CT flaw detection auxiliary method and equipment based on instance segmentation model and storage medium

Abstract

The invention provides a CT flaw detection auxiliary method and equipment based on an instance segmentation model and a storage medium, and belongs to the technical field of flaw detection measurement. The method is used for calculating the internal section defect size of the engine grain, and comprises the following steps: establishing an instance segmentation model; acquiring a section CT scanning image of the engine grain; analyzing the section CT scanning image through the instance segmentation model to obtain a defect type; calculating the section defect size according to the defect type; and generating an internal section defect evaluation result of the engine grain according to the section defect size and the defect type. According to the method, the instance segmentation model is established based on the YOLOv8 network, and the defect types are judged by utilizing the lightweight characteristic and the characteristic of integrating image segmentation, defect detection and target positioning, so that corresponding size calculation is conveniently carried out according to segmented images for different defect types; therefore, the precision and efficiency of defect size calculation are effectively improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of flaw detection and measurement, and in particular relates to a CT flaw detection auxiliary method, device and storage medium based on an instance segmentation model. Background Art

[0002] The defect location technology through machine vision has a wide range of applications in the industrial field. It is usually used to locate and detect defects on the external surface of the structure. However, since products such as engine grains are located inside the structure, it is difficult to achieve defect location and detection through machine vision defect location technology. In this regard, obtaining product CT images through CT flaw detection for defect detection can well solve the above problem.

[0003] At present, although the traditional CT flaw detection method can realize defect detection and size calculation to a certain extent, it can be seen from the Chinese invention patent with application number "201410428442.6" and the Chinese invention patent with application number "202310399158.X" that CT flaw detection requires artificial design of defect features of CT images, and it takes a long time to calculate the defect size, and the detection accuracy and efficiency are poor. Summary of the invention

[0004] The technical problem to be solved by the present invention is how to improve the accuracy and efficiency of defect size calculation through CT. In view of the shortcomings of the prior art, a CT flaw detection auxiliary method, device and storage medium based on an instance segmentation model are provided.

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is: In a first aspect, the present invention provides a CT flaw detection auxiliary method based on an instance segmentation model, which is used for calculating the size of internal cross-sectional defects of an engine grain. The CT flaw detection auxiliary method based on an instance segmentation model comprises the following steps: S1. Establish instance segmentation model; S2, obtaining a cross-sectional CT scan image of the engine charge; S3, analyzing the cross-sectional CT scan image by using the instance segmentation model to obtain the defect type; S4, calculating the cross-sectional defect size according to the defect type; S5. Generate an internal cross-sectional defect evaluation result of the engine grain according to the cross-sectional defect size and the defect type.

[0006] Compared with the prior art, the beneficial effects of the CT flaw detection auxiliary method based on the instance segmentation model of the present invention include: processing the obtained cross-sectional CT scan image of the engine charge based on the instance segmentation model, wherein an instance segmentation model is established based on the YOLOv8 network, and the lightweight characteristics of the YOLOv8 network and the characteristics of being able to integrate image segmentation, defect detection and target positioning are utilized, which not only facilitates the edge computing deployment of the instance segmentation model, thereby reducing the integration cost of the instance segmentation model, but also can segment and analyze the cross-sectional CT scan image, quickly generate evaluation results, and effectively improve the accuracy and efficiency of image segmentation and defect analysis; on this basis, the cross-sectional CT scan image is analyzed by using the instance segmentation model, so as to realize the determination of the defect type, and then the corresponding cross-sectional defect size calculation is performed according to the defect type, so as to facilitate the corresponding size calculation according to the segmented image for different defect types, thereby effectively improving the accuracy and efficiency of the defect size calculation.

[0007] Optionally, the defect type includes at least one of a crack defect and a debonding defect, and S4 specifically includes the following steps: S41, if the defect type is a crack defect, calculating the annular length l and the axial length d of the crack defect according to the cross-sectional CT scan image; S42, if the defect type is a debonding defect, obtaining all positions of the debonding defect according to the cross-sectional CT scan image; S43. Calculate the debonding areas S of the debonding defects at different positions according to the position information of the debonding defects.

[0008] Optionally, the S41 specifically includes the following steps: S411, obtaining crack defects of all cross-sectional layers on the cross-sectional CT scan image; S412, sequentially calculating the longest value l of the crack defect in each layer i , where i is a natural number and is the label of the cross-section layer; S413, according to l=max(l i ) obtaining the annular length l of the crack defect; S414, sequentially calculating the axial depth d of the crack defect in each layer i ; S415, the axial depth d of the crack defect of all cross-section layers i The axial length d of the crack defect is obtained by adding them together.

[0009] Optionally, the S42 specifically includes the following steps: S421, obtaining debonding defects of all cross-sectional layers on the cross-sectional CT scan image; S422, obtaining the contour number of the debonding defect in each of the cross-section layers; S423, overlapping the debonding defects of two adjacent cross-section layers to obtain the contour number of the debonding defects after the overlapping process; S424, judging whether the number of contours of the debonding defect after the overlapping process is greater than the number of defect contours of a single cross-section layer; S425, if yes, it is determined that the positions of the debonding defects on the two overlapping cross-section layers are different, and the position information of the debonding defects is recorded; S426: If not, determine that the positions of the debonding defects on the two overlapping cross-section layers are the same, record the position information of the debonding defects and repeat S423.

