Structural fatigue life prediction method, device, equipment and storage medium

Abstract

The present invention discloses a method, apparatus, device, and storage medium for predicting the fatigue life of a structure. The method comprises obtaining real crack data and multiple sets of crack growth parameters under the same number of load cycles; calculating weights based on the real crack data and the virtual crack data in each set of crack growth parameters, thereby determining the retained crack growth parameters; determining whether the virtual crack data in the retained crack growth parameters reaches the critical crack data, and if so, outputting the fatigue life of the structure; otherwise, expanding the retained crack growth parameters using a probabilistic fatigue crack growth model to obtain expanded crack growth parameters, and repeating the steps of calculating weights, determining the retained expanded crack growth parameters, and determining whether the critical crack data has been reached. The present invention achieves an effective association between the real crack data and the probabilistic fatigue crack growth model, thereby improving the prediction accuracy of crack fatigue life.

Description

Technical Field

[0001] The present invention belongs to the technical field of structural fatigue performance evaluation, and in particular relates to a structural fatigue life prediction method, device, equipment and storage medium integrating computer vision and fracture mechanics. Background Art

[0002] Traditional structural crack detection methods include X-rays, ultrasonics, and magnetic flux leakage. The results of these methods are significantly affected by the subjectivity of the inspector. With the rapid development of computer vision technology, computer vision-based structural crack detection methods can extract image features of surface cracks, avoiding the subjectivity and inefficiency of manual inspection and showing promising prospects for automation and intelligentization. However, structural cracks are characterized by their microscopic, multiple, and hidden nature. Furthermore, structural heterogeneity exacerbates the uncertainty of crack initiation and propagation, leading to significant technical bottlenecks in computer vision-based structural crack detection methods. There is an urgent need to develop an accurate and efficient structural fatigue crack detection method based on computer vision.

[0003] The fatigue crack growth model (FCG model) has been widely used for fatigue crack growth analysis and fatigue life prediction in aerospace, automotive, mechanical engineering and other fields. However, the structural fatigue crack growth behavior is highly stochastic, resulting in a significant deviation between the fatigue life prediction results based on the deterministic FCG model and the actual fatigue life. Based on this, the probabilistic fatigue crack growth model (PFCG model) was proposed to characterize the random crack growth behavior. Although the PFCG model considers the uncertainty of fatigue crack growth parameters through prior information, the calculated fatigue life still suffers from a strong discreteness problem, and the evaluation results fail to meet the standards for practical application. Therefore, due to the lack of effective correlation with real crack data, the fatigue life prediction accuracy of the PFCG model still has considerable room for improvement.

[0004] Despite the clear advantages of computer vision in real-time crack data monitoring, the lack of a mechanism to correlate this with fatigue crack growth prevents inspectors from making real-time structural assessments and maintenance decisions based on crack image information. This significantly limits the potential of computer vision for structural fatigue performance assessment. Therefore, integrating measured crack image data into PFCG models is a key research topic in structural fatigue performance assessment. Indeed, the application of computer vision technology in structural fatigue life prediction is extremely limited. Summary of the Invention

[0005] The purpose of the present invention is to provide a structural fatigue life prediction method, device, equipment and storage medium to solve the problem that the PFCG model lacks effective correlation with real crack data, resulting in low accuracy in crack fatigue life prediction.

[0006] The present invention solves the above technical problems through the following technical solutions: a method for predicting structural fatigue life, comprising:

[0007] Acquire real crack data and multiple sets of crack growth parameters under the same number of load cycles; wherein the virtual crack data in each set of crack growth parameters is obtained by expanding using a probabilistic fatigue crack growth model;

[0008] The weight of the corresponding virtual crack data is calculated based on the real crack data under the same number of load cycles and the virtual crack data in each set of crack growth parameters;

[0009] The retained crack growth parameters are determined according to the weights of all virtual crack data;

[0010] Determine whether the virtual crack data in the retained crack growth parameters reaches the critical crack data. If so, use the corresponding number of load cycles as the fatigue life of the structure. If not, use the probabilistic fatigue crack growth model to expand the retained crack growth parameters to obtain expanded virtual crack data and expanded crack growth parameters, and repeat the steps of calculating the weight of the expanded virtual crack data, determining the retained expanded crack growth parameters, and determining whether the critical crack data is reached.

[0011] Furthermore, the process of acquiring the real crack data includes:

[0012] Obtain a real crack image, use the crack segmentation model to identify the real crack image, and obtain the real crack;

[0013] A skeletonization algorithm is used to simplify the real crack into a crack line with a single pixel width, and the crack line in the pixel coordinate system is converted to a Cartesian coordinate system;

[0014] Multiple straight line segments are used to fit the crack line in the Cartesian coordinate system, and the pixel size of each straight line segment is calculated;

[0015] According to the scale factor of the real crack image, the pixel size of each straight line segment is converted to the physical size to obtain the real crack data.

[0016] The present invention detects fatigue cracks based on computer vision, provides real crack data in real time for fatigue life prediction, realizes effective correlation between fatigue life prediction and real crack data, and improves fatigue life prediction accuracy.

