Analysis method, system and medium for dynamic recrystallization microstructure of titanium alloy based on machine learning

Abstract

The present invention discloses a method, system and medium for analyzing the dynamic recrystallization microstructure of titanium alloys based on machine learning, which belongs to the technical field of quantitative characterization of the microstructure of metal materials. The titanium alloy to be tested is subjected to a Gleeble hot compression test and multimodal data is collected. Based on the DRX segmentation model, a DRX probability map is output, and then the features of DRX are extracted. The features of DRX are respectively input into the primary classifier and PCA dimensionality reduction processing, the predicted probability is spliced ​​and fused with the features after dimensionality reduction, and input into the secondary classifier to output a two-dimensional DRX segmentation map. Based on the FIB‑SEM tomography sequence images, a three-dimensional model of the DRX grains is reconstructed to perform consistency verification on the two-dimensional DRX segmentation map; the correlation coefficient between the three-dimensional model and the two-dimensional DRX model is evaluated, and finally the spatial aggregation of the DRX grains is visualized in three dimensions. The present invention significantly improves the degree of automation and efficiency of dynamic recrystallization ratio and type identification.

Description

Technical Field

[0001] The present invention belongs to the technical field of quantitative characterization of the microstructure of metal materials, and specifically relates to an analysis method, system and medium for the dynamic recrystallization microstructure of titanium alloys based on machine learning. Background Art

[0002] Titanium alloys often experience dynamic recrystallization (DRX) during thermal deformation. However, varying deformation temperature, strain rate, and strain often lead to different DRX phenomena, including discontinuous dynamic recrystallization (DDRX), continuous dynamic recrystallization (CDRX), and geometric dynamic recrystallization (GDRX). Quantitatively analyzing the proportions and types of DRX is a complex and important task.

[0003] Traditional analysis methods rely on metallographic analysis and EBSD local misorientation (GOS / KAM) threshold segmentation to assess the proportion and type of dynamic recrystallization (DRX) in titanium alloys. However, manually setting thresholds is highly subjective, with errors as high as 10%-15%, and it is difficult to distinguish DRX types. Specifically, manual analysis cannot accurately determine and identify grain boundary and subgrain boundary misorientation within each grain. However, grain boundary misorientation is a key factor in determining whether a grain is DDRX, CDRX, or GDRX. DDRX grains exhibit characteristics such as bowed grain boundaries and newly formed equiaxed grains; CDRX grains exhibit subgrain structure, elongated grains, and subgrain torsion; and GDRX grains exhibit elongated grains with aspect ratios greater than 5, and generally exhibit shear band distribution and grain boundary serrations. However, manual processing of this data, due to the grain boundary angle threshold setting, makes it difficult to determine the type of dynamic recrystallization grain observed, and accurate determination of the proportion of dynamic recrystallization is also impossible.

[0004] In recent years, with the development of computer vision and artificial intelligence technologies, researchers have begun to apply machine learning algorithms to the identification and analysis of material microstructures. Machine learning combined with digital image recognition and processing technology is expected to solve the efficiency and consistency issues of manual quantitative analysis of dynamic recrystallization ratios and types. In particular, machine learning technology has the ability to automatically extract image features according to a set program and has achieved great success in complex image pattern recognition. In the field of materials research, there are also studies applying deep learning to the microstructure analysis of materials. However, at present, research on machine learning for identifying the ratios and types of DDRX, CDRX, and GDRX is still blank. Most existing automated microstructure analysis methods still have the following shortcomings:

[0005] On the one hand, traditional machine learning methods (such as support vector machines (SVMs)) often require manual design of image features, resulting in small model sizes that make it difficult to fully express the complex characteristics of microstructures, thus limiting recognition accuracy and robustness. Even convolutional neural networks (CNNs), which have been widely used in recent years, are limited in their receptive field and primarily focus on extracting local image features, sometimes failing to account for information that requires global discrimination.

[0006] On the other hand, the data acquisition and annotation of microstructure images is expensive. Different thermal deformation conditions in titanium alloys can produce a variety of microstructures, requiring the model to have good generalization capabilities. Existing small deep learning models are prone to overfitting when training data is limited, resulting in unstable performance across different material batches or when imaging conditions vary.

[0007] In summary, a new technical solution is currently needed that uses more advanced machine learning architectures and strategies to efficiently and accurately identify the proportion and type of dynamic recrystallization in titanium alloys with limited training data and provide quantitative analysis results. Summary of the Invention

[0008] The purpose of the present invention is to provide a method, system and medium for analyzing the dynamic recrystallization microstructure of titanium alloys based on machine learning, aiming to improve the degree of automation and efficiency of dynamic recrystallization ratio and type identification and reduce the influence of human factors.

[0009] The present invention is mainly achieved through the following technical solutions:

[0010] The analysis method of dynamic recrystallization microstructure of titanium alloy based on machine learning includes the following steps:

[0011] Step S1: performing a Gleeble hot compression test on the titanium alloy to be tested and then cooling the alloy;

[0012] Step S2: Acquiring multimodal data: obtaining orientation imaging IPF images, grain orientation difference KAM images, TEM images, and FIB-SEM tomographic sequence images of the titanium alloy to be tested;

[0013] Step S3: extracting a two-dimensional DRX segmentation map based on the two-dimensional DRX model;

[0014] Step S31: inputting the three-channel fusion data of the orientation imaging IPF map, the grain orientation difference KAM map and the TEM image into the DRX segmentation model, marking the areas of complete recrystallization, partial recrystallization and non-recrystallization, and outputting the DRX probability map;

[0015] Step S32: extracting DRX features based on the DRX probability map;

[0016] Step S33: Input the DRX features into a first-level classifier to distinguish DRX areas from non-DRX areas and output predicted probabilities; perform PCA dimensionality reduction on the output DRX features and retain the first three principal component features;

[0017] Step S34: The predicted probability is concatenated with the three principal component features and input into a secondary classifier to distinguish DDRX, CDRX, and GDRX types, and output a two-dimensional DRX segmentation map;

[0018] Step S4: reconstructing a three-dimensional model of the DRX grain based on the FIB-SEM tomographic sequence images;

[0019] Step S5: Based on the three-dimensional model of the DRX grain, perform consistency check on the two-dimensional DRX segmentation map; evaluate the correlation coefficient between the three-dimensional model and the two-dimensional DRX model; if the consistency check is successful and the correlation coefficient meets the set conditions, the spatial clustering of the DRX grains is visualized in three dimensions.

[0020] In order to better realize the present invention, further, the titanium alloy to be tested Phase region and The phase region is subjected to a Gleeble hot compression test and a rapid cooling treatment is performed, and the cooling rate is greater than or equal to 200°C / s; The deformation temperature of the phase region is 900-1000°C; The deformation temperature of the phase region is 750-880℃; the low strain rate of the Gleeble hot compression test is 0.001-0.1 , high strain rate of 1-10 .

[0021] In order to better implement the present invention, further, in the step S31, first, data preprocessing is performed: the orientation imaging IPF map and the grain orientation difference KAM map are weightedly fused, and the fusion weight of the RGB three channels of the orientation imaging IPF map is 1:1:1, and the fusion weight of the grain orientation difference KAM map is 0.8; the TEM image is histogram equalized.

