Underground pipe network facility service life prediction system and method based on multi-parameter time sequence analysis

Abstract

The invention relates to the technical field of digital operation and maintenance, in particular to an underground pipe network facility life prediction system and method based on multi-parameter time sequence analysis, and the method comprises the steps: generating a holographic state set through a multi-scale window and a tensor filling algorithm, building a pipe network topology through three-layer logic verification, and paying attention to network quantification risk conduction characteristics based on a space-time diagram. And performing state recursion and continuous integration by using a particle filter algorithm to predict the remaining life. According to the method, heterogeneous monitoring data is converted into a low-rank tensor model, and sparse graph matrix operation is combined, so that the calculation complexity and memory overhead in a high-dimensional spatial-temporal feature extraction process are effectively reduced; meanwhile, based on a numerical integration strategy of a Bayesian sequence Monte Carlo method, the particle degradation problem in nonlinear system state estimation is solved, and convergence and numerical stability of a residual life prediction result in a computer simulation environment are guaranteed.

Description

Technical Field

[0001] This invention relates to the field of digital operation and maintenance technology, specifically to a system and method for predicting the lifespan of underground pipeline facilities based on multi-parameter time series analysis. Background Technology

[0002] Existing underground pipeline life prediction technologies mostly rely on corrosion rate calculations using a single physical model or simple regression analysis based on discrete monitoring data; However, in practical applications, pipeline network data exhibits high heterogeneity and fragmentation. High-frequency time-series data such as flow and pressure, discrete records from manual inspections, and static geographic information are often isolated "data silos," making effective integration difficult. More critically, existing methods typically treat each pipe segment in the network as an independent entity for isolated prediction, ignoring the complex topological network of the pipeline, where nodes have close interactions such as physical connections, media flow, and electrochemical corrosion. This fails to quantify the cumulative effect of cascading failure risks from adjacent nodes through the pipeline network topology. Furthermore, traditional models mostly rely on static parameter fitting, making it difficult to dynamically correct prediction trajectories based on real-time environmental load changes, resulting in significant deviations in the prediction results for the remaining life of the pipeline network under complex operating conditions.

[0003] To address the technical problem of inaccurate predictions of the remaining life of underground pipeline facilities due to the cascading effects of single-point failures, a life prediction system and method based on multi-parameter time series analysis is proposed. Summary of the Invention

[0004] This invention aims to provide a system and method for predicting the lifespan of underground pipeline facilities based on multi-parameter time-series analysis. By constructing a holographic data mapping and spatiotemporal risk transmission network, it achieves accurate dynamic prediction of the remaining lifespan of pipeline facilities under the coupling effect of multi-dimensional physical fields.

[0005] To achieve the above objectives, the present invention provides the following technical solution: A system and method for predicting the lifespan of underground pipeline facilities based on multi-parameter time series analysis, including: The acquisition module collects and completes high-frequency time-series data from the flow and pressure monitoring system, discrete records from manual inspections, and static vector data from the geographic information system to generate a holographic status set of the pipeline network. The knowledge graph construction module collects historical maintenance logs, geological exploration reports, and construction records. It uses natural language processing to extract entities and their relationships, including pipe sections, valves, soil types, and construction parties, to construct a pipeline network knowledge graph. It also performs logical verification and implicit relationship deduction to establish a pipeline network topology. Network construction module: Maps the holographic state set to the corresponding nodes of the pipeline topology network, constructs a spatiotemporal graph attention network, dynamically allocates the aggregation weight of each node through the attention mechanism, and calculates the risk transmission characteristics of a single point failure to adjacent nodes; Prediction module: Constructs a probability transfer model that includes environmental load, corrosion mechanism and material properties; using the risk transmission characteristics as external observation variables, and based on the probability transfer model, uses a particle filter algorithm to recursively update the state probability distribution of each node on the time axis, performs continuous time integration on the pipeline state, and outputs the remaining life prediction value.

[0006] Preferably, the acquisition module generates a holographic state set of the pipeline network, specifically including: By aligning the time granularity of the high-frequency time-series data, discrete records, and static vector data through a multi-scale sliding window, heterogeneous data is mapped into a multimodal array. The tensor filling algorithm is used to mine the correlation and time period similarity of the pipeline network spatial topology, infer and fill the numerical missing values ​​in the multimodal array, and generate the holographic state set.

[0007] Preferably, the knowledge graph construction module for constructing the pipeline network specifically includes: Named entity recognition algorithms are used to extract three types of node entities from the historical maintenance logs, geological exploration reports, and construction records: asset nodes including pipe sections and valves, environmental nodes including soil type and groundwater level, and management nodes including construction parties and maintenance records. Directed edges connecting these node entities are constructed based on engineering semantic logic. These directed edges include at least: topological connectivity edges representing the direction of medium flow and physical connections, environmental coupling edges representing the corrosive effect of environmental factors on pipe materials, and construction traceability edges representing construction quality and responsibility attribution. The asset nodes, environmental nodes, and management nodes are instantiated and connected through these topological connectivity edges, environmental coupling edges, and construction traceability edges to generate the pipeline network knowledge graph.

[0008] Preferably, the step of performing logical verification and implicit relationship deduction to establish the pipeline network topology specifically includes a three-layer verification system: Pressure rating system logic verification: Based on the gas pipeline network pressure rating standard, the connection relationship of pipeline nodes is traversed to identify and block abnormal topologies in which high and low pressure pipeline sections are directly interconnected without pressure regulating facilities, ensuring that physical connections comply with pressure reduction and rating system rules; Electrocorrosion connectivity derivation: For spatial discontinuities in geographic information system data, cluster analysis is performed using the spatial continuity characteristics of cathodic protection potential gradient to derive physically separated but electrochemically conductive implicit connections, thus completing the current loop logic used for corrosion rate calculation. Hydraulic topology flow direction correction: Based on the law of conservation of nodal flow and pipe diameter constraints, the contradiction between medium flow direction and pipe diameter gradient is detected, reverse flow or loop deadlock path is corrected, and a pipe network topology network with complete hydraulic and electrical logic is generated.

[0009] Preferably, the network construction module calculates the risk transmission characteristics specifically including: Based on the node identifiers of the pipeline network knowledge graph, the multimodal array in the holographic state set is mapped one by one to the asset nodes, environment nodes, and management nodes of the pipeline network topology to form a spatiotemporal embedding vector. A spatiotemporal graph attention network is constructed, in which the time series features of each node are used as input. The spatiotemporal dependencies are captured through a multi-head attention layer, and the neighborhood aggregation weight is dynamically calculated. The weight allocation considers the topological distance between nodes, the environmental coupling strength, and the similarity of historical failures. A single-point failure scenario is simulated, and the failure signal is propagated using the aggregation weight. The risk transmission characteristics of adjacent nodes are quantified. The risk transmission characteristics include the failure probability increase, the propagation path attenuation coefficient, and the cascading failure threshold.

