CN117150838A

CN117150838A - Crack damage intelligent assessment method based on visual information and physical fusion

Info

Publication number: CN117150838A
Application number: CN202310845567.8A
Authority: CN
Inventors: 龙湘云; 姜潮; 丁鑫烽; 廖望望
Original assignee: Hunan University
Current assignee: Hunan University
Priority date: 2023-07-11
Filing date: 2023-07-11
Publication date: 2023-12-01

Abstract

The invention discloses a crack damage intelligent assessment method based on visual information and physical fusion, which relates to the technical field of material fatigue strength, and comprises the steps of automatically identifying and segmenting pixel-level information of fatigue cracks through a mask area convolutional neural network, calculating crack coordinate information by using camera calibration information to generate a digital geometric model containing cracks, fusing the perception information and the crack geometric model together, continuously providing data by using an off-line high-precision fracture mechanics physical model, evaluating crack damage states in real time by using a crack damage evaluation deep learning model, and finally carrying out load optimization to reduce load working conditions to a safe area; according to the method, the automatic detection, damage evaluation and external load optimization of the fatigue cracks are integrated into one system, so that the real-time evaluation and load optimization of the on-line crack damage are realized.

Description

Crack damage intelligent assessment method based on visual information and physical fusion

Technical Field

The invention relates to the technical field of material fatigue strength, in particular to an intelligent crack damage assessment method based on visual information and physical fusion.

Background

Mechanical damage such as fatigue cracks and the like widely exist in a mechanical structure due to factors such as extreme service environment, manufacturing defects and the like, and a large number of fracture accidents such as aircraft blade fracture, oil pipe fracture, rail fracture and the like are caused. Along with the rapid development of technology, informatization and intellectualization of equipment detection and maintenance have become development trends, such as carrier rockets, on-orbit satellites, outer star detectors, nuclear power systems and other large-scale high-end complex equipment often need to operate in unmanned environments such as vacuum, irradiation and the like, and urgent demands are put forward on automation and intellectualization of equipment fatigue crack damage state monitoring. Information strength theory with informatization, intelligence and quick response capability based on artificial intelligence, edge calculation and the like has become a research hot spot in the field of modern mechanical strength.

In order to evaluate the safety state of a crack structure, many scholars have studied the crack tip strength of a crack-containing member. Classical methods are based on fracture mechanics to evaluate the damage status of crack structures. In the 50 s of the 20 th century, irwin assessed the magnitude of stress concentration at the tip of a crack by introducing a stress intensity factor, and laid the foundation of the line elastic fracture mechanics. Subsequently, in order to solve the engineering fracture problem, a number of fracture analysis numerical calculation methods, such as a fracture finite element method, an extended finite element method, a boundary finite element method, a proportional boundary finite element, and the like, have been developed to calculate the crack tip stress intensity factor. However, these methods typically involve the processing of crack tip singular cells and grid repartition, and lack fusion with perceptual information, making it difficult to accomplish automated and real-time assessment of crack tip damage. Aiming at cracks identified in the service stage, how to intelligently embed the cracks into a fracture analysis model for damage evaluation, so as to realize automatic identification, evaluation and load regulation of fatigue cracks, and the method is still one of the serious difficulties in the research of the field.

In recent years, with the development of artificial intelligence technology, deep learning methods have been widely used in the fields of computer vision, natural language processing, and the like. Due to their advantages of being able to automatically extract features and being conveniently integrated in mobile devices, deep learning methods are also increasingly being applied in the field of fatigue and fracture. However, these efforts have mainly studied whether or not a crack exists in the structure, and cannot evaluate the current safety state of the crack member. The fusion of information and fracture mechanics is an effective way to solve the problem of damage assessment of a crack structural member. The united states space agency has earlier proposed a digital twinning (digital twinning) based technology, and adopts an ultra-high fidelity simulation and health management system to realize the safety and reliability level in the service process of the aircraft. Overall, the field is still under exploration, especially in terms of fusion of perceptual information with physical models. Aiming at the intelligent damage evaluation problem of cracks suddenly appearing on a piece without cracks, how to automatically identify the cracks and effectively fuse crack perception information into a physical model to monitor the damage state of the cracks in real time, and research on related aspects still needs to be developed.

