A three-dimensional crack real-time monitoring and reconstruction method fusing multi-source information

Abstract

The application discloses a kind of fusion multi-source information's three-dimensional crack real-time monitoring and reconstruction method, it is related to crack analysis technical field, including the following steps: step 1: mechanism data are obtained by simulation;Step 2: guided wave monitoring data set is constructed;Step 3: weighted fusion data set is obtained;Step 4: initial predicted crack length is obtained by inputting fusion data set into Gaussian process regression model, and the most crack length is obtained by correction;Step 5: crack front coordinates are obtained by conversion;Step 6: crack front coordinates are input into three-dimensional crack propagation model, and the stress intensity factor of basic crack propagation mode and crack direction parameter are obtained;Step 7: material parameters C and m in Paris formula are calculated according to predicted crack length, and finally fatigue life is calculated by Paris formula.The method of the application realizes three-dimensional crack accurate inversion and the continuous monitoring of non-penetrating crack propagation.

Description

Technical Field

[0001] This invention relates to the field of crack analysis technology, and in particular to a method for real-time monitoring and reconstruction of three-dimensional cracks that integrates multi-source information. Background Technology

[0002] Marine engineering structures are key platforms for the development and utilization of marine resources. During long-term service, fatigue damage under complex stress fields is the most prominent failure cause. Fatigue cracks often initiate at critical welded joints, and the high-salt, high-humidity marine environment significantly accelerates crack propagation, easily leading to catastrophic failure, causing significant economic losses and environmental pollution. Therefore, early, accurate, and continuous fatigue crack monitoring and evolution assessment are core requirements for ensuring structural safety, extending service life, and reducing operation and maintenance costs.

[0003] The existing technology has the following shortcomings: 1. Traditional non-destructive testing (DIC, electromagnetic eddy current, infrared thermal imaging) is only suitable for controlled laboratory conditions and cannot achieve real-time monitoring of marine field and complex node cracks, nor can it integrate multi-source information to invert crack morphology; moreover, it is highly dependent on manual labor, inefficient, and difficult to monitor continuously over long distances, resulting in serious lack of field applicability.

[0004] 2. Although ultrasonic guided waves (UGW) are highly sensitive to micro-cracks, adaptable to complex structures, have long propagation distances and low attenuation, existing feature extraction methods are difficult to reliably and stably separate crack-sensitive features from the original guided wave signals. They rely too much on wave propagation mechanisms and a large number of numerical simulations, lack adaptability, and have poor robustness and weak generalization ability under complex marine noise and structural interference.

[0005] 3. Existing models and algorithms also have significant shortcomings: Although pure data-driven neural networks can fit data, they lack mechanical mechanism constraints, feature extraction is blind, physical meaning is vague, generalization is poor, they are prone to overfitting, and they are difficult to adapt to the complex nonlinearity of marine environment and crack evolution.

[0006] 4. Pure fracture mechanics mechanism models rely on fixed parameters, are highly dependent on empirical values ​​for calculation, have low computational efficiency, and cannot dynamically adjust the inversion of three-dimensional morphology based on experimental data.

[0007] 5. Current neural networks struggle to automatically extract effective features of crack evolution in this scenario, are sensitive to noise, and have weak cross-condition transfer capabilities, failing to achieve continuous and accurate reconstruction of 3D crack morphology. In particular, surface crack propagation lifetime accounts for more than 80% of the total lifetime, while existing methods cannot effectively monitor the overall 3D morphology before it penetrates, highlighting a significant gap in key technologies.

[0008] 6. Existing three-dimensional reconstruction methods rely on calibrated samples, are sensitive to the environment, lack real-time adaptive capabilities, and have poor versatility and robustness.

[0009] The above content is only used to help understand the technical solution of the present invention and does not represent an admission that the above content is prior art. Summary of the Invention

[0010] The purpose of this invention is to address the shortcomings of existing technologies by proposing a real-time monitoring and reconstruction method for three-dimensional cracks that integrates multi-source information. This method is based on UGW, mechanism-data dual-drive, and uncertainty modeling to detect and reconstruct three-dimensional fatigue cracks. It specifically addresses the core pain points of existing technologies, such as unreliable feature extraction, poor generalization of neural networks, and disconnect between mechanism and data, thereby achieving accurate three-dimensional crack inversion and continuous monitoring of non-penetrating crack propagation.