[0017] Furthermore, the crack segmentation model is obtained by training an improved DeepLabV3+ model using a crack sample dataset; the improved DeepLabV3+ model replaces the Xception convolution layer in the original DeepLabV3+ model with MobileNetV3, and replaces the ASPP module in the original DeepLabV3+ model with an improved ASPP module; the improved ASPP module retains the first branch and the fifth branch of the ASPP module, and redesigns the middle second to fourth branches into an initialization window size layer, a first batch normalization layer, a first activation layer, a depth convolution layer, a second batch normalization layer and a second activation layer connected in sequence.

[0018] The lightweight MobileNetV3 is used as the backbone network to balance recognition accuracy and computational efficiency, improving the adaptability of the model to multi-scale targets; the ASPP module in the encoder is redesigned and optimized to further reduce the complexity of the model.

[0019] Furthermore, the specific steps of constructing the probabilistic fatigue crack growth model include:

[0020] Obtain stress intensity factors at different crack sizes;

[0021] The stress intensity factors under different crack sizes are used to train the multilayer perceptron to obtain a stress intensity factor prediction model;

[0022] A probabilistic fatigue crack growth model is constructed based on the stress intensity factor prediction model and the Paris formula.

[0023] Furthermore, the calculation formula for the weight of each virtual crack data is:

[0024] ;

[0025] ;

[0026] ;

[0027] in, represents the standard deviation of the difference between the real crack data and each virtual crack data, represents the standard deviation function, represents the real crack data at the kth moment, represents the i-th virtual crack data at the k-th moment, represents the likelihood function, represents the weight of the i-th virtual crack data at the k-th moment, and n represents the number of virtual crack data.

[0028] Furthermore, the determination of the retained crack propagation parameters according to the weights of all virtual crack data includes:

[0029] Calculate n cumulative distribution function values ​​according to the weights of all virtual crack data; the specific calculation formula for the mth cumulative distribution function value is:

[0030] ;

[0031] in, represents the cumulative distribution function value of the first m virtual crack data, that is, the mth cumulative distribution function value; represents the weight of the i-th virtual crack data at the k-th moment; n represents the number of virtual crack data;

[0032] Generate n uniformly distributed random numbers in the interval [0,1];

[0033] When the jth random number is between the mth cumulative distribution function value and the m+1th cumulative distribution function value, the m+1th virtual crack data is retained, and thus the m+1th group of crack growth parameters is retained.

[0034] Based on the same concept, the present invention also provides a structural fatigue life prediction device, comprising:

[0035] An acquisition unit, configured to acquire real crack data and multiple sets of crack growth parameters under the same number of load cycles; wherein the virtual crack data in each set of crack growth parameters is obtained by expansion using a probabilistic fatigue crack growth model;

[0036] A calculation unit, used for calculating the weight of corresponding virtual crack data according to the real crack data under the same number of load cycles and the virtual crack data in each set of crack growth parameters;

[0037] a determination unit for determining a retained crack growth parameter according to weights of all virtual crack data;

[0038] A judgment unit is used to judge whether the virtual crack data in the retained crack growth parameters reaches the critical crack data. If so, the corresponding number of load cycles is used as the fatigue life of the structure; if not, the acquisition unit, calculation unit, determination unit and judgment unit are called in sequence;

[0039] An acquisition unit is used to obtain real crack data under the same number of load cycles and multiple groups of expanded crack growth parameters; wherein, the virtual crack data in each group of expanded crack growth parameters are obtained by expanding the virtual crack data in the retained crack growth parameters using a probabilistic fatigue crack growth model; a calculation unit is used to calculate the weight of the corresponding expanded virtual crack data based on the real crack data under the same number of load cycles and the virtual crack data in each group of expanded crack growth parameters; a determination unit is used to determine the retained expanded crack growth parameters based on the weights of all expanded virtual crack data; a judgment unit is used to judge whether the virtual crack data in the retained expanded crack growth parameters reaches the critical crack data, and if so, the corresponding number of load cycles is used as the fatigue life of the structure; if not, the acquisition unit, the calculation unit, the determination unit and the judgment unit are continued to be called in sequence.

[0040] Based on the same concept, the present invention also provides an electronic device, including a memory, a processor, and a computer program / instruction stored in the memory, wherein the processor executes the computer program / instruction to implement the structural fatigue life prediction method as described above.

[0041] Based on the same concept, the present invention also provides a computer-readable storage medium having a computer program / instruction stored thereon. When the computer program / instruction is executed by a processor, the structural fatigue life prediction method as described above is implemented.

[0042] Compared with the prior art, the advantages of the present invention are:

[0043] The present invention realizes the effective association between real crack data and probabilistic fatigue crack growth model by acquiring real crack data and virtual crack data under the same number of load cycles, thereby realizing the dynamic update of uncertain parameters of the probabilistic fatigue crack growth model and improving the prediction accuracy of crack fatigue life. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the technical solution of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only one embodiment of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0045] Figure 1 This is a flow chart of a method for predicting structural fatigue life in an embodiment of the present invention;

[0046] Figure 2 is a schematic diagram of some samples in the crack sample data set according to an embodiment of the present invention;

[0047] Figure 3is a schematic diagram of a crack image with a calibration plate in an embodiment of the present invention;

[0048] Figure 4 Schematic diagram of the structure of the original DeepLabV3+ model in an embodiment of the present invention;

[0049] Figure 5 2 is a schematic diagram of the structure of MobileNetV3 in an embodiment of the present invention;

[0050] Figure 6 2 is a schematic structural diagram of an improved ASPP module in an embodiment of the present invention;