[0010] Optionally, the S43 specifically includes the following steps: S432, sequentially calculating the length L of each debonding defect contour of each cross-section layer t , where t is a natural number and is the label of the cross-section layer; S433, sequentially calculating the axial depth D of each of the debonding defect profiles of each of the cross-section layers t ; S434, through S t =L t *D t The area S of each debonding defect of each cross-section layer is calculated. t ; S435, according to the location information of the debonding defect and the S t , and add up to get the total area S of the debonding defect at each position.

[0011] Optionally, the S1 specifically includes the following steps: S11, build the initial instance segmentation model using YOLOv8 as the baseline network; S12, optimizing the initial instance segmentation model in a preset optimization method, wherein the preset optimization method includes at least one of improving the backbone network by CBAM, improving the neck network by ASF, and improving the segmentation detection head by adding a feature extraction process; S13, inputting the training data set of the cross-sectional CT scan image into the initial instance segmentation model, and performing accelerated optimization through edge computing technology to obtain the instance segmentation model; S14: Perform a pruning operation on the optimized instance segmentation model and deploy the instance segmentation model to an edge computing terminal.

[0012] Optionally, the S5 specifically includes: S51, selecting a preset defect standard according to the defect type, and determining whether the cross-sectional defect size meets the preset defect standard; S52, if yes, generate a conclusion that a defect fault exists; S53. If not, generate a conclusion that the defect meets the acceptance criteria; S54. Generate the internal section defect evaluation result of the engine grain according to the conclusion that the defect failure exists and the conclusion that the defect meets the acceptance standard.

[0013] Optionally, after S54, the method further includes: S55, selecting a preset number of images from the cross-sectional CT scan image by a preset defect degree sorting algorithm, and attaching them to the internal cross-sectional defect evaluation result of the engine grain to form an internal cross-sectional defect evaluation report of the engine grain; And / or, S56, annotating defects on the cross-sectional CT scan image, uploading the annotated cross-sectional CT scan image to a folder of a preset storage model, and naming the folder according to the product number and upload time.

[0014] In a second aspect, the present invention also provides a CT flaw detection auxiliary device based on an instance segmentation model, comprising a human-computer interaction terminal, a support hinge and a suitcase, wherein the two ends of the support hinge are respectively connected to the human-computer interaction terminal and the suitcase so that the human-computer interaction terminal can move in and out of the suitcase, and the human-computer interaction terminal is used to communicate with a CT host computer to implement the CT flaw detection auxiliary method based on the instance segmentation model as described above.

[0015] Compared with the prior art, the beneficial effects of the CT flaw detection auxiliary device based on the instance segmentation model of the present invention are the same as the beneficial effects of the CT flaw detection auxiliary method based on the instance segmentation model as described above, and will not be repeated here.

[0016] In a third aspect, the present invention further provides a computer storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the CT flaw detection auxiliary method based on the instance segmentation model as described above is implemented.

[0017] Compared with the prior art, the beneficial effects of the computer storage medium of the present invention are the same as the beneficial effects of the CT flaw detection auxiliary method based on the instance segmentation model as described above, and will not be repeated here. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The present invention is further described in detail below in conjunction with the accompanying drawings.

[0020] Figure 1: Flow chart of the CT flaw detection auxiliary method based on the instance segmentation model in an embodiment of the present invention; Figure 2 : Figure 1 Sub-flow chart of S4 in the figure; Figure 3 : Figure 2 The sub-flow chart of S41; Figure 4 : Figure 2 The sub-flow chart of S42; Figure 5 : Figure 2 The sub-flow chart of S43; Figure 6 : Figure 1 Sub-flow chart of S1 in the figure; Figure 7 : Figure 1 Sub-flow chart of S5 in the figure; Figure 8 : Workflow diagram of the improved segmentation detection head in the embodiment of the present invention; Fig. 9 : A schematic diagram of the structure of a CT flaw detection auxiliary device based on an instance segmentation model in an embodiment of the present invention; Fig.10 : Communication connection structure diagram of the human-computer interaction terminal in an embodiment of the present invention.

[0021] Among them, 1-human-computer interaction terminal, 2-support hinge, 3-suitcase. DETAILED DESCRIPTION

[0023] In order to better understand the present invention, the content of the present invention is further clearly set forth in conjunction with the embodiments below, but the protection content of the present invention is not limited to the following embodiments. In the following description, a large number of specific details are provided in order to provide a more thorough understanding of the present invention. However, it is obvious to those skilled in the art that the present invention can be implemented without one or more of these details.

[0024] It should be noted that the Z axis in the drawings represents the vertical direction, that is, the up and down position, and the positive direction of the Z axis represents the top, and the reverse direction of the Z axis represents the bottom; the Y axis in the drawings represents the horizontal direction, and is designated as the front and back position, and the positive direction of the Y axis represents the front side, and the reverse direction of the Y axis represents the back side; the X axis in the drawings represents the left and right position, and the positive direction of the X axis represents the right side, and the reverse direction of the X axis represents the left side. It should also be noted that the aforementioned Z axis, Y axis, and X axis are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation to the present invention.

