A method and system for managing full life cycle data of an oil and gas station pipeline

Abstract

The application relates to the technical field of pipeline data management, and discloses an oil and gas station pipeline whole life cycle data management method and system, which comprises the following steps: extracting pipeline geographic information, extracting spatial correlation features through a graph neural network, combining pressure time sequence data to determine a pipeline conduction mode through a preset long short-term memory network; quantitatively evaluating the conduction mode according to historical maintenance and abnormal data, extracting topological nodes if the value exceeds a threshold, combining a preset attention mechanism to obtain a space-time coupling vector; extracting node correlation features from the space-time coupling, quantitatively fusing to obtain a heterogeneous node set, generating an abnormal atlas through graph convolution; extracting risk indexes, if upstream conduction effects are shown, obtaining a damage prediction sequence through a preset recurrent neural network, fusing topological and time sequence data to determine a health score vector, analyzing the matching degree of the health score vector with historical records, and generating a risk traceability warning if the matching degree is lower than a threshold. The method can realize accurate positioning and evaluate the overall health state of a pipeline system.

Description

Technical Field

[0001] This invention relates to the field of pipeline data management technology, and in particular to a method and system for managing the entire lifecycle data of oil and gas station pipelines. Background Technology

[0002] In the oil and gas energy industry, the safe and stable operation of pipeline networks at stations is a core infrastructure for ensuring the continuity of energy transmission and public safety. Accurate health status assessment and full lifecycle data management are crucial aspects of operation and maintenance management. Achieving deep integration and dynamic risk prediction of multi-source heterogeneous data under complex pipeline network structures is a core requirement for current oil and gas station pipeline operation and management.

[0003] In one existing technology, pipeline monitoring is carried out through regular manual inspections and threshold alarm modes of isolated sensors. The collected pressure and temperature data are compared with preset safety thresholds. When the data exceeds the threshold, an independent alarm and maintenance scheduling are performed on the single pipe section or equipment to achieve monitoring of the operating status of single-point equipment.

[0004] However, existing technologies only conduct independent monitoring and threshold judgment for single devices or isolated pipe sections, without incorporating big data processing technology to systematically mine and analyze pipeline network operation data. This makes it impossible to deeply integrate the static topological connections of the pipeline network with the dynamic temporal changes of sensor data, hindering the capture of the spatial transmission effects of pressure fluctuations in complex pipeline structures and their upstream and downstream correlations. Consequently, during pipeline anomaly diagnosis, it becomes impossible to distinguish between normal operational fluctuations and structural damage signs, and to trace the source of slow-down phenomena in the spatial network and their temporal accumulation patterns. When multiple points in the pipeline network exhibit parallel anomalies, some pipe sections are in a high-risk operating state due to the superposition of risk transmission, while other sections are in monitoring blind spots due to data fragmentation, resulting in an imbalance in load distribution and risk control across the overall energy transmission network.

[0005] In summary, existing technologies result in low overall operational stability and energy efficiency of oil and gas station pipeline systems, making it difficult to achieve accurate diagnosis and early warning from anomaly detection to tracing the root cause of risks. Summary of the Invention

[0006] This invention provides a method and system for managing the entire lifecycle data of oil and gas station pipelines, so as to achieve accurate positioning and assessment of the overall health status of the pipeline system.

[0007] Firstly, in order to solve the above-mentioned technical problems, the present invention provides a method for full life-cycle data management of oil and gas station pipelines, including: Obtain pipeline geographic information entities, extract node coordinates from the pipeline geographic information entities, construct topological relationships, and extract spatial network association features from the topological relationships through a preset graph neural network; Based on the spatial network association characteristics, pressure time series data is obtained, and the pressure time series data is fused using a preset long short-term memory network to determine the pipeline conduction mode; Historical fault data is acquired, maintenance frequency features and fluctuation attenuation features are extracted from the historical fault data, and the pipeline conduction mode is quantitatively evaluated based on the maintenance frequency features and fluctuation attenuation features. If the quantitative evaluation value exceeds the preset abnormal threshold of the conduction mode, spatial topology nodes are extracted from the pipeline conduction mode, and pipeline coupling vectors are extracted from the spatial topology nodes through a preset attention mechanism. Node association features are extracted from the pipeline coupling vector, and the maintenance frequency features and the fluctuation attenuation features are concatenated with the node association features to obtain a heterogeneous node set. Based on the heterogeneous node set, graph convolution operation is performed on the pipeline conduction mode and the spatial topology nodes to determine the pipeline anomaly map. Risk tracing indicators are extracted from the pipeline anomaly map. If the risk tracing indicators show an upstream transmission effect, the evolution trend is simulated and evaluated through a preset recurrent neural network to obtain a pipeline damage prediction sequence. Based on the pipeline damage prediction sequence, the pipeline health status score is obtained by fusing the topological relationship with the pressure time series data. The matching degree between the pipeline conduction mode and the historical fault data is analyzed based on the pipeline health status score. If the matching degree is lower than the preset matching threshold, a pipeline risk alarm is generated.

[0008] Secondly, the present invention provides an oil and gas station pipeline full life cycle data management device for the aforementioned oil and gas station pipeline full life cycle data management method, comprising: The geographic construction module is used to acquire pipeline geographic information entities, extract node coordinates from the pipeline geographic information entities, construct topological relationships, and extract spatial network association features from the topological relationships through a preset graph neural network. The pressure transmission module is used to acquire pressure time series data based on the spatial network association characteristics, and to fuse the pressure time series data using a preset long short-term memory network to determine the pipeline transmission mode. The spatiotemporal coupling module is used to acquire historical fault data, extract maintenance frequency features and fluctuation attenuation features from the historical fault data, and quantitatively evaluate the pipeline conduction mode based on the maintenance frequency features and fluctuation attenuation features. If the quantitative evaluation value exceeds the preset conduction mode abnormality threshold, spatial topology nodes are extracted from the pipeline conduction mode, and pipeline coupling vectors are extracted from the spatial topology nodes through a preset attention mechanism. An anomaly graph module is used to extract node association features from the pipeline coupling vector, and to concatenate the maintenance frequency features and the fluctuation attenuation features with the node association features to obtain a heterogeneous node set. Based on the heterogeneous node set, a graph convolution operation is performed on the pipeline conduction mode and the spatial topology nodes to determine the pipeline anomaly graph. The damage prediction module is used to extract risk tracing indicators from the pipeline anomaly map. If the risk tracing indicators show an upstream transmission effect, the evolution trend is evaluated by simulating the evolution trend through a preset recurrent neural network to obtain a pipeline damage prediction sequence. The health scoring module is used to obtain a pipeline health status score by fusing the topological relationship and the pressure time series data based on the pipeline damage prediction sequence. The risk warning module is used to analyze the matching degree between the pipeline conduction mode and the historical fault data based on the pipeline health status score. If the matching degree is lower than a preset matching threshold, a pipeline risk alarm is generated.

[0009] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention uses graph convolution operations to process heterogeneous node sets to determine the pipeline anomaly map, and combines it with a pre-set recurrent neural network to evaluate the evolution trend and generate a pipeline damage prediction sequence. The graph convolution layer aggregates the dynamic states of neighboring nodes to ensure the continuity and logical consistency of anomaly features in the complex topological space, while the pre-set recurrent neural network generates intermediate simulation sequences to predict the evolution trajectory of current anomaly features in future cycles. This process can accurately locate the coordinates of the risk source and predict the damage level, providing a high-dimensional quantitative decision-making basis for the preventive maintenance of pipelines in the station.

[0010] (2) This invention extracts spatial network correlation features through graph neural network and uses a preset long short-term memory network to process pressure time series data to determine pipeline conduction mode. It solves the limitation of single-point monitoring by isolated sensors, can accurately identify the real attenuation law and propagation trend of pressure waves on a specific path, and improves the capture accuracy of small abnormal signals in complex pipeline network environment.

[0011] (3) This invention extracts maintenance frequency features and fluctuation attenuation features from historical fault data, aligns historical operating data with real-time transmission patterns using a preset attention mechanism to obtain a pipeline coupling vector, and uses the preset attention mechanism to weight the time alignment matrix to achieve matching and fusion of historical dependencies and real-time features. This processing method places instantaneous physical fluctuations in the context of the entire pipeline life cycle for verification, realizes quantitative assessment of pipeline health status, and improves the objectivity and reliability of the judgment results. Attached Figure Description

[0012] Figure 1This is a schematic diagram of the process for managing the full lifecycle data of oil and gas station pipelines provided in the first embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of the oil and gas station pipeline full life cycle data management system provided in the second embodiment of the present invention. Detailed Implementation

[0013] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0014] Reference Figure 1 The first embodiment of the present invention provides a full lifecycle data management system for oil and gas station pipelines, including the following steps: S11, Obtain pipeline geographic information entities, extract node coordinates from the pipeline geographic information entities, construct topological relationships, and extract spatial network association features from the topological relationships through a preset graph neural network; S12, Based on the spatial network association characteristics, pressure time series data is obtained, and the pressure time series data is fused using a preset long short-term memory network to determine the pipeline conduction mode; S13, acquire historical fault data, extract maintenance frequency features and fluctuation attenuation features from the historical fault data, and quantitatively evaluate the pipeline conduction mode based on the maintenance frequency features and fluctuation attenuation features. If the quantitative evaluation value exceeds the preset abnormal threshold of the conduction mode, extract spatial topology nodes from the pipeline conduction mode, and extract pipeline coupling vectors from the spatial topology nodes through a preset attention mechanism. S14, extract node association features from the pipeline coupling vector, and concatenate the maintenance frequency features and the fluctuation attenuation features with the node association features to obtain a heterogeneous node set. Based on the heterogeneous node set, perform graph convolution operation on the pipeline conduction mode and the spatial topology nodes to determine the pipeline anomaly map. S15, extract risk tracing indicators from the pipeline anomaly map. If the risk tracing indicators show an upstream transmission effect, then use a preset recurrent neural network to simulate the evolution trend assessment to obtain a pipeline damage prediction sequence. S16, Based on the pipeline damage prediction sequence, the topological relationship and the pressure time series data are fused to obtain the pipeline health status score; S17. Analyze the matching degree between the pipeline conduction mode and the historical fault data based on the pipeline health status score. If the matching degree is lower than a preset matching threshold, generate a pipeline risk alarm.