[0051] Figure 7 is the comparison result between the crack length detected in the embodiment of the present invention and the actual crack length;

[0052] Figure 8 This is a flowchart of the interaction technology between ABAQUS and FRANC3D in an embodiment of the present invention; the data in the figure represent dimensions, unit: mm;

[0053] Figure 9 is a stress intensity factor prediction model for the mid-axis point of the crack front in an embodiment of the present invention;

[0054] Figure 10 is a stress intensity factor prediction model for the crack front length point in an embodiment of the present invention;

[0055] Figure 11 It is a probabilistic fatigue crack growth model constructed by combining the stress intensity factor prediction model and the Paris formula in an embodiment of the present invention;

[0056] Figure 12 This is the crack length update process when R=0.1 in the embodiment of the present invention;

[0057] Figure 13 This is the crack length update process when R=0.2 in the embodiment of the present invention;

[0058] Figure 14 This is the crack length update process when R=0.3 in the embodiment of the present invention;

[0059] Figure 15 This is the fatigue life prediction result when R=0.1 in the embodiment of the present invention;

[0060] Figure 16 This is the fatigue life prediction result when R=0.2 in the embodiment of the present invention;

[0061] Figure 17 This is the fatigue life prediction result when R=0.3 in the embodiment of the present invention. DETAILED DESCRIPTION

[0062] The following is a clear and complete description of the technical solutions of the present invention in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the embodiments described 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 are within the scope of protection of the present invention.

[0063] The following specific embodiments are used to describe the technical solution of the present invention in detail. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0064] Example 1

[0065] like Figure 1 As shown, the structural fatigue life prediction method provided by the present invention includes the following steps:

[0066] Step S1: Acquire real crack data and multiple sets of crack growth parameters under the same number of load cycles.

[0067] In order to effectively associate real crack data with the fatigue crack growth process, the present invention is based on computer vision and obtains real crack data based on real crack images. In a specific embodiment of the present invention, the process of obtaining real crack data includes:

[0068] Step A1: Obtain a real crack image, use the crack segmentation model to identify the real crack image, and obtain the real crack.

[0069] The image acquisition system captures the cracks on the structure to obtain a true crack image. The number of load cycles corresponding to the true crack image can be obtained by using a fatigue testing machine or other load cycle recording method. The structure of this embodiment is a steel structure. The crack segmentation model is obtained by training the improved DeepLabV3+ model using a crack sample dataset. The specific training process includes:

[0070] Step A1.1: Construct a crack sample dataset.

[0071] The crack sample dataset includes multiple samples, each of which includes an input and an output. The input is a scene image containing cracks (i.e., a crack image), and the output is a segmentation mask used to identify whether each pixel in the input image belongs to the crack category, such as Figure 2 All images were uniformly processed to a fixed resolution of 512×512 and divided into training and validation sets in a ratio of 9:1. The training set includes 3919 crack images and the validation set includes 436 crack images.

[0072] In order to avoid the lack of reference measurement objects for the sample, which makes it impossible to convert the pixel size of the crack into the physical size (i.e., the real-world size), the present invention conducts fatigue crack growth analysis on compact tensile specimens (i.e., CT specimens) and collects a large number of images with calibration plates, such as Figure 3 shown.

[0073] Specifically, the CT specimens were made of Q345qd steel with an elastic modulus of 2.10 × 105 MPa and a Poisson's ratio of 0.3. The experimental equipment primarily consisted of a fatigue testing machine and an image acquisition system. The CT specimens were connected to the clamping end of the brake via a prefabricated fixture and bolt pins. A 7Hz sinusoidal fatigue cyclic load was applied using a force-controlled method. The stress ratio (R) was controlled by fixing the upper fatigue load limit and adjusting the lower fatigue load limit, with a maximum load of 5kN. The image acquisition system primarily consisted of a CCD camera, an LED light source, and interactive software. To ensure clear crack image capture, the camera was positioned 0.5m from the specimen surface, with a resolution of 4096 × 3000 pixels and a frame rate of 40fps. To accelerate fatigue crack formation, a 2mm initial defect was prefabricated on each CT specimen using wire cutting technology.

[0074] Step A1.2: Build the improved DeepLabV3+ model.

[0075] Although the original DeepLabV3+ model can be directly used for structural crack detection, it is difficult for the original DeepLabV3+ model to meet the needs of lightweight detection. Therefore, the present invention improves the original DeepLabV3+ model. The structure of the original DeepLabV3+ model is as follows Figure 4 As shown in Figure 1, on the one hand, the lightweight MobileNetV3 is used as the backbone network to replace the original Xception convolution layer to balance recognition accuracy and computational efficiency, thereby improving the adaptability of the model under multi-scale targets. The structure of MobileNetV3 is as follows Figure 5 shown.

[0076] On the other hand, the ASPP module (atrous spatial pyramid pooling module) of the original DeepLabV3+ model is redesigned, such as Figure 6 Specifically, the first and fifth branches of the ASPP module are retained, while the intermediate branches (second through fourth) are redesigned to sequentially connect an initialization window size layer, a first batch normalization layer, a first activation layer, a depthwise convolutional layer, a second batch normalization layer, and a second activation layer. The redesigned ASPP module focuses only on the local correlation of the current window and its neighboring windows, ignoring the global relationship of the entire image. This effectively reduces the global-to-local computational overhead and enhances the model's scalability.