[0022] In order to better implement the present invention, further, in step S31, the loss function of the DRX segmentation model is the weighted cross entropy loss of complete recrystallization, partial recrystallization and non-recrystallization, and the corresponding weighted weight ratio is 1:2:1.5.

[0023] In order to better implement the present invention, further, step S4 includes the following steps:

[0024] Step S41: inputting FIB-SEM tomographic sequence images and aligning inter-layer images using SIFT feature matching;

[0025] Step S42: pre-processing the image;

[0026] Step S43: Adopting an adaptive threshold segmentation algorithm to perform adaptive threshold segmentation to obtain a binary image, separating the DRX grains from the background; performing a morphological closing operation on the binary image to fill the pores and smooth the grain boundaries;

[0027] Step S44: using a watershed algorithm based on distance transformation and local gradient to segment the adhered grains;

[0028] Step S45: using the Marching Cubes algorithm, converting the segmentation result of step S44 into a three-dimensional mesh model.

[0029] In order to better implement the present invention, further, step S5 includes the following steps:

[0030] Step S51: obtaining a two-dimensional DRX segmentation map, and extracting a three-dimensional cross section at the same position from the three-dimensional model of the DRX die;

[0031] Step S52: extract key points from the two-dimensional DRX segmentation map and the three-dimensional cross section, calculate the affine transformation matrix, and perform spatial registration;

[0032] Step S53: performing consistency check on the 2D DRX segmentation map: calculating the overlap ratio between the 2D DRX segmentation map and the DRX region in the 3D cross section. If the overlap ratio is greater than a set threshold, the region consistency meets the standard.

[0033] Step S54: In the three-dimensional model of the DRX grain, N three-dimensional cross sections are taken at equal intervals along the ND direction, and the DRX area ratio of each three-dimensional cross section is calculated;

[0034] Step S55: The calculation formula of the correlation coefficient r is:

[0035] ;

[0036] in: For the The DRX area ratio of each three-dimensional cross section;

[0037] is the average of the DRX area proportions of N three-dimensional sections;

[0038] For the DRX area ratio of a two-dimensional DRX segmentation map;

[0039] is the average value of the DRX area proportions of N two-dimensional DRX partition maps.

[0040] In order to better realize the present invention, further, if , or r<0.97, the abnormal section is automatically marked and the review process is carried out; check whether the segmentation threshold of the three-dimensional model of the DRX grain is adapted to the current section. If not, repeat steps S51 and S55 after re-aligning or locally optimizing the segmentation parameters; verify whether the training data of the two-dimensional DRX model covers the characteristics of the region.

[0041] The present invention is mainly achieved through the following technical solutions:

[0042] A machine learning-based titanium alloy dynamic recrystallization microstructure analysis system is used to implement the above-mentioned machine learning-based titanium alloy dynamic recrystallization microstructure analysis method, including a multimodal data acquisition module, a two-dimensional model detection module, a three-dimensional model construction module and a verification module;

[0043] The multimodal data acquisition module is used to obtain the orientation imaging IPF map, grain orientation difference KAM map, TEM image and FIB-SEM tomography sequence image of the titanium alloy to be tested;

[0044] The two-dimensional model detection module is used to output a two-dimensional DRX segmentation map based on the two-dimensional DRX model, and includes a DRX segmentation module, a DRX feature extraction module, a primary classifier, a PAC dimension reduction module, a feature fusion module, and a secondary classifier, which are arranged in order from the front to the back; the DRX feature extraction module is respectively connected to the primary classifier and the PAC dimension reduction module, and the primary classifier and the PAC dimension reduction module are respectively connected to the feature fusion module;

[0045] The three-dimensional model building module is used to reconstruct the three-dimensional model of the DRX grain based on the FIB-SEM tomographic sequence images and using the threshold segmentation and watershed algorithm;

[0046] The verification module is used to perform consistency verification on the two-dimensional DRX segmentation map based on the three-dimensional model of the DRX grain; and evaluate the correlation coefficient between the three-dimensional model and the two-dimensional DRX model.

[0047] In order to better implement the present invention, further, the DRX segmentation model includes an encoder and a decoder, the encoder is used to extract multi-scale features of microscopic tissues through multi-level downsampling, and the decoder is used to restore details through upsampling and skip connections;

[0048] The encoder includes a first coding layer to a fourth coding layer connected sequentially from top to bottom, wherein the first coding layer includes a 7×7 convolution layer, a ReLU activation function, and a 2×2 maximum pooling layer arranged sequentially from front to back, for capturing low-level features of grain boundaries and subgrain boundary textures; the second coding layer includes three residual blocks, a ReLU activation function, and a 2×2 maximum pooling layer arranged sequentially from front to back, for extracting morphological features of grain boundary curvature and bow height; the third coding layer includes four residual blocks, a ReLU activation function, and a 2×2 maximum pooling layer arranged sequentially from front to back, for learning orientation difference distribution and DRX area distribution trend; the fourth coding layer includes six residual blocks, a ReLU activation function, and a 2×2 maximum pooling layer arranged sequentially from front to back, for encoding global semantics;

[0049] The decoder includes a fourth decoding layer to a first decoding layer sequentially jump-connected from top to bottom, and the fourth decoding layer to the second decoding layer respectively include an upsampling layer and an SE attention module sequentially arranged from front to back;

[0050] The outputs of the first to fourth encoding layers are respectively connected to 1×1 convolutions to adjust the number of channels; and are spliced ​​with the corresponding decoding layers.

[0051] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the above-mentioned method for analyzing the dynamic recrystallization structure morphology of titanium alloy based on machine learning.

[0052] The beneficial effects of the present invention are as follows:

[0053] (1) The present invention uses a DRX segmentation model to perform pixel-level segmentation on multimodal data from EBSD and TEM. The present invention uses a primary classifier to distinguish DRX from non-DRX regions, and a secondary classifier to discriminate between DDRX, CDRX, and GDRX. The input features of the secondary classifier are the predicted probability of the primary classifier XGBoost + the first three principal components of the DRX features (explained variance ≥ 85% after PCA dimensionality reduction), which improves the accuracy of distinguishing DRX regions and types and integrates grain boundary morphology, subgrain boundary density, and orientation gradient features. The present invention reconstructs a three-dimensional model of DRX grains using FIB-SEM tomographic sequence images and verifies the two-dimensional segmentation results using the three-dimensional model. The deviation between the calculated volume fraction and the two-dimensional segmentation result can be achieved within 2%, ensuring the credibility of the analysis results. The three-dimensional model can also intuitively display the connectivity and aggregation of DRX grains, providing spatial dimension guidance for process optimization.

[0054] (2) The present invention combines EBSD orientation information, TEM subgrain boundary information, and three-dimensional tomography data to comprehensively capture the DRX microevolution mechanism and independently calculate the proportion and type of DRX in dual-phase titanium alloys. Combined with a reinforcement learning algorithm, it can automatically recommend optimal hot working parameters (such as temperature and strain rate) based on the DRX prediction results, promoting the realization of intelligent manufacturing. The present invention can automatically analyze titanium alloy microstructure images obtained by optical microscopy and electron backscatter diffraction, quickly and accurately identifying the dynamic recrystallization ratio and the type of each dynamically recrystallized grain in the image; significantly improving the automation and efficiency of dynamic recrystallization ratio and type identification, reducing the influence of human factors, and providing standardized data support for the study of the microstructure and properties of metal materials and production process control.