[0015] It should be noted that all the preset models involved in this invention are trained based on historical station operation data. The sources of historical data include the following three categories: The first category is pipeline geographic information data, which is obtained from the station GIS database. It includes pipeline node coordinates, pipe segment lengths, connection relationships, and node attribute information, and is used to construct topological relationships and adjacency matrices.

[0016] The second category is sensor time-series data, acquired from the station's data acquisition and monitoring control system. This includes continuous monitoring data of parameters such as pressure, temperature, flow rate, and vibration at each monitoring point over the past three years, with a sampling frequency of twice per second. Pressure data serves as the core input, while other parameters are used as auxiliary features.

[0017] The third category is historical maintenance and fault records, extracted from the station maintenance management system. These records include the time, location, and type of each maintenance operation, as well as tagged information for each fault event, such as fault type, occurrence time, impact range, and repair measures. Fault types include minor leaks, sudden leaks, valve failures, and corrosion perforation, which are used to provide supervisory signals for model training.

[0018] Before model training, the above data was preprocessed. For sensor time-series data, a sliding window was used to divide the samples, with a window length of 100 time steps (corresponding to 50 seconds) and a step size of 20 time steps (corresponding to 10 seconds), with each window serving as a training sample. For missing data, linear interpolation was used to complete the data. For pressure time-series data, the minmax normalization method was used to map the values ​​to the interval between 0 and 1. For fault records, each sample was labeled with a fault type based on the correspondence between the fault occurrence time and the sensor data window.

[0019] In step S11, the pipeline geographic information entity is obtained, the node coordinates in the pipeline geographic information entity are extracted, a topological relationship is constructed, and spatial network association features are extracted from the topological relationship through a preset graph neural network, including: Obtain the pipeline geographic information entity, extract the node coordinates in the pipeline geographic entity, and construct the topological relationship; Generate an adjacency matrix and an initial feature matrix based on the aforementioned topological relationships; The adjacency matrix and the initial feature matrix are input into a preset graph neural network for feature aggregation to obtain deep node features; If the dimension of the deep node feature is lower than a preset dimension threshold, then the deep node feature is subjected to dimensionality up-mapping to obtain the target dimension feature, and spatial network association features are extracted from the target dimension feature.

[0020] In this invention, the station's GIS geographic information database is accessed to identify and extract the latitude and longitude data of key physical locations of the pipeline. For example, the intersection of the main pipeline and branch lines, the locations of pressure regulating valves at all levels, and the installation points of pressure sensors are selected as nodes. The coordinates of node A are obtained as (116.39, 39.91) and the coordinates of node B are obtained as (116.40, 39.92). These coordinates form the basis of spatial topology analysis.

[0021] It is worth noting that an adjacency matrix A is constructed based on the actual physical connection between the pipelines. If there is a direct pipeline connection between node 1 and node 2, the corresponding element in the matrix is ​​set to 1; otherwise, it is 0. Simultaneously, real-time operating data from each node collected by sensors is retrieved to generate an initial feature matrix X. The feature vector of each node includes real-time pressure value, wall thickness reduction, instantaneous flow rate, and historical average corrosion rate. For example, the feature vector of node 1 can be specifically quantified as [0.8MPa, 20mm, 150m³ / h, 0.02mm / a].

[0022] In this invention, the adjacency matrix and the initial feature matrix are input into a preset graph neural network for feature aggregation to obtain deep node features. The preset graph neural network is used to extract spatial network association features from pipeline topology relationships. For example, node A may be influenced by neighboring nodes B and C, and its features will combine the attributes and connection strength of both to form a deep node feature representation.

[0023] This network employs a three-layer graph convolutional structure, with each layer containing a graph convolutional kernel. The input and output feature dimensions are both 64, and ReLU activation is used between layers. Training data is a combination of pipeline geographic information data and sensor time-series data. For each time window, the adjacency matrix and node feature matrix (containing features such as pressure, temperature, and flow rate) are used as inputs, and the fault labels of each node within that window are used as supervision signals. The training objective is node classification, i.e., identifying whether each node is in an abnormal state. The loss function is cross-entropy loss, the optimization algorithm is Adam, the learning rate is set to 0.001, and the training runs for 200 epochs. Training stops when the validation set loss does not decrease for 10 consecutive epochs. After training, the weight parameters of the graph convolutional layers are fixed and used as a pre-defined graph neural network for spatial feature extraction during the method execution phase. During model execution, the adjacency matrix and initial feature matrix of the current pipeline network are used as inputs. The spatial dependencies between nodes are encoded through the inter-layer weight parameters, ultimately outputting deep node feature vectors that characterize the upstream and downstream transmission effects and spatial coupling features of the pipeline network.

[0024] Finally, if the dimension of the deep node features is lower than a preset dimension threshold, the deep node features are subjected to dimensionality upscaling to ensure information integrity. Here, the preset dimension threshold is an empirical value pre-set based on the complexity of the pipeline network, for example, set to 64 dimensions. If the deep features output by the model are only 32 dimensions, a preset fully connected linear transformation layer is invoked to extend the feature vector to the target dimension through nonlinear mapping in the high-dimensional space. Extracting spatial network correlation features from the target dimension features yields a static connection representation of the pipeline system.

[0025] Based on the core requirements of pipeline network analysis, feature components related to node connection strength, pipeline transmission efficiency, upstream and downstream dependence, and key node coupling are extracted from the target dimension features to form an initial spatial correlation feature subset. For the feature components in this subset, components belonging to the same spatial correlation dimension are normalized and combined into a multi-dimensional feature vector, rather than compressed into a single value. For example, for the overall spatial dependence of a node, a three-dimensional sub-vector composed of upstream dependence, downstream dependence, and peer dependence can be constructed as the spatial dependence feature representation of the node. All aggregated spatial correlation features are organized according to the pipeline network structure hierarchy: taking nodes as the basic unit, the spatial correlation feature vectors of each node are arranged in order of pipeline segment connection relationship to form a node-level feature sequence; then, taking the entire pipeline network as the unit, all node feature sequences are combined into a pipeline network-level spatial network correlation feature matrix. Finally, minmax normalization is performed on each feature component in this matrix to uniformly map it to the [0,1] interval, resulting in standardized spatial network correlation features.

[0026] In step S12, the step of acquiring pressure time-series data based on the spatial network association characteristics and fusing the pressure time-series data using a preset long short-term memory network to determine the pipeline conduction mode includes: Based on the spatial network association characteristics, key monitoring nodes and static connections are located, and pressure time-series data on the key monitoring nodes are obtained. The pressure time series data is iteratively calculated using a preset long short-term memory network to obtain fused time series features; A dynamic propagation path matrix is ​​constructed by combining the fusion time-series features and the static connection, and the pressure fluctuation transmission vector is extracted from the dynamic propagation path matrix. If the dimension of the pressure fluctuation transmission vector is higher than a preset dimension threshold, then the pressure fluctuation transmission vector is subjected to dimension reduction projection to obtain a simplified transmission path vector. The simplified transmission path vector is matched with the spatial network association features, and the wave pattern classification result is obtained based on the matching result. The information from the dynamic propagation path matrix is ​​fused using graph convolution operations on the fluctuation pattern classification results to determine the pipeline conduction mode.

[0027] In this invention, based on the spatial network correlation features extracted in the early stage, key monitoring nodes can be located and pressure time series data on these nodes can be obtained. Pressure data streams from sensors at each monitoring point in the station can be obtained in real time, forming pressure time series data arranged according to time step.

[0028] A pre-defined Long Short-Term Memory (LSTM) network is used to fuse pressure time-series data to determine pipeline conduction patterns. This network employs a two-layer LSTM structure, with 128 hidden units in the first layer and 64 hidden units in the second. The input sequence length is 100 time steps, and the output is a fused time-series feature vector for each time step. A fully connected layer is added after the LSTM layer to map the time-series features to a 32-dimensional feature vector. Training data comes from pressure monitoring sequences in sensor time-series data. Pressure data from 100 consecutive time steps are used as the input sequence, and the corresponding fault label is used as the supervision signal. The training objective is a sequence classification task, i.e., identifying the operating condition type of the sequence (normal operation, pressure regulation, minor leak, sudden leak, etc.). The loss function is cross-entropy loss, the optimization algorithm is Adam, the learning rate is set to 0.001, the training epochs are 150, and an early stopping strategy is used to prevent overfitting. After training, the weight parameters of the LSTM layer and the fully connected layer are fixed and used as the pre-defined LSTM network for time-series feature extraction during the method execution phase.

