Intelligent identification and modeling method for prefabricated body in aviation composite special-shaped component

Abstract

The invention discloses an intelligent identification and modeling method for a prefabricated body in an aviation composite material special-shaped component, and belongs to the technical field of aviation composite material microstructure digital reconstruction and modeling, and the method comprises the following steps: S1, data acquisition; s2, data preprocessing; s3, model training preparation; s4, performing intelligent segmentation; s5, performing three-dimensional reconstruction; s6, carrying out grid optimization; s7, carrying out topology processing; and S8, outputting the model. According to the intelligent identification and modeling method for the prefabricated body in the aviation composite material special-shaped component, aiming at the problems of topological errors and grid conformality which are likely to occur in explicit modeling, grid intersection and discontinuous areas are accurately processed in combination with a grid Boolean operation and post-processing repair algorithm, and the modeling accuracy is improved. Sealing defects such as grid gaps and missing surfaces introduced by Boolean operation are repaired, and node conformality, topological connectivity and space sealing performance of substructure grids of the prefabricated body are guaranteed; and meanwhile, manual participation is reduced, and the modeling efficiency and stability are improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of digital reconstruction and modeling of aviation composite microstructure, and particularly relates to an intelligent identification and modeling method for a preform in an aviation composite special-shaped component. BACKGROUND

[0002] Digital reconstruction and modeling of aviation composite microstructure refers to converting the internal microstructure of aviation composites, such as continuous fiber toughened ceramic matrix composites, into a digital form through technical means. Automatic identification of the microstructure of the preform in the special-shaped aviation composite component can provide a basis for subsequent three-dimensional voxel reconstruction and refined modeling.

[0003] However, the prior art has the following disadvantages: complex implementation, large amount of calculation, high requirements for programming and mathematical foundation, and high engineering application threshold; limited support for Boolean operation of complex microstructure, difficult effective extraction and processing of multi-component coplanar regions, leading to difficulty in completely closing the fiber bundle region grid, affecting the accuracy of subsequent simulation analysis; a large amount of manual intervention is required in part of the process, lacking integrated and automated processing means, reducing modeling efficiency and stability.

[0004] Therefore, a new method is urgently needed. SUMMARY

[0005] The purpose of the application is to provide an intelligent identification and modeling method for a preform in an aviation composite special-shaped component, which can accurately process grid intersections and discontinuous regions, repair closed defects such as grid gaps and missing surfaces introduced by Boolean operation, and ensure the node conformality, topological connectivity and spatial closure of the preform substructure grid, while reducing manual intervention and improving modeling efficiency and stability.

[0006] To achieve the above purpose, the application provides an intelligent identification and modeling method for a preform in an aviation composite special-shaped component, comprising:

[0007] S1, CT scanning of a target region of a special-shaped aviation composite component to obtain original CT digital slice data;

[0008] S2, preprocessing of the original CT digital slice data in S1 to obtain a preprocessed slice sequence;

[0009] S3, selecting 20% of the slice samples from the preprocessed slice sequence in S2, and manually labeling to construct a training set;

[0010] S4, the U-Net neural network framework based on Python is used to realize automatic instance segmentation of fiber bundles for all preprocessed slice sequences in S2, and a segmentation label map is obtained;

[0011] S5, three-dimensional voxel model reconstruction is performed on the segmentation label map in S4, and an STL grid file is obtained;

[0012] S6, the STL grid file in S5 is imported into the preform surface grid model, and the preform inner substructure surface grid is automatically coated and smoothly reconstructed to obtain a smooth surface grid model;

[0013] S7, based on the smooth surface grid model in S6, Boolean operation is performed on the preform substructure grid, and the grid intersection and discontinuous area are processed to obtain a topologically connected grid model;

[0014] S8, based on the topologically connected grid model in S7, the common interface area of each component generated by the Boolean operation is extracted, the spatial closure of the partition is completed, and the closed triangular surface grid domain is tetrahedral meshed to obtain a preform model.

[0015] Preferably, S2 includes the following steps:

[0016] S201, batch import original CT digital slice data, set the target region XYZ three-axis space range, and focus on the representative volume element region containing the preform through the cropping formula; the cropping formula is:

[0017] I crop (x,y)=I(x0+x,y0+y),x∈[0,w],y∈[[0,h];

[0018] Where (x0, y0) is the left upper corner coordinate of the cropping window; h is the height of the window; w is the width of the window;

[0019] The adaptive histogram equalization method is used to improve the recognition degree of the fiber bundle, and the local contrast is enhanced through histogram equalization transformation; the transformation formula is:

[0020]

[0021] Where r k is the kth gray level; n j is the number of pixels of gray level j; L is the number of gray levels; MN is the image size;

[0022] S202, Gaussian filter smoothing is used to process random noise and high-frequency artifacts, and the formula is:

[0023]

[0024] Wherein, G(i,j) is a two-dimensional Gaussian kernel function;

[0025] S203, using morphological open-close operation to eliminate small isolated noise and micropore, the formula is:

[0026]

[0027] Wherein, I is the original binary image; B is the structural element; is dilatation operation; is erosion operation.