[0055] (3) This invention combines EBSD, TEM, and FIB-SEM techniques to obtain microstructural images of varying resolutions, encompassing complete information from the macroscopic to the microscopic level. It also performs weighted fusion of the EBSD IPF and KAM maps and histogram equalization of the TEM images. Weighted fusion and histogram equalization enhance image contrast, particularly the visibility of subgrain boundaries and dislocation structures. This invention ensures the comprehensiveness and quality of data input, providing a reliable foundation for automated analysis.

[0056] (4) The DRX segmentation model combines the channel attention mechanism (SE module) to enhance the segmentation ability of small features such as subgrain boundaries; and the loss function adopts weighted cross entropy loss, with the weight distribution of fully recrystallized: partially recrystallized: unrecrystallized = 1:2:1.5, which improves the model's adaptability to different data distributions and its detection sensitivity to small targets. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 Flowchart of the method for analyzing the dynamic recrystallization structure morphology of titanium alloy based on machine learning of the present invention;

[0058] Figure 2 Schematic diagram of the structure of the DRX segmentation model;

[0059] Figure 3 Flowchart for 3D reconstruction;

[0060] Figure 4 Flowchart for verifying 2D DRX segmentation map based on 3D model of DRX die;

[0061] Figure 5 This is a schematic diagram of the analysis system for the dynamic recrystallization structure morphology of titanium alloy based on machine learning of the present invention. DETAILED DESCRIPTION

[0062] Example 1:

[0063] The machine learning-based analysis method of dynamic recrystallization microstructure of titanium alloys combines machine learning with three-dimensional tomography to achieve automated analysis of the proportion and type of dynamic recrystallization (DRX) of titanium alloys, such as Figure 1 As shown, the following steps are included:

[0064] Step S1: Phase region (900~1100℃) and Gleeble hot compression tests were carried out in the phase region (700-900℃), with strain rates ranging from 0.001 to 10s-1 and true strains from 0.2 to 1.5. Water cooling was used (cooling rate ≥ 200℃ / s) to suppress the occurrence of static recrystallization and dynamic recovery.

[0065] Step S2: Collect multimodal data to obtain the orientation imaging IPF map, grain orientation difference KAM map, TEM image and FIB-SEM tomographic sequence image of the titanium alloy to be tested. Specifically, when performing multi-scale data acquisition, the scanning step size of EBSD is 0.1~1 , obtain orientation imaging map (IPF) and grain orientation difference map (KAM).

[0066] The bright field image resolution of TEM data is ≤2nm, which can locate subgrain boundaries and dislocation structures.

[0067] FIB-SEM tomographic sequence images, using FIB ion beam to cut along the vertical sample surface, each layer is 50nm thick, and the total thickness is ≥10 ; Each layer is polished after SEM imaging and cycled until the target thickness is 10~20 .

[0068] The present invention obtains samples through hot compression testing and uses rapid cooling to retain the high-temperature deformed structure. Subsequently, data is collected through EBSD, TEM, FIB-SEM and other means.

[0069] Step S3: extracting a two-dimensional DRX segmentation map based on the two-dimensional DRX model;

[0070] Step S31: Figure 2 As shown in the figure, a deep learning-driven DRX segmentation model is constructed, and the three-channel fusion data of the orientation imaging IPF map, the grain orientation difference KAM map and the TEM image are input into the DRX segmentation model to output the DRX probability map.

[0071] Specifically, when constructing the model framework, the U-Net model was improved, the encoder adopted the ResNet-50 pre-trained weights, and the decoder integrated the channel attention mechanism (SE module) to improve the segmentation accuracy of small features (such as subgrain boundaries); the DRX segmentation model was trained and optimized using loss functions (weighted cross entropy damage) and data enhancement (random rotation, elastic deformation, Gaussian noise injection).

[0072] Specifically, the DRX segmentation model was trained. When constructing the training dataset, more than 2000 EBSD / TEM images were annotated with labels including “completely recrystallized”, “partially recrystallized”, “unrecrystallized” and “DDRX / CDRX / GDRX” types.

[0073] The DRX segmentation model is mainly used to segment microstructure images, mark areas such as "completely recrystallized", "partially recrystallized", and "non-recrystallized", and provide a basis for subsequent extraction of DRX features.

[0074] Step S32: Extract DRX features based on the DRX probability map; specifically, extract features such as grain boundary bow height (OpenCV contour convex hull calculation), grain aspect ratio (EBSD grain ellipse fitting), subgrain boundary density (TEM image processing), dislocation wall spacing (FFT frequency domain analysis), local standard deviation of the KAM map, and cumulative distribution of orientation differences as DRX features.

[0075] DRX features include, but are not limited to, grain boundary arch height, grain aspect ratio, subgrain boundary density, dislocation wall spacing, local standard deviation of the KAM plot, and cumulative distribution of misorientation. DRX features are not direct outputs of the DRX segmentation model, but are physical features further extracted based on the segmentation results.

[0076] Step S33: Input the DRX features into a first-level classifier, use the first-level classifier (XGBoost) to distinguish DRX and non-DRX areas, and output the predicted probability;

[0077] Step S34: Perform PCA dimensionality reduction on the output DRX features, retaining the first three principal component features. The predicted probabilities are concatenated with the three principal component features and fed into a secondary classifier. The secondary classifier (SVM) uses the support vector machine kernel function to distinguish DDRX, CDRX, and GDRX types, and cross-validates its accuracy. Combining the predicted probabilities from the primary classifier (XGBoost) with the PCA-derived features improves the accuracy of the secondary classifier (SVM) in distinguishing DDRX, CDRX, and GDRX types.

[0078] Step S4: Based on the FIB-SEM tomographic sequence images, SIFT feature matching is used to align the interlayer images, and the three-dimensional model of the DRX grain is reconstructed based on threshold segmentation and watershed algorithm;

[0079] Step S5: Calculate the volume proportion of the DRX grains in the 3D model and evaluate the correlation coefficient r between the 3D model and the 2D DRX model. If r ≥ 0.97, the spatial aggregation of the 3D visualized DRX grains is determined.

[0080] Preferably, in step S1, Phase region (900~1100℃) and The Gleeble hot compression test is carried out in the phase region (700-900℃), with a strain rate ranging from 0.001 to 10s-1 and a true strain of 0.2 to 1.5. The deformation temperature of the phase region is preferably 900-1000°C to ensure complete recrystallization and avoid overheating of the grains; The deformation temperature of the phase region is preferably 750-880°C to balance Phase stability and DRX driving force. Preferably, in the Gleeble hot compression test, the low strain rate is preferably 0.001-0.1 (promote DDRX grain boundary bowing); high strain rate is preferably 1-10 (GDRX-induced shear thinning).

[0081] Preferably, in step S1, for For samples thermally deformed in the two-phase region, water cooling is preferred to preserve the high-temperature deformed structure.

