Xct image-based method and system for automatic recognition and classification of defects in cfpr

Abstract

The application discloses an XCT image-based CFRP defect automatic identification and classification method and system, and belongs to the field of carbon fiber reinforced composite material manufacturing and processing quality detection.The application obtains the three-dimensional features of defects through binarization three-dimensional label statistics classification, obtains the two-dimensional features of defects through manual labeling based on the three-dimensional label of defects, and after fusing the three-dimensional features and the two-dimensional features of defects, combines the features with a deep learning method to obtain an efficient real defect identification and classification model, which is simple and efficient and effectively improves the CFRP defect identification precision.The application uses the three-dimensional morphological parameters of defects when labeling the data set, and avoids the unreliability introduced by directly labeling on two-dimensional slices.The application trains a semantic segmentation model for identifying and classifying CFRP defects based on a small amount of two-dimensional CT images, and the semantic segmentation model is used for predicting internal defects of similar materials, so that a large amount of tedious and time-consuming manual image processing for obtaining internal defect data of materials is avoided, and a large amount of labor cost is saved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of manufacturing and processing quality detection of carbon fiber reinforced composites, and more particularly relates to a CFRP defect automatic identification and classification method and system based on XCT images. BACKGROUND

[0002] Carbon fiber reinforced composites (CFRP) are high-performance composites with carbon fibers as reinforcing phase and resins and other materials as matrix. CFRP has the characteristics of lightweight, high strength, high temperature resistance, and excellent mechanical properties. In recent years, CFRP has been widely used in fields with strict requirements on density, stiffness, fatigue characteristics, service temperature, and chemical stability, especially in the fields of aviation, wind power, and nuclear energy. In practical applications, various material defects (such as pores, cracks, delamination, and inclusions) directly affect the performance and service life of CFRP. The content and morphology of these defects are key parameters that affect the macroscopic physical and mechanical properties of CFRP. Quantitative description of the characteristics of defects in materials is of great significance for optimizing the production and manufacturing process of carbon fiber composites.

[0003] There are mainly two non-destructive testing methods for CFRP defects, namely ultrasonic testing and X-ray computed tomography (XCT). Ultrasonic testing is simple to operate and can measure the location of defects but cannot quantitatively describe and distinguish defects. XCT has the advantage that the three-dimensional images (two-dimensional image sequences) obtained by scanning can accurately describe the location and morphology of defects in the material. However, there are a large number of defects and noise in the CT images of CFRP. Two-dimensional image slices only reflect the projection of defects in a certain direction, and the gray value ranges of various defects are roughly the same. It is not realistic to simply identify and classify defects by visual inspection of the slices, and manual image segmentation and classification based on morphological characteristics require a large amount of manual labor. SUMMARY

[0004] In view of the defects of the prior art, the present application aims to provide a CFRP defect automatic identification and classification method and system based on XCT images, which solves the problem that the existing methods cannot automatically and efficiently identify defects due to the characteristics of CFRP images.

[0005] To achieve the above-mentioned purpose, in a first aspect, the present application provides a CFRP defect automatic identification and classification method based on XCT images, comprising:

[0006] Converting the original XCT three-dimensional image into a two-dimensional image sequence and inputting it into a trained semantic segmentation model to obtain the automatic identification and classification results of the carbon fiber composite defects; wherein,

[0007] The semantic segmentation model is trained by the following method:

[0008] The three-dimensional images of the relevant CFRP sample set are acquired by XCT, the sample set images are converted into two-dimensional image sequences, and filtering processing is performed;

[0009] All defects, substrates and background binary labels are segmented using threshold segmentation operation and Boolean logic operation, and the binary labels are stacked into three-dimensional labels through visualization;

[0010] Based on the three-dimensional morphological characteristics of the defects, the defects are classified;

[0011] The two-dimensional slices of the sample set original image are labeled using the classified three-dimensional defect labels to obtain a multi-channel image containing background, substrate and various defects;

[0012] After data enhancement of the multi-channel image, an enhanced training set is obtained;

[0013] The enhanced training set is input into the semantic segmentation network based on the convolutional neural network for training to obtain a trained semantic segmentation model.

