Underwater concrete structure crack detection method based on edge calculation and deep learning

Abstract

The invention discloses an underwater concrete structure crack detection method based on edge calculation and deep learning, and the method comprises the steps: carrying out the pixel preprocessing of underwater concrete structure crack image data in an image database, and obtaining a preprocessed image; constructing a lightweight image segmentation network to perform image enhancement processing on the preprocessed image to obtain an image enhancement data set; constructing an ASF-ODConv-YOLOv8 network model, and performing network training and verification according to the image enhancement data set to obtain an intelligent crack recognition algorithm model; based on the crack intelligent identification algorithm model, obtaining a predicted crack category label of the underwater concrete structure crack according to the test set; and constructing a digital twinborn model of the underwater concrete structure crack so as to detect the underwater concrete structure crack. The problems that a traditional recognition method has limitations in the aspects of space-time coverage, environmental adaptability, recognition efficiency and precision, depends on subjective judgment and is prone to causing missing recognition and wrong recognition are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of underwater concrete structure crack detection, and in particular to an underwater concrete structure crack detection method based on edge computing and deep learning. Background Art

[0002] During operation, underwater bridge structures are subjected to various loads and high thermal stresses, which often cause surface concrete cracking. Underwater bridge structures are exposed to even harsher environments, and concrete cracking is more severe under extreme loads or the coupling of multiple factors. Concrete cracking seriously affects the safety and durability of the structure.

[0003] The presence of cracks not only affects the service performance and safety of underwater bridge structures, but the identification of crack parameters such as width, length, and angle is also of great significance for scientifically evaluating the current service performance of bridge structures. However, the underwater environment makes traditional manual inspection extremely challenging and requires practitioners to have superb diving skills. It also faces multiple challenges such as limited operating time and harsh environmental conditions, making it difficult to meet the complex and changing needs of underwater inspection. In addition, traditional identification methods have limitations in terms of temporal and spatial coverage, environmental adaptability, recognition efficiency, and accuracy, and rely on subjective judgment, which can easily lead to omissions and misidentifications. Summary of the Invention

[0004] The present invention provides a method for detecting cracks in underwater concrete structures based on edge computing and deep learning to overcome the above technical problems.

[0005] In order to achieve the above object, the technical solution of the present invention is:

[0006] A method for detecting cracks in underwater concrete structures based on edge computing and deep learning, specifically comprising the following steps:

[0007] S1: Collect and obtain underwater concrete structure crack image data under various environmental interference factors, mark the underwater concrete structure crack image data with rectangular frames, and define and name the crack category labels to build an image database;

[0008] The image data of underwater concrete structure cracks include crack length, crack width and crack offset angle;

[0009] S2: performing pixel preprocessing on the underwater concrete structure crack image data in the image database to obtain a preprocessed image, wherein the preprocessing includes defogging, noise reduction, and histogram equalization processing;

[0010] S3: Improve the PSP Net network to build a lightweight image segmentation network and deploy it to edge devices to perform image enhancement on preprocessed images to obtain image enhancement datasets;

[0011] And divide the image enhancement dataset into training set, validation set and test set according to the preset ratio;

[0012] S4: Improve the YOLOv8 network to build the ASF-ODConv-YOLOv8 network model;

[0013] The network is trained and verified based on the training set and the validation set to obtain a crack intelligent recognition algorithm model for underwater concrete structure crack detection;

[0014] Based on the crack intelligent identification algorithm model, the predicted crack category labels of underwater concrete structure cracks are obtained according to the test set;

[0015] S5: A finite element model of the underwater concrete structure is established using the Abaqus software. By simulating the virtual stress conditions of the underwater concrete structure and mapping the crack parts identified by the intelligent crack identification algorithm model onto the finite element model, a digital twin model of the underwater concrete structure cracks that combines mechanics with actual scenarios is obtained to realize the detection of cracks in the underwater concrete structure.

[0016] Furthermore, the S1 specifically includes the following steps:

[0017] S11: Based on the ROS system, the camera on the underwater robot is controlled to capture and obtain images of several cracks in the underwater concrete structure under the service environment with various environmental interference factors;

[0018] And based on the labelimg annotation tool, each crack image is annotated with a rectangular frame;

[0019] S12: Create a VOC folder;

[0020] And under the VOC folder, create a JPEGImages subfolder for storing the annotated crack images and an xml tag file for storing the crack tags;

[0021] S13: Import the crack images marked in the JPEGImages subfolder into the crack category labels corresponding to the names of the xml label files, and store them in a preset txt folder in the format of the YOLO dataset to build an image database.

[0022] Furthermore, the improved PSP Net network described in S3 specifically includes:

[0023] The input feature layer of the original PSP Net pyramid pooling module is divided into four feature sub-layers with different convolution kernels;

[0024] Each feature sub-layer is used to extract features from the preprocessed image through the backbone feature extraction network of the original PSP Net network to obtain a feature map, and perform average pooling operations at different scales.

[0025] Furthermore, the PSP Net network is improved in S3 to construct a lightweight image segmentation network, including an improved PSP Net network module, a single-stage target detection algorithm module, a classification network module, a regression network module, a segmentation network module, a comprehensive generation module, and a verification submodule;

[0026] The improved PSP Net network module is used to obtain and output an initial enhanced crack image based on the pre-processed image processed by the input feature layer;

[0027] The single-stage target detection algorithm module is used to extract crack features from the initial enhanced crack image and input the crack features into the classification network module, the regression network module and the segmentation network module respectively;

[0028] The crack characteristics include crack length characteristics, crack width characteristics and crack angle characteristics;

[0029] The classification network module is used to obtain and output a first crack feature map containing crack category labels based on the extracted crack features;

[0030] The regression network module is used to obtain and output a second crack feature map containing a plurality of crack position positioning bounding boxes based on the extracted crack features;

[0031] The segmentation network module is used to obtain and output a third crack feature map including a plurality of semantic segmentation connected domains based on the extracted crack features;

[0032] The semantic segmentation connected domain is a crack position rectangular area or a noise rectangular area;

[0033] The comprehensive generation module is used to fuse the first crack feature map and the second crack feature map to obtain a crack target detection positioning map;

[0034] The verification submodule is used to verify the semantic segmentation connected domain of the third crack feature map according to the crack target detection positioning map based on the verification strategy to obtain an image enhancement data set.