[0029] From the spatial network association characteristics, the fixed physical connection relationships of the pipeline network are extracted. For example, node A connects to node B, and node B connects to node C, forming static connections, which are presented in the form of adjacency matrix and initial feature matrix.

[0030] It should be noted that after obtaining the fused temporal characteristics of each monitoring node through a pre-set long short-term memory network, a dynamic propagation path matrix is ​​constructed by combining it with the static topological connectivity extracted from pipeline geographic information entities. This matrix is ​​used to quantify the spatial transmission intensity and direction of pressure fluctuations in the pipeline network over time.

[0031] Based on the adjacency matrix in static topological relationships Identify nodes with physical pipe connections. Subsequent conduction strength calculations are performed only for these node pairs to exclude spurious conduction paths without physical connections. For each node pair with a physical connection... Fusion of temporal features based on LSTM output and (length is) Time window, ), calculate conduction strength The calculation formula is as follows: in The Pearson correlation coefficient is used to measure the correlation between nodes. With nodes Waveform similarity after considering theoretical conduction delay; The theoretical conduction delay is based on the pipe length. With pressure wave propagation speed (Take the sound velocity of the medium in the pipeline, such as natural gas pipeline) )calculate: ; The peak time delay of the actual observed waveform is calculated using the cross-correlation function; The time delay attenuation coefficient is taken as an empirical value. This is used to penalize conduction relationships where the actual delay deviates too much from the theoretical delay, preventing false correlations from being misjudged as valid conduction.

[0032] For each time step According to all nodes Conductivity Construct a Dynamic propagation path matrix ,in This represents the total number of nodes. The matrix elements are defined as follows: in To set the minimum conduction strength threshold, take This is used to filter weak correlations with extremely low relevance or excessive time delay deviations, ensuring the sparsity and physical interpretability of the matrix. The dynamic propagation path matrices of all time steps are stacked along the time axis to obtain a three-dimensional dynamic propagation tensor. ,in The length of the time window (e.g., taking...) (Each sampling point corresponds to 100 seconds). This tensor fully records the propagation path and intensity changes of pressure fluctuations in the pipeline network over time.

[0033] Suppose there are three nodes A, B, and C, where A is physically connected to B, and B is physically connected to C. Within a certain time window, A experiences a sudden pressure drop. The fused timing features extracted by LSTM show that the waveform correlation between A and B is 0.92, the actual delay is 0.52 seconds, and the theoretical delay is 0.50 seconds. The calculated... The correlation between B and C is 0.41, the actual delay is 0.85 seconds, and the theoretical delay is 0.55 seconds. The calculated... Therefore, only the propagation intensity from A to B is retained in the dynamic propagation path matrix, while the intensity from B to C is set to zero because it is below the threshold. This accurately reflects the physical fact that the pressure wave propagates along A to B and does not further propagate to C. For the three-dimensional dynamic propagation tensor... A weighted average is calculated along the time dimension, with the weights using an exponential decay function. , To highlight the recent transmission state, the weighted average matrix is ​​expanded row-wise to obtain... dimensional pressure fluctuation transmission vector The node pairs corresponding to the non-zero elements in this vector represent the active pressure transmission paths in the current system, and their numerical values ​​reflect the transmission strength.

[0034] After constructing the dynamic propagation path matrix, based on the requirements of pressure transmission analysis, feature components characterizing propagation intensity, propagation speed, attenuation coefficient, transmission direction, and pressure temporal change rate are extracted from the matrix to form an initial transmission feature set. The selection criterion for each feature component is its mutual information value with historical leakage event tags; feature components with mutual information values ​​higher than a preset threshold are retained, while redundant components with values ​​lower than the threshold are removed.

[0035] Based on the spatial topological relationship between pipeline nodes and pipe segments, the retained core feature values ​​are organized into an initial transmission feature matrix according to the node-pipe segment structure. Based on the betweenness centrality of each pipe segment in the topological structure, a linear mapping method is used to calculate the transmission weight coefficient. After normalizing the betweenness centrality to the [0,1] interval, it is directly used as the weight coefficient for that pipe segment; the higher the betweenness centrality, the larger the weight coefficient. The feature values ​​of each node and pipe segment at the same time step are weighted and aggregated with their corresponding weight coefficients to obtain a multidimensional pressure fluctuation transmission feature set that integrates spatial distribution and pressure temporal changes. This feature set is then normalized using minmax and used as input for subsequent dimensionality reduction operations. The extracted pressure fluctuation transmission vector can quantify the energy distribution and flow direction of fluctuations in space.

[0036] For example, the aforementioned pressure fluctuation transmission feature set reaches 128 dimensions in a complex pipeline network environment. To reduce the computational complexity of subsequent graph convolution operations and suppress redundant information, dimensionality reduction processing is required. The preset dimensionality threshold is set to 64 dimensions, determined through feature redundancy analysis of historical station operation data. Principal component analysis is used to perform dimensionality reduction projection on the current pressure fluctuation transmission feature set. First, 128-dimensional historical transmission feature samples with the same dimension as the current feature, accumulated during the historical operation phase of the pipeline network area, are retrieved to construct a 128×128 covariance matrix to measure the correlation between dimensions. Subsequently, eigenvalue decomposition is performed to calculate the eigenvalues ​​and corresponding eigenvectors of the matrix, and they are sorted in descending order of eigenvalues. The magnitude of the eigenvalues ​​reflects the data variance in each principal component direction.

[0037] During this projection process, the cumulative variance contribution rate is calculated, and the sum of the variances represented by the first 32 principal components is compared with the original total variance. Experimental data shows that these 32 core features contain more than 95% of the effective fluctuation energy in the original signal, while the remaining 96 dimensions with extremely small feature values ​​are identified as redundant components. In this process, redundant fluctuation components with a contribution rate of less than 5% to the overall health diagnosis can be accurately filtered out.

[0038] Specifically, these filtered features typically manifest as subtle pressure pulsations in remote terminal branches or thermal noise signals generated by the on-site electromagnetic environment. Their distribution in the feature space is extremely chaotic and their energy is weak, offering no substantial contribution to determining the conduction mode of the main pipeline. For example, when a dimensionality reduction operation removed features from dimensions 100 to 128, which contributed only 0.3%, it successfully eliminated background high-frequency vibration interference caused by the operation of the station's air conditioning system. Finally, by retaining the top 32 features with the highest energy concentration, a simplified conduction path vector was obtained.

[0039] In the matching operation, the simplified 32-dimensional transmission path vector obtained after dimensionality reduction is input into a pre-defined classifier. This classifier consists of a fully connected layer and a softmax activation function, and is pre-trained using historical fault data samples. The training samples include pressure fluctuation transmission feature vectors and their corresponding pattern labels under typical operating conditions such as normal operation, pressure regulation, minor leaks, and sudden leaks. Pre-defined fluctuation patterns include types such as "branch line minor leakage," "main line sudden pressure drop," "normal pressure regulation," and "pump group periodic fluctuation." The fully connected layer of the classifier performs a linear transformation between the input feature vector and the pre-defined pattern feature matrix, and calculates the matching degree score with each pre-defined fluctuation pattern. During execution, the 32-dimensional feature vector after principal component analysis (PCA) is first obtained. This vector represents the core transmission characteristics of the current pressure fluctuation in the spatial topology. Subsequently, a pre-trained weight matrix is ​​used to map this vector. Each row of this weight matrix essentially corresponds to an ideal feature template of a pre-defined fluctuation pattern, such as "branch line minor leakage" or "normal pressure regulation." The matching score can be calculated as z = Pq + b, where q is the real-time 32-dimensional input vector, P is the weight matrix of the pattern features, and b is the bias term used to correct system bias.

[0040] In actual implementation, assuming that the extracted simplified transmission path vector, after matching calculation, yields matching scores of 5.2, 1.1, and 0.4 for the three preset modes of "branch micro-leakage," "normal pressure regulation," and "pump set periodic fluctuation," respectively, a preset safety threshold of 3.5 is set to quantify the degree of matching and determine the reliability of the current state. For example, when the calculated matching score between the current simplified transmission path vector and historical leakage evolution characteristics reaches 5.2, because it significantly exceeds the preset safety threshold of 3.5, it is determined that the current abnormal transmission mode "branch micro-leakage" exists.

[0041] It is worth noting that the setting of the preset safety threshold of 3.5 is primarily based on the statistical distribution of historical samples and the trade-off between false alarms and missed alarms. Firstly, analysis of historical fault data and long-term operational data from the station revealed that the characteristic matching degree triggered by normal pressure regulation and other operations is typically at a low level, while the matching degree for actual damage or leakage conditions usually reaches above 3.8. Setting the threshold to 3.5 effectively covers the redundancy range of normal fluctuations and prevents invalid alarms. Secondly, based on the analysis of the receiver operating characteristic (ROC) curve, a threshold of 3.5 achieves optimal capture rates for abnormal transmission patterns such as minor leaks with an extremely low false alarm rate, thus preventing small problems from escalating into major malfunctions.

[0042] In this invention, based on the classification results, the dynamic propagation path matrix and the pattern identifier output by the classifier are jointly input into a preset graph convolutional network. This graph convolutional network uses the dynamic propagation path matrix as the adjacency relationship and the pressure fluctuation transmission characteristics of each node as the node features. It aggregates the feature information of neighboring nodes through graph convolutional layers to perform spatial consistency verification on the preliminary classification results. During the verification process, based on the principle of mass conservation, the theoretical values ​​and actual observed values ​​of pressure attenuation gradients and flow changes at upstream and downstream nodes are compared. If the deviation exceeds a preset physical constraint threshold, the preliminary classification results are corrected; for example, a misclassified "branch micro-leakage" is corrected to "upstream pressure regulation operation." The finally determined pipeline transmission pattern includes a pattern type identifier, pressure wave propagation velocity, and anomaly source location information, providing maintenance personnel with actionable risk tracing evidence.

