A crack damage intelligent evaluation method based on visual information and physical fusion

Abstract

This invention discloses an intelligent crack damage assessment method based on the fusion of visual information and physical properties, relating to the field of material fatigue strength technology. First, a masked region convolutional neural network is used to automatically identify and segment pixel-level information of fatigue cracks. Crack coordinate information is calculated using camera calibration information to generate a digital geometric model containing the crack. Then, the perceived information and the crack geometric model are fused together. An offline high-precision fracture mechanics physical model continuously provides data, and a deep learning model for crack damage assessment evaluates the crack damage state in real time. Finally, load optimization is performed to reduce the load condition to a safe range. This invention integrates automatic fatigue crack detection, damage assessment, and external load optimization into a single system, achieving real-time online crack damage assessment and load optimization.

Description

Technical Field

[0001] This invention relates to the field of material fatigue strength technology, and mainly to an intelligent assessment method for crack damage based on the fusion of visual information and physical properties. Background Technology

[0002] Due to extreme service environments and manufacturing defects, mechanical damage such as fatigue cracks is widespread in mechanical structures, causing numerous fracture accidents such as aircraft blade breakage, oil pipe rupture, and rail cracking. With rapid technological advancements, the informatization and intelligentization of equipment inspection and maintenance have become a development trend. Many large, high-end, and complex pieces of equipment, such as launch vehicles, on-orbit satellites, extraterrestrial probes, and nuclear power systems, often operate in unmanned environments such as vacuum and irradiation, creating an urgent need for automated and intelligent monitoring of fatigue crack damage. Information strength theories based on artificial intelligence, edge computing, and other technologies with informatization, intelligentization, and rapid response capabilities have become a research hotspot in the field of modern mechanical strength.

[0003] To assess the safety status of cracked structures, many scholars have studied the crack tip strength of cracked components. The classic approach is to assess the damage state of cracked structures based on fracture mechanics. In the 1950s, Irwin laid the foundation for linear elastic fracture mechanics by introducing a stress intensity factor to assess the magnitude of stress concentration at the crack tip. Subsequently, to solve engineering fracture problems, numerous numerical methods for fracture analysis, such as the fracture finite element method, extended finite element method, boundary finite element method, and scaled boundary finite element method, have been developed to calculate the stress intensity factor at the crack tip. However, these methods typically involve handling singular elements at the crack tip and mesh re-division, and lack integration with sensory information, making it difficult to automate and perform real-time assessment of crack tip damage. For cracks identified during service, how to intelligently embed them into fracture analysis models for damage assessment, and achieve automatic identification, assessment, and load control of fatigue cracks, remains one of the key and challenging research areas in this field.

[0004] In recent years, with the development of artificial intelligence technology, deep learning methods have been widely applied in fields such as computer vision and natural language processing. Due to their advantages of automatically extracting features and being easily integrated into mobile devices, deep learning methods are also gradually being applied to fatigue and fracture studies. However, these studies mainly investigate the presence of cracks in structures and cannot assess the current safety status of cracked components. Integrating information with fracture mechanics physics is an effective way to solve damage assessment of cracked structural components. NASA was among the first to propose a high-fidelity simulation and health management system based on digital twin technology, achieving safety and reliability levels during spacecraft service. Overall, this field is still in the exploratory stage, especially in the integration of perceived information and physical models. For the problem of intelligent damage assessment of sudden cracks appearing on crack-free components, how to automatically identify cracks and effectively integrate crack perception information into the physical model for real-time monitoring of crack damage status remains a research area that urgently needs further development. Summary of the Invention

[0005] The purpose of this invention is to address the problems of insufficient fusion of sensory information in traditional crack structure safety assessment, making it difficult to automatically and in real-time assess crack tip damage. This invention proposes an intelligent damage assessment and analysis method for crack structures, which enables automatic identification, assessment, and load optimization of surface cracks, thereby preventing structural fractures during operation and ensuring normal operation during service.

[0006] The technical solution of this invention is: to provide an intelligent crack damage assessment method based on the fusion of visual information and physical properties, the method comprising:

[0007] Step 1: Construct a deep learning model for crack recognition and segmentation to automatically recognize and segment the input image;

[0008] Step 2: Using the detected calibration information, the crack location information is converted into the actual crack location information on the surface of the structural component. The actual crack location information is then interacted with the offline fracture mechanics model to construct the crack geometry model.