[0010] Preferably, the construction of the probabilistic transfer model including environmental load, corrosion mechanism, and material properties specifically includes: A state transition equation for pipe wall degradation based on material properties and service life is established, and a power-law function is used to simulate the natural decay process of pipe wall thickness over time. An external disturbance state equation based on soil corrosivity and ground dynamic load is established to quantify the nonlinear accelerating effect of environmental factors on the pipe wall degradation rate. The external disturbance state equation is coupled to the pipe wall degradation state transition equation as a random disturbance term to form a probabilistic transfer model describing the evolution of the health status of pipeline nodes over time.

[0011] Preferably, the prediction module outputs the remaining lifetime prediction value specifically including: A set of random particles representing the potential health status of pipeline network nodes is generated. Forward prediction of the particle set is performed at time steps based on the probability transfer model. The likelihood between the forward prediction result and the risk transmission characteristic as an observed variable is calculated. The particle weights are updated and resampled based on the likelihood, and the state probability density function of each node is corrected. A pipeline failure threshold is set, and the corrected state probability density function is integrated over time. The cumulative probability distribution of the state curve reaching the failure threshold is calculated, and the moment when the cumulative probability reaches a preset confidence interval is determined as the remaining lifetime prediction value.

[0012] Preferably, the prediction module outputs the remaining lifetime prediction value specifically including: generating a set of random particles representing the current health status of pipeline nodes; performing single-step forward prediction based on the probability transfer model; correcting particle weights using the risk transmission feature and performing resampling to obtain a posterior particle distribution reflecting the true state at the current moment, thus completing the real-time update of the state probability distribution; using the posterior particle distribution as the initial condition, performing multi-step recursive iteration on the time axis using the probability transfer model, and generating a trajectory cluster describing the pipe wall degradation trend of each particle within a future time window through numerical integration; setting a pipeline failure threshold, statistically analyzing the time points when each trajectory in the trajectory cluster reaches the failure threshold, calculating the probability distribution interval of the time points, and determining the expected value of the interval as the remaining lifetime prediction value; establishing an index mapping relationship between the remaining lifetime prediction value and the corresponding pipeline node identifier, and writing the mapped data into the storage partition of a non-relational database for persistent storage.

[0013] Preferably, the system further includes a vulnerability identification module, which specifically performs the following steps: using the physical attribute data in the holographic state set to calculate the endogenous health index of the node; based on the risk transmission characteristics and combined with the dynamically allocated aggregation weights, quantifying the weighted impact value of neighborhood risks on the central node, and generating an exogenous stress index; constructing a risk comparison discrimination function to screen out nodes that meet the conditions of the endogenous health index being higher than the health threshold and the exogenous stress index being higher than the warning threshold, marking them as vulnerable nodes, and generating inspection and maintenance instructions for vulnerable nodes.

[0014] A method for predicting the lifespan of underground pipeline facilities based on multi-parameter time series analysis includes: Collect and supplement high-frequency time-series data from flow and pressure monitoring systems, discrete records from manual inspections, and static vector data from geographic information systems to generate a holographic status set of the pipeline network; Historical maintenance logs, geological exploration reports, and construction records are collected. Natural language processing is used to extract entities and their interrelationships, including pipe sections, valves, soil types, and construction parties, to construct a pipeline network knowledge graph. Logical verification and implicit relationship deduction are performed to establish a pipeline network topology. The holographic state set is mapped to the corresponding node of the pipeline network topology to construct a spatiotemporal graph attention network. The aggregation weight of each node is dynamically allocated through the attention mechanism to calculate the risk transmission characteristics of a single point failure to adjacent nodes. A probability transfer model incorporating environmental load, corrosion mechanism, and material properties is constructed. Using the risk transmission characteristics as external observation variables, and based on the probability transfer model, the state probability distribution of each node is recursively updated on the time axis using a particle filter algorithm. The pipeline state is then continuously integrated over time to output the predicted remaining life value.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By aligning time granularity through multi-scale sliding windows and utilizing tensor filling algorithms to mine spatiotemporal correlations, the problem of inconsistent frequencies and missing values ​​among traffic monitoring, manual inspection, and GIS data is effectively solved, generating a standardized holographic state set, which provides a highly reliable data foundation for subsequent model calculations.

[0016] 2. By combining a three-layer logical topology network constructed with a knowledge graph and a spatiotemporal graph attention mechanism, it can identify implicit correlations brought about by physical connections and media flow direction, dynamically calculate the risk impact weight of single-point failure on adjacent nodes, break through the limitations of traditional isolated pipe segment prediction, and significantly improve the network risk identification capability.

[0017] 3. By constructing a probabilistic transfer model that couples environmental load and corrosion mechanism, and by using a particle filtering algorithm to incorporate real-time risk transmission characteristics into the state recursion process, continuous time integral deduction from current state estimation to future degradation trajectory is realized, ensuring that the remaining life prediction value can be corrected in real time as the actual operating conditions of the pipeline network change.

[0018] 4. By utilizing the existing fiber optic network resources of telecommunications operators for neural sensing monitoring, real-time capture of external intrusion events (such as excavation and vehicle vibration) along the pipeline was achieved at low cost. This integrated sensor data source effectively solves the problem of difficulty in obtaining external load data in real time in traditional models, providing accurate dynamic environmental boundary conditions for the probabilistic transfer model. Without changing the original topology computing architecture, it significantly enhances the system's ability to predict situations where third-party damage leads to a sudden reduction in pipeline lifespan. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the steps of the underground pipeline facility life prediction system and method based on multi-parameter time series analysis of the present invention. Figure 2 This is a schematic diagram of the process for generating a holographic state set of the pipeline network according to the present invention; Figure 3 This is a schematic diagram of the process for outputting the predicted remaining lifetime value of this invention. Detailed Implementation

[0020] 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.