[0043] In step S13, historical fault data is acquired, and maintenance frequency features and fluctuation attenuation features are extracted from the historical fault data. Based on the maintenance frequency features and fluctuation attenuation features, the pipeline conduction mode is quantitatively evaluated. If the quantitative evaluation value exceeds a preset conduction mode anomaly threshold, spatial topology nodes are extracted from the pipeline conduction mode, and pipeline coupling vectors are extracted from the spatial topology nodes using a preset attention mechanism, including: Historical fault data is acquired, maintenance frequency characteristics and fluctuation attenuation characteristics are extracted from the historical fault data, and the pipeline conduction mode is quantitatively evaluated based on the maintenance frequency characteristics and fluctuation attenuation characteristics. If the quantitative evaluation value exceeds the preset abnormal threshold of the conduction mode, spatial topology nodes are extracted from the pipeline conduction mode. The historical fault data is mapped using the spatial topology nodes to obtain a time alignment matrix; A preset attention mechanism is used to assign the influence weights of the historical fault data to the time alignment matrix to determine the coupled feature tensor. For the coupling feature tensor, the spatial topology nodes are matched and fused with the temporal features of the historical fault data to obtain the pipeline coupling vector.

[0044] In this invention, the maintenance frequency characteristic is defined as the intervention frequency of a specific node or pipe section per unit time. For example, for node A, the total number of key valve replacements and sealing tests in the past 12 months is extracted and converted into a monthly average maintenance frequency characteristic value. The fluctuation attenuation characteristic is used to quantify the energy loss of pressure waves during propagation. Based on pressure drop events recorded in the historical abnormal fluctuation database, the attenuation rate of the pressure signal over time or the energy attenuation gradient over propagation distance is extracted. For example, the average energy attenuation gradient of the pressure wave per kilometer in a certain pipe section during historical leakage events is 0.15 MPa. Both reflect the health status of the pipeline from two dimensions: maintenance history and physical attenuation characteristics.

[0045] In this invention, when calculating the comprehensive score of pipeline transmission mode, the maintenance frequency feature and the fluctuation attenuation feature are first normalized by minmax, and then the comprehensive score is calculated by weighted summation. The weight coefficient of the maintenance frequency feature is set to 0.4, and the weight coefficient of the fluctuation attenuation feature is set to 0.6. This weight allocation is based on the regression analysis results of historical fault samples, in which the fluctuation attenuation feature contributes more to the identification of abnormal transmission mode.

[0046] During execution, if the comprehensive score of the transmission mode reaches 0.7, it is determined that the current pipeline transmission mode is abnormal, and the node whose comprehensive score exceeds the threshold, along with the topological neighboring nodes that have a physical connection to that node, are identified as the core spatial coordinates for subsequent spatiotemporal coupling analysis. This comprehensive score threshold of 0.7 was determined based on statistical analysis of a large number of historical operation samples and failure cases at the station. Under normal operating conditions, the comprehensive score is distributed in the range of 0 to 0.6, while the score is often higher than 0.7 during failures such as leakage, abnormal transmission, and structural defects.

[0047] A time alignment matrix is ​​constructed, with rows corresponding to continuous time steps within the current monitoring period, totaling 600 time steps, each spaced 1 second apart. Columns correspond to the historical maintenance attributes of spatial topology nodes, including maintenance event type, maintenance timestamp, and recovery time after maintenance. Historical maintenance events of locked nodes over the past year are mapped to the corresponding time step positions in the matrix according to their timestamps. If multiple maintenance events exist within the same time step, the maintenance event type, maintenance timestamp, and recovery time after maintenance for each event are stored in different columns of the matrix, forming a multi-dimensional feature representation.

[0048] A pre-defined attention mechanism is used to weight the time alignment matrix to determine the coupling feature tensor. This pre-defined attention mechanism is used to extract pipeline coupling vectors from spatial topology nodes. The attention mechanism employs a self-attention network structure with two attention heads, each with a dimension of 32, an input feature dimension of 64, and an output feature dimension of 64. Training data comes from aligned samples of historical maintenance records and sensor time-series data. Historical maintenance event features (maintenance type, maintenance timestamp, and recovery time after maintenance) corresponding to spatial topology nodes are used as key-value inputs, the pressure time-series features of the current time window are used as query inputs, and whether the node will fail within the next week is used as a supervision signal. The training objective is a binary classification task, i.e., predicting whether a node has a failure risk within the next week. The loss function is binary cross-entropy loss, the optimization algorithm is Adam, the learning rate is set to 0.001, and the training epochs are 150. After training, the weight parameters of the attention network are fixed and used as the pre-defined attention mechanism for weighted fusion of historical data during the method execution phase.

[0049] The query vector is composed of fused temporal features extracted from the current pressure time-series data via a Long Short-Term Memory (LSTM) network. The key and value vectors are composed of historical maintenance feature vectors corresponding to each time step in the time-aligned matrix; each row of the matrix serves as a feature vector for a given time step, acting as the input for the key and value, respectively. The attention weight distribution is obtained by calculating the dot product of the query and key vectors, followed by scaling and softmax normalization. This weight distribution is then weighted and summed with the value vector to generate a coupled feature tensor. The dimension of this coupled feature tensor is (number of nodes × time step × feature dimension), where the node dimension corresponds one-to-one with the spatial topology nodes, and the time dimension retains the contribution weights of historical maintenance events to the current abnormal state.

[0050] For example, for the current pressure reduction signal of node A, the attention mechanism calculates that its attention weight with the pipeline anti-corrosion coating failure event 12 months ago is 0.7, and its attention weight with the routine inspection event half a month ago is 0.1. The feature vectors corresponding to each historical event are weighted and summed according to the attention weights to generate a coupled feature tensor that integrates the key historical evolution trends.

[0051] For the feature vectors of each spatial topology node in the coupled feature tensor, a weighted average aggregation is performed along the time dimension. The attention weights at each time step are used as weighting coefficients to calculate the weighted average feature vector of each node on the time axis, forming a node-level comprehensive feature representation. This comprehensive feature representation simultaneously integrates pipeline spatial topology information and long-term impact information of historical faults, forming a pipeline coupling vector, which serves as the input for subsequent anomaly map construction.

[0052] In step S14, node association features are extracted from the pipeline coupling vector, and the maintenance frequency features and the fluctuation attenuation features are concatenated with the node association features to obtain a heterogeneous node set. Based on the heterogeneous node set, a graph convolution operation is performed on the pipeline conduction mode and the spatial topology nodes to determine the pipeline anomaly map, including: Based on the heterogeneous node set, graph convolution operation is performed on the pipeline conduction mode and the spatial topology nodes to construct a time matrix, and the preset conduction threshold is aligned with the time matrix to determine the basic graph of coupling anomaly distribution; If the dependency weight in the coupled anomaly distribution base map exceeds a preset dependency weight threshold, then the pressure anomaly fluctuation and physical spatial location are extracted from the coupled feature tensor, and the pressure anomaly fluctuation and the physical spatial location are mapped and fused to obtain the distribution extension map. Based on the distribution expansion map, the node association features and the abnormal pressure fluctuations are fused to obtain the abnormal distribution map; The abnormal distribution map is processed by graph convolution iteration. When the spatial distribution and temporal transmission law of the abnormal distribution map meet the preset transmission mode abnormality threshold, the pipeline abnormal map is determined.

[0053] In practice, the pipeline coupling vector is split along the node dimension. Each node's feature vector contains multiple pre-defined feature segments, one of which characterizes the connection strength between the node and its neighboring nodes. This connection strength feature segment is extracted from the node's feature vector and used as the node association feature. This feature segment is then concatenated with the node's historical maintenance frequency and energy decay gradient to form a heterogeneous node set. In this way, each spatial topology node evolves into a multi-dimensional heterogeneous node containing physical attributes, real-time fluctuation status, and historical health background.

[0054] Based on the heterogeneous node set, a graph convolution operation is performed on the pipeline transmission pattern and spatial topology nodes. Using the adjacency matrix as the topological relationship and the heterogeneous features of each node as input features, a graph convolutional network aggregates the feature information of each node and its neighboring nodes, outputting an updated node anomaly feature vector. Based on the updated node anomaly feature vector, the anomaly transmission dependency weights between node pairs are calculated, forming a transmission dependency weight matrix. The dependency weight between any two nodes with a physical connection is calculated by concatenating the anomaly feature vectors of the two nodes and inputting the result into a fully connected network. The fully connected network is used to calculate the anomaly transmission dependency weights between node pairs.

[0055] The network adopts a three-layer fully connected structure with an input dimension of 128 (the concatenation of the abnormal feature vectors of two nodes), hidden layer dimensions of 64 and 32 respectively, and an output dimension of 1. The activation function is ReLU, and the output layer uses the Sigmoid activation function to map the output value to the range of 0 to 1.

