 is shown in the appendix. Figure 4 As shown.

[0157] From 600 sets of corrected virtual data, 480 sets were randomly selected as the training set, 60 sets as the validation set, and 60 sets as the test set; the mean squared error (MSE) was used as the loss function; the Adam optimizer was used with a learning rate of 1×10⁻⁶. -3 The batch size is 32, and the maximum number of training epochs is set to 200. The MSE is monitored on the validation set. If the validation loss does not decrease significantly for 20 consecutive epochs, the early stopping strategy is triggered.

[0158] Training results show that, on the test set, the root mean square error of the predicted ODF pixel intensity is controlled within 8%, and the deviations between the predicted location and intensity of the local principal texture peaks are within acceptable limits.

[0159] The ODF microstructure was predicted using the model constructed in Example 3 of this study for a normalization process at 1100℃ for 2 minutes. ODF comparisons are attached. Figure 5 As shown, the position deviation of the main Goss peak is less than 3°, the intensity error of the Goss peak is about 7%, and the overall pixel-level MSE is comparable to the average level of the test set. The evaluation method is as follows:

[0160] First, a reference ODF slice is calculated from measured EBSD or XRD orientation data, and a predicted ODF slice is output by the model. Then, the local main peak positions and peak intensities of the measured and predicted slices are searched within a preset Goss target orientation neighborhood. The main Goss peak position deviation is defined as the minimum orientation angle difference between the predicted and measured main peak positions; the Goss peak intensity error is defined as the percentage difference between the predicted and measured main peak intensities relative to the measured main peak intensity. Preferably, the peak position search is performed within a preset tolerance angle range, and the peak intensity is represented by the local maximum value in the ODF slice. These evaluation metrics are used to characterize the model's accuracy in predicting the positions and intensities of key texture components.

[0161] Those skilled in the art should recognize that the above embodiments are merely illustrative of the present invention and are not intended to limit the present invention. Any variations or modifications to the above embodiments that are within the spirit and essence of the present invention will fall within the scope of the claims of the present invention.

Claims

1. A method for predicting and controlling the normalization microstructure of oriented electrical steel based on multilayer perceptual deep learning networks and cellular automata, characterized in that: The two-dimensional normalization process parameters, including the holding temperature T and the holding time t, are used as inputs, and an ODF slice image with a pixel size of 64*64 is used as the output. Virtual labeled data of "(t,T)-ODF" are generated in batches within a preset process design space using a cellular automata model driven by physical mechanisms, thus constructing a dataset. Then, a deep learning proxy model is used to realize the end-to-end nonlinear mapping from the two-dimensional normalization process parameters to the ODF slice diagram, thereby achieving rapid prediction and visualization of normalized tissues without having to call the cellular automata model one by one. The method for predicting and controlling the normalized microstructure of grain-oriented electrical steel includes the following steps: S1, Multi-source data acquisition and preprocessing; Electron backscatter diffraction data of hot-rolled raw materials were collected, covering the surface, intermediate layer and core. The initial grain topology, crystallographic orientation information, grain boundary information, and micro-region texture features are extracted from the electron backscatter diffraction data. The two-dimensional normalization process parameters are expressed as a two-dimensional vector u=(t,T); All collected data were standardized, normalized, and missing value processed to obtain a normalized process vector. ; S2, Dataset generation based on the cellular automata model; A cellular automaton (CA) model based on physical mechanisms was constructed to simulate recovery, recrystallization, grain growth, and texture evolution during the normalization process. The cellular automata model integrates the following physical fields: Grain boundary mobility: expressed as Where θ is the grain boundary orientation difference, T is the temperature, M0 is the pre-exponential factor, Q is the activation energy, and R is the gas constant; Nucleation rate I and growth rate G: are related to local energy storage distribution and real-time temperature history; By running the cellular automata model in batches within a preset process design space, a large-scale "virtual-labeled" data pair is generated from the initial state to the normalized orientation distribution function (ODF) characteristics; and a small amount of measured electron backscatter diffraction data is used to calibrate the systematic bias of the cellular automata model. S3, Construction and training of the deep learning agent model; The deep learning proxy model employs a deep neural network in the form of a multilayer perceptron deep learning network to directly learn the mapping relationship between the two-dimensional normalized process parameters and the ODF slice map, thereby realizing the learning from... arrive The end-to-end nonlinear regression; the deep learning proxy model uses mean squared error as the basic loss function and introduces L1 loss and structural similarity index as auxiliary terms to impose additional constraints on the local contrast and texture preservation of the predicted ODF slice image. S4, under the condition of the two-dimensional normalized process parameters, rapid prediction and visualization of the ODF slice diagram; Based on the prediction results of the ODF slice diagram, the operator screens and optimizes the normalizing process parameters to achieve control over the normalizing microstructure of oriented electrical steel.