[0035] Furthermore, the verification strategy is specifically as follows:

[0036] Confirming the first center coordinates of the connected domain of the third crack feature map and the second center coordinates of the crack target detection positioning map;

[0037] After the first center coordinates are overlapped with the second center coordinates, an area difference between the rectangular box area of ​​the connected domain of the third crack feature map and the rectangular area of ​​the crack position positioning boundary box of the crack target detection positioning map is obtained;

[0038] Determine whether the area difference meets a preset error threshold range;

[0039] If it is confirmed that the area difference meets the preset error threshold range, the connected domain is determined to be the crack position and the connected domain is retained; otherwise, the connected domain is determined to be a noise position and is deleted.

[0040] Furthermore, the ASF-ODConv-YOLOv8 network model constructed in S4 specifically includes a backbone network unit, a neck network unit, and a head network unit;

[0041] The backbone network unit is used to extract disease detection feature maps of different scales;

[0042] The backbone network unit includes a first CBL module, a second CBL module, a first C2f1 module, a third CBL module, a full-dimensional dynamic convolution module, a fourth CBL module, a second C2f1 module, a fifth CBL module, a third C2f1 module and an SPPF module connected in sequence;

[0043] The first CBL module is used to extract the crack features of the underwater concrete structure from the sample image in the input image enhancement dataset to obtain a first feature map;

[0044] The second CBL module is used to perform a convolution operation on the first feature map to obtain a second feature map;

[0045] The first C2f1 module is used to perform cross-channel feature fusion on the second feature map to obtain a third feature map; the third CBL module is used to perform a convolution operation on the third feature map to obtain a fourth feature map;

[0046] The full-dimensional dynamic convolution module is used to perform a full-dimensional dynamic convolution operation on the fourth feature map to obtain a fifth feature map;

[0047] The fourth CBL module is used to perform a convolution operation on the fifth feature map to obtain a sixth feature map;

[0048] The second C2f1 module is used to perform cross-channel feature fusion on the sixth feature map to obtain the seventh feature map;

[0049] The fifth CBL module is used to perform a convolution operation on the seventh feature map to obtain an eighth feature map;

[0050] The third C2f1 module is used to perform cross-channel feature fusion on the eighth feature map to obtain a ninth feature map;

[0051] The SPPF module is used to perform feature compression and feature fusion on the ninth feature map to obtain the tenth feature map;

[0052] The neck network unit is used to perform feature fusion on the disease detection feature maps of different scales to obtain a disease crack fusion feature map; the disease detection feature maps of different scales are the third feature map, the fifth feature map, the seventh feature map, and the tenth feature map;

[0053] The neck network unit includes a scale sequence fusion SSFF module, a first TFE+CSP combination module, a second TFE+CSP combination module, a channel / position attention mechanism CPAM module, a first Concat+CSP combination module, and a second Concat+CSP combination module;

[0054] The scale sequence fusion SSFF module is used to perform sequence feature fusion on the fifth feature map, the seventh feature map and the tenth feature map to obtain an eleventh feature map;

[0055] The first TFE+CSP combination module is used to perform feature segmentation on the seventh feature map and the tenth feature map respectively, and further extract features from the feature maps after feature segmentation to obtain a twelfth feature map;

[0056] The second TFE+CSP combination module is used to perform feature segmentation on the third feature map, the fifth feature map, and the twelfth feature map, and further extract features from the feature maps after feature segmentation to obtain a thirteenth feature map;

[0057] The channel / position attention mechanism CPAM module is used to splice the thirteenth feature map and the eleventh feature map to obtain a first spliced ​​feature map;

[0058] The first Concat+CSP combination module is used to splice the twelfth feature map and the first spliced ​​feature map to obtain a second spliced ​​feature map, and further extract features from the second spliced ​​feature map to obtain a fourteenth feature map;

[0059] The second Concat+CSP combination module is used to splice the fourteenth feature map and the tenth feature map to obtain a third spliced ​​feature map, and further extract features from the third spliced ​​feature map to obtain a fifteenth feature map;

[0060] The head network unit is used to include a first detection head module, a second detection head module, a third detection head module and a prediction and recognition module; and the first detection head module is connected to the channel / position attention mechanism CPAM module, and is used to detect and recognize according to the first spliced ​​feature map to obtain a first output feature map;

[0061] The second detection head module is connected to the first Concat+CSP combination module, and is used to detect and identify the fourteenth feature map to obtain a second output feature map;

[0062] The third detection head module is connected to the second Concat+CSP combination module, and is used to detect and identify the fifteenth feature map to obtain a third output feature map;

[0063] The prediction and recognition module is used to fuse the first output feature map, the second output feature map and the third output feature map to achieve prediction of crack categories of underwater concrete structure cracks.

[0064] Furthermore, the method for obtaining a crack intelligent identification algorithm model for underwater concrete structure crack detection in S4 specifically includes the following steps:

[0065] S41: Perform model training on the constructed ASF-ODConv-YOLOv8 network model according to the training set to obtain the trained ASF-ODConv-YOLOv8 network model:

[0066] S42: Use EIoU function as the model loss function;

[0067] And perform model evaluation on the trained ASF-ODConv-YOLOv8 network model based on the test set to determine whether the network output of the trained ASF-ODConv-YOLOv8 network model converges;

[0068] If so, the trained ASF-ODConv-YOLOv8 network model is the crack intelligent identification algorithm model;

[0069] Otherwise, the weight parameters of the trained ASF-ODConv-YOLOv8 network model are adaptively adjusted based on the back propagation method, and step S41 is repeated.

[0070] Furthermore, the method for constructing the digital twin model of underwater concrete structure cracks in S5 specifically includes the following steps:

[0071] S51: Establish a finite element model of underwater concrete structure using abaqus software;

[0072] The virtual stress of the underwater concrete structure is simulated by using the Abaqus software according to the actual stress conditions of the underwater concrete structure, and the crack parts identified by the intelligent crack identification algorithm model are mapped onto the finite element model;

[0073] S52: Construct a digital twin model of underwater concrete structure cracks that combines mechanics with actual scenarios. The model expression is:

[0074]

[0075] Where: a represents the crack length; N represents the number of load cycles for underwater concrete structures; K max Indicates the peak stress intensity factor; K C represents fracture toughness; C, m, p, q represent design parameters; ξ, R, ΔK th ,ΔK,ΔK max ,ΔK C They represent the crack opening coefficient, stress ratio, threshold stress intensity factor amplitude, stress intensity factor amplitude, stress intensity factor peak amplitude and fracture toughness amplitude respectively.

