A multi-stage fatigue crack growth prediction method for deep-sea pressure vessels with variable operating conditions

Abstract

The present invention provides a method for predicting fatigue crack propagation in a deep-sea pressure-resistant tank under variable working conditions and multiple stages, which relates to the technical field of fatigue detection of deep-sea pressure-resistant tanks. The method establishes a finite element model of the entire deep-sea pressure-resistant tank, analyzes weak links, and performs fine modeling. A stress intensity factor-Gaussian process proxy model is established. Fatigue cracks in the deep-sea pressure-resistant tank are predicted based on a dynamic Bayesian best-worst network model. Based on a small-time-scale fatigue life prediction model, a sensitivity analysis is performed on the uncertainty parameters in the dynamic Bayesian network model. A parameter inference algorithm based on a hybrid adaptive resampling intelligent particle filter is used to reduce the influence of uncertainty parameters during crack propagation.

Description

Technical Field

[0001] The present invention relates to the technical field of fatigue detection of deep-sea pressure-resistant cabins, and in particular to a method for predicting fatigue crack propagation of deep-sea pressure-resistant cabins under variable working conditions and multiple stages. Background Art

[0002] Deep-sea pressure vessels are critical equipment for ensuring the safety of deep-sea workers. They experience complex alternating loads, increasing during descent and decreasing during ascent. These loads not only lead to significant stress concentrations but also can induce creep. Under these extreme conditions, the internal microstructure of the material can change. Cracks are particularly susceptible to initiation and propagation in the welds of the pressure vessel. Therefore, determining the location and propagation trends of cracks in deep-sea pressure vessels is of great economic and safety significance.

[0003] To predict the lifespan of critical components subjected to complex alternating loads, such as aircraft engines, CN118536404A proposes a method for predicting fatigue crack growth under variable-amplitude loads based on dynamic Bayesian updating. This method combines the Huang model with a dynamic Bayesian network model for crack growth prediction. Kalman filtering theory combines observational and model information to calculate a theoretical optimal estimate. The model is then corrected based on the optimal estimate for crack prediction, improving computational accuracy. However, this method does not involve training the prediction model, making it unsuitable for fatigue crack growth prediction under other models.

[0004] To accurately determine the current crack status at key structural locations in complex equipment and analyze and predict their growth trends, CN114282709B proposes a method for predicting fatigue crack growth in structural components based on digital twins. This method establishes and trains a Gaussian process proxy model, combined with a dynamic Bayesian network model and a particle filter algorithm. This method can predict cracks in complex equipment and reduce the impact of uncertainties on the crack growth process. However, this method, specifically for deep-sea pressure tanks, fails to consider the impact of changing uncertainties and load characteristics under their operating conditions, severely limiting its application in this field.

[0005] (1) Poor prediction accuracy. In the process of fatigue crack prediction, there are influencing factors such as material properties, load conditions and initial cracks. Especially under the working characteristics of deep-sea pressure chambers, the loading and overloading stages need to mainly consider the changes in the plastic zone size and material yield strength, while the load-bearing stage needs to pay attention to its creep characteristics. As time accumulates, considering only a single case will cause the prediction results to deviate from reality. The dynamic Bayesian best-worst method adopted in this patent is suitable for dealing with uncertainty problems in complex systems. It analyzes the uncertain influencing factors at each stage and assigns different weights. After prediction and comparison, the dominant influencing factors at different stages are obtained. By using the data-model fusion method, under the premise of assuming that the crack detection or monitoring data is correct, the model correction of different working stages is implemented to improve the prediction accuracy.

[0006] (2) Low simulation efficiency. At present, the training data is highly redundant and poorly representative, which reduces the accuracy of model training. This patent proposes a method of nearest neighbor sampling based on unilateral constraints for model training. It can generate representative sample data and quickly and accurately complete the training of fatigue crack prediction models. It also divides the particle generation range according to the weight of the uncertain factors, which can effectively improve the efficiency and accuracy of particle filtering.

[0007] In summary, the current problem of fatigue crack growth prediction in deep-sea pressure vessels is that although the improved formula of Paris can take into account the situation of variable amplitude load, for the pressure vessel research object, the load variation amplitude is relatively slow, which makes the load characteristics inconsistent with the working conditions assumed by the pairs formula. In addition, the influence of creep makes the traditional method inapplicable. To this end, a fatigue crack growth prediction technology at a small time scale is proposed. By considering the crack growth process in any small interval, it can effectively deal with the short-term overload and load maintenance problems at a certain depth in the pressure vessel load cycle lasting several hours. However, when using this model, the prediction deviation caused by multiple uncertain factors such as materials, loads, and models during service is not considered. As a result, the crack prediction results are different from the actual situation, which in turn affects the safety and reliability assessment of the pressure vessel. Summary of the Invention

[0008] Currently, fatigue crack predictions are mostly based on the number of load cycles. However, under the operating conditions of deep-sea pressure vessels, due to the complex loads and short time periods, traditional prediction methods are computationally complex and have reduced accuracy. The present invention uses small time scales as units, integrating various loading conditions into a small-time proxy model. The proxy model is continuously trained with data, and the model parameters are then modified after comparison with the actual crack conditions, thereby improving the model's prediction accuracy. On the other hand, the crack propagation condition is calculated based on time rather than the number of load cycles. This allows for real-time and efficient monitoring of the crack propagation condition of the deep-sea pressure vessel, thereby determining whether its operating conditions are safe.

[0009] The present invention provides a method for predicting fatigue crack growth in a deep-sea pressure-resistant cabin under variable working conditions and multiple stages, comprising:

[0010] S1. Use the shell assumption to establish a finite element model of the pressure cabin, import the operating conditions to formulate boundary conditions, analyze the stress distribution characteristics and conduct detailed modeling of weak links.

[0011] S2. Construct samples for the stress intensity factor-Gaussian process proxy model, bring the generated samples into the finite element model, calculate the stress intensity factor at the crack tip, and establish a stress intensity factor sample database;

[0012] S3. Based on Gaussian process theory and the stress intensity factor sample database, a stress intensity factor-Gaussian process proxy model is established;

[0013] S4. Prediction of fatigue cracks in deep-sea pressure-resistant tanks based on the dynamic Bayesian best-worst network model;

[0014] S5. Based on the small time scale fatigue life calculation model, sensitivity analysis of uncertainty parameters in the dynamic Bayesian best-worst network model is performed;

[0015] S6. A parameter inference algorithm based on hybrid adaptive resampling intelligent particle filtering is used to reduce the influence of uncertain parameters during crack propagation.

[0016] In a preferred embodiment, in step S2, the spatial format of the sample of the stress intensity factor-Gaussian process proxy model is defined as:

[0017] ;

[0018] in, is the load magnitude during the loading and overloading phases, is the load size during the load holding phase, , are the creep constant and the second creep stage index, respectively, is the stress-strain state constraint factor, is the stress intensity factor threshold value, is the crack extension value in the loading and overloading stages, is the crack extension value during the load holding stage.

[0019] In a preferred embodiment, a method based on unilateral restriction of the nearest neighbor sampling is used to generate a sample database. Based on the simulated crack or predicted crack, the distance d above and below it is taken as the sample generation space. The distance between each selected data point and the sample is Multiples of , N groups of data above and below, a total of 2N groups of data.

[0020] In a preferred embodiment, in step S3, the square exponential covariance function is selected as the covariance kernel function of the stress intensity factor-Gaussian process proxy model. :

[0021] ;

[0022] in, is the kernel signal variance that controls the local correlation of the input variables; is the length scale parameter, x i is the i-th input data, x j is the jth input data;

[0023] The selected sample data is used to train the stress intensity factor-Gaussian process surrogate model, and the hyperparameters of the covariance kernel function are estimated. The hyperparameters of the covariance kernel function are optimized and adjusted through maximum likelihood estimation to minimize the prediction error. The trained stress intensity factor-Gaussian process surrogate model is used as the surrogate model to predict the stress intensity factor according to the input crack length and load conditions.

[0024] In a preferred embodiment, step S4 includes:

[0025] S41. Determine the weight of each influencing factor through BWM, manually score the influencing factors based on prior experience, and finally determine the weight;

[0026] S42. Based on the influencing factors and their weights, a DBN model is constructed, where each node represents a risk factor and the arrows represent the causal relationship;

[0027] S43, dynamically updating the weight distribution in the DBN model through monitoring data;

[0028] S44. When the prediction data of the dynamic Bayesian best-worst network model deviates, the working state of the pressure cabin is quickly determined through the reverse reasoning of the DBN model, so as to adjust the parameters of the dynamic Bayesian best-worst network model to adapt to the crack propagation characteristics of the current stage.

[0029] In a preferred embodiment, in step S5, the small time scale fatigue life calculation model is:

[0030] ;

[0031] ;

[0032] ;

[0033] in, is the crack increment during the load holding stage in a loading cycle, is the crack increment during the loading phase of a loading cycle, is the initial crack length, , , Secondary creep index function of the creep constant and the second creep stage index, θ maxo is the maximum crack opening angle, is the stress-strain state constraint factor, is the elastic modulus of the material, is the yield strength of the material, is the maximum stress intensity factor, is the stress intensity factor threshold value; 、 、 , stress-strain state constraint factor and stress intensity factor threshold is considered as an uncertain node, and other variables are considered as deterministic fixed nodes.

[0034] In a preferred embodiment, after setting an initial crack length in the dynamic Bayesian best-worst network model, the actual detected load data is input into the dynamic Bayesian best-worst network model, thereby calculating the crack growth of the structure based on the time series;

[0035] When the crack is detected at time t-1, the observed value The crack length at the next moment is To combine the observed crack length and the prior value of crack length The optimal estimate is calculated; if no crack observation is detected at time t-1, the crack length at the next moment is .

[0036] In a preferred embodiment, after the particle state estimation is completed, according to the small time scale prediction model, the state equation of the fatigue crack growth process is:

[0037] ;

[0038] in is the crack at time t-1, and The various parameters are determined by resampling. Through the Gaussian process proxy model, the stress intensity factor at any time can be accurately predicted, which is the key parameter for crack propagation analysis. Finally, the predicted crack at the time can be output. .

[0039] Compared with the prior art, the present invention has the following beneficial technical effects:

[0040] A sample data generation method based on unilaterally restricted nearest neighbor sampling is proposed, which can generate high-quality sample data, avoid redundancy, cover key areas, improve model training efficiency and prediction accuracy, provide a good foundation for model optimization, and enable the model to predict crack propagation more stably and accurately in complex environments, significantly improving the reliability and efficiency of deep-sea pressure chamber crack prediction; a dynamic Bayesian best-worst network model with variable working conditions and multiple stages is constructed. By analyzing the influence of various uncertain factors under different working conditions, the most accurate prediction method for each stage is determined, which is more in line with the actual working conditions of deep-sea pressure chambers than traditional prediction methods; a hybrid adaptive resampling intelligent particle filter is used as the inference algorithm of the dynamic Bayesian best-worst. According to the real-time situation of the system state and particle distribution, the resampling strategy is adaptively adjusted to better cope with the particle degradation problem. It has strong robustness, can improve the accuracy of state estimation, and is less likely to experience a significant performance degradation.

[0041] The dominant influencing factors of the deep-sea pressure chamber are different at different stages. The dynamic Bayesian best-worst network model is used to find out the dominant influencing factors of each stage. In the subsequent particle filtering, the error range can be adjusted according to the weight. The particle generation range of the dominant factor is appropriately increased, and the particle generation range of the secondary factor is appropriately reduced to improve efficiency and increase accuracy, so as to determine the most suitable prediction accuracy for fatigue cracks in different time periods.

[0042] Through hybrid adaptive resampling intelligent particle filtering, the particle degradation degree, effective sample size, particle diversity and other indicators are monitored in real time, the resampling process is automatically optimized, and the resampling parameters and method combinations are dynamically changed. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 A method for generating nearest neighbor sampling based on unilateral restriction according to the present invention;

[0044] Figure 2 This is the crack calculation flow chart for the stress intensity factor-Gaussian process proxy model;

[0045] Figure 3 This is the flow chart of the hybrid adaptive resampling intelligent particle filter algorithm; DETAILED DESCRIPTION

[0046] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0047] Deep-sea pressure chambers are important equipment for developing marine resources and conducting underwater operations. Fatigue cracks are prone to initiation at the connections of their hulls, which is very detrimental to the safety of personnel and equipment. To solve this problem, this patent proposes a multi-stage fatigue crack prediction method for variable working conditions of pressure chambers.

[0048] In order to obtain the stress intensity factor, a key parameter in fatigue crack prediction, while improving computational efficiency and reducing computational costs, the present invention establishes a proxy model based on the Gaussian process of the stress intensity factor of nearest neighbor sampling, realizes the calculation of the stress intensity factor at the crack tip, distinguishes the sample space format by the time points of transition between different loading stages, adopts the nearest neighbor sampling method based on unilateral restriction to generate the sample data space, and uses typical representative data to train the proxy model to improve its accuracy and credibility.

[0049] The specific steps are:

[0050] S1. Using the shell assumption, a finite element model of the entire pressure cabin was established. Operating conditions were imported to establish boundary conditions, and stress and deformation were analyzed to identify weak links. Solid elements were used to model these weak links, creating a fine mesh and coupling the elements. A precise geometric model was then created using SolidWorks software and imported into Abaqus for meshing.

[0051] S2. Construct samples for the stress intensity factor-Gaussian process proxy model, bring the generated samples into the finite element model, calculate the stress intensity factor at the crack tip, and establish a stress intensity factor sample database.

[0052] According to the working environment characteristics of the deep-sea pressure chamber, the crack propagation characteristics remain unchanged when it is in the respective loading, holding and overloading stages. In order to better construct and apply the Gaussian process proxy model to efficiently predict the distribution and changes of the stress intensity factor (SIF), the spatial format of the sample of the stress intensity factor-Gaussian process proxy model is defined as:

[0053] ;

[0054] in, is the load magnitude during the loading and overloading phases, is the load size during the load holding phase, , are creep constant, second creep stage index, is the stress-strain state constraint factor, is the stress intensity factor threshold value, is the crack extension value in the loading and overloading stages, is the crack extension value during the load holding stage.

[0055] The generated samples are brought into the finite element model, the stress intensity factor at the crack tip is calculated, and a stress intensity factor sample database is established.

[0056] In this step, in order to obtain sample data that is consistent with the actual crack propagation and improve the credibility of the proxy model, the following methods are proposed: Figure 1 The method shown is to generate a sample database.

[0057] Select reference points: First, select multiple reference points from the existing crack path. These reference points can be specific locations on the crack path, such as the starting point, inflection point, or end point of the crack.

[0058] Determine the center and radius: Draw a circle with each reference point as the center and a predetermined length above or below it as the radius. This length can be set based on actual needs and the expected pattern of crack growth.

[0059] Generate new sample data: uniformly or randomly select a point on each circle as a new sample data point. These points represent the possible crack propagation locations near the current reference point.

[0060] Connecting new sample data: These newly generated sample data points are connected from the reference point to the new point to form new crack sample paths. These paths can be straight lines, curves, or other shapes, depending on the physical mechanism and expected pattern of crack propagation.

[0061] Expanding sample data: Using each reference point as the center, continue drawing circles outward by a set length to obtain more new sample data. This process can be repeated multiple times to generate enough sample data for training and validating the crack prediction model.

[0062] Generating a Crack Sample Set: Through the above steps, we can generate a set of crack samples that include crack growth at different locations and directions. These samples can be used to build and train a Gaussian process surrogate model to predict crack growth paths and behaviors.

[0063] In order to improve the accuracy of the proxy model, the nearest neighbor sampling method based on unilateral restriction is used to generate the sample database. Based on the simulated crack or predicted crack, the distance d above and below it is taken as the sample generation space. The distance between each selected data point and the sample is Multiples of, N groups of data above and below, a total of 2N groups of data, where d can be 1~2mm, It can be 0.01~0.02mm, and N can be 50~60.

[0064] In a preferred embodiment, 1 mm is taken above and below as sample generation space, and the space is expanded outward in units of 0.02 mm, with 50 sets of data above and below, for a total of 100 sets of data.

[0065] S3. Based on Gaussian process theory and the stress intensity factor sample database, a stress intensity factor-Gaussian process proxy model is established. Specifically, it includes:

[0066] S31. Select the square exponential covariance function as the covariance kernel function of the stress intensity factor-Gaussian process proxy model :

[0067] ;

[0068] in, is the kernel signal variance that controls the local correlation of the input variables; is the length scale parameter, x i is the i-th input data, x j is the jth input data.

[0069] S32. Use the selected sample data to train the stress intensity factor-Gaussian process proxy model (GP model), estimate the hyperparameters of the covariance kernel function, and optimize and adjust the hyperparameters of the covariance kernel function through maximum likelihood estimation to minimize the prediction error.

[0070] S33. The trained GP model is used as a proxy model, which can predict the stress intensity factor based on the input crack length and loading conditions.

[0071] Traditional finite element analysis requires repeated calculations for each crack state when predicting crack growth, which is computationally very time-consuming. However, the SIF-GP proxy model uses machine learning methods to rapidly predict the SIF value for new crack states, significantly improving computational efficiency. Deep-sea pressure vessels are subject to various uncertainties in their actual operating environments. The SIF-GP proxy model can integrate these uncertainties to provide more accurate crack growth predictions. It also dynamically updates and predicts crack growth states based on real-time monitoring data, which is crucial for structural health management and maintenance decisions.

[0072] S4. Based on the dynamic Bayesian network model, fatigue cracks in deep-sea pressure chambers are predicted.

[0073] Fatigue life assessment of deep-sea pressure hulls requires proper consideration of their load characteristics, including short-term overloads and load maintenance at a certain depth during load cycles lasting several hours. If the influencing factors are not analyzed in a targeted manner at each stage, the predicted results of crack growth will gradually deviate significantly from the actual crack situation over time. Therefore, if Figure 2 A dynamic Bayesian best-worst network model is proposed. Based on the traditional Bayesian approach, it can compare the influencing factors and accurately predict fatigue cracks in deep-sea pressure chambers. The key steps in establishing this dynamic Bayesian best-worst network model are:

[0074] S41. Determine the weight of each influencing factor through BWM, manually score the influencing factors based on prior experience, and finally determine the weight.

[0075] In crack prediction, the weights of various influencing factors must first be determined through BWM to provide the model with initial parameter estimates, which will affect the model's prediction accuracy. Manually score these uncertainty parameters based on prior experience. For example, the stress-strain state constraint factor is generally used in the loading and overloading stages. As the main influencing factor, the creep constant in the load holding stage As the main influencing factors, the relative importance of each factor is determined by combining historical data analysis.

[0076] S42. Based on the influencing factors and their weights, a dynamic Bayesian network model (DBN) is constructed, where each node represents a risk factor and the arrows represent causal relationships.

[0077] Based on the influencing factors and their weights, a DBN model is constructed. Each node represents a risk factor, and the arrows represent the causal relationship. A model framework is created that can represent the interaction of various factors in the crack propagation process. The arrows represent how these factors affect the formation and propagation of cracks.

[0078] S43. Dynamically update the weight distribution in the dynamic Bayesian network model through monitoring data.

[0079] In practice, the DBN's weight distribution is dynamically updated based on monitoring data. Large changes in load indicate a loading or overloading phase, while less significant changes indicate a load-holding phase. This step enables the model to adapt to changing operating conditions, improving prediction accuracy and reliability. For example, if a sudden increase in load is detected, the model can adjust the weights to reflect this change's impact on crack growth.

[0080] S44. When the prediction data of the dynamic Bayesian network model deviates, the working state of the pressure cabin is quickly determined through the reverse reasoning of the DBN, so as to adjust the parameters of the dynamic Bayesian network model to adapt to the crack propagation characteristics of the current stage.

[0081] When the dynamic Bayesian network model's predicted data deviates, reverse reasoning using the DBN can quickly determine the operating state of the pressure chamber, thereby adjusting the dynamic Bayesian network model parameters to adapt to the current crack growth characteristics. This step improves the model's adaptability and accuracy, ensuring that it accurately reflects the actual crack growth process. For example, if the model's predicted crack growth rate does not match the actual observed data, the model can be optimized by adjusting the weights of relevant nodes to better reflect the actual situation.

[0082] S5. Based on the small time scale fatigue life calculation model, sensitivity analysis of the uncertainty parameters in the dynamic Bayesian network model is performed.

[0083] To identify uncertainty parameters in the dynamic Bayesian network model, a sensitivity analysis of uncertainty parameters was performed on its fatigue life calculation model based on small time scales. This module not only accurately calculates crack growth under long-cycle loading, taking into account overload and sustained load effects, but also integrates plastic zone and creep effects, enabling a comprehensive assessment of crack propagation behavior in deep-sea pressure vessels.

[0084] Fatigue life calculation model based on small time scale:

[0085] ;

[0086] ;

[0087] ;

[0088] in is the crack increment during the load holding phase of a loading cycle, t is the time increment at any moment, is the crack increment during the loading phase of a loading cycle (the overloading phase is included in the loading phase), is the initial crack length, , , Secondary creep index Function of creep constant, second creep stage index, , is a constant coefficient determined by the material itself and the load ratio. is the crack opening angle, is the stress-strain state constraint factor, is the elastic modulus of the material, is the yield strength of the material, is the maximum stress intensity factor, is the stress intensity factor threshold value.

[0089] In the dynamic Bayesian best-worst network model, we identified several key uncertain parameters: 、 、 , stress-strain state constraint factor and stress intensity factor threshold These parameters are highly sensitive in the dynamic Bayesian network model of crack growth and are considered as uncertain nodes. Other variables in the model are considered as deterministic fixed nodes.

[0090] After setting an initial crack length in the dynamic Bayesian best-worst network model, the actual detected load data is input into the dynamic Bayesian best-worst network model to calculate the crack extension of the structure based on the time series.

[0091] When the observed crack length detected at time t-1 is The crack length at the next moment is To combine the observed crack length and the prior value of crack length The optimal estimate is calculated; if no crack observation is detected at time t-1, the crack length at the next moment is .

[0092] S6. A parameter inference algorithm based on hybrid adaptive resampling intelligent particle filtering is used to reduce the influence of uncertain parameters during crack propagation.

[0093] In order to reduce the influence of uncertainty parameters during crack growth, such as Figure 3 The paper shows a parameter inference algorithm based on a hybrid adaptive resampling intelligent particle filter. This algorithm effectively addresses the problem of insufficient particle diversity by combining the dynamic inference mechanism of the particle filter with an adaptive resampling strategy.

[0094] The following are the specific steps of the algorithm:

[0095] 1) Prior distribution N initial particles are extracted from , where N is the number of particles;

[0096] 2) From the importance probability density function Extract particles from the equation Calculate the particle weights and normalize the weights to get the particle set ,in is the particle weight, Indicates a known state Under the condition of The conditional probability of is the state value of the particle, is the normalized particle weight;

[0097] 3) Gather particles According to the following formula, it is divided into a set of high-weight particles and a collection of low-weight particles , the high-weight particle set is retained, and each particle in the low-weight particle set is resampled according to steps (3)-(6);

[0098] ;

[0099] Where: weight threshold , Indicates rounding up. is the weight set after descending order, is the effective sample size, which can be calculated according to the following formula:

[0100] ;

[0101] 4) Based on probability (0< <1) Select the adaptive Gaussian mutation strategy to The probability of choosing the crossover strategy is λ. λ is calculated according to the following formula:

[0102] ;

[0103] Generate a random number r∈[0,1], when r≤ Execute step (5), otherwise go to step (6), where is the probability of selecting the adaptive Gaussian mutation strategy;

[0104] 5) According to Generate particles using adaptive Gaussian mutation strategy , and calculate the weight ,in is the particle after Gaussian mutation, , is a collection of low-weight particles The number of particles in is the particle before mutation, , is a set of high-weight particles The number of particles, is the mean vector and , is the covariance matrix;

[0105] 6) According to Generate particles with adaptive crossover strategy , and calculate the weight ;: is the particle after crossing, , is the number of particles in the low-weight particle set, is a particle in the low-weight particle set, is a particle in the high-weight particle set, , is the number of particles in the high-weight particle set, random number is the crossover rate, Available To calculate;

[0106] 7) According to Accept or reject the newly generated particles to obtain particles ,in are the particles after acceptance and rejection after resampling, For particles The weights before normalization, For particles Weights before normalization;

[0107] 8) New particles Add high-weight particle collection Get a new particle set , when the number of resampled particles is less than When is the number of particles in the low-weight particle set, go to step (4), otherwise go to step (9);

[0108] 9) According to Compute the final state estimate.

[0109] (8) After the particle state estimation is completed, the parameters of the small time scale prediction model are obtained, and the state equation of the fatigue crack growth process is:

[0110] ;

[0111] in is the crack length at time t-1, which can be determined by hybrid intelligent resampling and Gaussian process agent model and The various parameters and the key factor of stress intensity factor can finally output the predicted crack length at time t .

[0112] If the predicted crack growth data differs from the actual monitoring data, the DBN-BWM model will correct the prediction by adjusting its parameters. This adjustment is based on the best-worst approach, that is, according to the difference between the predicted and actual monitoring results, the best and worst parameter combinations are selected for adjustment. In this way, the DBN-BWM model can accurately track the growth of cracks in real time, providing decision support for the maintenance and safe operation of the pressure chamber. This dynamic adjustment mechanism enables the model to adapt to the ever-changing working environment and improve the accuracy and reliability of the prediction.

[0113] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above and that the invention can be embodied in other specific forms without departing from the spirit or essential characteristics of the invention. The embodiments should therefore be considered illustrative and non-restrictive, and the scope of the invention is defined by the appended claims, not the foregoing description, and all variations within the meaning and range of equivalents of the claims are intended to be encompassed therein. Any reference sign in a claim should not be construed as limiting the claim to which it relates.

Claims

1. A method for predicting fatigue crack growth in deep-sea pressure-resistant tanks under variable working conditions and multiple stages, characterized in that: include: S1. Establish a finite element model of the entire pressure cabin, import operating conditions to formulate boundary conditions, analyze stress distribution characteristics, and identify weak links; S2. constructing a sample for a stress intensity factor-Gaussian process proxy model, bringing the generated sample into the finite element model, calculating the stress intensity factor at the crack tip, and establishing a stress intensity factor sample database; S3. Based on Gaussian process theory, a stress intensity factor-Gaussian process proxy model is established according to the stress intensity factor sample database; the spatial format of the sample of the stress intensity factor-Gaussian process proxy model is defined as: ； in, is the load magnitude during the loading and overloading phases, is the load size during the load holding phase, , are the creep constant and the second creep stage index, respectively, is the stress-strain state constraint factor, is the stress intensity factor threshold value, is the crack extension value in the loading and overloading stages, is the crack extension value in the load-holding stage, m is the number of rows in the spatial format; S4. Prediction of fatigue cracks in deep-sea pressure-resistant tanks based on dynamic Bayesian network model; S5. Based on a small time scale fatigue life calculation model, perform sensitivity analysis on the uncertainty parameters in the dynamic Bayesian network model; S6. A parameter inference algorithm based on hybrid adaptive resampling intelligent particle filtering is used to reduce the influence of uncertain parameters during crack propagation.

2. The variable working condition multi-stage deep-sea pressure chamber fatigue crack growth prediction method according to claim 1 is characterized in that: The sample database is generated by the nearest neighbor sampling method based on unilateral restriction. Based on the simulated crack or predicted crack, d is taken above and below it as the sample generation space. The distance between each selected data point and the sample is Multiples of , N groups of data above and below, a total of 2N groups of data.

3. The variable working condition multi-stage deep-sea pressure chamber fatigue crack growth prediction method according to claim 1 is characterized in that: The step S4 includes: S41. Determine the weight of each influencing factor, score the influencing factors based on prior experience, and determine the final weight; S42. Based on the influencing factors and their weights, a dynamic Bayesian network model is constructed, where each node represents a risk factor and the arrows represent the causal relationship; S43, dynamically updating the weight distribution in the dynamic Bayesian network model through monitoring data; S44. When the prediction data of the dynamic Bayesian network model deviates, the working state of the pressure cabin is quickly determined through reverse reasoning of the dynamic Bayesian network model, so as to adjust the parameters of the dynamic Bayesian network model to adapt to the crack propagation characteristics of the current stage.

4. The variable working condition multi-stage deep-sea pressure chamber fatigue crack growth prediction method according to claim 1 is characterized in that: In step S5, the small time scale fatigue life calculation model is: ； ； ； in, is the crack increment during the load holding stage in a loading cycle, is the crack increment during the loading phase of a loading cycle, is the initial crack length, Secondary creep index function of the creep constant and the second creep stage index, is the maximum crack opening angle, is the stress-strain state constraint factor, is the elastic modulus of the material, is the yield strength of the material, is the maximum stress intensity factor, is the stress intensity factor threshold value.

5. The variable working condition multi-stage deep-sea pressure chamber fatigue crack growth prediction method according to claim 1 is characterized in that: After setting an initial crack length in the dynamic Bayesian network model, the actual detected load data is input into the dynamic Bayesian network model to calculate the crack growth of the structure based on the time series; When the observed crack length detected at time t-1 is The crack length at the next moment is To combine observations and the prior value of crack length The optimal estimate is calculated; if no crack observation is detected at time t-1, the crack length at the next moment is .

6. The variable working condition multi-stage deep-sea pressure chamber fatigue crack growth prediction method according to claim 5 is characterized in that: Based on the parameter inference algorithm of hybrid adaptive resampling intelligent particle filter, the particle state estimation is performed, and then the state equation of the fatigue crack growth process is calculated as follows: ； in is the crack length at time t-1, which can be determined by hybrid intelligent resampling and Gaussian process agent model and The various parameters and stress intensity factors are finally output, and the predicted crack length at time t is .

Images