[0028] Preferably, S4 comprises the following steps:

[0029] S401, the labeled data set in S3 and the corresponding preprocessed slice sequence in S2 are paired and normalized, and the image and label are center cropped and boundary padded to ensure that the input size is adapted to the network structure and the network input size;

[0030] S402, using U-Net encoding-decoding structure, the encoding path extracts multi-scale features through convolution and pooling, the decoding path restores features and spatial positioning through upsampling and skip connection, and outputs multi-channel prediction map, represented as:

[0031] f(x)=Decoder(Encoder(x)+Skip Connections);

[0032] Wherein, Encoder represents extracting high-dimensional semantic features, Decoder represents gradually restoring spatial structure, and Skip Connections represents guaranteeing the fusion of features and position information;

[0033] S403, using multi-class cross entropy as the main loss, and introducing Dice coefficient and Tversky loss index; the multi-class cross entropy loss is represented as:

[0034]

[0035] Wherein, y i,k is the true label of pixel i; P i,k is the prediction probability; L CE is the multi-class cross entropy loss;

[0036] The Dice coefficient is represented as:

[0037]

[0038] Wherein, y i is the true label; p i is the prediction probability; ∈ is a smoothing term;

[0039] S404. The Adam optimizer is used for training. The Dice coefficient is monitored through the validation set. The EarlyStopping and ReduceLROnPlateau mechanisms are used to prevent overfitting and adaptively adjust the learning rate. After training, the model weights are saved.

[0040] S405: Perform batch inference on all CT slices and output pixel-level segmentation label maps;

[0041] S406. Automatically predict the segmentation labels of batch CT slices and save the segmentation labels as standard image formats.

[0042] Preferably, S5 includes the following steps:

[0043] S501: Batch import 2D segmentation masks, identify materials according to RGB thresholds, and combine them into a voxel array V in three-dimensional space. (x,y,z) Each voxel is assigned a material label M. (x,y,z) in{0,1,2,3}; correspond to the background, matrix, fiber bundle A, and fiber bundle B, respectively.

[0044] S502. Isolated noise points are eliminated through three-dimensional morphological opening and closing operations, as shown below:

[0045]

[0046] Combined with connected component analysis, components with volumes smaller than the minimum connected component volume threshold S are removed. min The minute components, retaining the main structure, are represented as:

[0047]

[0048] Among them, V filtered (x,y,z) represents the output value of the voxel at position (x,y,z); S min The minimum connected component volume threshold is V(x,y,z); V(x,y,z) is the original value.

[0049] S503. Merge each material region into a single three-dimensional voxel model according to priority, and add a background boundary to the outer layer of the voxel.

[0050] S504. Traverse the voxel array. For each non-background volume (x, y, z), sequentially detect its adjacent voxels (x', y', z') in the six positive and negative axis directions. If the material numbers are different, generate an interface patch between them. The vertex coordinates of the patch are determined according to the voxel center and facing direction, and the patch is classified into the corresponding material interface.

[0051] S505. Each interface panel is generated as a quadrilateral, then split into two triangular faces; the formula for decomposing a quadrilateral into a triangle is:

[0052] Face quad =[v1,v2,v3,v4]→{[v1,v2,v3],[v1,v3,v4]};

[0053] Among them, Face quad v1, v2, v3, v4 are the four vertices of the quadrilateral face; {[v1, v2, v3], [v1, v3, v4]} are the two triangular faces obtained after splitting.

[0054] S506. After the voxel mesh coordinates are expanded by the boundary, they need to be translated as a whole to restore the original spatial position. Export each type of interface triangular mesh as a standard STL file to obtain the STL mesh file.

[0055] Preferably, S8 includes the following steps:

[0056] S801. For each fiber bundle surface mesh, using the minimum outer cube parameter of the matrix surface mesh, determine whether all vertices of its faces satisfy the following criteria:

[0057] A_min≤V≤A_max;

[0058] Where [A_min, A_max] is the minimum bounding cube of the region boundary;

[0059] If the condition is met, then the face piece is retained;

[0060] S802. For the prefabricated components after Boolean operations, extract the common interface between them to ensure the internal mesh is completely closed. Specifically:

[0061] Use a KD tree to search the vertex space and locate the common nodes between A and B; check whether the common nodes are on the same edge and identify the boundary loop edge; based on the connected component partitioning, divide A and B into several patches with the common boundary loop edge as the boundary, and extract the interface patch except for the main patch; merge all the interface patches and export them, assign the patches to the missing face component, and construct a closed internal interface mesh.

[0062] S803. For each substructure of the prefabricated body and the outer boundary of the region, perform the common interface identification and patch partitioning operation of S802, extract all common interface patches related to the boundary and merge and export them, so that each substructure of the prefabricated body is completely closed to the outside world.

[0063] S804. After all substructures within the prefabricated body are closed and topologically continuous, tetrahedral volume meshing is performed using the closed boundary as the constraint surface to generate the prefabricated body mesh model.

