Composite material CT image performance identification method based on deep learning

Abstract

The invention relates to a composite material CT image performance identification method based on deep learning, and belongs to the field of industrial CT image processing. Comprising the following steps: acquiring CT image data of a composite material sample to perform gray value calibration, and inputting a trained U-net model to extract images of different semantic regions so as to obtain a corresponding gray matrix; constructing a finite element model of the composite material sample, and performing linear static analysis on the model to obtain a stiffness matrix of the model; obtaining gray matrixes corresponding to images of different semantic regions in the CT image of the composite material sample to be detected, inputting the gray matrixes into the trained twin-tower neural network model to obtain corresponding stiffness matrixes, and calculating the elastic modulus and Poisson's ratio of the composite material sample to be detected based on the stiffness matrixes. Therefore, the performance detection of the to-be-detected composite material sample is realized. According to the method, the detection from the composite material CT image to the composite material performance is realized, and the problems of low efficiency and low accuracy of the existing composite material performance detection are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial CT image processing, and in particular to a method for identifying composite material CT image performance based on deep learning. Background Art

[0002] Composite materials are widely used in aerospace, automotive, and other industries due to their excellent mechanical properties and lightweight characteristics. However, due to the influence of numerous factors, composite materials' performance will change over time. Therefore, testing the performance of composite materials is particularly important.

[0003] Currently, performance testing of composite materials mostly relies on manual testing methods, which is not only time-consuming and labor-intensive, but also relies on manual experience and subjective judgment, resulting in large errors in test results and reducing the accuracy of testing.

[0004] It can be seen that providing a composite material performance identification method to improve the efficiency and accuracy of detection is an urgent problem to be solved. Summary of the Invention

[0005] In view of the above analysis, the present invention aims to provide a composite material CT image performance recognition method based on deep learning to solve the current problems of low efficiency and low accuracy in composite material performance detection.

[0006] The present invention provides a method for identifying composite material CT image properties based on deep learning, the method comprising the following steps: Acquiring CT image data of the composite material sample, and performing grayscale value calibration on the CT image data; The calibrated CT image data is input into the trained U-net model to extract images of different semantic regions in the CT image, and the corresponding grayscale matrix is ​​obtained based on the extracted images of different semantic regions; Constructing a finite element model of the composite material sample, and performing linear static analysis on the finite element model to obtain its stiffness matrix; Constructing a twin-tower neural network model, and training the twin-tower neural network model based on the grayscale matrix and the stiffness matrix to obtain a trained twin-tower neural network model; The calibrated CT image data of the composite material sample to be tested is input into the trained U-net model, and images of different semantic areas in the CT image are extracted to obtain the corresponding grayscale matrix; the grayscale matrix is ​​input into the trained double-tower neural network model to obtain the corresponding stiffness matrix, and the elastic modulus and Poisson's ratio of the composite material sample to be tested are calculated based on the stiffness matrix, thereby realizing the performance detection of the composite material sample to be tested.

[0007] Furthermore, the grayscale values ​​of the CT image data are calibrated by using a histogram matching method.

[0008] Furthermore, inputting the calibrated CT image data into the trained U-net model to extract images of different semantic regions in the CT image includes: The calibrated CT image data is input into the trained U-net model to identify images of different semantic regions in the CT image; Create corresponding masks for the images of different semantic areas identified; By performing bitwise operations on the CT image and the corresponding mask, the images of the corresponding different semantic regions are extracted.

[0009] Furthermore, obtaining corresponding grayscale matrices based on the extracted images of different semantic areas includes: For each extracted image of different semantic areas: calculate the grayscale value of each pixel of the image, and arrange the grayscale values ​​of all pixels of the image in the form of a matrix to obtain the grayscale matrix corresponding to the image.

[0010] Furthermore, a finite element model of the composite material sample was constructed using Digimat software, and a linear static analysis was performed on the finite element model using Abaqus software to obtain stress, strain, reaction force and displacement data of the finite element model, and the stiffness matrix of the finite element model was obtained based on the stress, strain, reaction force and displacement data.