[0076] Beneficial effects: The present invention provides a method for detecting cracks in underwater concrete structures based on edge computing and deep learning, which includes collecting an image database of underwater concrete structure crack defects, processing image data, intelligent crack identification, and constructing a digital twin model. A self-developed underwater robot defect detection device collects defect data in real time under a service environment with multiple environmental interference factors to construct a database. Based on a data enhancement strategy, histogram equalization technology is used to dehaze and reduce noise on the image of underwater concrete structure crack defects. A lightweight image segmentation network is constructed by improving the PSP Net network and transmitting it to an edge device setting to perform image enhancement processing on the preprocessed image to obtain an image enhancement data set. An ASF-ODConv-YOLOv8 network architecture is designed to deeply analyze image features and realize high-precision intelligent crack identification based on multi-source data. Based on the feedback of crack detection data of bridge structures in actual service environments, a digital twin model of underwater concrete structure cracks that combines mechanics with actual scenarios is constructed. This embodiment optimizes the underwater concrete structure crack detection method based on the construction of a digital twin model to achieve the purpose of high-precision intelligent detection of underwater concrete structure cracks, and improves the detection efficiency and intelligence level. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following is a brief introduction to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0078] Figure 1 This is a flow chart of the underwater concrete structure crack detection method based on edge computing and deep learning of the present invention;

[0079] Figure 2Schematic diagram of the ASF-ODConv-YOLOv8 network architecture in this embodiment;

[0080] Figure 3 Schematic diagram of the structure of the full-dimensional dynamic convolution ODConv in this embodiment;

[0081] Figure 4 This is a flow chart of improving the PSP Net network in this embodiment;

[0082] Figure 5 This is a schematic diagram of the underwater robot disease detection system pre-established in this embodiment. DETAILED DESCRIPTION

[0083] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0084] This embodiment provides a method for detecting cracks in underwater concrete structures based on edge computing and deep learning. Figure 1 As shown, the specific steps include:

[0085] S1: Collect and obtain underwater concrete structure crack image data under various environmental interference factors, annotate the underwater concrete structure crack image data with rectangular frames, and define and name crack category labels to build an image database; underwater concrete structure crack image data includes crack length, crack width, and crack offset angle;

[0086] This embodiment also includes a pre-established underwater robot disease detection system, such as Figure 5As shown, it utilizes a motor propeller for propulsion and is equipped with a binocular system based on structured light technology for real-time image data acquisition. The robot device comprises a camera (traction), lighting, a waterproof housing, a motor propeller, power accessories, and a central processing unit as its hardware platform, and a real-time information transmission module (traction) and a data processing module as its software platform. The waterproof housing has an inner and outer layer, with the inner shell being waterproof and the outer shell being pressure-resistant, and the inner and outer shells tightly connected by high-strength bolts. The motor propeller comprises a motor unit and a propeller, each of which consists of four symmetrical drive assemblies, each consisting of four blades and a metal protective cover. The blade protective cover is connected to the robot body via a bearing. When the motor is powered by the power accessories, it drives the blades to rotate, driving the robot device to float up and down and propel forward and backward. The power accessories are located inside the waterproof housing, connected to the motor propeller at one end and the lighting at the other. The camera uses an RG B-binocular structured light camera for real-time video recording. In dimly lit areas, the robot's built-in lighting can be used to increase brightness. However, in underwater environments, especially in deep water areas, the camera vision will be significantly disturbed due to the influence of optical refraction. For this reason, this embodiment performs image correction on the camera based on high-order distortion correction polynomials to eliminate distortion. Compared with monocular imaging, binocular imaging technology can better reflect the connection between three-dimensional points and pixel points. It uses dual cameras to shoot the detection target and obtains target information based on the principle of parallax and triangulation. The main technical support for the real-time information transmission module in the above-mentioned software platform is the edge processing platform. The construction of this platform integrates the Raspberry Pi system and the ROS (Robot Operating System) system, and uses an embedded industrial-grade Raspberry Pi 3rd generation Revolution Pi device based on ARM core as the core controller of the edge processing platform. After the burning is successful, the Raspberry Pi system is called by the binocular camera to collect images of concrete structure defects.The camera is serially connected to the CSI (Camera Serial Interface) of the Raspberry Pi 4B platform. The industrial Raspberry Pi 3 is used as the main body and is equipped with the Raspbian system. Remote control and multi-data monitoring of the underwater robot are achieved through the ROS system, realizing multiple functions of the robot such as free propulsion, snorkeling, and lighting. After pre-implementing the above functions, this embodiment can use the wireless transmission module to transmit wireless signals to the computer. The specific implementation process is as follows: inserting the SD card into the computer, reading the Raspbian system, installing the Raspberry Pi, configuring the WiFi connection, writing code in the VIM editor environment, and controlling the Raspberry Pi GPIO terminal by calling the wiringPi.GPIO library. The code is then converted into an executable text file to realize the program's control of the hardware platform; after the device collects data, it compiles its format and stores its disease data, including crack length, width, and angle information, in a MySQL database; using the Raspberry Pi SHH (Secure Shell) module, pre-write WiFi configuration information into the code, and map the Raspberry Pi's IP address to the public network to ensure that the data can be stably transmitted to the remote computer via WiFi;

[0087] The test stand is constructed with aluminum alloy profiles and is designed as a double-layer structure. The profiles on both sides of the upper layer are responsible for stabilizing the test block placement table, while the lower layer assumes the supporting function. The inner side of the upper profile is equipped with a ruler for precise distance measurement, and the camera sealed cabin is fixed to one end of the test stand with a crossbeam. The test block placement table is constructed of a base plate, a cover plate and a support rod. The base plate is slightly larger than the test block; the cover plate and the base plate are tightly attached to the upper and lower surfaces of the test block for fixation, and the support rod is threaded. It serves as a control device when installing the lifting ring. An ordered shooting numbering method is used to enable the robot to shoot according to the specified running trajectory. The shooting process should keep the camera optical axis direction perpendicular to the observation surface to reduce the calibration error caused by the tilt of the shooting line of sight. The images taken by the underwater robot generally contain a small amount of crack information, defect information and noise. The grayscale value of the noise generally overlaps with part of the grayscale value of the crack. The defect information will affect the extraction of the crack edge contour during the edge detection process. The crack image is converted from pixel units to actual physical distance through size calibration;

[0088] The extraction of the crack edge contour refers to reducing the noise other than the crack information in the captured image. The specific method is: when the connected domain is highly consistent with the center coordinates of the positioning map and the area of ​​the rectangular frame is similar, the connected domain is determined to be a crack; otherwise, it is regarded as noise and deleted immediately to ensure that only the connected domain containing the crack is retained.

