Intelligent prediction method for fatigue crack growth in near-threshold value area of welding joint

Abstract

The invention belongs to the technical field of structure prediction, and particularly relates to an intelligent prediction method for fatigue crack propagation of a near-threshold value region of a welding joint, which comprises the steps of S1, collecting fatigue crack propagation data of the near-threshold value region of the welding joint, S2, constructing a physically-guided multi-source fatigue crack propagation data set, S3, training a neural network prediction model introducing physical consistency constraint, and S4, predicting the fatigue crack propagation of the near-threshold value region of the welding joint. S4, monitoring working condition data and loading history in the equipment service process; and S5, establishing the fatigue crack propagation digital twin based on the model. According to the intelligent prediction method for the fatigue crack growth of the near-threshold value region of the welding joint, through physically guided data fusion and physical consistency constraint learning, the model is endowed with the capability of keeping predicted physical rationality and engineering availability under the condition of data scarcity; according to the constructed multi-source data set, through fusion coding of physical information and experimental data and fusion of material performance and tissue characteristics, the model can process fatigue behavior differences of different areas, and therefore the digital twinborn body is constructed.

Description

Technical Field

[0001] This invention relates to the field of structural integrity assessment and intelligent diagnosis, specifically to the technical field of cross-integration of fracture mechanics and artificial intelligence, and particularly to an intelligent prediction method for fatigue crack propagation in the near-threshold zone of welded joints. Background Technology

[0002] Welded structures are widely used in major engineering fields such as shipbuilding, nuclear power equipment, pressure vessels, and the manufacturing of large-scale engineering equipment, and are one of the most common connection forms in engineering structures. However, due to the thermal cycling effect introduced during the welding process, welded structures exhibit significant inhomogeneities in terms of microstructure and mechanical properties. This makes welded joints more susceptible to crack initiation and propagation during service, and their fatigue resistance directly determines the safety and reliability of the engineering structure. Regarding fatigue issues, the fatigue crack propagation threshold value is a key parameter in damage tolerance design and structural safety evaluation. Especially under long-life service conditions, accurately understanding the fatigue crack propagation behavior near the threshold value region is of great significance for the safety assessment and remaining service life prediction of welded structures.

[0003] Traditional methods for obtaining fatigue crack propagation behavior near the threshold region mainly include experimental testing and theoretical modeling. Experimental testing can directly obtain crack propagation behavior, but the testing cycle is long and costly, and due to the complexity of the near-threshold region mechanism and the inherent inhomogeneity of welded joints, the experimental results often exhibit significant dispersion. Physical models built based on experimental data incorporate crack propagation mechanisms to some extent, but their prediction accuracy and applicability are still limited by various complex influencing factors. In recent years, data-driven methods such as machine learning have been gradually applied to the prediction of fatigue crack propagation behavior. For example, the methods disclosed in patents CN120387245A, CN115169240A, and CN120429744A have achieved prediction of crack propagation behavior to a certain extent. However, existing related technologies are mainly aimed at homogeneous materials and are difficult to apply to structural forms such as welded joints with significant regional differences and inhomogeneous characteristics. Especially when experimental data is scarce, their predictive ability and reliability are significantly limited.

[0004] Therefore, how to effectively combine the physical mechanism of fatigue crack propagation with data-driven methods under limited experimental data conditions, and establish a method capable of reliably predicting fatigue crack propagation behavior in different regions of welded joints, especially in the near-threshold region, has become a key technical problem urgently needing to be solved in the field of welded structure safety assessment and life prediction. Solving this problem will provide important technical support for damage tolerance design, structural safety assessment, and remaining life prediction of welded structures. Summary of the Invention

[0005] Based on existing technical problems, this invention proposes an intelligent prediction method for fatigue crack propagation in the near-threshold region of welded joints.

[0006] The present invention proposes an intelligent prediction method for fatigue crack propagation in the near-threshold region of a welded joint, comprising S1, collecting fatigue crack propagation data in the near-threshold region of the welded joint, and recording the stress intensity factor range and the corresponding crack propagation rate.

[0007] S2. Construct a physically guided multi-source fatigue crack extended dataset.