[0082] Preferably, in step S31, the preferred configuration of the U-Net model is improved, and the SE module compression ratio is preferably 16 (the number of channels is compressed from 64 to 4), balancing the computational efficiency and the feature enhancement effect.

[0083] The EBSD IPF map (HCP structure color coding) and KAM map (grayscale gradient) are preferably fused at a weighting of 1:1:1:0.8. In the input data of the DRX segmentation model, the IPF map (inverse pole figure, RGB three channels) is weighted at 1:1:1 for crystal orientation information, and the KAM map (nuclear average misorientation, single channel) is weighted at 0.8 for local strain information. Purpose: Combining crystal orientation and local strain enhances the physical basis for DRX region detection.

[0084] TEM bright field images are preferably pre-processed by histogram equalization to enhance the subgrain boundary contrast.

[0085] The loss function of the DRX segmentation model is preferably weighted cross entropy loss, with the weight distribution being fully recrystallized: partially recrystallized: unrecrystallized = 1:2:1.5, to improve the detection sensitivity of small targets (such as partially recrystallized grains). The loss function of the DRX segmentation model is:

[0086] ;

[0087] in: , , are the weight coefficients corresponding to complete recrystallization, partial recrystallization, and no recrystallization, and are 1, 2, and 1.5 respectively.

[0088] , , are the probabilities of the predictions output by the DRX segmentation model being fully recrystallized, partially recrystallized, and not recrystallized, respectively;

[0089] , , are the true labels corresponding to complete recrystallization, partial recrystallization and no recrystallization, respectively.

[0090] Preferably, in step S32, when the grain boundary bow height is ≥0.2 Or grain boundary curvature ≥ 0.25 When , it is determined to be a DDRX chip;

[0091] When the subgrain boundary misorientation gradually increases from <2° to >15° or the local gradient standard deviation of the KAM diagram is <0.8°, it is identified as a CDRX grain;

[0092] When the grain aspect ratio is ≥5:1 or the proportion of sharp grain boundaries is ≥60%, it is identified as GDRX grain.

[0093] Preferably, in step S33, the maximum tree depth of the first-level classifier XGBoost is preferably max_depth=5 to prevent overfitting; the learning rate is preferably learning_rate=0.1, and the subsampling rate subsample=0.8;

[0094] In step S34, the kernel function of the secondary classifier SVM is preferably the RBF kernel, and the regularization parameter C=10 to balance the classification interval and the misclassification penalty to avoid overfitting. = 0.1 to control the kernel width and adapt to the distribution complexity of the 6-dimensional features. The input features are preferably the predicted probabilities of the first-level classifier XGBoost plus the first three principal components of the DRX features after PCA dimensionality reduction. The feature transfer mechanism is as follows: the predicted probabilities of the first-level classifier XGBoost provide the class confidence, and the features after PCA dimensionality reduction provide the physical model supplement. For example, if the first-level classifier XGBoost predicts a region as "partially recrystallized" (P_partial = 0.8), but the PCA features show a high subgrain boundary density, the SVM is more likely to classify it as CDRX.

[0095] First-level classifier XGBoost: adjust the weight of the minority class (such as partially recrystallized area) through scale_pos_weight.

[0096] Secondary classifier SVM: Class weight parameter class_weight='balanced', automatically adjust the loss function.

[0097] Preferably, PCA dimensionality reduction refers to the dimensionality reduction of DRX features through principal component analysis (PCA); this primarily involves converting high-dimensional data into low-dimensional data through linear transformations while retaining the data's essential information. PCA dimensionality reduction can reduce redundant information between features, lower model complexity, and avoid overfitting. Specifically, principal component analysis (PCA) is performed on the original 6-dimensional features, retaining the first three principal component features (cumulative variance ≥ 85%). Physical significance: Extracts the most discriminative global patterns (such as the principal direction of the strain gradient). The 3-dimensional probability output by the first-level classifier, XGBoost, is concatenated with the 3-dimensional features after PCA dimensionality reduction to form a 6-dimensional fused feature. Advantage: Combines model prediction confidence with the original physical features to enhance discriminative information.

[0098] Preferably, in step S34, after PCA dimensionality reduction is performed on the DRX features, the top three principal components with a cumulative explained variance of ≥ 85% are selected. These principal components are linear combinations of the DRX features and can maximize the preservation of DRX feature information while reducing the data dimension and improving the efficiency and performance of the model.

[0099] Preferably, in step S2, the slice thickness of the FIB-SEM is preferably 50 nm to balance the resolution (subgrain boundaries are visible) and the total imaging time.

[0100] Preferably, in step S4, the segmentation algorithm is preferably an adaptive threshold segmentation algorithm (Otsu algorithm), combined with morphological closing operation to reduce pore misjudgment; DRX grains preferably have an equivalent diameter ≥ 0.5 , eliminating noise interference.

[0101] In step S4, the FIB-SEM tomographic image needs to be segmented to reconstruct the 3D model of the DRX grain. However, due to the complexity of the microstructure image, the following problems may occur:

[0102] Pore ​​interference: There may be pores or noise in the image, which may lead to misjudgment in the segmentation results;

[0103] Grain boundary blur: The grain boundary may not be clear enough, affecting the accuracy of segmentation;

[0104] Non-uniformity: The grayscale distribution of the image is uneven, resulting in poor segmentation effect using a single threshold.

[0105] Therefore, in order to solve these problems, it is preferred to use a solution combining adaptive threshold segmentation (Otsu algorithm) with morphological closing operation.

[0106] Specifically, the principle of the adaptive threshold segmentation algorithm (Otsu algorithm) is as follows:

[0107] The Otsu algorithm is an adaptive threshold segmentation method based on the image grayscale histogram. It determines the optimal segmentation threshold by maximizing the inter-class variance. This method can automatically adapt to the grayscale distribution of the image and is suitable for images with uneven grayscale.

[0108] Implementation steps: Calculate the grayscale histogram of the image; traverse all possible thresholds and calculate the inter-class variance under each threshold; select the threshold that maximizes the inter-class variance as the optimal segmentation threshold; perform binary segmentation on the image based on the threshold, dividing the image into foreground (DRX grains) and background (non-DRX area).

[0109] Advantages: Strong adaptability, suitable for images with uneven grayscale distribution; no need to manually set the threshold, high degree of automation.

[0110] Specifically, the principle of morphological closing operation is as follows:

[0111] The morphological closing operation is to perform a dilation operation first, followed by an erosion operation. The dilation operation can fill small holes and disconnected areas, while the erosion operation can restore the basic shape of the target area.

[0112] Implementation steps: Define a structural element (such as a circular or square kernel) for dilation and erosion operations. Dilate the binary image to fill small holes and connect broken grain boundaries. Erode the dilated image to restore the basic shape of the grains.

[0113] Advantages: It can effectively fill pores and reduce misjudgment. It can also smooth grain boundaries and improve segmentation accuracy.