[0011] Furthermore, the twin-tower neural network model includes an input layer, a twin-tower structure, and an output layer; Wherein, the input layer is used to receive the input grayscale matrix and stiffness matrix; The dual-tower structure includes two parallel tower structures and an information fusion layer. The tower structures are used to extract the eigenvectors of the grayscale matrix and the eigenvectors of the stiffness matrix respectively. The information fusion layer adopts a cross-attention mechanism to fuse and splice the eigenvectors of the grayscale matrix and the eigenvectors of the stiffness matrix, and sends the fused and spliced ​​eigenvectors to the output layer. The output layer includes a fully connected layer for mapping the fused and concatenated feature vectors back to a matrix to generate a predicted stiffness matrix, and outputting the predicted stiffness matrix.

[0012] Furthermore, each tower structure includes two fully connected modules connected in series, and the fully connected module includes a fully connected layer and a ReLU activation function connected in series.

[0013] Furthermore, the loss function of the dual-tower neural network model is expressed as follows: , in, represents the value of the loss function, represents the element in row i and column j of the input stiffness matrix, represents the element in row i and column j of the predicted stiffness matrix, Represents the dimension of the stiffness matrix.

[0014] Furthermore, the elastic modulus of the composite material to be tested is calculated by the following formula: , Wherein, E1 is the elastic modulus of the composite material to be tested in the x direction, E2 is the elastic modulus of the composite material to be tested in the y direction, and E3 is the elastic modulus of the composite material to be tested in the z direction; is the element in row i and column i in the stiffness matrix, A 11 Represents the positive stiffness of the composite material to be tested in the x direction, A 22 Represents the positive stiffness of the composite material to be tested in the y direction, A 33 Represents the positive stiffness of the composite material to be tested in the z direction.

[0015] Furthermore, the Poisson's ratio of the composite material to be tested is calculated by the following formula: , in, is the Poisson's ratio of the composite material to be tested, which represents the transverse deformation in the j direction when the composite material to be tested is stretched in the i direction, wherein i is 1 for the x direction, i is 2 for the y direction, i is 3 for the z direction, j is 1 for the transverse deformation in the x direction, j is 2 for the transverse deformation in the y direction, and j is 3 for the transverse deformation in the z direction; is the element in the i-th row and j-th column of the stiffness matrix.

[0016] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects: 1. The present invention extracts the grayscale matrices corresponding to different semantic regions in the CT image of the composite material sample to be tested, obtains the corresponding stiffness matrix through a dual-tower neural network model, and then performs performance testing on the composite material sample to be tested. This realizes the detection of composite material properties from composite material CT images, improves detection efficiency and accuracy, and solves the current problems of low efficiency and low accuracy in composite material performance testing. The invention can be widely used in the performance testing of most composite materials before and after service.

[0017] 2. The present invention uses the U-net model to identify images of different semantic regions in the CT image of the composite material sample to be tested, thereby improving the efficiency and accuracy of recognition.

[0018] 3. The present invention uses a histogram matching method to calibrate the grayscale values ​​of images acquired by different CT scanners or by different scans of the same CT scanner, thereby achieving the unification of the grayscale values ​​of CT images, avoiding errors caused by grayscale value differences, reducing noise and unevenness in CT images, and thus improving the quality and efficiency of CT image processing.

[0019] 4. The present invention utilizes Digimat software and Abaqus software to complete a complete set of processes from finite element model construction to linear static analysis of composite materials, which is easy to operate and reduces dependence on the expertise of R&D personnel.

[0020] In the present invention, the above-mentioned technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of the present invention will be described in the following description, and some advantages will become apparent from the description or be learned through practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the contents particularly pointed out in the description and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] The accompanying drawings are only used for the purpose of illustrating specific embodiments and are not to be considered as limiting the present invention. Like reference symbols denote like components throughout the accompanying drawings. Figure 1 This is a flow chart of a composite material CT image performance recognition method based on deep learning according to an embodiment of the present invention; Figure 2 This is a CT image of a braided composite material sample according to an embodiment of the present invention; FIG3( a ) is a diagram showing the result of prediction and recognition of a CT image according to an embodiment of the present invention; FIG3( b ) is a diagram showing the result of regional division of a CT image according to an embodiment of the present invention; FIG3( c ) is a diagram showing the result of bitwise operation of a CT image according to an embodiment of the present invention; FIG3( d ) is a diagram showing the results of grayscale matrix extraction of a CT image according to an embodiment of the present invention. DETAILED DESCRIPTION

[0022] The preferred embodiments of the present invention will be described in detail below in conjunction with the accompanying drawings, wherein the accompanying drawings constitute a part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, and are not used to limit the scope of the present invention.