[0014] Preferably, the two-dimensional image sequence converted to uint8 format is obtained by the following operation:

[0015] The maximum element x max in the image two-dimensional matrix X0∈R x×y is found, and the following operation is performed on each element in the matrix:

[0016]

[0017] Where [] is the floor symbol, x ij is the i-th row and j-th column matrix element.

[0018] Preferably, the filtering processing is specifically as follows:

[0019] If the noise scatter points are distributed or there are single-pixel bad points, median filtering is used;

[0020] If the noise is uniform and continuous, non-local mean filtering is used.

[0021] Preferably, the threshold value of the threshold segmentation operation is selected by the following steps:

[0022] (1) Determine the defect segmentation grayscale threshold range such that the difference between the upper limit and the lower limit is not more than 20, select a series of values g j not less than ten points in this range, and calculate the defect volume fraction V dj under the threshold value:

[0023]

[0024] (2) adopt four order polynomial to fit the relationship, wherein, a, b, c, d, e are fitting coefficients:

[0025]

[0026] (3) take the second order derivative of the four order polynomial, and take the gray value g t as the threshold value of defect segmentation.

[0027] Preferably, the three-dimensional morphological characteristics of the defects are used to classify the defects, specifically as follows:

[0028] If the sphericity is in the interval of 0.8-1.5, further judge the volume:

[0029] 1) the volume is greater than or equal to 1mm 3 , which belongs to a circular hole;

[0030] 2) the volume is less than 1mm 3 , which belongs to loose;

[0031] If the sphericity is not in the interval of 0.8-1.5, further judge the volume and the Feret shape factor:

[0032] 1) the Feret shape factor is less than 5, and the volume is greater than or equal to 1mm 3 , which belongs to a pore;

[0033] 2) the Feret shape factor is less than 5, and the volume is less than 1mm 3 , which belongs to fiber debonding;

[0034] 3) the Feret shape factor is greater than or equal to 5, and the Feret length-thickness ratio is less than 5, which belongs to delamination;

[0035] 4) the Feret shape factor is greater than or equal to 5, and the Feret length-thickness ratio is greater than or equal to 5, which belongs to a crack. Preferably, the volume is the volume represented by the voxels occupied by the defect, and the calculation formula is as follows:

[0036] V 3d = R 3 C

[0037] Wherein, R is the spatial resolution, and C is the total number of defect voxels;

[0038] The sphericity is used to describe the approximation degree of the target defect to a sphere, and the closer to 1 in value indicates that the defect is closer to a sphere, and the calculation formula is as follows:

[0039]

[0040] Wherein, A 3d is the surface area of the defect, and V 3d is the volume of the defect;

[0041] The Feret shape factor is the ratio of the maximum Feret diameter in the orthogonal plane to the minimum Feret diameter, and is greater than or equal to 1 in value.

[0042] Preferably, the semantic segmentation network of the convolutional neural network is a U-net model.

[0043] Preferably, the loss function in the training process is specifically as follows:

[0044]

[0045] Wherein, TP is true positive, FP is false positive, and FN is false negative.

[0046] Preferably, the method further comprises obtaining the content of each type of defect by counting voxels or pixels.

[0047] To achieve the above object, in a second aspect, the application provides a CFRP defect automatic identification and classification system based on XCT images, comprising: a processor and a memory; the memory is used for storing computer execution instructions; the processor is used for executing the computer execution instructions, so that the method of the first aspect is executed.

