A method for predicting and analyzing the fatigue life of wellhead equipment

Abstract

This invention provides a method for predicting and analyzing the fatigue life of wellhead equipment, relating to the field of downhole engineering technology. The method includes the following steps: collecting load spectrum, material degradation, environmental corrosion, and historical maintenance data of the wellhead equipment to establish a four-dimensional damage factor matrix; preprocessing and feature fusion of the aforementioned multi-source data; and constructing a self-attention LSTM network. This invention incorporates four key factors—load spectrum, material degradation, environmental corrosion, and historical maintenance—into a unified analytical framework through the four-dimensional damage factor matrix, solving the problem of the one-sidedness of traditional single-factor analysis. The self-attention LSTM network can automatically learn the attention weights between multi-source data, significantly improving the ability to extract the correlation features of multi-physics parameters under complex service environments. Combined with Monte Carlo simulation to generate fatigue life probability distribution curves, it provides prediction results including confidence intervals, offering a more scientific quantitative basis for equipment maintenance decisions.

Description

Technical Field

[0001] This invention relates to the field of downhole engineering technology, and in particular to a method for predicting and analyzing the fatigue life of wellhead equipment. Background Technology

[0002] Wellhead equipment is a general term for key equipment installed on top of the drilling casing in oil and gas extraction. It mainly consists of three major components: casing head, tubing head, and Christmas tree. It undertakes core functions such as sealing the wellhead, controlling fluid flow, suspending the downhole tubing string, monitoring pressure, and implementing injection and production operations. Its working environment must withstand the combined effects of high pressure, corrosive media, and complex loads. It is the first line of defense to ensure the safe production of oil and gas wells.

[0003] Wellhead equipment, as a key pressure-bearing device in oil and gas extraction, is subjected to complex service environments such as alternating loads, media corrosion, and temperature changes over long periods. Its fatigue failure risk directly affects safe production and economic benefits. Traditional fatigue life prediction methods are usually based on single load spectrum analysis or simple environmental parameter correction, which have the following technical bottlenecks: lack of a multi-physics coupled failure factor system, making it difficult to quantify the interaction of multiple dimensions such as load spectrum, material degradation, environmental corrosion, and historical maintenance; reliance on manual experience to build data association models, resulting in insufficient ability to mine deep features of multi-source heterogeneous data; and lack of probabilistic life prediction result expression, failing to provide quantitative support for confidence intervals for risk classification decisions. Therefore, this invention proposes a fatigue life prediction and analysis method for wellhead equipment to solve the problems existing in the prior art. Summary of the Invention

[0004] To address the aforementioned problems, this invention proposes a fatigue life prediction and analysis method for wellhead equipment, which provides a more scientific quantitative basis for equipment maintenance decisions.

[0005] To achieve the objectives of this invention, the invention is implemented through the following technical solution: a method for predicting and analyzing the fatigue life of wellhead equipment, comprising the following steps:

[0006] S1: Collect load spectrum, material degradation, environmental corrosion and historical maintenance data of wellhead equipment, and establish a four-dimensional damage factor matrix;

[0007] S2: Perform preprocessing and feature fusion on the above multi-source data;

[0008] S3: Construct a self-attention LSTM network to establish a mapping relationship between multi-source data and fatigue life;

[0009] S4: Combine Monte Carlo simulation to output fatigue life probability distribution curve;

[0010] S5: Make fatigue risk classification decisions based on probability distribution results.

[0011] A further improvement is that S1 includes the following steps:

[0012] Establish a four-dimensional damage factor matrix D that includes load spectrum P, material degradation M, environmental corrosion E, and historical maintenance H:

[0013]

[0014] Wherein: the load spectrum factor P is composed of the dynamic load sequence collected by the wellhead pressure sensor, denoted as P = [p1, p2, ..., p n ], where n is the number of load sampling points, p i This represents the load value at the i-th sampling point;

[0015] Material degradation M is obtained through material hardness testing and metallographic analysis, and includes parameters such as yield strength degradation and elastic modulus degradation, denoted as M = [m1, m2, ..., m k ], k is the dimension of material degradation characteristics;

[0016] Environmental corrosion E includes parameters such as the concentration of the corrosive medium, temperature, and humidity, denoted as E = [e1, e2, ..., e l ], l represents the environmental corrosion characteristic dimension;

[0017] Historical maintenance data (H) is composed of the number of maintenance visits and the stress concentration factor correction value of the maintenance location extracted from maintenance records, denoted as H = [h1, h2, ..., h m ], m is the dimension of historical maintenance characteristics.

[0018] A further improvement is made in S2, where the multi-source data is divided into time-series data and non-time-series data. The time-series data is the load spectrum P, and the non-time-series data includes material degradation M, environmental corrosion E, and historical maintenance H.

[0019] A further improvement is that S2 includes the following steps:

[0020] The time-series data load spectrum P is normalized.

[0021] Perform unique thermal coding or numerical conversion on non-time-series data such as material degradation (M), environmental corrosion (E), and historical maintenance (H);

[0022] Form an input vector of uniform dimension X = [X_P, X_M, X_E, X_H];

[0023] Where X_P is the load spectrum feature vector, X_M is the material degradation feature vector, X_E is the environmental corrosion feature vector, and X_H is the historical maintenance feature vector.

[0024] A further improvement is that, in S3, the self-attention LSTM network includes an input layer, an LSTM layer, a self-attention layer, a fully connected layer, and an output layer;

[0025] The input layer receives the preprocessed feature vector X, with dimensions d = n + k + l + m;

[0026] The LSTM layer contains L memory units, performs sequence modeling on temporal features, and outputs a hidden state sequence H = [h1, h2, ..., ht], where t is the time step.

[0027] The self-attention layer calculates the attention weights αij between features of different dimensions to uncover deep correlations between multi-source data and outputs an attention-weighted feature z, as shown in the formula:

[0028]

[0029] Where αij represents the attention weight of the hidden state hi at time step i to the hidden state hj at time step j (0≤αij≤1); t represents the time step of the LSTM layer output; score(hi,hj) is the attention scoring function, which measures the degree of correlation between hi and hj; k is the summation index, which iterates through all time steps (k=1,2,…,t);

[0030]

[0031] Where αij is the attention weight, hj is the hidden state of the LSTM layer at the j-th time step, and t is the time step size.

[0032] A further improvement is made in S3, where the fully connected layer maps the attention-weighted features z to the output dimension; the output layer uses a regression function to predict the fatigue life L, as shown in the formula:

[0033] L=σ(W·z+b)

[0034] Where z is the output attention weighted feature, W is the weight matrix, b is the bias vector, and σ is the activation function.

[0035] A further improvement is that S4 includes the following steps:

[0036] Define the probability distribution types of random variables such as load spectrum fluctuation range, material degradation rate, and environmental corrosion parameters;

[0037] S different four-dimensional damage factor matrices D^(s) (s=1,2,...,S) were generated through Monte Carlo simulation.

[0038] Input the trained self-attention LSTM network to obtain the life prediction value L^(s), and construct the fatigue life sample set {L^(1),L^(2),...,L^(S)};

[0039] Calculate the mean μ and standard deviation σ of the sample, and plot the fatigue life probability distribution curve;

[0040] Determine the lifetime confidence intervals [L_low, L_high] at 90% and 95% confidence levels.

[0041] A further improvement is made in S5, where a three-level risk grading standard is established based on the fatigue life probability distribution results:

[0042] Low risk: The lower limit of the lifetime confidence interval L_low > the design lifetime L_d, and the lifetime L_mode corresponding to the peak of the probability distribution > 1.2L_d;

[0043] Medium risk: L_d / 2≤L_low≤L_d, and L_mode∈[0.8L_d,1.2L_d];

[0044] High risk: L_low < L_d / 2, or the distribution curve showing a multi-peak pattern indicates the existence of multiple failure mode coupling.

[0045] A further improvement is that, in S5, after the risk classification decision, a three-dimensional visualization report containing probability density curves, confidence intervals, and risk levels is generated.

[0046] The beneficial effects of this invention are as follows:

[0047] 1. This invention incorporates four key factors—load spectrum, material degradation, environmental corrosion, and historical maintenance—into a unified analytical framework through a four-dimensional damage factor matrix, solving the problem of the one-sidedness of single-factor analysis in traditional methods. By using a self-attention LSTM network, it can automatically learn the attention weights between multi-source data, significantly improving the ability to extract the correlation features of multi-physics parameters under complex service environments. Combined with Monte Carlo simulation to generate fatigue life probability distribution curves, it provides prediction results including confidence intervals, providing a more scientific quantitative basis for equipment maintenance decisions.

[0048] 2. This invention introduces historical maintenance factors, enabling the model to quantify the impact of maintenance measures on fatigue life. By analyzing the impact of historical maintenance records on current fatigue life, it can avoid over-maintenance or under-maintenance, reduce maintenance costs, and also connect to the online monitoring data of the wellhead equipment in real time to dynamically update the four-dimensional damage factor matrix, thereby realizing real-time prediction of fatigue life and tracking of risk status. The attention weight matrix output from the attention layer can intuitively reflect the degree of influence of different dimensions of data on fatigue life, which is of great significance for engineering technicians to analyze the causes of equipment failure.

[0049] 3. This invention optimizes the computation process of the self-attention mechanism through matrix operations and combines parallel computing technology to achieve efficient execution of Monte Carlo simulation. While ensuring prediction accuracy, it can control the prediction time of a single sample to within 50ms, meeting the real-time analysis needs of engineering. Attached Figure Description

[0050] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0051] To enhance understanding of the present invention, the present invention will be further described in detail below with reference to embodiments. These embodiments are only used to explain the present invention and do not constitute a limitation on the scope of protection of the present invention.

[0052] Example 1

[0053] according to Figure 1 As shown in the figure, this embodiment proposes a method for predicting and analyzing the fatigue life of wellhead equipment, including the following steps:

[0054] Data on load spectrum, material degradation, environmental corrosion, and historical maintenance of the wellhead equipment were collected to establish a four-dimensional damage factor matrix; a four-dimensional damage factor matrix D was also established, comprising load spectrum P, material degradation M, environmental corrosion E, and historical maintenance H.

[0055]

[0056] Wherein: the load spectrum factor P is composed of the dynamic load sequence collected by the wellhead pressure sensor, denoted as P = [p1, p2, ..., p n ], where n is the number of load sampling points, p i This represents the load value at the i-th sampling point;

[0057] Material degradation M is obtained through material hardness testing and metallographic analysis, and includes parameters such as yield strength degradation and elastic modulus degradation, denoted as M = [m1, m2, ..., m k ], k is the dimension of material degradation characteristics;

[0058] Environmental corrosion E includes parameters such as the concentration of the corrosive medium, temperature, and humidity, denoted as E = [e1, e2, ..., e l ], l represents the environmental corrosion characteristic dimension;

[0059] Historical maintenance data (H) is composed of the number of maintenance visits and the stress concentration factor correction value of the maintenance location extracted from maintenance records, denoted as H = [h1, h2, ..., h m ], m is the dimension of historical maintenance characteristics.

[0060] The aforementioned multi-source data undergoes preprocessing and feature fusion. The multi-source data is divided into time-series and non-time-series data. Time-series data is represented by the load spectrum P, while non-time-series data includes material degradation M, environmental corrosion E, and historical maintenance H. The load spectrum P of the time-series data is normalized. The material degradation M, environmental corrosion E, and historical maintenance H of the non-time-series data are either uniquely encoded or numerically converted. A unified-dimensional input vector X = [X_P, X_M, X_E, X_H] is formed, where X_P is the load spectrum feature vector, X_M is the material degradation feature vector, X_E is the environmental corrosion feature vector, and X_H is the historical maintenance feature vector.

[0061] A self-attention LSTM network is constructed to establish a mapping relationship between multi-source data and fatigue life. The self-attention LSTM network includes an input layer, an LSTM layer, a self-attention layer, a fully connected layer, and an output layer.

[0062] The input layer receives the preprocessed feature vector X, with dimensions d = n + k + l + m;

[0063] The LSTM layer contains L memory units, performs sequence modeling on temporal features, and outputs a hidden state sequence H = [h1, h2, ..., ht], where t is the time step.

[0064] The self-attention layer calculates the attention weights αij between features of different dimensions to uncover deep correlations between multi-source data and outputs an attention-weighted feature z, as shown in the formula:

[0065]

[0066] Where αij represents the attention weight of the hidden state hi at time step i to the hidden state hj at time step j (0≤αij≤1); t represents the time step of the LSTM layer output; score(hi,hj) is the attention scoring function, which measures the degree of correlation between hi and hj; k is the summation index, which iterates through all time steps (k=1,2,…,t);

[0067]

[0068] Where αij is the attention weight, hj is the hidden state of the LSTM layer at the j-th time step, and t is the time step size;

[0069] The fully connected layer maps the attention-weighted features z to the output dimension; the output layer uses a regression function to predict the fatigue life L, as shown in the formula:

[0070] L=σ(W·z+b)

[0071] Where z is the output attention weighted feature, W is the weight matrix, b is the bias vector, and σ is the activation function.

[0072] Traditional LSTM models capture temporal dependencies solely through memory units, failing to distinguish the differences in contributions of data from different dimensions to fatigue life (e.g., the coupling effect of load abrupt changes and material degradation may be treated equally). The self-attention layer of this invention achieves the following innovations: αij automatically quantifies the importance of features at different time steps (e.g., a high load value at a certain time step has a higher correlation weight with subsequent material degradation), overcoming the "equal attention" deficiency of traditional models; the scoring function directly utilizes the inherent correlation of the LSTM hidden states, eliminating the need for additional feature engineering (e.g., manually setting feature weights), and can simultaneously handle cross-dimensional correlations of load spectrum (temporal data), material degradation (static parameters), environmental corrosion (slowly varying parameters), and historical maintenance (discrete events); the attention weight matrix α = [αij]t × t intuitively reflects key failure causes (e.g., if α5 and 10 are significantly higher than other values, it indicates a strong correlation between the load peak at time step 5 and material degradation at time step 10, which can be identified as a key monitoring period).

[0073] The generation process of z is essentially a "dynamic focusing" of multi-source data: by assigning weights to the hidden state hj at each time step through αij, the model automatically focuses on feature combinations that have a greater impact on fatigue life. For example, if hj at a certain time step j simultaneously contains high load values ​​(load spectrum factor P), low material yield strength (material degradation factor M), and high Cl... - If the concentration (environmental corrosion factor E) increases, αij will increase significantly, making hj account for a higher proportion in z, reflecting the key influence of multi-factor coupling on fatigue life; if historical maintenance records show that stress concentration correction (historical maintenance factor H) was performed at a certain time step k, then αik corresponding to hk will decrease, indicating that the maintenance measures effectively alleviated subsequent fatigue damage. The model quantifies this mitigation effect through the weighted result of z. Compared to traditional LSTM models, this invention achieves the following breakthroughs: Traditional models rely only on local information from the latest time step, while z integrates the hidden states of all time steps (j=1 to j=t), capturing long-term dependencies (such as the cumulative effect of early load peaks on later material degradation); each dimension of z is composed of a weighted sum of multiple damage factors (for example, the third dimension of z may be contributed by high load values ​​of the load spectrum, low elastic modulus of material degradation, and high humidity of environmental corrosion), enabling the model to explicitly express the coupling mechanism of "load-material-environment"; by suppressing the influence of noise features (such as occasional sensor anomalies) on z through attention weights, experiments show that on a dataset containing 10% noise, the prediction error of the model of this invention is reduced by more than 25% compared to traditional LSTM.

[0074] Combining the fatigue life probability distribution curve output by Monte Carlo simulation; setting the probability distribution types of random variables such as load spectrum fluctuation range, material degradation rate, and environmental corrosion parameters; generating S different four-dimensional damage factor matrices D^(s) (s=1,2,...,S) through Monte Carlo simulation; inputting the trained self-attention LSTM network to obtain the life prediction value L^(s), constructing a fatigue life sample set {L^(1),L^(2),...,L^(S)}; calculating the mean μ and standard deviation σ of the samples, and plotting the fatigue life probability distribution curve, the formula is:

[0075]

[0076] Determine the lifetime confidence intervals [L_low, L_high] at 90% and 95% confidence levels;

[0077] Fatigue risk classification decisions are made based on probability distribution results; a three-level risk classification standard is established based on the fatigue life probability distribution results.

[0078] Low risk: The lower limit of the lifetime confidence interval L_low > the design lifetime L_d, and the lifetime L_mode corresponding to the peak of the probability distribution > 1.2L_d;

[0079] Medium risk: L_d / 2≤L_low≤L_d, and L_mode∈[0.8L_d,1.2L_d];

[0080] High risk: L_low < L_d / 2, or the distribution curve showing a multi-peak pattern indicates the existence of multiple failure mode coupling.

[0081] After risk classification decision-making, a three-dimensional visualization report is generated, which includes probability density curves, confidence intervals, and risk levels.

[0082] Example 2

[0083] according to Figure 1 As shown in the figure, this embodiment proposes a method for predicting and analyzing the fatigue life of wellhead equipment, including the following steps:

[0084] Deploy a fiber Bragg grating sensor network to synchronously collect stress-strain, corrosion potential, and maintenance record data;

[0085] The D matrix is ​​constructed in real time through edge computing units, with a data refresh rate of 1Hz.

[0086] The Adam optimizer and a dynamic learning rate adjustment strategy are adopted, with a training set to validation set ratio of 8:2.

[0087] Deploy a CUDA-accelerated Monte Carlo engine to achieve 10 5 Parallel simulations per second;

[0088] Generate a 3D visualization report that includes probability density curves, confidence intervals, and risk levels.

[0089] Multi-dimensional coupled modeling: By incorporating four key factors—load spectrum, material degradation, environmental corrosion, and historical maintenance—into a unified analytical framework through a four-dimensional damage factor matrix, the one-sidedness of traditional single-factor analysis methods is resolved. For example, when calculating the load spectrum factor, not only the dynamic load sequence itself is considered, but the influence of changes in material mechanical properties on fatigue damage is also corrected through the material degradation factor, making the model closer to actual service conditions.

[0090] Deep correlation mining capability: The self-attention LSTM network can automatically learn the attention weights between multi-source data, significantly improving the ability to extract the correlation features of multi-physics parameters in complex service environments compared to the traditional LSTM model. Experiments show that, on the same dataset, the prediction accuracy of the SA-LSTM model is 15%-20% higher than that of the ordinary LSTM model, especially when dealing with the interaction between historical maintenance records and current load status.

[0091] Probabilistic Prediction and Risk Quantification: By combining Monte Carlo simulation to generate fatigue life probability distribution curves, this invention provides prediction results including confidence intervals, offering a more scientific quantitative basis for equipment maintenance decisions. Traditional methods only provide a single life prediction value, while this invention can output life ranges at different confidence levels. For example, at a 95% confidence level, it can clearly inform users that the wellhead equipment has a 95% probability of failure within the range [L_low, L_high], helping decision-makers to more accurately assess risks.

[0092] Maintenance decision support optimization: The introduction of historical maintenance factors enables the model to quantify the impact of maintenance measures on fatigue life. For example, the correlation analysis between the number of maintenance operations and the stress concentration factor correction value can guide the formulation of more reasonable maintenance strategies. By analyzing the impact of historical maintenance records on current fatigue life, problems of over-maintenance or under-maintenance can be avoided, reducing maintenance costs.

[0093] Adaptability and Generalization: The data preprocessing module supports standardized processing of various data types, making the method applicable to wellhead equipment of different models and service environments, exhibiting strong engineering versatility. Whether it's newly commissioned equipment or long-serving older equipment, a corresponding fatigue life prediction model can be quickly established by adjusting the parameters in the four-dimensional matrix.

[0094] Multidisciplinary Innovation: Integrating technologies from multiple disciplines such as materials science (material degradation characterization), environmental engineering (corrosion parameter modeling), reliability engineering (Monte Carlo simulation), and machine learning (self-attention mechanism), a comprehensive cross-domain prediction system is formed, breaking through the limitations of traditional single-discipline methods. For example, obtaining material degradation factors requires hardness testing and metallographic analysis techniques from materials science, while modeling environmental corrosion factors relies on corrosion mechanism research from environmental engineering. Combining these with machine learning models enables multidisciplinary collaborative prediction of the fatigue life of wellhead equipment.

[0095] Dynamic update capability: The system can access online monitoring data from wellhead equipment in real time, dynamically updating the four-dimensional damage factor matrix to achieve real-time prediction of fatigue life and tracking of risk status. With the development of IoT technology, sensors on wellhead equipment can collect large amounts of data in real time. The method of this invention can process this data in real time, update model prediction results promptly, and provide support for real-time equipment monitoring and fault early warning.

[0096] Enhanced Feature Interpretability: The attention weight matrix output from the self-attention layer can intuitively reflect the degree of influence of different dimensions of data on fatigue life. For example, the high attention weights of the load spectrum and environmental corrosion parameters indicate that the coupling effect between the two is the main cause of failure, thus improving the interpretability of the model. This is of great significance for engineers to analyze the causes of equipment failure. They can understand which factors have a greater impact on fatigue life through the attention weights, and thus take targeted preventive measures.

[0097] Computational efficiency optimization: By optimizing the computational process of the self-attention mechanism through matrix operations and combining it with parallel computing technology, the Monte Carlo simulation is executed efficiently. While ensuring prediction accuracy, the prediction time for a single sample is controlled within 50ms, meeting the real-time analysis needs of engineering projects. Traditional methods often suffer from low computational efficiency when dealing with large-scale data and complex models. This invention, through technological optimization, ensures high efficiency in practical engineering applications.

[0098] This fatigue life prediction and analysis method for wellhead equipment incorporates four key factors—load spectrum, material degradation, environmental corrosion, and historical maintenance—into a unified analytical framework using a four-dimensional damage factor matrix. This overcomes the limitations of traditional single-factor analysis. By employing a self-attention LSTM network, it automatically learns the attention weights between multi-source data, significantly improving the ability to extract the correlation characteristics of multi-physics parameters under complex service environments. Combined with Monte Carlo simulation to generate fatigue life probability distribution curves, it provides prediction results including confidence intervals, offering a more scientific quantitative basis for equipment maintenance decisions. Furthermore, the introduction of historical maintenance factors allows the model to quantify the impact of maintenance measures on fatigue life. By analyzing the influence of historical maintenance records on current fatigue life, it avoids over-maintenance or under-maintenance, reducing maintenance costs. It can also access real-time online monitoring data from the wellhead equipment to dynamically update the four-dimensional damage factor matrix, enabling real-time fatigue life prediction and risk status tracking. The attention weight matrix output by the self-attention layer intuitively reflects the degree of influence of different dimensions of data on fatigue life, which is of great significance for engineers analyzing the causes of equipment failure. Meanwhile, this invention optimizes the computation process of the self-attention mechanism through matrix operations and combines parallel computing technology to achieve efficient execution of Monte Carlo simulation. While ensuring prediction accuracy, it can control the prediction time of a single sample to within 50ms, meeting the real-time analysis needs of engineering.

[0099] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for predicting and analyzing the fatigue life of wellhead equipment, characterized in that, Includes the following steps: S1: Collect load spectrum, material degradation, environmental corrosion, and historical maintenance data of the wellhead equipment to establish a four-dimensional damage factor matrix. This includes the following steps: Establish a four-dimensional damage factor matrix D that includes load spectrum P, material degradation M, environmental corrosion E, and historical maintenance H: Wherein: the load spectrum factor P is composed of the dynamic load sequence collected by the wellhead pressure sensor, denoted as P=[p1,p2,...,pn], where n is the number of load sampling points and p is the load value of each sampling point; Material degradation M is obtained through material hardness testing and metallographic analysis, and includes parameters such as yield strength degradation and elastic modulus degradation, denoted as M=[m1,m2,...,mk], where k is the dimension of material degradation characteristics; Environmental corrosion E includes parameters such as the concentration of corrosive medium, temperature, and humidity, denoted as E=[e1,e2,...,el], where l is the characteristic dimension of environmental corrosion. Historical maintenance H is composed of the number of maintenances and the stress concentration factor correction value of the maintenance location extracted from the maintenance records, denoted as H=[h1,h2,...,hm], where m is the characteristic dimension of historical maintenance; S2: Preprocessing and feature fusion of load spectrum, material degradation, environmental corrosion and historical maintenance data; S3: Construct a self-attention LSTM network to establish a mapping relationship between multi-source data and fatigue life; S4: Combine Monte Carlo simulation to output fatigue life probability distribution curve; S5: Fatigue risk classification decision based on probability distribution results. A three-level risk classification standard is established based on the fatigue life probability distribution results. Low risk: Lower limit of the lifespan confidence interval >Design life And the lifetime corresponding to the peak of the probability distribution > ; Medium risk: L_d / 2 ≤ ≤ ,and ; High risk: < The presence of multiple peaks in the distribution curve indicates the existence of multiple coupled failure modes.

2. The fatigue life prediction and analysis method for wellhead equipment according to claim 1, characterized in that: In S2, the multi-source data is divided into time-series data and non-time-series data. The time-series data is the load spectrum P, and the non-time-series data includes material degradation M, environmental corrosion E, and historical maintenance H.

3. The fatigue life prediction and analysis method for wellhead equipment according to claim 2, characterized in that: S2 includes the following steps: The time-series data load spectrum P is normalized. Perform unique thermal coding or numerical conversion on non-time-series data such as material degradation (M), environmental corrosion (E), and historical maintenance (H); Form a uniform-dimensional input vector ; in For load spectrum eigenvectors, This is the feature vector of material degradation. This is the characteristic vector of environmental corrosion. This represents the historical maintenance feature vector.

4. The fatigue life prediction and analysis method for wellhead equipment according to claim 3, characterized in that: In S3, the self-attention LSTM network includes an input layer, an LSTM layer, a self-attention layer, a fully connected layer, and an output layer. The input layer receives the preprocessed feature vector X, with dimensions d=n+k+l+m; The LSTM layer contains L memory units, performs sequence modeling on temporal features, and outputs a hidden state sequence H=[h1,h2,...,ht], where t is the time step. The self-attention layer calculates attention weights between features of different dimensions. To uncover deep correlations among multi-source data, the attention-weighted feature z is output, with the formula: Where hi represents the hidden state at time step i, and hj represents the hidden state at time step j. This represents the attention weight of hi on hj, 0 ≤ ≤1; t represents the time step of the LSTM layer output; score(hi,hj) is the attention scoring function, which measures the correlation between hi and hj; k is the summation index, which iterates through all time steps, k=1,2,…,t; in, Here, hj represents the attention weight, hj represents the hidden state of the LSTM layer at time step j, and t represents the time step size.

5. The fatigue life prediction and analysis method for wellhead equipment according to claim 4, characterized in that: In S3, the fully connected layer maps the attention-weighted features z to the output dimension; the output layer uses a regression function to predict the fatigue life L, as shown in the formula: ， Where z is the output attention-weighted feature, W is the weight matrix, and b is the bias vector. This is the activation function.

6. The fatigue life prediction and analysis method for wellhead equipment according to claim 5, characterized in that: S4 includes the following steps: Define the probability distribution types of random variables such as load spectrum fluctuation range, material degradation rate, and environmental corrosion parameters; S groups of different four-dimensional damage factor matrices were generated using Monte Carlo simulation. , The lifetime prediction is obtained by inputting the trained self-attention LSTM network. Construct a fatigue life sample set ; Calculate the sample mean Standard deviation Plot the fatigue life probability distribution curve; Determine the lifetime confidence intervals at 90% and 95% confidence levels. .

7. The fatigue life prediction and analysis method for wellhead equipment according to claim 1, characterized in that: In step S5, after the risk classification decision is made, a three-dimensional visualization report containing probability density curves, confidence intervals, and risk levels is generated.

Images