[0023] A specific embodiment of the present invention discloses a method for identifying composite material CT image properties based on deep learning. Figure 1 As shown, the method includes the following steps: Step S1, obtaining CT image data of a composite material sample, and performing grayscale value calibration on the CT image data; Step S2: input the calibrated CT image data into the trained U-net model, extract images of different semantic regions in the CT image, and obtain corresponding grayscale matrices based on the extracted images of different semantic regions; Step S3: constructing a finite element model of the composite material sample, and performing linear static analysis on the finite element model to obtain its stiffness matrix; Step S4: constructing a twin-tower neural network model, and training the twin-tower neural network model based on the grayscale matrix and the stiffness matrix to obtain a trained twin-tower neural network model; Step S5: Input the calibrated CT image data of the composite material sample to be tested into the trained U-net model, extract images of different semantic areas in the CT image, and then obtain a corresponding grayscale matrix; input the grayscale matrix into the trained double-tower neural network model to obtain a corresponding stiffness matrix, and calculate the elastic modulus and Poisson's ratio of the composite material sample to be tested based on the stiffness matrix, thereby realizing performance testing of the composite material sample to be tested.

[0024] Specifically, in step S1, CT image data of several composite material samples of the same type as the composite material to be tested is obtained through CT scanning. Each CT scan can obtain a series of two-dimensional slice CT images of a composite material sample, and all two-dimensional slice CT images of the composite material sample constitute its CT image data.

[0025] It should be noted that a three-dimensional composite material sample is scanned by CT to obtain a series of two-dimensional slice CT images, each of which represents a cross-section of the composite material sample; all two-dimensional slice CT images of the composite material sample constitute its three-dimensional volume CT image, i.e., CT image data.

[0026] Specifically, the resolution of the CT scan is 2 microns, and the format of the CT image data is TIFF.

[0027] It should be noted that TIFF (Tagged Image File Format) is a standard format for storing and exchanging raster images. The original image data generated by CT scanning may not be in TIFF format. It can be converted to TIFF format through software to facilitate the sharing and processing of CT image data between different platforms and software.

[0028] For example, a CT scan is performed on a certain woven composite material sample, and the resolution of the CT scanner is 2 microns. The CT image data thereof is obtained by CT scanning, such as Figure 2 As shown, the CT image data is output in TIFF format.

[0029] Furthermore, the grayscale values ​​of the CT image data are calibrated by using a histogram matching method.

[0030] Specifically, a composite material sample is randomly selected, and a two-dimensional slice CT image is randomly selected from the CT image data of the composite material sample as a reference image, and all two-dimensional slice CT images in the CT image data of the remaining composite material samples are used as source images, and each source image is histogram matched with the reference image.

[0031] Specifically, each source image is histogram matched with the reference image by the following method: Calculate the grayscale histograms of the reference image and the source image respectively; Calculating the cumulative distribution functions corresponding to the reference image and the source image respectively based on the grayscale histograms of the reference image and the source image; For each grayscale value in the source image, find the closest value in the cumulative distribution function corresponding to the reference image through the cumulative distribution function corresponding to the source image, thereby obtaining the corresponding grayscale value in the reference image; Each grayscale value in the source image is mapped to the corresponding grayscale value in the reference image for histogram matching to obtain a matched source image.

[0032] It should be noted that even when the same CT scanner is used for CT scanning, the grayscale values ​​of the two-dimensional slice CT images obtained from a single scan are relatively small, while the grayscale values ​​of the two-dimensional slice CT images obtained from different scans are relatively large, requiring grayscale value unification. This application uses Python code to adjust the grayscale distribution of the image using histogram matching, that is, to calibrate the grayscale values ​​of the image to match the grayscale values ​​of the reference image.

[0033] It can be understood that the grayscale value calibration of images acquired by different CT scanners or different scans of the same CT scanner is performed by using histogram matching, which achieves the unification of the grayscale values ​​of CT images, avoids errors caused by grayscale value differences, reduces the noise and unevenness of CT images, and thus improves the quality and efficiency of CT image processing.