[0064] The application also provides an intelligent identification and modeling device for an inner preform of an aviation composite special-shaped component, comprising a data acquisition module configured to perform CT scanning on a target region of the aviation composite special-shaped component to obtain original CT digital slice data;

[0065] A data preprocessing module is connected with the data acquisition module and configured to preprocess the original CT digital slice data to obtain a slice sequence after preprocessing;

[0066] A training set construction module is connected with the data preprocessing module and configured to select 20% of slice samples from the slice sequence after preprocessing and manually label to construct a training set;

[0067] An intelligent segmentation module is connected with the training set construction module and configured to implement automatic instance segmentation of fiber bundles on all slice sequences after preprocessing based on a U-Net neural network framework built by Python to obtain a segmentation label graph;

[0068] A three-dimensional reconstruction module is connected with the intelligent segmentation module and configured to perform three-dimensional voxel model reconstruction on the segmentation label graph to obtain an STL grid file;

[0069] A grid optimization module is connected with the three-dimensional reconstruction module and configured to import the STL grid file into a preform surface grid model, automatically coat and smooth reconstruct a preform inner substructure surface grid to obtain a smooth surface grid model;

[0070] A topology processing module is connected with the grid optimization module and configured to perform Boolean operation on the preform substructure grid based on the smooth surface grid model, process grid intersection and discontinuous regions to obtain a topologically connected grid model;

[0071] A model output module is connected with the topology processing module and configured to extract each component common interface region generated after the Boolean operation based on the topologically connected grid model, complete spatial closure of the partition, and perform tetrahedral mesh division on the closed triangular surface grid domain to obtain a preform model.

[0072] Therefore, the intelligent identification and modeling method for the inner preform of the aviation composite special-shaped component has the following beneficial effects compared with the prior art:

[0073] (1) The explicit geometry modeling is adopted, the problems of complex implementation, large calculation amount and high engineering application threshold of the implicit geometry modeling are overcome, the operation is intuitive, the efficiency is high, the method is convenient for integration with subsequent simulation processes, and the difficulty of engineering application is reduced;

[0074] (2) The application improves the segmentation accuracy through the U-Net neural network, combines the Boolean operation function of the open source tool called by the Python script, overcomes the problems of the fiber bundle appearing staggered error topology and the grid continuity being destroyed due to the segmentation accuracy error of artificial intelligence in the explicit modeling, and ensures the consistency and topological connectivity of the substructure grid boundary nodes through the grid post-processing algorithm for processing the grid intersection and discontinuous area;

[0075] (3) The application integrates CT slice preprocessing, deep learning segmentation, three-dimensional reconstruction and other links, overcomes the problems of low modeling efficiency and stability caused by the lack of integrated automatic processing means in the prior art, and achieves the technical effects of reducing manual participation and improving modeling efficiency and stability.

[0076] The technical solutions of the application will be further described in detail below with reference to the drawings and embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0077] Figure 1 The flowchart of an aviation composite special-shaped component inner preform intelligent recognition and modeling method embodiment of the application;

[0078] Figure 2 The microstructure schematic diagram obtained by CT scanning of the target area of a turbine blade in an aviation composite special-shaped component inner preform intelligent recognition and modeling method embodiment of the application;

[0079] Figure 3 The training diagram input into the neural network in an aviation composite special-shaped component inner preform intelligent recognition and modeling method embodiment of the application; wherein, Figure 3 (a) in the above (a) represents an original image; wherein Figure 3 (b) in the above (b) represents a labeled image;

[0080] Figure 4 The result diagram of the model prediction segmentation after training in an aviation composite special-shaped component inner preform intelligent recognition and modeling method embodiment of the application; wherein, Figure 4 (a) in the above (a) represents an original image; wherein Figure 4 (b) in the above (b) represents a segmentation result image;

[0081] Figure 5 The preform voxel model schematic diagram of three-dimensional reconstruction in an aviation composite special-shaped component inner preform intelligent recognition and modeling method embodiment of the application;

[0082] Figure 6 The preform surface grid model schematic diagram in an aviation composite special-shaped component inner preform intelligent recognition and modeling method embodiment of the application;

[0083] Figure 7 A preform surface grid reconstruction and smoothed model schematic diagram of an embodiment of the present application is shown in the figure.

[0084] Figure 8 A preform substructure mutual intersection area schematic diagram of an embodiment of the present application is shown in the figure.

[0085] Figure 9 A preform post-processing schematic diagram after grid Boolean operation of an embodiment of the present application is shown in the figure; wherein, Figure 9 (a) in the figure represents a structure Boolean operation open surface; wherein Figure 9 (b) in the figure represents an extracted red coplanar interface.

[0086] Figure 10 A preform three-dimensional body grid model schematic diagram of an embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0087] In order to make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. All other embodiments obtained by a person of ordinary skill in the art without creative labor fall within the scope of protection of the present application. Unless otherwise defined, the technical terms or scientific terms used in the present application should be understood as the usual meanings understood by persons having ordinary skills in the art to which the present application belongs.

[0088] Embodiment one

[0089] As shown in the figure, the present application is a kind of aviation composite profiled component internal preform intelligent identification and modeling method, comprising the following steps: Figures 1-10

[0090] S1, CT scanning is performed on the target area of the profiled aviation composite material component to obtain original CT digital slice data, which contains initial image information of the preform structure;

[0091] S2, the original CT digital slice data in S1 is preprocessed using a self-developed Python script, specifically:

[0092] S201, the original CT digital slice data is imported from a specified folder in batches, the XYZ three-axis space range of the target area is set, the representative volume element (RVE) area containing the preform is focused through the cropping formula, irrelevant background is excluded, and the data processing amount is reduced; the cropping formula is:​

[0093] I crop (x,y)=I(x0+x,y0+y),x∈[0,w],y∈[[0,h];