[0114] Specifically, the combination of Otsu algorithm and morphological closing operation is as follows:

[0115] Preprocess FIB-SEM tomographic images (e.g., denoising, contrast enhancement);

[0116] Use Otsu algorithm for adaptive threshold segmentation to obtain a binary image;

[0117] Perform morphological closing operation on the binary image to fill pores and smooth grain boundaries;

[0118] A 3D model of the DRX grains is reconstructed based on the segmentation results. The segmented image shows clear DRX grain boundaries, with significantly reduced pores and noise. The reconstructed 3D model is more accurate, providing reliable data for subsequent volume fraction calculations and spatial clustering analysis. Preferably, in the morphological closing operation, the size of the structural element is selected based on the average grain size and pore size, typically 3×3 or 5×5 pixels. The number of iterations is selected based on the image characteristics, typically 1–2.

[0119] Advantages: strong adaptability, suitable for images with different grayscale distributions; can effectively reduce the interference of pores and noise, and improve segmentation accuracy; high degree of automation, reducing human intervention.

[0120] Preferably, step S4 includes the following steps:

[0121] Step S41: data preparation;

[0122] Input data: FIB-SEM tomographic sequence images: cut along the vertical sample surface (ND direction), each layer thickness 50nm, total thickness ≥10 Inter-layer alignment data: Sub-pixel alignment (error < 50 nm) is achieved through SIFT feature matching.

[0123] Step S42: Image preprocessing: Histogram equalization: Enhance the contrast of inter-layer images and reduce the impact of cutting marks.

[0124] Step S43: Adaptive threshold segmentation (Otsu algorithm): Separate DRX grains from the background (pores, inclusions). Morphological closing operation: Fill small pores and smooth grain boundaries.

[0125] Step S44: grain segmentation and marking;

[0126] Watershed algorithm: Based on distance transformation and local gradient, it can segment the adhered grains. Filtering condition: Only retain the grains with equivalent diameter ≥ 0.5 of grains (excluding noise interference).

[0127] Step S45: generating a three-dimensional model;

[0128] Marching Cubes algorithm: convert the segmentation results into a 3D mesh model (isosurface extraction). Voxelization: based on the layer thickness (50 nm) and pixel resolution (such as 0.1 / pixel), calculate the number of voxels in each DRX grain.

[0129] Preferably, step S5 includes the following steps:

[0130] Step S51: data pairing;

[0131] 2D data source: 2D DRX segmentation map (complete / partial / unrecrystallized regions). Corresponding 3D cross section: A cross section (RD-TD plane) extracted from the 3D model at the same location as the 2D image.

[0132] Step S52: 2D-3D alignment verification;

[0133] Spatial registration: SIFT feature matching: Extract key points (such as grain boundary intersections and pore positions) from the 2D image and the 3D cross-section, and calculate the affine transformation matrix. Error control: Root mean square error (RMSE) after registration < 0.1 .

[0134] Step S53: regional consistency check;

[0135] Overlap ratio calculation: Compare the overlapping area ratio of the DRX region in the 2D prediction and the 3D cross-section, which must be ≥95%.

[0136] Step S54: Data extraction: Take N cross sections at equal intervals along the ND direction in the three-dimensional model (e.g., every 2 Take a cross section (5 in total). Calculate the DRX area ratio for each cross section (same as the 2D prediction method).

[0137] Step S55: Correlation coefficient calculation;

[0138] The correlation coefficient r is:

[0139] ;

[0140] in: For the The DRX area ratio of each three-dimensional cross section;

[0141] is the average of the DRX area proportions of N three-dimensional sections;

[0142] For the DRX area ratio of a two-dimensional DRX segmentation map;

[0143] is the average value of the DRX area proportions of N two-dimensional DRX partition maps.

[0144] Judgment criteria: r≥0.97.

[0145] Bland-Altman analysis: The mean deviation and 95% limits of agreement (LoA) of the DRX area ratios between 2D and 3D images were calculated. Acceptable range: mean deviation ≤ 2%, LoA within ±5%.

[0146] Deviation exceeding limit trigger condition: If , or r < 0.97, automatically mark abnormal sections. Manual review process: Check whether the segmentation threshold of the 3D reconstruction is appropriate for the current section. Verify whether the training data of the 2D model covers the characteristics of the region. Re-register or locally optimize the segmentation parameters and recalculate.

[0147] Example 2:

[0148] An analysis method of dynamic recrystallization microstructure of titanium alloy based on machine learning is used to realize the dynamic recrystallization microstructure of Ti-6Al-4V alloy. DRX analysis of the phase region automatically identifies the proportion and type of dynamically recrystallized grains, including the following steps:

[0149] 1) The titanium alloy to be tested was subjected to Gleeble hot compression test, where the deformation temperature was 900℃ and the strain rate was 0.1 , the true strain is 0.8; then, water-cooled quenching;

[0150] 2) Set the EBSD scanning step size to 0.3 , the resolution of TEM images is 1.5 nm, and the thickness of FIB-SEM tomography is 30 , the number of layers is 300; obtain the orientation imaging IPF map, grain orientation difference KAM map, TEM image and FIB-SEM tomography sequence image of the titanium alloy to be tested;

[0151] 3) A DRX segmentation model based on an improved U-Net is used, in which the encoder introduces a channel attention mechanism (SE module). The input is a three-channel fusion data of the orientation imaging IPF map, the grain orientation difference KAM map, and the TEM image, and the output is a pixel-level DRX probability map;

[0152] 4) A stacked classifier based on the first-level classifier XGBoost and the second-level classifier SVM, with input features including grain boundary curvature, subgrain density, orientation gradient and grain aspect ratio, outputs the classification results of DDRX, CDRX and GDRX, and obtains a two-dimensional DRX segmentation map.

[0153] 5) Use Avizo software to perform 3D reconstruction of FIB-SEM tomographic sequence images, calculate the DRX volume fraction, and compare the 2D segmentation results with the 3D volume fraction.

[0154] Example 3:

[0155] An analysis method of dynamic recrystallization microstructure of titanium alloy based on machine learning is used to realize the dynamic recrystallization microstructure of Ti-6Al-4V alloy. DRX analysis of the phase region automatically identifies the proportion and type of dynamically recrystallized grains, including the following steps:

[0156] 1) The titanium alloy to be tested is subjected to a Gleeble hot compression test, wherein the deformation temperature is 950°C, the strain rate is 1s-1, and the true strain is 0.9; then, the alloy is water-quenched;

[0157] 2) Set the EBSD scanning step to 0.5 , the resolution of TEM images is 1.5 nm, and the thickness of FIB-SEM tomography is 40 , the number of layers is 200; obtain the orientation imaging IPF map, grain orientation difference KAM map, TEM image and FIB-SEM tomography sequence image of the titanium alloy to be tested;

[0158] 3) A DRX segmentation model based on an improved U-Net is used, in which the encoder introduces a channel attention mechanism (SE module). The input is a three-channel fusion data of the orientation imaging IPF map, the grain orientation difference KAM map, and the TEM image, and the output is a pixel-level DRX probability map;

[0159] 4) A stacked classifier based on the first-level classifier XGBoost and the second-level classifier SVM, with input features including grain boundary curvature, subgrain density, orientation gradient and grain aspect ratio, outputs the classification results of DDRX, CDRX and GDRX, and obtains a two-dimensional DRX segmentation map.

[0160] 5) Use Avizo software to perform 3D reconstruction of FIB-SEM tomographic sequence images, calculate the DRX volume fraction, and compare the 2D segmentation results with the 3D volume fraction.