[0008] S3. Train a neural network prediction model with physical consistency constraints. Train the prediction model with physical consistency constraints based on the dataset.

[0009] S4. Monitor the operating conditions and loading history of the equipment during its service life to form dynamic input data that can drive the operation of the prediction model and twin.

[0010] S5. Establish a model-based digital twin of fatigue crack propagation to analyze the evolution of crack state over time in real time.

[0011] S6. Conduct a structural safety assessment oriented towards damage tolerance based on the prediction results to determine whether the critical crack size has been reached.

[0012] Preferably, in step S1, compact tensile specimens with pre-existing cracks are prepared for the base metal region, weld region, and heat-affected zone of the welded joint, under different stress ratios. Tests were conducted under specific conditions, and after the tests were completed, the range of stress intensity factors in different regions under different working conditions was recorded. Corresponding crack propagation rate .

[0013] Preferably, in step S2, a unified multi-source fatigue crack propagation dataset is established based on the data points obtained in step S1, combined with the Paris formula and material parameters of each region.

[0014] Preferably, the data points are the stress ratios of the base metal region, weld region, and heat-affected zone of the weld joint. Below .

[0015] The stress ratio is determined by combining the Paris formula. Crack propagation rate in each set of experimental data Through the Paris formula Using stress ratio Paris parameter = 0.9 and Perform inverse operation Get and corresponding and will obtain As one of the inputs to the dataset.

[0016] The material-related parameters for each region include the Vickers hardness of the base material region. ,tensile strength and average grain size .

[0017] Vickers hardness of weld zone ,tensile strength and average grain size .

[0018] Vickers hardness of heat-affected zone ,tensile strength and average grain size .

[0019] Calculate the ratios of each parameter relative to the base material parameters in different regions to obtain the hardness ratio. Strength ratio and size ratio Then integrate all the information to As input, As a label, a multi-source information fusion fatigue crack extended dataset containing experimental data, physical information, and material information was established.

[0020] Preferably, in step S3, the model is trained, validated, or tuned using the dataset constructed in step S2.

[0021] Preferably, during training, an optimizer is selected, and the learning rate, training epochs, and batch size hyperparameters are set.

[0022] Physical constraints are embedded to ensure the physical consistency of the model predictions. The physical model used for these physical constraints is a modified Zhu-Xuan model.

[0023] .

[0024] in, The correction factor was derived based on experimental data. The hardness ratio, strength ratio, and size ratio are weighted by a factor. The weighted average is obtained.

[0025] Preferably, the Zhu-Xuan model is evaluated using the mean absolute error (MAE), mean square error (MSE), root mean square error (RMSE), and coefficient of determination (R²). 2Evaluation metrics are used, and if the metrics do not meet expectations, the hyperparameters are adjusted. When the accuracy of the Zhu-Xuan model meets the requirements, the trained model is stored in the form of a parameter file, and the parameter information required for input data normalization is saved simultaneously.

[0026] Preferably, in step S4, the working condition data of the equipment welding joint is acquired in real time or periodically through a load monitoring system or a structural health monitoring system, and recorded and saved as a loading history.

[0027] The operating data includes, but is not limited to, axial force, bending moment, internal pressure, or combined load information.

[0028] Preferably, in step S5, the neural network prediction model trained and saved in step S3 is deployed in the digital twin system, the service data stream obtained in step S4 is received, a digital twin of fatigue crack propagation in the near-threshold region of the welded joint is constructed, and the crack propagation state is continuously updated.

[0029] Preferably, in step S6, based on the crack length evolution results obtained from the continuous updating of the digital twin in step S5, and combined with the damage tolerance design criteria, the safety status and remaining service capability of the welded joint structure are evaluated.