[0048] Overall, compared with the prior art, the above technical solutions conceived by the application have the following beneficial effects:

[0049] The application proposes a CFRP defect automatic identification and classification method and system based on XCT images, obtains three-dimensional features of defects through binarization three-dimensional label statistics classification, obtains two-dimensional features of defects through manual labeling based on three-dimensional labels of defects, and after fusing the three-dimensional features and the two-dimensional features of defects, combines the features with a deep learning method to obtain an efficient real defect identification and classification model, realizes simplicity and efficiency, and effectively improves CFRP defect identification precision. The application uses three-dimensional morphological parameters of defects when labeling a data set, avoids unreliability introduced by direct labeling on two-dimensional slices, and compared with ultrasonic detection and identification of defects, the method has higher authenticity and reliability in identifying defects. The application trains a semantic segmentation model for identifying and classifying CFRP defects based on a small amount of two-dimensional CT images, and uses the model to predict internal defects of similar materials, avoids developing a large amount of tedious and time-consuming manual image processing to obtain internal defect data of materials, and this saves a large amount of labor cost. BRIEF DESCRIPTION OF DRAWINGS

[0050] Figure 1 A CFRP defect automatic identification and classification method flowchart based on XCT images is provided in the application.

[0051] Figure 2The CT image after median filtering of the CFRP sample provided by the embodiment of the present application has a resolution of 0.7 microns.

[0052] Figure 3 The schematic diagram of the Ferrite diameter provided by the embodiment of the present application.

[0053] Figure 4 The schematic diagram of the multi-channel image synthesis process of the CFRP sample identification classification training set provided by the embodiment of the present application.

[0054] Figure 5 The comparison diagram of the prediction result and the manual segmentation effect provided by the embodiment of the present application. DETAILED DESCRIPTION

[0055] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application is further described in detail below in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0056] As shown in Figure 1 , the present application provides a CFRP defect automatic identification classification method based on XCT image, which comprises the following steps:

[0057] Step one, use XCT equipment to scan CFRP test piece, ensure that the defects are identified as much as possible, and perform three-dimensional reconstruction and image preprocessing on the obtained scanning image.

[0058] Obtain the three-dimensional image of the related CFRP sample set through XCT, convert the sample set image into a two-dimensional image sequence in uint8 format, and then perform filtering to filter noise.

[0059] The method for preprocessing the CFRP image comprises the following steps:

[0060] (1.1) In order to enable the image to be recognized by the neural network, the image is converted into uint8 format, and the maximum element x x×y in the two-dimensional matrix X0 of the image is found. max The following operation is performed on each element in the matrix:

[0061]

[0062] In the formula, [] is the down rounding symbol, and x ij is the element in the matrix.

[0063] (1.2) Perform filtering on the converted image to eliminate the influence of part of the noise. When noise scatter points are distributed or single-pixel bad points exist, median filtering is adopted, and when noise is uniform and continuous, non-local mean filtering is adopted.

[0064] Different filtering algorithms are used in different situations, which can effectively reduce the misrecognition caused by image noise in subsequent threshold segmentation operation.

[0065] In this embodiment, the internal defects of the CFRP sample with a resolution of 0.7 microns are automatically classified and recognized, and the image after median filtering processing is as shown in the figure. Figure 2

[0066] Step two, threshold segmentation operation is performed on all two-dimensional images to ensure that the inclusion and void type defects (holes, cracks, delamination, and fiber debonding) are segmented by the same threshold value, the binary label of the matrix and the image background is segmented by using Boolean operation, and the two-dimensional label is stacked into three-dimensional label through visualization.

[0067] The defect threshold segmentation method includes the following steps:

[0068] (2.1) The range of the defect segmentation gray threshold value is roughly determined, so that the difference between the upper limit and the lower limit is not more than 20, and a series of values g j are selected in this range, and the defect volume fraction V dj under the threshold value is calculated:

[0069]

[0070] (2.2) A fourth-order polynomial is used to fit the relationship:

[0071]

[0072] (2.3) The second derivative of the above fourth-order polynomial is calculated, and when the second derivative is 0, the threshold sensitivity of the defect is the lowest, and the gray value g t at this time is taken as the threshold value of the defect segmentation.

[0073]

[0074] Through the above steps, the image will be divided into high-intensity inclusion defects, matrix, void type defects and image background.