[0056] Training data is derived from anomalous propagation chains recorded in historical failure events. For each historical failure event, the failure source node and the affected downstream node pair are extracted, their anomalous feature vectors are concatenated, and used as input. The actual propagation intensity (calculated based on the pressure fluctuation arrival time difference and amplitude attenuation ratio) is used as the supervision label. The mean squared error loss function is used, the Adam optimization algorithm is employed, the learning rate is set to 0.001, and the training iterations are 100. After training, the network's weight parameters are fixed and used as a preset weight calculation network for dependency weight calculation during the method execution phase.

[0057] Simultaneously, a time matrix is ​​constructed. The rows of this matrix correspond to spatial topological nodes, and the columns correspond to consecutive time steps, each time step consistent with the sensor sampling interval. For each physically connected node pair, the theoretical propagation delay is calculated based on its physical connection distance and the pressure wave propagation speed. This delay is divided by the sampling interval and rounded to the nearest integer to obtain the corresponding time step index. A 1 is then filled into the corresponding position in the time matrix, indicating that the node pair has a propagation possibility within that time step. If the node pair has no physical connection, or the theoretical propagation delay cannot match any time step within the current time window, a 0 is filled into the corresponding position. After the matrix elements are filled, the time matrix is ​​compared element-by-element with a preset propagation threshold. If the physical propagation delay between a node pair exceeds the preset safe propagation threshold, the node pair is marked as having a time anomaly.

[0058] When generating the basic graph of coupling anomaly distribution, spatial topological nodes are used as vertices, and only node pairs that simultaneously satisfy the following two conditions are retained as edges: first, the corresponding dependency weight in the transitive dependency weight matrix is ​​greater than a preset dependency weight threshold; second, the node pair is marked as a temporal anomaly in the temporal matrix. The dependency weight of each edge is taken from the corresponding value in the transitive dependency weight matrix. This basic graph reflects the anomaly transitive relationships that simultaneously satisfy both transitive dependency strength and temporal logic constraints.

[0059] A dependency weight threshold is set, which is determined based on the weight distribution of effective transmission chains in the station's historical anomaly events. Typically, a higher statistical quantile is chosen; for example, if the weights of true anomaly transmission chains in historical data are generally higher than 0.6, the threshold can be set to 0.65. Each edge in the base graph of coupled anomaly distribution is traversed. If its dependency weight is greater than or equal to the threshold, the edge is marked as a high-confidence anomaly transmission chain and will be the focus of subsequent extended analysis. Edges below the threshold are retained in the base graph to maintain the integrity of the graph structure but are not spatially expanded.

[0060] For each upstream and downstream node pair corresponding to a high-confidence anomaly propagation chain, pressure anomaly fluctuation parameters are extracted from the previously generated coupling feature tensor. The extracted parameters include pressure fluctuation amplitude, fluctuation attenuation degree, and anomaly frequency characteristics. Pressure fluctuation amplitude refers to the peak amplitude of the pressure fluctuation at the node, reflecting the energy intensity of the anomaly; fluctuation attenuation degree refers to the proportion of amplitude attenuation during the propagation of the pressure wave from the upstream node to the downstream node, used to determine the effectiveness of propagation; anomaly frequency characteristics refer to the energy proportion of the pressure fluctuation within a specific frequency band, used to assist in identifying the anomaly type. These parameters are calculated in real-time using a sliding time window, forming a parameter sequence aligned with the monitoring time.

[0061] The physical spatial coordinates of upstream and downstream nodes are extracted from the pipeline geographic information entity to determine the spatial orientation of the pipeline segment. Using this pipeline segment as the axis, a sequence of spatial interpolation points is generated at fixed intervals to continuously represent abnormal fluctuation parameters along the pipeline segment. Each interpolation point records its relative position information on the pipeline segment.

[0062] The extracted pressure anomaly fluctuation parameters are mapped onto spatial interpolation points to form a distribution extension map. The mapping fusion employs a distance-weighted allocation method: for any interpolation point on the pipe segment, its anomaly intensity is calculated by weighting its distance from upstream and downstream nodes. Points closer to a node are more significantly affected by the node's anomaly fluctuations, while the influence gradually diminishes at greater distances. Simultaneously, considering the physical law of pressure wave energy decay during propagation, the theoretical time required for the fluctuation to propagate from the upstream node to the current point is calculated based on the pressure wave propagation speed. Combined with the pressure fluctuation decay law determined in step S12, the anomaly intensity of the interpolation points is adjusted to ensure that the spatial distribution of the anomaly intensity conforms to the physical propagation law. The anomaly parameters at each time step are arranged chronologically to form a temporal evolution sequence for each interpolation point, reflecting the dynamic development process of the anomaly on the pipe segment. Through this fusion, the anomaly information, originally existing only at the nodes, is extended to the entire pipe segment, generating a distribution extension map. This map uses the pipe segment as the basic unit, with each spatial location carrying the anomaly fluctuation intensity and its temporal evolution information.

[0063] The distribution expansion map is fused with node association features. These features include maintenance frequency and fluctuation attenuation characteristics. Maintenance frequency reflects the historical intervention level of the pipe segment, while fluctuation attenuation reflects the inherent response characteristics of the pipe segment to pressure fluctuations. For each interpolation point, the maintenance frequency and fluctuation attenuation characteristics of its associated pipe segment are correlated. The intensity of abnormal fluctuations is combined with the aforementioned node association features to form a comprehensive feature vector for that point. All interpolation points are organized according to pipe segment connections to form a preliminary anomaly distribution map.

[0064] After obtaining the preliminary anomaly distribution map, iterative processing is performed using a graph convolutional network (GCNN). This GCNN uses each interpolation point in the preliminary anomaly distribution map as a processing unit, and the spatial connectivity of the interpolation points as adjacency relationships. It consists of three GCNN layers, each taking the comprehensive features and adjacency relationships of the interpolation points in the current map as input and outputting an updated feature vector. After each iteration, the spatial distribution characteristics and temporal transmission patterns of the anomaly map are calculated and compared with preset anomaly map judgment thresholds. These preset thresholds include: a spatial anomaly coverage ratio of no less than 30%, where the spatial anomaly coverage ratio is defined as the proportion of the number of anomaly interpolation points to the total number of interpolation points; a pressure transmission delay deviation of no less than 200 milliseconds; and an anomaly dependency weight between interpolation points of no less than 0.6. When both the spatial distribution and temporal transmission patterns of the anomaly distribution map meet the above preset anomaly map judgment thresholds, the map is determined as a pipeline anomaly map for overall system diagnosis.

[0065] In step S15, risk tracing indicators are extracted from the pipeline anomaly map. If the risk tracing indicators show an upstream transmission effect, the evolution trend is simulated and evaluated using a preset recurrent neural network to obtain a pipeline damage prediction sequence, including: Risk tracing indicators are extracted based on the abnormal distribution map, and the node association features and pipeline transmission patterns are fused through spatiotemporal correlation to determine the preliminary quantitative value of the upstream transmission effect. If the initial quantification value exceeds the preset quantification threshold, a preset recurrent neural network is used to evaluate the trend of the risk tracing indicator to obtain an intermediate simulation sequence. Based on the intermediate simulation sequence, the basic damage trend is determined, and the basic damage trend is aligned and fused with the spatial distribution and the temporal transmission law. If the fusion result is determined to meet the preset damage diffusion judgment rule, the damage probability, damage degree and transmission correlation of different pipe segments are quantified according to the time and space dimensions to generate an extended risk distribution set. The pipeline damage prediction sequence is obtained by iteratively processing the risk tracing indicators using the extended risk distribution set.

[0066] Specifically, risk tracing indicators are extracted from the anomaly distribution map. First, nodes in the map whose outliers exceed a preset pressure threshold are identified, forming an anomalous node set. The preset pressure threshold is set according to the normal operating pressure range of the pipeline network, for example, 45MPa to 50MPa. For each node in the anomalous node set, its anomaly intensity is quantified, including the percentage of pressure deviation from the threshold and the duration of the anomaly. Then, based on the temporal sequence of the anomalous nodes, the anomaly propagation trajectory is reconstructed. Following the upstream-to-downstream pressure propagation pattern, starting from the starting node of the trajectory, the topological connection weights between each node are extracted. These weights are derived from the aforementioned propagation dependency weight matrix.

[0067] Subsequently, correlation analysis was performed by combining node functional attributes with anomaly types. Anomaly types were matched with the functions of the associated equipment (e.g., compressors, valves) to extract related indicators such as real-time operating load and recent maintenance records. Finally, native anomaly nodes and transmissible anomaly nodes were distinguished by anomaly contribution. Anomaly contribution was calculated using a weighted summation method, with node anomaly intensity accounting for 0.7 and temporal position accounting for 0.3. The node with the highest contribution was identified as the native anomaly node. The anomaly type, deviation threshold, associated equipment operating status, and topology connection weights of the native anomaly nodes were extracted to construct a risk tracing indicator set.

[0068] It should be noted that the preliminary quantification value of the upstream transmission effect is determined by fusing spatiotemporal correlation features with the pressure transmission pattern. For node correlation features such as temperature and vibration, the percentage of the current value of each feature deviating from the normal operating range is calculated as the quantification value of a single feature, with a value range of [0,1]. The arithmetic mean of all single feature quantification values ​​is then taken to obtain the mean value of the node correlation features.

