A multi-parameter weld fatigue life prediction method considering the influence of manufacturing process

Abstract

The present invention requests to disclose a multi-parameter weld fatigue life prediction method considering the influence of manufacturing process, aiming to solve the problem of difficulty in predicting weld fatigue life under complex process conditions caused by too many influencing factors in the welding manufacturing process, and determine the most effective welding process combination on this basis. Specifically, the present invention proposes a life prediction method in combination with feature analysis methods and conventional data-driven models. The method uses feature component analysis to reduce the impact of changes in weld fatigue performance caused by different weld process parameters on the model, enhances the generalization ability of the model under complex process conditions, and then uses existing multi-condition data to assist life prediction under target conditions, reducing data requirements; further, the method constructs a hyperparameter optimization index based on K-means clustering, optimizes process characteristics through parameter optimization algorithms, thereby enhancing the optimization effect between complex conditions, while ensuring prediction accuracy.

Description

Technical Field

[0001] The invention belongs to the field of fatigue life prediction, and in particular relates to a weld multi-axis fatigue life prediction method taking into account the influence of welding manufacturing process. Background Art

[0002] With the increase in manufacturing demand and the rapid development of artificial intelligence technology, intelligent manufacturing has emerged widely around the world. By realizing the full interconnection and automation of the production process, intelligent manufacturing has improved production efficiency and product quality, met the needs of personalized customization, and promoted the transformation of the manufacturing industry towards efficiency, precision and intelligence. In the context of intelligent manufacturing, traditional welding technology has achieved innovation and optimization by integrating new intelligent manufacturing methods, significantly improving the accuracy of welding processes, and showing higher application potential and technical advantages in multiple high-tech fields such as marine engineering and aerospace.

[0003] In engineering applications, fatigue failure is one of the common and dangerous forms of failure. As the weak link of the structure, the weld is particularly prone to fatigue failure, so the study of the multi-axial fatigue performance of the weld is of great significance. At the same time, the welding process parameters have a significant effect on the fatigue performance of the weld. However, since the welding process involves many process parameters, their interactions are complex and it is difficult to determine the optimal combination. This brings great challenges to the analysis of fatigue performance and also limits the exploration of the optimal combination of process parameters. In summary, there is an urgent need to establish a method that can be applied to complex process conditions and accurately predict the fatigue life of welds, so as to promote the widespread application and development of new welding manufacturing technologies in engineering.

[0004] Traditional fatigue life prediction models are mostly based on classical theories such as fracture mechanics and continuous damage mechanics, and mainly focus on the relationship between the size, location and other microstructure characteristics of material defects and fatigue life. However, this type of model is difficult to fully consider the impact of changes in welding process parameters on fatigue life, resulting in low prediction accuracy and reliability when working conditions change, limiting its versatility in different process scenarios, and unable to provide effective guidance for the selection of optimal process parameters. Among the existing models that can establish the "process parameter-fatigue life" relationship, the method based on finite element numerical simulation has low computational efficiency and requires a large amount of experimental data for parameter calibration. In addition, models based on conventional data-driven methods also rely on a large amount of data for training, which further increases the difficulty of engineering applications. Summary of the invention

[0005] The purpose of the present invention is to overcome the above shortcomings and provide a multi-parameter weld fatigue life prediction method that takes into account the influence of manufacturing process. Feature analysis is used to reduce the influence of weld fatigue performance changes caused by different process parameters on the model, and the generalization ability of the model under complex process conditions is enhanced. Then, the existing multi-condition data is used to assist the life prediction under the target condition and reduce the data demand. Furthermore, the method constructs a hyperparameter optimization index based on K-means clustering, optimizes the feature analysis through a parameter optimization algorithm, thereby enhancing the coordination effect between complex conditions and ensuring prediction accuracy.

[0006] The purpose of the present invention is achieved through the following technical solution: a multi-parameter weld fatigue life prediction method considering the influence of manufacturing process, the steps comprising:

[0007] S1. Determine the welding manufacturing process parameters and weld fatigue life data that need to be predicted, and record them as the source domain and target domain respectively;

[0008] S2. Based on the target domain parameters, a blind hole method test is carried out with the aid of welding specimens to obtain the residual stress of the material under the target domain data;

[0009] S3. Based on the material sample in step S2, a small amount of fatigue test is carried out, and the loading condition and test life are recorded;

[0010] S4, taking the material working condition parameters and residual stress in step S2 and the loading data in step S3 as model input, and the life data in step S3 as model output, recorded as the target domain including life data;