[0075] Step three, based on the three-dimensional morphological characteristics of the defects such as sphericity, Feret length-thickness ratio and Feret shape factor, the void type defects are manually classified into holes (circular holes, pores and loose), cracks, delamination and fiber debonding models.

[0076] The morphological characteristics in the application include surface area, volume, sphericity, Feret diameter, Feret length-thickness ratio and Feret shape factor.

[0077] (3.1) Surface area: describes the sum of the surface area of the voxels occupied by the defect.

[0078] ​

[0079] where R is the spatial resolution, i is the number of faces that the voxel contributes to the surface area, ranging from 0 to 6, C is the number of voxels that contribute to the surface area as face i. i

[0080] (3.2) Volume: describes the volume represented by the voxels that the defect occupies.

[0081] V = R 3d 3 C = R

[0082] where R is the spatial resolution, C is the total number of defect voxels.

[0083] (3.3) Sphericity: describes the degree of approximation of the target defect to a sphere, numerically closer to 1 means the defect is closer to a sphere.

[0084]

[0085] where A is the defect surface area, V is the defect volume. 3d 3d

[0086] (3.4) Feret diameter: the distance between two parallel tangent planes of the measured object, as shown in Figure 3

[0087] (3.5) Feret aspect ratio: the ratio of the maximum Feret diameter to the maximum Feret diameter in its orthogonal plane, numerically greater than or equal to 1.

[0088] (3.6) Feret shape factor: the ratio of the maximum Feret diameter in the orthogonal plane of the minimum Feret diameter to the minimum Feret diameter, numerically greater than or equal to 1.

[0089] The defect classification criteria are as follows:

[0090] If the sphericity is in the interval of 0.8-1.5:

[0091] Volume greater than or equal to 1 mm 3 , belongs to circular hole;

[0092] Volume less than 1 mm 3 , continuous, belongs to loose;

[0093] If the sphericity is not in the interval of 0.8-1.5:

[0094] Volume greater than 1 mm 3 , Feret shape factor less than 5, belongs to air hole;

[0095] Volume less than 1 mm 3 , Feret shape factor less than 5, belongs to fiber debonding;​​​​​

[0096] Feret shape factor greater than or equal to 5, and Feret length-thickness ratio less than 5, belongs to delamination;

[0097] Feret shape factor greater than or equal to 5, and Feret length-thickness ratio greater than or equal to 5, belongs to crack.

[0098] Step four, select no less than 50 slices as a training set, 10 slices in the training set as a validation set, and the profile of each type of defect model at the slice obtained in step three is used as a binary label and synthesized into a multi-channel image with the matrix and background, and the multi-channel image is required to be the same as the number of two-dimensional slices in the training set and one-to-one correspondence. The multi-channel image is enhanced by symmetrical, flip and other methods. The multi-channel image synthesis process is as shown in Figure 4 .

[0099] Step five, the multi-channel image obtained in step four and the binary label corresponding to step three, step two and step one are imported into a convolutional neural network such as a U-net model for training, and the corresponding training parameters are set. After the Dice loss is less than 0.005, the training is completed, and a semantic segmentation model capable of segmenting all defects of CFRP under the scanning accuracy is obtained.

[0100] The loss function Dice loss is defined as:

[0101]

[0102] In the formula, TP is true positive (predicted as positive, actual as true value), FP is false positive (predicted as positive, actual as non-true), and FN is false negative (predicted as negative, actual as true value).

[0103] Step six, the CT image of CFRP under similar scanning accuracy is imported into the model obtained in step four for prediction, and the image classification result is obtained: each voxel / pixel is assigned a label of a component, and the image should contain the prediction labels of background, matrix and each type of defect. Finally, the content of each type of defect is obtained by counting the voxels / pixels.

[0104] As shown in Figure 5 , the content calculated by the method of the application is compared with manual segmentation, and the error is not more than 0.5%.