[0030] The beneficial effects of this invention are as follows:

[0031] By employing physics-guided data fusion and physical consistency constraint learning, the model is endowed with the ability to maintain the physical rationality and engineering usability of its predictions even under data scarcity. The constructed multi-source dataset, through the fusion encoding of physical information and experimental data, and the integration of material properties and microstructure characteristics, enables the model to handle differences in fatigue behavior across different regions, thereby constructing a digital twin. Intelligent prediction utilizes service loads and historical operating conditions to drive the digital twin, achieving continuous characterization of crack behavior under actual service conditions. This is further used for crack length evolution estimation and remaining service life assessment, effectively improving the reliability and applicability of fatigue crack propagation prediction for welded joints, and providing crucial support for structural safety assessment and health management of engineering equipment. Attached Figure Description

[0032] Figure 1 The flowchart shows a method for intelligent prediction of fatigue crack propagation in the near-threshold region of a welded joint, as proposed in this invention.

[0033] Figure 2 This is a neural network architecture diagram of an intelligent prediction method for fatigue crack propagation in the near-threshold region of a welded joint proposed in this invention.

[0034] Figure 3 This is a comparison of the neural network prediction effects of the intelligent prediction method for fatigue crack propagation in the near-threshold region of a welded joint proposed in this invention.

[0035] Figure 4 This is a model prediction diagram of the heat-affected zone R=0.7 under the condition of intelligent prediction method for fatigue crack propagation in the near threshold region of welded joints proposed in this invention. Detailed Implementation

[0036] 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.

[0037] Reference Figures 1-4 A method for intelligent prediction of fatigue crack propagation in the near-threshold region of a welded joint includes S1, collecting fatigue crack propagation data in the near-threshold region of the welded joint. The fatigue crack propagation data is collected specifically as follows:

[0038] For the base metal zone, weld zone, and heat-affected zone of the welded joint, compact tensile specimens with pre-existing cracks were prepared according to ASTM E647 standard, and subjected to different stress ratios. The test was conducted under the specified conditions, including a stress ratio of 0.9. The experimental procedure is well-known in the art, and its specific steps will not be detailed here. After the test, the stress intensity factor range for different regions under different working conditions was recorded. Corresponding crack propagation rate .

[0039] In one embodiment, samples were prepared for the base metal zone, weld zone, and heat-affected zone of a welded joint of a low-alloy high-strength marine steel, respectively, under stress ratios... Tests were conducted at room temperature with stresses of 0.1, 0.3, 0.5, 0.7, and 0.9. The loading frequency was 100 Hz throughout the testing. The load was gradually reduced by 5%, and crack propagation at each load level remained within 4 to 6 times the plastic zone size of the previous level. Crack length was measured using an optical microscope with a resolution of 0.01 mm. Based on the specimen geometry, real-time crack length, and cyclic loading conditions, the maximum and minimum values ​​of the stress intensity factor were calculated according to the standard formula for fracture mechanics, thus obtaining the stress intensity factor range. Simultaneously, the crack propagation rate was calculated using the adjacent multi-point difference method, thereby obtaining at least five values ​​for different regions of the welded joint within the near-threshold region under each condition. Experimental data points.

[0040] S2. Construct a physically-guided multi-source fatigue crack extension dataset. Specifically, it is constructed as follows:

[0041] Based on the data points obtained in S1, and combined with the Paris formula and material-related parameters for each region, a unified dataset for multi-source fatigue crack propagation is established. Specifically, the data points are the stress ratios of the base metal region, weld region, and heat-affected zone of the weld joint. Below The aforementioned combination with the Paris formula refers to the case where R=0.9, for other stress ratios. Crack propagation rate in each set of experimental data Through the Paris formula Using the Paris parameter with R=0.9 and Perform inverse operation Get and corresponding and will obtain As one of the inputs to the dataset. The material-related parameters for each region refer to the Vickers hardness of the parent material region. ,tensile strength and average grain size Vickers hardness of weld zone ,tensile strength and average grain size Vickers hardness in the heat-affected zone ,tensile strength and average grain size Based on this, the ratios of each parameter relative to the base material parameters in different regions are calculated to obtain the hardness ratio. Strength ratio and size ratio Integrate all information to As input, As a label, a multi-source information fusion fatigue crack extended dataset containing experimental data, physical information, and material information was established.