[0069] For the pressure transmission mode, the quantized value of the transmission mode is calculated based on the attenuation of the pressure wave along the transmission path. The pressure attenuation rate is calculated by subtracting the terminal pressure from the initial pressure, and then dividing the difference by the initial pressure. This attenuation rate is compared with the baseline attenuation rate under historical normal operating conditions. The baseline attenuation rate is obtained by averaging the pressure attenuation data of the same pipe section during the station's historical normal operation. When the actual attenuation rate exceeds the baseline attenuation rate, the quantized value of the transmission mode is taken as the ratio of the actual attenuation rate to the baseline attenuation rate, with an upper limit of 1. Finally, the mean of the node association characteristics and the quantized value of the pressure transmission mode are weighted and summed, with all weighting coefficients set to 0.5, to obtain the preliminary quantized value of the upstream transmission effect.

[0070] It is worth noting that the preset quantization threshold is set to 0.7, which is determined based on the statistical analysis of the distribution of preliminary quantization values ​​in the station's historical operational data. Under normal production fluctuations, the preliminary quantization values ​​are mostly distributed in the range of 0.3 to 0.6. Using 0.7 as a statistical cutoff point can effectively eliminate background noise and identify abnormal deviations.

[0071] When the initial quantification value exceeds 0.7, the time-series features from the risk tracing indicators are input into a pre-defined recurrent neural network for evolution trend assessment. Input features include pressure fluctuation time-series data from the past 48 hours, anomaly dependency weights between nodes, and node maintenance frequency characteristics. The pre-defined recurrent neural network is used to simulate the assessment of abnormal evolution trends and generate pipeline damage prediction sequences. This network employs a two-layer LSTM structure, with 128 hidden units in the first layer and 64 hidden units in the second layer. The input sequence length is 48 time steps (corresponding to 48 hours), and the output sequence length is also 48 time steps. Each output corresponds to a damage probability prediction for the next time step.

[0072] Training data is derived from the evolution of historical fault events. For each historical fault event, sensor data sequences from the 48 hours prior to the event are extracted as input, and damage severity labels (assessed based on a combination of actual maintenance records and sensor data) at each time point within the 48 hours following the event are extracted as output sequences. The training objective is sequence-to-sequence prediction. The loss function is mean squared error, the optimization algorithm is Adam, the learning rate is set to 0.001, the training epochs are 200, and an early stopping strategy is employed. After training, the weight parameters of the LSTM layer are fixed and used as a pre-defined recurrent neural network for predicting damage evolution trends during the method execution phase.

[0073] In practice, if the intermediate simulation sequence predicts that the pressure drop rate of a certain risk node will continue to increase within the next 24 hours, and the monthly average maintenance frequency of that node is higher than a preset threshold, then a damage propagation trend is identified. Using the risk node as the spatial center, a Gaussian kernel function is used for risk probability mapping. The risk probability distribution with spatial distance follows a Gaussian distribution, and the standard deviation is determined statistically based on the damage propagation range in historical fault samples. The risk probability at the risk node is set to 1, and the risk propagation radius is defined as the spatial distance at which the risk probability drops to 0.5. Within the risk propagation radius, the risk coverage probability is calculated based on the spatial distance between each pipe segment and the risk node. The potential damage level is then quantified by combining factors such as pipe segment material and service life, generating an extended risk distribution set.

[0074] The intermediate simulation sequence generated by the pre-defined recurrent neural network is mapped to the pipeline topology space, and an extended risk distribution set is used as a dynamic spatial constraint for iterative correction. For each time step, the predicted value of the recurrent neural network is compared with the upper limit of the maximum physical damage energy level corresponding to the spatial location in the extended risk distribution set. If the predicted value exceeds the upper limit, it is corrected to the upper limit; if the predicted value is lower than the upper limit, the original predicted value is retained. Through iterative correction at each time step, the prediction results are ensured to conform to the physical carrying capacity of the pipeline.

[0075] The final output pipeline damage prediction sequence is represented as an N×48 risk prediction matrix, where N is the total number of affected critical nodes, 48 ​​is the prediction time step, and each step is 1 hour apart. Each element in the matrix represents the pipeline health status score of the corresponding node at a specific future time point; the lower the score, the higher the risk of damage.

[0076] In step S16, based on the pipeline damage prediction sequence, the topological relationship and the pressure time series data are fused to obtain a pipeline health status score, including: Based on the pipeline damage prediction sequence and the topological relationship, a spatiotemporal coupling tensor is constructed; The frequency domain features and state decay gradient of the pressure time series data are extracted using the spatiotemporal coupling tensor. If the state attenuation gradient is greater than a preset attenuation threshold, then the signal frequency domain features and the topological relationship are fused to determine the node vulnerability. Based on the node vulnerability, the pipeline damage prediction sequence is weighted to obtain health status characteristics; The health status features are quantified and converted using a preset scoring mapping rule to obtain a pipeline health status score.

[0077] In the specific operation, the pipeline damage prediction sequence generated in step S15 is retrieved. This sequence is represented as an N×48 risk prediction matrix, where N is the total number of monitoring nodes, 48 ​​is the number of prediction time steps, each step is 1 hour apart, and the matrix elements represent the damage probability value of the corresponding node at a specific future time point. This risk prediction matrix is ​​fused with the adjacency matrix describing the physical connection relationship of the pipeline to construct a spatiotemporal coupling tensor. Specifically, for each time step, the node risk vector corresponding to that time step in the risk prediction matrix is ​​multiplied bitwise with the adjacency matrix to form the spatial risk distribution matrix under that time step. The spatial risk distribution matrices of all time steps are stacked along the time axis to obtain a three-dimensional spatiotemporal coupling tensor with dimensions N×N×48, where the first two dimensions represent the spatial correlation between nodes, and the third dimension represents the temporal evolution.

[0078] Spatial correlation information in the spatiotemporal coupling tensor is used as spatiotemporal weights to window the real-time acquired pressure signals. For each node, the pressure signals of neighboring nodes are weighted and aggregated based on the correlation strength between that node and other nodes in the spatiotemporal coupling tensor, enhancing the signal components related to abnormal propagation paths. Subsequently, a fast Fourier transform is used to extract frequency domain features such as the power spectral density of the aggregated signal to identify specific frequency components related to micro-leaks. Simultaneously, the slope of the pressure envelope is calculated to determine the state decay gradient. For example, taking an oil pipeline as an example, the normal operating pressure is 6.0 MPa. When a small leak occurs, the pressure at the monitoring node drops from 6.00 MPa to 5.94 MPa within 0.5 seconds, a decrease of 0.06 MPa, and the calculated state decay gradient is 0.12 MPa / s. This gradient index directly reflects the rate of deterioration of the pipeline's current physical state.

[0079] It is worth noting that the preset attenuation threshold is set to 0.10 MPa / s based on the station's historical operating baseline. This threshold is a statistical boundary point determined by the probability distribution of historical pressure monitoring data. During normal station operations, the state attenuation gradient of pressure fluctuations is typically distributed in a stable range of 0.02 MPa / s to 0.05 MPa / s. Setting the threshold to 0.10 MPa / s allows for the exclusion of benign pressure fluctuations caused by PID regulation, routine flow switching, etc., with a confidence level of over 95%, ensuring that subsequent signal frequency domain feature fusion is triggered only for sudden performance degradation with destructive potential. When the calculated state attenuation gradient exceeds this threshold, significant physical performance degradation is determined. At this point, the extracted signal frequency domain features are fused with the betweenness centrality of the node in the topology to calculate the node vulnerability.

[0080] The frequency domain distribution of pressure fluctuations was analyzed using Fast Fourier Transform (FFT), and the energy proportion of a specific high-frequency band was extracted as the signal frequency domain feature. This feature was then normalized using minmax to obtain a signal feature score. Simultaneously, the betweenness centrality of the target node in the pipeline network map was calculated using topological relationships to measure the node's pivotal importance in the global path. This was also normalized using minmax to obtain a topological feature score. Taking the core node A of the main road as an example, pressure fluctuation signals from node A were collected using a high-frequency acoustic emission sensor. FFT analysis extracted the energy proportion in the 3-5kHz high-frequency band as the signal frequency domain feature. The energy proportion in this band jumped from a baseline of 0.15 to 0.78, exhibiting significant leakage fingerprint characteristics. After normalization, the signal feature score was 0.78. Topological calculation showed that the betweenness centrality of node A was 0.92, and after normalization, the topological feature score was also 0.92. The node vulnerability was calculated by weighting the results using preset weights, with the signal frequency domain weight set to 0.6 and the topological importance weight set to 0.4.

[0081] After obtaining the node vulnerability, the pipeline health status score is calculated using the pipeline damage prediction sequence generated in step S15 and a preset scoring mapping rule. Specifically, the average damage probability of each node within a future time window is extracted from the pipeline damage prediction sequence. This average is multiplied by the vulnerability of the corresponding node to obtain a corrected damage characterization value. Then, a piecewise linear mapping function is used to map this value to a standardized range of 0 to 100, where 100 represents complete health and below 60 represents serious risk.