[0077] Step A1.3: Use the crack sample dataset constructed in step A1.1 to train and verify the improved DeepLabV3+ model to obtain the crack segmentation model.

[0078] The improved DeepLabV3+ model is trained using the training set and verified using the validation set. During the training process, the hyperparameters are set as follows: Adam is selected as the optimizer and the initial learning rate is 5e -4 , to accelerate convergence and maintain training stability. The momentum factor was set to 0.9 to further enhance convergence and reduce oscillations during training. The Adam optimizer has an adaptive learning rate feature that dynamically adjusts the learning rate based on the training process. To prevent overfitting, the weight decay coefficient was set to 0, the batch size was 8, and the number of training epochs was 100. A TXT file was generated to record and observe the changes in model loss on the training and validation sets. The cross-entropy loss was used as the loss function, which is widely used in crack segmentation tasks to optimize crack identification accuracy.

[0079] Step A1.4: Evaluate the performance of the crack segmentation model.

[0080] This embodiment uses indicators such as intersection over union (IoU), recall, and precision to evaluate the crack recognition accuracy of the original DeepLabV3+ model and the improved DeepLabV3+ model. IoU represents the ratio of the intersection and union of the crack prediction value and the true value, which intuitively reflects the crack segmentation effect. Recall measures the proportion of the number of samples correctly identified as positive to all positive samples, and is used to evaluate the model's recognition coverage ability for positive samples. Precision is the proportion of samples predicted to be positive that are actually positive, reflecting the accuracy of the model's prediction results. In addition, the number of parameters, FLOPs, and training time are key indicators for evaluating the lightweightness of the model, which can measure the computational complexity and operational efficiency of the model.

[0081] This example is based on the PyTorch deep learning framework. The crack sample dataset constructed in step A1.1 is used to train and verify the original DeepLabV3+ model and the improved DeepLabV3+ model in a GPU environment. The performance of the verified original DeepLabV3+ model and the improved DeepLabV3+ model is then quantitatively evaluated using evaluation indicators. The evaluation results are shown in Table 1, where mIoU, mRecall, and mPrecision represent the average values ​​of intersection over union, recall, and precision, respectively. As shown in Table 1, the improved DeepLabV3+ model has significantly improved in lightweight crack recognition compared to the DeepLabV3+ model, with the number of parameters, FLOPs, and training time reduced by 93.18%, 73.17%, and 35.55%, respectively. In addition, the improved DeepLabV3+ model can maintain a high crack recognition accuracy while significantly reducing the computing resource requirements, among which mIoU, mRecall and mPrecision reach 88.57%, 94.00% and 93.15% respectively.

[0082] Table 1 Structural crack identification results of different models

[0083]

[0084] Step A2: Use a skeletonization algorithm to simplify the real crack obtained in step A1 into a crack line with a single pixel width, and convert the crack line in the pixel coordinate system to the Cartesian coordinate system.

[0085] In order to calculate the real crack data (such as the real crack length and the real crack angle), this embodiment uses a skeletonization algorithm to refine the real cracks identified by the crack segmentation model into crack lines with a single pixel width, and removes small noise areas through connected domain analysis.

[0086] Step A3: Use multiple straight line segments to fit the crack line in the Cartesian coordinate system, and calculate the pixel size of each straight line segment.

[0087] Because single-pixel crack lines are extremely irregular in Cartesian coordinates, multiple straight line segments are used to fit the crack lines. During the fitting process, the error between the skeleton points and the straight line segments is minimized using the Nelder-Mead optimization algorithm, gradient descent, or least squares method to ensure fitting accuracy. After fitting, the pixel size information of each straight line segment is calculated.

[0088] The calculation formula for the pixel length of each straight line segment is:

[0089] (1)

[0090] The calculation formula for the pixel angle of each straight line segment is:

[0091] (2)

[0092] Where L represents the pixel length of the current straight line segment, θ represents the pixel angle between the current straight line segment and the previous straight line segment, Indicates the starting point of the previous straight line segment. Indicates the end point of the previous straight line segment and the starting point of the current straight line segment. Indicates the end point of the current line segment.

[0093] Step A4: According to the scale factor of the real crack image, the pixel size of each straight line segment is converted into a physical size to obtain the real crack data.

[0094] The calibration plate corner detection method is used to calculate the scale factor of the real crack image. The pixel size is converted to the physical size through the scale factor, thus realizing the conversion of the crack pixel size to the real world size. The specific calculation formula of the scale factor is:

[0095] (3)

[0096] Where A represents the scale factor of the real crack image, Indicates the physical length of the calibration plate, Therefore, the pixel length of the straight line segment can be converted into a physical length by the scaling factor. The real crack data of this embodiment includes the real crack physical length, that is, the real crack length.

[0097] In order to verify the effectiveness of crack data acquisition, a ruler is used for calibration when collecting crack images in the crack sample dataset, such as Figure 7 As shown. Figure 7 The detected crack length is highly consistent with the actual crack length, with the overall recognition error within ±0.5 mm and an average error of 0.31 mm. This demonstrates the reliability and practicality of the crack data acquisition method in engineering applications, and can provide real-time, accurate true crack length information for the digital twin (DT) driving framework.

[0098] The present invention utilizes the improved DeepLabV3+ model to identify structural cracks and calculates crack data based on the identified cracks, which can provide real-time and accurate structural real crack data for the digital twin driving framework.