Disclosure of Invention

The invention aims at: the intelligent damage assessment analysis method for the crack structure solves the problems that the conventional crack structure is insufficient in perceived information fusion, is difficult to automatically and timely assess crack tip damage and the like, and realizes automatic identification, assessment and load optimization of the surface cracks of the structure, so that the structure is prevented from breaking in the operation process, and the normal operation of the service process is ensured.

The technical scheme of the invention is as follows: the utility model provides a crack damage intelligent assessment method based on visual information and physical fusion, which comprises the following steps:

step 1: constructing a crack recognition and segmentation deep learning model, and automatically recognizing and segmenting an input image;

step 2: converting the position information of the crack into real crack position information of the surface of the structural member by using the detected calibration information, and interacting the real crack position information to an offline fracture mechanical model to construct a crack geometric model;

step 3: constructing a crack damage evaluation deep learning model, predicting an online crack geometric model in real time, and continuously enriching a crack damage evaluation deep learning model database by using the crack geometric model constructed by the offline fracture mechanical model;

step 4: and (5) carrying out load optimization, and reducing the load working condition to a safe area.

Further, the step 1 of constructing the crack recognition and segmentation deep learning model comprises the following steps:

step 1.1: selecting experimental test pieces with uniform size and materials to prepare crack test pieces with different crack lengths, collecting effective crack pictures, labeling cracks and square calibration plates in the pictures, forming a group of data by the crack pictures and the labels, and forming a database of a Mask R-CNN crack recognition and segmentation deep learning model by the collection of all the data, wherein the database is divided into a training set and a testing set;

step 1.2: training the crack recognition deep learning model based on the database in the step 1.1, determining the size of the trained batch_size according to the data quantity, and adjusting the initial learning rate, the super-parameter momentum model and the weight attenuation weight_decay to ensure that the training model is quickly converged to a better value;

step 1.3: when the intersection ratio IoU of the recognition and segmentation results in the training model is larger than a set threshold, the model correctly recognizes and segments the target, the model is evaluated by calculating the accuracy and the recall rate, and after the evaluation reaches the set standard, the crack recognition and segmentation deep learning model is constructed;

in step 2, the calibration information is obtained by calculating crack Mask pixel information obtained by Mask R-CNN crack recognition and segmentation of a deep learning model, and the manner of converting the position information of the crack tip into real crack position information of the surface of the structural member is as follows:

firstly, initial geometric information of a structural part geometric model is determined, wherein the initial geometric information comprises structural geometric information without cracks and position information of a calibration plate, L is the diagonal length of the calibration plate, and x is ₀ And y ₀ Respectively obtaining Cartesian coordinates of the center of the calibration plate in the x and y directions, and then obtaining pixel coordinates (u, v) of cracks detected by the Mask R-CNN crack recognition and segmentation deep learning model, pixel length l of the diagonal line of the calibration plate and pixel coordinates (u) of the center position of the calibration plate ₀ ,v ₀ ) Finally, the Cartesian coordinates (x, y) of the crack are obtained, and the conversion relation is as follows:

obtaining geometric position information of a crack in the structure under a Cartesian coordinate system through the method, and obtaining coordinates of two end points of the crack, thereby determining real crack position information of the surface of the structural member;

the X-FEM is used for digitally modeling a crack-containing structure, the real end point position information of the crack is interacted to an offline fracture mechanical model through secondary development by adopting Abaqus software, and a crack structure geometric model comprising crack position information, structure geometric information, material attribute parameters, boundary conditions, grid division and the like is constructed.