[0021] Please see Figures 1 to 3This invention provides a system and method for predicting the lifespan of underground pipeline facilities based on multi-parameter time series analysis, referring to... Figure 1 The flowchart and technical solution are as follows: The acquisition module collects and completes high-frequency time-series data from the flow and pressure monitoring system, discrete records from manual inspections, and static vector data from the geographic information system to generate a holographic status set of the pipeline network. The knowledge graph construction module collects historical maintenance logs, geological exploration reports, and construction records. It uses natural language processing to extract entities and their relationships, including pipe sections, valves, soil types, and construction parties, to construct a pipeline network knowledge graph. It also performs logical verification and implicit relationship deduction to establish a pipeline network topology. Network construction module: Maps the holographic state set to the corresponding nodes of the pipeline topology network, constructs a spatiotemporal graph attention network, dynamically allocates the aggregation weight of each node through the attention mechanism, and calculates the risk transmission characteristics of a single point failure to adjacent nodes; Prediction module: Constructs a probability transfer model that includes environmental load, corrosion mechanism and material properties; using the risk transmission characteristics as external observation variables, and based on the probability transfer model, uses a particle filter algorithm to recursively update the state probability distribution of each node on the time axis, performs continuous time integration on the pipeline state, and outputs the remaining life prediction value.

[0022] Example 1 This embodiment describes the implementation scenario of life prediction for urban medium-pressure gas pipeline facilities. The system's acquisition module holographically integrates heterogeneous monitoring data, the map construction module constructs and verifies a topological network with hydraulic and electrical logic, the network construction module quantifies the neighborhood risk transmission characteristics, and the prediction module performs remaining life prediction based on a probability transfer model. This enables full-process control of underground gas pipeline facilities, from multi-dimensional state perception and digital topology reconstruction to failure evolution simulation.

[0023] First, the acquisition module obtains high-frequency flow and pressure monitoring data with a sampling frequency of 1Hz through the preset interface of the industrial control system. The specific values ​​cover the real-time operating pressure of 0.4MPa and the instantaneous flow rate sequence of 5200 cubic meters per hour. At the same time, it retrieves discrete records of manual inspections, including pipe wall thickness reduction of 0.12 mm and anti-corrosion layer peel strength of 120 Newtons per centimeter, as well as static vector data from the geographic information system, including pipe material L360 steel, burial depth of 1.8 meters, and three-dimensional topological coordinates. In addition, in order to make up for the blind spots in spatial coverage of traditional point sensors, the acquisition module also connects to external environmental monitoring data based on distributed acoustic sensing technology. This data uses the existing underground communication optical cables in the city as the sensing carrier, without the need to lay new sensing equipment along the pipeline. The system uses optical fiber sensing equipment to demodulate the backscattered signal sensed along the optical cable to obtain high-frequency vibration time series data. The specific content collected includes two categories: First, vibration acoustic signature data: collecting vibration amplitude, frequency, and phase information distributed along the pipeline, with the sampling rate adaptable to system requirements (e.g., feature values ​​after edge computing from a 100MHz fiber optic sampling rate); Second, external operation event data: collecting the business event classification results after processing by edge computing nodes, specifically including external operation events that may cause mechanical damage to the pipeline, such as excavators hitting the ground, pneumatic drills impacting concrete, cutting machine operations, and vehicles crossing potholes. This data includes the type of event, occurrence time, duration, and corresponding optical cable channel location (distance).

[0024] Reference Figure 2 A flowchart illustrating the generation of the holographic state set of the pipeline network is provided. Further, a multi-scale sliding window operator is used to perform time-granular consistency alignment on the acquired high-frequency time-series data, discrete records from manual inspections, and static vector data. By setting sliding sampling step sizes of 1 second, 1 hour, and 1 day, the high-frequency pressure fluctuation data is non-linearly aligned with the low-frequency updated inspection records, thereby mapping the heterogeneous data from different systems into a multi-modal feature array with dimensions of 3600×24×12. For the introduced fiber optic sensing data, the system also uses a multi-scale sliding window for alignment processing. Due to the massive amount of data generated by the fiber optic sensing device (e.g., 50MB per second), the acquisition module prioritizes reading the CSV format result file after edge computing processing, mapping discrete abnormal events (e.g., excavation damage alarms) to low-frequency time windows (e.g., days or hours), and mapping continuous environmental vibration energy values ​​to high-frequency time windows (e.g., seconds). By calibrating the relationship between the optical cable channel number and the GIS coordinates of the pipeline network, the linear vibration data sensed by the optical fiber is accurately mapped to the corresponding pipe segment nodes in the pipeline network topology, thereby expanding the physical dimension of the holographic state set to include dynamic environmental load information around the pipeline.

[0025] Specifically, in the process of multi-scale alignment, in response to the time span problem in discrete inspection records, the algorithm uses Lagrange interpolation logic to convert the monthly updated corrosion depth values ​​to a second-level time coordinate system synchronized with the flow and pressure monitoring data, and establishes an index based on the globally unique identifier of the pipeline in the static vector data, so that each time node slice in the multimodal array contains complete mechanical load parameters, material loss parameters, and pipe segment geometric parameters.

[0026] The holographic state set is specifically stored in the form of a three-dimensional tensor. The three-dimensional tensor is discretely sampled in the time dimension at second-level or hour-level time steps T, indexed by the unique identifier ID of the pipe segment or node in the spatial dimension, and features physical quantities directly related to the pipeline network's operating status as characteristic components in the attribute dimension. These physical quantities include at least operating pressure, instantaneous flow rate, pipe wall thinning amount, anti-corrosion layer insulation potential, burial depth, and soil corrosivity indicators. For different engineering scenarios, those skilled in the art can select specific physical quantities related to the pipeline network's health status to fill the attribute dimensions while keeping the three-dimensional structure of "time step-T / space-ID / physical attribute" unchanged, thereby reconstructing the corresponding holographic state set under a unified data structure framework.

[0027] Furthermore, a tensor imputation algorithm is used to intelligently infer and fill in the missing values ​​in the multimodal array. This algorithm adopts a tensor imputation model architecture based on a deep autoencoder. The model architecture specifically includes an input layer containing 1200 neurons, three consecutive symmetric hidden layers containing 512, 256, and 512 neurons respectively, and an output layer. The hidden layers use the Adam optimizer to update the weights and set a neuron dropout probability of 0.15. In actual implementation, the deep autoencoder uses historical multimodal array fragments of complete monitoring sections as training samples and adopts an offline training method to obtain stable network weights. The loss function during training consists of two linear combinations: one is a reconstruction error term, which is used to constrain the numerical deviation between the output tensor and the input tensor at the observed positions; the other is a physical constraint regularization term, which is used to constrain the imputed flow data to satisfy Kirchhoff's flow conservation law. The physical constraint regularization term calculates the residual of "total inflow minus total outflow" for each node within a given time step, and proportionally accumulates any residual exceeding a preset tolerance range into the loss function. After training convergence, at spatiotemporal locations with missing data, the system only needs to input the corresponding tensor slice into the trained autoencoder, using the reconstruction result of the output layer as the missing value imputation result, thus balancing statistical features and physical equilibrium constraints without altering the tensor structure.