[0034] For example, the CT image data of woven composite material samples are read by Python's OpenCV and NumPy libraries, and grayscale histogram and cumulative distribution function are calculated and mapped to achieve grayscale value calibration of the CT image data.

[0035] Specifically, in step S2, inputting the calibrated CT image data into the trained U-net model and extracting images of different semantic regions in the CT image includes: The calibrated CT image data is input into the trained U-net model to identify images of different semantic regions in the CT image; Create corresponding masks for the images of different semantic areas identified; By performing bitwise operations on the CT image and the corresponding mask, the images of the corresponding different semantic regions are extracted.

[0036] Specifically, the calibrated CT image data is fed into a trained U-net model, which identifies the images of different semantic regions within each 2D slice CT image, representing each region with a different color. Python code is written to convert the color space of the images of different semantic regions into the HSV color space, and a mask corresponding to each semantic region is created based on the HSV color space. For each 2D slice CT image, an AND operation is performed on each pixel in the CT image with the corresponding pixel in the mask to extract the image of the corresponding semantic region.

[0037] It should be noted that different semantic region images are converted from the BGR (or RGB) color space to the HSV color space, which includes hue (H), saturation (S), and value (V). Hue represents color information, namely, the position of the spectral color; saturation indicates the color's purity, namely, how close the color is to the spectral color; and value indicates how bright the color is. The value range of the HSV color space is defined based on the color of the identified semantic region images. For example, assuming the color of the identified semantic region image is blue, the hue value range in the HSV color space is defined as 100-140, the saturation value range is defined as 0-1, and the value range is defined as 0-1. A mask corresponding to the semantic region image is created based on the value range of the HSV color space. If the pixel value in the CT image is within the value range of the HSV color space, the corresponding pixel value in the mask is set to the same value. Otherwise, the corresponding pixel value in the mask is set to 0, resulting in a blue mask. By performing image bitwise operations on the two-dimensional slice CT image and the corresponding mask, images of different semantic areas corresponding to the two-dimensional slice CT image are extracted.

[0038] As can be understood, the HSV color space includes hue, saturation, and value, making it more convenient for color analysis and processing. Furthermore, by creating corresponding masks for images of different semantic regions, it is helpful to extract regions of specified colors (i.e., regions of different semantics) in CT images. The present invention uses a U-net model to identify images of different semantic regions in CT images of composite material samples to be tested, improving recognition efficiency and accuracy.

[0039] Furthermore, obtaining corresponding grayscale matrices based on the extracted images of different semantic areas includes: For each extracted image of different semantic areas: calculate the grayscale value of each pixel of the image, and arrange the grayscale values ​​of all pixels of the image in the form of a matrix to obtain the grayscale matrix corresponding to the image.

[0040] For example, the labeled CT image data of a woven composite material sample was input into a trained U-net model to identify the warp and weft regions in the CT image. The recognition result is shown in Figure 3(a). The warp region is represented by red, and the weft region is represented by green. Python code was written to convert the color space of the warp and weft regions into the HSV color space. Based on the red color representing the warp region, the value range of the HSV color space corresponding to red was defined, and a mask corresponding to the warp region was established based on the HSV color space corresponding to red. Similarly, based on the green color representing the weft region, the value range of the HSV color space corresponding to green was defined, and a mask corresponding to the weft region was established based on the HSV color space corresponding to green. The masks corresponding to the warp and weft regions are shown in Figure 3(b). The image of the warp region was extracted by performing a bitwise AND operation on the CT image of the woven composite material sample and the mask corresponding to the warp region. Similarly, the image of the weft region was extracted by performing a bitwise AND operation on the CT image of the woven composite material sample and the mask corresponding to the weft region. The extracted images of the warp and weft regions are shown in Figure 3(c). The grayscale value of each pixel in the extracted warp and weft region images was calculated to obtain the grayscale matrices corresponding to the warp and weft region images. The image represented by these grayscale matrices is shown in Figure 3(d).

[0041] Furthermore, the trained U-net model is obtained by the following method: Constructing a U-net model, obtaining CT image data of the composite material sample based on step S1, performing grayscale value calibration on the CT image data, and annotating images of different semantic areas in the calibrated CT image data to construct a first data set; Inputting the first data set into the U-net model for training, wherein the U-net model recognizes images of different semantic regions; The value of the U-net model loss function is calculated by the error between the labeled and identified images of different semantic regions, and a trained U-net model is obtained when the value of the U-net model loss function reaches a threshold.