Further, in step 3, the size and material properties of the geometric model of the crack structure in the offline simulation are kept consistent with those of the experimental test piece, and the steps of constructing the crack damage evaluation deep learning model are as follows:

step 3.1: the method comprises the following steps of performing off-line mass simulation, randomly defining the starting point and the end point of a crack and the load of a force, randomly generating crack combinations of different positions and loads of different sizes, wherein the abscissa and the ordinate of the starting point and the end point of the crack are defined, the load sizes are uniformly distributed, calculating stress intensity factors corresponding to the crack combinations by combining the X-FEM parameterized modeling technology with the different combinations, and constructing a database for training a crack damage evaluation deep learning model, wherein the data are divided into a training set, a verification set and a test set;

step 3.2: based on the database obtained in the step 3.1, constructing a crack damage evaluation deep learning model, wherein a neural network in the model is a multi-layer perceptron MLP, wherein the input is crack endpoint position and load size, the output is stress intensity factor value, an activation function is selected as a ReLU function, an initial learning rate is set to ensure rapid convergence to an optimal value, the learning rate is gradually reduced along with the training, and the prediction capacity of the evaluation model is calculated according to the following formula:

R ² variance of model, y _i Is the true value of the i-th data,for predictive value +.>R is the average value of ² The value of (2) ranges from 0 to 1, with a value closer to 1 indicating a better model.

Further, the offline fracture mechanics model continuously builds a crack geometric model through online interactive data, provides a sample for the online deep learning damage assessment model, enriches the database of the model, and simultaneously completes real-time prediction of the previous section of crack geometric model through the online deep learning damage assessment model.

Further, in step 4, load optimization is performed on the basis of real-time crack damage evaluation, so that the structure is prevented from breaking in the running process, the load working condition required by the task is ensured, and the load optimization expression is as follows:

wherein f is a safety coefficient, the minimum load exists when the W executes the task, an uncertainty factor exists in the load optimization process, and the f value is set to 0.7, so that the safety service of the crack structural member is ensured.

The beneficial effects of the invention are as follows:

(1) According to the invention, the automatic detection, damage evaluation and external load optimization of the fatigue crack are integrated into one system, the automatic damage evaluation and load optimization are carried out on the crack structure in real time, the working state of the equipment is monitored in time, the load working condition is reduced to a safe area, the stable running process is ensured, and the running risk is reduced.

(2) The invention fuses the perception information and the physical geometric model, automatically recognizes the digital model of the crack, and then reconstructs the crack, and models the crack by adopting an expansion finite element analysis method which is convenient for digital modeling, and once the crack information is perceived in the recognition process, the geometric model with the crack structure which is consistent with the reality is established.

(3) The method realizes real-time evaluation and load optimization of the on-line crack damage, and the off-line physical model improves the prediction precision and generalization of the deep learning model, meets the real-time predictability requirement, and enriches the database of the deep learning model by using the simulation data of the off-line physical model so as to enhance the prediction generalization capability of the deep learning model.

(4) According to the intelligent damage evaluation analysis method for the crack structure, which is provided by the invention, the intelligent damage evaluation is performed by embedding the intelligent damage evaluation method into the fracture analysis model, so that the intelligent damage recognition evaluation, evaluation and load optimization of the flat crack structural member can be realized within seconds, the work can be rapidly completed, and the efficiency is improved.

Drawings

FIG. 1 is a schematic diagram of a crack damage intelligent assessment method based on visual information and physical fusion;

FIG. 2 is a schematic diagram of crack segmentation of an intelligent crack damage assessment method based on visual information and physical fusion;

FIG. 3 shows training loss and accuracy of a test set in a crack segmentation training process based on a crack damage intelligent assessment method based on visual information and physical fusion;

FIG. 4 is a schematic diagram of a crack model reconstruction of an intelligent crack damage assessment method based on visual information and physical fusion;

FIG. 5 is a schematic diagram of a machine learning damage assessment database of a physical calculation model of a crack damage intelligent assessment method based on visual information and physical fusion;

FIG. 6 is a schematic diagram of a visual information and physical fusion based damage assessment of a crack damage intelligent assessment method based on the physical fusion of perception information;

fig. 7 is a training loss and test set result in the MLP model training process of the crack damage intelligent evaluation method based on visual information and physical fusion.