[0028] Specifically, the tensor filling algorithm first reorganizes the mapped multimodal array into a third-order tensor containing three dimensions: time step, spatial pipe segment ID, and physical attribute features. It then uses the spatial topology correlation operator to calculate the pressure transmission correlation coefficient of 0.88 between the target faulty pipe segment and each node within 50 meters of it. Combined with the time period similarity operator, it extracts the average flow load features of the pipe segment in the same time period over the past 30 calendar days. These features are then input into the deep autoencoder model as prior conditions. Through the model's hidden layer, the hidden variables of the missing features are reconstructed, and the pressure gap of 0.35 MPa caused by the sensor communication interruption is filled in the array. Finally, a holographic state set covering the monitoring points throughout the entire life cycle of the pipeline network is generated.

[0029] Furthermore, the method of inferring and filling numerical missing values ​​using the tensor imputation algorithm specifically includes: constructing a three-dimensional tensor containing time dimension, spatial node dimension, and physical attribute dimension; introducing a physical constraint regularization term into the loss function of tensor decomposition; the physical constraint regularization term is constructed based on Kirchhoff's flow conservation law; when the residual between the imputed node flow and the sum of the flow of adjacent nodes exceeds the physical tolerance threshold, a penalty is applied to the gradient descent direction of the objective function until a complete dataset that satisfies the hydraulic balance condition of the pipeline network is generated. For low-pressure gas distribution networks with only sparse pressure monitoring points, the system constructs a three-dimensional tensor including the number of nodes, time step, and attribute type. When performing tensor normalization multidimensional decomposition, a physical penalty term is added to the optimization objective. For any target node, its physical constraint is defined as follows: the sum of the inflow to the node minus the sum of the outflow should equal the actual gas consumption of the node. If the algorithm predicts the pressure value of a missing node, resulting in an excessive pressure difference with its adjacent nodes, and thus causing the calculated flow to violate the conservation law, the penalty term increases sharply. The regularization coefficient is set to 0.5, and the physical tolerance threshold is set to 0.05 MPa. This threshold is a judgment standard set based on the pipeline pressure balance requirements and engineering measurement accuracy. When using algorithms to fill in missing pressure monitoring data, the system uses this value as a critical point to measure whether the filling result conforms to common sense. If the filled pressure value causes an excessive pressure difference between adjacent nodes (exceeding 0.05 MPa), it will trigger a logical data conflict. In this case, the system will automatically apply a penalty and correct the value. This setting effectively solves the problem that pure data interpolation may produce "values ​​that look good but do not conform to actual operating rules," ensuring that the filled virtual pressure conforms to statistical laws and strictly matches the actual operating conditions of the pipeline network. By using the projection gradient descent method for iterative solution, the filled virtual pressure value not only conforms to statistical laws but also strictly satisfies fluid dynamic constraints. This solves the defect of pure data-driven interpolation algorithms producing data that "violates physical common sense" in sparse monitoring networks. On a test set with a monitoring point missing rate of 40%, the physical consistency error rate of the filled data was reduced by 25.3%. Furthermore, because the input data satisfies hydraulic balance, the convergence speed of the subsequent lifetime prediction model is improved by 18%.

[0030] By aligning the temporal granularity of heterogeneous data using a multi-scale sliding window and deeply mining the implicit correlation between pipeline spatial topology and time period through a tensor filling algorithm, the problem of missing and inconsistent monitoring data caused by sensor failure or inspection interruption is effectively solved, providing a standardized holographic data foundation for subsequent models.

[0031] The original text dataset containing 5,000 historical maintenance logs, 200 geological exploration reports, and 150 construction records is obtained through a preset database interface, and then the dataset is uniformly converted into a UTF-8 format character sequence and input into the named entity recognition model.

[0032] Furthermore, the named entity recognition model adopts a deep learning network based on a transformer architecture. Its bottom layer contains 12 self-attention encoding layers, each layer has 768 hidden units and 12 parallel attention heads, which map the input character sequence into a 768-dimensional dense semantic vector. Then, the vector is input into a bidirectional long short-term memory network layer containing 256 hidden units. This layer is used to bidirectionally aggregate the contextual features in the text to obtain a feature matrix that incorporates the context.

[0033] Specifically, the feature matrix is ​​input into the conditional random field layer. This layer maintains a 9×9 label transition probability matrix and calculates the globally optimal label prediction path, thereby accurately extracting asset entity nodes, environmental entity nodes, and management entity nodes from the text. For example, the identified asset node includes a steel pipe with a nominal diameter of 300 mm numbered RQ-GD001, the environmental node includes an acidic soil area with a pH value of 4.2 and a resistivity of 15 ohm-meters, and the management node includes a pipeline engineering company with a construction qualification level of Level 1.

[0034] Furthermore, based on the pipeline connectivity index table provided by the geographic information system database, the physical connection relationship between each asset entity node is extracted. By reading the coordinate values ​​of the starting end of the pipeline asset node (116.39, 39.91) and the coordinate values ​​of the ending end (116.40, 39.92), a topological connection edge is established between the asset entities associated with the two coordinate points. The direction of this edge is from the high-pressure end (0.4 MPa) in the pressure monitoring data to the low-pressure end (0.2 MPa) to characterize the actual flow direction of the gas medium.

[0035] Specifically, an environmental coupling edge is constructed using a spatial geometric judgment algorithm. First, the 3D center coordinates of the asset entity nodes are acquired, and the coordinate sequence of the polygon vertices covered by the environmental nodes in the geological exploration report is retrieved. The shortest Euclidean distance from the center coordinate point to each boundary of the polygon is calculated—that is, the square root of the sum of the squares of the differences in the horizontal and vertical coordinates of two coordinate points. When this distance is less than a preset 5-meter influence radius threshold, which is a judgment standard based on the safety interference radius of the pipeline space and the influence range of the geological environment, this setting references engineering experience in the permeability of the soil's physical and chemical properties around underground pipelines and ensures that the model can accurately capture the nonlinear accelerating effect of environmental loads on pipe wall degradation. An environmental coupling edge is created between the corresponding asset node and the soil environment node, and values ​​based on the corrosion rate model, such as an average annual wall thickness reduction of 0.05 mm, are injected into this edge.