[0025] The term "including" and its variations used in this document are open inclusions, that is, "including but not limited to"; the term "based on" means "based at least in part on"; the term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one other embodiment"; the term "some embodiments" means "at least some embodiments"; the term "optionally" means "optional embodiments". The relevant definitions of other terms will be given in the following description. It should be noted that the concepts of "first", "second", etc. mentioned in the present invention are only used to distinguish different devices, modules or units, and are not used to limit the order or interdependence of the functions performed by these devices, modules or units.

[0026] It should be noted that the modifications of "one" and "plurality" mentioned in the present invention are illustrative rather than restrictive, and those skilled in the art should understand that, unless otherwise clearly indicated in the context, it should be understood as "one or more".

[0027] In the first aspect, an embodiment of the present invention provides a CT flaw detection auxiliary method based on an instance segmentation model, which is used for calculating the internal cross-sectional defect size of an engine grain. The CT flaw detection auxiliary method based on an instance segmentation model includes the following steps: S1. Establishing an instance segmentation model; S2. Obtaining a cross-sectional CT scanning image of the engine grain; S3. Analyzing the cross-sectional CT scanning image through the instance segmentation model to obtain the defect type; S4. Calculating the cross-sectional defect size according to the defect type; S5. Generating an internal cross-sectional defect evaluation result of the engine grain according to the cross-sectional defect size and the defect type.

[0028] Specifically, the instance segmentation model is established based on the YOLOv8 network. The YOLOv8 network consists of a backbone network (Backbone), a neck network (Neck) and a detection head (Head). The instance segmentation model can not only segment the collected images, separate the foreground and background, and extract the contour information and grayscale features of the foreground to improve the accuracy of the image, but also identify defects in the candidate areas. Different defects have different shape characteristics. After algorithm calculation, different defects can be identified.

[0029] In this embodiment, Figure 1As shown, the cross-sectional CT scan image of the engine grain is processed based on the instance segmentation model, wherein the instance segmentation model is established based on the YOLOv8 network, and the lightweight characteristics of the YOLOv8 network and the characteristics of being able to integrate image segmentation, defect detection and target positioning are utilized, which not only facilitates the edge computing deployment of the instance segmentation model, thereby reducing the integration cost of the instance segmentation model, but also can segment and analyze the cross-sectional CT scan image, quickly generate evaluation results, and effectively improve the accuracy and efficiency of image segmentation and defect analysis; on this basis, the instance segmentation model is used to analyze the cross-sectional CT scan image, thereby realizing the determination of the defect type, and then the corresponding cross-sectional defect size calculation is performed according to the defect type, which facilitates the corresponding size calculation according to the segmented image for different defect types, thereby effectively improving the accuracy and efficiency of the defect size calculation.

[0030] Optionally, the defect type includes at least one of a crack defect and a debonding defect, and S4 specifically includes the following steps: S41. If the defect type is a crack defect, calculate the annular length l and the axial length d of the crack defect according to the cross-sectional CT scanning image; S42. If the defect type is a debonding defect, obtain all positions of the debonding defect according to the cross-sectional CT scanning image; S43. According to the position information of the debonding defect, calculate the debonding area S of the debonding defect at different positions.

[0031] In this optional embodiment, in order to improve the pertinence of defect size calculation, as shown in FIG. Figure 2 As shown in the figure, if the defect detection result obtained by the instance segmentation model after performing image segmentation on the cross-sectional CT scan image is a crack defect, the annular length l and axial length d of the crack defect can be calculated according to the cross-sectional CT scan image, so as to accurately calculate the size of the crack defect from the annular and axial directions, so as to facilitate the subsequent accurate judgment of whether the internal structure of the engine grain meets the acceptance criteria according to the calculation results; on this basis, if the defect detection result obtained by the instance segmentation model after performing image segmentation on the cross-sectional CT scan image is a debonding defect, all positions of the debonding defect can also be obtained according to the cross-sectional CT scan image. Since there may be multiple debonding defects, debonding defects at different positions need to be judged separately. Therefore, the debonding area S of the debonding defects at different positions can be calculated according to the position information of the debonding defect, so as to finally generate an accurate debonding fault conclusion according to the calculated debonding area S.

[0032] Optionally, S41 specifically includes the following steps: S411, obtaining crack defects of all cross-sectional layers on the cross-sectional CT scan image; S412, sequentially calculating the longest value l of the crack defect of each layer; i , where i is a natural number and is the label of the cross-section layer; S413, according to l=max(l i) obtain the annular length l of the crack defect; S414, calculate the axial depth d of each layer of crack defect in sequence i ; S415, the axial depth d of the crack defects of all cross-section layers i The addition gives the axial length d of the crack defect.