[0009] Step 3: Construct a deep learning model for crack damage assessment, perform real-time prediction on the online crack geometry model, and continuously enrich the crack damage assessment deep learning model database by constructing the crack geometry model from the offline fracture mechanics model.

[0010] Step 4: Optimize the load and reduce the load conditions to a safe range.

[0011] Furthermore, the construction of the deep learning model for crack recognition and segmentation in step 1 includes the following steps:

[0012] Step 1.1: Select experimental specimens of uniform size and material to make crack specimens with different crack lengths, collect effective crack images, and label the cracks and square calibration plates in the images. The crack images and labels together form a set of data. The collection of all data constitutes the database of the Mask R-CNN crack recognition and segmentation deep learning model. The database is divided into training set and test set.

[0013] Step 1.2: Train the deep learning model for crack recognition based on the database described in Step 1.1. Determine the batch size for training according to the amount of data, and adjust the initial learning rate, hyperparameter momentum, and weight decay to ensure that the training model converges to a better value quickly.

[0014] Step 1.3: When the Intersection over Union (IoU) of the identification and segmentation results in the training model is greater than the set threshold, the model correctly identifies and segments the target. The precision and recall are calculated to evaluate the model. Once the evaluation meets the set standard, the deep learning model for crack identification and segmentation is completed.

[0015] Furthermore, in step 2, the calibration information is obtained by calculating the crack mask pixel information obtained from the Mask R-CNN crack recognition and segmentation deep learning model. The method for converting the position information of the crack tip into the actual crack position information on the surface of the structural component is as follows:

[0016] First, determine the initial geometric information of the structural component's geometric model, including the structural geometry without cracks and the position information of the calibration plate, where L is the length of the diagonal of the calibration plate, and x0 and y0 are the Cartesian coordinates of the calibration plate's center in the x and y directions, respectively. Then, obtain the pixel coordinates (u, v) of the crack detected by the Mask R-CNN crack recognition and segmentation deep learning model, the pixel length l of the calibration plate's diagonal, and the pixel coordinates (u0, v0) of the calibration plate's center position. Finally, derive the crack's Cartesian coordinates (x, y), with the transformation relationship as follows:

[0017]

[0018] The above formula is used to obtain the geometric location information of the crack in the structure in the Cartesian coordinate system, and the coordinates of the two ends of the crack are obtained, thereby determining the actual crack location information on the surface of the structural component.

[0019] X-FEM was used to digitally model the cracked structure. By using Abaqus software for secondary development, the actual crack endpoint location information was exchanged with the offline fracture mechanics model to construct a geometric model of the cracked structure that includes crack location information, structural geometry information, material property parameters, boundary conditions, mesh generation, etc.

[0020] Furthermore, in step 3, the size and material properties of the crack structure geometric model are kept consistent with those of the experimental specimen during offline simulation. The steps for constructing the deep learning model for crack damage assessment are as follows:

[0021] Step 3.1: Extensive offline simulations were conducted. By randomly defining the crack initiation and termination points and the magnitude of the applied load, different combinations of cracks at different locations and loads of different magnitudes were randomly generated. The x and y coordinates of the crack initiation and termination points and the load magnitudes were defined to conform to a uniform distribution. For different combinations, the stress intensity factor was calculated using X-FEM parametric modeling technology. A database for training a deep learning model for crack damage assessment was constructed, and the data was divided into training set, validation set, and test set.

[0022] Step 3.2: Construct a deep learning model for crack damage assessment based on the database obtained in Step 3.1. The neural network in the model is a multilayer perceptron (MLP), where the input is the crack endpoint location and load magnitude, and the output is the stress intensity factor value. The activation function is the ReLU function. An initial learning rate is set to ensure rapid convergence to the optimal value. The learning rate is gradually reduced as training progresses. The predictive ability of the model is calculated and evaluated using the following formula:

[0023]

[0024] R 2 y represents the variance of the model. i For the true value of the i-th data point, For predicted values, R is the average value. 2 The value of is between 0 and 1, and the closer it is to 1, the better the model is.