[0089] In digital image processing, the number of pixels corresponding to the long side of the grid of 100mm is 1411 pixels, and the actual length corresponding to 1 pixel is L0=100 / 1411mm;

[0090] The number of pixels of the crack area in the image is known as b, which is converted into a physical distance of b*L0;

[0091] The specific operation process of the acquisition platform is as follows: First, complete the assembly of the various parts of the acquisition platform, connect the transmission line and fix the shooting equipment; then, pass the rope through the reserved lifting ring, and steadily lift the test platform into the water, and ensure that the test platform remains balanced by adjusting the length of the front and rear lifting ropes. Next, place the pre-prepared concrete disease test block on the test block placement table, fix it with a cover plate, and pass the rope through the four upper lifting rings to fix the placement table in the slide of the test platform. As the test platform slowly descends, the test block placement table is synchronously lifted into the water; by manipulating the rope above the placement table, it can be moved back and forth in the slide, which is convenient for shooting and collecting at different distances. In addition, the test block is prepared manually, and the barnacles are bonded to the previously prepared concrete test block with an adhesive to simulate the attachment of barnacle biological diseases on the concrete surface;

[0092] The specific steps include:

[0093] S11: Based on the ROS system, the camera on the underwater robot is controlled to capture and obtain images of several cracks in the underwater concrete structure under the service environment with various environmental interference factors;

[0094] And based on the labelimg annotation tool, each crack image is annotated with a rectangular frame;

[0095] S12: Create a VOC folder;

[0096] And under the VOC folder, create a JPEGImages subfolder for storing the annotated crack images and an xml tag file for storing the crack tags;

[0097] S13: Import the crack images marked in the JPEGImages subfolder into the crack category labels corresponding to the names of the xml label files, and store them in a preset txt folder in the format of the YOLO dataset to build an image database.

[0098] In this embodiment, an underwater robot is controlled on the ROS system side to be within 30 cm of the structure to be tested. The robot is equipped with a camera to take 6,000 pictures of cracks, and the dataset images are placed in a VOC folder. An Annotations subfolder is created in the folder to store the labeled xml label files, and a JPEGImages subfolder is created to store the dataset images to be labeled in this article. The name of the image should correspond one-to-one to the name of the xml label file, and a txt folder is created to store the label files converted into the YOLO dataset format. The operation of labeling the dataset is as follows: first, in a preset training environment, use the pip command to install the labelimg software, and use the software to complete the labeling of the dataset; then, label, select the objects to be labeled in turn and name and classify the labels. The labeled xml files will be saved accordingly in the established folder. Finally, the labeled dataset is distributed according to the ratio of 8:1:1 among the training set, validation set and test set. The dataset distributed in this ratio is used for network training of underwater structural diseases.

[0099] S2: performing pixel preprocessing on the underwater concrete structure crack image data in the image database to obtain a preprocessed image, wherein the preprocessing includes defogging, noise reduction, and histogram equalization processing;

[0100] This embodiment considers the sample quantity and annotation quality of underwater concrete structure crack image data, covering a variety of environmental interference factors (such as pitting, rust cracks, erosion, porosity, and marine organism attachment on the concrete surface, multiple backgrounds with similar colors to the cracks, and thin lines, wires, and marks of similar shapes, including staggered cracks and multiple parallel cracks). The number of annotated images reaches 6,000, and covers a variety of environmental interference factors. The dataset is subsequently expanded through data enhancement to ensure the sample quantity. Pixel preprocessing is also performed on the underwater concrete structure crack image data to obtain a data-enhanced dataset to ensure annotation quality.

[0101] The content of pixel preprocessing in this embodiment is as follows:

[0102] Obtain a grayscale image from the original R, G, and B foggy images, and use the average method to calculate the Gray value (an integer between 0 and 255):

[0103] Gray=(R+G+B) / 3

[0104] Where: R, B, G matrix sizes are all 500*338;

[0105] Stretch the pixels of the decomposed R, B, and G channel images, and adjust the grayscale of the original image based on the linear transformation in grayscale transformation. The linear transformation formula is:

[0106] g(x,y)=k*f(x,y)+b

[0107] Where: g(x,y) represents the target pixel value after transformation, f(x,y) represents the original pixel value, k represents the slope, b represents the intercept, and the value of k and the transformation effect have the following situations:

[0108] When k>1, it can be used to enhance the image contrast. The pixel values ​​of the image are all increased after the transformation, and the overall display effect is enhanced.

[0109] When k = 1, it is often used to adjust image brightness;

[0110] When k<1, the effect is opposite to k>1, and the contrast and overall effect of the image are weakened;

[0111] When k < 0, the brighter areas of the source image will become darker, and the darker areas will become brighter; if k = -1 and b = 255, the image can be inverted).

[0112] S3: Improve the PSP Net network to build a lightweight image segmentation network and transmit it to the edge device setting to perform image enhancement on the preprocessed image to obtain an image enhancement dataset. The image enhancement dataset is then divided into a training set, a validation set, and a test set according to a preset ratio.

[0113] The improved PSP Net network, such as Figure 4 As shown, specifically including:

[0114] The input feature layer of the original PSP Net pyramid pooling module is divided into four feature sub-layers with different convolution kernels. Each feature sub-layer is used to extract features from the preprocessed image through the backbone feature extraction network of the original PSP Net network to obtain feature maps, and then perform average pooling operations at different scales.

[0115] This example performs image enhancement processing on the acquired underwater concrete structure crack image data to obtain the original R, G, and B foggy images. The input image data is passed through the optimized PSP Net network, and after the backbone features are extracted, it is passed to the pyramid pooling module. The pyramid pooling module divides the input feature layer into four sub-layers, namely 1×1, 2×2, 4×4, and 8×8 regions, and performs average pooling operations on each of them, thereby reducing the robustness of the input image in semantic segmentation.