[0105] Those skilled in the art will readily understand that the above description is only a preferred embodiment of the present application and is not intended to limit the present application. Any modification, equivalent replacement and improvement made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. An XCT image-based automatic recognition and classification method for CFRP defects, characterized in that, The method comprises the following steps: The original XCT three-dimensional image is converted into a two-dimensional image sequence, which is input into a trained semantic segmentation model to obtain the automatic recognition and classification result of the CFRP defect; wherein The semantic segmentation model is trained by the following method: Three-dimensional images of the relevant CFRP sample set are obtained by XCT, the sample set images are converted into two-dimensional image sequences, and filtering processing is performed; Threshold segmentation operation and Boolean logic operation are used to segment all defects, matrix and background binary labels, and the binary labels are stacked into three-dimensional labels through visualization; Based on the three-dimensional morphological characteristics of the defects, the defects are classified; The two-dimensional slices of the original sample set image are labeled using the classified three-dimensional defect labels to obtain a multi-channel image containing background, matrix and various defects; After data enhancement, the training set is obtained; The enhanced training set is input into a semantic segmentation network based on a convolutional neural network for training to obtain a trained semantic segmentation model; The two-dimensional image sequence in uint8 format is converted by the following operations: Find the maximum element in a two-dimensional matrix of images For each element in the matrix, do the following:​ wherein [ ] is a floor symbol, is the row of the column matrix element; The threshold value of the threshold segmentation operation is selected by the following steps: (1) determining a range of threshold values for the segmentation of the defects such that the difference between the upper and lower bounds does not exceed 20, selecting a series of values not less than ten points in this range , calculating the volume fraction of defects at this threshold value : (2) The relationship is fitted using a fourth order polynomial, where, are fitting coefficients: (3) taking the second derivative of the fourth order polynomial, and taking the gray value when the second derivative is 0 as the threshold value of defect segmentation as the threshold value of defect segmentation.

2. The method of claim 1, wherein, The filtering processing is as follows: If the noise scatter points are distributed or there are single-pixel bad points, median filtering is used; If the noise is uniform and continuous, non-local mean filtering is used.

3. The method of claim 1, wherein, Based on the three-dimensional morphological characteristics of the defects, the defects are classified as follows: If the sphericity is in the interval of 0.8-1.5, further judge the volume: 1) Volume greater than or equal to 1 mm 3 , which belongs to a circular hole; 2) less than 1 mm in volume 3 Loose; If the sphericity is not in the interval of 0.8-1.5, further judge the volume and the Feret shape factor: 1) Feret shape factor less than 5 and volume greater than or equal to 1 mm 3 , belongs to the group of pores; 2) Feret shape factor less than 5 and volume less than 1 mm 3 , which belongs to fiber debonding; 3) The Feret shape factor is greater than or equal to 5, and the Feret length-thickness ratio is less than 5, which belongs to delamination; 4) The Feret shape factor is greater than or equal to 5, and the Feret length-thickness ratio is greater than or equal to 5, which belongs to crack.

4. The method of claim 3, wherein, The volume is the volume represented by the voxels occupied by the defect, and the calculation formula is as follows: wherein, for spatial resolution, is the total number of defect voxels counted; The sphericity is used to describe the approximation degree of the target defect to a sphere, and the closer to 1 in value indicates that the defect is closer to a sphere, and the calculation formula is as follows: wherein is the surface area of the defect, is the volume of the defect; The Feret shape factor is the ratio of the maximum Feret diameter in the orthogonal plane to the minimum Feret diameter, and the value is greater than or equal to 1.

5. The method of claim 1, wherein, The semantic segmentation network of the convolutional neural network is a U-net model.

6. The method of claim 1, wherein, The loss function in the training process is as follows: Wherein, TP is true positive, FP is false positive, and FN is false negative.

7. The method according to any one of claims 1 to 6, characterized in that, The method further comprises obtaining the content of each type of defect by counting voxels or pixels.

8. An XCT image-based CFRP defect automatic identification classification system, characterized in that, The method comprises the following steps: A processor and a memory; The memory is used to store computer execution instructions; The processor is used to execute the computer execution instructions, so that the method of any one of claims 1 to 7 is executed.

Images