[0161] Comparative Example 1:

[0162] An analysis method of dynamic recrystallization microstructure of titanium alloy based on machine learning is used to realize the dynamic recrystallization microstructure of Ti-6Al-4V alloy. DRX analysis of the phase region automatically identifies the proportion and type of dynamically recrystallized grains, including the following steps:

[0163] 1) The titanium alloy to be tested was subjected to Gleeble hot compression test, where the deformation temperature was 920℃ and the strain rate was 0.2 , the true strain is 0.9; then, water-cooled quenching;

[0164] 2) Set the EBSD scanning step size to 0.4 , the resolution of TEM images is 1.8 nm, and the thickness of FIB-SEM tomography is 35 , the number of layers is 320; obtain the orientation imaging IPF map, grain orientation difference KAM map, TEM image and FIB-SEM tomography sequence image of the titanium alloy to be tested;

[0165] 3) Using the traditional U-Net-based DRX segmentation model, the input is the three-channel fusion data of the orientation imaging IPF map, the grain orientation difference KAM map and the TEM image, and the output is the DRX probability map;

[0166] 4) A stacked classifier based on the first-level classifier XGBoost and the second-level classifier SVM, with input features including grain boundary curvature, subgrain density, orientation gradient and grain aspect ratio, outputs the classification results of DDRX, CDRX and GDRX, and obtains a two-dimensional DRX segmentation map.

[0167] 5) Use Avizo software to perform 3D reconstruction of FIB-SEM tomographic sequence images, calculate the DRX volume fraction, and compare the 2D segmentation results with the 3D volume fraction.

[0168] Comparative Example 2:

[0169] An analysis method of dynamic recrystallization microstructure of titanium alloy based on machine learning is used to realize the dynamic recrystallization microstructure of Ti-6Al-4V alloy. DRX analysis of the phase region automatically identifies the proportion and type of dynamically recrystallized grains, including the following steps:

[0170] 1) The titanium alloy to be tested was subjected to Gleeble hot compression test, where the deformation temperature was 930℃ and the strain rate was 0.3 , the true strain becomes 1; then, water-cooled quenching;

[0171] 2) Set the EBSD scanning step to 0.5 , the resolution of TEM images is 2 nm, and the thickness of FIB-SEM tomography is 33 , the number of layers is 250; obtain the orientation imaging IPF map, grain orientation difference KAM map, TEM image and FIB-SEM tomography sequence image of the titanium alloy to be tested;

[0172] 3) A DRX segmentation model based on an improved U-Net is used, in which the encoder introduces a channel attention mechanism (SE module). The input is a three-channel fusion data of the orientation imaging IPF map, the grain orientation difference KAM map, and the TEM image, and the output is a DRX probability map;

[0173] 4) Based on a single first-level classifier XGBoost, the input features include grain boundary curvature, subgrain density, orientation gradient and grain aspect ratio, and the classification results of DDRX, CDRX and GDRX are output to obtain a two-dimensional DRX segmentation map.

[0174] 5) Use Avizo software to perform 3D reconstruction of FIB-SEM tomographic sequence images, calculate the DRX volume fraction, and compare the 2D segmentation results with the 3D volume fraction.

[0175] Comparative Example 3:

[0176] An analysis method of dynamic recrystallization microstructure of titanium alloy based on machine learning is used to realize the dynamic recrystallization microstructure of Ti-6Al-4V alloy. DRX analysis of the phase region automatically identifies the proportion and type of dynamically recrystallized grains, including the following steps:

[0177] 1) The titanium alloy to be tested was subjected to Gleeble hot compression test, where the deformation temperature was 930℃ and the strain rate was 0.3 , the true strain becomes 1; then, water-cooled quenching;

[0178] 2) Set the EBSD scanning step size to 0.4 , the resolution of TEM images is 3 nm, and the thickness of FIB-SEM tomography is 40 , the number of layers is 200; obtain the orientation imaging IPF map, grain orientation difference KAM map, TEM image and FIB-SEM tomography sequence image of the titanium alloy to be tested;

[0179] 3) A DRX segmentation model based on an improved U-Net is used, in which the encoder introduces a channel attention mechanism (SE module). The input is a three-channel fusion data of the orientation imaging IPF map, the grain orientation difference KAM map, and the TEM image, and the output is a DRX probability map;

[0180] 4) Based on a single classifier SVM, the input features include grain boundary curvature, subgrain density, orientation gradient and grain aspect ratio, and the classification results of DDRX, CDRX and GDRX are output to obtain a two-dimensional DRX segmentation map.

[0181] As shown in Table 1, the encoders of the DRX segmentation models in Examples 2 and 3 both incorporate a channel attention mechanism (SE module) and employ a stacked classifier consisting of an XGBoost primary classifier and an SVM secondary classifier. Compared to Comparative Examples 1-3, the final DRX recognition accuracy in Examples 2 and 3 exceeds 90%, significantly improving the recognition readiness rate.

[0182] Table 1

[0183]

[0184] Example 4:

[0185] The analysis system of dynamic recrystallization microstructure of titanium alloy based on machine learning includes a multimodal data acquisition module, a two-dimensional model detection module, a three-dimensional model construction module and a verification module.

[0186] The multimodal data acquisition module is used to obtain the orientation imaging IPF map, grain orientation difference KAM map, TEM image and FIB-SEM tomography sequence image of the titanium alloy to be tested;

[0187] The two-dimensional model detection module is used to output a two-dimensional DRX segmentation map based on the two-dimensional DRX model, and includes a DRX segmentation module, a DRX feature extraction module, a primary classifier, a PAC dimension reduction module, a feature fusion module, and a secondary classifier, which are arranged in order from the front to the back; the DRX feature extraction module is respectively connected to the primary classifier and the PAC dimension reduction module, and the primary classifier and the PAC dimension reduction module are respectively connected to the feature fusion module;

[0188] The three-dimensional model building module is used to reconstruct the three-dimensional model of the DRX grain based on the FIB-SEM tomographic sequence images and using the threshold segmentation and watershed algorithm;

[0189] The verification module is used to perform consistency verification on the two-dimensional DRX segmentation map based on the three-dimensional model of the DRX grain; and evaluate the correlation coefficient between the three-dimensional model and the two-dimensional DRX model.

[0190] Preferably, if Figure 2 As shown in the figure, the DRX segmentation model adopts a symmetrical encoder-decoder structure. The encoder extracts features through multi-level downsampling, and the decoder restores details through upsampling and skip connections. The core improvements include ResNet-50 backbone, SE attention mechanism and multimodal input fusion.

[0191] Preferably, a pre-trained ResNet-50 is used as the encoder, removing the fully connected layers and retaining the convolutional layers. Advantages: Accelerated convergence: Utilizing ImageNet pre-trained weight initialization allows for rapid adaptation to microstructural features. Deep feature extraction: Residual connections mitigate vanishing gradients and support deeper network structures. Multi-scale capture: Four encoding layers progressively extract local to global features (from grain boundary details to DRX regional distribution).