[0094] wherein (x0,y0) is the left upper corner coordinate of the clipping window; h is the height of the window; w is the width of the window;

[0095] Adopting adaptive histogram equalization (CLAHE) and other methods to improve the fiber bundle recognition, through histogram equalization transformation to enhance the local contrast, avoiding the global over-enhanced artifacts; the transformation formula is:

[0096]

[0097] wherein r k is the kth gray level; n j is the pixel number of the gray level j; L is the gray level number; MN is the image size;

[0098] S202, adopting Gaussian filter smoothing processing random noise and high frequency artifacts, the formula is:

[0099]

[0100] wherein G(i,j) is a two-dimensional Gaussian kernel function;

[0101] S203, adopting morphological opening and closing operation to eliminate small isolated noise points and micropores, the formula is:

[0102]

[0103] wherein I is the original binary image; B is the structural element; is the dilation operation; is the erosion operation;

[0104] Through region clipping, CLAHE contrast enhancement, Gaussian filtering and morphological opening and closing operation, the saliency and segmentation accuracy of the preform structure in the CT slice are significantly enhanced, the noise and structure artifacts are effectively suppressed, and the preprocessed slice sequence is obtained;

[0105] S3, according to the equal interval along the Z axis, 20% of the slice samples are selected from the preprocessed slice sequence in S2, and the fiber bundles of different orientations are manually labeled to obtain the labeled data set;

[0106] S4, based on the U-Net neural network framework built by Python, the automatic instance segmentation of the fiber bundle is realized for all the preprocessed slice sequences in S2, specifically:

[0107] S401, normalize the pairing of the labeled data set in S3 and the corresponding preprocessed slice sequence in S2, and center crop and boundary padding are performed on the image and label to ensure that the input size is adapted to the network structure and the network input size;

[0108] S402, adopt the U-Net encoding-decoding structure, the encoding path extracts multi-scale features through convolution and pooling, the decoding path restores features and spatial positioning through upsampling and skip connection, and outputs a multi-channel prediction map, represented as:

[0109] f(x) = Decoder(Encoder(x) + Skip Connections);

[0110] Wherein, Encoder represents extracting high-dimensional semantic features, Decoder represents gradually restoring spatial structure, and Skip Connections represents ensuring the fusion of feature and position information;

[0111] S403, take multi-class cross-entropy (Sparse Categorical Crossentropy) as the main loss, and introduce Dice coefficient, Tversky loss and other indicators to improve the segmentation performance of a small number of fiber bundles; the multi-class cross-entropy loss is represented as:

[0112]

[0113] Wherein, y i,k is the true label of pixel i; P i,k is the predicted probability; L CE is the multi-class cross-entropy loss;

[0114] The Dice coefficient is represented as:

[0115]

[0116] Wherein, y i is the true label; p i is the predicted probability; ∈ is a smoothing term;

[0117] S404, use Adam optimizer for training, monitor Dice coefficient and other indicators through validation set, and use EarlyStopping and ReduceLROnPlateau mechanisms to prevent overfitting and adaptively adjust learning rate; after training, save the model weight;

[0118] S405, batch inference on full-quantity CT slices, and output pixel-level segmentation label map;

[0119] S406, automatically predict batch CT slices to obtain a segmentation label map; the segmentation label map is saved in standard picture format;

[0120] S5, independently develop a Python script to reconstruct a three-dimensional voxel model for the segmentation label map in S4, and extract the surface triangular mesh of each component, specifically:

[0121] S501, batch import 2D segmentation mask, identify materials (such as fiber bundles A, B, matrix) according to RGB threshold, and combine them into a voxel array V in three-dimensional space (x,y,z) ; each voxel is assigned a material label M (x,y,z) in{0,1,2,3}; corresponding to background, matrix, fiber bundle A (warp direction), and fiber bundle B (weft direction);

[0122] S502, eliminate isolated noise points by three-dimensional morphological opening and closing operation, represented as:

[0123]

[0124] Combine connected component analysis to remove small components with a volume less than the minimum connected component volume threshold S min , and retain the main structure, represented as:

[0125]

[0126] where V filtered (x,y,z) is the output value of the voxel at (x,y,z); S min is the minimum connected component volume threshold; and V(x,y,z) is the original value;

[0127] S503, merge each material area into a single three-dimensional voxel model according to priority, and add a background boundary to the outer layer of the voxel to avoid losing model structure during subsequent boundary surface extraction;

[0128] S504, traverse the voxel array, and for each non-background voxel (x,y,z), sequentially detect its adjacent voxels (x',y',z') in six positive and negative axis directions; if the material numbers are different, an interface patch is generated between them; the patch vertex coordinates are determined according to the voxel center and the facing direction, and the patch is classified into the corresponding material interface;

[0129] S505, each interface patch is generated in quadrilateral form, and then split into two triangular faces. Use spatial coordinate hashing to remove duplicate vertices and eliminate redundancy; through triangular mesh optimization operation, improve mesh quality and output efficiency, and use the formula for decomposing quadrilaterals into triangles:

[0130] Face quad= [v1, v2, v3, v4] -> {[v1, v2, v3], [v1, v3, v4]} ;

[0131] wherein Face quad is a quadrilateral facet; v1, v2, v3, v4 are 4 vertices of the quadrilateral facet; {[v1, v2, v3], [v1, v3, v4]} are two triangular facets obtained after splitting;