FIG. 8 is a schematic diagram of a partially reconstructed finite element model of a crack damage intelligent assessment method based on visual information and physical fusion;

fig. 9 is a load optimization result comparison chart of one embodiment of a crack damage intelligent assessment method based on visual information and physical fusion.

Detailed Description

In order that the above-recited objects, features and advantages of the present invention will be more clearly understood, a more particular description of the invention will be rendered by reference to the appended drawings and appended detailed description. It should be noted that, without conflict, embodiments of the present invention and features in the embodiments may be combined with each other.

In the following description, numerous specific details are set forth in order to provide a thorough understanding of the present invention, however, the present invention may be practiced in other ways than those described herein, and the scope of the invention is therefore not limited to the specific embodiments disclosed below.

The embodiment provides an intelligent crack damage assessment method based on visual information and physical fusion, which comprises the following steps:

step 1: and collecting a test piece photo for training to form a training database, training a crack recognition and segmentation deep learning model, automatically recognizing and segmenting an input image, and detecting calibration information.

Step 1.1: design of 10 plate experiments with different initial crack positionsThe length of the test piece is 100mm, the width is 40mm, and the thickness is 3mm. The test piece is made of 316L steel, the elastic modulus is 206GPa, the Poisson ratio is 0.3, the yield strength is 176MPa, the tensile strength is 485MPa, and the fracture toughness isThe two ends of the experimental test piece are clamped. And (3) aiming at each test piece, manufacturing crack test pieces with different crack lengths through fatigue cyclic loading. And shooting the crack test piece by using a camera to obtain crack pictures with different lengths in different situations. The pixels of the picture taken by the camera are 6000 x 4000, and 305 effective crack images are collected. The crack and square calibration plate of the picture collected in the test process are further marked with labels, the crack picture and the labels form a group of data, the collection of all the data forms a database of Mask R-CNN crack recognition and segmentation deep learning models, the database is randomly divided into a training set and a testing set, the training set is 275 pictures, and the testing set is 30 pictures.

Step 1.2: as shown in fig. 2, the Mask R-CNN based crack automatic recognition and segmentation basic content includes three main parts, and an input picture of 6000×4000 pixels is first scaled to 1200×800 pixels. In order to improve prediction accuracy, a backbone network backbone based on a Microsoft COCO dataset pre-training ResNet-50 network is adopted to perform Feature extraction on picture information, and Feature maps with the sizes of 304×200×256, 152×100×256, 76×50×256, 38×25×256 and 19×13×256 are generated respectively; next, regional advice network RPN is used to generate regional probes of interest. The RPN network generates a plurality of anchor boxes by convolving the input feature map and assigns a score to each anchor box. The score indicates that the anchor box may contain a score of the detection target, and the anchor box with high score is selected as the region of interest; finally, these generated regions of interest are pooled using the RoIAlign layer to generate feature vectors of 7×7×256 and 14×14×256 sizes, respectively. The former is used for classifying objects and detecting box branches, and can detect and classify the objects; the latter is used to mask branches, which may generate masking processes. The loss function value and learning rate in the model training process are shown in fig. 3 (a), and fig. 3 (b) shows the training accuracy of the model. In the training process, the size of the trained batch_size is set to 4, the initial learning rate is set to 4E-3, and the super-parametric momentum and weight decay weight are set to 0.9 and 1E-04, respectively. To converge to the optimum value as soon as possible, the learning rate decays to 4E-4 at the 20 th training period and to 4E-5 at the 40 th training period. The loss function value and the learning rate in the model training process are shown in fig. 3. The results show that the loss function value reached convergence after 60 epochs. Thus, 60 epochs were used in the subsequent tests. Identification of pictures in the database the Crack Crack and calibration plate Calibration board are first identified by a suggestion box. On this basis, whether the object class is detected by judging each pixel in the bounding box. And if the detection target is the detection target, masking the pixel to achieve the segmentation effect, so that the crack in the boundary box and the calibration plate information are segmented at the pixel level.