[0082] The preset scoring mapping rule uses a piecewise linear mapping function, mapping segments based on the corrected damage characterization value x. When x∈[0,0.2), the mapping formula is f(x)=100-50x, indicating the pipeline is in a very low-risk state, with a score falling in the [90,100] interval, representing good structural integrity. When x∈[0.2,0.5), the mapping formula is f(x)=110-100x, corresponding to slight performance degradation or potential hazards in the pipeline, with a score falling in the [60,90) interval, reminding maintenance personnel to increase inspection frequency. When x≥0.5, the mapping formula is f(x)=120-120x, corresponding to severe anomalies or structural damage, with a score falling between [0,60). The threshold setting of this segmented mapping rule is based on objective statistical results of signal-to-noise ratio distribution and physical characteristics of engineering failure. Setting 0.2 as the first threshold is to define a safety redundancy zone. Analysis of historical operating benchmarks for the station revealed that interference scores from conventional fluid turbulence or regulation fluctuations are typically below 0.15. Setting a gentle slope for the area below 0.2 effectively filters background noise and maintains a stable score above 90. Setting 0.5 as the severe risk trigger point is based on historical leakage test data. When the damage characterization value reaches this critical value, the pipeline energy decay rate is usually more than three times that under normal operating conditions, marking a physical transition from performance degradation to structural damage. The logic of the slope in the piecewise function gradually increasing from 50 to 120 is to implement a non-linear risk penalty mechanism, ensuring that once the damage intensity crosses the physical critical point of 0.5, the score will plummet and directly fall below the 60-point alarm threshold, thus widening the numerical gap between sub-optimal and severe risk.

[0083] In step S17, the matching degree between the pipeline conduction pattern and the historical fault data is analyzed based on the pipeline health status score. If the matching degree is lower than a preset matching threshold, a pipeline risk alarm is generated, including: By fusing the spatiotemporal coupling tensor with the signal frequency domain features through the pipeline health status score, a comprehensive feature is obtained. From the comprehensive feature, the state decay gradient and node vulnerability correlation sequence in the pipeline conduction mode are extracted. The matching index is calculated by comparing the vulnerability association sequence with the damage evolution rate in the historical failure data. If the matching degree index is lower than the preset matching threshold, preliminary alarm data is generated by combining the topological relationship and the health status characteristics. By integrating the preliminary alarm data with the topological relationship and the pressure time series data, the alarm signal strength is determined. By combining the alarm signal strength with the preset scoring mapping rules, a pipeline risk alarm is generated.

[0084] For example, using the health score vector determined by S16 as mask weights, element-wise dot product operations are performed on the high-dimensional spatiotemporal coupling tensor, thereby enhancing the features of nodes with lower scores in the feature space. Subsequently, the enhanced tensor is concatenated with the signal frequency domain features (such as the energy spectrum of a specific frequency band) to form a comprehensive feature that can characterize the global attributes of the current fluctuation. The correlation sequence extracted from the comprehensive feature records the dynamic relationship between the state decay gradient (e.g., 0.12 MPa / s) and node vulnerability (e.g., 0.85) of each risk node as the spatial topology evolves.

[0085] In practice, the damage evolution rate of the node marked in historical fault data is extracted from the database; for example, the average decay rate of the pipe segment over the past 12 months is 0.08 rating units per hour. The Pearson correlation coefficient between the current state decay trend and the historical evolution rate is calculated and converted into a matching index between 0 and 1. For example, if the current rate of pressure drop far exceeds the steady wear evolution logic in historical records, the calculated matching index is 0.42. This index quantifies the extent to which the current abnormal fluctuation deviates from the normal historical evolution pattern, thus determining whether there are any sudden new risks.

[0086] It is worth noting that the preset matching threshold mentioned here is set to 0.6 (in physical logic, 0.6 is defined as the boundary between the characteristic correlation of known evolutionary logic and newly emerging abnormal patterns. Through cosine similarity analysis of a large number of historical maintenance cases, it was found that the correlation strength between normal maintenance evolutionary characteristics and their actual signals is usually stable above 0.65, hence 0.6 is used as the threshold). When the matching degree of 0.42 mentioned above is lower than this threshold, it is determined that the current abnormal fluctuation is not a product of historically known evolution and belongs to a newly emerging risk that needs to be traced back to its root cause immediately. At this time, using topological relationships (adjacency matrix), a breadth-first search is performed, starting from the most severely affected node, to trace the starting position of the pressure wave propagation in reverse. For example, if the pressure wave is identified as originating from "branch tee node C", and combined with the health status characteristics of this node, such as the high vulnerability score due to corrosion, preliminary alarm data including fault location, suspected type, and preliminary risk level is generated.

[0087] After generating preliminary alarm data, the alarm signal strength is determined by integrating the topological relationship with the pressure time series data. Specifically, the initially identified risk location is first reprojected onto the original pressure time series data axis, and the pressure jump amplitude at the current moment is extracted, for example, 0.5 MPa. Based on the elastic limit and operational safety margin of the pipe material, a pressure sensitivity coefficient is pre-calibrated. This coefficient is used to convert the pressure jump amplitude into a dimensionless intensity component reflecting the physical impact energy level. The calibration of the pressure sensitivity coefficient is based on the station pipeline design parameters and a historical pressure fluctuation database. It is determined through statistical analysis of the pressure jump amplitude and corresponding damage degree in typical leakage events, ensuring that abnormal fluctuations at different pressure levels can be mapped to a unified intensity dimension. Taking a certain station as an example, based on its pipeline material and operating pressure level, the pressure sensitivity coefficient calibration value is 14. Therefore, the physical intensity component is 0.5 multiplied by 14, resulting in a calculation of 7.0.

[0088] Meanwhile, by retrieving topological relationships, the number of key nodes to which the fluctuation spreads through physical connections is statistically analyzed; for example, the number of spreading nodes is 4. Based on the pipeline network topology, a node expansion coefficient is pre-calibrated. This coefficient is used to convert the number of nodes affected by the fluctuation into a dimensionless expansion component characterizing the spatial damage coverage area. The calibration of the node expansion coefficient is based on the statistical distribution of the average connectivity of the core nodes of the station and the spatial diffusion range in historical risk events. It is determined through comparative analysis of the fluctuation propagation range under normal and abnormal operating conditions, enabling risk events of different scales to quantify their threat range in the form of spatial expansion components. Taking this station as an example, based on the statistical analysis of its average pipeline node degree and risk coverage area, the node expansion coefficient is calibrated to 0.375. Therefore, the spatial expansion component is 4 multiplied by 0.375, resulting in a calculation of 1.5.

[0089] Finally, the physical intensity component and the spatial expansion component are superimposed to determine the alarm signal strength as 8.5, ensuring that the generated alarm data can reflect both the instantaneous destructive power of abnormal fluctuations and their real-time threat range in complex pipeline networks.

[0090] In this invention, a pipeline risk alarm is generated by combining the alarm signal strength with a preset scoring mapping rule. The preset scoring mapping rule divides continuous signal strength into different levels of warning actions. The mapping rule is as follows: the strength value is divided into four quantitative ranges: blue (normal, 0-2.0), yellow (caution, 2.0-5.0), orange (alert, 5.0-8.0), and red (high risk, 8.0-10.0). The final generated pipeline risk alarm is output in the form of a visual chart, clearly marking the root cause of the risk as "negative pressure wave caused by micro-leakage at the branch tee," and includes a 24-hour evolution prediction suggestion based on a damage prediction sequence.

[0091] In summary, this invention discloses a method for managing the entire lifecycle data of oil and gas station pipelines, which enables precise location and assessment of the overall health status of the pipeline system.

[0092] Reference Figure 2 The second embodiment of the present invention provides a data management system for the entire life cycle of oil and gas station pipelines, including: The geographic construction module is used to acquire pipeline geographic information entities, extract node coordinates from the pipeline geographic information entities, construct topological relationships, and extract spatial network association features from the topological relationships through a preset graph neural network. The pressure transmission module is used to acquire pressure time series data based on the spatial network association characteristics, and to fuse the pressure time series data using a preset long short-term memory network to determine the pipeline transmission mode. The spatiotemporal coupling module is used to acquire historical fault data, extract maintenance frequency features and fluctuation attenuation features from the historical fault data, and quantitatively evaluate the pipeline conduction mode based on the maintenance frequency features and fluctuation attenuation features. If the quantitative evaluation value exceeds the preset conduction mode abnormality threshold, spatial topology nodes are extracted from the pipeline conduction mode, and pipeline coupling vectors are extracted from the spatial topology nodes through a preset attention mechanism. An anomaly graph module is used to extract node association features from the pipeline coupling vector, and to concatenate the maintenance frequency features and the fluctuation attenuation features with the node association features to obtain a heterogeneous node set. Based on the heterogeneous node set, a graph convolution operation is performed on the pipeline conduction mode and the spatial topology nodes to determine the pipeline anomaly graph. The damage prediction module is used to extract risk tracing indicators from the pipeline anomaly map. If the risk tracing indicators show an upstream transmission effect, the evolution trend is evaluated by simulating the evolution trend through a preset recurrent neural network to obtain a pipeline damage prediction sequence. The health scoring module is used to obtain a pipeline health status score by fusing the topological relationship and the pressure time series data based on the pipeline damage prediction sequence. The risk warning module is used to analyze the matching degree between the pipeline conduction mode and the historical fault data based on the pipeline health status score. If the matching degree is lower than a preset matching threshold, a pipeline risk alarm is generated.

[0093] It should be noted that the oil and gas station pipeline full life cycle data management system provided in this embodiment of the invention is used to execute all process steps of the oil and gas station pipeline full life cycle data management method in the above embodiment. The working principle and beneficial effects of the two are one-to-one, so they will not be described again.