[0011] To achieve the above objectives, the present invention adopts the following technical solution: A method for real-time monitoring and reconstruction of three-dimensional cracks by integrating multi-source information includes the following steps: Step 1: Obtain mechanism data through joint simulation using ABAQUS and FRANC3D, preprocess the mechanism data, and construct the initial mechanism dataset X. mech-raw ; Step 2: Conduct cyclic loading tests on a fatigue testing machine. Acquire raw signals through an ultrasonic acquisition system with two receivers. After preprocessing the raw signals, perform weighted summation using a dual-channel convolutional neural network to construct the guided wave monitoring dataset X. data ; Step 3: Convert the initial mechanism dataset X mech-raw A nonlinear dimensionality transformation is performed to obtain the guided wave monitoring dataset X. data Dimensionally Consistent Mechanism Dataset X mech Mechanism dataset X mech And guided wave monitoring dataset X data The weighted fusion dataset X is obtained fusion ; Step 4: Merge dataset X fusion Input the Gaussian process regression model to obtain the initial predicted crack length a. pre The crack length a is obtained by setting its confidence interval and adjusting it. corr ; Step 5: Using the initial mechanism dataset X mech-raw , obtain and The corresponding crack depth h, and according to The crack front coordinates are obtained by converting h; Step 6: Input the crack front coordinates into the three-dimensional crack propagation model to obtain the stress intensity factor K of the basic crack propagation mode. I K II K III And crack direction parameters e1, e2, e3; Step 7: Stress intensity factor K for the basic crack propagation mode I K II K III Calculate the stress intensity factor ΔK based on the predicted crack length a. pre.i The material parameters C and m in the Paris formula are calculated, and the fatigue life is finally calculated using the Paris formula.

[0012] Furthermore, in step 1, the geometric and material parameters of the specimen are input into ABAQUS software to complete the static stress analysis of the structure and obtain the stress distribution results. Then, the stress distribution results obtained from the analysis are imported into FRANC3D software, and an initial crack is inserted in a preset area of ​​the specimen. The crack propagation process is simulated step by step with a preset step size to generate an initial mechanism dataset. In the crack propagation area, a local finite element mesh is generated using a mesh template. The crack leading edge area is arranged with three layers of ring elements. The innermost layer uses C3D15 pentahedral elements, and the outer two layers use C3D20 hexahedral elements to ensure accurate calculation of stress singularity.

[0013] Furthermore, in step 1, the preprocessing of the mechanism data includes two steps: First, normalize the maximum and minimum values ​​of all feature variables to uniformly map the data to the interval [-1, 1] and eliminate dimensional differences; Second, use principal component analysis to reduce the dimensionality of the normalized data and project the high-dimensional data into a low-dimensional orthogonal subspace.

[0014] Furthermore, in step 2, the original signal undergoes time-domain signal processing through normalization, baseline correction, and crosstalk suppression to highlight crack-related features; then, continuous wavelet transform is performed using Morlet wavelet basis functions to obtain time-frequency features, converting the time-domain signal into a two-dimensional time-frequency graph.

[0015] Furthermore, in step 2, h1 and h2 are obtained separately through a dual-channel convolutional neural network, and weighting is achieved using the following formula: ; in, , These are the weights of a two-channel convolutional neural network.

[0016] Furthermore, in step 3, the initial mechanism dataset X is processed using the following formula. mech-raw Perform non-linear dimensional transformation: ; in, As weight, For bias; Mechanism Dataset X mech And guided wave monitoring dataset X data Weighted summation, the formula is as follows: ; in, ; Where T is the stress triaxiality, ΔK is the stress intensity factor, and k is the weighting adjustment coefficient. The weights of the mechanism dataset after dimensionality transformation. Weights for the guided wave monitoring dataset.

[0017] Furthermore, in step 4, the initial predicted crack length a is calculated using the following formula. pre Make corrections: ; Where m is the weight adjustment coefficient, a exp The crack size is calibrated in real time during ultrasonic guided wave testing of a flat plate specimen under tensile conditions at a standstill. f(T) is the stress triaxiality correction term, with a value range of [0.5, 1.5].

[0018] Furthermore, in step 6, the three-dimensional crack propagation model is an MLP neural network.

[0019] Furthermore, in step 7, the stress intensity factor K of the basic crack propagation mode I K II K III The stress intensity factor ΔK is obtained through the following formula: ; Where α is the adjustment parameter, It is Poisson's ratio.

[0020] Furthermore, in step 7, the material parameters C and m in the Paris formula are calculated using the following formula: ; ; Where n is the total number of samples in the training set, and i is the sample traversal index. To predict crack length, To measure the crack length, The consistency coefficient of mechanism-data feature fusion; and The parameters of Paris are updated after the (k+1)th iteration. and This represents the current parameter value at the k-th iteration. and The learning rate step size is used to control the magnitude of each update, prevent oscillations, and ensure convergence. Let C be the partial derivative of the loss function with respect to C. The gradient is the partial derivative of the loss function with respect to m, representing the magnitude of the influence of the current parameters on the total error. The larger the gradient, the worse the current parameters are. and These are the initial Paris parameters. is the regularization coefficient.