[0099] Using the interactive technology of ABAQUS and FRANC3D for fatigue crack growth (FCG) analysis can generate virtual crack data. However, a single FCG analysis cycle takes over 70 minutes, resulting in an extremely low virtual crack data generation rate, making it difficult to meet the real-time update requirements of uncertain FCG parameters (i.e., crack growth parameters). To address this technical issue, the present invention constructs a probabilistic fatigue crack growth model (PFCG model). Using the PFCG model, the crack growth model is applied to the input FCG parameters to generate virtual crack data.

[0100] In a specific embodiment of the present invention, the specific steps of constructing the probabilistic fatigue crack growth model include:

[0101] Step B1: Obtain stress intensity factors at different crack sizes.

[0102] Based on the CT specimen in step A1.1, by adjusting the inserted crack size, the corresponding crack front stress intensity factor SIF is calculated based on the M-integral method using the interactive technology of ABAQUS and FRANC3D, and the stress intensity factor under different crack sizes can be obtained, such as Figure 8 The crack size of this embodiment includes a and h, where a represents the length of the crack center axis and h represents the crack length. Figure 8 The c in the figure represents the length of the semi-major axis of the crack.

[0103] A stress intensity factor dataset is constructed based on the stress intensity factors for different crack sizes. Each sample in the stress intensity factor dataset includes an input and an output: the inputs are the crack sizes a and h, and the output is the corresponding stress intensity factor (SIF). The stress intensity factor dataset in this example includes 3,500 samples and is used to train a multilayer perceptron (MLP).

[0104] Step B2: Use the stress intensity factors under different crack sizes to train the MLP to obtain a stress intensity factor prediction model.

[0105] This embodiment uses machine learning methods to establish a high-precision stress intensity factor prediction model, namely the SIF prediction model, to capture the nonlinear relationship between crack size a and h and SIF. The stress intensity factor prediction model is tested using 1500 input quantities, and the test results are as follows: Figure 9 and Figure 10 As shown, represents the stress intensity factor at the mid-axis point of the crack front, Represents the stress intensity factor at the crack front length point. Figure 9 and Figure 10 It can be seen that the stress intensity factor prediction model can quickly and accurately output the corresponding SIF based on the given crack size.

[0106] Step B3: Construct a probabilistic fatigue crack growth model based on the stress intensity factor prediction model and the Paris formula.

[0107] Based on the prior information, a probability distribution model of the initial crack growth parameters (a, h, C, m) of the probabilistic fatigue crack growth model is established. Subsequently, the probability distribution model is sampled to generate each set of initial crack growth parameters. In this embodiment, the initial crack growth parameters include crack sizes a and h, and material-related parameters C and m. It is assumed that the material-related parameter C follows a normal distribution (3.3×10 -13 , 1.6×10 -13 ), the material-related parameter m follows a uniform distribution (2.80, 3.10), and the material-related parameters C and m are obtained based on prior information. By simulating the crack growth process of each sample using the stress intensity factor prediction model and the Paris formula, virtual crack data under any number of load cycles can be output, namely, crack sizes a and h, such as Figure 11 As shown, represents the stress intensity factor range threshold, 、 and They represent the crack median length, crack length and number of load cycles at the current moment respectively.

[0108] In the PFCG model, the Paris formula is used to describe the relationship between the stress intensity factor range and the crack growth rate. The specific formula is:

[0109] , (4)

[0110] Where N is the number of load cycles, ΔK is the stress intensity factor range (obtained directly from the SIF prediction model), and They represent the stress intensity factor ranges of the crack mid-axis point and crack length point as the number of load cycles changes.

[0111] The crack growth increment under a given number of load cycles can be expressed according to the integration method as:

[0112] (5)

[0113] (6)

[0114] in, represents the crack extension increment in the direction of the crack centerline, represents the crack length extension increment, Indicates the increment of load cycles. The machine learning-assisted PFCG model uses the crack length at the current moment Reaching critical crack length As the termination condition for fatigue crack growth analysis. It is worth noting that on a workstation configured with 20 cores (Intel i7-14700KF), the time for a single fatigue crack growth analysis cycle was shortened to less than 1 second. Compared with the 70 minutes required for pure ABAQUS and FRANC3D interactive technology, the computational efficiency has been significantly improved, and the generation of crack data has been accelerated through the PFCG model. The crack data generated by the PFCG model is defined as virtual crack data. For example, the median length of the crack generated by the PFCG model is defined as the virtual median length of the crack, both represented by the character a; the crack length generated by the PFCG model is defined as the virtual crack length, both represented by the character h.

[0115] Since the real crack median length cannot be obtained through steps A1 to A4, the real crack data and virtual crack data in this embodiment both refer to the crack length, that is, the real crack length and the virtual crack length.

[0116] Step S2: Calculate the weight of the corresponding virtual crack data based on the real crack data under the same number of load cycles and the virtual crack data in each set of crack growth parameters.

[0117] The present invention adopts a dynamic Bayesian network (DBN) to establish a synchronous connection between real crack data and crack propagation parameters under the same number of load cycles, and adopts a particle filtering algorithm as an inference method of the dynamic Bayesian network (DBN).