[0042] Furthermore, images of different semantic areas in the calibrated CT image data are divided, the divided images are named and labeled, and the labeled CT image data is used as the first data set.

[0043] Specifically, the images of different semantic regions in the calibrated CT image data are segmented. Each region represents a material with specific characteristics. Each region is named in the form of a serial number and labeled with a corresponding material name label.

[0044] Exemplarily, the images of the warp yarn area and the weft yarn area in the calibrated CT image data are divided, the divided warp yarn images and weft yarn images are named and labeled respectively, and the labeled CT image data is used as the first data set.

[0045] Specifically, a sample in the first dataset is a two-dimensional slice CT image of an annotated composite material sample. The first dataset includes samples corresponding to multiple composite materials of the same type, with each composite material sample corresponding to multiple samples. When training the U-net model, the training step size is set to 64, the number of categories is set to 3, and the batch size is set to 50, 100, or 200. The Adam optimizer is used for learning rate descent, and the model parameters are iteratively adjusted based on the calculated loss function value.

[0046] Specifically, in step S3, a finite element model of the composite material sample is constructed using Digimat software, and a linear static analysis is performed on the finite element model using Abaqus software to obtain stress, strain, reaction force and displacement data of the finite element model, and a stiffness matrix of the finite element model is obtained based on the stress, strain, reaction force and displacement data.

[0047] Specifically, Digimat software is used to construct a geometric model of the composite material sample, the geometric model is meshed, and a file in .inp format is exported as the finite element model of the composite material sample; then Abaqus software is used to read the .inp file, set boundary conditions and loads are applied, and the solution is performed to obtain the stress and strain cloud map of the finite element model and the stress, strain, reaction force and displacement data and output them; by writing Python code, the stress, strain, reaction force and displacement data output by the Abaqus software are parsed, the nodes and units in the finite element model are traversed, and the stiffness matrix of the finite element model is extracted.

[0048] Exemplarily, Digimat software is used to establish the matrix and fiber of a woven composite material sample, and the elastic modulus and Poisson's ratio of the matrix and fiber are input respectively; for the fiber, parameters such as fiber density, fiber diameter, fiber length, etc. are input, and the close arrangement degree of the warp and weft yarns is selected to establish the geometric model of the woven composite material sample. A hexahedral grid is selected to mesh the geometric model, and a file in the .inp format is exported as the finite element model of the woven composite material sample. Abaqus software is used to read the .inp file, the analysis type is set, the boundary conditions and the magnitude of the load are applied to the finite element model, and the Abaqus software is submitted for solution to obtain the stress, strain cloud diagram, and stress, strain, reaction force and displacement data of the finite element model. Python code is written to traverse the nodes and units in the finite element model and extract the stiffness matrix of the finite element model.

[0049] It can be understood that the present invention utilizes Digimat software and Abaqus software to complete a complete set of processes from finite element model construction to linear static analysis of composite materials, which is easy to operate and reduces dependence on the expertise of R&D personnel.

[0050] Specifically, in step S4, the twin-tower neural network model includes an input layer, a twin-tower structure, and an output layer; Wherein, the input layer is used to receive the input grayscale matrix and stiffness matrix; The dual-tower structure includes two parallel tower structures and an information fusion layer. The tower structures are used to extract the eigenvectors of the grayscale matrix and the eigenvectors of the stiffness matrix respectively. The information fusion layer adopts a cross-attention mechanism to fuse and splice the eigenvectors of the grayscale matrix and the eigenvectors of the stiffness matrix, and sends the fused and spliced ​​eigenvectors to the output layer. The output layer includes a fully connected layer for mapping the fused and concatenated feature vectors back to a matrix to generate a predicted stiffness matrix, and outputting the predicted stiffness matrix.