[0116] In a specific embodiment, the PSP Net network is improved in S3 to construct a lightweight image segmentation network, including an improved PSP Net network module, a single-stage target detection algorithm module, a classification network module, a regression network module, a segmentation network module, a comprehensive generation module, and a verification submodule;

[0117] The improved PSP Net network module is used to obtain and output an initial enhanced crack image based on the pre-processed image processed by the input feature layer;

[0118] The single-stage target detection algorithm module is used to extract crack features from the initial enhanced crack image and input the crack features into the classification network module, the regression network module and the segmentation network module respectively;

[0119] The crack characteristics include crack length characteristics, crack width characteristics and crack angle characteristics;

[0120] The classification network module is used to obtain and output a first crack feature map containing crack category labels based on the extracted crack features;

[0121] The regression network module is used to obtain and output a second crack feature map containing a plurality of crack position positioning bounding boxes based on the extracted crack features;

[0122] The segmentation network module is used to obtain and output a third crack feature map including a plurality of semantic segmentation connected domains based on the extracted crack features;

[0123] The semantic segmentation connected domain is a crack position rectangular area or a noise rectangular area;

[0124] The comprehensive generation module is used to fuse the first crack feature map and the second crack feature map to obtain a crack target detection positioning map;

[0125] The verification submodule is used to verify the semantic segmentation connected domain of the third crack feature map according to the crack target detection positioning map based on the verification strategy to obtain an image enhancement data set;

[0126] The verification strategy is specifically as follows:

[0127] Confirm the first center coordinate of the connected domain of the third crack feature map and the second center coordinate of the crack target detection and positioning map; after coinciding the first center coordinate with the second center coordinate, obtain the area difference between the rectangular box area of ​​the connected domain of the third crack feature map and the rectangular area of ​​the crack position positioning boundary box of the crack target detection and positioning map; determine whether the area difference meets the preset error threshold range; if it is confirmed that the area difference meets the preset error threshold range, determine that the connected domain is the crack position and retain the connected domain; otherwise, determine that the connected domain is a noise position and delete the connected domain;

[0128] This embodiment is based on target detection and semantic segmentation, and integrates and constructs a lightweight image segmentation module to optimize the intelligent crack recognition algorithm, perform high-precision detection on crack images, and obtain a feature-enhanced data set: First, the backbone network of the single-stage target detection algorithm is used to efficiently extract crack features, and these feature information is accurately passed to the classification network, regression network, and segmentation network modules. Each network branch module outputs accurate classification results, accurate positioning results, and detailed segmentation results. Subsequently, the results of classification and positioning are used to comprehensively generate a crack target detection positioning map, and this positioning map is finely checked with the connected domain in the semantic segmentation results. Under the action of the verification submodule, only the connected domain containing crack information is retained, thereby significantly improving the accuracy and efficiency of crack segmentation. Compared with the semantic segmentation results, classification and target detection have significant advantages in reducing the probability of false detection. Therefore, the verification submodule fully utilizes all the connected domain information output by semantic segmentation and performs a deep comparison with the crack localization map information output by classification and localization. When the center coordinates of the connected domain and the localization map are highly consistent and the rectangular box area is similar, the connected domain is determined to be a crack; otherwise, it is considered noise and immediately deleted, ensuring that only the connected domain containing the crack is retained. Ultimately, the algorithm achieves the multi-task integrated goal of high-precision real-time classification, accurate detection, and fine segmentation of crack images, demonstrating excellent performance and practicality.

[0129] S4: Improve the YOLOv8 network to build the ASF-ODConv-YOLOv8 network model;

[0130] The network is trained and verified based on the training set and the validation set to obtain a crack intelligent recognition algorithm model for underwater concrete structure crack detection;

[0131] Based on the crack intelligent identification algorithm model, the predicted crack category labels of underwater concrete structure cracks are obtained according to the test set;

[0132] In a specific embodiment, the constructed ASF-ODConv-YOLOv8 network model is as follows: Figure 2 As shown, it specifically includes a backbone network unit, a neck network unit and a head network unit;

[0133] The backbone network unit is used to extract disease detection feature maps of different scales;

[0134] The backbone network unit includes a first CBL module, a second CBL module, a first C2f1 module, a third CBL module, a full-dimensional dynamic convolution module, a fourth CBL module, a second C2f1 module, a fifth CBL module, a third C2f1 module and an SPPF module connected in sequence;

[0135] The first CBL module is used to extract the crack features of the underwater concrete structure from the sample image in the input image enhancement dataset to obtain a first feature map;

[0136] The second CBL module is used to perform a convolution operation on the first feature map to obtain a second feature map;

[0137] The first C2f1 module is used to perform cross-channel feature fusion on the second feature map to obtain a third feature map; the third CBL module is used to perform a convolution operation on the third feature map to obtain a fourth feature map;

[0138] The full-dimensional dynamic convolution module is used to perform a full-dimensional dynamic convolution operation on the fourth feature map to obtain a fifth feature map;

[0139] The fourth CBL module is used to perform a convolution operation on the fifth feature map to obtain a sixth feature map;

[0140] The second C2f1 module is used to perform cross-channel feature fusion on the sixth feature map to obtain the seventh feature map;

[0141] The fifth CBL module is used to perform a convolution operation on the seventh feature map to obtain an eighth feature map;

[0142] The third C2f1 module is used to perform cross-channel feature fusion on the eighth feature map to obtain a ninth feature map;

[0143] The SPPF module is used to perform feature compression and feature fusion on the ninth feature map to obtain the tenth feature map;

[0144] The neck network unit is used to perform feature fusion on the disease detection feature maps of different scales to obtain a disease crack fusion feature map; the disease detection feature maps of different scales are the third feature map, the fifth feature map, the seventh feature map, and the tenth feature map;

[0145] The neck network unit includes a scale sequence fusion SSFF module, a first TFE+CSP combination module, a second TFE+CSP combination module, a channel / position attention mechanism CPAM module, a first Concat+CSP combination module, and a second Concat+CSP combination module;

[0146] The scale sequence fusion SSFF module is used to perform sequence feature fusion on the fifth feature map, the seventh feature map and the tenth feature map to obtain an eleventh feature map;