[0033] In this optional embodiment, in order to ensure the calculation accuracy of crack defects, Figure 3 As shown in the figure, in the process of calculating the size of crack defects, firstly, the crack defects of all cross-section layers are obtained according to the obtained cross-section CT scanning images, so as to ensure that the final crack defect calculation result will not be inaccurate due to the omission of the crack defects of a certain cross-section layer; the longest value l of each layer of crack defects can be calculated according to the order of the cross-section layers. i , where i is a natural number and is the label of the cross-section layer. For example, the longest value of the crack defect on the first cross-section layer image is l1. At the same time, the axial depth d of each crack defect is calculated according to the order of the cross-section layers. i For example, the axial depth of the crack defect on the first cross-sectional layer image is d1. This setting can not only facilitate the calculation according to l=max(l i ) to calculate the annular length l of the crack defect and the axial depth d of all crack defects i The axial length d is obtained by addition calculation, which enables accurate calculation of the crack defect size. In addition, a three-dimensional model of the crack defect can be constructed from the axial and circumferential dimensions according to the order of the cross-section layers, further improving the detection accuracy of the crack defect.

[0034] It should be noted that when calculating the length of crack defects, there may be multiple crack defects on the same cross-section layer. In this case, the number of contours of the crack defects on the cross-section layer can be obtained first, and then the images of two adjacent cross-section layers can be overlapped in sequence. When the number of contours after overlap does not increase, it means that the crack defects on the two cross-section layers are the same crack defects corresponding to each other. If the number of contours changes, it means that the types of crack defects on the two cross-section layers are different. In this case, the longest value l of different types of crack defects on each cross-section layer can be calculated separately. i Finally, according to the longest value l of each crack defect on all cross-section layers i Comparison and l=max(l i ), to obtain the annular length l of the crack defect.

[0035] Optionally, S42 specifically includes the following steps: S421, obtaining debonding defects of all cross-sectional layers on the cross-sectional CT scan image; S422, obtaining the contour number of debonding defects in each cross-sectional layer; S423, overlapping the debonding defects of two adjacent cross-sectional layers to obtain the contour number of debonding defects after the overlapping process; S424, judging whether the contour number of debonding defects after the overlapping process is greater than the defect contour number of a single cross-sectional layer; S425, if so, judging that the positions of the debonding defects on the two overlapping cross-sectional layers are different, and recording the position information of the debonding defects; S426, if not, judging that the positions of the debonding defects on the two overlapping cross-sectional layers are the same, recording the position information of the debonding defects and repeating S423.

[0036] In this optional embodiment, since the sizes of the debonding defects at different positions need to be calculated separately, before calculating the sizes of the debonding defects, it is necessary to first count the location information of the debonding defects. Specifically, Figure 4 As shown, first, the debonding defects of all section layers on the section CT scan image are obtained, and the contour number of the debonding defects on the section layer is obtained, so that the number of debonding defects on different section layers can be obtained. Then, the debonding defects of two adjacent section layers are overlapped to obtain the contour number of the debonding defects after the overlapping process. At this time, when the contour number of the debonding defects after the overlapping process is greater than the defect contour number of a single section layer, it means that the debonding defect contours of the two adjacent section layers do not belong to the same debonding defect, that is, the two debonding defect contours belong to debonding defects at two positions respectively. Correspondingly, when the contour number of the debonding defects after the overlapping process is equal to the defect contour number of a single section layer, it means that there is an overlapping area between the debonding defect contours on the two section layers, and they belong to the debonding defects at the same position. In this way, the two adjacent section layers are overlapped in sequence, so that the number and position of the debonding defects can be quickly and accurately determined, which is convenient for the subsequent calculation of the debonding defect size.

[0037] Optionally, S43 specifically includes the following steps: S432, sequentially calculating the length L of each debonding defect contour of each cross-section layer t , where t is a natural number and is the label of the cross-section layer; S433, sequentially calculating the axial depth D of each debonding defect profile of each cross-section layer t ; S434, through S t =L t *D t Calculate the area S of each debonding defect in each cross-section layer t ; S435, according to the location information of the debonding defect and S t , and add up to get the total area S of the debonding defect at each location.

[0038] In this optional embodiment, if Figure 5As shown in Figure 1, in order to ensure the calculation accuracy of the debonding defect, in the process of calculating the size of the debonding defect, the length L of the debonding defect contour of all cross-sectional layers is first calculated in sequence according to the obtained cross-sectional CT scan image. t , where t is a natural number and is the label of the cross-section layer. For example, the length L1 of the debonding defect contour on the first cross-section layer image is calculated. At the same time, the axial depth D of the debonding defect contour of each layer is calculated according to the order of the cross-section layers. t For example, the axial depth of the debonding defect contour on the first cross-sectional layer image is D1. This setting is not only convenient for t =L t *D t To calculate the area S of each debonding defect in each cross-section layer t , to achieve accurate calculation of the size of each debonding defect, and according to the location information of the debonding defect and S t The total area S of the debonding defect at each position is obtained by adding up to ensure the detection accuracy of the debonding defect.