2. The method for predicting and controlling the normalization microstructure of oriented electrical steel based on multilayer perceptual deep learning networks and cellular automata according to claim 1, characterized in that, In step S1, the scanning step size for collecting electron backscatter diffraction data of hot-rolled raw materials is 0.5~2.0μm.

3. The method for predicting and controlling the normalization microstructure of oriented electrical steel based on multilayer perceptual deep learning networks and cellular automata according to claim 1, characterized in that, In step S3, the deep neural network in the form of a multilayer perceptual deep learning network includes: Input layer: The input dimension is 2, corresponding to the normalized heat preservation time t and heat preservation temperature T; Several hidden layers: consisting of multiple fully connected layers and nonlinear activation functions stacked alternately, used to extract high-dimensional features of the process-texture mapping; Output layer: It is a fully connected layer with an output dimension of 4096, corresponding to the pixel-by-pixel intensity prediction value of the ODF slice image.

4. The method for predicting and controlling the normalization microstructure of oriented electrical steel based on multilayer perceptual deep learning networks and cellular automata according to claim 3, characterized in that: The output layer is followed by a Sigmoid or Tanh activation function to limit the predicted values ​​to a preset range, so as to maintain consistency with the normalization method of the target data in step S1.

5. The method for predicting and controlling the normalization microstructure of oriented electrical steel based on multilayer perceptual deep learning networks and cellular automata according to claim 1, characterized in that, In step S3, the basic loss function is used to measure the vector of the predicted ODF slice map. Vectors of ODF slices in the real dataset The difference between them can be expressed in the following form: 。 6. The method for predicting and controlling the normalization microstructure of oriented electrical steel based on multilayer perceptual deep learning networks and cellular automata according to claim 5, characterized in that: During the training process, the deep learning agent model adopts a gradient descent-based optimization algorithm. It iteratively optimizes the model parameters on the training set and monitors the loss changes on the validation set. When the loss on the validation set converges or no longer decreases significantly, the training ends and the final model weights are saved.

7. The method for predicting and controlling the normalization microstructure of oriented electrical steel based on multilayer perceptual deep learning networks and cellular automata according to claim 4, characterized in that, In step S4, after the deep learning agent model is trained, for any given combination of two-dimensional constant process parameters, the ODF slice image can be quickly predicted through the following steps: S41, apply the same standardization or normalization method as step S1 to the holding time t and holding temperature T to obtain the normalized process vector. ; S42, the normalized process vector The intensity vector of the predicted ODF slice map is obtained by inputting it into the pre-trained deep neural network. ; S43, the intensity vector of the ODF slice image Perform denormalization to restore the original ODF intensity dimensions; S44, the vector of the ODF slice image Reconstructing the texture into ODF slices using a 64*64 spatial layout enables the visualization and prediction of texture evolution results under specific two-dimensional normalized process parameters.

8. A system for predicting and controlling the normalizing microstructure of oriented electrical steel based on a multilayer perceptual deep learning network and cellular automata as described in any one of claims 1-7, characterized in that, include: The data acquisition and virtual experiment module is used to acquire or simulate the initial microstructure of hot rolling, two-dimensional normalization process parameters and corresponding ODF slice diagrams, and to complete the data generation of the cellular automata model and the calibration with the measured data. The data preprocessing and encoding module is used to normalize, flatten, and store the two-dimensional normalized process parameters and the ODF slice image to form a dataset that can be used for deep learning training. The surrogate model training and prediction module is used to build, train and deploy a feedforward neural network in the form of a multilayer perceptual deep learning network to achieve fast forward prediction from the two-dimensional normalized process parameters to the ODF slice map. The process scanning and visualization module is used to call proxy models in batches within a given process range, generate ODF prediction results and present them in the form of images or statistical features to assist engineers in process window evaluation and preliminary screening.

9. The system for predicting and controlling the normalization microstructure of oriented electrical steel based on multilayer perceptual deep learning networks and cellular automata according to claim 8, characterized in that, The system for predicting and controlling the normalized microstructure of grain-oriented electrical steel also includes: An integrated module that interfaces with production line control or process design software to embed prediction results into existing process development workflows.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed on a processor, it implements the steps of the method for predicting and controlling the normalization microstructure of oriented electrical steel based on multilayer perceptual deep learning networks and cellular automata as described in any one of claims 1-7.

Images