[0025] Furthermore, the offline fracture mechanics model continuously constructs crack geometry models through online interactive data, providing samples for the online deep learning damage assessment model and enriching its database. At the same time, the online deep learning damage assessment model completes real-time prediction of the crack geometry model from the previous section.

[0026] Furthermore, in step 4, load optimization is performed based on real-time crack damage assessment to prevent structural fracture during operation and ensure the load conditions required for the mission. The load optimization is expressed as follows:

[0027]

[0028] In the formula, f is the safety factor, W is the minimum load when performing the task. There are uncertainties in the load optimization process, so the value of f is set to 0.7 to ensure the safe service of the cracked structural component.

[0029] The beneficial effects of this invention are:

[0030] (1) This invention integrates automatic detection of fatigue cracks, damage assessment and external load optimization into one system, performs automatic damage assessment and load optimization on cracked structures in real time, monitors the working status of equipment in a timely manner, reduces the load conditions to a safe range, ensures smooth operation, and reduces the risk of operation.

[0031] (2) This invention integrates sensory information and physical geometric model, automatically identifies cracks and reconstructs the digital model of cracks, and uses the extended finite element analysis method which is convenient for digital modeling to model cracks. Once crack information is sensed during the identification process, a geometric model of the cracked structure consistent with the actual situation will be established.

[0032] (3) This invention realizes a dual-line framework of online real-time crack damage assessment and load optimization, and offline physical model to improve the prediction accuracy and generalization of deep learning model. It meets the real-time prediction requirements and uses the simulation data of offline physical model to enrich the database of deep learning model to enhance its prediction generalization ability.

[0033] (4) The intelligent damage assessment and analysis method for cracked structures proposed in this invention embeds intelligence into the fracture analysis model for damage assessment. It can realize intelligent damage identification, assessment and load optimization of flat cracked structural components within seconds, quickly complete the work and improve efficiency. Attached Figure Description

[0034] Figure 1 This is a schematic diagram of the framework for an intelligent crack damage assessment method based on the fusion of visual information and physical properties.

[0035] Figure 2 This is a schematic diagram of crack segmentation in an intelligent crack damage assessment method based on the fusion of visual information and physics.

[0036] Figure 3 The training loss and accuracy of the test set are considered in the crack segmentation training process of an intelligent crack damage assessment method based on the fusion of visual information and physics.

[0037] Figure 4 This is a schematic diagram of crack model reconstruction for an intelligent crack damage assessment method based on the fusion of visual information and physics.

[0038] Figure 5 This is a schematic diagram of a machine learning damage assessment database for a physical computation model of a crack damage intelligent assessment method based on the fusion of visual information and physics.

[0039] Figure 6 This is a schematic diagram of a crack damage intelligent assessment method based on the fusion of visual and physical information, specifically a physical fusion-based damage assessment method.

[0040] Figure 7 This is the training loss and test set results during the training process of an MLP model for an intelligent crack damage assessment method based on the fusion of visual information and physics.

[0041] Figure 8 This is a schematic diagram of a partially reconstructed finite element model of a crack damage intelligent assessment method based on the fusion of visual information and physics.

[0042] Figure 9 This is a comparison chart of load optimization results for an embodiment of an intelligent crack damage assessment method based on the fusion of visual information and physical properties. Detailed Implementation

[0043] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments of the present invention and the features thereof can be combined with each other.

[0044] In the following description, many specific details are set forth in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0045] This embodiment provides a crack damage intelligent assessment method based on the fusion of visual information and physical properties, including:

[0046] Step 1: Collect photos of training specimens to form a training database, train a deep learning model for crack recognition and segmentation, automatically recognize and segment the input images, and detect calibration information.

[0047] Step 1.1: Design 10 plate test specimens with different initial crack locations. Each specimen is 100 mm long, 40 mm wide, and 3 mm thick. The specimen material is 316L steel with an elastic modulus of 206 GPa, Poisson's ratio of 0.3, yield strength of 176 MPa, tensile strength of 485 MPa, and fracture toughness of [missing information]. The experimental specimens were clamped at both ends. For each specimen, fatigue cyclic loading was applied to create crack specimens with different crack lengths. The crack specimens were photographed using a camera to obtain images of cracks of different lengths under different conditions. The images captured by the camera were 6000×4000 pixels, and a total of 305 valid crack images were collected. During the experiment, the cracks and square calibration plates in the collected images were further labeled. The crack images and labels together formed a dataset. The collection of all data constituted the database for the Mask R-CNN crack recognition and segmentation deep learning model. This database was randomly divided into a training set (275 images) and a test set (30 images).