[0039] For example, a shooting start point A and an end point B are set on the engine axis. Starting from the start point A, a CT image is taken at every distance m, m>0, until the end point B. Starting from the first image, each image is recorded as image 1, image 2, image 3...image n, n≥1. The debonding area is calculated as follows: Figure 1Start to perform instance segmentation reasoning on the entire set of CT images, extract all debonding defect mask images in each image, record all debonding in image 1 as mask11, mask12...maskjk, j=1, k≥1, record all debonding in image 2 as mask21, mask22...maskjk, j=1, k≥1, and so on; use "mask comparison" to determine the areas of the same debonding defect in two adjacent images in turn, and form a list, such as [mask11, mask23, mask32], indicating that mask1 in image 1, mask3 in image 2, and mask2 in image 3 belong to the same debonding defect. Another example is [mask12, mask21, mask22, mask34], indicating that mask2 in image 1, mask1 and mask2 in image 2, and mask4 in image 3 belong to the same debonding defect; calculate the debonding area. For the debonding defects [mask11, mask23, mask32], the lengths of the three debondings are recorded as L11, L23, and L32 respectively, the axial depth of mask11 is recorded as D12, the axial depth of mask23 is recorded as D23, and the axial depth of mask32 is recorded as 0, then the debonding area is L11*D12+L23*D23+L32*0. For the debonding defects [mask12, mask21, mask22, mask34], the debonding length in image 2 is L21+L22; Mask comparison: Each debonding mask derived by instance segmentation reasoning has a unique contour. Take a debonding defect mask from each of the two adjacent CT images for overlapping processing. If the number of image contours after overlapping is still 1, it means that the two masks have overlapping areas and belong to the same debonding defect. Otherwise, it means that the two mask images do not have overlapping areas and do not belong to the same debonding defect.

[0040] Optionally, S1 specifically includes the following steps: S11, constructing an initial instance segmentation model using YOLOv8 as the baseline network; S12, optimizing the initial instance segmentation model using a preset optimization method, wherein the preset optimization method includes improving the backbone network through CBAM, improving the neck network through ASF, and improving at least one of the segmentation detection head by adding a feature extraction process; S13, inputting the training data set of the cross-sectional CT scan image into the initial instance segmentation model, and performing accelerated optimization through edge computing technology to obtain an instance segmentation model; S14, performing pruning operations on the optimized instance segmentation model, and deploying the instance segmentation model to the edge computing terminal.

[0041] Specifically, CBAM (Convolutional Block Attention Module) is a lightweight attention mechanism module that can enhance the feature representation ability in convolutional neural networks. It guides the learning process of feature maps by calculating the attention in the spatial dimension and the attention in the channel dimension respectively, and fuses these two attention maps, thereby improving the performance of the model; ASF (Attentional Scale Sequence Fusion) is a method for processing sequence data. It can fuse sequence information from multiple scales through attention mechanisms of different scales, thereby enhancing the model's ability to understand complex sequence patterns. It is suitable for processing sequence data containing multi-level information; Figure 8 As shown in the figure, improving the segmentation detection head by adding a feature extraction process specifically means that before the segmentation detection head predicts the feature map, Stem feature extraction is performed. Stem consists of two 3×3 convolutional layers, and then a series of convolution operations can be used to increase the feature dimension of the input data, so that subsequent layers can more easily capture complex features. At the same time, multiple small convolution kernels can extract different types of local features and enhance the model's ability to perceive details.

[0042] In this optional embodiment, if Figure 6 As shown, first, an initial instance segmentation model is constructed using YOLOv8 as a baseline network, specifically a YOLOv8-seg semantic segmentation model; then, the initial instance segmentation model is optimized in a preset optimization manner, wherein the preset optimization manner includes at least one of improving the backbone network by CBAM, improving the neck network by ASF, and improving the segmentation detection head by adding a feature extraction process. For example, the feature representation capability enhanced by adding a CBAM attention module to the backbone network of YOLOv8 enables CNN to more accurately identify and classify objects in the image, thereby improving the overall performance of the model, and the improved ASF is used to optimize the neck network of the YOLOv8 network to improve the overall performance of the network in detecting and segmenting smaller targets with small data sets, large amounts of information, and position correlation, that is, the Conv layer and the C2f layer are added to the ASF neck network to form an improved neck network structure ASFv8, and the robustness and generalization ability of the model can be improved by improving the segmentation detection head by adding a feature extraction process. The workflow of the segmentation detection head can be seen in FIG. Figure 8; Then, the training data set of the cross-sectional CT scan image is input into the initial instance segmentation model, that is, the engine propellant scanning tomographic image data set containing defects is collected through industrial CT, and the model is input for training to obtain the prediction model. The improved performance of the model is verified by comparing the accuracy with the original YOLOv8-seg model to ensure the subsequent use effect of the model. At the same time, the edge computing technology is used for accelerated optimization to obtain the instance segmentation model to ensure the efficiency of model construction; finally, the optimized instance segmentation model is pruned so that its network structure can efficiently adapt to the edge computing terminal, which is convenient for the subsequent deployment of the instance segmentation model to the edge computing terminal.

[0043] Optionally, S5 specifically includes: S51, selecting a preset defect standard according to the defect type, and determining whether the cross-sectional defect size meets the preset defect standard; S52, if so, generating a conclusion that a defect fault exists; S53, if not, generating a conclusion that the defect meets the acceptance standard; S54, generating an internal cross-sectional defect evaluation result of the engine propellant based on the conclusion that a defect fault exists and the conclusion that the defect meets the acceptance standard.