[0132] S506, the voxel grid coordinates after boundary expansion need to be translated as a whole to restore the original spatial position. Finally, each type of interface triangular facet grid is exported as a standard STL file to obtain the STL grid file; an input interface is provided for subsequent finite element grid partitioning and modeling integration; thereby realizing efficient and automated three-dimensional voxel reconstruction and interface surface meshing of the preform in the irregular aeronautical composite component, ensuring the continuity of the model;

[0133] S6, the STL grid file in S5 is imported into the preform surface grid model, and the AlphaWrap function of the open source tool Pymeshlab is called through a Python script to automatically wrap and smooth the surface grid of the preform substructure, generating a high-quality closed surface grid; although S6 optimizes the surface quality, there are still potential topological errors and mesh conforming defects at the intersection of complex substructures, which need to be repaired by subsequent methods;

[0134] S7, based on the smoothed surface grid model in S6, the MeshBoolean function of the open source tool Pymeshlab is called through a Python script to perform Boolean operations on the preform substructure grid, and redundant nodes and intersecting facets generated by the interference of substructures are detected and corrected in real time during the operation process, and the grid intersection and discontinuous area are processed to ensure the consistency and topological connectivity of the substructure grid boundary nodes;

[0135] S8, based on the topologically connected grid model in S7, each component interfacial region generated after Boolean operation is automatically extracted through a Python script, the spatial closure of all partitions is completed, and tetrahedral volume meshing is performed on the closed triangular facet grid domain to obtain the preform model; including the following steps:

[0136] S801, for each fiber bundle surface grid, the minimum enclosing cube parameters A_min and A_max of the matrix surface grid are used to determine whether all vertices of its facets satisfy ; the judgment condition is as follows:

[0137] A_min≤V≤A_max;

[0138] wherein [A_min, A_max] is the minimum enclosing cube of the region boundary;

[0139] If the judgment condition is met, the facet is retained, and the screening of the fiber bundle surface inside the preform region is realized;

[0140] S802, for the preform assembly after Boolean operation (such as fiber bundle A and B), extract the common interface between them, so that the internal grid is completely closed. Specifically: with the help of KD tree (k-dtree) for vertex space search, accurately locate the common nodes between A and B; check if the common nodes are on the common edge (edge), identify the intersection ring edge (ring edges); based on connected components partition, take the common interface ring edge as the boundary, divide A and B into several patches respectively, extract the intersection patch except the main patch; merge all the intersection patches and export, assign the patch to the missing surface component, and construct a closed internal interface grid;

[0141] S803, for each substructure and region of the preform, perform the common interface identification and patch partition operation of S802, extract all the common interface patches related to the boundary and merge and export them, so that each substructure of the preform is completely closed with the outside world;

[0142] S804, when all the substructures in the preform are closed and topologically continuous, perform tetrahedral meshing with the closed boundary as the constraint surface to generate the preform mesh model.

[0143] In this embodiment:

[0144] S1, select C f / SiC turbine guide vane blade as an implementation sample, CT scanning is carried out for the target region selected for the blade component, and CT image information containing the preform structure is obtained;

[0145] S2, develop a script based on Python (Ver 3.9), mainly rely on numpy, opencv-python, pydicom, matplotlib and other libraries, the specific processing process is as follows:

[0146] The slice data obtained by CT scanning is imported and subjected to gray scale normalization processing, so that the normalized pixel value is distributed in the interval [0, 1]; the preliminary functional unit region of the image to be processed is cropped, and the three-axis size of the model after cropping is 250x100x250, focusing on the key region containing the preform; in order to unify the size of the subsequent neural network training processing, the pad_to_multiple function is used to pad zeros for each slice, so that the height and width can be divided by 16, and the single-channel slice size after processing is (112, 256, 1); after the second processing of the slice size, the CLAHE (adaptive histogram equalization) function of OpenCV is called to enhance the contrast, and then Gaussian blur is used to denoise the slice, improve the quality of the slice, and obtain the processed slice image;

[0147] S3, from the slice image (a total of 251, used to construct a three-dimensional voxel model) processed in S2, in order to ensure that the sample used to train the neural network prediction model can cover the full target domain structure characteristics of the preform, about 20% of the slices (i.e. 50) are selected from the data set in an equal interval along the Z axis. With the help of the function of open source software ImageJ, different direction fiber bundles in the fiber bundle are manually labeled: red represents warp yarn (fiber bundle A), blue represents weft yarn (fiber bundle B), and black represents the background area, to generate a corresponding number of labeled data sets (GroundTruth);

[0148] S4, a U-Net neural network is built based on Python (Ver 3.9), mainly relying on numpy, opencv-python, pydicom, matplotlib, tensorflow, scikit-learn and other libraries, and the process is as follows:

[0149] The preprocessed slice is saved as a Numpy array, and the training set is generated in cooperation with the labeled image; the training set is divided into 80% for training and 20% for verification; the training set is subjected to data enhancement processing (such as rotation, translation, scaling, etc.) before being sent into the network, to improve the feature sample richness;