[0042] In one embodiment, tensile tests, hardness tests, and metallographic microstructure characterization were performed on the base metal region, weld region, and heat-affected zone of a low-alloy high-strength marine steel welded joint, respectively, to obtain the Vickers hardness, tensile strength, and average grain size of each region, and the hardness ratio, strength ratio, and size ratio were further calculated. Based on the stress ratio obtained in S1... For experimental data under conditions of 0.1, 0.3, 0.5, 0.7, and 0.9, the fitted stress ratio is... The Paris curve under the given conditions and the inverse solution under the remaining stress ratio conditions were used to determine the crack propagation rate corresponding to the experimental data points. A multi-source fatigue crack extended dataset was constructed, with the training and test sets divided in a 7:3 ratio. The input data was then analyzed. Perform Min-Max normalization.

[0043] S3. Train a neural network prediction model that incorporates physical consistency constraints.

[0044] A neural network architecture is established, and the model is trained, validated, or tuned using the dataset constructed in S2. Specifically, during training, an appropriate optimizer is selected, and suitable hyperparameters such as the learning rate, epochs, and batch size are set. The introduction of physical consistency constraints refers to embedding physical constraint terms into the loss function, in addition to the data-driven terms, to ensure the physical consistency of the model's predictions. Specifically, the physical constraint term uses a modified Zhu-Xuan model, whose expression is: .in, The correction factor was derived based on experimental data. The hardness ratio, strength ratio, and size ratio are weighted by a factor. The weighted average is obtained. The data-driven term and the physical constraint term together constitute the composite loss function, which calculates the deviation between the model's predicted values ​​and the dataset labels and backpropagates to update the model parameters. The model is evaluated using the mean absolute error (MAE), mean squared error (MSE), root mean square error (RMSE), and coefficient of determination (R²). 2 Evaluation metrics are used, and if the metrics do not meet expectations, the hyperparameter sizes are adjusted. When the model accuracy meets the requirements, the trained model is stored as a parameter file, and the parameter information required for input data normalization is saved simultaneously.

[0045] In one embodiment, for a welded joint of a certain low-alloy high-strength marine steel, based on test data, the following was taken: , , This determines the specific expression of the modified Zhu-Xuan model and calculates the physical constraint loss accordingly. This loss is then added to the data-driven MSE loss to form the total loss, and the model is trained on the dataset constructed in S2. The model consists of 45 stacked residual blocks, each containing two fully connected layers of dimension 64, connected by a ReLU activation function. A schematic diagram of the model structure is shown below. Figure 2 As shown in the figure, the trained network has high prediction accuracy, and the prediction performance comparison is shown in the figure below. Figure 3 As shown, on the test set The score reached 0.99, therefore the neural network model is considered successful. Based on the benchmark, arbitrary cross-weld joint area was achieved Equivalent mapping under certain conditions. Heat-affected zone. Model predictions under the given conditions, such as Figure 4 As shown. Save the trained neural network prediction model and normalization information for later use.

[0046] S4. Monitor the operating condition data and loading history of the equipment during its service life.

[0047] During equipment service, load monitoring systems or structural health monitoring systems acquire real-time or periodic operating condition data of the welded joints, which are recorded and saved as loading history. This operating condition data includes, but is not limited to, axial force, bending moment, internal pressure, or combined load information. In one embodiment, based on the collected operating condition data and the recorded loading history, the load history is processed by cycle counting to obtain the equivalent cycle number N and the corresponding maximum stress. Minimum stress Furthermore, by combining structural geometric parameters and a stress analysis model, the stress ratio at critical locations of the welded joint under each loading cycle was calculated. With the maximum value of stress intensity factor Minimum value And thus the range of stress intensity factors is obtained. The calculation results are then organized into equivalent loop counts. and its corresponding and The sequence forms a service data stream that can be used as input to a neural network prediction model. This service data stream is updated as monitored operating conditions change, driving subsequent fatigue crack propagation prediction and digital twin operation.

[0048] S5. Establish a model-based digital twin of fatigue crack propagation.