[0044] Specifically, the preset defect standards include crack defect standards and debonding defect standards, wherein the crack defect standard is that the circumferential length l>12mm or the axial depth d>12mm, and the debonding defect standard is S>40cm 2 .

[0045] In this optional embodiment, if Figure 7 As shown, in order to ensure the accuracy of the internal cross-sectional defect evaluation results of the engine grain, a preset defect standard is set. The preset defect standard can be selected according to the defect type, so that the calculated cross-sectional defect size is compared with the preset defect standard. When the preset defect standard is met, a conclusion that the defect meets the acceptance standard can be generated to ensure that the engine grain meets the production requirements. Correspondingly, when the preset defect standard is not met, a conclusion of defect failure can be generated, which is convenient for adjustment according to the specific content of the defect failure. With this setting, the code instantiation of the evaluation standard constraint conditions can be completed, and the calculated defect size can be effectively judged, which completely replaces manual interpretation and provides accurate conclusion recommendations.

[0046] Optionally, after S54, it also includes: S55, selecting a preset number of images from the cross-sectional CT scan images through a preset defect degree sorting algorithm, and attaching them to the internal cross-sectional defect evaluation results of the engine grain to form an internal cross-sectional defect evaluation report of the engine grain; and / or, S56, marking defects on the cross-sectional CT scan images, uploading the marked cross-sectional CT scan images to a preset storage model folder, and naming the folder according to the product number and upload time.

[0047] In this optional embodiment, in order to improve the efficiency of issuing the internal section defect evaluation report of the engine grain, as shown in FIG. Figure 7 As shown, by utilizing the function of the original CT flaw detection work card and designing a defect degree sorting algorithm, a preset number of images can be selected from the cross-sectional CT scan image through the preset defect degree sorting algorithm. For example, 4 typical flaw detection defect annotation images (including original CT flaw detection images) are selected as comparison samples and attached to the AI ​​inspection report, which greatly facilitates archiving and eliminates the need for operators to re-select typical defect images, thereby improving the efficiency of operators in issuing flaw detection reports.

[0048] In this optional embodiment or other optional embodiments of the present invention, in order to avoid operation omission, as Figure 7 As shown in the figure, for the automatic push of defect annotated images, without changing the original operating habits, the defect images annotated by CT flaw detection can be automatically pushed to a fixed folder of the CT host computer, and automatically named by time and product number, so that the operator can receive the pushed data "unconsciously" and avoid omissions in operation.

[0049] It should be noted that the instance segmentation model of the present invention shows a significant improvement in accuracy in target detection and segmentation tasks compared to the original YOLOv8 model. In particular, the improved model has made a breakthrough in segmentation accuracy. For example, the instance segmentation model of the present invention and the original YOLOv8 model are respectively tested on two data sets, one of which is a public crack segmentation data set officially provided by Ultralytics YOLO, which contains 3717 training images with a size of 411×411 pixels, and the other is an engine grain data set based on industrial CT scanning, which contains 1952 training images with a size of 2048×2048 pixels. Among them, through the analysis and comparison of the detection data, it can be found that the mAP segmentation accuracy of the original YOLOv8-seg model is usually slightly lower than the detection accuracy, while the instance segmentation model of the present invention has greatly improved the segmentation performance. In addition, comparing the accuracy improvement of the instance segmentation model and the original YOLOv8 model in the two datasets, when processing small-sized targets (small-sized cracks on large images) and smaller datasets (the number of some categories in the dataset is less than 500), the instance segmentation model shows more significant performance improvement in the accuracy of feature information extraction from large-sized images.

[0050] For example, on the one hand, the Roboflow crack segmentation dataset (Crack-seg) officially provided by Ultralytics YOLO is used as the experimental verification dataset of the instance segmentation model. The dataset contains 3717 training sets, 200 verification sets and 112 test sets. The image size of the dataset is 411×411 pixels. The computer environment is an Nvida GeForce RTX4090D graphics card. The training sets are respectively input into the original YOLOv8-seg model and the improved YOLOv8-ACE model for 500 rounds of iterative training. The model adopts the minimum model structure scale ratio "YOLOv8n". In the model training parameters, batchsize is set to 16, workers are set to 8, the input image size is 640, the data enhancement function is turned on, the optimizer selects SGD gradient descent, and other parameters are selected by default. It can be concluded from the final data that the instance segmentation model of the present invention has a significant improvement in segmentation effect than the original YOLOv8-seg model.