[0021] Compared with the prior art, the beneficial effects of this invention are as follows: It can integrate data-driven models and fracture mechanics mechanisms to construct a three-dimensional fatigue crack morphology reconstruction model, accurately capturing the progressive propagation morphology of non-penetrating surface cracks. Simulation verification shows a prediction error of less than 5%. A dual-channel adaptive attention fusion network is proposed, and the fused features are dynamically fused with mechanism-data features. The fused feature values ​​are combined with Gaussian process probability modeling to obtain the predicted crack length, realizing multi-sensor information complementarity and prediction uncertainty quantification. The maximum prediction error of crack length is 9.59%, and the prediction error of depth is less than 10%. Based on ultrasonic guided waves, it realizes non-contact, long-distance, real-time online monitoring, and can reconstruct the morphology of three-dimensional cracks. It is suitable for full-cycle monitoring of fatigue cracks in marine engineering steel structures, and has significant engineering application value. Attached Figure Description

[0022] Figure 1 A flowchart of a method for real-time monitoring and reconstruction of three-dimensional cracks that integrates multi-source information; Figure 2 This is a diagram showing the structural dimensions of the sample. Figure 3 Introducing and reconstructing the mesh for cracks; Figure 4 Stress intensity factor K I One-dimensional space representation; Figure 5 Stress intensity factor K I Two-dimensional spatial representation; Figure 6 A one-dimensional spatial representation of the crack front; Figure 7 This represents the two-dimensional space of the crack front. Figure 8 Stress intensity factor K I The first principal component loss curve; Figure 9 Stress intensity factor K I The second principal component loss curve; Figure 10 Stress intensity factor K I The third principal component loss curve; Figure 11 Diagram of the finite element model of an ultrasonic guided wave; Figure 12This is the time-frequency diagram of the crack signal after continuous wavelet transform; Figure 13 This is a fatigue life diagram of the sample; Figure 14 Comparison of crack propagation paths; Figure 15 Initial noise map of the receiving point; Figure 16 Signal diagrams of receiving points with different crack lengths; Figure 17 Simulation signal diagrams for different crack lengths; Figure 18 This is a comparison chart of experimental and simulated signals; Figure 19 The loss curves for model training and validation are shown. Figure 20 This is a prediction diagram of fatigue crack propagation probability. Figure 21 A comparison chart showing the predicted and experimental crack depths. Detailed Implementation

[0023] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0024] Example 1: A method for real-time monitoring and reconstruction of three-dimensional cracks that integrates multi-source information, such as Figure 1 As shown, it includes the following steps: Step 1: Obtain mechanism data through joint simulation using ABAQUS and FRANC3D, preprocess the mechanism data, and construct the initial mechanism dataset X. mech-raw .

[0025] In this embodiment, in step 1, the geometric and material parameters of the specimen are input into ABAQUS software to complete the static stress analysis of the structure and obtain the stress distribution results. Then, the stress distribution results obtained from the analysis are imported into FRANC3D software, and an initial crack is inserted in a preset area of ​​the specimen. The crack propagation process is simulated step by step with a preset step size to generate an initial mechanism dataset. In the crack propagation area, a local finite element mesh is generated using a mesh template. The crack leading edge area is arranged with three layers of ring elements. The innermost layer uses C3D15 pentahedral elements, and the outer two layers use C3D20 hexahedral elements to ensure accurate calculation of stress singularity.

[0026] In this embodiment, in step 1, the data obtained by the joint simulation of ABAQUS and FRANC3D includes the front coordinates and their corresponding stress triaxiality, and the crack length, crack depth, stress intensity factor and crack direction parameters are obtained by conversion, and the initial mechanism dataset is constructed in sequence.

[0027] In this embodiment, step 1, the preprocessing of the mechanism data includes two steps: First, normalize the maximum and minimum values ​​of all feature variables to uniformly map the data to the interval [-1, 1], eliminate the difference in dimensions, and accelerate the convergence speed of the subsequent model; Second, use principal component analysis (PCA) to reduce the dimensionality of the normalized data and project the high-dimensional data to a low-dimensional orthogonal subspace.

[0028] In this embodiment, in step 1, principal component analysis is used to reduce the data dimensionality, which can retain more than 99% of the original information, while reducing redundancy and computational overhead, and improving the efficiency of subsequent model training and prediction stability.