[0048] Step 1.2: As Figure 2 As shown, the basic content of automatic crack recognition and segmentation based on Mask R-CNN includes three main parts. First, the input image of 6000×4000 pixels is scaled down to 1200×800 pixels. To improve prediction accuracy, a pre-trained ResNet-50 network based on the Microsoft COCO dataset is used as the backbone network to extract features from the image information, generating feature maps of sizes 304×200×256, 152×100×256, 76×50×256, 38×25×256, and 19×13×256. Second, a Region Proposal Network (RPN) is used to generate regions of interest (ROIs). The RPN network generates several anchor boxes by performing convolution operations on the input feature maps and assigns a score to each anchor box. The score indicates the likelihood that the anchor box contains the detected target; anchor boxes with higher scores are selected as regions of interest (ROIs). Finally, these generated ROIs are pooled using the RoIAlign layer, producing feature vectors of sizes 7×7×256 and 14×14×256, respectively. The former is used for the target classification (class) and detection (box) branches, enabling target detection and classification; the latter is used for the mask branch, generating mask processing. The loss function value and learning rate during model training are as follows: Figure 3 As shown in (a). Figure 3 (b) Demonstrates the model's training accuracy. During training, the batch size was set to 4, the initial learning rate was set to 4E-3, and the hyperparameters momentum and weight decay were set to 0.9 and 1E-04, respectively. To converge to an optimal value as quickly as possible, the learning rate was decayed to 4E-4 at the 20th training epoch and to 4E-5 at the 40th training epoch. The loss function value and learning rate during model training are shown below. Figure 3 As shown, the results indicate that the loss function converged after 60 epochs. Therefore, 60 epochs were used in subsequent tests. Image recognition in the database first identifies cracks and calibration boards using bounding boxes. Based on this, each pixel within the bounding box is analyzed to determine if it belongs to the target category. If it is a detected target, the pixel is masked to achieve segmentation, thus performing pixel-level segmentation of the crack and calibration board information within the bounding box.

[0049] Step 1.3: When the Intersection over Union (IoU) of the identification and segmentation results is greater than a certain threshold, the target is considered to have been correctly identified and segmented. Here, the threshold is set to 0.5. The model is evaluated by calculating its precision and recall. Precision refers to the proportion of correctly identified and segmented targets out of all predicted targets, while recall refers to the proportion of correctly identified and segmented targets out of all true targets. The average precision (AP) for this category can be calculated based on the area under the precision-recall curve. The mean precision (mAP) is obtained by averaging the AP values ​​of all categories. The final crack identification model has an mAP of 96.4%, indicating that the cracks and calibration plates segmented by the proposed method are consistent with reality.

[0050] Step 2: Using the detected crack and calibration plate information, the location information of the crack tip is converted into the actual crack location information on the surface of the structural component. This location information is then interacted with the offline fracture mechanics model module, and a geometric model containing the crack is constructed in combination with the geometric model database.

[0051] Information about the crack and calibration plate was obtained from the captured images. The coordinates of the crack's initiation and termination points were calculated using the crack mask pixel information obtained from the Mask R-CNN crack recognition and segmentation deep learning model. First, the difference between the maximum and minimum values ​​of the crack mask information was calculated in both the vertical and horizontal directions. Then, edge pixels in the directions with larger differences were selected, and the average pixel coordinates of these pixels were obtained as the endpoint pixel coordinates of the crack.

