Deep sea riser fatigue life prediction method based on physical information driving

Abstract

The invention provides a deep sea riser fatigue life prediction method based on physical information driving. The deep sea riser fatigue life prediction method comprises the following steps: 1) collecting original data of a riser fatigue test; 2) calculating physical information parameters of the riser; 3) taking material performance, vertical pipe geometric parameters, crack geometric parameters, stress strain, load parameters and crack fatigue physical information parameters as model input, taking a crack growth rate as model output, and performing normalization and dimensionless processing to form a deep sea vertical pipe characteristic data set; 4) performing data segmentation and recombination based on the time sequence characteristics of the data, and building a long-short memory deep learning model containing physical constraints based on the characteristics of crack propagation; 5) obtaining an optimal physical information driven deep learning model; and 6) performing fatigue life prediction through the optimal physical information driving model, and performing uncertainty evaluation. According to the method, probability prediction and uncertainty quantification of the fatigue life of the riser can be realized, and a reference can be provided for reliability evaluation of an in-service riser structure.

Description

Technical Field

[0001] The present invention relates to the field of marine engineering technology, and in particular to a physical information driven deep-sea riser fatigue life prediction method. Background Art

[0002] The deep sea is an important way to transport oil and gas in deepwater oil and gas projects. Due to the complexity of the deep sea environment, the riser is affected by cyclic stress and suffers fatigue damage, especially the girth weld of the pipeline is very prone to cracking. The common fatigue life prediction methods currently include experimental method, numerical simulation method and data-driven method. The experimental method includes standard part experiment and full-scale experiment, and its core is to fit the fatigue damage model based on physical quantities such as strain, stress or energy. The numerical simulation method focuses on simulating the physical state of the pipeline through numerical methods and boundary conditions, and combines the SN curve to predict fatigue life. Although the above methods have been successfully applied to the research related to riser fatigue, they still have some defects and shortcomings, mainly including: (1) Theoretical methods require a large number of physical assumptions, which greatly reduces their accuracy in reflecting actual physical phenomena. (2) The numerical simulation method cannot fully reflect the real changes of the model due to the limitations of the model itself. In practical applications, the interaction between factors such as size, material, manufacturing process and operating environment makes structural safety uncertain.

[0003] Unlike experimental and numerical methods that focus on causal reasoning, data-driven methods focus on revealing correlations within the data. However, it should be noted that since data-driven models are essentially black-box systems, their predictive performance is heavily dependent on the quantity and quality of available data. The high cost of fatigue testing and the scarcity of data points, coupled with finite element data that has not yet achieved high-fidelity accuracy, pose challenges to training powerful data-driven models. Chinese patent CN115081321B discloses a corrosion fatigue life prediction method for marine welded structures, which uses edge oversampling to perform data enhancement, and then combines extreme gradient boosting algorithms and deep convolutional neural networks to achieve life prediction of marine welded structures. However, due to the lack of description of cracks in the input features of the data, and the lack of consideration of the impact of parameter uncertainty, there are significant limitations.

[0004] In order to solve the problems existing in the data-driven method, it is necessary to introduce physical information and physical constraints into the model to improve the accuracy and interpretability of the model and reduce the model's dependence on high-cost data. Chinese patent CN117993304B discloses a physical information-driven machine learning prediction method for notch fatigue life of metal materials. This method adds the geometric parameters and physical information parameters of the notch to the input features of the data, and effectively builds a fatigue life machine learning model with strong generalization ability. However, all the selected data come from standard parts experiments, and there is a certain deviation between standard parts experiments and full-scale experiments, especially under high stress conditions. Therefore, this model has certain limitations.

[0005] Machine learning has been widely used in the field of deep-sea riser fatigue life prediction, but there are few reports on the fatigue life prediction of girth welds under complex stress states. On the other hand, a unified life model for multiple sizes and a wide life range has not yet been established. The machine learning model driven by physical information provides a new idea for building a unified life prediction model. Summary of the invention

[0006] In view of the technical problems existing in the above-mentioned background technology, the present invention proposes a deep-sea riser fatigue life prediction method driven by physical information. The method has a reasonable conception and can evaluate the crack growth rate of the girth weld under multi-size and multi-stress conditions, and can realize the probabilistic prediction and uncertainty quantification of the riser fatigue life, thus providing a reference for the reliability assessment of the in-service riser structure.

[0007] In order to solve the above technical problems, the present invention provides a deep-sea riser fatigue life prediction method based on physical information driving, which mainly includes the following steps:

[0008] 1) Collect the original data of the riser fatigue test, including material properties, riser geometry parameters, crack geometry parameters, stress strain, load parameters and crack growth rate;

[0009] 2) Calculate the physical information parameters of the riser

[0010] The maximum circumferential stress and generalized strain at the riser crack are obtained by analytical method, and the damage parameters at the riser crack are further obtained.

[0011] 3) taking five types of parameters, namely, material properties, riser geometry parameters, crack geometry parameters, stress-strain, load parameters, and crack fatigue physical information parameters, as model inputs, and the crack growth rate as model output, the original data of the riser fatigue test in the above step 1) and the physical information parameters of the riser in the above step 2) are normalized and dimensionless respectively to form a deep-sea riser characteristic data set;

[0012] 4) Data segmentation and reorganization are performed based on the temporal characteristics of the riser fatigue test data. The riser fatigue test data set is divided into a training set and a test set in a ratio of 7:3. A long-short memory deep learning model with physical constraints is built based on the characteristics of crack propagation.

[0013] 5) Use the training set in the riser fatigue test data set to train the long-short memory deep learning model, obtain the optimal parameters of the long-short memory deep learning model through the Bayesian optimization method, and compare different models to obtain the optimal physical information driven deep learning model;

[0014] 6) Use the optimal physical information to drive the deep learning model for fatigue life prediction and uncertainty assessment.

[0015] The deep-sea riser fatigue life prediction method based on physical information drive, wherein: the material properties in step 1) include elastic modulus E, Poisson's ratio v, yield strength limit σ limit The geometric parameters include the riser geometric parameters and the crack geometric parameters. The riser geometric parameters include the inner diameter d, the outer diameter D, the wall thickness t and the tube length L. The crack geometric parameters include the crack depth a and the crack width c. The stress and strain include the maximum principal stress σ max , hydrostatic pressure σ m 、Von Mises stress σ eq , strain range Δε t and stress intensity factor K, load parameters include stress ratio R σ and frequency f.

[0016] The deep-sea riser fatigue life prediction method based on physical information driving, wherein, in the step 2), assuming that the crack of the riser extends radially, the circumferential stress σ at the riser crack is rmax Expressed by the stress intensity factor:

[0017]

[0018] In the above formula (1), K is the force intensity factor, r is the crack length;

[0019] Generalized strain ε at the riser crack f Using the material's stress triaxiality and critical strain ε o To express:

[0020]

[0021] In the above formula (2), is the hydrostatic pressure; σ eq is the Von Mises stress; σ1, σ2 and σ3 are the three components of the principal stress, ε ois the critical strain for the initial crack of the material.

[0022] The deep-sea riser fatigue life prediction method based on physical information drive, wherein in the step 2), the damage parameter at the riser crack is obtained by the following formula (3):

[0023] SWT=σ rmax Δε t / twenty three);

[0024] In the above formula (3), Δε t is the total strain range around the girth weld.

[0025] The deep-sea riser fatigue life prediction method based on physical information drive, wherein the material properties in step 3) are dimensionless and normalized using the material properties parameters of the material at room temperature:

[0026]

[0027] In the above formula (4), M is the dimensionless and normalized performance parameter of the material; M d is the dimensionless and normalized performance parameter of the material; M s is the material performance parameter at room temperature;

[0028] The riser geometric parameters in step 3) are normalized using the dimensions of a standard pipeline, and the crack geometric parameters are normalized using a standard threshold:

[0029]

[0030] In the above formula (5), G and G d is the dimensionless and normalized geometric parameters. G includes the riser geometric parameters and crack geometric parameters. th Represents the dimensionless and normalized threshold value, which is selected based on the size of the standard pipe and the maximum value measured in the crack fatigue test;

[0031] The stress and strain parameters in step 3) are normalized using standard thresholds:

[0032]

[0033] In the above formula (6), φ is the dimensionless and normalized stress-strain parameter; φ d is the dimensionless and normalized stress-strain parameter; φ th represents the dimensionless and normalized threshold value, which is selected based on the maximum value measured in the crack fatigue test;

[0034] The load parameters in step 3) are normalized using the following formula (7):

[0035]

[0036] In the above formula (7), R σ represents the stress ratio, represents the ratio of dimensionless to normalized stress, f represents the load frequency, and f d represents the dimensionless and normalized frequency, f th Threshold values ​​specified in the riser fatigue test standards;

[0037] The physical information parameters in step (3) are normalized using the following formula (8):

[0038]

[0039] In the above formula (8), σ rmax represents the maximum hoop stress, represents the dimensionless and normalized maximum hoop stress, ε f represents the generalized strain, ε d represents the dimensionless and normalized strain, SWT represents the damage parameter at the crack, SWT d represents the dimensionless and normalized damage parameter; σ th , ε th and SWT th Select the maximum value among the results calculated by analytical method;

[0040] The crack growth rate in step 3) is normalized using the following formula (9):

[0041]

[0042] In the above formula (9), CGR represents the crack growth rate, CGR d represents the dimensionless and normalized crack growth rate, CGR th The selection is based on the maximum value of the experimental data.

[0043] The deep-sea riser fatigue life prediction method based on physical information drive, wherein in the step 4), data segmentation and reorganization are performed based on the temporal characteristics of the riser fatigue test data, and each subset is first divided into a training set and a test set using a stratified sampling method; then physical constraints are introduced into the loss function of the long short-term memory deep learning model to improve the interpretability of the model, and the loss function L(θ) consists of two parts, the first part is the deviation between the predicted value and the true value, and the second part is the deviation of the predicted result under the physical constraint, which can be specifically expressed as:

[0044] L(θ)=L d (θ)+λL p(θ) (10);

[0045] In the above formula (10), L d (θ) represents the deviation between the predicted value and the true value; L p (θ) represents the deviation between the predicted value and the physical law; λ is the penalty factor, which is used to balance the impact of the two parts on the loss function; L d (θ) is calculated using RMSE, L p (θ) is determined by calculating the derivative of the crack growth rate;

[0046] The crack growth process is divided into three stages: initiation stage, stable stage and rapid growth stage; p The expression of (θ) is:

[0047]

[0048] In the above formula (11), N represents the number of cycles, a represents the crack depth, and I(A>B) is a state indicator function. When A is greater than B, its value is 1, indicating that the physical constraint is violated, otherwise it is equal to 0, indicating that it is consistent with the physical constraint.

[0049] The deep-sea riser fatigue life prediction method based on physical information drive, wherein the specific process of step 5) is as follows:

[0050] 5.1) Select 6 cross validation R 2 The average As the objective function for optimization, determine the number of layers N of the long short-term memory deep learning model layer , the number of units per layer N units and the optimization range of the learning rate η, combined with Bayesian optimization and K-fold cross validation to obtain the optimal parameters of the long short-term memory deep learning model. The objective function selected in the optimization process is K-fold cross validation R 2 The average

[0051]

[0052] In the above formula (12), R i 2 is the coefficient of determination of the i-th cross validation; is the average value of K cross validations.

[0053] 5.2) Select the three most critical hyperparameters, namely the number of layers N, from the long short-term memory deep learning model layer , the number of units per layer N units and learning rate η, and use the Bayesian optimization method to obtain the final hyperparameter optimization results;

[0054] 5.3) Compare the prediction results of the optimal physical information driven deep learning model with those of the traditional model to obtain the optimal model.

[0055] The method for predicting fatigue life of deep-sea riser based on physical information driving, wherein the specific process of step 6) is: using the trained physical information to drive the deep learning model to predict fatigue life, comparing the prediction results with the deep learning model without introducing physical parameters and the traditional physical fatigue life model, and quantitatively evaluating the life uncertainty.

[0056] The deep-sea riser fatigue life prediction method based on physical information drive, wherein the specific steps of quantitatively evaluating the life uncertainty are as follows:

[0057] 6.1) The crack depth a, crack width c, and maximum principal stress σ of the deep-sea riser feature data set were collected by Latin hypercube sampling. max , strain range Δε t Sampling is performed to obtain a sufficient number of samples; the sampling range is limited to the first 80% of the deep-sea riser characteristic data set; it is assumed that the data of the first 80% of the deep-sea riser characteristic data set obeys a normal distribution, in which the experimental measurement value represents the mean value and the standard deviation is determined by the measurement error;

[0058] 6.2) Segment the first 80% of the data samples in the deep-sea riser feature data set to meet the input form of the physical information driven deep learning model, and then use the trained physical information driven deep learning model to make predictions to obtain the corresponding crack growth data;

[0059] 6.3) Calculate the riser crack depth based on the crack extension data. The calculation formula for the crack depth is as follows:

[0060]

[0061] In the above formula (13), a i represents the crack depth at time i, a i+1 represents the crack depth at time i+1, represents the crack growth rate at time i, and dN represents the number of cycles;

[0062] 6.4) Using the failure crack depth measured in the test as the criterion, calculate the number of cycles N when this crack depth is reached in the above formula (13), and this N is the fatigue life;

[0063] 6.5) Plot the PSN curve based on the fatigue life results to achieve quantitative assessment of uncertainty.

[0064] By adopting the above technical solution, the present invention has the following beneficial effects:

[0065] The deep-sea riser fatigue life prediction method based on physical information driven by the present invention is well conceived. The introduction of physical information and physical constraints improves the accuracy and interpretability of the model, and solves the problems of poor interpretability and poor robustness of pure data-driven deep learning models. The well-trained physical information-driven deep learning model can evaluate the crack growth rate of the girth weld under multi-size and multi-stress conditions, and has good extrapolation ability and robustness; by combining LHS with the physical information-driven deep learning model, the probability prediction and uncertainty quantification of the riser fatigue life can be achieved, which can provide a reference for the reliability assessment of the in-service riser structure. In addition to the riser structure, the present invention can also be extended to all crack growth problems involving time series data, providing a new idea for the safety prediction of deep-sea structures. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings in the following description are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0067] Figure 1 It is a flow chart of the deep-sea riser fatigue life prediction method driven by physical information according to the present invention;

[0068] Figure 2 This is a prediction result diagram of the data-driven model used in the deep-sea riser fatigue life prediction method based on physical information drive of the present invention;

[0069] Figure 3 This is a prediction result diagram of the physical information driven model used in the deep-sea riser fatigue life prediction method based on physical information driven according to the present invention;

[0070] Figure 4 It is a PSN curve diagram of an SCR riser drawn according to the results of a physical information driven model in the deep-sea riser fatigue life prediction method based on physical information driven according to the present invention;

[0071] Figure 5 It is an uncertainty quantification diagram drawn according to the results of the physical information driven model in the deep-sea riser fatigue life prediction method based on physical information driven according to the present invention;

[0072] Figure 6 It is an SN curve diagram drawn according to the physical information driven model results and industry standards in the deep-sea riser fatigue life prediction method based on physical information driven of the present invention. DETAILED DESCRIPTION

[0073] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0074] The present invention is further explained below in conjunction with specific implementation modes.

[0075] like Figure 1 As shown, this embodiment provides a deep-sea riser fatigue life prediction method based on physical information drive, taking the API X65 SCR fatigue life test as a case to evaluate the SCR (crack growth rate) fatigue life, including the following steps:

[0076] S100, collect the original data of fatigue test of full-size, small-size and standard parts of SCR risers, including material properties, riser geometry parameters, crack geometry parameters, stress strain, load parameters and crack growth rate; among them, the original data are collected from the relevant experiments of the "Deep Sea No. 1" sub-project "Deepwater Riser Test and Development System Construction" and domestic and foreign publicly published literature and material manuals. The aforementioned material properties include elastic modulus E, Poisson's ratio v, yield strength limit σ limit The geometric parameters include the riser geometric parameters and the crack geometric parameters. The riser geometric parameters include the inner diameter d, the outer diameter D, the wall thickness t and the tube length L. The crack geometric parameters include the crack depth a and the crack width c. The stress and strain include the maximum principal stress σ max , hydrostatic pressure σ m 、VonMises stress σ eq , strain range Δε t and stress intensity factor K, load parameters include stress ratio R σ and frequency f.

[0077] S200, calculate the physical information parameters of the SCR riser

[0078] The maximum circumferential stress and generalized strain at the crack are obtained by analytical method, and the damage parameters at the crack are further obtained. Assuming that the crack of the riser expands radially, the circumferential stress σ at the crack of the riser is rmax Expressed by the stress intensity factor:

[0079]

[0080] In the above formula (1), K is the force intensity factor, r is the crack length;

[0081] Generalized strain ε at the riser crack f Using the material's stress triaxiality and critical strain ε o To express:

[0082]

[0083] In the above formula (2), is the hydrostatic pressure, σ eq is the Von Mises stress, σ1, σ2 and σ3 are the principal stresses in three directions respectively, ε o is the critical strain for the initial crack of the material.

[0084] The damage parameter SWT at the riser crack is obtained by the following formula (3):

[0085] SWT=σ rmax Δε t / twenty three);

[0086] In the above formula (3), Δε t is the total strain range around the girth weld.

[0087] S300, taking five types of parameters, namely, material properties, riser geometry parameters, crack geometry parameters, stress-strain, load parameters and crack fatigue physical information parameters, as model inputs, and taking crack propagation rate as model outputs, normalizing and dimensionlessly processing the original data of the riser fatigue test in the above step S100 and the physical information parameters of the riser in the above step S200, respectively, to form a deep-sea riser characteristic data set.

[0088] The material properties mentioned above are dimensionless and normalized using the material properties parameters of the SCR material at room temperature, as shown in the following formula (4):

[0089]

[0090] In the above formula (4), M and M d are the dimensionless and normalized properties of the material, M s are the material properties at room temperature.

[0091] The above-mentioned riser geometric parameters are normalized using the dimensions of the SCR standard pipe, and the crack geometric parameters are normalized using the standard threshold value obtained from the SCR experiment, as shown in the following formula (5):

[0092]

[0093] In the above formula (5), G and G d are the geometric parameters before and after normalization; G includes the riser geometric parameters and crack geometric parameters; G th It represents the dimensionless and normalized threshold value, which is selected based on the size of the standard pipe and the maximum value measured in the crack fatigue test.

[0094] The above stress-strain parameters are normalized using the standard threshold value obtained from the SCR experiment, as shown in the following formula (6):

[0095]

[0096] In the above formula (6), φ is the dimensionless and normalized stress-strain parameter; φ d Dimensionless and normalized stress-strain parameters; φ th Represents the threshold value used for dimensionless and normalization, which is selected based on the maximum value measured in the crack fatigue test.

[0097] The above load parameters are normalized using the following formula (7):

[0098]

[0099] In the above formula (7), R σ represents the stress ratio, represents the ratio of dimensionless to normalized stress, f represents the load frequency, and f d represents the dimensionless and normalized frequency, f th Threshold value specified in the riser fatigue test standard.

[0100] The above physical information parameters are normalized using the method shown in the following formula (8):

[0101]

[0102] In the above formula (8), σ rmax represents the maximum hoop stress, represents the dimensionless and normalized maximum hoop stress, ε f represents the generalized strain, ε d represents the dimensionless and normalized strain, SWT represents the damage parameter at the crack, SWT d represents the dimensionless and normalized damage parameter; σ th , ε th and SWT th The maximum value among the analytical calculation results is selected.

[0103] The above crack growth rate is normalized using the method shown in the following formula (9):

[0104]

[0105] In the above formula (9), CGR represents the crack growth rate, CGR d represents the dimensionless and normalized crack growth rate, CGR thThe selection is based on the maximum value of the experimental data.

[0106] The specific input and output of the data set are shown in Table 1.

[0107] Table 1: Dataset input features

[0108]

[0109]

[0110] S400, based on the time series characteristics of the riser fatigue test data, data segmentation and reorganization are performed. The riser fatigue test data set is divided into a training set and a test set in a ratio of 7:3. Based on the characteristics of crack propagation, a long-short memory deep learning model with physical constraints is built.

[0111] The data in this embodiment includes crack extension data under different stress levels. The crack data distribution under high, medium and low stress levels are 10%, 30% and 60% respectively. The data subsets are divided according to different stress levels. Based on the principle of 20 sample points per group, the 60,000 data points are re-divided. The division process ensures that the sample points are from the same data subset and the order remains unchanged. Finally, 300 sets of new data are obtained.

[0112] The above-mentioned data segmentation and reorganization based on the time series characteristics of the riser fatigue test data adopts a stratified sampling method to perform stratified sampling on each subset according to the ratio of training:test=7:3 to obtain the training set and test set for the deep learning model to train and test.

[0113] Then, physical constraints are introduced into the loss function of the long short-term memory deep learning model to improve the interpretability of the model. The loss function L(θ) consists of two parts. The first part is the deviation between the predicted value and the true value, and the second part is the deviation of the predicted result under the physical constraint, which can be specifically expressed as:

[0114] L(θ)=L d (θ)+λL p (θ) (10);

[0115] In the above formula (10), L d (θ) represents the deviation between the predicted value and the true value, L p (θ) represents the deviation between the predicted value and the physical law, and λ is the penalty factor, which is used to balance the impact of the two parts on the loss function; L d (θ) is calculated using RMSE, L p (θ) is determined by calculating the derivative of the crack growth rate;

[0116] The crack growth process is divided into three stages: initiation stage, stable stage and rapid growth stage; the loss term expression is:

[0117]

[0118] In the above formula (11), N represents the number of cycles, a represents the crack depth, and I(A>B) is a state indicator function. When A is greater than B, its value is 1, indicating that the physical constraint is violated, otherwise it is equal to 0, indicating that it is consistent with the physical constraint.

[0119] S500, using the training set in the riser fatigue test data set to train the long short memory deep learning model, using the Bayesian optimization method to obtain the optimal parameters of the long short memory deep learning model, and comparing different models to obtain the optimal physical information driven deep learning model; the specific process is:

[0120] S501, select 6 cross validation R 2 The average As the objective function of optimization, determine the number of layers N of the long short-term memory deep learning model LSTM layer , the number of units per layer N units and the optimization range of the learning rate η, combined with Bayesian optimization and K-fold cross validation to obtain the optimal parameters of the long short-term memory deep learning model. The objective function selected in the optimization process is N cross validation R 2 The average

[0121] In the above formula (12), R i 2 is the coefficient of determination of the i-th cross validation; is the average value of K cross validations.

[0122] S502, select three most critical hyperparameters from the long short-term memory deep learning model LSTM for optimization. They are the number of layers N of the long short-term memory model LSTM layer , the number of units per layer N units and learning rate η, and the Bayesian optimization method is used to obtain the final hyperparameter optimization results; Table 2 lists the optimization ranges of these hyperparameters in detail.

[0123] Table 2 Hyperparameter optimization range

[0124]

[0125] The final hyperparameter optimization results are shown in Table 3. It should be noted that this result is the result after the hyperparameter optimization converges. The maximum value indicates that the prediction ability of the model is the most stable under this hyperparameter combination. The number of optimizations needs to be determined according to the size of the database. In this embodiment, 1000 times are selected.

[0126] Table 3 Hyperparameter optimization results

[0127]

[0128]

[0129] S503. Compare the prediction results of the optimal physical information driven deep learning model with those of the traditional model to obtain the optimal model.

[0130] This example compares the performance of a data-driven model without physical information and physical constraints and a physical information-driven model. For the data-driven model, this example also performs hyperparameter optimization. The prediction results of the two models are shown in Figure 2 , Figure 3 It is not difficult to see that for the LSTM model, only 75% of the predictions are within the ±1.5 error range and 63% of the predictions are within the ±2 error range. Figure 2 Some outliers with poor predictions are shown marked with circles. This indicates potential improvements in prediction stability and generalization using LSTM. Figure 3 Detailed prediction results using the PI-LSTM model are provided. More than 95% of the PI-LSTM model predictions are within the ±1.5 error range, and all predictions are within the ±2 error range. Compared with the LSTM model, the prediction accuracy of the PI-LSTM model is significantly improved, effectively alleviating Figure 2 The results show that incorporating physical information as prior knowledge significantly improves the accuracy and stability of the LSTM model.

[0131] S600, using the trained physical information to drive the deep learning model to predict fatigue life, comparing the prediction results with those of the deep learning model without introducing physical parameters and the traditional physical fatigue life model, and quantitatively evaluating the life uncertainty; specifically comprising the following steps:

[0132] S601, Latin hypercube sampling (LHS) is used to collect crack depth a, crack width c, maximum principal stress σ of deep-sea riser characteristic data. max , strain range Δε t Sampling is performed to obtain enough samples. The sampling range is limited to the first 80% of the deep-sea riser characteristic data set, where the data basically conforms to the Paris theorem. It is assumed that the first 80% of the deep-sea riser characteristic data set follows a normal distribution, where the experimental measurement value represents the average value, the standard deviation of the crack depth is 0.05 mm, the standard deviation of the crack width is 0.1 mm, and the maximum principal stress σmax The standard deviation is 0.5MPa, and the strain range is Δε t The standard deviation is 0.1 and the total number of samples is 500.

[0133] S602, the first 80% of the data samples in the deep-sea riser characteristic data set are divided into groups of 20 sample points to meet the input form of the physical information driven deep learning model, and then the trained physical information driven deep learning model is used for prediction to obtain the corresponding crack propagation data.

[0134] S603. Calculate the riser crack depth based on the crack extension data. The calculation formula for the crack depth is as follows:

[0135]

[0136] In the above formula (13), a i represents the crack depth at time i, a i+1 represents the crack depth at time i+1, represents the crack growth rate at time i, and dN represents the number of cycles;

[0137] S604. Using the failure crack depth measured in the test as a criterion, calculate the number of cycles N when the crack depth is reached in the above formula (13), and this N is the fatigue life.

[0138] S605. Draw a PSN curve based on the fatigue life results to achieve quantitative evaluation of uncertainty. Figure 4 The PSN curve of the riser is given. Figure 4 It shows that there is a strong correlation between the experimental life and the predicted fatigue life at a survival probability of 50%. In addition, the experimental life is between 99% and 1% survival probability, and the variation of the fatigue life of the girth weld is mainly attributed to the geometric inconsistency of the weld. Figure 5 It shows that the uncertainty range of the predicted life is within the error range of ±1.5, indicating that the uncertainty quantification results of the PI-LSTM model are in good agreement with the experimental life data. In order to further verify the physical consistency of the PI-LSTM model, Figure 6 The SN curve of the pipeline girth weld under the BS7608 standard is given. It is not difficult to see that the PI-LSTM model is very close to the BS7608 D curve.

[0139] The present invention has a reasonable concept and can evaluate the crack growth rate of the girth weld under multi-size and multi-stress conditions, can realize the probabilistic prediction and uncertainty quantification of the fatigue life of the riser, and can provide a reference for the reliability evaluation of the in-service riser structure.

[0140] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A deep-sea riser fatigue life prediction method based on physical information drive, characterized in that: The main steps include: 1) Collect the original data of the riser fatigue test, including material properties, riser geometry parameters, crack geometry parameters, stress strain, load parameters and crack growth rate; 2) Calculate the physical information parameters of the riser The maximum circumferential stress and generalized strain at the riser crack are obtained by analytical method, and the damage parameters at the riser crack are further obtained. 3) taking five types of parameters, namely, material properties, riser geometry parameters, crack geometry parameters, stress-strain, load parameters, and crack fatigue physical information parameters, as model inputs, and the crack growth rate as model output, the original data of the riser fatigue test in the above step 1) and the physical information parameters of the riser in the above step 2) are normalized and dimensionless respectively to form a deep-sea riser characteristic data set; 4) Data segmentation and reorganization are performed based on the temporal characteristics of the riser fatigue test data. The riser fatigue test data set is divided into a training set and a test set in a ratio of 7:

3. A long-short memory deep learning model with physical constraints is built based on the characteristics of crack propagation. 5) Use the training set in the riser fatigue test data set to train the long-short memory deep learning model, obtain the optimal parameters of the long-short memory deep learning model through the Bayesian optimization method, and compare different models to obtain the optimal physical information driven deep learning model; 6) Use the optimal physical information to drive the deep learning model for fatigue life prediction and uncertainty assessment.

2. The deep-sea riser fatigue life prediction method based on physical information drive according to claim 1, characterized in that: The material properties in step 1) include elastic modulus E, Poisson's ratio v, yield strength limit σ limit The geometric parameters include the riser geometric parameters and the crack geometric parameters. The riser geometric parameters include the inner diameter d, the outer diameter D, the wall thickness t and the tube length L. The crack geometric parameters include the crack depth a and the crack width c. The stress and strain include the maximum principal stress σ max , hydrostatic pressure σ m 、Von Mises stress σ eq , strain range Δε t and stress intensity factor K, load parameters include stress ratio R σ and frequency f.

3. The deep-sea riser fatigue life prediction method based on physical information drive according to claim 1, characterized in that: In step 2), assuming that the crack of the riser expands radially, the circumferential stress σ at the crack of the riser is rmax Expressed by the stress intensity factor: In the above formula (1), K is the force intensity factor, r is the crack length; Generalized strain ε at the riser crack f Using the material's stress triaxiality and critical strain ε o To express: In the above formula (2), is the hydrostatic pressure; σ eq is the Von Mises stress; σ1, σ2 and σ3 are the three components of the principal stress, ε o is the critical strain for the initial crack of the material.

4. The deep-sea riser fatigue life prediction method based on physical information drive according to claim 1, characterized in that: In step 2), the damage parameter at the riser crack is obtained by the following formula (3): SWT=s rmax No t / 2 (3); In the above formula (3), Δε t is the total strain range around the girth weld.

5. The deep-sea riser fatigue life prediction method based on physical information drive according to claim 1, characterized in that: The material properties in step 3) are dimensionless and normalized using the material property parameters at room temperature: In the above formula (4), M is the dimensionless and normalized performance parameter of the material; M d M is the dimensionless and normalized performance parameter of the material; s is the material performance parameter at room temperature; The riser geometric parameters in step 3) are normalized using the dimensions of a standard pipeline, and the crack geometric parameters are normalized using a standard threshold: In the above formula (5), G and G d is the dimensionless and normalized geometric parameters. G includes the riser geometric parameters and crack geometric parameters. th Represents the dimensionless and normalized threshold value, which is selected based on the size of the standard pipe and the maximum value measured in the crack fatigue test; The stress and strain parameters in step 3) are normalized using standard thresholds: In the above formula (6), φ is the dimensionless and normalized stress-strain parameter; φ d is the dimensionless and normalized stress-strain parameter; φ th represents the dimensionless and normalized threshold value, which is selected based on the maximum value measured in the crack fatigue test; The load parameters in step 3) are normalized using the following formula (7): In the above formula (7), R σ Represents stress ratio, R σd represents the ratio of dimensionless to normalized stress, f represents the load frequency, and f d represents the dimensionless and normalized frequency, f th Threshold values ​​specified in the riser fatigue test standards; The physical information parameters in step (3) are normalized using the following formula (8): In the above formula (8), σ rmax represents the maximum hoop stress, σ rmaxd represents the dimensionless and normalized maximum hoop stress, ε f represents the generalized strain, ε d represents the dimensionless and normalized strain, SWT represents the damage parameter at the crack, SWT d represents the dimensionless and normalized damage parameter; σ th , ε th and SWT th Select the maximum value among the results calculated by analytical method; The crack growth rate in step 3) is normalized using the following formula (9): In the above formula (9), CGR represents the crack growth rate, CGR d represents the dimensionless and normalized crack growth rate, CGR th The selection is based on the maximum value of the experimental data.

6. The deep-sea riser fatigue life prediction method based on physical information drive according to claim 1, characterized in that: In step 4), data segmentation and reorganization are performed based on the temporal characteristics of the riser fatigue test data. First, each subset is divided into a training set and a test set using a stratified sampling method; then, physical constraints are introduced into the loss function of the long short-term memory deep learning model to improve the interpretability of the model. The loss function L(θ) consists of two parts. The first part is the deviation between the predicted value and the true value, and the second part is the deviation of the predicted result under the physical constraint, which can be specifically expressed as: L(θ)=L d (θ)+λ·L p (i) (10); In the above formula (10), L d (θ) represents the deviation between the predicted value and the true value; L p (θ) represents the deviation between the predicted value and the physical law; λ is the penalty factor, which is used to balance the impact of the two parts on the loss function; L d (θ) is calculated using RMSE, L p (θ) is determined by calculating the derivative of the crack growth rate; The crack growth process is divided into three stages: initiation stage, stable stage and rapid growth stage; p The expression of (θ) is: In the above formula (11), N represents the number of cycles, a represents the crack depth, and I(A>B) is a state indicator function. When A is greater than B, its value is 1, indicating that the physical constraint is violated, otherwise it is equal to 0, indicating that it is consistent with the physical constraint.

7. The deep-sea riser fatigue life prediction method based on physical information drive according to claim 1, characterized in that: The specific process of step 5) is as follows: 5.1) Select 6 cross validation R 2 The average As the objective function for optimization, determine the number of layers N of the long short-term memory deep learning model layer , the number of units per layer N units and the optimization range of the learning rate η, combining Bayesian optimization with K-fold cross validation to obtain the optimal parameters of the long short-term memory deep learning model. The objective function selected in the optimization process is K-fold cross validation R 2 The average In the above formula (12), R i 2 is the coefficient of determination of the i-th cross validation; is the average value of K cross validations. 5.2) Select the three most critical hyperparameters, namely the number of layers N, from the long short-term memory deep learning model layer , the number of units per layer N units and learning rate η, and use the Bayesian optimization method to obtain the final hyperparameter optimization results; 5.3) Compare the prediction results of the optimal physical information driven deep learning model with those of the traditional model to obtain the optimal model.

8. The deep-sea riser fatigue life prediction method based on physical information drive according to claim 1, characterized in that: The specific process of step 6) is: using the trained physical information to drive the deep learning model to predict fatigue life, comparing the prediction results with the deep learning model without introducing physical parameters and the traditional physical fatigue life model, and quantitatively evaluating the life uncertainty.

9. The deep-sea riser fatigue life prediction method based on physical information drive according to claim 8, characterized in that: The specific steps for quantitatively evaluating the lifetime uncertainty are as follows: 6.1) The crack depth a, crack width c, and maximum principal stress σ of the deep-sea riser feature data set were collected by Latin hypercube sampling. max , strain range Δε t Sampling is performed to obtain a sufficient number of samples; the sampling range is limited to the first 80% of the deep-sea riser characteristic data set; it is assumed that the data of the first 80% of the deep-sea riser characteristic data set obeys a normal distribution, in which the experimental measurement value represents the mean value and the standard deviation is determined by the measurement error; 6.2) Segment the first 80% of the data samples in the deep-sea riser feature data set to meet the input form of the physical information driven deep learning model, and then use the trained physical information driven deep learning model to make predictions to obtain the corresponding crack growth data; 6.3) Calculate the riser crack depth based on the crack extension data. The calculation formula for the crack depth is as follows: In the above formula (13), a i represents the crack depth at time i, a i+1 represents the crack depth at time i+1, represents the crack growth rate at time i, and dN represents the number of cycles; 6.4) Using the failure crack depth measured in the test as the criterion, calculate the number of cycles N when this crack depth is reached in the above formula (13), and this N is the fatigue life; 6.5) Plot the PSN curve based on the fatigue life results to achieve quantitative assessment of uncertainty.

Images