[0147] The first TFE+CSP combination module is used to perform feature segmentation on the seventh feature map and the tenth feature map respectively, and further extract features from the feature maps after feature segmentation to obtain a twelfth feature map;

[0148] The second TFE+CSP combination module is used to perform feature segmentation on the third feature map, the fifth feature map, and the twelfth feature map, and further extract features from the feature maps after feature segmentation to obtain a thirteenth feature map;

[0149] The channel / position attention mechanism CPAM module is used to splice the thirteenth feature map and the eleventh feature map to obtain a first spliced ​​feature map;

[0150] The first Concat+CSP combination module is used to splice the twelfth feature map and the first spliced ​​feature map to obtain a second spliced ​​feature map, and further extract features from the second spliced ​​feature map to obtain a fourteenth feature map;

[0151] The second Concat+CSP combination module is used to splice the fourteenth feature map and the tenth feature map to obtain a third spliced ​​feature map, and further extract features from the third spliced ​​feature map to obtain a fifteenth feature map;

[0152] The head network unit is used to include a first detection head module, a second detection head module, a third detection head module and a prediction and recognition module; and the first detection head module is connected to the channel / position attention mechanism CPAM module, and is used to detect and recognize according to the first spliced ​​feature map to obtain a first output feature map;

[0153] The second detection head module is connected to the first Concat+CSP combination module, and is used to detect and identify the fourteenth feature map to obtain a second output feature map;

[0154] The third detection head module is connected to the second Concat+CSP combination module, and is used to detect and identify the fifteenth feature map to obtain a third output feature map;

[0155] The prediction and recognition module is used to fuse the first output feature map, the second output feature map, and the third output feature map to achieve prediction of crack categories of underwater concrete structure cracks;

[0156] This embodiment designs the ASF-ODConv-YOLOv8 network architecture, that is, adding full-dimensional dynamic convolution, namely ODConv convolution block, to the backbone network. Figure 3As shown in the figure, the neck network in the YOLO architecture is improved; by introducing the ODConv convolution block, the fourth convolution layer in the backbone network is replaced with a full-dimensional dynamic convolution, which can parse the features of the input image from different dimensions when training the network, perform linear weighting on multiple convolution kernels, and use the multi-dimensional attention mechanism and parallel strategy to learn the attention of the convolution kernel along the four dimensions of the kernel space in any convolution layer, while maximizing the detection accuracy and speed without increasing the amount of calculation, so as to meet the needs of disease detection and recognition under different conditions; secondly, the neck of the YOLOv8 network is improved, combining the Scale Sequence Feature Fusion (SSFF) module and the Triple Feature Encoder (TFE) module, and through the CPAM module (Channel and PositionAttention) containing the channel and position attention mechanism The CPAM (CPAM) combines the feature information of the SSFF and TFC modules to further improve the accuracy of instance segmentation. When training the dataset, the bounding box position loss is optimized by minimizing the difference in width and height between the bounding box and the anchor box using EIoU. In the preset post-processing stage, Soft Non Maximum Suppression (Soft-NMS) is used to improve image overlap.

[0157] The scale sequence fusion SSFF module scales the input feature map using the following formula:

[0158] F σ (w,h)=G σ (w,h)×f(w,h)

[0159]

[0160] Where: f(w,h) represents the input two-dimensional image; w represents the image width; h represents the image length; σ represents the scaling parameter of the standard deviation of the two-dimensional Gaussian filter used for convolution; G σ (w,h) represents the intermediate parameter; F σ(w,h) represents the output of the scale sequence fusion (SSFF) module. This embodiment extracts scale sequence features by horizontally stacking feature maps of different scales and using 3D convolution. The feature extraction process of the scale sequence fusion (SSFF) module is as follows: the dimension of each feature layer is increased by using the unsqueeze method, converting it from a 3D tensor (height, width, channel) to a 4D tensor (depth, height, width, channel). The 4D feature maps are concatenated along the depth dimension to form a 3D feature map for subsequent convolution. Scale sequence feature extraction is completed using 3D convolution, 3D batch normalization, and the silu activation function.

[0161] The triple feature encoder TFE module is used to segment the size features (Large, Middle, Small), amplify the feature image of the Large image to improve the detailed feature information, and adjust the number of feature channels to perform feature encoding so that it is consistent with the main scale features. After the convolution module processes the Large image, its number of channels is adjusted to 1C, and then a hybrid structure of maximum pooling and average pooling is used for downsampling to retain the effectiveness and diversity of high-resolution features and cell images; for the Small image, the same operation is performed to adjust the number of channels, and then the nearest neighbor interpolation method is used for upsampling; finally, the multi-size feature maps are convolved once and then spliced ​​in the channel dimension. The expression of the triple feature encoder TFE module is:

[0162] F TFE =Concat(F l ,F m ,F s )

[0163] Where: F TFE Represents the output concatenated feature map; F l ,F m ,F s They represent the feature maps of three sizes obtained by size segmentation through the triple feature encoder TFE module;

[0164] The path aggregation network PA Net structure consists of two input channels, one is the channel attention network that receives TFE input, and the other is the position attention network that receives the superposition input of the channel attention network and the SSFF output. In this structure, the attention mechanism is introduced to perform global average pooling on the channel, so that the interaction between each individual channel and k adjacent channels is realized through one-dimensional convolution, where the k value is linearly related to the channel dimension c, and its expression is:

[0165] C=ψ(k)=2 (γ×k-b)

[0166]

[0167] Where: γ,b represents the design parameters;

[0168] In a specific embodiment, the method for obtaining a crack intelligent identification algorithm model for underwater concrete structure crack detection in S4 specifically includes the following steps:

[0169] S41: Perform model training on the constructed ASF-ODConv-YOLOv8 network model according to the training set to obtain the trained ASF-ODConv-YOLOv8 network model:

[0170] S42: Use EIoU function as the model loss function;

[0171] And perform model evaluation on the trained ASF-ODConv-YOLOv8 network model based on the test set to determine whether the network output of the trained ASF-ODConv-YOLOv8 network model converges;

[0172] If so, the trained ASF-ODConv-YOLOv8 network model is the crack intelligent identification algorithm model;

[0173] Otherwise, the weight parameters of the trained ASF-ODConv-YOLOv8 network model are adaptively adjusted based on the back-propagation method, and step S41 is repeated;

[0174] S5: A finite element model of the underwater concrete structure is established using Abaqus software. By simulating the virtual stress conditions of the underwater concrete structure and mapping the cracks identified by the intelligent crack identification algorithm model onto the finite element model, a digital twin model of the underwater concrete structure cracks is obtained that combines mechanics with actual scenarios to achieve crack detection in the underwater concrete structure.