[0011] S5, taking the working condition parameters, residual stress and loading data to be predicted of the material in step S2 as model input, recorded as target domain data to be predicted;

[0012] S6, fusing the existing process parameter source domain data in step S1 and the target domain life data in step S4 as training data;

[0013] S7, input the model input of the training data in step S6 and the target domain data to be predicted in step S5 into the characteristic component analysis to obtain the mapped training data and the mapped predicted data, only mapping the model input without changing the model output;

[0014] S8, training the neural network model using the training data mapped in step S7;

[0015] S9. Input the predicted data mapped in step S7 into the neural network model trained in step S8 to obtain the output predicted lifespan.

[0016] A further improvement of the present invention is that the specific steps of optimizing the characteristic component analysis in step S7 include:

[0017] S71, performing feature component analysis and mapping on the training data and the target domain data to be predicted to obtain mapped training data and mapped predicted data;

[0018] S72, based on K-means clustering, respectively perform cluster analysis on the mapped training data and the mapped prediction data in step A to obtain a cluster center of the training data and a cluster center of the prediction data;

[0019] S73. Calculate the distance between the training data cluster center and the prediction data cluster center in step B. If it is greater than s, update the hyperparameters in the feature analysis and repeat the above steps. When the distance is less than s, the optimal feature mapping is obtained.

[0020] A further improvement of the present invention is that the specific steps include:

[0021] A. Record the existing weld fatigue life test data under different working conditions as source domain D S ={XS, YS}; where X S is the model input, including welding process parameters, residual stress and loading parameters, Y S is the model output, i.e., the fatigue life under multiaxial loading conditions;

[0022] B. Record the welding manufacturing condition to be predicted as the target domain, and obtain D after the blind hole method test T ={XT};

[0023] C. Conduct a small amount of fatigue test on the target domain condition, denoted as D G ={XG, YG};

[0024] D. S With D G Fusion, get training data D train ={X SG , Y SG}, the data to be predicted in the target domain without fatigue test is recorded as D test ={X TG};

[0025] E. Model input for training data and target domain data to be predicted (X SG and X TG ) Perform characteristic component analysis and mapping;

[0026] F. Optimize the hyperparameters in characteristic component analysis;

[0027] H. training the neural network ANN model using the mapped training data obtained by mapping the optimal feature components in step E;

[0028] I. Input the mapped prediction data obtained by mapping the optimal characteristic components in step E into the ANN model trained in step H to obtain the predicted lifespan.

[0029] A further improvement of the present invention is that the specific steps of feature analysis and mapping in step E include:

[0030] E1. Feature component analysis assumes the existence of feature mapping , making the marginal distributions of the mapped source domain and target domain in the new feature space as similar as possible;

[0031] E2. To solve the feature map , the feature analysis uses the Neural Tangent Kernel (NTK) method to measure the distance between the source domain and the target domain:

[0032] (1)

[0033] Among them, n s , n t , n g are the number of samples in the source domain, target domain, and target domain containing life span data, respectively. , , ω is the weight parameter, ω0 is the arithmetic mean, ▽ is the Jacobian matrix, represents the reproducing kernel Hilbert space norm;

[0034] E3. Use the kernel function matrix K and the coefficient matrix L to simplify the operation and obtain NTK=tr(KL):

[0035] (2)

[0036] (3)

[0037] tr(·) represents the trace of the matrix;

[0038] E4. The feature optimization target is:

[0039] (4)

[0040] in is a regularization parameter used to prevent overfitting;

[0041] E5. Solve equation (4) to get the mapping matrix W, and multiply it with the kernel function matrix W T K, then the model input of the mapped training data and the mapped prediction data ( and ) are W T The first (ns +ng) and last (nt-ng) rows of K.

[0042] A further improvement of the present invention is that: step F performs parameter optimization on the hyperparameter l in the characteristic component analysis, and the specific steps include:

[0043] F1, K-means clustering is used to and Perform clustering and obtain their respective cluster centers c s and c t ;

[0044] F2. Calculate c s and c t Distance ;

[0045] F3. Compare the CD values ​​after feature analysis mapping with different l values, and search for the l value that minimizes the CD value in a given large range through the hyperparameter optimization algorithm.

[0046] A further improvement of the present invention is that in step F3, the hyperparameter optimization algorithm is a relatively stable grid search method, which ultimately determines the optimal welding process parameter combination.