[0150] The classical U-Net symmetric structure is adopted, including down-sampling and up-sampling paths, each layer containing convolution, BatchNorm (batch normalization) and ReLU activation function, and the down-sampling is realized by maximum pooling and the up-sampling is realized by transposed convolution. The main structure parameters are as follows: the input layer size is (112, 256, 1); the down-sampling stage is four levels, and the convolution kernel numbers are 64, 128, 256 and 512 in turn; the up-sampling stage is four levels, and the convolution kernel numbers are set in the reverse order of 512, 256, 128 and 64; the output layer is 1x1 convolution, and outputs 3-channel logits (respectively corresponding to background, fiber bundle A and fiber bundle B);

[0151] The network adopts Sparse Categorical Crossentropy loss (from_logits=True), and monitors comprehensive indicators such as Sparse_Categorical_Crossentropy, Dice coefficient, Tversky Focal Loss, etc. in real time to balance segmentation accuracy and class balance; wherein the Dice coefficient and Tversky Focal Loss are expressed as:

[0152]

[0153] Loss = (1-Tversky) γ ;

[0154] Wherein, A is the real labeled region; B is the segmentation region predicted by the model; TP is the predicted pair region; FP is the false positive; FN is the false negative; a is the weight of FP; β is the weight of FN; γ is the focus parameter;

[0155] Early stopping (early stopping) and ReduceLROnPlateau (learning rate decay) are used to automatically adjust the learning rate and early termination strategy, and the key parameter settings are as follows: using Adam optimizer, the initial learning rate is 1e -3 (adjust dynamically), the batch size is 8, and the maximum iteration is 300 epochs;

[0156] After training, the optimal model is saved, and the full CT slice sequence is automatically predicted and segmented, and the segmentation label with the same size as the original mask is output and saved in PNG format; the segmentation result is visualized in background-red-blue three colors;

[0157] S5, develop Python script to process 2D slice label graph of U-Net segmentation output, the specific process is as follows:

[0158] Load slice label graph (size 251x101, slice number 251), automatically identify materials based on label information according to color threshold: red (R>128, G<50, B<50) is judged as fiber bundle A (value 2), blue (B>128, R<50, G<50) is judged as fiber bundle B (value 3), and the rest is automatically identified as matrix (value 1) or background (value 0); reconstruct the preform three-dimensional voxel model through the self-defined function load_color_masks;

[0159] Perform three-dimensional opening and closing operation on the obtained voxel model, and each group of voxels performs opening and closing operation operation of 9x9x9 structure element twice, to remove isolated noise generated by segmentation;

[0160] The current voxel model is three-dimensionally voxel coded, fiber bundle A, fiber bundle B and the matrix are combined into a three-dimensional integer voxel array, a background area is introduced outside the region, and a four-value voxel system is constructed; six types of interfaces of the voxel array (background-matrix, background-fiber bundle A / B, matrix-fiber bundle A / B, A-B interface) are automatically identified, the interface of adjacent voxels of different types is determined voxel by voxel, and a four-node rectangular patch is output, which is then split into triangular patches;

[0161] All the interfaces are split into triangular patches, and independent three-dimensional STL grid files are generated for each type of interface;

[0162] S6, after importing the prefabricated body surface grid model, a loop function is used to traverse each substructure of the prefabricated body, and the AlphaWrap function in the open source tool Pymeshlab library is called to realize the reconstruction of the surface grid (function parameters: alpha is selected as 0.002, offset is selected as 0.0005); after the surface reconstruction is completed, the HCLaplacian Smooth function is called to complete the surface grid smoothing processing, and a high-quality closed surface grid is generated, and the processed prefabricated body model is output;

[0163] S7, since the substructures inside the prefabricated body are in contact with each other, after the surface coating grid reconstruction and smoothing, the contact surface will change into an area intersecting with each other, therefore, the MeshBoolean function of the open source tool Pymeshlab is called through the Python script to perform Boolean operation on the prefabricated body substructure grid; during the operation, a "node consistency checking mechanism" is introduced to detect and correct redundant nodes and cross patches generated due to the interference of substructures in real time, effectively solving the problems of "operation not converging" and "boundary misplacement" that are prone to occur in traditional Boolean operation, and finally obtaining a topologically connected grid model;

[0164] S8, based on Python (mainly relying on Trimesh and Scipy libraries), a script is developed to post-process the prefabricated body model after the grid level Boolean operation, and the process is as follows:

[0165] Load the prefabricated body grid STL file, use the space clipping algorithm, and based on the bounding box of the matrix region, respectively judge the grid vertices of the substructures such as fiber bundle, and retain the substructure triangular faces completely located in the matrix bounding box;

[0166] All connected domains of the fiber bundle A or B main triangular grid inside the prefabricated body are obtained by using Trimesh splitting, and the KDTree (k-d tree) is used to quickly search the near neighbors of the vertices of the other substructure (such as fiber bundle B or matrix) (spatial tolerance tol = 1e -6) ; for each connected domain, identify the triangular faces with common nodes and common edges with another substructure, and divide the patch in a boundary "cut-off" manner (patch as a subset of the original structure, retaining the node index and face index of the original structure) ; for each patch, automatically detect the area, and regard the main patch with the largest area as the main region, and output all the secondary patches as the interface patches for merging, to ensure that all gaps are filled; the extracted patches are assigned to the substructure missing the region for common node processing, forming an internal space closed domain;

[0167] The above-mentioned patch identification process is repeated for the region boundary outer packaging box and the surfaces of the fiber bundles A and B to find all patches at the interface between the external region to prevent the closed domain from being missing; all "patch" triangular faces between the partitions and the external are output to provide the necessary boundary for the spatial domain closure;

[0168] The main surface of each partition, the patch and the external interface patch are summarized to form a closed triangular mesh body of the whole space; the mesh merging and deduplication mechanism of Trimesh is used to ensure the uniqueness of the nodes and the continuity of the patches; the closed triangular face domain of the preform is taken as input, the Delaunay is used for automatic subdivision to generate a mesh (the target mesh size is 1mm), and finally the preform mesh model is output.