[0036] Furthermore, a construction traceability edge is constructed by executing a text fuzzy matching algorithm. Specifically, the contract batch number string in the construction record entity and the enterprise unified social credit code in the management node entity are compared to calculate the matching degree. The cosine similarity algorithm is used to calculate the cosine value of the angle between the feature vectors of the two strings. When this score exceeds 0.95, it is set to 0.95 to ensure that a directed association is established between the construction unit management node and the contracted pipe segment asset node only when the string feature vectors highly overlap. This high score setting can effectively filter out interference caused by input errors or format differences, ensuring that the logical chain of "construction quality and responsibility attribution" in the knowledge graph has extremely high engineering credibility. A directed construction traceability edge is established between the construction unit management node and the contracted pipe segment asset node.

[0037] Specifically, the aforementioned asset nodes, environment nodes, and management nodes are used as vertices, and the topological connectivity edges, environmental coupling edges, and construction traceability edges are used as associations. All of these are uniformly loaded into the graph database for instantiation and storage.

[0038] Furthermore, during the instantiation process, each asset node is configured with 12 attribute fields, including a globally unique identifier, pipe size (nominal diameter), pipe material grade, pipe wall thickness, pipe burial depth, installation / service date, three-dimensional topological coordinates, operating pressure rating, real-time / instantaneous flow rate, anti-corrosion layer attributes, material physical and mechanical parameters, and material corrosion coefficient / power law exponent. Each topological connection edge is also configured with a dynamic attribute containing an instantaneous flow rate of 5200 cubic meters per hour.

[0039] Specifically, by using the graph computing operators of the graph database to perform the node-relationship-node association mapping, the semantic alignment of data across business systems is completed, and finally a knowledge graph of underground gas pipeline network with complete engineering logic constraints is generated, providing underlying topological support for subsequent failure evolution analysis.

[0040] By using natural language processing technology, unstructured engineering documents are transformed into semantic networks containing three-dimensional entities of assets, environment, and management. By establishing directed associations of topological connectivity, environmental coupling, and construction tracing, multidimensional risk factors hidden in historical records are revealed, realizing a structured mapping from text records to knowledge logic.

[0041] The first-level pressure rating system logic verification is performed on the pre-constructed pipeline network knowledge graph. The topological connectivity edges between all asset nodes are traversed using a graph computing engine, and the rated pressure attribute value of each starting node and ending node is extracted. This value consists of a high-pressure rating of 0.8 MPa, a medium-pressure rating of 0.4 MPa, and a low-pressure rating of 0.02 MPa.

[0042] Furthermore, during the verification process, the system constructs a logical judgment operator to search for all adjacent node pairs in the graph that have different pressure attributes, and detects whether there are asset entities marked as pressure regulating stations or pressure regulating boxes connected in series between these pairs of nodes. If it is detected that a certain connecting edge directly connects a 0.8MPa pipe segment node and a 0.4MPa pipe segment node and there is no pressure regulating entity in between, it is determined that the local topology violates the graded rule of gradually decreasing gas pressure.

[0043] Specifically, for topological connections that are determined to be abnormal, the system automatically records their spatial coordinates 116.39 and 39.91 and outputs them to the logical conflict queue. Then, it cuts off the abnormal connected edge by executing the graph pruning algorithm, and re-searches for the nearest pressure regulating facility node based on the original vector path in the geographic information system to perform logical completion, thereby ensuring that the generated pipeline topology network fully complies with the physical operation specifications at the pressure level.

[0044] Furthermore, for spatial breakpoints ranging from 0.5 meters to 2.0 meters in length caused by signal blind spots or data loss in the geographic information system data, a second layer of electrical corrosion connectivity derivation is performed. This process first retrieves the pipe-to-ground potential data transmitted back from the monitoring piles along the entire line from the cathodic protection management platform. This data contains potential values ​​in the range of -1.2V to -0.85V, and correlates them with the three-dimensional coordinates of the center points of the pipe sections on both sides of the breakpoint.

[0045] Specifically, a latent relation derivation model based on a multilayer perceptron is constructed. The model architecture includes a 5-dimensional input layer to receive a feature vector consisting of the difference in horizontal distance (0.8 m), vertical depth (0.1 m), pipe diameter (100 mm), resistivity difference of anti-corrosion layer (15000 ohm square meters), and cathodic protection potential difference (0.05 V) between the pipe segments on both sides of the breakpoint; a feature extraction layer containing 64 hidden neurons; and an output layer.

[0046] Furthermore, the hidden layer of the multilayer perceptron model uses the ReLU activation function for nonlinear transformation, and introduces the Adam optimizer to calculate the contribution weight of the input features to electrical connectivity in each training cycle. By performing a fully connected mapping on the weights, the output layer uses the Sigmoid function to calculate the probability score of two pipe segment nodes belonging to the same electrochemical circuit.

[0047] Specifically, when the probability score exceeds a preset confidence threshold of 0.98, this value serves as a stringent filtering condition for confirming the existence of a hidden connection. Only when the model integrates multiple features such as burial depth difference, pipe diameter difference, and cathodic protection potential difference, and the output conduction probability reaches 98% or higher, does the system determine that the two physically separated asset nodes are in the same circuit in terms of electrochemical properties, and automatically completes the current loop logic. It determines that although the asset nodes on both sides of the spatial breakpoint are physically separated, they are in a conductive state in terms of electrochemical properties, and automatically creates a hidden electrical connection edge between these two nodes in the pipeline network knowledge graph, injecting virtual electrical parameters including a loop current of 0.5A and a polarization potential of -0.95V.

[0048] Furthermore, the derived implicit connectivity was verified by spatial clustering using the K-means clustering algorithm. This algorithm uses the cathodic protection potential gradient as the clustering feature and groups adjacent pipe segment nodes with potential differences within 0.01V into the same electrically connected domain, thereby completing the corrosion current loop logic that was originally interrupted by spatial discontinuities and providing closed circuit topology support for subsequent corrosion rate calculations for specific pipe segments.

[0049] Furthermore, the third layer of hydraulic topology flow direction correction processing is performed. By calling the preset hydraulic balance calculation engine, the real-time flow data of each asset node in the pipeline topology network is extracted in the holographic state set as 5200 cubic meters per hour and the pressure attribute value of 0.4 MPa.