[0052] Specific examples Figure 4 As shown, the black and white square plate in the center of the geometric model is the calibration plate, using a checkerboard design. The left side displays the initial geometric information of the model, including the structural geometry without cracks and the position information of the calibration plate, where L is the length of the diagonal of the calibration plate, and x0 and y0 are the Cartesian coordinates of the calibration plate's center in the x and y directions, respectively. The center displays the pixel information of the cracks obtained through the Mask R-CNN recognition and segmentation model, the pixel length l of the calibration plate's diagonal, and the pixel coordinates (u0, v0) of the calibration plate's center. The transformation relationship between the crack's Cartesian coordinates (x, y) and its pixel coordinates (u, v) can be derived as follows:

[0053]

[0054] The above formula can be used to obtain the geometric location information of the crack in the structure in the Cartesian coordinate system, and the coordinates of the two ends of the crack can be obtained. Table 1 shows the actual crack tip coordinates of some samples and the crack tip coordinates detected by the crack recognition and segmentation deep learning model. The average error between the coordinates extracted by the Mask R-CNN crack recognition and segmentation deep learning model and the actual coordinates is 2.22%. It can be found that the generated digital model has good consistency with the actual structure.

[0055] Table 1. Crack tip detection table for partially reconstructed models

[0056]

[0057] Based on the crack location information mentioned above, X-FEM is used to digitally model the cracked structure. Through secondary development using Abaqus software, the actual crack endpoint location information is written into the geometric model, establishing a finite element geometric model that includes crack location information, structural geometric information, material property parameters, boundary conditions, mesh generation, etc. Through this process, once the offline fracture mechanics physical model detects crack information, it can establish a geometric model of the cracked structure consistent with reality, such as... Figure 8 As shown in the figure, the corresponding crack geometry model is generated.

[0058] Step 3: Conduct extensive offline simulation analysis of crack fracture mechanics models to obtain stress intensity factors under different crack locations and loads, forming a database for training deep learning models of crack damage. A deep learning model of crack damage is constructed using this database to predict the stress intensity factors corresponding to the online fracture geometry model in real time. The crack geometry models constructed from the offline fracture mechanics models continuously enrich the database of deep learning models for crack damage assessment. The specific steps are as follows:

[0059] Step 31: During the offline simulation, the geometric model of the cracked structure is generated, and the structural geometry and material properties remain consistent with the crack-free model from Step 2. For example... Figure 5As shown, simulation analysis was performed using a crack fracture mechanics model. By randomly defining the crack initiation and termination points and the applied load size, different combinations of crack locations and load sizes were randomly generated. The x and y coordinates of the crack initiation and termination points and the load size were defined to conform to a uniform distribution. Based on the actual geometry and load information of the studied structure, the generated x and y coordinate ranges of the crack initiation and termination points were (0, 40) and (20, 80), respectively, and the generated load size range was (10, 22). For different combinations, the crack tip stress intensity factor under different crack geometry and load conditions was obtained using XFEM parametric modeling technology. During the calculation, a dataset of 12,120 samples was created to form a database for training the crack damage depth assessment learning model. The input was the crack endpoint location and load size, and the output was the crack tip stress intensity factor value. 70% of the dataset was randomly selected as the training set, 20% as the validation set, and 10% as the test set, which were used to train, calibrate, and validate the crack damage depth assessment learning model, respectively.

[0060] Step 32: Based on the obtained database, a deep learning model for crack damage assessment is constructed to achieve rapid intelligent assessment. The neural network in the deep learning model is a multilayer perceptron (MLP), which consists of an input layer, a hidden layer, and an output layer, with full connectivity between different layers. In this paper, the input layer is a 5-dimensional vector, representing the two-dimensional coordinates and load magnitudes of the two crack endpoints; the hidden layer consists of two layers of 128 neurons; and the output layer is 1-dimensional, outputting the stress intensity factor. The ReLU function is selected as the activation function. During the damage function optimization process, to better and faster converge to the optimal value, the initial learning rate is set to 0.001, and as training progresses, the learning rate is gradually reduced, decaying to 1E-04 in the 150th cycle, 1E-05 in the 264th cycle, and 1E-06 in the 285th cycle. Figure 7(a) shows the loss function and learning rate during the training process, and the test results are as follows. Figure 7 As shown in (b), the mean squared error (MSE) on the test set is 1.42E-04. The predictive ability of this damage assessment model is evaluated by calculation, and the formula for this indicator is as follows:

[0061]

[0062] R 2 y represents the variance of the model. i For the true value of the i-th data point, For predicted values, R is the average value. 2 The value ranges from 0 to 1. The closer it is to 1, the better the model is. The value calculated on the test set is 0.989, which indicates that the evaluation model has high prediction accuracy and stability.