[0169] Therefore, the aviation composite special-shaped component inner preform intelligent identification and modeling method adopts the above-mentioned method, which is aimed at the topological error and mesh conforming problem easily occurring in explicit modeling, combines the grid Boolean operation and post-processing repair algorithm, accurately processes the grid intersection and discontinuous area, repairs the mesh gap, missing face and other closed defects introduced by the Boolean operation, and guarantees the node conforming, topological connectivity and spatial closure of the preform substructure mesh; at the same time, the manual participation is reduced, and the efficiency and stability of modeling are improved.

[0170] Finally, it should be noted that: the above examples are only used to illustrate the technical solutions of the present application but not to limit it, although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can still be modified or equivalently replaced, and these modifications or equivalent replacements cannot make the modified technical solutions deviate from the spirit and scope of the technical solutions of the present application.

Claims

1. A method for intelligent identification and modeling of prefabricated bodies within aerospace composite irregular-shaped components, characterized in that, Includes the following steps: S1. Perform CT scans on the target area of ​​the irregularly shaped aerospace composite material component to obtain the original CT digital slice data. S2. Preprocess the original CT digital slice data in S1; Obtain the preprocessed slice sequence; S3. Select 20% of the slice samples from the preprocessed slice sequence in S2, and manually label them to construct the training set; S4. Based on the U-Net neural network framework built with Python, automatic instance segmentation of fiber bundles is achieved on all preprocessed slice sequences in S2 to obtain segmentation label maps; S5. Reconstruct the 3D voxel model of the segmentation label map in S4 to obtain the STL mesh file; S6. Import the STL mesh file from S5 into the prefabricated body surface mesh model, and automatically wrap and smooth the surface mesh of the substructure within the prefabricated body to obtain a smoothed surface mesh model. S7. Based on the smoothed surface mesh model in S6, Boolean operations are performed on the prefabricated substructure mesh. During the operation, redundant nodes and cross patches caused by factor structure interference are detected and corrected in real time. Mesh intersections and discontinuous regions are processed to obtain a topologically connected mesh model. S8. Based on the topologically connected mesh model in S7, extract the common interface regions of each component generated after Boolean operation, complete the spatial closure of the partition, and perform tetrahedral volume mesh subdivision on the closed triangular mesh domain to obtain the prefabricated volume model.

2. The intelligent identification and modeling method for prefabricated bodies within aerospace composite irregular-shaped components according to claim 1, characterized in that, S2 includes the following steps: S201. Batch import raw CT digital slice data, set the target region's XYZ three-axis spatial range, and focus on the representative volume element region containing the preform using a clipping formula; the clipping formula is: I crop (x,y)=I(x0+x,y0+y),x∈[0,w],y∈[[0,h]; Where (x0, y0) are the coordinates of the top left corner of the clipping window; h is the height of the window; and w is the width of the window. An adaptive histogram equalization method is used to improve fiber bundle identification, and local contrast is enhanced through histogram equalization transformation; the transformation formula is: Where, r k n represents the k-th gray level; j L is the number of pixels at gray level j; L is the number of gray levels; MN is the image size. S202. Gaussian filtering is used to smooth out random noise and high-frequency artifacts. The formula is: Where G(i,j) is a two-dimensional Gaussian kernel function; S203. Morphological opening and closing operations are used to eliminate tiny isolated noise points and micropores. The formula is as follows: Where I is the original binary image; B is the structuring element; ⊕ is the dilation operation; For erosion calculation.

3. The intelligent identification and modeling method for prefabricated bodies within aerospace composite irregular-shaped components according to claim 1, characterized in that, S4 includes the following steps: S401. The labeled dataset in S3 is paired with the corresponding preprocessed slice sequence in S2 and normalized. The image and label are then cropped at the center and filled at the boundary to ensure that the input size is adapted to the network structure and the network input size. S402. A U-Net encoder-decoder structure is adopted. The encoding path extracts multi-scale features through convolution and pooling, while the decoding path restores features and spatial localization through upsampling and skip connections, outputting a multi-channel prediction map, represented as follows: f(x)=Decoder(Encoder(x)+SkipConnections); Among them, Encoder represents extracting high-dimensional semantic features, Decoder represents gradually restoring the spatial structure, and SkipConnections represents ensuring the fusion of features and location information; S403 uses multi-class cross-entropy as the main loss, and introduces the Dice coefficient and Tversky loss index; the multi-class cross-entropy loss is expressed as: Among them, y i,k P is the true label for pixel i; i,k For predicting probabilities; L CE For multi-class cross-entropy loss; The Dice coefficient is expressed as: Among them, y i For real labels; p i ∈ represents the predicted probability; ∈ represents the smoothing term; S404. The Adam optimizer is used for training. The Dice coefficient is monitored through the validation set. The EarlyStopping and ReduceLROnPlateau mechanisms are used to prevent overfitting and adaptively adjust the learning rate. After training, the model weights are saved. S405: Perform batch inference on all CT slices and output pixel-level segmentation label maps; S406. Automatically predict the segmentation labels of batch CT slices and save the segmentation labels as standard image formats.