[0050] Specifically, based on the law of conservation of flow at nodes, a dynamic balance check of inflow and outflow is performed on each node in the pipeline network topology. By reading the direction attributes and flow values ​​of all topological connecting edges connected to a specific node, the difference between the total flow into and out of that node is calculated. If this difference exceeds a preset 5% error threshold, which is a logical verification standard set based on the law of conservation of flow at nodes and engineering fault tolerance experience, the system extracts real-time flow data from each node in the pipeline network during hydraulic balance verification and calculates the numerical difference between inflow and outflow. If the proportion of this flow difference exceeds the preset 5% error range, the system determines that the flow direction logic of that node is abnormal, i.e., it is identified as reverse flow or a loop deadlock path. This setting takes into account the normal reading fluctuations of sensors during high-frequency acquisition and ensures that the pipeline network topology strictly conforms to the medium dynamics characteristics of the actual gas transmission and distribution system; therefore, it is determined that the flow direction logic associated with that node is abnormal.

[0051] Furthermore, a flow determination algorithm based on the feature extraction module is introduced during the verification process. This algorithm first constructs a model containing an input layer, a hidden layer, and an output layer. The input layer receives a 12-dimensional feature vector, and the hidden layer contains 128 neurons and uses the Sigmoid activation function to process the feature weights.

[0052] Specifically, the 12-dimensional feature vector includes " Geometric features (3D): Pipe diameter (300mm), wall thickness (8mm), pipe length (150m); Spatial features (3D): starting elevation, ending elevation, node height difference (1.5m); Hydraulic operation characteristics (4 dimensions): operating pressure (0.4MPa), outlet pressure (0.38MPa), pressure gradient, instantaneous flow rate (5200m³ / h); Material / environmental characteristics (2D): roughness coefficient, medium viscosity (or Reynolds number related parameters).

[0053] Specifically, by calculating the inner product of the pressure gradient vector and the pipe diameter gradient vector between adjacent pipe sections, it is determined whether the flow direction of the medium conforms to the engineering logic of flowing from the large pipe to the small pipe diameter and from the high pressure end to the low pressure end. When an abnormal gradient is detected where the pipe diameter changes from 200 mm to 300 mm and the flow direction is opposite to the pressure drop direction, it is determined to be a flow direction reversal contradiction.

[0054] Furthermore, the depth-first search algorithm is used to traverse the ring structure in the pipeline network topology to detect whether there is a ring deadlock path consisting of three or more connected edges and the sum of the flow vectors is zero. The identified closed loop information and its associated asset node ID RQ-GD105 are then output to the correction module.

[0055] Specifically, the correction module performs a flip operation on the direction attributes of topological connected edges that are determined to be contradictory, based on the calculated flow direction confidence weight, or forcibly reconstructs the ring topology into a directed acyclic graph structure that conforms to the dynamic characteristics of the medium by introducing logical breakpoints with unidirectional flow constraints in the deadlock path, and finally generates a pipe network topology with complete hydraulic and electrical logic.

[0056] A three-layer verification system was constructed that integrates pressure level compliance, electrochemical corrosion circuit and hydraulic flow conservation. It corrects spatial breakpoints and flow direction logic conflicts in GIS data from both physical and chemical dimensions, ensuring that the generated pipeline topology strictly conforms to the actual physical laws of gas transmission and distribution system operation.

[0057] Based on the globally unique identifier strings of asset entities, environmental entities, and management entities stored in the pipeline network knowledge graph, a 3600×24×12-dimensional multimodal array corresponding to the holographic state set is extracted. Using a hash index algorithm, dynamic data such as real-time operating pressure of 0.42MPa, instantaneous flow rate of 5200 cubic meters per hour, and pipe wall thinning of 0.12 mm at each time sampling step are concatenated column by column into the initial feature matrix of asset nodes, environmental nodes, and management nodes.

[0058] Furthermore, the initial feature matrix is ​​mapped to a 128-dimensional continuous spatiotemporal embedding vector using a fully connected layer. This fully connected layer contains 12 input units and 128 output neurons, and L2 regularization is used to suppress data noise introduced by sensor sampling errors, thereby providing standardized feature inputs for subsequent graph computation.

[0059] Specifically, a spatiotemporal graph attention network is constructed. This network model consists of two layers of temporal-gated convolutional layers and three layers of spatial multi-head graph attention layers stacked alternately. The temporal-gated convolutional layers contain 64 convolutional kernels with a kernel size of 3, which are used to extract the pressure fluctuation trend and flow evolution characteristics of the pipeline nodes in the past 24 consecutive sampling periods.

[0060] Furthermore, eight independent attention heads are set in the spatial multi-head graph attention layer, with the hidden feature dimension of each head set to 32. By mapping the feature vectors output by the temporal convolutional layer to a high-dimensional space, the spatial interaction effects between nodes of different pipe segments are captured.

[0061] Specifically, during the calculation of the multi-head attention layer, the algorithm extracts the spatiotemporal embedding vectors of the source node and the target node, and simultaneously introduces three values ​​of 1 for each dimension as edge auxiliary features, including the physical topological distance between nodes of 150 meters, the environmental coupling strength coefficient of 0.75, and the historical failure similarity score of 0.82 calculated based on the Jaccard similarity algorithm.

[0062] Furthermore, the source node vector, the target node vector, and the three auxiliary feature vectors mentioned above are concatenated using tensors to form a joint feature vector with a dimension of 259. This joint feature vector is then input into a scoring function containing a single-layer feedforward neural network. This feedforward network consists of 259 input units and 1 output neuron and is processed using the LeakyReLU activation function.

[0063] Specifically, the calculated raw score is normalized within the range of all neighboring nodes of the current target node using the Softmax function. This process divides the score of each neighboring node by the sum of the scores of all nodes in the neighborhood, thereby obtaining the aggregate weight of each neighboring node's risk contribution to the central node.

[0064] Furthermore, in the spatiotemporal graph attention network that has completed parameter training, a single-point failure scenario is simulated. The initial failure probability feature value of a specific asset node is forcibly modified from 0.01 to 1.0, and the modified feature matrix is ​​input into the network model for forward inference calculation.

[0065] Specifically, the failure signal is transmitted to the neighboring asset nodes at all levels along the dynamic aggregate weight calculated in the aforementioned steps. Each transmission process involves a dot product operation between the node feature matrix and the aggregate weight matrix, and the calculation results are superimposed layer by layer into the feature vectors of the adjacent nodes.