[0063] To achieve online crack damage identification, crack damage assessment comprises two phases: online and offline. In the offline phase, a deep learning model for crack damage assessment is constructed based on a fracture mechanics model simulation database to meet real-time predictive requirements. This machine learning model connects possible crack damage scenarios with their severity. In the online phase, the deep learning model for crack damage assessment performs real-time evaluation of the crack geometry model from the previous section. Simultaneously, the fracture mechanics model constructs a crack geometry model using the acquired online data to enrich the database of the deep learning model for crack damage assessment, thereby enhancing its predictive generalization ability. Figure 6 As shown, offline fracture mechanics model analysis continuously provides samples for online deep learning model, while online crack deep learning damage assessment model is used to complete real-time prediction of crack damage.

[0064] Step 4: Crack damage analysis based on the fracture damage module. The load is optimized using an optimization algorithm, and the optimization results are fed back to the crack structure control system to prevent structural fracture during operation while ensuring the required load conditions for the mission. The load conditions are reduced to a safe range. This load optimization can be expressed as follows:

[0065]

[0066] In the formula, f represents the safety factor, and W represents the minimum load during the task. Through the above optimization, the component can be prevented from fracturing even when fatigue cracks appear, and its normal operation during service can be guaranteed. During load optimization, considering the existence of uncertainties, the safety factor f is set to 0.7 to ensure the safe service of the cracked structural component.

[0067] The results of automatic crack identification, evaluation, and load optimization for 10 real cracked structures are as follows: Figure 9As shown in the figures, Figure (a) compares the load magnitudes before and after adjustment, with the circle representing the initial applied load and the triangle representing the adjusted load. Figure (b) shows the comparison of SIF before and after adjustment, with the circle representing the evaluated SIF value and the triangle representing the adjusted SIF value. The upper dashed line represents the threshold line, and the lower dashed line represents the set safety line. It can be seen from the figures that the critical crack component 2 was significantly reduced, thus ensuring the current operational safety of the structure. During the load assessment of the 10 cracked structures, the times for completing the entire process of intelligent crack identification, assessment, and load optimization were: 0.672 seconds, 0.537 seconds, 0.528 seconds, 0.537 seconds, 0.526 seconds, 0.530 seconds, 0.533 seconds, 0.529 seconds, 0.530 seconds, and 0.528 seconds, respectively. Statistical analysis shows that the average assessment time for each structure is 0.545 seconds, enabling rapid damage assessment and load optimization of cracked structures.

[0068] In summary, the intelligent crack damage assessment method based on the fusion of visual information and physical model enables intelligent damage assessment of cracks in specimens at arbitrary crack locations. By fusing perceptual information and physical models, it achieves automatic crack identification and automatic reconstruction of the digital crack model, realizing a dual-track framework of online real-time crack damage assessment and load optimization, and offline physical model improvement of deep learning model prediction accuracy and generalization.

[0069] Although the invention has been disclosed in detail with reference to the accompanying drawings, it should be understood that these descriptions are merely exemplary and not intended to limit the application of the invention. The invention is not limited to the specific embodiments described above. Those skilled in the art can implement the invention using various other specific embodiments based on the disclosed content of the embodiments and drawings. Therefore, any design that adopts the design structure and concept of the invention and makes some simple modifications or alterations falls within the scope of protection of the invention.