[0029] Step 2: Conduct cyclic loading tests on a fatigue testing machine. Acquire raw signals through an ultrasonic acquisition system with two receivers. After preprocessing the raw signals, perform weighted summation using a dual-channel convolutional neural network to construct the guided wave monitoring dataset X. data .

[0030] In this embodiment, step 2 involves constructing an ultrasonic guided wave acquisition system using a signal generator, oscilloscope, 40dB amplifier, and three PZT-5H piezoelectric crystals. One crystal serves as the excitation end, and the other two as the receiving ends. These crystals are firmly attached to the specimen surface using specialized adhesive to ensure stable acoustic coupling. The experimental environment is maintained at 25°C and 40% relative humidity. Cyclic loading tests are conducted on a fatigue testing machine with a loading frequency of 10Hz, a load ratio of 0.1, and a minimum load of 6kN. Loading is paused every 5000 cycles to maintain static tension while acquiring guided wave signals. The acquired raw signals are preprocessed, including signal normalization, baseline correction, and crosstalk suppression. The signal from a crack-free specimen under the same load is used as a benchmark. The crack signal is then compared with the benchmark signal to highlight the signal changes caused by crack propagation.

[0031] In this embodiment, in step 2, the original signal undergoes time-domain signal processing through normalization, baseline correction, and crosstalk suppression to highlight crack-related features. Then, continuous wavelet transform is performed using Morlet wavelet basis functions to obtain time-frequency features, converting the time-domain signal into a two-dimensional time-frequency graph. The transformed two-dimensional time-frequency graph clearly shows that the signal energy is concentrated near the excitation frequency, and as the crack continues to expand, the signal's scattering characteristics, spectral components, and energy distribution will exhibit identifiable continuous changes, providing highly discriminative features for crack length prediction.

[0032] In this embodiment, in step 2, h1 and h2 are obtained respectively through a dual-channel convolutional neural network, and weighting is achieved using the following formula: ; in, , These are the weights of a two-channel convolutional neural network.

[0033] In this embodiment, in step 2, the dual-channel convolutional neural network and the weighting process are trained end-to-end. The training dataset can be obtained through joint simulation of ABAQUS and FRANC3D, or through experiments.

[0034] Step 3: Convert the initial mechanism dataset X mech-raw A nonlinear dimensionality transformation is performed to obtain the guided wave monitoring dataset X. data Dimensionally Consistent Mechanism Dataset X mech Mechanism dataset X mech And guided wave monitoring dataset X data The weighted fusion dataset X is obtained fusion .

[0035] In this embodiment, in step 3, the initial mechanism dataset X is processed using the following formula. mech-raw Perform non-linear dimensional transformation: ; in, As weight, For bias.

[0036] In this embodiment, in step 3, the mechanism dataset X mech And guided wave monitoring dataset X data Weighted summation, the formula is as follows: ; in, ; Where T is the stress triaxiality, ΔK is the stress intensity factor, and k is a weighting adjustment coefficient, which can be taken as 0.8, used to couple the stress triaxiality and the stress intensity factor. The weights of the mechanism dataset after dimensionality transformation. Weights for the guided wave monitoring dataset.

[0037] In this embodiment, in step 3, the higher the weight of the mechanism feature, the stronger the fracture mechanistic mechanism constraint of the crack; as the crack continues to expand, the adaptive balance between mechanism and data is dynamically achieved.

[0038] Step 4: Merge dataset X fusion Input the Gaussian process regression model to obtain the initial predicted crack length a.pre The crack length a is obtained by setting its confidence interval and adjusting it. corr .

[0039] In this embodiment, the training dataset for the Gaussian process regression model in step 4 can be obtained through joint simulation using ABAQUS and FRANC3D, or it can be obtained experimentally. The fused dataset X... fusion Input the Gaussian process regression model to obtain the initial predicted crack length a. pre It uses its confidence interval to achieve high-precision and robust quantitative assessment of crack length.

[0040] In this embodiment, in step 4, the initial predicted crack length a is calculated using the following formula. pre Make corrections: ; Where m is the weight adjustment coefficient, a exp The crack size, calibrated in real time during ultrasonic guided wave testing of a flat plate specimen under tensile conditions at rest, is the core data constraint for crack size correction. It is generally obtained by direct reading through a high-magnification microscope and is used to correct the deviation of the initial predicted crack length. f(T) is the stress triaxiality correction term, with a value range of [0.5, 1.5], used to characterize the damage effect of stress triaxiality on crack propagation. That is, the higher the stress triaxiality, the faster the plastic damage accumulates at the crack tip of the flat plate. The larger the value of this correction term, the corresponding adjustment of the correction range is made to ensure that the correction result conforms to the fracture mechanics mechanism.