[0118] The probability distribution model of the initial crack growth parameters of the PFCG model is sampled to generate 100 sets of initial crack growth parameters (a, h, C, m). The particle swarm then consists of 100 particles, each corresponding to a set of initial crack growth parameters (a, h, C, m), and each particle has the same weight. The state equations for fatigue crack growth analysis are established based on the PFCG model (Equations (7) and (8)):

[0119] (7)

[0120] (8)

[0121] in, represents the crack median length at the current moment or the kth moment (i.e., the virtual crack median length), represents the crack median length at the previous moment or the k-1th moment, represents the state noise, represents the crack length at the current moment or the kth moment (i.e., virtual crack length), represents the crack length at the previous moment or the k-1th moment, represents the actual crack length at the current moment (obtained according to steps A1 to A4), Represents the weight of the crack length.

[0122] The true crack length The calculated value of the state equation (i.e. the output value of the PFCG model, i.e. the virtual crack length ) to compare, calculate the error between each virtual crack length and the real crack length, and calculate the likelihood function value through Gaussian distribution to update the weight of each particle at the current moment:

[0123] (9)

[0124] (10)

[0125] (11)

[0126] in, Indicates the true crack length With each virtual crack length The standard deviation of the difference, represents the standard deviation function, represents the likelihood function, represents the i-th virtual crack length or the weight of the i-th particle at the k-th moment, n represents the virtual crack length or the number of particles, and n=100 in this embodiment.

[0127] Specifically, based on the fatigue crack growth analysis experiment in step A1.1, the image acquisition system collects crack images according to 5000 load cycles, and the acquisition cycle is appropriately shortened only when the crack approaches the critical length. Considering that when the actual crack length reaches about 14 mm during the experiment, the CT specimen will break rapidly in a very short time. Therefore, 14 mm is determined as the critical crack length. Subsequently, the actual crack length is extracted from the crack image. Correspondingly, the PFCG model also calculates the virtual crack length according to 5000 load cycles. In this embodiment, three groups of fatigue crack growth analysis experiments were carried out under stress ratios R = 0.1, 0.2 and 0.3, respectively. The actual crack length and the virtual crack length output by the PFCG model were compared at intervals of 20000, 30000 and 40000 load cycles, respectively. Then, the matching degree between the virtual crack length and the actual crack length was calculated according to formulas (9) to (11), thereby determining the weight of each particle or each virtual crack length.

[0128] Step S3: Determine the retained crack propagation parameters according to the weights of all virtual crack data.

[0129] In a specific embodiment of the present invention, the virtual crack data refers to the virtual crack length, and the retained crack growth parameters are determined according to the weights of all virtual crack lengths (i.e., particles), including:

[0130] Step S3.1: Calculate n cumulative distribution function values ​​based on the weights of all virtual crack lengths; the specific calculation formula for the mth cumulative distribution function value is:

[0131] (12)

[0132] in, represents the cumulative distribution function value of the first m virtual crack lengths, that is, the mth cumulative distribution function value; represents the i-th virtual crack length or the weight of the i-th particle at the k-th moment; n represents the virtual crack length or the number of particles. When m is 1, the first cumulative distribution function value is obtained; when m is 2, the second cumulative distribution function value is obtained; and so on, when m is n, the n-th cumulative distribution function value is obtained.

[0133] Step S3.2: Generate n uniformly distributed random numbers in the interval [0,1].

[0134] Step S3.3: When the jth random number is between the mth cumulative distribution function value and the m+1th cumulative distribution function value, retain the m+1th virtual crack length, and further retain the m+1th set of crack growth parameters.

[0135] The retained crack growth parameters are determined based on the cumulative distribution function value and the random number. The number of retained crack growth parameters remains n. The larger the weight of the virtual crack length or particle, the greater the probability of being selected; conversely, the smaller the weight of the virtual crack length or particle, the smaller the probability of being selected, or even not being selected at all.

[0136] In this embodiment, the updating process of material related parameters C and m in the three groups of fatigue crack growth analysis experiments is shown in Table 2, and the updating process of crack length h is shown in Table 2. Figures 12 to 14 The results show that by continuously integrating the measured crack length (i.e., the true crack length), the crack growth parameters are dynamically updated, while the uncertainty of the crack growth parameters is reduced.

[0137] Table 2 Update process of material related parameters C and m

[0138]

[0139] Step S4: Determine whether the virtual crack data in the retained crack growth parameters reaches the critical crack data. If so, the corresponding number of load cycles is used as the fatigue life of the structure. If not, the probabilistic fatigue crack growth model is used to perform crack growth on the retained crack growth parameters to obtain the expanded virtual crack data and the expanded crack growth parameters, and repeat steps S2 to S4.

[0140] The average value of all retained virtual crack lengths (i.e., the virtual crack lengths in the retained crack extension parameters) is taken as the current crack length h k , determine the current crack length h k Whether the critical crack length h is reached f If yes, the corresponding number of load cycles N is used as the fatigue life of the structure; if not, the retained crack growth parameters are input into the PFCG model, and the FCG analysis at time k+1 is performed to obtain the virtual crack length after expansion and the crack growth parameters after expansion. Then, steps S2 to S4 are repeated until the critical crack length is reached, and the fatigue life of the structure is output.