Step 1.3: when the intersection ratio IoU of the recognition and division results is greater than a certain threshold, the target is considered to be correctly recognized and divided, where the threshold is set to 0.5. The model is evaluated by calculating the accuracy of the model, which refers to the proportion of the correctly identified and segmented target number of the model to all the predicted target numbers, and the recall, which refers to the real proportion of the correctly identified and segmented target number of the model to all the real target numbers. The average precision value AP for that class can be calculated from the area under the precision-recall curve. And calculating the average value of the AP values of all the categories to obtain an average accuracy average value mAP. The mAP of the final crack recognition model is 96.4%, which shows that the crack and the calibration plate segmented by the method have consistency with the actual.

Step 2: and converting the position information of the crack tip into the real crack position information of the surface of the structural member by using the detected crack and the calibration plate information, and interacting the position information to an offline fracture mechanical model module to construct a geometric model containing cracks by combining a geometric model database.

And obtaining the information of the crack and the calibration plate from the shot picture, and calculating the coordinate information of the starting end point of the crack and the position of the crack terminal through Mask R-CNN crack identification and crack Mask pixel information obtained by dividing a deep learning model. Firstly, respectively calculating the difference value between the maximum value and the minimum value of the crack mask information in the vertical direction and the horizontal direction, then selecting edge side pixel points in the direction with larger difference value, and obtaining the average value of the pixel coordinates of the pixel points to be used as the endpoint pixel coordinates of the crack.

Specifically, as shown in fig. 4, a square plate with black and white alternate in the middle of the geometric model is a calibration plate, adopts a checkerboard design, and the left side is initial geometric information of the geometric model, including structural geometric information without cracks and position information of the calibration plate, wherein L is the diagonal length of the calibration plate, and x is ₀ And y ₀ The Cartesian coordinates of the calibration plate center in the x and y directions, respectively. The middle is the pixel information of the crack obtained by the mask-CNN recognition and segmentation model, the pixel length l of the diagonal line of the calibration plate and the pixel coordinate (u) of the central position of the calibration plate ₀ ,v ₀ ). The conversion between the Cartesian coordinates (x, y) of the crack and its pixel coordinates (u, v) can be deduced as follows:

the geometric position information of the crack in the structure under the Cartesian coordinate system can be obtained through the method, and the coordinates of two endpoints of the crack are obtained. Table 1 shows the actual detected crack tip coordinates of a part of the test sample and the crack tip coordinates detected by the crack recognition and segmentation deep learning model, and the average error between the coordinates extracted based on the Mask R-CNN crack recognition and segmentation deep learning model and the actual coordinates was 2.22%. The generated digital model can be found to have better consistency with the actual structure.

Table 1 partial reconstruction model crack tip test table

Based on the crack location information, the crack-containing structure is digitally modeled using an X-FEM. By adopting Abaqus software to carry out secondary development, the real end point position information of the crack is written into the geometric model, and a finite element geometric model comprising the crack position information, the structural geometric information, the material attribute parameters, the boundary conditions, the grid division and the like is established. Through the above flow, once crack information is perceived by the offline fracture mechanics physical model, a crack-containing structural geometric model which is consistent with reality can be established, as shown in fig. 8, which is illustrated as generating a corresponding crack geometric model.

Step 3: and carrying out simulation analysis on a large number of crack fracture mechanical models under the line to obtain stress intensity factors at different crack positions and loads, and forming a database for training a crack damage deep learning model. And constructing a crack damage deep learning model through a database, and predicting the stress intensity factor corresponding to the on-line fracture geometric model in real time. Crack geometry model constructed by the offline fracture mechanics model continuously enriches a crack damage assessment deep learning model database. The method comprises the following specific steps:

step 31: in the generation process of the crack structure geometric model sample in the off-line simulation, the structure geometric model and the material property are kept consistent with those of the crack-free model in the step 2. As shown in fig. 5, simulation analysis is performed on a crack fracture mechanical model, and by randomly defining the starting point and the end point of the crack and the magnitude of the load L load applied, crack combinations with different positions and different magnitudes of loads are randomly generated, wherein the abscissa and ordinate defining the starting point and the end point of the crack and the magnitude of the load are consistent with uniform distribution. The crack initiation and termination abscissa and the crack termination abscissa are generated in the ranges of (0, 40) and (20, 80), respectively, and the load size is generated in the ranges of (10, 22) according to the actual geometry and load information of the structure under study. For different combinations, combining XFEM parametric modeling technology to obtain crack tip stress intensity factors under different crack geometries and load conditions. In the calculation process, a data set formed by 12120 samples is made to form a database for training a crack damage depth evaluation learning model, wherein the input is the crack endpoint position and the load size, and the output is the crack tip stress intensity factor value. 70% of the data set is randomly selected to be used as a training set, 20% is used as a verification set, and 10% is used as a test set, and training, calibration and verification crack damage depth evaluation learning models are respectively applied.

Step 32: based on the obtained database, a crack damage depth evaluation learning model is constructed, so that rapid intelligent evaluation is realized. The neural network in the deep learning model is a multi-layer perceptron MLP, and the multi-layer perceptron consists of an input layer, a hidden layer and an output layer, wherein the different layers are fully connected. The input layer is a 5-dimensional vector, and the input layer is two-dimensional coordinates and load sizes of two crack endpoints respectively; the hidden layer is composed of two layers 128 of neurons; the dimension of the output layer is 1 dimension, and the size of the stress intensity factor is output. The activation function is selected as the ReLU function. In the damage function optimization process, for better and faster convergence to an optimal value, the initial learning rate is set to be 0.001, and as training is carried out, the learning rate is gradually reduced, the attenuation is 1E-04 in the 150 th period, the attenuation is 1E-05 in the 264 th period, and the attenuation is 1E-06 in the 285 th period. 7 (a) shows the loss function and learning rate during training, the test results are shown in fig. 7 (b), the mean square error MSE over the test set is 1.42E-04, and the predictive power of the damage assessment model is evaluated by calculation, and the calculation formula of the index is as follows:

R ² variance of model, y _i Is the true value of the i-th data,for predictive value +.>R is the average value of ² The value range of (1) is 0 to 1, the closer to 1 the model is, the better the model is, and the calculation results show that the evaluation model has higher prediction precision and stability by calculating to obtain 0.989 on the test set.

In order to realize the online identification of crack damage, the crack damage assessment comprises two stages of online and offline, the offline stage is used for constructing a crack damage assessment deep learning model on the basis of a fracture mechanics model simulation database so as to meet the real-time predictive requirement, and the machine learning model is used for connecting the relation between the possible situation of the crack damage and the damage degree. And in the online stage, the crack geometric model of the previous section is evaluated in real time by using the crack damage evaluation deep learning model, and meanwhile, the fracture mechanics model constructs the crack geometric model through the acquired online data and is used for enriching a database of the crack damage evaluation deep learning model so as to enhance the prediction generalization capability of the crack damage evaluation deep learning model. As shown in fig. 6, a sample is continuously provided for the on-line deep learning model through the analysis of the off-line fracture mechanics model, and meanwhile, the on-line crack deep learning damage assessment model is utilized to complete the real-time prediction of crack damage.

Step 4: based on crack damage analysis of the crack damage module, load is optimized through an optimization algorithm, and a load optimization result is fed back to a crack structure control system, so that the structure is prevented from being broken in the operation process, and meanwhile, the load working condition required by a task is ensured. The load condition is reduced to a safe region. The load optimization can be expressed as follows:

the safety factor in formula f is the minimum load of W when executing the task. Through the optimization, the component can be prevented from being broken under the condition that fatigue cracks appear, and normal operation in the service process is ensured. In the load optimization process, the existence of uncertainty factors is considered, and the safety factor f value is set to be 0.7 so as to ensure the safe service of the crack structural member.