[0051] On the other hand, the CT scan cross-sectional images of a certain type of rocket engine grain were used as experimental samples for this model. The cross-sectional images of the engine grain containing defects were collected by industrial CT tomography technology, with a total of 2,636 CT scan images, and the image size was 2048*2048 pixels. Among them, there were 1515 images of debonding defects, 249 images of axial crack defects, and 1893 images of circumferential crack defects. The data set was randomly divided into 1952 training sets, 480 validation sets, and 204 test sets. The computer training environment was Nvidia GeForce RTX 4090D. The training sets were input into the original YOLOv8-seg model and the instance segmentation model for training, and the model network structure adopted the minimum model structure scale ratio "YOLOv8n". Due to the limitation of computer performance, the input image size was set to 1280 (after testing, the improvement effect of using 2048 pixels compared to 1280 pixels was limited, and the increase between 640 pixels and 1280 pixels had a greater impact on the accuracy improvement effect). The optimizer uses gradient descent (SGD), and the batch size and workers are 8. After 500 iterations, the loss function gradually converges. According to the performance and prediction effect of the two models finally obtained, it can be seen that the instance segmentation model of the present invention has significant improvements in re-labeling, missing labels, and segmentation accuracy compared with the original YOLOv8-seg model.

[0052] It should be noted that the YOLOv8 model is known for its lightweight characteristics and contains 5 scales of different scale ratios. Even when the model with the largest model structure scale ratio Yolov8n is used, its accuracy remains above 0.8, and the model weight size is only 7MB. Even the model with the largest model structure scale ratio Yolov8x has a weight size of only 140MB, which can be easily integrated into edge computing terminals. In sharp contrast, the deployment of large network models such as Mask-RCNN and U-net often requires expensive high-performance computing resources. Therefore, the instance segmentation model based on the improvement of the YOLOv8 model of the present invention has significant cost advantages in industrial applications.

[0053] In a second aspect, the present invention also provides a CT flaw detection auxiliary device based on an instance segmentation model, comprising a human-computer interaction terminal 1, a support hinge 2 and a suitcase 3, wherein both ends of the support hinge 2 are respectively connected to the human-computer interaction terminal 1 and the suitcase 3 so that the human-computer interaction terminal 1 can move in and out of the suitcase 3, and the human-computer interaction terminal 1 is used to communicate with a CT host computer to implement the above-mentioned CT flaw detection auxiliary method based on the instance segmentation model.

[0054] Specifically, Fig.10 As shown, the human-computer interaction terminal 1 includes a touch screen display and a built-in CT flaw detection auxiliary system based on an instance segmentation model. The CT flaw detection auxiliary system based on an instance segmentation model includes a human-computer interaction module, a CT digital image defect recognition module, a defect logic discrimination module and a data information access module, wherein the touch screen display can be used to receive human-computer interaction instruction injection, and the human-computer interaction module, the CT digital image defect recognition module, the defect logic discrimination module and the data information access module can work together, and according to the human-computer interaction instruction, they are respectively communicated and connected with the CT flaw detection host computer and the EMS system through a network cable and a switch, so as to facilitate the cooperation with the CT flaw detection equipment and ensure the efficient and accurate defect size calculation process of CT flaw detection.

[0055] In this embodiment, Fig. 9 As shown, the human-computer interaction terminal 1 can receive human-computer interaction instructions and communicate with the CT host computer, so as to realize the above-mentioned CT flaw detection auxiliary method based on the instance segmentation model. At the same time, the human-computer interaction terminal 1 can be moved into and out of the suitcase 3 through the support hinge 2, which is convenient for the rapid movement of the CT flaw detection auxiliary equipment based on the instance segmentation model, and is easy to carry. It can also be cascaded with the flaw detection host computers of different products through the network cable, realizing the "non-destructive installation" of the intelligent auxiliary decision-making function, and facilitating the expansion of the non-destructive detection analysis function for different objects or methods, so as to form a universal non-destructive detection intelligent auxiliary decision-making device.

[0056] It should be noted that in order to solve the time synchronization problem between the CT flaw detection auxiliary system based on the instance segmentation model and the CT flaw detection host computer, the NTP synchronization technology was introduced to achieve cross-system time synchronization, effectively solving the problem of inaccurate time recording in flaw detection reports. At the same time, it standardized the naming format of introducing timestamps to autonomously annotate pictures of flaw detection defects, ensuring the traceability of data.

[0057] It should be noted that the CT flaw detection auxiliary equipment based on the instance segmentation model of the present invention is a portable design with high flexibility and simplicity. It can be carried with you for rapid deployment and use in various working conditions. Its simple assembly process and portability enable the detection project to be easily carried out in various on-site environments, greatly improving the efficiency and adaptability of flaw detection and analysis work. At the same time, this technology can also be extended to the intelligent upgrade process of digital X-ray photography system (Digital Radiography, DR) equipment to form an intelligent DR auxiliary decision system, realizing portable real-time rapid auxiliary analysis of DR detection results.

[0058] In a third aspect, the present invention further provides a computer storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the above-mentioned CT flaw detection auxiliary method based on the instance segmentation model is implemented.

[0059] The technical effect of the computer storage medium in this embodiment is similar to the technical effect of the above-mentioned CT flaw detection auxiliary method based on the instance segmentation model, and will not be repeated here.

[0060] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A CT flaw detection auxiliary method based on an instance segmentation model, characterized in that: The CT flaw detection auxiliary method based on the instance segmentation model for calculating the size of the internal section defect of the engine grain comprises the following steps: S1. Establish instance segmentation model; S2, obtaining a cross-sectional CT scan image of the engine charge; S3, analyzing the cross-sectional CT scan image by using the instance segmentation model to obtain the defect type; S4, calculating the cross-sectional defect size according to the defect type; S5. Generate an internal cross-sectional defect evaluation result of the engine grain according to the cross-sectional defect size and the defect type.