[0141] In this embodiment, three groups of fatigue crack growth analysis experiments were respectively subjected to three DBN inferences (i.e., each group of experiments compared the actual crack length (i.e., the experimental value) with the virtual crack length output by the PFCG model three times). The predicted structural fatigue life is as follows: Figures 15 to 17 The results show that DBN inference establishes a synchronous connection between real crack data and the PFCG model. By continuously integrating real crack data, the uncertainty of the crack growth parameters in the PFCG model is effectively reduced, significantly improving the fatigue life prediction accuracy, with an average error of 6.94%.

[0142] This paper integrates computer vision and fracture mechanics theory to propose a digital twin-driven framework for real-time prediction of structural fatigue life. Based on an improved DeepLabV3+ model, this framework implements real-time structural crack detection, providing real-time, accurate real-world crack data for the digital twin-driven framework. A machine learning-based PFCG model is established to accelerate the generation of virtual crack data. Furthermore, a dynamic Bayesian network (DBN) is used to synchronize virtual and real crack data in digital space, allowing for real-time updating of the uncertain parameters of the PFCG model. This improves the fatigue life prediction accuracy of steel structures, reducing the average fatigue life prediction error from 52.33% to 6.94%. Experimental studies validate the effectiveness of the proposed digital twin-driven framework.

[0143] Example 2

[0144] The structural fatigue life prediction device provided by the present invention comprises an acquisition unit, a calculation unit, a determination unit and a judgment unit.

[0145] An acquisition unit is configured to acquire real crack data and multiple sets of crack growth parameters under the same number of load cycles; wherein the virtual crack data in each set of crack growth parameters is obtained by expansion using a probabilistic fatigue crack growth model. In this embodiment, the real crack data and virtual crack data refer to the real crack length and virtual crack length, respectively.

[0146] The calculation unit is used to calculate the weight of the corresponding virtual crack data based on the real crack data under the same number of load cycles and the virtual crack data in each set of crack growth parameters, see formula (9) to formula (11) in embodiment 1.

[0147] A determination unit is used to determine the retained crack growth parameters according to the weights of all virtual crack data.

[0148] The judging unit is used to judge whether the virtual crack data in the retained crack growth parameters has reached the critical crack data. If so, the corresponding number of load cycles is used as the fatigue life of the structure. If not, the acquiring unit, calculating unit, determining unit and judging unit are called in sequence. The acquiring unit is used to obtain the real crack data under the same number of load cycles and multiple sets of expanded crack growth parameters. The virtual crack data in each set of expanded crack growth parameters is obtained by expanding the virtual crack data in the retained crack growth parameters using a probabilistic fatigue crack growth model. The calculating unit is used to calculate the weight of the corresponding expanded virtual crack data based on the real crack data under the same number of load cycles and the virtual crack data in each set of expanded crack growth parameters. The determining unit is used to determine the retained expanded crack growth parameters based on the weights of all expanded virtual crack data. The judging unit is used to judge whether the virtual crack data in the retained expanded crack growth parameters has reached the critical crack data. If so, the corresponding number of load cycles is used as the fatigue life of the structure. If not, the acquiring unit, calculating unit, determining unit and judging unit are continued to be called in sequence.

[0149] In some specific embodiments of the present invention, the structural fatigue life prediction device may be combined with the features of the structural fatigue life prediction method in the first embodiment of the present invention, and vice versa.

[0150] Example 3

[0151] An embodiment of the present invention further provides an electronic device comprising: a memory, a processor, and a computer program / instructions stored in the memory, wherein the processor executes the computer program / instructions to implement the structural fatigue life prediction method in the first embodiment of the present invention.

[0152] Although not shown, the electronic device includes a processor that can perform various appropriate operations and processes based on programs and / or data stored in a read-only memory (ROM) or programs and / or data loaded from a storage portion into a random access memory (RAM). The processor can be a multi-core processor or can include multiple processors. In some embodiments, the processor can include a general-purpose main processor and one or more special coprocessors, such as a central processing unit, a graphics processing unit (GPU), a neural network processor (NPU), a digital signal processor (DSP), etc. Various programs and data required for device operation are also stored in the RAM. The processor, ROM, and RAM are connected to each other via a bus. An input / output (I / O) interface is also connected to the bus.

[0153] The processor and memory are used together to execute the program / instructions stored in the memory. When the program / instructions are executed by the computer, the methods, steps or functions described in the above embodiments can be implemented.

[0154] Although not shown, an embodiment of the present invention further provides a computer-readable storage medium having a computer program / instruction stored thereon. When the computer program / instruction is executed by a processor, the structural fatigue life prediction method in the first embodiment of the present invention is implemented.

[0155] Computer-readable storage media include permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. The information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassettes, magnetic disk storage or other magnetic storage devices, or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer-readable media does not include transitory media such as modulated data signals and carrier waves.

[0156] The above disclosure is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with this technical field can easily think of changes or modifications within the technical scope disclosed in the present invention, and they should all be covered by the scope of protection of the present invention.

Claims

1. A method for predicting structural fatigue life, characterized in that: The prediction method comprises: Acquire real crack data and multiple sets of crack growth parameters under the same number of load cycles; wherein the virtual crack data in each set of crack growth parameters is obtained by expanding using a probabilistic fatigue crack growth model; The weight of the corresponding virtual crack data is calculated based on the real crack data under the same number of load cycles and the virtual crack data in each set of crack growth parameters; The retained crack growth parameters are determined according to the weights of all virtual crack data; Determine whether the virtual crack data in the retained crack growth parameters reaches the critical crack data. If so, use the corresponding number of load cycles as the fatigue life of the structure. If not, use the probabilistic fatigue crack growth model to expand the retained crack growth parameters to obtain expanded virtual crack data and expanded crack growth parameters, and repeat the steps of calculating the weight of the expanded virtual crack data, determining the retained expanded crack growth parameters, and determining whether the critical crack data is reached.