[0051] Specifically, each tower structure includes two fully connected modules connected in series, each of which includes a fully connected layer and a ReLU activation function connected in series. The information fusion layer uses a cross-attention mechanism, which uses the eigenvector of the grayscale matrix as the query vector and the eigenvector of the stiffness matrix as the key vector and value vector. This enables the dual-tower neural network model to dynamically focus on the features of the stiffness matrix based on the features of the grayscale matrix, thereby strengthening the correlation between the two features. The fully connected layer in the output layer maps the fused and concatenated eigenvectors back into a two-dimensional matrix to output the predicted stiffness matrix.

[0052] Specifically, the expression of the ReLU activation function is as follows: , in, represents the output of the activation function, represents the value input to the activation function, Indicates taking and The maximum of the two.

[0053] Furthermore, the trained dual-tower neural network model is obtained by the following method: Constructing a second data set based on the grayscale matrix obtained in steps S1-S2 and the stiffness matrix obtained in step S3; Inputting the second data set into the twin-tower neural network model for training, and the twin-tower neural network model predicts the stiffness matrix; The value of the loss function of the twin-tower neural network model is calculated by the error between the input and predicted stiffness matrices, and a trained twin-tower neural network model is obtained when the value of the loss function of the twin-tower neural network model reaches a threshold.

[0054] Specifically, a sample in the second data set includes a grayscale matrix and its corresponding stiffness matrix obtained from the labeled CT image data of the composite material sample. The second data set includes samples corresponding to multiple composite materials of the same type, and each composite material sample corresponds to multiple samples. When training the twin-tower neural network model, the batch size is set to 50, 100, or 200, and the learning rate is reduced using the Adam optimizer. The parameters of the twin-tower neural network model are optimized by calculating the loss function value of the twin-tower neural network model. The training of the twin-tower neural network model is completed by continuous iteration until the loss function converges.

[0055] Specifically, the loss function of the twin-tower neural network model is expressed as follows: , in, represents the value of the loss function, represents the element in row i and column j of the input stiffness matrix, represents the element in row i and column j of the predicted stiffness matrix, Represents the dimension of the stiffness matrix.

[0056] It should be noted that the dimensions of the input grayscale matrix, stiffness matrix, and predicted stiffness matrix are all the same. By training the twin-tower neural network model, the trained twin-tower neural network model establishes a mapping relationship between the grayscale matrix and the stiffness matrix, and stores this mapping relationship in the model through model parameters to achieve prediction from the grayscale matrix to the stiffness matrix.

[0057] Specifically, in step S5, the CT image data of the composite material sample to be tested is obtained based on step S1, and the grayscale value is calibrated; the calibrated CT image data of the composite material sample to be tested is input into the trained U-net model, and images of different semantic areas in the CT image are extracted, and corresponding grayscale matrices are obtained based on the extracted images of different semantic areas; the trained dual-tower neural network model is called by writing code, and the grayscale matrix is ​​input into the trained dual-tower neural network model to obtain the corresponding stiffness matrix.

[0058] It should be noted that several corresponding grayscale matrices are obtained from the CT image data of the composite material sample to be tested. Each grayscale matrix obtains the stiffness matrix predicted by the grayscale matrix through the trained dual-tower neural network model. The stiffness matrix predicted by all grayscale matrices is taken as the arithmetic average to obtain the stiffness matrix corresponding to the composite material sample to be tested. When calling the trained dual-tower neural network model to predict the stiffness matrix, the tower structure and information fusion layer used to extract the eigenvector of the stiffness matrix are shielded. The eigenvector of the input grayscale matrix is ​​extracted through the tower structure that extracts the eigenvector of the grayscale matrix. The corresponding stiffness matrix is ​​generated through the mapping relationship between the grayscale matrix represented by the fully connected layer and the stiffness matrix.

[0059] Specifically, the elastic modulus of the composite material to be tested is calculated by the following formula: , Wherein, E1 is the elastic modulus of the composite material to be tested in the x direction, E2 is the elastic modulus of the composite material to be tested in the y direction, and E3 is the elastic modulus of the composite material to be tested in the z direction; is the element in row i and column i in the stiffness matrix, A 11 Represents the positive stiffness of the composite material to be tested in the x direction, A 22 Represents the positive stiffness of the composite material to be tested in the y direction, A 33 Represents the positive stiffness of the composite material to be tested in the z direction.

[0060] Specifically, the Poisson's ratio of the composite material to be tested is calculated by the following formula: , in, is the Poisson's ratio of the composite material to be tested, which represents the transverse deformation in the j direction when the composite material to be tested is stretched in the i direction, wherein i is 1 for the x direction, i is 2 for the y direction, i is 3 for the z direction, j is 1 for the transverse deformation in the x direction, j is 2 for the transverse deformation in the y direction, and j is 3 for the transverse deformation in the z direction; is the element in the i-th row and j-th column of the stiffness matrix.

[0061] It should be noted that composite materials are composed of multiple materials, and the mechanical properties of different materials are usually anisotropic; for anisotropic composite materials, the stiffness matrix It can be expressed as: ,in, is the volume fraction of the first material in the composite material, , is the stiffness matrix of the first material in the composite material, and m is the number of different materials in the composite material. Therefore, based on the stiffness matrix of the composite material sample to be tested, the elastic modulus and Poisson's ratio of the composite material sample to be tested can be calculated. Based on the elastic modulus and Poisson's ratio of the composite material sample to be tested, the composite material sample to be tested can be tested for properties such as tensile strength, compressive strength, and fatigue resistance, thereby enabling performance testing based on CT images of the composite material sample to be tested. CT scanning of the composite material can be used to test the properties of the composite material using the above method, improving detection efficiency and accuracy.

[0062] It can be understood that the present invention extracts the grayscale matrix corresponding to the images of different semantic areas in the CT image of the composite material sample to be tested, obtains the corresponding stiffness matrix through the dual-tower neural network model, and then performs performance testing of the composite material sample to be tested, thereby realizing the detection of composite material performance from composite material CT image, improving detection efficiency and accuracy, solving the current problems of low efficiency and low accuracy in composite material performance testing, and can be widely used in performance testing of most composite materials before and after service.

[0063] Compared with the existing technology, the composite material CT image performance identification method based on deep learning provided by the present invention has the following beneficial effects: 1. The present invention extracts the grayscale matrix corresponding to the image of different semantic areas in the CT image of the composite material sample to be tested, obtains the corresponding stiffness matrix through the double-tower neural network model, and then performs performance testing of the composite material sample to be tested. It realizes the detection of composite material performance from composite material CT image, improves the detection efficiency and accuracy, solves the current problems of low efficiency and low accuracy in composite material performance testing, and can be widely used in the performance testing of most composite materials before and after service.

[0064] 2. The present invention uses the U-net model to identify images of different semantic areas in the CT image of the composite material sample to be tested, thereby improving the efficiency and accuracy of recognition.

[0065] 3. The present invention uses a histogram matching method to calibrate the grayscale values ​​of images acquired by different CT scanners or by different scans of the same CT scanner, thereby achieving the unification of the grayscale values ​​of CT images, avoiding errors caused by grayscale value differences, reducing noise and unevenness in CT images, and thus improving the quality and efficiency of CT image processing.

[0066] 4. The present invention utilizes Digimat software and Abaqus software to complete a complete set of processes from finite element model construction to linear static analysis of composite materials, which is easy to operate and reduces dependence on the expertise of R&D personnel.

[0067] Those skilled in the art will appreciate that all or part of the process steps of the above-described embodiments can be implemented by instructing related hardware through a computer program, and the program can be stored in a computer-readable storage medium, such as a magnetic disk, an optical disk, a read-only memory, or a random access memory.

[0068] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by any technician familiar with this technical field within the technical scope disclosed by the present invention should be covered by the scope of protection of the present invention.

Claims

1. A composite material CT image performance recognition method based on deep learning, characterized in that: The method comprises the following steps: Acquiring CT image data of the composite material sample, and performing grayscale value calibration on the CT image data; The calibrated CT image data is input into the trained U-net model to extract images of different semantic areas in the CT image, and the corresponding grayscale matrix is ​​obtained based on the extracted images of different semantic areas; Constructing a finite element model of the composite material sample, and performing a linear static analysis on the finite element model to obtain its stiffness matrix; Constructing a twin-tower neural network model, and training the twin-tower neural network model based on the grayscale matrix and the stiffness matrix to obtain a trained twin-tower neural network model; The calibrated CT image data of the composite material sample to be tested is input into the trained U-net model, and the images of different semantic areas in the CT image are extracted to obtain the corresponding grayscale matrix; the grayscale matrix is ​​input into the trained double-tower neural network model to obtain the corresponding stiffness matrix, and the elastic modulus and Poisson's ratio of the composite material sample to be tested are calculated based on the stiffness matrix, thereby realizing the performance detection of the composite material sample to be tested.