[0047] A further improvement of the present invention is that in step F3, the given larger range is 10 -4 ~100.

[0048] Compared with the prior art, the present invention has the following advantages:

[0049] The present invention aims to solve the problem of difficulty in predicting weld fatigue life under complex process conditions caused by too many influencing factors in the welding manufacturing process, and on this basis determine the most effective welding process combination. Specifically, the present invention proposes a life prediction method by combining feature analysis methods and conventional data-driven models. This method uses feature component analysis to reduce the impact of changes in weld fatigue performance caused by different weld process parameters on the model, enhances the generalization ability of the model under complex process conditions, and then uses existing multi-condition data to assist in life prediction under target conditions, reducing data requirements; further, this method constructs a hyperparameter optimization index based on K-means clustering, optimizes process characteristics through parameter optimization algorithms, thereby enhancing the optimization effect between complex conditions while ensuring prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 The present invention is a schematic flow chart of a multi-parameter weld fatigue life prediction method that takes into account the influence of manufacturing process.

[0051] Figure 2 Schematic diagram of source task and target task in the present invention.

[0052] Figure 3 This is a schematic diagram of optimizing the characteristic component analysis in step S7 of the present invention. DETAILED DESCRIPTION

[0053] In order to deepen the understanding of the present invention, the present invention will be further described in detail below in conjunction with embodiments and drawings. The embodiments are only used to explain the present invention and do not constitute a limitation on the protection scope of the present invention.

[0054] A multi-parameter weld fatigue life prediction method considering the influence of manufacturing process, referring to Figure 1 , the steps include,

[0055] S1. Determine the welding manufacturing process parameters and weld fatigue life data that need to be predicted, and record them as the source domain and target domain respectively;

[0056] S2. Based on the target domain parameters, a blind hole method test is carried out with the aid of welding specimens to obtain the residual stress of the material under the target domain data;

[0057] S3. Based on the material sample in step S2, a small amount of fatigue test is carried out, and the loading condition and test life are recorded;

[0058] S4, taking the material working condition parameters and residual stress in step S2 and the loading data in step S3 as model input, and the life data in step S3 as model output, recorded as the target domain including life data;

[0059] S5, taking the working condition parameters, residual stress and loading data to be predicted of the material in step S2 as model input, recorded as target domain data to be predicted;

[0060] S6, fusing the existing process parameter source domain data in step S1 and the target domain life data in step S4 as training data;

[0061] S7, input the model input of the training data in step S6 and the target domain data to be predicted in step S5 into the characteristic component analysis to obtain the mapped training data and the mapped predicted data, only mapping the model input without changing the model output;

[0062] S8, training the neural network model using the training data mapped in step S7;

[0063] S9. Input the predicted data mapped in step S7 into the neural network model trained in step S8 to obtain the output predicted lifespan.

[0064] Reference Figure 2In the context of intelligent manufacturing, welding manufacturing involves a variety of process parameters (such as welding power, welding speed, welding gun angle, etc.). In order to deeply study the influence of each process parameter on welding quality, it is necessary to conduct repeated welding experiments under a variety of process parameter conditions. The present invention uses feature analysis to enhance the generalization ability of conventional data-driven models, so that it can use the data in the source task to assist in cross-condition life prediction modeling, and greatly reduce the data demand for the target task, reducing the experimental cost.

[0065] In order to enhance the feature analysis effect, the present invention optimizes the feature analysis process in step S7, and the specific steps include:

[0066] S71, performing feature component analysis and mapping on the training data and the target domain data to be predicted to obtain mapped training data and mapped predicted data;

[0067] S72, based on K-means clustering, respectively perform cluster analysis on the mapped training data and the mapped prediction data in step A to obtain a cluster center of the training data and a cluster center of the prediction data;

[0068] S73. Calculate the distance between the training data cluster center and the prediction data cluster center in step B. If it is greater than s, update the hyperparameters in the feature map and repeat the above steps. When the distance is less than s, the optimal feature map is obtained.