[0041] Step 5: Using the initial mechanism dataset X mech-raw , obtain and The corresponding crack depth h, and according to The coordinates of the crack leading edge are obtained by converting h.

[0042] Step 6: Input the crack front coordinates into the three-dimensional crack propagation model to obtain the stress intensity factor K of the basic crack propagation mode. I K II K III And crack direction parameters e1, e2, e3.

[0043] In this embodiment, in step 6, K I K is the stress intensity factor for an open crack. II K is the stress intensity factor for a slip crack. III Let e1 be the stress intensity factor for a tearing crack, e3 be the unit vector of the local crack leading edge normal, e2 be the unit vector of the local crack leading edge tangent, and e3 be the cross product of the two.

[0044] In this embodiment, in step 6, the three-dimensional morphology of the crack is inverted using a three-dimensional crack propagation model. The corresponding crack depth is obtained by combining the crack propagation mechanism, achieving a complete reconstruction of the crack's three-dimensional morphology. Subsequently, the crack front is physically marked using the beach marking method, and the predicted crack length and depth are compared with the measured values ​​from the beach markings.

[0045] In this embodiment, in step 6, the three-dimensional crack propagation model is an MLP neural network, and the training dataset can be obtained through joint simulation using ABAQUS and FRANC3D. During training, the optimizer can be Adam, with 200 training epochs. The model training loss function is mean squared error, ensuring that the training and validation loss curves closely match. The loss curves of the training set and the validation set continuously decrease and remain close, without overfitting, and eventually converge and stabilize. For K... I The training results of the three principal components show that the relative error between the model's predicted values ​​and the simulation data is less than 2%, demonstrating good prediction accuracy and generalization ability.

[0046] Step 7: Stress intensity factor K for the basic crack propagation mode I K II K III Calculate the stress intensity factor ΔK based on the predicted crack length a. pre.i The material parameters C and m in the Paris formula are calculated, and the fatigue life is finally calculated using the Paris formula.

[0047] In this embodiment, in step 7, the stress intensity factor K of the basic crack propagation mode I K II K III The stress intensity factor ΔK is obtained through the following formula: ; Where α is an adjustment parameter, which can be 1. It is Poisson's ratio.

[0048] In this embodiment, in step 7, the material parameters C and m in the Paris formula are calculated using the following formula: ; ; Where n is the total number of samples in the training set, and i is the sample traversal index. To predict crack length, To measure the crack length, The consistency coefficient of mechanism-data feature fusion; and The parameters of Paris are updated after the (k+1)th iteration. and This represents the current parameter value at the k-th iteration. and The learning rate step size is used to control the magnitude of each update, prevent oscillations, and ensure convergence. Let C be the partial derivative of the loss function with respect to C. The gradient is the partial derivative of the loss function with respect to m, representing the magnitude of the influence of the current parameters on the total error. The larger the gradient, the worse the current parameters are. and These are the initial Paris parameters. is the regularization coefficient.

[0049] In this embodiment, in step 7, K can be obtained through a Gaussian process regression model, or from a three-dimensional crack propagation model. I K II K III It was obtained through the median extension method.

[0050] In this embodiment, in step 7, the Paris formula is as follows: ; in, The fatigue crack propagation rate is given by ΔK, and the fatigue life can be obtained by accumulating the fatigue crack propagation rate. ΔK is the stress intensity factor. C represents the stress intensity factor threshold value, which varies depending on the structural material. C and m are material constants.

[0051] In this embodiment, in step 7, the crack propagation rate is calculated using the Paris formula, the torsional angle and extension direction of crack propagation are determined based on the maximum circumferential stress criterion, and the crack propagation increment is determined by combining the median propagation method, thereby achieving stable prediction of the three-dimensional crack propagation path.

[0052] This embodiment presents a three-dimensional crack real-time monitoring and reconstruction method that integrates multi-source information. It breaks through the dilemma of "data-driven methods relying solely on experimental data and mechanism-driven methods relying solely on fracture mechanics theory derivation". It constructs a dual-drive crack monitoring method that integrates fracture mechanics mechanism and experimental data. It integrates ultrasonic guided wave characteristic dynamic prediction, and through ultrasonic guided wave signal acquisition, dual-channel adaptive attention fusion, Gaussian process probability prediction, data-driven crack morphology reconstruction, crack size correction and update, fatigue crack propagation parameter update and life prediction, it realizes accurate inversion and continuous monitoring of the three-dimensional morphology of the entire process from non-penetrating surface fatigue crack to through crack.

[0053] This embodiment of a three-dimensional crack real-time monitoring and reconstruction method integrating multi-source information can obtain crack depths that are difficult to measure in actual sea conditions by other non-destructive testing methods, as well as fatigue life parameters that conform to the actual situation at that time. This enables the determination of whether the crack has penetrated the wall thickness, the remaining life of the structure, and whether it meets the conditions for safe operation and the design life.

[0054] Addressing the shortcomings of existing technologies: To address shortcoming 1: This solution uses a combination of high-fidelity co-simulation and experimentation, combining physical mechanisms with simulation data, to ensure high-precision crack morphology inversion while enabling continuous monitoring of the entire crack propagation process.

[0055] Addressing shortcoming 2: This solution integrates neural network models and physical mechanism information, fuses experimental data and simulation data for training, and combines multi-source information to accurately capture crack features. This enables adaptive adjustment of multi-source information at different crack propagation stages, resulting in high computational efficiency and high-precision inversion of three-dimensional crack morphology.

[0056] To address shortcoming 3: This solution establishes a dataset through co-simulation, which includes three-dimensional crack morphology data. It integrates fracture mechanics theory into the three-dimensional crack morphology reconstruction model, enhancing the interpretability of crack propagation iteration. At the same time, when monitoring crack length with ultrasonic guided waves, it integrates mechanism data with monitoring data, improving monitoring accuracy and enhancing the robustness and generalization of the ultrasonic guided wave crack length prediction model.

[0057] Addressing shortcoming 4: This solution integrates the fracture mechanics mechanism of three-dimensional crack propagation into the MLP-based three-dimensional crack propagation model, enabling automatic propagation of three-dimensional cracks and inversion of their corresponding morphology. Simultaneously, by fusing the mechanism data with experimental values ​​from guided wave detection, the crack length results input into the three-dimensional reconstruction model are more accurate and closer to the actual crack length data, thereby greatly improving the inversion accuracy of the three-dimensional crack morphology.

[0058] Addressing shortcoming 5: This solution constructs a data-mechanism dual-driven feature fusion model to effectively mine sensitive features of crack evolution, suppress complex marine noise interference, significantly improve computational efficiency and prediction accuracy, accurately capture the three-dimensional morphological evolution features of fatigue cracks from the surface to the entire process, realize continuous and accurate reconstruction of three-dimensional crack morphology, and make up for the technical shortcomings of full-process monitoring of the three-dimensional morphology of non-penetrating cracks.

[0059] Addressing shortcoming 6: This solution introduces an uncertainty adaptive correction strategy to break free from the constraints of traditional fixed calibration samples. It dynamically optimizes model parameters by combining real-time on-site monitoring data, adaptively adapts to complex time-varying working conditions, effectively reduces reconstruction errors caused by environmental disturbances, and significantly improves the real-time performance, versatility, and environmental robustness of the three-dimensional crack reconstruction method.

[0060] Verification of Examples To verify the effectiveness of the method in Example 1, the following specific examples were conducted: First, a model of the steel plate structure used in the experiment is created, such as... Figure 2 As shown, the specimen dimensions are 210 mm in length, 35 mm in width, and 6 mm in thickness, with a material density of 7847 kg / m³. 3 The specimen has an elastic modulus of 210,000 MPa, Poisson's ratio of 0.3, tensile strength of 573 MPa, and yield strength of 377 MPa. These geometric and material parameters were input into ABAQUS software to perform static stress analysis and obtain the stress distribution results. The resulting file was then imported into FRANC3D software, where an initial crack was inserted within a 4 mm wide region at the center of the specimen. The crack propagation process was simulated step-by-step with a fixed step size of 0.5 mm. The stress analysis results and crack propagation data were combined to form a fracture mechanistic dataset.

[0061] Based on the sample size, stress analysis was performed in ABAQUS, and then... Figure 3 The crack introduction and mesh reconstruction diagram shows the process of inserting an initial crack in FRANC3D and refining the local mesh at the crack leading edge—the crack leading edge uses three layers of ring elements (inner layer C3D15, outer layer C3D20), which provides a high-precision calculation basis for subsequent crack propagation simulation.

[0062] After completing the simulation dataset construction, structural mechanism data such as crack front coordinates and stress triaxiality are projected into a low-dimensional space using principal component analysis (PCA), such as... Figure 4-7 As shown, the first two principal components can retain more than 99% of the original data variance information, intuitively presenting the stress characteristic distribution under different crack states.

[0063] Based on these low-dimensional features, a three-dimensional crack propagation model was constructed to establish a nonlinear mapping relationship between the crack state and the stress intensity factor. For example... Figure 8-10 As shown, the model's performance in predicting the stress intensity factor K is verified. I The convergence performance of each principal component was observed, and the model stabilized after 200 training epochs. K I The relative error of principal component prediction is less than 2%, with no obvious overfitting.

[0064] In order to obtain ultrasonic guided wave signal data, establish Figure 11The ultrasonic guided wave finite element model was used to simulate the guided wave propagation process under different crack states using a 250kHz, 5-cycle Hanning window excitation signal. Figure 12 This demonstrates the result of converting the time-domain signal into a time-frequency diagram using continuous wavelet transform (Morlet wavelet basis). The energy is concentrated at the excitation center frequency, clearly showing the signal scattering and energy distribution changes caused by crack propagation.

[0065] During the experiment, Figure 13 The fatigue life curve of the specimen was recorded. During the test, the crack length at the corresponding life was continuously measured. Under the conditions of 10Hz loading frequency, 0.1 load ratio and 6kN minimum load, the total fatigue life of the specimen was about 96,660 cycles, which provided a basis for the division of crack propagation stages. The experimental records were fitted to obtain the theoretical value of fatigue life. Figure 14 The comparison results of the crack propagation path prediction of the three-dimensional crack reconstruction model show that the paths are very similar, and the simulation verification shows that the prediction error is less than 5%.

[0066] Signal acquisition phase, Figure 15 The pure noise collected during the experiment provides a benchmark for subsequent signal preprocessing (baseline correction, crosstalk suppression); Figure 16 and Figure 17 The test signals and simulation signals (pure noise before wavelet transform, which is converted into a time-frequency image after wavelet transform) under different crack lengths are presented respectively, which intuitively reflects the influence of crack propagation on the amplitude and phase of the guided wave signal; Figure 18 The experimental and simulated signals were then compared, and the waveforms and characteristic peak positions of the two were basically consistent, which verified the accuracy of the finite element model. This verifies that the simulation data can be used as a supplement to the experimental data to participate in the model training.

[0067] Model training phase, Figure 19 The training and validation loss curves of the dual-channel fusion network are presented. The model uses the Adam optimizer with a batch size of 10. After 200 training rounds, the loss steadily decreases and tends to converge, with no overfitting. Figure 18 The fatigue crack propagation probability prediction plot shows the model's prediction results for crack length and its uncertainty distribution. Test results from three sets of specimens show that the maximum error in crack length prediction is 9.59%, and the root mean square error is as low as 0.093. This figure shows the prediction results from the Gaussian process regression model, including the crack length and confidence interval at the corresponding fatigue life; finally, Figure 20 By comparing the predicted crack depth with the measured data of the beach marking method, it can be shown that during the experiment, when the crack length is monitored by ultrasonic guided wave, the crack morphology can be marked on the fracture surface by the beach strip method. After the experiment, the fracture surface can be compared with the three-dimensional crack morphology reconstruction model to verify the reconstruction accuracy of the three-dimensional reconstruction model. Figure 21 The crack depth is derived from the data calibrated by the mark lines during the experiment. The maximum error in crack depth prediction is less than 10%, verifying the prediction accuracy of the three-dimensional crack morphology reconstruction model and completing the full-process verification from signal acquisition to three-dimensional crack morphology inversion. Subsequently, the crack propagation parameters C and m are updated, and the remaining life of the structure is calculated using the updated propagation parameters. The updated parameter C is 2.6 × 10⁻⁶. -15 The value of parameter m is 3.58.