[0049] The neural network prediction model trained and saved in S3 is deployed in the digital twin system. It receives the service data stream acquired in S4, constructs a digital twin of the fatigue crack propagation in the near-threshold region of the weld joint, and continuously updates the crack propagation state. In one embodiment, during the digital twin initialization phase, the initial crack length of the weld joint is set. The initial crack length can be the actual crack length obtained from non-destructive testing, or it can be the equivalent initial crack length assumed based on damage tolerance design. The digital twin, based on service data streams and equipment material information, invokes a neural network prediction model to predict crack propagation and dynamically update crack propagation evolution. The equipment material information includes the tensile strength, hardness, and average grain size of the material in each region of the welded joint, as well as the tensile strength, hardness, and average grain size of the base material region. Sparse near-threshold region obtained under the condition Experimental data points are used as input to the neural network prediction model. Under physical consistency constraints, the neural network prediction model uses stress ratios... Based on the crack propagation behavior under the given conditions, the current stress ratio is used to achieve the desired crack propagation behavior. Under certain conditions, the mapping prediction of fatigue crack propagation behavior in the base metal zone, weld zone, and near-threshold zone of the heat-affected zone of the welded joint is obtained, thereby obtaining the corresponding... Predicted crack propagation rate under the given conditions Subsequently, based on the predicted crack propagation rate... and the corresponding equivalent number of loops Calculate the crack length increment and through Update the current crack length. In this way, the crack length is gradually evolved over service time and load history, thereby continuously characterizing the fatigue crack propagation state of the welded joint in digital space.

[0050] S6. Conduct damage-tolerant structural safety assessments based on prediction results.

[0051] Based on the crack length evolution results obtained from continuous updates of the digital twin in S5, and combined with damage tolerance design criteria, the safety status and remaining service capability of the welded joint structure are assessed. Specifically, the critical crack length under the corresponding operating condition is determined according to the material fracture toughness, component geometry, and service load characteristics of the structure where the welded joint is located. The critical crack length can be calculated using linear elastic fracture mechanics or elastoplastic fracture mechanics criteria, or a safety factor can be introduced; this invention does not limit this. In one embodiment, the current crack length is updated in real time by the digital twin. With critical crack length By comparison, it can be determined whether the crack propagation state is close to or has reached the structural failure threshold. When the structure is in a healthy state, it is considered to be in a safe operating condition; when Approaching or exceeding When the structure is deemed to be at risk of failure, a corresponding safety warning or maintenance prompt is triggered.

[0052] In another implementation, based on the crack propagation rate prediction obtained from a neural network prediction model and the current crack length of the digital twin, the crack length evolution trend can be calculated. Based on this, the remaining number of cycles or remaining service time required for the crack to propagate from its current length to the critical crack length can be estimated, thereby achieving a quantitative assessment of the remaining fatigue life of the welded joint. The remaining life assessment results can serve as an important basis for determining structural maintenance cycles, adjusting operating loads, and making maintenance decisions.

[0053] Through the above embodiments, this invention achieves intelligent prediction and dynamic updating of crack propagation behavior during service under conditions of non-uniform welded joint materials, multi-regional attributes, and sparse data in the near-threshold region. It constructs a data-model-evaluation integrated technical route, which significantly improves the reliability and practicality of fatigue crack propagation analysis of welded joints in engineering applications.

[0054] The above results demonstrate that this invention, by constructing an intelligent model for predicting fatigue crack propagation behavior in the near-threshold region of welded joints, comprehensively considers the data sparsity and mechanism complexity in the near-threshold region, the performance differences in different areas of the welded joint, and the influence of service loads and loading history on crack propagation behavior. This achieves intelligent prediction and engineering evaluation of fatigue crack propagation behavior in the near-threshold region of different areas of welded joints under small sample conditions. This method has advantages in terms of data requirements, predictive adaptability, and correlation with actual service conditions, and can provide technical support for structural safety assessment and health management of engineering equipment.

[0055] 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 intelligent prediction of fatigue crack propagation in the near-threshold region of a welded joint, characterized in that: This includes S1, collecting fatigue crack propagation data in the near-threshold region of the welded joint, and recording the stress intensity factor range and corresponding crack propagation rate; S2. Construct a physically guided multi-source fatigue crack extension dataset; S3. Train a neural network prediction model that incorporates physical consistency constraints; S4. Monitor the operating condition data and loading history of the equipment during its service life; S5. Establish a model-based digital twin of fatigue crack propagation; S6. Conduct damage-tolerant structural safety assessments based on prediction results.