[0175] The specific steps include:

[0176] S51: Establish a finite element model of underwater concrete structure using abaqus software;

[0177] The virtual stress of the underwater concrete structure is simulated by using the Abaqus software according to the actual stress conditions of the underwater concrete structure, and the crack parts identified by the intelligent crack identification algorithm model are mapped onto the finite element model;

[0178] S52: Construct a digital twin model of underwater concrete structure cracks that combines mechanics with actual scenarios. The model expression is:

[0179]

[0180] Where: a represents the crack length; N represents the number of load cycles for underwater concrete structures; K max Indicates the peak stress intensity factor; K C represents fracture toughness; C, m, p, q represent design parameters; ξ, R, ΔK th ,ΔK,ΔK max ,ΔK C They represent the crack opening coefficient, stress ratio, threshold stress intensity factor amplitude, stress intensity factor amplitude, stress intensity factor peak amplitude and fracture toughness amplitude respectively.

[0181] Beneficial effects: This embodiment uses a self-developed underwater robot defect detection device to collect defect data in a service environment with various environmental interference factors in real time to build a database; based on the data enhancement strategy, the histogram equalization technology is used to dehaze and reduce noise on the underwater concrete structure crack defect image, and a lightweight image segmentation network is constructed by improving the PSP Net network, and it is transmitted to the edge device setting to perform image enhancement processing on the pre-processed image to obtain an image enhancement data set; by designing the ASF-ODConv-YOLOv8 network architecture, the image features are deeply analyzed to achieve high-precision intelligent crack recognition based on multi-source data; based on the feedback of crack detection data of bridge structures in actual service environments, a digital twin model of underwater concrete structure cracks that combines mechanics and actual scenarios is constructed. This embodiment optimizes the underwater concrete structure crack detection method based on the construction of a digital twin model to achieve the purpose of high-precision intelligent detection of underwater concrete structure cracks, and improves the detection efficiency and intelligence level.

[0182] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for detecting cracks in underwater concrete structures based on edge computing and deep learning, characterized in that: The specific steps include: S1: Collect and obtain underwater concrete structure crack image data under various environmental interference factors, mark the underwater concrete structure crack image data with rectangular frames, and define and name the crack category labels to build an image database; The image data of underwater concrete structure cracks include crack length, crack width and crack offset angle; S2: performing pixel preprocessing on the underwater concrete structure crack image data in the image database to obtain a preprocessed image, wherein the preprocessing includes defogging, noise reduction, and histogram equalization processing; S3: Improve the PSP Net network to build a lightweight image segmentation network and deploy it to edge devices to perform image enhancement on preprocessed images to obtain an image enhancement dataset; And divide the image enhancement dataset into training set, validation set and test set according to the preset ratio; S4: Improve the YOLOv8 network to build the ASF-ODConv-YOLOv8 network model; The network is trained and verified based on the training set and the validation set to obtain a crack intelligent recognition algorithm model for underwater concrete structure crack detection; Based on the crack intelligent identification algorithm model, the predicted crack category labels of underwater concrete structure cracks are obtained according to the test set; S5: A finite element model of the underwater concrete structure is established using the Abaqus software. By simulating the virtual stress conditions of the underwater concrete structure and mapping the crack parts identified by the intelligent crack identification algorithm model onto the finite element model, a digital twin model of the underwater concrete structure cracks that combines mechanics with actual scenarios is obtained to realize the detection of cracks in the underwater concrete structure.

2. The method for detecting cracks in underwater concrete structures based on edge computing and deep learning according to claim 1, characterized in that: The S1 specifically includes the following steps: S11: Based on the ROS system, the camera on the underwater robot is controlled to capture and obtain images of several cracks in the underwater concrete structure under the service environment with various environmental interference factors; And based on the labelimg annotation tool, each crack image is annotated with a rectangular frame; S12: Create a VOC folder; And under the VOC folder, create a JPEGImages subfolder for storing the annotated crack images and an xml tag file for storing the crack tags; S13: Import the crack images marked in the JPEGImages subfolder into the crack category labels corresponding to the names of the xml label files, and store them in a preset txt folder in the format of the YOLO dataset to build an image database.

3. The method for detecting cracks in underwater concrete structures based on edge computing and deep learning according to claim 2, characterized in that: The method for improving the PSP Net network described in S3 specifically includes: The input feature layer of the original PSP Net pyramid pooling module is divided into four feature sub-layers with different convolution kernels; Each feature sub-layer is used to extract features from the preprocessed image through the backbone feature extraction network of the original PSP Net network to obtain a feature map, and perform average pooling operations at different scales.

4. The method for detecting cracks in underwater concrete structures based on edge computing and deep learning according to claim 3, characterized in that: The lightweight image segmentation network constructed by improving the PSP Net network in S3 includes an improved PSP Net network module, a single-stage target detection algorithm module, a classification network module, a regression network module, a segmentation network module, a comprehensive generation module, and a verification submodule; The improved PSP Net network module is used to obtain and output an initial enhanced crack image based on the pre-processed image processed by the input feature layer; The single-stage target detection algorithm module is used to extract crack features from the initial enhanced crack image and input the crack features into the classification network module, the regression network module and the segmentation network module respectively; The crack characteristics include crack length characteristics, crack width characteristics and crack angle characteristics; The classification network module is used to obtain and output a first crack feature map containing crack category labels based on the extracted crack features; The regression network module is used to obtain and output a second crack feature map containing a plurality of crack position positioning bounding boxes based on the extracted crack features; The segmentation network module is used to obtain and output a third crack feature map including a plurality of semantic segmentation connected domains based on the extracted crack features; The semantic segmentation connected domain is a crack position rectangular area or a noise rectangular area; The comprehensive generation module is used to fuse the first crack feature map and the second crack feature map to obtain a crack target detection positioning map; The verification submodule is used to verify the semantic segmentation connected domain of the third crack feature map according to the crack target detection positioning map based on the verification strategy to obtain an image enhancement data set.