The results of automatic crack identification evaluation and load optimization for 10 real crack structures are shown in fig. 9, wherein the graph (a) is the comparison of the load sizes before and after adjustment, the circle represents the initial applied load size, and the triangle is the adjusted load size. Graph (b) shows a comparison of SIF before and after adjustment, circles represent estimated SIF values, triangles are adjusted SIF values, upper dashed lines represent threshold lines, and lower Fang Xuxian represents set security lines. It can be seen from the figure that the dangerous crack device 2 is significantly lowered, thereby ensuring the current operational safety of the structure. In the process of 10 crack structure load assessment, the complete process of crack intelligent identification, assessment and load optimization is completed by the following steps: 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, 0.528 seconds. By statistics, the average evaluation time of the evaluation time of each structure is 0.545 seconds, and crack structure damage evaluation and load optimization can be rapidly realized.

In summary, by the crack damage intelligent evaluation method based on the visual information and the physical fusion, the intelligent damage evaluation of cracks of the sample at any crack position is realized, the automatic identification of the cracks and the automatic reconstruction of the crack digital model are completed by fusing the perception information and the physical model, and the double-line framework of the real-time evaluation of the crack damage on line, the load optimization and the improvement of the prediction precision and the generalization of the deep learning model by the off-line physical model is realized.

Although the invention has been disclosed in detail with reference to the accompanying drawings, it is to be understood that such description is merely illustrative and is not intended to limit the application of the invention. The present invention is not limited to the above embodiments, and those skilled in the art can implement the present invention in various other embodiments according to the examples and the disclosure of the drawings, so that the design of the present invention is simply changed or modified while adopting the design structure and concept of the present invention, and the present invention falls within the scope of protection.

Claims

1. The crack damage intelligent assessment method based on visual information and physical fusion is characterized by comprising the following steps of:

2. The method for intelligently evaluating crack damage based on visual information and physical fusion according to claim 1, wherein the step 1 of constructing a crack recognition and segmentation deep learning model comprises the following steps:

step 1.1: selecting experimental test pieces with uniform size and materials to prepare different crack test pieces, collecting effective crack pictures, marking cracks in the pictures and square calibration plates, forming a group of data by the crack pictures and the labels, and forming a database of a Mask R-CNN crack recognition and segmentation deep learning model by the collection of all the data, wherein the database is divided into a training set and a testing set;

step 1.3: and when the intersection ratio IoU of the recognition and segmentation results in the training model is larger than a set threshold, the training model correctly recognizes and segments the target, the accuracy rate and the recall rate are calculated to evaluate the training model, and after the evaluation reaches the set standard, the crack recognition and segmentation deep learning model is constructed.

3. The method for intelligently evaluating crack damage based on visual information and physical fusion according to claim 1, wherein the step 2 comprises the following steps:

step 2.1: the calibration information is obtained by calculating crack Mask pixel information obtained by Mask R-CNN crack identification and segmentation deep learning models, and the mode of converting the position information of the crack tip into real crack position information of the surface of the structural part is as follows:

obtaining geometric position information of a crack in the structural member under a Cartesian coordinate system through the method, and obtaining coordinates of two end points of the crack, thereby determining real crack position information of the surface of the structural member;

step 2.2: the secondary development is carried out by adopting Abaqus software, the digital modeling is carried out on the crack-containing structure by using the extended finite element X-FEM, the real end point position information of the crack is interacted to an off-line fracture mechanical model, and a crack structure geometric model comprising the crack position information, the structural geometric information, the material attribute parameters, the boundary conditions and the grid division is constructed by combining a geometric model database.

4. The method for intelligently evaluating crack damage based on visual information and physical fusion according to claim 1, wherein in the step 3, the size and material properties of a crack structure geometric model in offline simulation are kept consistent with those of an experimental test piece, and the steps of constructing a crack damage evaluation deep learning model are as follows:

5. The crack damage intelligent assessment method based on visual information and physical fusion according to claim 4 is characterized in that an offline fracture mechanics model continuously builds a crack geometric model through online interactive data, a sample is provided for an online deep learning damage assessment model, a database is enriched, and meanwhile the online deep learning damage assessment model completes real-time prediction of a previous section of crack geometric model.

6. The intelligent crack damage assessment method based on visual information and physical fusion according to claim 1, wherein in step 4, load optimization is performed on the basis of real-time crack damage assessment, so that the structure is prevented from breaking in the running process, the load working condition required by a task is ensured, and the load optimization is expressed as follows: