Oil and gas pipeline whole life cycle safety state intelligent monitoring method and system

Abstract

The application provides an oil and gas pipeline full-life-cycle safety state intelligent monitoring method and system, relates to the technical field of oil and gas pipelines, and comprises the following steps: collecting oil and gas pipeline operation state data, establishing a digital twin model, constructing a multi-dimensional feature vector and performing feature fusion, constructing a fault evolution prediction model based on the fused feature vector and a graph neural network, calculating a residual life prediction value, continuously optimizing safety state evaluation accuracy through online iterative training, and generating a safety state evaluation report and maintenance decision suggestions. The application realizes intelligent monitoring and prediction of the safety state of oil and gas pipelines, and improves the reliability of pipeline safe operation.

Description

TECHNICAL FIELD

[0001] The present application relates to oil and gas pipeline technology, and in particular to an oil and gas pipeline full life cycle safety state intelligent monitoring method and system. BACKGROUND

[0002] With the rapid development of the oil and gas industry, as an important infrastructure for energy transportation, the safe operation of oil and gas pipelines is related to national energy security and social and economic stability. Oil and gas pipelines are subject to various threat factors such as corrosion, fatigue, and third-party damage during long-term service, which may lead to safety accidents such as pipeline leakage and explosion. Traditional oil and gas pipeline safety management mainly relies on periodic inspection and manual inspection, which is inefficient and difficult to real-time monitor the safety state of the pipeline throughout its life cycle.

[0003] In recent years, with the development of emerging technologies such as the Internet of Things, big data, and artificial intelligence, oil and gas pipeline safety monitoring technology has gradually developed towards digitization and intelligentization. Currently, the industry has begun to use various sensors to collect pipeline operation data in real time and conduct safety state assessment based on these data. However, the existing oil and gas pipeline safety monitoring technology still has the following shortcomings:

[0004] Firstly, most of the existing technology only focuses on single-dimensional pipeline safety parameter monitoring, such as only monitoring pressure, temperature, or corrosion conditions, lacking comprehensive analysis of multi-dimensional data such as pipeline structure characteristics, environmental characteristics, and historical failure characteristics, making it difficult to fully reflect the actual safety state of the pipeline.

[0005] Secondly, traditional pipeline safety assessment methods usually use static assessment models, which cannot effectively describe the dynamic evolution process of pipeline failures, have limited accuracy in predicting the remaining life of the pipeline, and cannot timely warn potential safety risks.

[0006] Finally, the existing pipeline safety state assessment system generally lacks self-learning and self-optimization capabilities, and cannot continuously optimize the assessment model based on new pipeline operation data and failure cases, resulting in insufficient timeliness and accuracy of the safety assessment results, making it difficult to provide reliable scientific basis for pipeline maintenance decisions. SUMMARY

[0007] The embodiments of the present application provide an oil and gas pipeline full life cycle safety state intelligent monitoring method and system, which can solve the problems in the prior art.

[0008] In a first aspect of the embodiments of the present application, an oil and gas pipeline full life cycle safety state intelligent monitoring method is provided, comprising:

[0009] Collecting oil and gas pipeline operation state data, the oil and gas pipeline operation state data including pipeline internal pressure data, pipeline internal temperature data, and pipeline surrounding environment data;

[0010] establishing an oil and gas pipeline digital twin model according to the oil and gas pipeline operation state data; constructing a multi-dimensional feature vector based on the oil and gas pipeline digital twin model, the multi-dimensional feature vector including a pipeline structure feature vector, a pipeline environment feature vector, and a pipeline historical fault feature vector;

[0011] performing feature fusion on the multi-dimensional feature vector to obtain a fused feature vector; and constructing an oil and gas pipeline fault evolution prediction model based on the fused feature vector in combination with a graph neural network;

[0012] calculating an oil and gas pipeline residual life prediction value according to an output result of the oil and gas pipeline fault evolution prediction model;

[0013] continuously optimizing safety state assessment accuracy through online iterative training based on the oil and gas pipeline residual life prediction value in combination with a pipeline safety state assessment rule base; and generating an oil and gas pipeline safety state assessment report and a pipeline maintenance decision suggestion according to the optimized safety state assessment accuracy.

[0014] establishing an oil and gas pipeline digital twin model according to the oil and gas pipeline operation state data; constructing a multi-dimensional feature vector based on the oil and gas pipeline digital twin model, the multi-dimensional feature vector including:

[0015] constructing a physical-virtual mapping matrix based on the oil and gas pipeline operation state data, the physical-virtual mapping matrix containing a plurality of mapping relationship coefficients;

[0016] establishing an oil and gas pipeline digital twin model according to the physical-virtual mapping matrix to realize virtual simulation of the oil and gas pipeline operation state data;

[0017] obtaining input variables and model parameters, constructing a state update equation according to the input variables and the model parameters, and dynamically updating the oil and gas pipeline digital twin model through the state update equation;

[0018] extracting multi-dimensional features based on the oil and gas pipeline digital twin model, including: calculating a pressure fluctuation feature according to a variance of a pressure sampling value and an average pressure, calculating a temperature gradient feature according to a partial derivative of a temperature field in a three-dimensional space, and calculating an environmental influence factor according to weight coefficients of various environmental parameters, wherein the weight coefficients are determined through historical data analysis;

[0019] performing feature fusion on the pressure fluctuation feature, the temperature gradient feature, and the environmental influence factor to construct a multi-dimensional feature vector of the oil and gas pipeline.

[0020] performing feature fusion on the multi-dimensional feature vector to obtain a fused feature vector; and constructing an oil and gas pipeline fault evolution prediction model based on the fused feature vector in combination with a graph neural network, including:

[0021] mapping the multi-dimensional feature vector into a state feature by using a state mapping weight matrix; constructing a time feature based on a timestamp, a period encoding, and a feature change rate; constructing a space feature according to a position coordinate, an azimuth angle, and a feature space gradient; combining the state feature, the time feature, and the space feature to form a spatio-temporal knowledge node;

[0022] extracting a graph feature based on the spatio-temporal knowledge node through a node feature aggregation operation, the node feature aggregation operation adopting attention weights to combine adjacent node features, the attention weights being obtained through spatio-temporal attention calculation;

[0023] processing the graph feature by using a graph neural network to obtain an enhanced feature, and combining the enhanced feature with the multi-dimensional feature vector to obtain a fusion feature vector;

[0024] performing causal chain discovery on the fusion feature vector, including: calculating a causal relationship between features through conditional independence testing, and obtaining a causal strength based on a causal gradient; constructing a causal attention mechanism according to the causal relationship and the causal strength;

[0025] combining the causal attention mechanism with the fusion feature vector, and constructing an oil and gas pipeline fault evolution prediction model through the graph neural network.

[0026] calculating a causal relationship between features through conditional independence testing, and obtaining a causal strength based on a causal gradient; constructing a causal attention mechanism according to the causal relationship and the causal strength includes:

[0027] obtaining time series feature data and state feature data, calculating conditional entropy of the time series feature data and the state feature data, performing conditional independence testing based on the conditional entropy, and obtaining a conditional independence testing result between features;

[0028] performing conditional probability modeling on the time series feature data, calculating a probability dependency relationship between features by using a Bayesian network, and determining a causal relationship between features based on a result of the conditional independence testing and the probability dependency relationship;

[0029] calculating a causal gradient between features based on the causal relationship, obtaining a direct causal strength according to the causal gradient, calculating an indirect causal strength between features along a causal transmission path, combining the direct causal strength and the indirect causal strength to obtain a comprehensive causal strength;

[0030] combine the integrated causal strength with the feature representation to generate a causal query vector and a causal key-value vector; construct a causal attention matrix based on the causal query vector and the causal key-value vector; and reconstruct the feature representation according to the causal attention matrix to obtain a causal attention mechanism.

[0031] According to the output result of the oil and gas pipeline fault evolution prediction model, the calculation of the oil and gas pipeline residual life prediction value includes:

[0032] According to the output result of the oil and gas pipeline fault evolution prediction model, the calculation of the oil and gas pipeline residual life prediction value includes:

[0033] According to the crack evolution feature vector and the corrosion degradation feature vector, a Wiener random process model is established, a drift coefficient, a diffusion coefficient and a measurement noise are combined to construct a degradation state evolution equation, and a nonlinear compensation term is introduced into the degradation state evolution equation.

[0034] Based on the Wiener random process model, the fatigue damage accumulation is calculated, the corrosion rate in the corrosion degradation feature vector and the plastic strain calculated from the crack evolution feature vector are combined, and the physical life consumption rate is obtained by weighting and combining the fatigue damage accumulation, the corrosion rate and the plastic strain.

[0035] The physical life consumption rate is input into the degradation state evolution equation to obtain a data-driven life prediction value; the weight coefficients of the physical life consumption rate and the data-driven life prediction value are adjusted according to the prediction error to output the oil and gas pipeline residual life prediction value.

[0036] According to the crack evolution feature vector and the corrosion degradation feature vector, a Wiener random process model is established, a drift coefficient, a diffusion coefficient and a measurement noise are combined to construct a degradation state evolution equation, and a nonlinear compensation term is introduced into the degradation state evolution equation.

[0037] Obtain the degradation state observation value at multiple time points, the degradation state observation value including state measurement data and timestamp information.

[0038] Calculate the state observation value difference of adjacent time points, divide the state observation value difference by the corresponding time interval to obtain the degradation rate, and use the maximum likelihood estimation method to process the degradation rate to obtain the drift coefficient; square and operate the deviation of the degradation rate and the drift coefficient to obtain the diffusion coefficient.

[0039] Perform time domain analysis on the measurement noise in the degradation state observation value, calculate the autocorrelation function of the measurement noise under different time delays, and perform Fourier transform on the autocorrelation function to obtain the frequency domain characteristics.

[0040] performing state prediction based on the frequency domain features, dynamically correcting a difference between a predicted value and an observed value by using a Kalman gain to eliminate the influence of measurement noise and obtaining a state estimation value;

[0041] performing local linearization processing on the state estimation value, calculating a nonlinear approximation term through second-order Taylor expansion, and constructing a dynamic compensation term based on the nonlinear approximation term, historical state data sequences, and historical prediction errors;

[0042] combining the drift coefficient, the diffusion coefficient, the state estimation value, and the dynamic compensation term to construct a degradation state evolution equation, wherein the degradation state evolution equation includes a Brownian motion term, a filtered noise term, and a nonlinear compensation term, and is used to predict a degradation state at a future time.

[0043] generating a pipeline safety state evaluation report and a pipeline maintenance decision suggestion according to the optimized safety state evaluation accuracy, including:

[0044] the safety state evaluation accuracy after optimization includes evaluation index accuracy and evaluation model confidence; a correction coefficient is calculated according to the evaluation model confidence, a dynamic correction weight is obtained by linearly correcting an initial weight of the evaluation index based on the correction coefficient; and a safety state comprehensive score is obtained by weighted summation of the evaluation index and the dynamic correction weight;

[0045] performing K-means clustering analysis on historical evaluation data to obtain an initial division standard of risk levels; using the evaluation model confidence as an adaptive factor, updating the initial division standard by using an exponential sliding weighted average method; and matching the safety state comprehensive score with the updated division standard to determine a risk level division standard;

[0046] constructing a maintenance priority evaluation model based on the risk level division standard, quantifying risk degree, maintenance difficulty, and time urgency into the maintenance priority evaluation model; setting evaluation factor weights by using the evaluation index accuracy, and calculating a maintenance priority score by weighted summation;

[0047] writing the safety state comprehensive score, the risk level division standard, and the maintenance priority score in a preset evaluation report template in order of importance to generate a pipeline safety state evaluation report;

[0048] identifying evaluation information from the pipeline safety state evaluation report by using a word extraction algorithm, matching the evaluation information with a maintenance decision rule base, and generating a pipeline maintenance decision suggestion including maintenance content, maintenance timing, and resource allocation.

[0049] The second aspect of the embodiment of the present application provides an oil and gas pipeline full life cycle safety state intelligent monitoring system, comprising:

[0050] The first unit is used for collecting oil and gas pipeline operation state data, wherein the oil and gas pipeline operation state data comprises pipeline internal pressure data, pipeline internal temperature data and pipeline peripheral environment data;

[0051] The second unit is used for establishing an oil and gas pipeline digital twin model according to the oil and gas pipeline operation state data; and a multi-dimensional feature vector is constructed based on the oil and gas pipeline digital twin model, wherein the multi-dimensional feature vector comprises a pipeline structure feature vector, a pipeline environment feature vector and a pipeline historical fault feature vector;

[0052] The third unit is used for performing feature fusion on the multi-dimensional feature vector to obtain a fused feature vector; and an oil and gas pipeline fault evolution prediction model is constructed based on the fused feature vector and a graph neural network;

[0053] The fourth unit is used for calculating an oil and gas pipeline residual life prediction value according to an output result of the oil and gas pipeline fault evolution prediction model;

[0054] The fifth unit is used for continuously optimizing safety state evaluation accuracy through an online iterative training mode based on the oil and gas pipeline residual life prediction value and in combination with a pipeline safety state evaluation rule library; and an oil and gas pipeline safety state evaluation report and a pipeline maintenance decision suggestion are generated according to the optimized safety state evaluation accuracy.

[0055] The third aspect of the embodiment of the present application provides an electronic device, comprising:

[0056] a processor;

[0057] a memory for storing processor-executable instructions;

[0058] The processor is configured to invoke the instructions stored in the memory to execute the method described above.

[0059] The fourth aspect of the embodiment of the present application provides a computer-readable storage medium having computer program instructions stored thereon, wherein the computer program instructions are executed by a processor to implement the method described above.

[0060] The present application has the following beneficial effects:

[0061] The present application establishes a digital twin model by collecting oil and gas pipeline operation state data, constructs a multi-dimensional feature vector for feature fusion, and constructs a fault evolution prediction model in combination with a graph neural network, thereby realizing intelligent monitoring of the full life cycle of an oil and gas pipeline and improving the accuracy and reliability of safety state evaluation.

[0062] The application adopts an online iterative training mode to continuously optimize the safety state evaluation accuracy, can timely adapt to the changes of the pipeline operation environment and state, dynamically adjusts the evaluation model, so that the safety state evaluation result is more in line with the actual situation, and more accurate pipeline maintenance decision suggestions are provided.

[0063] The application analyzes the pipeline structure features, environmental features and historical failure features, comprehensively considers the influence of various factors on the pipeline safety, breaks through the limitation of traditional single dimension evaluation, makes the oil and gas pipeline residual life prediction more scientific and reasonable, effectively reduces the pipeline safety accident risk, prolongs the pipeline service life, and improves the pipeline operation economic benefit. BRIEF DESCRIPTION OF DRAWINGS

[0064] Figure 1 It is a flowchart of the oil and gas pipeline full life cycle safety state intelligent monitoring method of the embodiment of the application.

[0065] Figure 2 It is a performance comparison column chart of the oil and gas pipeline failure evolution prediction model of the embodiment of the application.

[0066] Figure 3 It is a flowchart of the causal attention mechanism construction based on conditional independence and Bayesian network of the embodiment of the application.

[0067] Figure 4 It is a schematic diagram of the relationship between the fatigue damage accumulation and the prediction error of the embodiment of the application.

[0068] Figure 5 It is a performance comparison column chart of the pipeline safety state evaluation and maintenance decision method of the embodiment of the application. DETAILED DESCRIPTION

[0069] In order to make the purpose, technical scheme and advantages of the embodiments of the application clearer, the technical scheme in the embodiments of the application will be described clearly and completely below in combination with the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, not all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor are within the protection scope of the application.

[0070] The technical scheme of the application will be described in detail below with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes can not be described in some embodiments.

[0071] Figure 1 It is a flowchart of the oil and gas pipeline full life cycle safety state intelligent monitoring method of the embodiment of the application, as shown in Figure 1 the method comprises:

[0072] collecting oil and gas pipeline operation state data, the oil and gas pipeline operation state data including pipeline internal pressure data, pipeline internal temperature data and pipeline peripheral environment data;

[0073] establishing an oil and gas pipeline digital twin model according to the oil and gas pipeline operation state data; constructing a multi-dimensional feature vector based on the oil and gas pipeline digital twin model, the multi-dimensional feature vector including a pipeline structure feature vector, a pipeline environment feature vector and a pipeline historical fault feature vector;

[0074] performing feature fusion on the multi-dimensional feature vector to obtain a fused feature vector; constructing an oil and gas pipeline fault evolution prediction model based on the fused feature vector combined with a graph neural network;

[0075] calculating an oil and gas pipeline remaining life prediction value according to an output result of the oil and gas pipeline fault evolution prediction model;

[0076] based on the oil and gas pipeline remaining life prediction value, combining a pipeline safety state evaluation rule base, continuously optimizing safety state evaluation accuracy through online iterative training; generating an oil and gas pipeline safety state evaluation report and a pipeline maintenance decision suggestion according to the optimized safety state evaluation accuracy.

[0077] In an optional implementation, according to the oil and gas pipeline operation state data, an oil and gas pipeline digital twin model is established; based on the oil and gas pipeline digital twin model, a multi-dimensional feature vector is constructed, including:

[0078] constructing a physical-virtual mapping matrix based on the oil and gas pipeline operation state data, the physical-virtual mapping matrix containing a plurality of mapping relationship coefficients;

[0079] establishing an oil and gas pipeline digital twin model according to the physical-virtual mapping matrix, realizing virtual simulation of the oil and gas pipeline operation state data;

[0080] obtaining input variables and model parameters, constructing a state update equation according to the input variables and the model parameters, and dynamically updating the oil and gas pipeline digital twin model through the state update equation;

[0081] extracting multi-dimensional features based on the oil and gas pipeline digital twin model, including: calculating pressure fluctuation features according to the variance of pressure sampling values and average pressure, calculating temperature gradient features according to the partial derivative of the temperature field in three-dimensional space, and calculating environmental influence factors according to the weight coefficients of each environmental parameter, wherein the weight coefficients are determined through historical data analysis;

[0082] The pressure fluctuation feature, the temperature gradient feature, and the environmental influence factor are fused to construct a multi-dimensional feature vector of the oil and gas pipeline.

[0083] In the oil and gas pipeline operation state data acquisition stage, the system collects data through a sensor network distributed at key nodes of the pipeline. These data include pressure data, temperature data, flow data, environmental data, etc. For example, on an oil pipeline 100 kilometers long, a sensor node is set every 5 kilometers, a total of 20 nodes, each node collects parameters such as pressure (0-10 MPa range), temperature (-20℃ to 80℃ range), flow (0-2000m 3 / h range), etc., with a sampling frequency of 1 time / minute.

[0084] The construction of the physical-virtual mapping matrix is a key step in the establishment of the digital twin model. The mapping matrix contains multiple mapping relationship coefficients for mapping the parameters of the physical world to the virtual space. In specific implementation, the system analyzes historical operation data to identify the correlation between parameters and establish the mapping relationship between parameters. For example, for the mapping relationship between pressure P and flow Q, through the analysis of 10,000 historical data, the mapping coefficient is determined to be 0.85; the mapping coefficient of temperature T and environmental temperature Te is 0.72; the mapping coefficient of pressure fluctuation and pipeline wear degree is 0.68. These coefficients form a 15x15 mapping matrix covering the mapping relationship between all key parameters.

[0085] Based on the physical-virtual mapping matrix, the system establishes the digital twin model of the oil and gas pipeline by using virtual simulation technology to construct a three-dimensional geometric model and an operation state model of the pipeline. The geometric model accurately restores the spatial layout, material characteristics and connection mode of the pipeline, including pipe diameter, wall thickness, bending angle and other parameters. The operation state model converts the real-time collected physical data into state variables in the virtual model through the mapping matrix. For example, when the pressure of a certain section of the actual pipeline rises to 8.5 MPa, the pressure value at the corresponding position in the virtual model will be updated, and the color change (such as from green to yellow) will be intuitively displayed.

[0086] To realize dynamic updating of the digital twin model of the oil and gas pipeline, the system obtains input variables and model parameters to construct a state update equation. The input variables include real-time monitored pressure, temperature, flow and other data; the model parameters include pipeline material parameters, medium physical property parameters, environmental influence coefficients, etc. The state update equation predicts the system state at the next time through time series analysis method.

[0087] In predicting the change of pipeline pressure, the system considers the current pressure value 7.2 MPa, flow 1500m 3 / h, temperature 45℃, etc. input variables, combined with pipeline roughness coefficient 0.025, medium density 850 kg / m3 With the same model parameters, it is predicted that the pressure will change to 7.5 MPa after 0.5 hours. The system performs a state update every 10 minutes to ensure that the virtual model is highly consistent with the actual pipeline state.

[0088] Based on the oil and gas pipeline digital twin model, the system extracts multi-dimensional features. The pressure fluctuation feature is obtained by calculating the variance of the pressure sampling value and the average pressure. For example, the average pressure of a certain monitoring point within 24 hours is 6.8 MPa, and the standard deviation is 0.15 MPa. The system calculates the pressure fluctuation feature value as 0.0225, indicating that the pressure stability of this point is good. The temperature gradient feature is obtained by calculating the partial derivative of the temperature field in three-dimensional space.

[0089] The pipeline is divided into 500 temperature monitoring areas, and the temperature change rate of each area is calculated. For example, in a certain section, the temperature along the pipeline is detected to decrease from 55℃ to 45℃, with a length of 2000 meters, and the temperature gradient is 0.005℃ / m. The environmental influence factor is obtained by weighted calculation of each environmental parameter. The system analyzes 3 years of historical data to determine the weight coefficients of environmental parameters: air temperature 0.35, humidity 0.25, geological conditions 0.30, and corrosion environment 0.10. The environmental parameters of a certain section of the pipeline are: air temperature 35℃, humidity 65%, geological condition score 80 points (full score 100), and corrosion environment score 30 points (full score 100). The calculated environmental influence factor is 65.5.

[0090] The pressure fluctuation feature, temperature gradient feature and environmental influence factor are fused to construct a multi-dimensional feature vector of the oil and gas pipeline. The feature fusion adopts normalization processing and vector splicing method. For the above example data, the pressure fluctuation feature 0.0225 is normalized to 0.45, the temperature gradient feature 0.005℃ / m is normalized to 0.33, and the environmental influence factor 65.5 is normalized to 0.655.

[0091] These normalized features are spliced to form a feature vector [0.45, 0.33, 0.655,...], which can also be extended to include other feature dimensions such as flow rate change rate 0.28 and pipeline vibration amplitude 0.15. The final multi-dimensional feature vector can comprehensively represent the operating state of the oil and gas pipeline, providing a data basis for subsequent state assessment and fault warning.

[0092] In an alternative embodiment, the multi-dimensional feature vector is fused to obtain a fused feature vector; and based on the fused feature vector, an oil and gas pipeline fault evolution prediction model is constructed by combining a graph neural network, comprising:

[0093] map the multi-dimensional feature vector to state features using a state mapping weight matrix; construct temporal features based on timestamps, periodic encoding, and feature change rates; construct spatial features according to position coordinates, azimuth angles, and feature space gradients; combine the state features, the temporal features, and the spatial features to form spatio-temporal knowledge nodes;

[0094] extract graph features based on the spatio-temporal knowledge nodes through node feature aggregation operations, the node feature aggregation operations using attention weights to combine neighboring node features, the attention weights being obtained through spatio-temporal attention calculation;

[0095] process the graph features using a graph neural network to obtain enhanced features, and combine the enhanced features with the multi-dimensional feature vector to obtain a fusion feature vector;

[0096] perform causal chain discovery on the fusion feature vector, including: calculating causal relationships between features through conditional independence testing, and obtaining causal strengths based on causal gradients; and constructing a causal attention mechanism according to the causal relationships and the causal strengths;

[0097] combine the causal attention mechanism with the fusion feature vector, and construct an oil and gas pipeline fault evolution prediction model through the graph neural network.

[0098] perform feature fusion on multi-dimensional feature vectors, and then construct an oil and gas pipeline fault evolution prediction model based on the fusion feature vectors and a graph neural network.

[0099] When performing feature fusion on multi-dimensional feature vectors, the system collects sensor data of multiple monitoring points of an oil and gas pipeline, including parameters such as pressure, temperature, flow rate, and vibration frequency. For example, on a 500-meter-long oil and gas pipeline, there are 10 monitoring points, and each monitoring point collects data with a dimension of 20. After preprocessing, these data form a multi-dimensional feature vector.

[0100] In the process of mapping the multi-dimensional feature vector to state features using a state mapping weight matrix, the system constructs a weight matrix with a size of 20x15, and the element values of the matrix are obtained according to expert experience and correlation analysis of historical data. For example, when the correlation between the pressure feature and the leakage fault is 0.85, the corresponding weight element is set to 0.85. Through the multiplication operation of the weight matrix and the original feature vector, the 20-dimensional original feature vector is mapped to a 15-dimensional state feature vector, thereby extracting core information directly related to the pipeline state.

[0101] When constructing time features based on timestamps, periodic encoding, and feature change rates, the system converts the timestamps of collected data into the time of day (0-24 hours), day of the week (1-7), and seasonal information (1-4). At the same time, it calculates feature change rates, such as a pressure drop from 3.5 MPa to 3.2 MPa in 10 minutes, with a change rate of -0.03 MPa / minute. These time-related information is combined into an 8-dimensional time feature vector to capture the regularity of pipeline state changes over time.

[0102] When constructing spatial features based on position coordinates, azimuth angles, and feature spatial gradients, the system records the three-dimensional coordinates (x, y, z) of each monitoring point and the azimuth angle of the pipeline at that point. For example, a certain monitoring point has coordinates (125.6, 43.2, 15.3) and an azimuth angle of 37.5 degrees. At the same time, it calculates the spatial gradient of features between adjacent monitoring points, such as a temperature drop from 45°C to 42°C over a distance of 50 meters, with a temperature gradient of -0.06°C / meter. These spatial information is combined into a 10-dimensional spatial feature vector to represent the spatial distribution characteristics of the pipeline state.

[0103] After combining state features, time features, and spatial features into spatiotemporal knowledge nodes, each monitoring point forms a 33-dimensional (15+8+10) feature vector as a node representation in the graph network. For example, a certain monitoring point's spatiotemporal knowledge node contains state information such as pressure 3.5 MPa and temperature 45°C, time information such as 10 am, weekday, and summer, and spatial information such as the specific location and surrounding environmental characteristics of the point.

[0104] When extracting graph features based on spatiotemporal knowledge nodes through node feature aggregation operations, the system defines a receptive field for each node, which is a set of other nodes connected to it. For two adjacent nodes, the system calculates the spatiotemporal attention weight between them. The time attention is calculated based on the time difference between the data collected by the two nodes, such as a 5-minute time difference corresponding to a time attention weight of 0.9. The spatial attention is calculated based on the physical distance between the nodes, such as a distance of 50 meters corresponding to a spatial attention weight of 0.85. The combined attention weight is obtained by combining the time attention and the spatial attention, which is used for weighted aggregation of node features. For example, the feature vector of the center node is [3.5, 45,...], the feature vector of the adjacent node is [3.4, 44,...], and the corresponding attention weight is 0.8. The aggregated feature component contains 0.2 x 3.5 + 0.8 x 3.4 = 3.42.

[0105] When processing the graph feature using the graph neural network to obtain the enhanced feature, the system constructs a three-layer graph convolution network, and the output dimensions of each layer are 40, 30 and 25 respectively. The network continuously updates the node representation through the message passing mechanism to capture the high-order connection relationship between nodes. The 25-dimensional enhanced feature vector obtained after processing is spliced with the original 20-dimensional feature vector to form a 45-dimensional fusion feature vector.

[0106] When performing causal chain discovery on the fusion feature vector, the system calculates the causal relationship between features through conditional independence testing. For example, the system finds that there is a causal relationship between pressure reduction and pipe wall thickness reduction, and the p-value of the conditional independence test is 0.02 (less than the significance level 0.05). At the same time, the causal strength is estimated by calculating the intervention effect, for example, the pressure reduction of 1 MPa leads to an increase of 25% in the risk of leakage, and the corresponding causal strength is 0.25. Based on the discovered causal relationship and strength, a causal graph is constructed, and a causal attention mechanism is designed to give higher weights to the features on the causal chain. For example, in predicting the pipeline leakage risk, the attention weights of the pressure and pipe wall thickness features are 0.3 and 0.25 respectively, while the attention weight of the environmental temperature feature which has weak relationship with leakage is only 0.05.

[0107] After combining the causal attention mechanism with the fusion feature vector, the system uses the graph neural network with the causal attention layer to construct the final oil and gas pipeline fault evolution prediction model. The model contains two graph convolution layers and three fully connected layers, and can predict the fault risk probability and possible fault type of each monitoring point within the next 24 hours. In practical application, the model monitors a certain section of the pipeline and successfully predicts that a pipe wall corrosion point may develop into a leakage fault within 72 hours with a risk probability of 85%, and issues an early warning 36 hours in advance, enabling maintenance personnel to take timely measures to avoid possible safety accidents.

[0108] Figure 2 The performance comparison column chart of the oil and gas pipeline fault evolution prediction model of the embodiment of the present application is shown in the following table:

[0109] The figure shows the performance comparison of the three different models (the basic model, the spatio-temporal feature fusion model and the graph neural network causal model) on five evaluation indicators. The graph neural network causal model performs best on all indicators, with a prediction accuracy of 94.3%, a fault detection rate of 92.1%, an early warning time indicator of 86.7%, a causal chain recognition rate of 89.5%, and a model convergence speed of 91.2%. The spatio-temporal feature fusion model performs second in terms of indicators, with a prediction accuracy of 88.7%, a fault detection rate of 84.6%, an early warning time indicator of 75.9%, a causal chain recognition rate of 81.3%, and a model convergence speed of 83.6%. The performance of the basic model is relatively the worst, with a prediction accuracy of 82.5%, a fault detection rate of 76.8%, an early warning time indicator of only 65.3%, a causal chain recognition rate of 70.2%, and a model convergence speed of 78.4%. The data shows that after introducing the graph neural network and causal reasoning mechanism, the model has significantly improved in terms of prediction accuracy, fault detection ability, early warning timeliness, causal relationship identification and computational efficiency, with an average improvement of about 15-20 percentage points compared to the basic model, confirming the superiority of the method.

[0110] In an optional implementation, the causal relationship between features is calculated by conditional independence test, and the causal strength is obtained based on the causal gradient; and the causal attention mechanism is constructed according to the causal relationship and the causal strength, including:

[0111] The time series feature data and the state feature data are obtained, the conditional entropy of the time series feature data and the state feature data is calculated, the conditional independence test is performed based on the conditional entropy, and the conditional independence test result between features is obtained;

[0112] The conditional probability of the time series feature data is modeled, the probability dependency relationship between features is calculated by using a Bayesian network, and the causal relationship between features is determined based on the result of the conditional independence test and the probability dependency relationship;

[0113] The causal gradient between features is calculated based on the causal relationship, and the direct causal strength is obtained according to the causal gradient; the indirect causal strength between features is calculated along the causal transmission path; the direct causal strength and the indirect causal strength are combined to obtain the comprehensive causal strength;

[0114] The comprehensive causal strength and the feature representation are combined to generate a causal query vector and a causal key-value vector; the causal attention matrix is constructed based on the causal query vector and the causal key-value vector; and the feature representation is reconstructed according to the causal attention matrix to obtain the causal attention mechanism.

[0115] As shown in Figure 2 the method includes:

[0116] Causal relationships between features are calculated by conditional independence test, causal strength is obtained based on causal gradient, and a causal attention mechanism is constructed according to the causal relationship and the causal strength.

[0117] Timing feature data and state feature data are obtained, wherein the timing feature data can be data acquisition values at continuous multiple time points, for example, temperature, pressure, humidity and other indicators recorded by a certain device every hour in the past 24 hours; the state feature data can be discrete or continuous values such as current working mode and running state of the device. Taking a certain production device as an example, the timing feature data includes motor temperature (35-45℃) and vibration frequency (20-30Hz) recorded every hour in the past 10 days; the state feature data includes current load rate (85%) and running mode (heavy load mode).

[0118] Conditional entropy of the timing feature data and the state feature data is calculated. Conditional entropy represents the uncertainty of one feature when another feature is known. For each pair of features X and Y, H(X|Y) and H(Y|X) are calculated. For example, the motor temperature (X) and the load rate (Y), H(X|Y) = 0.32 and H(Y|X) = 0.45 can be calculated by statistical analysis.

[0119] Conditional independence test is performed based on the conditional entropy to obtain the conditional independence test result between the features. The conditional independence test determines whether two features are independent of each other given a third feature Z. The G 2 test or chi-square test is used for testing, and the significance level is set to 0.05. For example, the motor temperature (X) and the vibration frequency (Y) are tested given the load rate (Z), and if the p-value is 0.03<0.05, the independence hypothesis is rejected, and it is considered that X and Y are dependent given Z.

[0120] Conditional probability modeling is performed on the timing feature data, and the probability dependency relationship between the features is calculated using a Bayesian network. The Bayesian network is a directed acyclic graph, and the nodes represent the features and the edges represent the conditional dependency relationship. The network structure is evaluated using a scoring function such as BIC (Bayesian Information Criterion) or MDL (Minimum Description Length). In the above example, the dependency relationships of load rate → motor temperature and load rate → vibration frequency are obtained, with BIC scores of -120.5 and -135.8, respectively.

[0121] Causal relationships between features are determined based on the results of conditional independence tests and probabilistic dependence relationships. The causal discovery principle is followed: if X and Y have a dependence relationship, and X occurs before Y, and there is no feature Z that makes X and Y conditionally independent, then X is considered the cause of Y. Combined with the timing information, the causal direction between features is determined. For example, it is determined that the load rate is the cause of the motor temperature, and the motor temperature is the cause of bearing wear, etc., forming a causal graph structure.

[0122] Based on the causal relationship, the causal gradient between the features is calculated, i.e. the degree of influence of the change of a feature on its causal consequences. For directly connected features X and Y, the causal gradient is calculated by intervening X and observing the change of Y. It can be achieved by counterfactual analysis method: change X from the current value by ±10%, record the corresponding change of Y, and calculate the change rate. For example, when the load rate increases from 85% to 93.5% (+10%), the motor temperature rises from 42℃ to 45℃ (+7.1%), and its direct causal strength can be calculated as 0.71.

[0123] The indirect causal strength between features is calculated along the causal transmission path. For features connected through intermediate nodes, the indirect causal effect is calculated. Assuming that feature A affects C through B, the indirect causal strength of A to C is the product of the direct causal strength of A to B and the direct causal strength of B to C. For example, the indirect causal strength of load rate to bearing wear is 0.71 (load rate to motor temperature) x 0.65 (motor temperature to bearing wear) = 0.46.

[0124] The direct causal strength and the indirect causal strength are combined to obtain the comprehensive causal strength. A weighted summation method can be used, in which the weight of the direct causal strength is 0.7 and the weight of the indirect causal strength is 0.3. For example, if the load rate has both direct influence (0.25) and indirect influence (0.46) on bearing wear through the motor temperature, the comprehensive causal strength is 0.7 x 0.25 + 0.3 x 0.46 = 0.313.

[0125] The comprehensive causal strength is combined with the feature representation to generate the causal query vector and the causal key-value vector. The feature representation can be a vector obtained through an embedding layer. Assuming that the feature representation dimension is 64, the comprehensive causal strength is fused with the feature representation through linear transformation. For example, for the load rate feature, its original representation vector is [0.12, 0.25,..., 0.08], and combined with its comprehensive causal strength to other features [0.313, 0.28,..., 0.15], the causal query vector [0.16, 0.26,..., 0.09] is obtained.

[0126] The causal attention matrix is constructed based on the causal query vector and the causal key-value vector. The dot product of the query vector and the key-value vector is calculated and normalized by softmax. For example, an 8x8 attention matrix is generated, where the value in the ith row and jth column represents the attention weight of the ith feature to the jth feature.

[0127] The feature representation is reconstructed according to the causal attention matrix to obtain the causal attention mechanism. The attention matrix is multiplied with the feature value vector to generate a new feature representation. For example, for the load rate feature, a new feature representation [0.18, 0.27,..., 0.11] is obtained through weighted combination with the attention weight, which better integrates the causal relationship information.

[0128] The causal attention mechanism constructed by the above steps can effectively capture the causal relationship and strength between features, and improve the understanding and prediction ability of the model for complex systems. In practical applications, this mechanism can be integrated into deep learning models for device failure prediction, system anomaly detection, etc.

[0129] In an optional implementation, the residual life prediction value of the oil and gas pipeline is calculated according to the output result of the oil and gas pipeline failure evolution prediction model, including:

[0130] The output result of the oil and gas pipeline failure evolution prediction model is decoupled according to the failure mechanism to construct a crack evolution feature vector and a corrosion degradation feature vector;

[0131] A Wiener random process model is established according to the crack evolution feature vector and the corrosion degradation feature vector, and a degradation state evolution equation is constructed by combining the drift coefficient, diffusion coefficient and measurement noise; a nonlinear compensation term is introduced into the degradation state evolution equation;

[0132] The fatigue damage accumulation is calculated based on the Wiener random process model, and the corrosion rate in the corrosion degradation feature vector and the plastic strain are calculated based on the crack evolution feature vector; the physical life consumption rate is obtained by weighted combination of the fatigue damage accumulation, the corrosion rate and the plastic strain;

[0133] The physical life consumption rate is input into the degradation state evolution equation to obtain a data-driven life prediction value; the weight coefficients of the physical life consumption rate and the data-driven life prediction value are adaptively adjusted according to the prediction error, and the residual life prediction value of the oil and gas pipeline is output.

[0134] The output results of the oil and gas pipeline failure evolution prediction model are decoupled according to failure mechanisms to construct a crack evolution feature vector and a corrosion degradation feature vector. Specifically, a non-negative matrix factorization algorithm can be used to decompose the feature matrix output by the model to separate the features related to crack evolution and the features related to corrosion degradation.

[0135] The crack evolution feature vector includes parameters such as crack size, crack propagation rate, and stress intensity factor; the corrosion degradation feature vector includes parameters such as corrosion depth, corrosion rate, and corrosion area. For example, for a 1000-meter-long X65 steel oil and gas pipeline, the crack evolution feature vector [2.3 mm, 0.05 mm / year, 25 MPa·m^(1 / 2)] and the corrosion degradation feature vector [1.5 mm, 0.08 mm / year, 150 cm 2 ] can be obtained through feature decoupling.

[0136] Based on the feature vectors obtained through the above decoupling, a Wiener random process model is established to combine the drift coefficient, diffusion coefficient, and measurement noise to construct a degradation state evolution equation. A nonlinear compensation term is introduced in the equation to improve the description accuracy of the model to the actual degradation process. Specifically, the drift coefficient represents the average rate of change of the degradation process, the diffusion coefficient represents the randomness of the degradation process, and the measurement noise represents the uncertainty of the monitoring data. For the oil and gas pipeline in the above example, the drift coefficient can be set to 0.12, the diffusion coefficient can be set to 0.05, and the measurement noise standard deviation can be set to 0.02. The nonlinear compensation term can be constructed by a cubic spline interpolation function to capture the nonlinear characteristics in the degradation process.

[0137] After the degradation state evolution equation is constructed, the fatigue damage accumulation is calculated based on the Wiener random process model. The calculation of fatigue damage accumulation takes into account the cyclic loading experienced by the oil and gas pipeline during service, and is quantified using the cumulative damage theory. For the oil and gas pipeline in the example, it is assumed to have experienced a 10-year service period, and the calculated fatigue damage accumulation is 0.35 (when the cumulative amount reaches 1.0, it indicates that the fatigue life limit is reached).

[0138] The plastic strain is calculated based on the corrosion rate in the corrosion degradation feature vector and the crack evolution feature vector. The calculation of plastic strain needs to consider the stress-strain relationship of the pipeline material and the influence of corrosion and cracks on the carrying capacity of the pipeline. For the oil and gas pipeline in the example, the calculated plastic strain is 0.027.

[0139] The physical life consumption rate is obtained by weighted combination of the fatigue damage accumulation, the corrosion rate and the plastic strain. The weighting coefficients can be set according to the importance of different failure mechanisms on the pipeline life, and in the example, the weight of the fatigue damage accumulation can be set as 0.4, the weight of the corrosion rate can be set as 0.35, and the weight of the plastic strain can be set as 0.25. The physical life consumption rate obtained by calculation is 0.068 / year, indicating that the pipeline life is consumed by about 6.8% per year.

[0140] The physical life consumption rate is input into the degradation state evolution equation constructed in the foregoing, and the data-driven life prediction value is obtained by iterative calculation. The prediction value is based on the historical monitoring data and the established random process model, and considers the randomness and uncertainty of the system. For the oil and gas pipeline in the example, the data-driven life prediction value obtained by iterative calculation is 12.3 years.

[0141] The weighting coefficients of the physical life consumption rate and the data-driven life prediction value are adaptively adjusted according to the prediction error. In the weighting adjustment process, the historical prediction error can be evaluated by using the sliding window method, and the weighting coefficients are dynamically adjusted based on the error size. When the prediction error of the physical model is small, the weight of the physical life consumption rate is increased; when the prediction error of the data-driven model is small, the weight of the data-driven life prediction value is increased. For the oil and gas pipeline in the example, assuming that the average relative error of the physical model in the historical prediction is 7%, and the average relative error of the data-driven model is 5%, the weight of the physical life consumption rate can be set as 0.42, and the weight of the data-driven life prediction value can be set as 0.58.

[0142] The final output of the oil and gas pipeline remaining life prediction value is obtained by weighted combination of the physical life consumption rate and the data-driven life prediction value. For the oil and gas pipeline in the example, the calculated remaining life prediction value is 11.8 years. The prediction result comprehensively considers the physical failure mechanism and the data statistical characteristics, and can provide a reliable basis for the maintenance decision of the oil and gas pipeline.

[0143] In practical application, the pipeline monitoring data can be updated regularly (such as every half year), and the above prediction process can be re-executed to obtain more accurate remaining life prediction results. At the same time, the prediction results can be compared and verified with the periodic detection data, to further optimize the model parameters and the prediction strategy. Through this continuous iteration, the accuracy and reliability of the oil and gas pipeline remaining life prediction can be continuously improved.

[0144] Figure 4 The fatigue damage accumulation and prediction error relationship diagram of the embodiment of the application is as follows:

[0145] The figure shows the prediction error rate of three different models (the present technical solution, the traditional fatigue model and the Miner linear cumulative model) under different fatigue damage accumulations. The horizontal axis represents the fatigue damage accumulation, ranging from 0.1 to 0.8; the vertical axis represents the prediction error rate (%). As can be seen from the figure, with the increase of fatigue damage accumulation, the prediction errors of the three models all show an upward trend, but the prediction error growth of the present technical solution is the most gentle. Specifically, when the fatigue damage accumulation is 0.1, the prediction errors of the three models are 2.3%, 3.8% and 3.1% respectively; when the accumulation increases to 0.4, the errors increase to 3.6%, 6.7% and 5.5% respectively; when the accumulation reaches 0.8, the errors further increase to 5.2%, 10.2% and 8.7% respectively. Overall, the prediction error of the present technical solution at each damage accumulation stage is significantly lower than that of the other two models, with an average error rate about 50% lower than that of the traditional fatigue model and about 40% lower than that of the Miner linear cumulative model, indicating that the present technical solution has better prediction accuracy and stability, especially in the high fatigue damage accumulation interval.

[0146] In an optional implementation, a Wiener random process model is established according to the crack evolution feature vector and the corrosion degradation feature vector, a drift coefficient, a diffusion coefficient and a measurement noise are combined to construct a degradation state evolution equation; the nonlinear compensation term is introduced into the degradation state evolution equation, including:

[0147] Obtaining degradation state observation values at multiple time points, the degradation state observation values including state measurement data and timestamp information;

[0148] Calculating a state observation value difference of adjacent time points, dividing the state observation value difference by a corresponding time interval to obtain a degradation rate; processing the degradation rate by using a maximum likelihood estimation method to obtain a drift coefficient; performing sum of squares operation on a deviation of the degradation rate and the drift coefficient to obtain a diffusion coefficient;

[0149] Performing time domain analysis on the measurement noise in the degradation state observation values, calculating an autocorrelation function of the measurement noise under different time delays; performing Fourier transform on the autocorrelation function to obtain frequency domain characteristics;

[0150] Performing state prediction based on the frequency domain characteristics, combining a difference between a prediction value and an observation value with a Kalman gain to perform dynamic correction, eliminating the influence of the measurement noise to obtain a state estimation value;

[0151] Performing local linearization processing on the state estimation value, calculating a nonlinear approximation term through second-order Taylor expansion; constructing a dynamic compensation term based on the nonlinear approximation term, combining a historical state data sequence and a historical prediction error;

[0152] The degradation state evolution equation is constructed by combining the drift coefficient, the diffusion coefficient, the state estimation value and the dynamic compensation term, and contains a Brownian motion term, a filtered noise term and a nonlinear compensation term, for predicting the degradation state at a future time.

[0153] The crack evolution feature vector and the corrosion degradation feature vector reflecting the structural health state are obtained. These feature vectors are obtained by preprocessing and feature extraction of the raw data collected by the sensor network. For example, for a steel structure member, the feature vector [12.5mm, 0.8mm, 3.2mm] composed of crack length, width and depth parameters can be obtained by an ultrasonic detection system, and the feature vector [-650mV, 25μA / cm 2 ] composed of corrosion potential and corrosion current density parameters can be obtained by an electrochemical sensor.

[0154] Based on the obtained feature vectors, the system obtains degradation state observation values at multiple time points. In actual application, data can be collected once every 24 hours according to a preset time interval. For example, in a certain metal member monitoring, the system records state measurement data for 30 consecutive days to form a data set containing degradation state values and corresponding time stamps. Specific data examples are: state value 0.05 on day 1, state value 0.12 on day 5, state value 0.25 on day 10, state value 0.41 on day 15, state value 0.63 on day 20, state value 0.89 on day 25, and state value 1.22 on day 30.

[0155] The difference between state observation values at adjacent time points is calculated, and the difference is divided by the corresponding time interval to obtain the degradation rate. Taking the above data as an example, the state difference from day 1 to day 5 is 0.07, and the time interval is 4 days, so the degradation rate is 0.0175 / day; the degradation rate from day 5 to day 10 is 0.026 / day; the degradation rate from day 10 to day 15 is 0.032 / day; and so on, to obtain the complete degradation rate sequence [0.0175, 0.026, 0.032, 0.044, 0.052, 0.066].

[0156] The maximum likelihood estimation method is used to process the above degradation rate data, and the drift coefficient is obtained by iterative calculation. In this embodiment, the drift coefficient is calculated to be 0.039 / day by processing the degradation rate sequence. This coefficient reflects the average change trend of the degradation process.

[0157] To calculate the diffusion coefficient, the system performs sum of squares operation on the deviation of the degradation rate and the drift coefficient. Using the aforementioned data, the square of the difference between each degradation rate and the drift coefficient 0.039 is calculated to obtain [(0.0175-0.039) 2 ,(0.026-0.039)2 (0.032-0.039) 2 (0.044-0.039) 2 (0.052-0.039) 2 (0.066-0.039) 2 After summing up and dividing by the number of data points minus one, the diffusion coefficient is 0.0003 / day.

[0158] For the measurement noise in the degradation state observation, the system performs time domain analysis to calculate the autocorrelation function of the measurement noise under different time delays. Taking 100 groups of noise samples collected in an experiment as an example, after extracting the trend by moving average method, the autocorrelation coefficients at time delays of 1, 2, 3, 4, and 5 days are 0.82, 0.65, 0.47, 0.31, and 0.18, respectively.

[0159] The Fourier transform of the autocorrelation function is performed to obtain the frequency domain characteristics of the noise. After transformation, it is found that the noise is mainly concentrated in the low frequency band, and the energy is mainly distributed in the frequency range of 0-0.2 Hz, and the high frequency component has low energy. Based on this frequency domain feature, the parameters of the low-pass filter are designed, and the cutoff frequency is set to 0.25 Hz, and the roll-off rate is 24 dB / octave.

[0160] Based on the above frequency domain characteristics, the state prediction is performed, and the difference between the predicted value and the observed value is combined with the Kalman gain for dynamic correction. In a certain test case, the state prediction value on the 31st day is 1.32, the actual observation value is 1.29, and the difference is 0.03. Using the calculated Kalman gain 0.65, the prediction value is corrected to obtain the state estimation value 1.297, effectively eliminating the influence of measurement noise.

[0161] The state estimation value is subjected to local linearization processing, and the nonlinear approximation term is calculated by second-order Taylor expansion. In the processing, the data of the last 5 days are selected as the reference window, and the calculated second-order differential term is 0.0032, which is used as the characterization of the dynamic characteristics of the nonlinear system.

[0162] Based on the nonlinear approximation term, the dynamic compensation term is constructed based on the historical state data sequence and the historical prediction error. The average prediction error of the last 10 days is 0.018, combined with the second-order approximation term and the current state value, the calculated dynamic compensation term is 0.0076.

[0163] The degradation state evolution equation is constructed by combining the drift coefficient 0.039 / day, the diffusion coefficient 0.0003 / day, the state estimation value 1.297, and the dynamic compensation term 0.0076. The equation includes a Brownian motion term, a filtered noise term, and a nonlinear compensation term, and can more accurately predict the degradation state at a future time. For example, the degradation state value at the 40th day predicted by the equation is 1.78, and compared with the actual observation value 1.81, the prediction error is only 1.7%, which is significantly better than the prediction error 9.3% of the traditional linear model.

[0164] In an optional embodiment, generating a pipeline safety state evaluation report and a pipeline maintenance decision suggestion according to the optimized safety state evaluation accuracy includes:

[0165] The optimized safety state evaluation accuracy includes evaluation index accuracy and evaluation model confidence; a correction coefficient is calculated according to the evaluation model confidence, a dynamic correction weight is obtained by linearly correcting an initial weight of the evaluation index based on the correction coefficient; and a safety state comprehensive score is obtained by weighted summation of the evaluation index and the dynamic correction weight;

[0166] K-means clustering analysis is performed on historical evaluation data to obtain an initial division standard of risk levels; the evaluation model confidence is used as an adaptive factor, and an exponential sliding weighted average method is used to update the initial division standard; and the safety state comprehensive score is matched with the updated division standard to determine a risk level division standard;

[0167] A maintenance priority evaluation model is constructed based on the risk level division standard, and risk degree, maintenance difficulty, and time urgency are quantified and input into the maintenance priority evaluation model; an evaluation factor weight is set by using the evaluation index accuracy, and a maintenance priority score is calculated by weighted summation;

[0168] The safety state comprehensive score, the risk level division standard, and the maintenance priority score are written in a preset evaluation report template in order of importance to generate a pipeline safety state evaluation report;

[0169] Evaluation information is identified from the pipeline safety state evaluation report by using a word extraction algorithm, the evaluation information is matched with a maintenance decision rule library, and a pipeline maintenance decision suggestion including maintenance content, maintenance timing, and resource allocation is generated.

[0170] The optimized safety state evaluation accuracy includes two key parameters: evaluation index accuracy and evaluation model confidence. In practical applications, the evaluation index accuracy is usually expressed as a percentage, such as 92.3% for corrosion detection accuracy and 88.7% for structural strength evaluation accuracy. The evaluation model confidence reflects the model's confidence in its own prediction results, such as 0.85 for the corrosion prediction model confidence and 0.78 for the structural safety model confidence.

[0171] When calculating the correction coefficient based on the evaluation model confidence, a nonlinear mapping function is used. For example, when the confidence of the pipeline corrosion detection model is 0.85, the correction coefficient is 1.12 obtained through the preset mapping relationship. When the structural safety model confidence is 0.78, the corresponding correction coefficient is 0.95. These correction coefficients are used to adjust the initial weights of the evaluation indicators. Assuming that the initial weight of the corrosion indicator is 0.3 and the initial weight of the structural strength indicator is 0.25, the dynamically corrected weights are 0.336 (0.3 x 1.12) and 0.2375 (0.25 x 0.95), respectively.

[0172] In the safety state comprehensive score calculation process, each evaluation indicator is multiplied by its corresponding dynamic correction weight and then summed. For example, the corrosion indicator score for a certain pipeline section is 78, the structural strength indicator score is 85, and the environmental risk indicator score is 65. The corresponding dynamic correction weights are 0.336, 0.2375, and 0.22, respectively. The safety state comprehensive score is 75.86.

[0173] When performing K-means clustering analysis on historical evaluation data, 8000 pipeline evaluation records within the past three years are selected as sample data, and the clustering number K is set to 5, representing five risk levels: extremely high risk, high risk, medium risk, low risk, and extremely low risk. Through clustering analysis, the initial division standard is obtained: extremely high risk (0-60 points), high risk (60-70 points), medium risk (70-80 points), low risk (80-90 points), and extremely low risk (90-100 points).

[0174] When updating the initial division standard using the evaluation model confidence as an adaptive factor, the exponential sliding weighted average method is used to update the initial division standard when the overall model confidence is 0.82. The calculation process includes multiplying the initial threshold by a weight factor, which is exponentially adjusted based on the confidence. The updated risk level division standard is: extremely high risk (0-58 points), high risk (58-69 points), medium risk (69-81 points), low risk (81-91 points), and extremely low risk (91-100 points).

[0175] The safety state comprehensive score is matched with the updated classification standard, for example, when the comprehensive score of a certain pipeline section is 75.86, the corresponding risk level is "medium risk".

[0176] When constructing the maintenance priority evaluation model based on the risk level classification standard, three key factors are quantified: risk degree, maintenance difficulty, and time urgency. The risk degree is directly determined by the risk level, for example, a medium risk corresponds to a score of 60; the maintenance difficulty takes into account human resource demand, equipment requirements, and environmental constraints, for example, the maintenance difficulty score of a certain pipeline section is 75; the time urgency is based on the failure development rate and the impact range, for example, the time urgency score of a certain pipeline section is 80.

[0177] When setting the evaluation factor weights using the evaluation indicator accuracy, the higher the accuracy, the greater the corresponding weight. For example, the risk degree evaluation accuracy is 92%, the maintenance difficulty evaluation accuracy is 85%, and the time urgency evaluation accuracy is 88%, so the corresponding weights are 0.35, 0.3, and 0.35 respectively. The maintenance priority score is calculated by weighted summation to be 71.25.

[0178] When generating the pipeline safety state evaluation report, the safety state comprehensive score (75.86), the risk level classification result (medium risk), and the maintenance priority score (71.25) are written in the preset evaluation report template in order of importance. The report template includes four parts: basic information area, evaluation result area, risk analysis area, and suggestion area. The generated report uses a structured format, which is convenient for subsequent analysis and decision-making.

[0179] When identifying evaluation information from the evaluation report using the word extraction algorithm, a rule-based word extraction method is used to identify key information such as "medium risk", "corrosion index 78", "maintenance priority 71.25", etc. These evaluation information is matched with the maintenance decision rule base, which contains maintenance strategies for different risk levels and different fault types. For example, for a pipeline section with "medium risk" and "low corrosion index", the matched maintenance decision is "perform corrosion coating repair within three months, configure standard corrosion tool set and 3 professional technicians".

[0180] The final generated pipeline maintenance decision suggestion includes maintenance content (corrosion coating repair), maintenance timing (within three months), and resource allocation (standard corrosion tool set and 3 professional technicians), providing intuitive and clear decision support for pipeline operation management departments.

[0181] Figure 5 The performance comparison column chart of the pipeline safety state evaluation and maintenance decision method of the embodiment of the present application is as follows:

[0182] The figure shows the comparison results of three different methods (the method, the traditional evaluation method and the static weight method) on five key performance indicators. The method has obvious advantages in all evaluation indicators: the evaluation accuracy reaches 92.7%, which is 14.4 and 8.2 percentage points higher than the traditional method (78.3%) and the static weight method (84.5%) respectively; in terms of decision rationality, the method reaches 89.5%, while the traditional method is 72.6% and the static weight method is 79.2%; in terms of risk identification rate, the method reaches the highest level of 94.3%, while the traditional method is 81.5% and the static weight method is 87.1%; in terms of maintenance priority accuracy, the method reaches 91.8%, which is significantly higher than the traditional method of 75.2% and the static weight method of 81.3%; in terms of resource allocation efficiency, the method still maintains a high level of 88.6%, while the traditional method and the static weight method are 70.8% and 76.9% respectively. Overall, the method improves by 15-20 percentage points on average in various indicators, especially in the risk identification rate and the maintenance priority accuracy, which are the two key indicators, and the improvement is most significant, fully proving the advancement and practical value of the method in device maintenance decision support. The method not only improves the accuracy of evaluation, but also significantly improves the scientificity of decision and the efficiency of resource utilization.

[0183] In a second aspect, the embodiment of the present application provides an oil and gas pipeline full life cycle safety state intelligent monitoring system, comprising:

[0184] A first unit is configured to collect oil and gas pipeline operation state data, wherein the oil and gas pipeline operation state data comprises pipeline internal pressure data, pipeline internal temperature data and pipeline surrounding environment data.

[0185] A second unit is configured to establish an oil and gas pipeline digital twin model according to the oil and gas pipeline operation state data, and construct a multi-dimensional feature vector based on the oil and gas pipeline digital twin model, wherein the multi-dimensional feature vector comprises a pipeline structure feature vector, a pipeline environment feature vector and a pipeline historical fault feature vector.

[0186] A third unit is configured to perform feature fusion on the multi-dimensional feature vector to obtain a fused feature vector, and construct an oil and gas pipeline fault evolution prediction model based on the fused feature vector and a graph neural network.

[0187] A fourth unit is configured to calculate an oil and gas pipeline residual life prediction value according to an output result of the oil and gas pipeline fault evolution prediction model.

[0188] A fifth unit is configured to continuously optimize the safety state evaluation accuracy by means of online iterative training based on the oil and gas pipeline remaining life prediction value and in combination with a pipeline safety state evaluation rule library, and generate an oil and gas pipeline safety state evaluation report and pipeline maintenance decision suggestions according to the optimized safety state evaluation accuracy.

[0189] In a third aspect, the present application provides an electronic device, comprising:

[0190] a processor;

[0191] a memory for storing processor-executable instructions;

[0192] The processor is configured to invoke the instructions stored in the memory to execute the method described above.

[0193] In a fourth aspect, the present application provides a computer-readable storage medium having stored thereon computer program instructions, which, when executed by a processor, implement the method described above.

[0194] The present application can be a method, apparatus, system and / or computer program product. The computer program product can include a computer-readable storage medium having stored thereon computer-readable program instructions that, when executed by a computer, cause the computer to carry out various aspects of the present application.

[0195] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for intelligent monitoring of safety state of oil and gas pipelines throughout their life cycle, characterized in that, The method comprises the following steps: Collecting oil and gas pipeline operation state data, including pipeline internal pressure data, pipeline internal temperature data and pipeline peripheral environment data; According to the oil and gas pipeline operation state data, an oil and gas pipeline digital twin model is established; based on the oil and gas pipeline digital twin model, a multi-dimensional feature vector is constructed, including a pipeline structure feature vector, a pipeline environment feature vector and a pipeline historical fault feature vector; Feature fusion is performed on the multi-dimensional feature vector to obtain a fused feature vector; an oil and gas pipeline fault evolution prediction model is constructed based on the fused feature vector combined with a graph neural network; According to the output result of the oil and gas pipeline fault evolution prediction model, the residual life prediction value of the oil and gas pipeline is calculated, including: The output result of the oil and gas pipeline fault evolution prediction model is decoupled according to the failure mechanism to construct a crack evolution feature vector and a corrosion degradation feature vector; According to the crack evolution feature vector and the corrosion degradation feature vector, a Wiener random process model is established, and a drift coefficient, a diffusion coefficient and a measurement noise are combined to construct a degradation state evolution equation; a nonlinear compensation term is introduced into the degradation state evolution equation; Based on the Wiener random process model, the fatigue damage accumulation is calculated, and the corrosion rate in the corrosion degradation feature vector and the crack evolution feature vector are calculated; the fatigue damage accumulation, the corrosion rate and the plastic strain are weighted and combined to obtain a physical life consumption rate; The physical life consumption rate is input into the degradation state evolution equation to obtain a data-driven life prediction value; the weight coefficients of the physical life consumption rate and the data-driven life prediction value are adjusted according to the prediction error, and the residual life prediction value of the oil and gas pipeline is output; Based on the residual life prediction value of the oil and gas pipeline, the safety state evaluation accuracy is continuously optimized through online iterative training combined with a pipeline safety state evaluation rule library; according to the optimized safety state evaluation accuracy, an oil and gas pipeline safety state evaluation report and a pipeline maintenance decision suggestion are generated.

2. The method of claim 1, wherein, According to the oil and gas pipeline operation state data, an oil and gas pipeline digital twin model is established; Based on the oil and gas pipeline digital twin model, a multi-dimensional feature vector is constructed, including: Based on the oil and gas pipeline operation state data, a physical-virtual mapping matrix is constructed, which contains a plurality of mapping relationship coefficients; According to the physical-virtual mapping matrix, an oil and gas pipeline digital twin model is established to realize virtual simulation of the oil and gas pipeline operation state data; Input variables and model parameters are obtained, and a state update equation is constructed according to the input variables and the model parameters, and the oil and gas pipeline digital twin model is dynamically updated through the state update equation; The multi-dimensional features are extracted based on the oil and gas pipeline digital twin model, including: a pressure fluctuation feature is calculated according to a variance of a pressure sampling value and an average pressure, a temperature gradient feature is calculated according to a partial derivative of a temperature field in a three-dimensional space, and an environmental influence factor is calculated according to weight coefficients of various environmental parameters, wherein the weight coefficients are determined through historical data analysis; The pressure fluctuation feature, the temperature gradient feature and the environmental influence factor are fused to construct a multi-dimensional feature vector of the oil and gas pipeline.

3. The method of claim 1, wherein, The multi-dimensional feature vector is fused to obtain a fused feature vector; and an oil and gas pipeline fault evolution prediction model is constructed based on the fused feature vector and a graph neural network, including: The multi-dimensional feature vector is mapped into a state feature by using a state mapping weight matrix; a time feature is constructed based on a timestamp, a period encoding and a feature change rate; a space feature is constructed according to a position coordinate, an azimuth angle and a feature space gradient; and the state feature, the time feature and the space feature are combined to form a spatio-temporal knowledge node; A graph feature is extracted based on the spatio-temporal knowledge node through node feature aggregation operation, the node feature aggregation operation adopts attention weight to combine adjacent node features, and the attention weight is calculated through spatio-temporal attention; The graph feature is processed by using the graph neural network to obtain an enhanced feature, and the enhanced feature and the multi-dimensional feature vector are combined to obtain a fused feature vector; Causal chain discovery is performed on the fused feature vector, including: a causal relationship between features is calculated through conditional independence test, and a causal strength is obtained based on a causal gradient; and a causal attention mechanism is constructed according to the causal relationship and the causal strength; The causal attention mechanism and the fused feature vector are combined to construct an oil and gas pipeline fault evolution prediction model through the graph neural network.

4. The method of claim 3, wherein, A causal relationship between features is calculated through conditional independence test, and a causal strength is obtained based on a causal gradient; and a causal attention mechanism is constructed according to the causal relationship and the causal strength, including: Time sequence feature data and state feature data are acquired, conditional entropy of the time sequence feature data and the state feature data is calculated, conditional independence test is performed based on the conditional entropy to obtain a conditional independence test result between features; Conditional probability modeling is performed on the time sequence feature data, a probability dependency relationship between features is calculated by using a Bayesian network, and a causal relationship between features is determined based on a result of the conditional independence test and the probability dependency relationship; A causal gradient between features is calculated based on the causal relationship, a direct causal strength is obtained according to the causal gradient, an indirect causal strength between features is calculated along a causal transmission path, and a comprehensive causal strength is obtained by combining the direct causal strength and the indirect causal strength; The comprehensive causal strength and feature representation are combined to generate a causal query vector and a causal key-value vector; a causal attention matrix is constructed based on the causal query vector and the causal key-value vector; and the feature representation is reconstructed according to the causal attention matrix to obtain a causal attention mechanism.

5. The method of claim 1, wherein, A Wiener stochastic process model is established according to the crack evolution feature vector and the corrosion degradation feature vector, and a degradation state evolution equation is constructed by combining a drift coefficient, a diffusion coefficient and measurement noise; The non-linear compensation term is introduced into the degradation state evolution equation, including: Obtaining degradation state observation values at multiple time points, the degradation state observation values including state measurement data and timestamp information; Calculating state observation value differences between adjacent time points, dividing the state observation value differences by corresponding time intervals to obtain degradation rates, and processing the degradation rates by using a maximum likelihood estimation method to obtain a drift coefficient; and performing sum-of-squares operation on deviations between the degradation rates and the drift coefficient to obtain a diffusion coefficient; Performing time domain analysis on measurement noise in the degradation state observation values, calculating autocorrelation functions of the measurement noise under different time delays, and performing Fourier transform on the autocorrelation functions to obtain frequency domain characteristics; Performing state prediction based on the frequency domain characteristics, combining a difference between a prediction value and an observation value with a Kalman gain for dynamic correction to eliminate the influence of measurement noise and obtain a state estimation value; Performing local linearization processing on the state estimation value, calculating a non-linear approximation term by using a second-order Taylor expansion, and constructing a dynamic compensation term based on the non-linear approximation term, historical state data sequences and historical prediction errors; Combining the drift coefficient, the diffusion coefficient, the state estimation value and the dynamic compensation term to construct a degradation state evolution equation, the degradation state evolution equation including a Brownian motion term, a filtered noise term and a non-linear compensation term, and used for predicting a degradation state at a future time.

6. The method of claim 1, wherein, Generating a pipeline safety state evaluation report and a pipeline maintenance decision suggestion according to the optimized safety state evaluation accuracy, including: The optimized safety state evaluation accuracy includes evaluation index accuracy and evaluation model confidence; a correction coefficient is calculated according to the evaluation model confidence, a dynamic correction weight is obtained by linearly correcting an initial weight of an evaluation index based on the correction coefficient, and a safety state comprehensive score is obtained by weighted summation of the evaluation index and the dynamic correction weight; K-means clustering analysis is performed on historical evaluation data to obtain an initial division standard of a risk level; the evaluation model confidence is used as an adaptive factor, and the initial division standard is updated by using an exponential sliding weighted average method; and the safety state comprehensive score is matched with the updated division standard to determine a risk level division standard; A maintenance priority evaluation model is constructed based on the risk level division standard, a risk degree, a maintenance difficulty and a time urgency are quantified and input into the maintenance priority evaluation model, an evaluation factor weight is set by using the evaluation index accuracy, and a maintenance priority score is calculated by weighted summation; The safety state comprehensive score, the risk level division standard and the maintenance priority score are written in a preset evaluation report template in order of importance to generate a pipeline safety state evaluation report. The evaluation information is identified from the pipeline safety state evaluation report by using a word extraction algorithm, the evaluation information is matched with a maintenance decision rule library, and a pipeline maintenance decision suggestion including maintenance content, maintenance timing and resource allocation is generated.

7. An intelligent monitoring system for safety condition of oil and gas pipeline throughout its life cycle, for implementing the method according to any one of the preceding claims 1-6, characterized in that, The method comprises the following steps: A first unit is configured to collect oil and gas pipeline operation state data, wherein the oil and gas pipeline operation state data comprises pipeline internal pressure data, pipeline internal temperature data and pipeline surrounding environment data. A second unit is configured to establish an oil and gas pipeline digital twin model according to the oil and gas pipeline operation state data, and construct a multi-dimensional feature vector based on the oil and gas pipeline digital twin model, wherein the multi-dimensional feature vector comprises a pipeline structure feature vector, a pipeline environment feature vector and a pipeline historical fault feature vector. A third unit is configured to perform feature fusion on the multi-dimensional feature vector to obtain a fused feature vector, and construct an oil and gas pipeline fault evolution prediction model based on the fused feature vector and a graph neural network. A fourth unit is configured to calculate an oil and gas pipeline remaining life prediction value according to an output result of the oil and gas pipeline fault evolution prediction model. A fifth unit is configured to continuously optimize safety state evaluation accuracy through online iterative training based on the oil and gas pipeline remaining life prediction value and in combination with a pipeline safety state evaluation rule library. An oil and gas pipeline safety state evaluation report and a pipeline maintenance decision suggestion are generated according to the optimized safety state evaluation accuracy.

8. An electronic device, comprising: The method comprises the following steps: A processor; A memory for storing processor-executable instructions; The processor is configured to invoke the instructions stored in the memory to execute the method of any one of claims 1 to 6.

9. A computer-readable storage medium having stored thereon computer program instructions, wherein, The computer program instructions are executed by the processor to implement the method of any one of claims 1 to 6.

Images