[0192] Preferably, an SE attention module (SE module) is embedded after each upsampling layer in the decoder. Advantages: Enhanced small object detection: Boosts channel weights for small structures such as subgrain boundaries (2-10 nm wide) and dislocation walls. Noise suppression: Reduces weights for background or blurred regions (such as dislocation noise in TEM images). Multimodal optimization: Differentially weights IPF (orientation) and KAM (strain) features. For example, KAM channel weights emphasize local strain information.

[0193] The specific contents are as follows:

[0194] (1) Input layer: multimodal data fusion;

[0195] Input data: IPF image (inverse pole figure): crystal orientation information, with the RGB channels weighted at 1:1:1. KAM image (nuclear average misorientation, single channel): local strain information, with a weight of 0.8. Purpose: Combines crystal orientation and local strain to enhance the physical basis for DRX region detection. Specifically, the RGB of the IPF image and the grayscale of the KAM image are fused into a four-channel input with weights of 1:1:1:0.8. Advantages: Physics-driven: IPF provides crystal orientation, and KAM reflects local strain, jointly guiding DRX region identification. Weight design: The KAM weight of 0.8 balances the amount of information in its single channel to avoid dominance by the three channels of IPF.

[0196] (2) The encoder consists of the first to fourth encoding layers connected sequentially from top to bottom, gradually extracting the multi-scale features of the microstructure.

[0197] First encoding layer: Primary feature extraction; Architecture: 7×7 convolutional layer with 64 channels, stride 2, ReLU activation. MaxPool 2×2 with stride 2 for fast resolution compression. Output: 64 channels, resolution 1 / 4 of the input (e.g., 256×256 → 64×64). Purpose: Captures low-level features such as grain boundaries and subgrain boundary textures.

[0198] Second encoding layer: Intermediate feature abstraction; Structure: Three residual blocks (Conv2_x): Each block contains two 3×3 convolutions (256 channels) with residual connections. Downsampling: Convolutions with a stride of 2 compress the resolution to 1 / 8. Output: 256 channels, resolution 32×32. Purpose: Extracts morphological features such as grain boundary curvature and bowing height.

[0199] The third encoding layer: extracts high-level semantic information. Structure: 4 residual blocks (Conv3_x): 3×3 convolutions (512 channels) with residual connections. Downsampling: Convolutions with stride 2 compress the resolution to 1 / 16. Output: 512 channels, resolution 16×16. Purpose: Learns orientation difference distribution (KAM gradient) and DRX regional distribution trends.

[0200] Fourth encoding layer: Global context modeling; Structure: 6 residual blocks (Conv4_x): 3×3 convolution (1024 channels), residual connections. Downsampling: Convolution with stride 2 compresses the resolution to 1 / 32. Output: 1024 channels, resolution 8×8. Purpose: Encodes global semantics (such as the boundary between DRX and non-DRX areas).

[0201] (3) The decoder includes a fourth decoding layer to a first decoding layer that are sequentially skip-connected from top to bottom. The decoder restores the resolution through upsampling and skip connections. The core improvement is the integration of SE attention modules (SE modules).

[0202] Fourth decoding layer, architecture: Upsampling layer: transposed convolution (2×2, stride 2), channel count from 1024 to 512. Sequencing module: recalibrates the channel weights of the 512-channel features. Skip connection: concatenates with the output of the third encoding layer (512 channels). Convolution layer: 3×3 convolution to fuse features. Output: 512 channels, resolution 16×16.

[0203] Third decoding layer, architecture: Upsampling layer: 512 → 256 channels, 32×32 resolution. SE module: Enhances key channels (such as subgrain boundary correlation). Skip connection: Fusion of the output of the second encoding layer (256 channels). Output: 256 channels, 32×32 resolution.

[0204] Second decoding layer: Upsampling layer: 256 → 128 channels, 64×64 resolution. SE module: Optimizes grain boundary curvature features. Skip connection: Fusion of the output of the first encoding layer (64 channels). Output: 128 channels, 64×64 resolution.

[0205] First decoding layer: Upsampling layer: 128 → 64 channels, 128×128 resolution. SE module: Final channel weight adjustment. Convolutional layer: Outputs four-category segmentation results (complete / partial / unrecrystallized / background). Output: 4-channel probability map, with the same resolution as the input.

[0206] Preferably, the SE attention module (SE module) is located after the upsampling of each decoding layer and before the skip connection. It includes:

[0207] Global pooling layer Squeeze: Global average pooling (GAP) compresses spatial information and generates a 1×1×C vector.

[0208] Fully connected layer FC1: compresses channels to C / r (r=16, such as C=512→32).

[0209] Activation function layer ReLU: introduces nonlinearity.

[0210] Fully connected layer FC2: restore the channel to C (32→512).

[0211] Sigmoid: Generates channel weights from 0 to 1.

[0212] Rescaling: Weights are multiplied by DRX features channel by channel to enhance key features. This function dynamically improves the channel response of microstructures such as subgrain boundaries and dislocation walls.

[0213] The output of each encoder layer undergoes a 1×1 convolution to adjust the number of channels and is then concatenated with the corresponding decoder layer. Purpose: Fusion of low-level details (such as grain boundary locations) with high-level semantics (such as DRX categories) improves segmentation boundary accuracy. Advantages: Detail recovery: Low-level features in the first encoding layer (such as grain boundary locations) directly assist in the final segmentation. Channel alignment: Avoids channel mismatches between the encoder and decoder (e.g., 1024 → 512).

[0214] Preferably, the input of the first-level classifier XGBoost is six types of features: grain boundary bow height, grain aspect ratio, subgrain boundary density, KAM local gradient standard deviation, orientation difference cumulative distribution, and dislocation wall spacing; specifically, the six input features need to be feature normalized (Z-score normalization) to ensure dimensional consistency.

[0215] The primary classifier, XGBoost, is used to initially distinguish DRX regions (complete, partial, or non-recrystallized). Key hyperparameters designed include max_depth = 5, which limits the tree depth to prevent overfitting. learning_rate = 0.1, which incorporates an early stopping strategy to optimize convergence speed. subsample = 0.8, which randomly samples 80% of the data to train a single tree, enhancing generalization. Probability output: Softmax is used to output three-category probabilities (summing to 1), preserving uncertainty between categories.

[0216] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any form. Any simple modification or equivalent change made to the above embodiment based on the technical essence of the present invention shall fall within the scope of protection of the present invention.

Claims

1. A method for analyzing the dynamic recrystallization structure of titanium alloy based on machine learning, characterized in that: The following steps are involved: Step S1: performing a Gleeble hot compression test on the titanium alloy to be tested and then cooling the alloy; Step S2: Acquiring multimodal data: obtaining orientation imaging IPF images, grain orientation difference KAM images, TEM images, and FIB-SEM tomographic sequence images of the titanium alloy to be tested; Step S3: extracting a two-dimensional DRX segmentation map based on the two-dimensional DRX model; Step S31: inputting the three-channel fusion data of the orientation imaging IPF map, the grain orientation difference KAM map and the TEM image into the DRX segmentation model, marking the areas of complete recrystallization, partial recrystallization and non-recrystallization, and outputting the DRX probability map; Step S32: extracting DRX features based on the DRX probability map; Step S33: Input the DRX features into a first-level classifier to distinguish DRX areas from non-DRX areas and output predicted probabilities; perform PCA dimensionality reduction on the output DRX features and retain the first three principal component features; Step S34: The predicted probability is concatenated with the three principal component features and input into a secondary classifier to distinguish DDRX, CDRX, and GDRX types, and output a two-dimensional DRX segmentation map; Step S4: reconstructing a three-dimensional model of the DRX grain based on the FIB-SEM tomographic sequence images; Step S5: Based on the three-dimensional model of the DRX die, perform consistency check on the two-dimensional DRX segmentation map; evaluate the correlation coefficient between the three-dimensional model and the two-dimensional DRX model; If the consistency check is successful and the correlation coefficient meets the set conditions, the spatial clustering of the DRX grains is visualized in three dimensions.