[0069] The technical solution of the present invention will be further described below in conjunction with a specific embodiment. A multi-parameter weld fatigue life prediction method considering the influence of manufacturing process, the specific steps include:

[0070] A. Record the existing weld fatigue life test data under different working conditions as source domain D S ={XS, YS}; where X S is the model input, including welding process parameters, residual stress and loading parameters, Y S is the model output, i.e., the fatigue life under multiaxial loading conditions;

[0071] B. Record the welding manufacturing condition to be predicted as the target domain, and obtain D after the blind hole method test T ={XT};

[0072] C. Conduct a small amount of fatigue test on the target domain working condition, denoted as D G ={XG, YG};

[0073] D. S With D G Fusion, get training data D train ={X SG , Y SG}, the data to be predicted in the target domain without fatigue test is recorded as Dtest ={X TG};

[0074] E. Model input for training data and target domain data to be predicted (X SG and X TG ) to perform characteristic component analysis and feature analysis mapping:

[0075] E1. Feature analysis assumes that there is a feature map , making the marginal distributions of the mapped source domain and target domain in the new feature space as similar as possible;

[0076] E2. To solve the feature map , the feature analysis uses the Neural Tangent Kernel (NTK) method to measure the distance between the source domain and the target domain:

[0077] (1)

[0078] Among them, n s , n t , n g are the number of samples in the source domain, target domain, and target domain containing life span data, respectively. , , ω is the weight parameter, ω0 is the arithmetic mean, ▽ is the Jacobian matrix, represents the reproducing kernel Hilbert space norm;

[0079] Description: Because the neural tangent kernel (NTK) method has obvious advantages in handling non-local operations, simplifying optimization problems, and theoretical convergence and generalization. It is able to handle complex data structures, such as multi-parameter data, and appears under broader standards without the need for infinite width limit conditions. In addition, the NTK method provides information about network weights during network training, especially when the network parameters are close to the initial state, which is crucial for understanding how the network evolves from a randomly initialized state to an optimized state. In contrast, related methods such as the maximum mean difference (MMD) method are mainly suitable for scenarios such as domain adaptation and generative adversarial networks, and are not as good as the NTK method when dealing with non-local or complex data structures. Therefore, the NTK method is recommended in the process of multi-parameter weld fatigue life prediction methods considering the influence of manufacturing processes.

[0080] E3. Use the kernel function matrix K and the coefficient matrix L to simplify the operation and obtain NTK=tr(KL):

[0081] (2)

[0082] (3)

[0083] tr(·) represents the trace of the matrix;

[0084] E4. The feature optimization target is:

[0085] (4)

[0086] in is a regularization parameter used to prevent overfitting;

[0087] E5. Solve equation (4) to get the mapping matrix W, and multiply it with the kernel function matrix W T K, then the model input of the mapped training data and the mapped prediction data ( and ) are W T the front (ns +ng) and back (nt-ng) rows of K;

[0088] F. Optimize the hyperparameter l in feature analysis. The specific steps include:

[0089] F1, K-means clustering is used to and Perform clustering and obtain their respective cluster centers c s and c t ;

[0090] F2. Calculate c s and c t Distance ;

[0091] F3, compare settings The CD value after feature analysis of the value is mapped, and the hyperparameter optimization algorithm is a grid search method. In a given large range (such as 10 -4 ~100) Search for the l value that minimizes the CD value;

[0092] H. training the neural network ANN model using the mapped training data obtained by the optimal feature analysis mapping in step E;

[0093] I. Input the mapped prediction data obtained by the optimal feature analysis mapping in step E into the ANN model trained in step H to obtain the predicted lifespan.

[0094] The present invention uses several sets of weld multiaxial fatigue data under different welding processes as the source domain, and uses the weld multiaxial fatigue data under the specific welding process to be predicted as the target domain. The number of samples containing life data in the target domain is n gA small number of samples are set, randomly selected, and the average is taken after multiple repetitions. The predicted value is compared with the test result, and the prediction accuracy of the model is evaluated using the coefficient of determination R². The results show that the model proposed in this invention has a significant improvement in accuracy compared to the direct use of the ANN model, showing good generalization ability.

[0095] Those skilled in the art should understand that the present invention is not limited to the above embodiments, and the above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, and these changes and improvements fall within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A multi-parameter weld fatigue life prediction method considering the influence of manufacturing process, characterized by: The steps include, S1. Determine the welding manufacturing process parameters and weld fatigue life data that need to be predicted, and record them as the source domain and target domain respectively; S2. Based on the target domain parameters, with the help of welding specimens, the blind hole method is carried out to obtain the residual stress of the material under the target domain data; S3. Based on the material sample in step S2, a small amount of fatigue test is carried out, and the loading condition and test life are recorded; S4, taking the material working condition parameters and residual stress in step S2 and the loading data in step S3 as model input, and the life data in step S3 as model output, recorded as the target domain including life data; S5, taking the working condition parameters, residual stress and loading data to be predicted of the material in step S2 as model input, recorded as target domain data to be predicted; S6, fusing the existing process parameter source domain data in step S1 and the target domain life data in step S4 as training data; S7, input the model input of the training data in step S6 and the target domain data to be predicted in step S5 into the characteristic component analysis to obtain the mapped training data and the mapped predicted data, only mapping the model input without changing the model output; S8, training the neural network model using the training data mapped in step S7; S9. Input the predicted data mapped in step S7 into the neural network model trained in step S8 to obtain the output predicted lifespan.

2. According to claim 1, a multi-parameter weld fatigue life prediction method considering the influence of manufacturing process is characterized in that: The specific steps of the characteristic component analysis optimization in step S7 include: S71, performing feature component mapping on the training data and the target domain data to be predicted to obtain mapped training data and mapped predicted data; S72, based on K-means clustering, respectively perform cluster analysis on the mapped training data and the mapped prediction data in step S71 to obtain a cluster center of the training data and a cluster center of the prediction data; S73. Calculate the distance between the training data cluster center and the prediction data cluster center in step S72. If it is greater than s, update the hyperparameters in the feature map and repeat the above steps. When the distance is less than s, the optimal feature map is obtained.

3. According to claim 2, a multi-parameter weld fatigue life prediction method considering the influence of manufacturing process is characterized in that: The specific steps include: A. Record the existing weld fatigue life test data under different working conditions as source domain D S ={X S , Y S }; where X S is the model input, including welding process parameters, residual stress and loading parameters, Y S is the model output, i.e., the fatigue life under multiaxial loading conditions; B. Record the welding manufacturing condition to be predicted as the target domain, and obtain D after the blind hole method test T ={XT}; C. Conduct a small amount of fatigue test on the target domain condition, denoted as D G ={XG, YG}; D. S With D G Fusion, get training data D train ={X SG , Y SG }, the data to be predicted in the target domain without fatigue test is recorded as D test ={X TG }; E. Model input X for training data and target domain data to be predicted SG and X TG Perform feature analysis mapping; F. Optimize the hyperparameters in feature analysis; H. training the neural network ANN model using the mapped training data obtained by the optimal feature analysis mapping in step E; I. Input the mapped prediction data obtained by the optimal feature analysis mapping in step E into the ANN model trained in step H to obtain the predicted lifespan.

4. According to claim 3, a multi-parameter weld fatigue life prediction method considering the influence of manufacturing process is characterized in that: The specific steps of feature analysis and mapping in step E include: E1. Feature analysis assumes that there exists a feature mapping φ, which makes the marginal distributions of the mapped source domain and target domain in the new feature space as similar as possible; E2. To solve the feature map φ, the feature analysis uses the neural tangent kernel method NTK to measure the distance between the source domain and the target domain: Among them, n s ,n t ,n g are the number of samples in the source domain, target domain, and target domain containing life span data, respectively. ω is the weight parameter, ω0 is the arithmetic mean, ▽ is the Jacobian matrix, and ||·|| represents the reproducing kernel Hilbert space norm; E3. Use the kernel function matrix K and the coefficient matrix L to simplify the operation and obtain NTK=tr(KL): tr(·) represents the trace of the matrix; E4. The feature optimization target is: s.t. W T KHKW=I m (4) Where λ is a regularization parameter used to prevent overfitting; E5. Solve equation (4) to get the mapping matrix W, and multiply it with the kernel function matrix W T K, then the model input of the mapped training data and the mapped prediction data and W T The first n of K s +n g Row and after n t -n g OK.

5. A multi-parameter weld fatigue life prediction method considering the influence of manufacturing process according to claim 4, characterized in that: The step F performs parameter optimization on the hyperparameter λ in the characteristic component analysis, and the specific steps include: F1, K-means clustering is used to and Perform clustering and obtain their respective cluster centers c s and c t ; F2. Calculate c s and c t The distance CD=||c s -c t || 2 ; F3. Compare the CD values ​​after feature analysis mapping with different λ values, and search for the λ value that minimizes the CD value in a given range through a hyperparameter optimization algorithm.

6. A multi-parameter weld fatigue life prediction method considering the influence of manufacturing process according to claim 5, characterized in that: In step F3, the hyperparameter optimization algorithm is a stable grid search method, which ultimately determines the optimal welding process parameter combination.

7. A multi-parameter weld fatigue life prediction method considering the influence of manufacturing process according to claim 6, characterized in that: In step F3, the given range is 10 -4 ~100.

Images