[0094] It should be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the device embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0095] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A method for managing the entire lifecycle data of oil and gas station pipelines, characterized in that, include: Obtain pipeline geographic information entities, extract node coordinates from the pipeline geographic information entities, construct topological relationships, and extract spatial network association features from the topological relationships through a preset graph neural network; Based on the spatial network association characteristics, pressure time series data is obtained, and the pressure time series data is fused using a preset long short-term memory network to determine the pipeline conduction mode; Historical fault data is acquired, maintenance frequency features and fluctuation attenuation features are extracted from the historical fault data, and the pipeline conduction mode is quantitatively evaluated based on the maintenance frequency features and fluctuation attenuation features. If the quantitative evaluation value exceeds the preset abnormal threshold of the conduction mode, spatial topology nodes are extracted from the pipeline conduction mode, and pipeline coupling vectors are extracted from the spatial topology nodes through a preset attention mechanism. Node association features are extracted from the pipeline coupling vector, and the maintenance frequency features and the fluctuation attenuation features are concatenated with the node association features to obtain a heterogeneous node set. Based on the heterogeneous node set, graph convolution operation is performed on the pipeline conduction mode and the spatial topology nodes to determine the pipeline anomaly map. Risk tracing indicators are extracted from the pipeline anomaly map. If the risk tracing indicators show an upstream transmission effect, the evolution trend is simulated and evaluated through a preset recurrent neural network to obtain a pipeline damage prediction sequence. Based on the pipeline damage prediction sequence, the pipeline health status score is obtained by fusing the topological relationship with the pressure time series data. The matching degree between the pipeline conduction mode and the historical fault data is analyzed based on the pipeline health status score. If the matching degree is lower than the preset matching threshold, a pipeline risk alarm is generated.

2. The method for full lifecycle data management of oil and gas station pipelines according to claim 1, characterized in that, Using a pre-defined graph neural network, spatial network association features are extracted from the topological relationships, including: Generate an adjacency matrix and an initial feature matrix based on the aforementioned topological relationships; The adjacency matrix and the initial feature matrix are input into a preset graph neural network for feature aggregation to obtain deep node features; If the dimension of the deep node feature is lower than a preset dimension threshold, then the deep node feature is subjected to dimensionality up-mapping to obtain the target dimension feature, and spatial network association features are extracted from the target dimension feature.

3. The method for full lifecycle data management of oil and gas station pipelines according to claim 1, characterized in that, The step of obtaining pressure time-series data based on the spatial network correlation characteristics and fusing the pressure time-series data using a preset long short-term memory network to determine the pipeline conduction mode includes: Based on the spatial network association characteristics, key monitoring nodes and static connections are located, and pressure time-series data on the key monitoring nodes are obtained. The pressure time series data is iteratively calculated using a preset long short-term memory network to obtain fused time series features; A dynamic propagation path matrix is ​​constructed by combining the fusion time-series features and the static connection, and the pressure fluctuation transmission vector is extracted from the dynamic propagation path matrix. If the dimension of the pressure fluctuation transmission vector is higher than a preset dimension threshold, then the pressure fluctuation transmission vector is subjected to dimension reduction projection to obtain a simplified transmission path vector. The simplified transmission path vector is matched with the spatial network association features, and the wave pattern classification result is obtained based on the matching result. The information from the dynamic propagation path matrix is ​​fused using graph convolution operations on the fluctuation pattern classification results to determine the pipeline conduction mode.

4. The method for full lifecycle data management of oil and gas station pipelines according to claim 1, characterized in that, The extraction of pipeline coupling vectors from the spatial topology nodes using a preset attention mechanism includes: The historical fault data is mapped using the spatial topology nodes to obtain a time alignment matrix; A preset attention mechanism is used to assign the influence weights of the historical fault data to the time alignment matrix to determine the coupled feature tensor. For the coupling feature tensor, the spatial topology nodes are matched and fused with the temporal features of the historical fault data to obtain the pipeline coupling vector.

5. The method for managing the entire lifecycle data of oil and gas station pipelines according to claim 4, characterized in that, The step of performing graph convolution operations on the pipeline conduction pattern and the spatial topology nodes based on the heterogeneous node set to determine the pipeline anomaly graph includes: Based on the heterogeneous node set, graph convolution operation is performed on the pipeline conduction mode and the spatial topology nodes to construct a time matrix, and the preset conduction threshold is aligned with the time matrix to determine the basic graph of coupling anomaly distribution; If the dependency weight in the coupled anomaly distribution base map exceeds a preset dependency weight threshold, then the pressure anomaly fluctuation and physical spatial location are extracted from the coupled feature tensor, and the pressure anomaly fluctuation and the physical spatial location are mapped and fused to obtain the distribution extension map. Based on the distribution expansion map, the node association features and the abnormal pressure fluctuations are fused to obtain the abnormal distribution map; The abnormal distribution map is processed by graph convolution iteration. When the spatial distribution and temporal transmission law of the abnormal distribution map meet the preset transmission mode abnormality threshold, the pipeline abnormal map is determined.

6. The method for managing the entire lifecycle data of oil and gas station pipelines according to claim 5, characterized in that, The step of extracting risk tracing indicators from the pipeline anomaly map, and if the risk tracing indicators show an upstream transmission effect, then simulating the evolution trend through a preset recurrent neural network to obtain a pipeline damage prediction sequence, includes: Risk tracing indicators are extracted based on the abnormal distribution map, and the node association features and pipeline transmission patterns are fused through spatiotemporal correlation to determine the preliminary quantitative value of the upstream transmission effect. If the initial quantification value exceeds the preset quantification threshold, a preset recurrent neural network is used to evaluate the trend of the risk tracing indicator to obtain an intermediate simulation sequence. Based on the intermediate simulation sequence, the basic damage trend is determined, and the basic damage trend is aligned and fused with the spatial distribution and the temporal transmission law. If the fusion result is determined to meet the preset damage diffusion judgment rule, the damage probability, damage degree and transmission correlation of different pipe segments are quantified according to the time and space dimensions to generate an extended risk distribution set. The pipeline damage prediction sequence is obtained by iteratively processing the risk tracing indicators using the extended risk distribution set.

7. The method for full lifecycle data management of oil and gas station pipelines according to claim 1, characterized in that, The step of obtaining a pipeline health status score by fusing the topological relationship and the pressure time series data based on the pipeline damage prediction sequence includes: Based on the pipeline damage prediction sequence and the topological relationship, a spatiotemporal coupling tensor is constructed; The frequency domain features and state decay gradient of the pressure time series data are extracted using the spatiotemporal coupling tensor. If the state attenuation gradient is greater than a preset attenuation threshold, then the signal frequency domain features and the topological relationship are fused to determine the node vulnerability. Based on the node vulnerability, the pipeline damage prediction sequence is weighted to obtain health status characteristics; The health status features are quantified and converted using a preset scoring mapping rule to obtain a pipeline health status score.

8. The method for full lifecycle data management of oil and gas station pipelines according to claim 7, characterized in that, The process involves analyzing the matching degree between the pipeline conduction pattern and the historical fault data based on the pipeline health status score. If the matching degree is lower than a preset matching threshold, a pipeline risk alarm is generated, including: By fusing the spatiotemporal coupling tensor with the signal frequency domain features through the pipeline health status score, a comprehensive feature is obtained. From the comprehensive feature, the state decay gradient and node vulnerability correlation sequence in the pipeline conduction mode are extracted. The matching index is calculated by comparing the vulnerability association sequence with the damage evolution rate in the historical failure data. If the matching degree index is lower than the preset matching threshold, preliminary alarm data is generated by combining the topological relationship and the health status characteristics. By integrating the preliminary alarm data with the topological relationship and the pressure time series data, the alarm signal strength is determined. By combining the alarm signal strength with the preset scoring mapping rules, a pipeline risk alarm is generated.

9. A data management system for the entire lifecycle of oil and gas station pipelines, used to implement the data management method for the entire lifecycle of oil and gas station pipelines as described in any one of claims 1 to 8, characterized in that, include: The geographic construction module is used to acquire pipeline geographic information entities, extract node coordinates from the pipeline geographic information entities, construct topological relationships, and extract spatial network association features from the topological relationships through a preset graph neural network. The pressure transmission module is used to acquire pressure time series data based on the spatial network association characteristics, and to fuse the pressure time series data using a preset long short-term memory network to determine the pipeline transmission mode. The spatiotemporal coupling module is used to acquire historical fault data, extract maintenance frequency features and fluctuation attenuation features from the historical fault data, and quantitatively evaluate the pipeline conduction mode based on the maintenance frequency features and fluctuation attenuation features. If the quantitative evaluation value exceeds the preset conduction mode abnormality threshold, spatial topology nodes are extracted from the pipeline conduction mode, and pipeline coupling vectors are extracted from the spatial topology nodes through a preset attention mechanism. An anomaly graph module is used to extract node association features from the pipeline coupling vector, and to concatenate the maintenance frequency features and the fluctuation attenuation features with the node association features to obtain a heterogeneous node set. Based on the heterogeneous node set, a graph convolution operation is performed on the pipeline conduction mode and the spatial topology nodes to determine the pipeline anomaly graph. The damage prediction module is used to extract risk tracing indicators from the pipeline anomaly map. If the risk tracing indicators show an upstream transmission effect, the evolution trend is evaluated by simulating the evolution trend through a preset recurrent neural network to obtain a pipeline damage prediction sequence. The health scoring module is used to obtain a pipeline health status score by fusing the topological relationship and the pressure time series data based on the pipeline damage prediction sequence. The risk warning module is used to analyze the matching degree between the pipeline conduction mode and the historical fault data based on the pipeline health status score. If the matching degree is lower than a preset matching threshold, a pipeline risk alarm is generated.

Images