2. The method for analyzing the dynamic recrystallization structure of titanium alloy based on machine learning according to claim 1, characterized in that: For titanium alloy to be tested Phase region and The phase region is subjected to a Gleeble hot compression test and a rapid cooling treatment is performed, and the cooling rate is greater than or equal to 200°C / s; The deformation temperature of the phase region is 900-1000°C; The deformation temperature of the phase region is 750-880℃; the low strain rate of the Gleeble hot compression test is 0.001-0.1 , high strain rate of 1-10 .

3. The method for analyzing the dynamic recrystallization structure of titanium alloy based on machine learning according to claim 1, characterized in that: In the step S31, first, data preprocessing is performed: the orientation imaging IPF map and the grain orientation difference KAM map are weightedly fused, and the fusion weight of the RGB three channels of the orientation imaging IPF map is 1:1:1, and the fusion weight of the grain orientation difference KAM map is 0.8; the TEM image is histogram equalized.

4. The method for analyzing the dynamic recrystallization structure of titanium alloy based on machine learning according to claim 3, characterized in that: In step S31, the loss function of the DRX segmentation model is a weighted cross entropy loss of complete recrystallization, partial recrystallization and non-recrystallization, and the corresponding weighted weight ratio is 1:2:1.

5.

5. The method for analyzing the dynamic recrystallization structure of titanium alloy based on machine learning according to claim 1, characterized in that: The step S4 comprises the following steps: Step S41: inputting FIB-SEM tomographic sequence images and aligning inter-layer images using SIFT feature matching; Step S42: pre-processing the image; Step S43: Adopting an adaptive threshold segmentation algorithm to perform adaptive threshold segmentation to obtain a binary image, separating the DRX grains from the background; performing a morphological closing operation on the binary image to fill the pores and smooth the grain boundaries; Step S44: using a watershed algorithm based on distance transformation and local gradient to segment the adhered grains; Step S45: using the Marching Cubes algorithm, converting the segmentation result of step S44 into a three-dimensional mesh model.

6. The method for analyzing the dynamic recrystallization structure of titanium alloy based on machine learning according to claim 5, characterized in that: The step S5 comprises the following steps: Step S51: obtaining a two-dimensional DRX segmentation map, and extracting a three-dimensional cross section at the same position from the three-dimensional model of the DRX die; Step S52: extract key points from the two-dimensional DRX segmentation map and the three-dimensional cross section, calculate the affine transformation matrix, and perform spatial registration; Step S53: performing consistency check on the 2D DRX segmentation map: calculating the overlap ratio between the 2D DRX segmentation map and the DRX region in the 3D cross section. If the overlap ratio is greater than a set threshold, the region consistency meets the standard. Step S54: In the three-dimensional model of the DRX grain, N three-dimensional cross sections are taken at equal intervals along the ND direction, and the DRX area ratio of each three-dimensional cross section is calculated; Step S55: The calculation formula of the correlation coefficient r is: ； in: For the The DRX area ratio of each three-dimensional cross section; is the average of the DRX area proportions of N three-dimensional sections; For the DRX area ratio of a two-dimensional DRX segmentation map; is the average value of the DRX area proportions of N two-dimensional DRX partition maps.

7. The method for analyzing the dynamic recrystallization structure of titanium alloy based on machine learning according to claim 6, characterized in that: like , or r<0.97, the abnormal section is automatically marked and the review process is carried out; check whether the segmentation threshold of the three-dimensional model of the DRX grain is adapted to the current section. If not, repeat steps S51 and S55 after re-aligning or locally optimizing the segmentation parameters; verify whether the training data of the two-dimensional DRX model covers the characteristics of the region.

8. A machine learning-based analysis system for the dynamic recrystallization microstructure of titanium alloys, used to implement the method according to any one of claims 1 to 7, characterized in that: It includes multimodal data acquisition module, two-dimensional model detection module, three-dimensional model construction module and verification module; The multimodal data acquisition module is used to obtain the orientation imaging IPF map, grain orientation difference KAM map, TEM image and FIB-SEM tomography sequence image of the titanium alloy to be tested; The two-dimensional model detection module is used to output a two-dimensional DRX segmentation map based on the two-dimensional DRX model, and includes a DRX segmentation module, a DRX feature extraction module, a primary classifier, a PAC dimension reduction module, a feature fusion module, and a secondary classifier, which are arranged in order from the front to the back; the DRX feature extraction module is respectively connected to the primary classifier and the PAC dimension reduction module, and the primary classifier and the PAC dimension reduction module are respectively connected to the feature fusion module; The three-dimensional model building module is used to reconstruct the three-dimensional model of the DRX grain based on the FIB-SEM tomographic sequence images and using the threshold segmentation and watershed algorithm; The verification module is used to perform consistency verification on the two-dimensional DRX segmentation map based on the three-dimensional model of the DRX grain; and evaluate the correlation coefficient between the three-dimensional model and the two-dimensional DRX model.

9. The titanium alloy dynamic recrystallization structure morphology analysis system based on machine learning according to claim 8, characterized in that: The DRX segmentation model includes an encoder and a decoder, wherein the encoder is used to extract multi-scale features of microstructures through multi-level downsampling, and the decoder is used to restore details through upsampling and skip connections; The encoder includes a first encoding layer to a fourth encoding layer connected sequentially from top to bottom, wherein the first encoding layer includes a 7×7 convolution layer, a ReLU activation function, and a 2×2 maximum pooling layer arranged sequentially from front to back, for capturing low-level features of grain boundaries and subgrain boundary textures; The second encoding layer includes three residual blocks, a ReLU activation function, and a 2×2 maximum pooling layer, which are arranged in sequence from front to back to extract the morphological features of grain boundary curvature and bow height; The third coding layer includes four residual blocks, a ReLU activation function, and a 2×2 maximum pooling layer, which are arranged in sequence from front to back, and is used to learn the orientation difference distribution and the DRX area distribution trend; the fourth coding layer includes six residual blocks, a ReLU activation function, and a 2×2 maximum pooling layer, which are arranged in sequence from front to back, and is used to encode global semantics; The decoder includes a fourth decoding layer to a first decoding layer sequentially jump-connected from top to bottom, and the fourth decoding layer to the second decoding layer respectively include an upsampling layer and an SE attention module sequentially arranged from front to back; The outputs of the first to fourth encoding layers are respectively connected to 1×1 convolutions to adjust the number of channels; and are spliced ​​with the corresponding decoding layers.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images