[0066] Furthermore, risk transmission features of adjacent nodes are extracted based on the calculation results of the network output layer. These features are represented as a risk vector containing 128 elements and mapped to specific physical quantitative indicators by a classifier. Specifically, the classifier adopts a multi-layer feedforward neural network structure. Its input is a 128-dimensional risk vector from the output layer of the spatiotemporal graph attention network, and its output consists of three continuous real-valued components, corresponding to the failure probability increase, propagation path attenuation coefficient, and cascade failure threshold, respectively. The training samples of the classifier are obtained by retrospectively analyzing historical accident data: for historical pipe sections that have experienced cascade failures, the changes in failure probability, path length attenuation coefficient, and whether cascade damage was triggered in the neighboring nodes before and after the accident are used as supervision signals, and paired with the 128-dimensional risk vector within the corresponding time window for training. After mastering the above-mentioned training mechanism of "128-dimensional input - 3-dimensional real-valued output - historical accident supervision signal", those skilled in the art can adjust the training sample size and the number of network layers according to specific engineering projects without affecting the definition of risk transmission features.

[0067] Specifically, the quantitative output of risk transmission characteristics includes a failure probability increase of 0.35 for adjacent pipe sections after the failure signal propagates, a propagation path attenuation coefficient of 0.68 representing the energy loss of risk propagation along a 150-meter pipeline path, and a cascading failure threshold of 0.85 to determine whether irreversible cascading damage will occur at adjacent nodes. Setting it to 0.85 is to define a physical boundary between "normal fluctuations" and "irreversible functional collapse." Risk increase data of adjacent nodes before and after damage in historical cascading failure accidents were extracted and trained using a multi-layer feedforward neural network. The risk vector component that triggers a chain reaction in neighboring nodes is labeled at the high-risk point of 0.85. Once the failure probability increase after propagation exceeds this value, the system marks the area as having an extremely high risk of cascading damage, thus providing a quantitative basis for accurately identifying vulnerable nodes. Finally, a systemic risk assessment result for the underground gas pipeline network is generated.

[0068] When constructing a spatiotemporal graph attention network, the neighborhood aggregation weight can be dynamically calculated through a multi-head attention layer. This weight is determined by the physical connection strength between nodes, the direction of media flow, and the real-time risk status. Furthermore, the system extracts the nominal diameter of the pipeline segment asset node (300 mm) and the elastic modulus parameter of L360 steel (206 GPa). It then calculates the cross-sectional stiffness to obtain a physical connection strength coefficient of 0.92. Simultaneously, based on the instantaneous flow direction of 5200 cubic meters per hour in the hydraulic topology network, a vector angle algorithm is used to generate a flow direction weight factor characterizing the unidirectional propagation of risk along the fluid. Specifically, the physical connection strength, flow direction weight factor, and the 0.85 intrinsic health index output by the vulnerability identification module are tensor-concatenated and input into an attention mechanism containing eight attention heads. This is transformed into a 64-dimensional query vector and key vector through a linear mapping layer, and the final aggregate weight is obtained using a dot product operator and Softmax normalization. Using this aggregate weight to weightedly propagate the failure signal, the system can accurately identify the specific asset node with the largest risk increase within 150 meters downstream of the damaged pipeline segment. This provides mechanically grounded quantitative characteristics to support the precise maintenance sequence of the pipeline network, ensuring the engineering reliability of life prediction results in complex topological environments.

[0069] By dynamically capturing the physical connection strength and medium flow bias between nodes through a spatiotemporal graph attention network, the nonlinear diffusion process of risk impact caused by single-point failure in the topology is accurately quantified, realizing the digital characterization of the degree of impact on adjacent nodes and making up for the deficiency of static models in calculating cascading risks.

[0070] A pipe wall degradation state transition equation based on the physical properties of L360 steel and a 10-year service life was established, and a power-law function was used to simulate the natural decay process of a gas pipe wall with an initial wall thickness of 8.0 mm over time.

[0071] Furthermore, when performing the pipe wall degradation state transition calculation, the system extracts the material corrosion coefficient of 0.05 and the power law exponent of 1.25 from the asset node attributes as preset parameters, and inputs the discrete time variable in units of years into the power law operator.

[0072] Specifically, the operator calculates the time variable to the power of 1.25 and multiplies it by 0.05 to output a baseline value representing the annual thinning of the pipe wall under standard service conditions. This value serves as a deterministic component of the state transition equation.

[0073] Furthermore, an external disturbance state equation was established based on the soil's corrosivity index of 4.2 and the 0.15 MPa dynamic load generated by 50 round trips per day by heavy vehicles on the ground. This was used to quantify the nonlinear accelerating effect of environmental factors on the pipe wall degradation rate. Specifically, the ground compressive stress generated by heavy vehicles was captured. In this embodiment, the average value of the original ground pressure was 250 kN / m². A classical vertical additional stress calculation model under the action of concentrated force on an elastic half-space surface (i.e., the Businesko theory model) was introduced from soil mechanics to simulate the spatial diffusion law of pressure within the soil. The point of application of the ground pressure was set as the origin of the coordinate system. In this embodiment, the target calculation point was located vertically below the origin at the pipe's central axis with a burial depth of 1.8 meters. At this point, the horizontal offset distance was 0. According to the Businesko theory, the vertical additional stress coefficient is governed by the reciprocal of the square of the depth.

[0074] Under ideal elastic body conditions with a horizontal offset of 0 and a vertical depth of 1.8 meters, the theoretical stress distribution coefficient at this point is calculated to be approximately 0.477. Considering that underground pipe networks in actual engineering are located in non-ideal elastic soil environments, the system introduces engineering experience correction operators, specifically including: Since the soil in this embodiment is high-density clay, compared to the loose medium in the theoretical model, high-density soil has stronger stress transfer efficiency and a lower spatial diffusion angle. Based on this, the system performs a 3% positive gain compensation on the theoretical stress concentration. Monitoring data shows... The soil in this area is at a critical saturation point. The presence of liquid components significantly reduces the internal frictional resistance between soil particles, leading to an increase in the effective stress component of the dynamic load in the vertical direction. Based on this, a 4% stress concentration correction is introduced. Considering that the soft soil layer has a certain damping absorption effect on high-frequency dynamic loads, which will consume some mechanical vibration energy, the system performs a 2% negative reduction correction based on the vibration attenuation constant obtained by fiber optic sensing, and performs nonlinear superposition calculation (i.e., 1.0 + 3% + 4% - 2% = 1.05). Therefore, a concentration correction of 1.05 times is applied to the above theoretical value.

[0075] The final diffusion attenuation coefficient, after correction based on engineering experience, was determined to be 0.5 (i.e., the product of the theoretical value of 0.477 and the engineering correction coefficient of 1.05). This coefficient represents that after the ground pressure penetrates 1.8 meters of soil, 50% of it is effectively transmitted to the top of the pipe.