2. The CT flaw detection auxiliary method based on the instance segmentation model according to claim 1 is characterized in that: The defect type includes at least one of a crack defect and a debonding defect, and S4 specifically includes the following steps: S41, if the defect type is a crack defect, calculating the annular length l and the axial length d of the crack defect according to the cross-sectional CT scan image; S42, if the defect type is a debonding defect, obtaining all positions of the debonding defect according to the cross-sectional CT scan image; S43. Calculate the debonding areas S of the debonding defects at different positions according to the position information of the debonding defects.

3. The CT flaw detection auxiliary method based on the instance segmentation model according to claim 2, characterized in that: The S41 specifically includes the following steps: S411, obtaining crack defects of all cross-sectional layers on the cross-sectional CT scan image; S412, sequentially calculating the longest value l of the crack defect in each layer i , where i is a natural number and is the label of the cross-section layer; S413, according to l=max(l i ) obtaining the annular length l of the crack defect; S414, sequentially calculating the axial depth d of the crack defect in each layer i ; S415, the axial depth d of the crack defect of all cross-section layers i The axial length d of the crack defect is obtained by adding them together.

4. The CT flaw detection auxiliary method based on the instance segmentation model according to claim 2, characterized in that: The S42 specifically includes the following steps: S421, obtaining debonding defects of all cross-sectional layers on the cross-sectional CT scan image; S422, obtaining the contour number of the debonding defect in each of the cross-section layers; S423, overlapping the debonding defects of two adjacent cross-section layers to obtain the contour number of the debonding defects after the overlapping process; S424, judging whether the number of contours of the debonding defect after the overlapping process is greater than the number of defect contours of a single cross-section layer; S425, if yes, it is determined that the positions of the debonding defects on the two overlapping cross-section layers are different, and the position information of the debonding defects is recorded; S426: If not, determine that the positions of the debonding defects on the two overlapping cross-section layers are the same, record the position information of the debonding defects and repeat S423.

5. The CT flaw detection auxiliary method based on the instance segmentation model according to claim 4, characterized in that: The S43 specifically includes the following steps: S431, sequentially calculating the length L of each debonding defect contour of each cross-section layer t , where t is a natural number and is the label of the cross-section layer; S432, sequentially calculating the axial depth D of each of the debonding defect profiles of each of the cross-section layers t ; S433, through S t =L t *D t The area S of each debonding defect of each cross-section layer is calculated. t ; S434, according to the location information of the debonding defect and the S t , and add up to get the total area S of the debonding defect at each position.

6. The CT flaw detection auxiliary method based on the instance segmentation model according to any one of claims 1 to 5, characterized in that: The S1 specifically includes the following steps: S11, build the initial instance segmentation model using YOLOv8 as the baseline network; S12, optimizing the initial instance segmentation model in a preset optimization method, wherein the preset optimization method includes at least one of improving the backbone network by CBAM, improving the neck network by ASF, and improving the segmentation detection head by adding a feature extraction process; S13, inputting the training data set of the cross-sectional CT scan image into the initial instance segmentation model, and performing accelerated optimization through edge computing technology to obtain the instance segmentation model; S14: Perform a pruning operation on the optimized instance segmentation model and deploy the instance segmentation model to an edge computing terminal.

7. The CT flaw detection auxiliary method based on the instance segmentation model according to any one of claims 1 to 5, characterized in that: The S5 specifically includes: S51, selecting a preset defect standard according to the defect type, and determining whether the cross-sectional defect size meets the preset defect standard; S52, if yes, generate a conclusion that a defect fault exists; S53. If not, generate a conclusion that the defect meets the acceptance criteria; S54. Generate the internal section defect evaluation result of the engine grain according to the conclusion that the defect failure exists and the conclusion that the defect meets the acceptance standard.

8. The CT flaw detection auxiliary method based on the instance segmentation model according to claim 7, characterized in that: After S54, the method further includes: S55, selecting a preset number of images from the cross-sectional CT scan image by a preset defect degree sorting algorithm, and attaching them to the internal cross-sectional defect evaluation result of the engine grain to form an internal cross-sectional defect evaluation report of the engine grain; And / or, S56, annotating defects on the cross-sectional CT scan image, uploading the annotated cross-sectional CT scan image to a folder of a preset storage model, and naming the folder according to the product number and upload time.

9. A CT flaw detection auxiliary device based on an instance segmentation model, characterized in that: The invention comprises a human-machine interaction terminal (1), a support hinge (2) and a suitcase (3), wherein two ends of the support hinge (2) are respectively connected to the human-machine interaction terminal (1) and the suitcase (3) so that the human-machine interaction terminal (1) can be moved into and out of the suitcase (3), and the human-machine interaction terminal (1) is used for communication connection with a CT host computer to implement the CT flaw detection auxiliary method based on an instance segmentation model as claimed in any one of claims 1 to 8.

10. A computer storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the CT flaw detection auxiliary method based on the instance segmentation model as claimed in any one of claims 1 to 8 is implemented.

Images