4. The intelligent identification and modeling method for prefabricated bodies within aerospace composite irregular-shaped components according to claim 1, characterized in that, S5 includes the following steps: S501: Batch import 2D segmentation masks, identify materials according to RGB thresholds, and combine them into a voxel array V in three-dimensional space. (x,y,z) Each voxel is assigned a material label M. (x,y,z) in{0,1,2,3}; correspond to the background, matrix, fiber bundle A, and fiber bundle B, respectively. S502. Isolated noise points are eliminated through three-dimensional morphological opening and closing operations, as shown below: Combined with connected component analysis, components with volumes smaller than the minimum connected component volume threshold S are removed. min The minute components, retaining the main structure, are represented as: Among them, V filtered (x,y,z) represents the output value of the voxel at position (x,y,z); S min The minimum connected component volume threshold is V(x,y,z); V(x,y,z) is the original value. S503. Merge each material region into a single three-dimensional voxel model according to priority, and add a background boundary to the outer layer of the voxel. S504. Traverse the voxel array. For each non-background voxel (x, y, z), sequentially detect its adjacent voxels (x, y, z) along the six positive and negative axes. ’ ,y ’ ,z ’ If the material numbers are different, an interface patch is generated between them; the vertex coordinates of the patch are determined by the voxel center and the facing direction, and the patch is classified into the corresponding material interface. S505. Each interface panel is generated as a quadrilateral, then split into two triangular faces; the formula for decomposing a quadrilateral into a triangle is: Face quad =[v1,v2,v3,v4]→{[v1,v2,v3],[v1,v3,v4]}; Among them, Face quad v1, v2, v3, v4 are the four vertices of the quadrilateral face; {[v1, v2, v3], [v1, v3, v4]} are the two triangular faces obtained after splitting. S506. After the voxel mesh coordinates are expanded by the boundary, they need to be translated as a whole to restore the original spatial position. Export each type of interface triangular mesh as a standard STL file to obtain the STL mesh file.

5. The intelligent identification and modeling method for prefabricated bodies within aerospace composite irregular-shaped components according to claim 1, characterized in that, S8 includes the following steps: S801. For each fiber bundle surface mesh, using the minimum outer cube parameter of the matrix surface mesh, determine whether all vertices of its faces satisfy the following criteria: A_min≤V≤A_max; Where [A_min, A_max] is the minimum bounding cube of the region boundary; If the condition is met, then the face piece is retained; S802. For the prefabricated components after Boolean operations, extract the common interface between them to ensure the internal mesh is completely closed. Specifically: Use a KD tree to search the vertex space and locate the common nodes between A and B; check whether the common nodes are on the same edge and identify the boundary loop edge; based on the connected component partitioning, divide A and B into several patches with the common boundary loop edge as the boundary, and extract the interface patch except for the main patch; merge all the interface patches and export them, assign the patches to the missing face component, and construct a closed internal interface mesh. S803. For each substructure of the prefabricated body and the outer boundary of the region, perform the common interface identification and patch partitioning operation of S802, extract all common interface patches related to the boundary and merge and export them, so that each substructure of the prefabricated body is completely closed to the outside world. S804. After all substructures within the prefabricated body are closed and topologically continuous, tetrahedral volume meshing is performed using the closed boundary as the constraint surface to generate the prefabricated body mesh model.

6. A device for intelligent identification and modeling of prefabricated bodies within aerospace composite irregular-shaped components, applied to the method for intelligent identification and modeling of prefabricated bodies within aerospace composite irregular-shaped components as described in any one of claims 1-5, characterized in that, include: The data acquisition module is used to perform CT scans on the target area of ​​irregularly shaped aerospace composite material components to obtain raw CT digital slice data. A data preprocessing module, connected to the data acquisition module, is used to preprocess the raw CT digital slice data; Obtain the preprocessed slice sequence; The training set construction module, connected to the data preprocessing module, is used to select 20% of the slice samples from the preprocessed slice sequence and manually label them to construct the training set. The intelligent segmentation module, connected to the training set construction module, is used to automatically segment fiber bundles from all preprocessed slice sequences using the U-Net neural network framework built in Python, and obtain segmentation label maps. The 3D reconstruction module, connected to the intelligent segmentation module, is used to reconstruct the 3D voxel model of the segmentation label map to obtain an STL mesh file. The mesh optimization module, connected to the 3D reconstruction module, is used to import the STL mesh file into the prefabricated surface mesh model, automatically cover and smooth the surface mesh of the substructure within the prefabricated body, and obtain a smoothed surface mesh model. The topology processing module, connected to the mesh optimization module, performs Boolean operations on the prefabricated substructure mesh based on the smoothed surface mesh model, processes mesh intersections and discontinuous regions, and obtains a topologically connected mesh model. The model output module is connected to the topology processing module. Based on the topologically connected mesh model, it extracts the common interface regions of each component generated after Boolean operation, completes the spatial closure of the partition, and performs tetrahedral volume mesh subdivision on the closed triangular mesh domain to obtain the prefabricated volume model.

Images