2. The method for composite material CT image performance recognition based on deep learning according to claim 1, characterized in that: The grayscale value of the CT image data is calibrated by using a histogram matching method.

3. The composite material CT image performance recognition method based on deep learning according to claim 1 is characterized in that: The step of inputting the calibrated CT image data into the trained U-net model and extracting images of different semantic areas in the CT image comprises: The calibrated CT image data is input into the trained U-net model to identify images of different semantic areas in the CT image; Create corresponding masks for images of different semantic areas identified; By performing bitwise operations on the CT image and the corresponding mask, the images of the corresponding different semantic areas are extracted.

4. The method for composite material CT image performance recognition based on deep learning according to claim 3 is characterized in that: The method of obtaining a corresponding grayscale matrix based on the extracted images of different semantic areas includes: For each image of different semantic areas extracted: calculate the grayscale value of each pixel of the image, and arrange the grayscale values ​​of all pixels of the image in the form of a matrix to obtain the grayscale matrix corresponding to the image.

5. The method for composite material CT image performance recognition based on deep learning according to claim 1, characterized in that: Digimat software is used to construct a finite element model of the composite material sample, and Abaqus software is used to perform linear static analysis on the finite element model to obtain stress, strain, reaction force and displacement data of the finite element model. Based on the stress, strain, reaction force and displacement data, the stiffness matrix of the finite element model is obtained.

6. The method for composite material CT image performance recognition based on deep learning according to claim 1, characterized in that: The double-tower neural network model includes an input layer, a double-tower structure and an output layer; Wherein, the input layer is used to receive the input grayscale matrix and stiffness matrix; The dual-tower structure includes two parallel tower structures and an information fusion layer, wherein the tower structures are used to extract the eigenvector of the grayscale matrix and the eigenvector of the stiffness matrix respectively; the information fusion layer adopts a cross attention mechanism to fuse and splice the eigenvector of the grayscale matrix with the eigenvector of the stiffness matrix, and sends the fused and spliced ​​eigenvector to the output layer; The output layer includes a fully connected layer, which is used to map the fused and concatenated feature vectors back to a matrix to generate a predicted stiffness matrix, and output the predicted stiffness matrix.

7. The method for composite material CT image performance recognition based on deep learning according to claim 6, characterized in that: Each tower structure includes two fully connected modules connected in series, and the fully connected module includes a fully connected layer and a ReLU activation function connected in series.

8. The method for composite material CT image performance recognition based on deep learning according to claim 1, characterized in that: The expression of the loss function of the twin-tower neural network model is as follows: ， in, represents the value of the loss function, represents the element in the i-th row and j-th column of the input stiffness matrix, represents the element in the i-th row and j-th column of the predicted stiffness matrix, Represents the dimension of the stiffness matrix.

9. The method for composite material CT image performance recognition based on deep learning according to claim 1, characterized in that: The elastic modulus of the composite material to be tested is calculated by the following formula: ， Wherein, E1 is the elastic modulus of the composite material to be tested in the x direction, E2 is the elastic modulus of the composite material to be tested in the y direction, and E3 is the elastic modulus of the composite material to be tested in the z direction; is the element in the i-th row and i-th column of the stiffness matrix, A 11 Represents the positive stiffness of the composite material to be tested in the x direction, A 22 Represents the positive stiffness of the composite material to be tested in the y direction, A 33 Represents the positive stiffness of the composite material to be tested in the z direction.

10. The method for composite material CT image performance recognition based on deep learning according to claim 9, characterized in that: The Poisson's ratio of the composite material to be tested is calculated by the following formula: ， in, is the Poisson's ratio of the composite material to be tested, indicating the transverse deformation in the j direction when the composite material to be tested is stretched in the i direction, wherein i is 1 for the x direction, i is 2 for the y direction, i is 3 for the z direction, j is 1 for the transverse deformation in the x direction, j is 2 for the transverse deformation in the y direction, and j is 3 for the transverse deformation in the z direction; is the element in the i-th row and j-th column of the stiffness matrix.

Images