5. The method for detecting cracks in underwater concrete structures based on edge computing and deep learning according to claim 4, characterized in that: The verification strategy is specifically: Confirming the first center coordinates of the connected domain of the third crack feature map and the second center coordinates of the crack target detection positioning map; After the first center coordinates are overlapped with the second center coordinates, an area difference between the rectangular box area of ​​the connected domain of the third crack feature map and the rectangular area of ​​the crack position positioning boundary box of the crack target detection positioning map is obtained; Determine whether the area difference meets a preset error threshold range; If it is confirmed that the area difference meets the preset error threshold range, the connected domain and the crack position are determined and the connected domain is retained; Otherwise, the connected domain is determined to be a noise location and is deleted.

6. The method for detecting cracks in underwater concrete structures based on edge computing and deep learning according to claim 5, characterized in that: The ASF-ODConv-YOLOv8 network model built in S4 includes the backbone network unit, the neck network unit, and the head network unit; The backbone network unit is used to extract disease detection feature maps of different scales; The backbone network unit includes a first CBL module, a second CBL module, a first C2f1 module, a third CBL module, a full-dimensional dynamic convolution module, a fourth CBL module, a second C2f1 module, a fifth CBL module, a third C2f1 module and an SPPF module connected in sequence; The first CBL module is used to extract the crack features of the underwater concrete structure from the sample image in the input image enhancement dataset to obtain a first feature map; The second CBL module is used to perform a convolution operation on the first feature map to obtain a second feature map; The first C2f1 module is used to perform cross-channel feature fusion on the second feature map to obtain a third feature map; the third CBL module is used to perform a convolution operation on the third feature map to obtain a fourth feature map; The full-dimensional dynamic convolution module is used to perform a full-dimensional dynamic convolution operation on the fourth feature map to obtain a fifth feature map; The fourth CBL module is used to perform a convolution operation on the fifth feature map to obtain a sixth feature map; The second C2f1 module is used to perform cross-channel feature fusion on the sixth feature map to obtain the seventh feature map; The fifth CBL module is used to perform a convolution operation on the seventh feature map to obtain an eighth feature map; The third C2f1 module is used to perform cross-channel feature fusion on the eighth feature map to obtain a ninth feature map; The SPPF module is used to perform feature compression and feature fusion on the ninth feature map to obtain the tenth feature map; The neck network unit is used to perform feature fusion on the disease detection feature maps of different scales to obtain a disease crack fusion feature map; the disease detection feature maps of different scales are the third feature map, the fifth feature map, the seventh feature map, and the tenth feature map; The neck network unit includes a scale sequence fusion SSFF module, a first TFE+CSP combination module, a second TFE+CSP combination module, a channel / position attention mechanism CPAM module, a first Concat+CSP combination module, and a second Concat+CSP combination module; The scale sequence fusion SSFF module is used to perform sequence feature fusion on the fifth feature map, the seventh feature map and the tenth feature map to obtain an eleventh feature map; The first TFE+CSP combination module is used to perform feature segmentation on the seventh feature map and the tenth feature map respectively, and further extract features from the feature maps after feature segmentation to obtain a twelfth feature map; The second TFE+CSP combination module is used to perform feature segmentation on the third feature map, the fifth feature map, and the twelfth feature map, and further extract features from the feature maps after feature segmentation to obtain a thirteenth feature map; The channel / position attention mechanism CPAM module is used to splice the thirteenth feature map and the eleventh feature map to obtain a first spliced ​​feature map; The first Concat+CSP combination module is used to splice the twelfth feature map and the first spliced ​​feature map to obtain a second spliced ​​feature map, and further extract features from the second spliced ​​feature map to obtain a fourteenth feature map; The second Concat+CSP combination module is used to splice the fourteenth feature map and the tenth feature map to obtain a third spliced ​​feature map, and further extract features from the third spliced ​​feature map to obtain a fifteenth feature map; The head network unit is used to include a first detection head module, a second detection head module, a third detection head module and a prediction and recognition module; and the first detection head module is connected to the channel / position attention mechanism CPAM module, and is used to detect and recognize according to the first spliced ​​feature map to obtain a first output feature map; The second detection head module is connected to the first Concat+CSP combination module, and is used to detect and identify the fourteenth feature map to obtain a second output feature map; The third detection head module is connected to the second Concat+CSP combination module, and is used to detect and identify the fifteenth feature map to obtain a third output feature map; The prediction and recognition module is used to fuse the first output feature map, the second output feature map and the third output feature map to achieve prediction of crack categories of underwater concrete structure cracks.

7. The method for detecting cracks in underwater concrete structures based on edge computing and deep learning according to claim 6, characterized in that: The method for obtaining a crack intelligent identification algorithm model for underwater concrete structure crack detection in S4 specifically includes the following steps: S41: Perform model training on the constructed ASF-ODConv-YOLOv8 network model according to the training set to obtain the trained ASF-ODConv-YOLOv8 network model: S42: Use EIoU function as the model loss function; And perform model evaluation on the trained ASF-ODConv-YOLOv8 network model based on the test set to determine whether the network output of the trained ASF-ODConv-YOLOv8 network model converges; If so, the trained ASF-ODConv-YOLOv8 network model is the crack intelligent identification algorithm model; Otherwise, the weight parameters of the trained ASF-ODConv-YOLOv8 network model are adaptively adjusted based on the back propagation method, and step S41 is repeated.

8. The method for detecting cracks in underwater concrete structures based on edge computing and deep learning according to claim 7, characterized in that: The method for constructing a digital twin model of underwater concrete structure cracks in S5 specifically includes the following steps: S51: Establish a finite element model of underwater concrete structure using abaqus software; The virtual stress of the underwater concrete structure is simulated by using the Abaqus software according to the actual stress conditions of the underwater concrete structure, and the crack parts identified by the intelligent crack identification algorithm model are mapped onto the finite element model; S52: Construct a digital twin model of underwater concrete structure cracks that combines mechanics with actual scenarios. The model expression is: Where: a represents the crack length; N represents the number of load cycles for underwater concrete structures; K max Indicates the peak stress intensity factor; K C represents fracture toughness; C, m, p, q represent design parameters; ξ, R, ΔK th ,ΔK,ΔK max ,ΔK C They represent the crack opening coefficient, stress ratio, threshold stress intensity factor amplitude, stress intensity factor amplitude, stress intensity factor peak amplitude and fracture toughness amplitude respectively.

Images