Claims

1. A method for intelligent assessment of crack damage based on the fusion of visual information and physical properties, characterized in that, The method includes: Step 1: Construct a deep learning model for crack recognition and segmentation to automatically recognize and segment the input image; Step 2: Using the detected calibration information, the crack location information is converted into the actual crack location information on the surface of the structural component. The actual crack location information is then fed into the offline fracture mechanics model to construct the crack geometry model; this includes the following steps: Step 2.1: The calibration information is obtained by calculating the crack mask pixel information obtained from the Mask R-CNN crack recognition and segmentation deep learning model. The method for converting the crack tip position information into the actual crack position information on the surface of the structural component is as follows: First, determine the initial geometric information of the structural component's geometric model, including the structural geometry information without cracks and the position information of the calibration plate, where L is the length of the diagonal of the calibration plate. and These are the Cartesian coordinates of the calibration board center in the x and y directions, respectively. Then, the pixel coordinates of the cracks detected by the Mask R-CNN crack recognition and segmentation deep learning model are obtained. Pixel length of the calibration board diagonal and pixel coordinates of the center position of the calibration board Finally, the Cartesian coordinates of the crack were obtained. The conversion relationship is as follows: ； The above formula is used to obtain the geometric location information of the crack in the structural component in the Cartesian coordinate system, and the coordinates of the two ends of the crack are obtained, thereby determining the actual crack location information on the surface of the structural component. Step 2.2: Use Abaqus software for secondary development, use extended finite element method X-FEM to digitally model the structure containing cracks, exchange the actual crack endpoint location information with the offline fracture mechanics model, and combine the geometric model database to construct a geometric model of the cracked structure that includes crack location information, structural geometric information, material property parameters, boundary conditions, and mesh generation. Step 3: Construct a deep learning model for crack damage assessment, perform real-time prediction on the online crack geometry model, and continuously enrich the crack damage assessment deep learning model database by constructing the crack geometry model from the offline fracture mechanics model. Step 4: Optimize the load and reduce the load conditions to a safe range.

2. The intelligent crack damage assessment method based on the fusion of visual information and physical properties according to claim 1, characterized in that, Step 1, which involves constructing a deep learning model for crack recognition and segmentation, includes the following steps: Step 1.1: Select experimental specimens of uniform size and material to make different crack specimens, collect effective crack images, label the cracks and square calibration plates in the images, and the crack images and labels together form a set of data. The collection of all data constitutes the database of the Mask R-CNN crack recognition and segmentation deep learning model. The database is divided into training set and test set. Step 1.2: Train the deep learning model for crack recognition based on the database described in Step 1.

1. Determine the batch size for training according to the amount of data, and adjust the initial learning rate, hyperparameter momentum, and weight decay to ensure that the training model converges to a better value quickly. Step 1.3: When the Intersection over Union (IoU) of the identification and segmentation results in the training model is greater than the set threshold, the training model correctly identifies and segments the target. The precision and recall are calculated to evaluate the training model. After the evaluation meets the set standard, the deep learning model for crack identification and segmentation is completed.

3. The intelligent crack damage assessment method based on the fusion of visual information and physical properties according to claim 1, characterized in that, In step 3, the size and material properties of the crack structure geometric model are kept consistent with the experimental specimen during offline simulation. The steps for constructing the deep learning model for crack damage assessment are as follows: Step 3.1: Extensive offline simulations were conducted. By randomly defining the crack initiation and termination points and the magnitude of the applied load, different combinations of cracks at different locations and loads of different magnitudes were randomly generated. The x and y coordinates of the crack initiation and termination points and the load magnitudes were defined to conform to a uniform distribution. For different combinations, the stress intensity factor was calculated using X-FEM parametric modeling technology. A database for training a deep learning model for crack damage assessment was constructed, and the data was divided into training set, validation set, and test set. Step 3.2: Construct a deep learning model for crack damage assessment based on the database obtained in Step 3.

1. The neural network in the model is a multilayer perceptron (MLP), where the input is the crack endpoint location and load magnitude, and the output is the stress intensity factor value. The activation function is the ReLU function. An initial learning rate is set to ensure rapid convergence to the optimal value. The learning rate is gradually reduced as training progresses. The predictive ability of the model is calculated and evaluated using the following formula: ； The variance of the model. For the first The true value of each data point. For predicted values, This is the average value. The value of is between 0 and 1, and the closer it is to 1, the better the model is.

4. The intelligent crack damage assessment method based on the fusion of visual information and physical properties according to claim 3, characterized in that, The offline fracture mechanics model continuously builds crack geometry models through online interactive data, providing samples for the online deep learning damage assessment model and enriching its database. At the same time, the online deep learning damage assessment model completes real-time prediction of the crack geometry model in the previous section.

5. The intelligent crack damage assessment method based on the fusion of visual information and physical properties according to claim 1, characterized in that, In step 4, load optimization is performed based on real-time crack damage assessment to prevent structural fracture during operation and ensure the load conditions required for the mission. The load optimization is expressed as follows: ； In the formula For safety, W represents the minimum load when performing the task. Uncertainties exist during load optimization. The value is set to 0.7 to ensure the safe service of cracked structural components.

Images