[0068] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for real-time monitoring and reconstruction of three-dimensional cracks by integrating multi-source information, characterized in that, Includes the following steps: Step 1: Obtain mechanism data through joint simulation using ABAQUS and FRANC3D, preprocess the mechanism data, and construct the initial mechanism dataset X. mech-raw ; Step 2: Conduct cyclic loading tests on a fatigue testing machine. Acquire raw signals through an ultrasonic acquisition system with two receivers. After preprocessing the raw signals, perform weighted summation using a dual-channel convolutional neural network to construct the guided wave monitoring dataset X. data ; Step 3: Convert the initial mechanism dataset X mech-raw A nonlinear dimensionality transformation is performed to obtain the guided wave monitoring dataset X. data Dimensionally Consistent Mechanistic Dataset X mech Mechanism dataset X mech And guided wave monitoring dataset X data The weighted fusion dataset X is obtained fusion The initial mechanism dataset X is analyzed using the following formula. mech-raw Perform non-linear dimensional transformation: ； in, As weight, For bias; Mechanism Dataset X mech And guided wave monitoring dataset X data Weighted summation, the formula is as follows: ； in, ； Where T is the stress triaxiality, ΔK is the stress intensity factor, and k is the weighting adjustment coefficient. The weights of the mechanism dataset after dimensional transformation. Weights for the guided wave monitoring dataset; Step 4: Merge dataset X fusion Input the Gaussian process regression model to obtain the initial predicted crack length a. pre The crack length a is obtained by setting its confidence interval and adjusting it. corr The initial predicted crack length a is calculated using the following formula. pre Make corrections: ； Where m is the weight adjustment coefficient, a exp The crack size is calibrated in real time during ultrasonic guided wave testing of a flat plate specimen under tensile conditions at a standstill. f(T) is the stress triaxiality correction term, with a value range of [0.5, 1.5]. Step 5: Using the initial mechanism dataset X mech-raw , obtain and The corresponding crack depth h, and according to The crack front coordinates are obtained by converting h; Step 6: Input the crack front coordinates into the three-dimensional crack propagation model to obtain the stress intensity factor K of the basic crack propagation mode. I K II K III And crack direction parameters e1, e2, e3; Step 7: Stress intensity factor K for the basic crack propagation mode I K II K III Calculate the stress intensity factor ΔK based on the predicted crack length a. pre.i The material parameters C and m in the Paris formula are calculated, and the fatigue life is finally calculated using the Paris formula.

2. The method for real-time monitoring and reconstruction of three-dimensional cracks by fusing multi-source information as described in claim 1, characterized in that, In step 1, the geometric and material parameters of the specimen are input into ABAQUS software to complete the static stress analysis of the structure and obtain the stress distribution results. Then, the stress distribution results obtained from the analysis are imported into FRANC3D software, and an initial crack is inserted in a preset area of ​​the specimen. The crack propagation process is simulated step by step with a preset step size to generate an initial mechanism dataset. In the crack propagation area, a mesh template is used to generate a local finite element mesh. The crack leading edge area is arranged with three layers of ring elements. The innermost layer uses C3D15 pentahedral elements, and the outer two layers use C3D20 hexahedral elements to ensure accurate calculation of stress singularity.

3. The method for real-time monitoring and reconstruction of three-dimensional cracks by fusing multi-source information according to claim 1, characterized in that, Step 1 involves two preprocessing steps for the mechanistic data: First, normalize all feature variables by their maximum and minimum values, mapping the data uniformly to the interval [-1, 1] to eliminate dimensional differences; Second, use principal component analysis to reduce the dimensionality of the normalized data, projecting the high-dimensional data into a low-dimensional orthogonal subspace.

4. The method for real-time monitoring and reconstruction of three-dimensional cracks by fusing multi-source information as described in claim 1, characterized in that, In step 2, the original signal is processed in the time domain by normalization, baseline correction and crosstalk suppression to highlight crack-related features; then, the time-frequency features are obtained by continuous wavelet transform using Morlet wavelet basis functions, and the time-domain signal is converted into a two-dimensional time-frequency graph.

5. The method for real-time monitoring and reconstruction of three-dimensional cracks by fusing multi-source information according to claim 1, characterized in that, In step 2, h1 and h2 are obtained separately through a dual-channel convolutional neural network, and weighting is achieved using the following formula: ； in, , These are the weights of a two-channel convolutional neural network.

6. The method for real-time monitoring and reconstruction of three-dimensional cracks by fusing multi-source information according to claim 1, characterized in that, In step 6, the three-dimensional crack propagation model is an MLP neural network.

7. The method for real-time monitoring and reconstruction of three-dimensional cracks by fusing multi-source information according to claim 1, characterized in that, In step 7, the stress intensity factor K of the basic crack propagation mode I K II K III The stress intensity factor ΔK is obtained through the following formula: ； Where α is the adjustment parameter, It is Poisson's ratio.

8. The method for real-time monitoring and reconstruction of three-dimensional cracks by fusing multi-source information according to claim 1, characterized in that, In step 7, the material parameters C and m in the Paris formula are calculated using the following formula: ； ； Where n is the total number of samples in the training set, and i is the sample traversal index. To predict crack length, To measure the crack length, The consistency coefficient of mechanism-data feature fusion; and The parameters of Paris are updated after the (k+1)th iteration. and This represents the current parameter value at the k-th iteration. and The learning rate step size is used to control the magnitude of each update, prevent oscillations, and ensure convergence. Let C be the partial derivative of the loss function with respect to C. The gradient is the partial derivative of the loss function with respect to m, representing the magnitude of the influence of the current parameters on the total error. The larger the gradient, the worse the current parameters are. and These are the initial Paris parameters. is the regularization coefficient.

Images