2. The structural fatigue life prediction method according to claim 1, characterized in that: The process of obtaining the real crack data includes: Obtain a real crack image, use the crack segmentation model to identify the real crack image, and obtain the real crack; A skeletonization algorithm is used to simplify the real crack into a crack line with a single pixel width, and the crack line in the pixel coordinate system is converted to a Cartesian coordinate system; Multiple straight line segments are used to fit the crack line in the Cartesian coordinate system, and the pixel size of each straight line segment is calculated; According to the scale factor of the real crack image, the pixel size of each straight line segment is converted to the physical size to obtain the real crack data.

3. The structural fatigue life prediction method according to claim 2, characterized in that: The crack segmentation model is obtained by training an improved DeepLabV3+ model using a crack sample dataset; the improved DeepLabV3+ model replaces the Xception convolution layer in the original DeepLabV3+ model with MobileNetV3, and replaces the ASPP module in the original DeepLabV3+ model with an improved ASPP module; the improved ASPP module retains the first branch and the fifth branch of the ASPP module, and redesigns the middle second to fourth branches into an initialization window size layer, a first batch normalization layer, a first activation layer, a depth convolution layer, a second batch normalization layer and a second activation layer connected in sequence.

4. The structural fatigue life prediction method according to claim 1, characterized in that: The specific steps of constructing the probabilistic fatigue crack growth model include: Obtain stress intensity factors at different crack sizes; The stress intensity factors under different crack sizes are used to train the multilayer perceptron to obtain a stress intensity factor prediction model; A probabilistic fatigue crack growth model is constructed based on the stress intensity factor prediction model and the Paris formula.

5. The structural fatigue life prediction method according to claim 4, characterized in that: The interactive technology between ABAQUS and FRANC3D is used to obtain the stress intensity factors under different crack sizes.

6. The structural fatigue life prediction method according to claim 1, characterized in that: The calculation formula of the weight of each virtual crack data is: ； ； ； in, represents the standard deviation of the difference between the real crack data and each virtual crack data, represents the standard deviation function, represents the real crack data at the kth moment, represents the i-th virtual crack data at the k-th moment, represents the likelihood function, represents the weight of the i-th virtual crack data at the k-th moment, and n represents the number of virtual crack data.

7. The structural fatigue life prediction method according to any one of claims 1 to 6, characterized in that: The method of determining the retained crack propagation parameters according to the weights of all virtual crack data includes: Calculate n cumulative distribution function values ​​according to the weights of all virtual crack data; the specific calculation formula for the mth cumulative distribution function value is: ； in, represents the cumulative distribution function value of the first m virtual crack data, that is, the mth cumulative distribution function value; represents the weight of the i-th virtual crack data at the k-th moment; n represents the number of virtual crack data; Generate n uniformly distributed random numbers in the interval [0,1]; When the jth random number is between the mth cumulative distribution function value and the m+1th cumulative distribution function value, the m+1th virtual crack data is retained, and thus the m+1th group of crack growth parameters is retained.

8. A structural fatigue life prediction device, characterized in that: The prediction device comprises: An acquisition unit, configured to acquire real crack data and multiple sets of crack growth parameters under the same number of load cycles; wherein the virtual crack data in each set of crack growth parameters is obtained by expansion using a probabilistic fatigue crack growth model; A calculation unit, used for calculating the weight of corresponding virtual crack data according to the real crack data under the same number of load cycles and the virtual crack data in each set of crack growth parameters; a determination unit for determining a retained crack growth parameter according to weights of all virtual crack data; A judgment unit is used to judge whether the virtual crack data in the retained crack growth parameters reaches the critical crack data. If so, the corresponding number of load cycles is used as the fatigue life of the structure; if not, the acquisition unit, calculation unit, determination unit and judgment unit are called in sequence; An acquisition unit is used to obtain real crack data under the same number of load cycles and multiple groups of expanded crack growth parameters; wherein, the virtual crack data in each group of expanded crack growth parameters are obtained by expanding the virtual crack data in the retained crack growth parameters using a probabilistic fatigue crack growth model; a calculation unit is used to calculate the weight of the corresponding expanded virtual crack data based on the real crack data under the same number of load cycles and the virtual crack data in each group of expanded crack growth parameters; a determination unit is used to determine the retained expanded crack growth parameters based on the weights of all expanded virtual crack data; a judgment unit is used to judge whether the virtual crack data in the retained expanded crack growth parameters reaches the critical crack data, and if so, the corresponding number of load cycles is used as the fatigue life of the structure; if not, the acquisition unit, the calculation unit, the determination unit and the judgment unit are continued to be called in sequence.

9. An electronic device comprising a memory, a processor, and a computer program / instruction stored in the memory, characterized in that: The processor executes the computer program / instructions to implement the structural fatigue life prediction method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program / instruction stored thereon, characterized in that: When the computer program / instruction is executed by a processor, the structural fatigue life prediction method according to any one of claims 1 to 7 is implemented.

Images