2. The intelligent prediction method for fatigue crack propagation in the near-threshold region of a welded joint according to claim 1, characterized in that: In step S1, compact tensile specimens with pre-existing cracks are prepared for the base metal region, weld region, and heat-affected zone of the welded joint, and subjected to different stress ratios. Tests were conducted under specific conditions, and after the tests were completed, the range of stress intensity factors in different regions under different working conditions was recorded. Corresponding crack propagation rate .

3. The intelligent prediction method for fatigue crack propagation in the near-threshold region of a welded joint according to claim 2, characterized in that: In S2, based on the data points obtained in S1, and combined with the Paris formula and material parameters of each region, a unified multi-source fatigue crack propagation dataset is established.

4. The intelligent prediction method for fatigue crack propagation in the near-threshold region of a welded joint according to claim 3, characterized in that: The data points are the stress ratios of the base metal region, weld region, and heat-affected zone of the welded joint. Below ; The stress ratio is determined by combining the Paris formula. Crack propagation rate in each set of experimental data Through the Paris formula Using stress ratio Paris parameter = 0.9 and Perform inverse operation Get and corresponding and will obtain As one of the inputs to the dataset; The material-related parameters for each region include the Vickers hardness of the base material region. ,tensile strength and average grain size ; Vickers hardness of weld zone ,tensile strength and average grain size ; Vickers hardness of heat-affected zone ,tensile strength and average grain size ; Calculate the ratios of each parameter relative to the base material parameters in different regions to obtain the hardness ratio. Strength ratio and size ratio Then integrate all the information to As input, As a label, a multi-source information fusion fatigue crack extended dataset containing experimental data, physical information, and material information was established.

5. The intelligent prediction method for fatigue crack propagation in the near-threshold region of a welded joint according to claim 4, characterized in that: In step S3, the model is trained, validated, or its parameters are tuned using the dataset constructed in step S2.

6. The intelligent prediction method for fatigue crack propagation in the near-threshold region of a welded joint according to claim 5, characterized in that: During training, select an optimizer and set the learning rate, epochs, and batch size hyperparameters. Physical constraints are embedded to ensure the physical consistency of the model predictions. The physical model used for these physical constraints is a modified Zhu-Xuan model. ； in, The correction factor was derived based on experimental data. The hardness ratio, strength ratio, and size ratio are weighted by a factor. The weighted average is obtained.

7. The intelligent prediction method for fatigue crack propagation in the near-threshold region of a welded joint according to claim 6, characterized in that: The Zhu-Xuan model was evaluated using mean absolute error (MAE), mean square error (MSE), root mean square error (RMSE), and coefficient of determination (R²). 2 Evaluation metrics are used, and if the metrics do not meet expectations, the hyperparameters are adjusted. When the accuracy of the Zhu-Xuan model meets the requirements, the trained model is stored in the form of a parameter file, and the parameter information required for input data normalization is saved simultaneously.

8. The intelligent prediction method for fatigue crack propagation in the near-threshold region of a welded joint according to claim 7, characterized in that: In step S4, the working condition data of the equipment welding joint is acquired in real time or periodically through the load monitoring system or structural health monitoring system, and recorded and saved as loading history. The operating data includes, but is not limited to, axial force, bending moment, internal pressure, or combined load information.

9. The intelligent prediction method for fatigue crack propagation in the near-threshold region of a welded joint according to claim 8, characterized in that: In step S5, the neural network prediction model trained and saved in step S3 is deployed in the digital twin system, the service data stream obtained in step S4 is received, a digital twin of fatigue crack propagation in the near-threshold region of the welded joint is constructed, and the crack propagation state is continuously updated.

10. The intelligent prediction method for fatigue crack propagation in the near-threshold region of a welded joint according to claim 9, characterized in that: In step S6, based on the crack length evolution results obtained from the continuous updating of the digital twin in step S5, and combined with the damage tolerance design criteria, the safety status and remaining service capability of the welded joint structure are evaluated.

Images