[0076] The system further combines the attenuated pipe top stress (i.e., 250 kN / m² multiplied by 0.5 to get 125 kN / m²) with a dynamic amplification factor of 1.2 to compensate for the instantaneous impact load during vehicle operation. The final calculation result is 125 kN / m² multiplied by 1.2 to get 150 kN / m². Through unit conversion, the dynamic environmental load acting on the pipe wall is specifically 0.15 MPa. Specifically, the algorithm inputs the pH value of 4.2 into the soil corrosivity mapping operator. The process of using the soil corrosivity mapping operator to transform the acidic environment into a corrosion enhancement coefficient of 1.2 is as follows: A neutral soil environment (pH 7.0) was set as the calculation baseline, with an initial corrosion enhancement coefficient of 1.0. This baseline represents the standard degradation state without additional chemical acceleration. The real-time soil pH value (pH 4.2 in this embodiment) was extracted from the holographic state monitoring. The acidity offset of the environment where the pipeline node is located was calculated as 2.8 (i.e., 7.0-4.2=2.8) by calculating the absolute difference between the baseline value and the measured value. For the L360 steel used in this embodiment, the sensitivity factor of the material to the acidic environment was extracted based on the pre-stored electrochemical corrosion experimental data. This factor characterizes the acceleration weight of the metal ion precipitation rate for each unit decrease in hydrogen ion concentration. In this embodiment, in order to achieve accurate coupling with the subsequent degradation model, the sensitivity factor is 0.0714.

[0077] The specific steps for determining the sensitivity factor are as follows: First, a simulated acidic soil solution is prepared and its pH is adjusted to the target level to reproduce the electrolyte corrosion environment of a specific area. Simultaneously, the pipeline steel undergoes standardized grinding and surface degreasing treatment. Electrochemical analysis equipment is used to measure the current response of the pipe material in the target acidic environment and the neutral reference environment. By fitting polarization curves, the anodic dissolution rate in the two environments is extracted and compared to obtain a relative growth rate of 7.14%. An extremely slow stretching rate is set, and the pipe is stretched until fractured in both the target acidic solution and a non-corrosive protective medium (such as vacuum or silicone oil). By comparing the difference in the reduction of area at fracture between the two groups of samples, the 7.14% additional plasticity loss of the pipe material under the coupled effect of stress and corrosion is measured. The 7.14% damage increment observed in the above experiments is converted into a dimensionless component and calibrated as a sensitivity factor of 0.0714.

[0078] Finally, multiplying the acidity offset of 2.8 by the sensitivity factor of 0.0714 yields an environmentally induced corrosion increment of 0.2. This corrosion increment of 0.2 is then summed with the initial baseline value of 1.0 to accurately output a corrosion enhancement coefficient of 1.2. This value of 1.2 serves as a nonlinear acceleration operator, directly applied to the natural decay equation of pipe wall degradation to quantify the accelerating effect of acidic soil on the pipe wall thinning process. This operator transforms the acidic environment into a corrosion enhancement coefficient of 1.2 (dimensionless).

[0079] Simultaneously, a dynamic load of 0.15 MPa is input into the stress acceleration operator, which acts as a "translator," converting external pressure fluctuations into the damage progress within the material (a dimensionless value of 0.15). Specifically, a simplified power-law model based on the Pamgren-Miner linear cumulative damage theory is used: employing a power-law relationship describing the fatigue performance of materials, the equivalent alternating stress borne by the pipe is compared with the fatigue limit of the material itself, and the damage increment generated by a single stress cycle is quantified through exponential power calculation; based on the finite element stress analysis of the pipeline after being impacted by a 0.15 MPa dynamic load from the ground, the actual equivalent alternating stress borne by the pipe is calculated to be 87 MPa. Meanwhile, the fatigue limit of the material is set to 40% of the original yield strength of the pipe (i.e., 140 MPa), and a constant of 4.0 is selected as the power law exponent based on the fatigue evolution characteristics of steel. The measured equivalent alternating stress is divided by the material fatigue limit to obtain a stress-intensity ratio of approximately 0.6214. Then, the ratio is raised to the power of 4.0 to calculate an intermediate value of approximately 0.1492. The significant digits of the calculation result are taken, and the final output value is a fatigue damage contribution value of 0.15, which is used to characterize the cumulative loss of the pipe section structure integrity due to external dynamic loads. The contribution value of stress cycle to fatigue damage of 0.15 (dimensionless value) is calculated using this operator.

[0080] By summing the two coefficients mentioned above, a degradation acceleration coefficient of 1.35 is output. The corrosion enhancement coefficient of 1.2 represents the rate of natural thinning of the pipe wall due to the soil chemical environment under static conditions. The contribution of stress cycle to fatigue damage of 0.15 represents the additional rate increment of microcrack accumulation inside the pipe due to external dynamic load cycles. Summing the two essentially constructs a comprehensive degradation operator to characterize the accelerated aging rate of the pipeline node relative to the ideal healthy state under the dual conditions of "chemical corrosion + physical vibration".

[0081] Furthermore, the external disturbance state equation is coupled as a random disturbance term to the pipe wall degradation state transition equation to form a dynamic probability transfer model describing the evolution of the health status of the pipeline network nodes over time.

[0082] Specifically, the system introduces a Gaussian random number with a mean of 0 and a variance of 0.01 as a random disturbance in the underground environment during the calculation process. By multiplying the baseline pipe wall thinning amount by the degradation acceleration coefficient of 1.35 and superimposing this random disturbance, a 100×100-dimensional state transition probability matrix is ​​constructed. When constructing the external disturbance state equation, this embodiment no longer relies solely on preset theoretical dynamic load parameters, but directly uses the collected fiber optic sensing event data as real-time input variables. For example, when the fiber optic monitoring system identifies a continuous "heavy vehicle crater" or "pneumatic impact" event around a pipe section, the model automatically increases the external stress parameters of that pipe section at the corresponding time step. The all-time, all-scale monitoring capability provided by fiber optic sensing technology enables the external disturbance state equation to be dynamically corrected based on the actual physical vibration amplitude (such as the average amplitude and amplitude set), significantly improving the sensitivity of the remaining life prediction model to sudden external destructive factors.

[0083] Furthermore, each row of the state transition probability matrix represents the probability that the pipe wall thickness is at a specific value at the current moment, and each column represents the probability that the pipe wall thickness will transition to the target value in the next year. This matrix serves as the core evolution logic for subsequent lifetime prediction tasks.

[0084] Furthermore, the prediction module uses the aforementioned dynamic probability transfer model to generate a set of 1000 random particles. Each particle stores a specific floating-point value to represent a potential distribution state of the current residual wall thickness of the pipeline node. For example, in the initial state, the particles are all distributed around 7.5 mm.

[0085] Specifically, refer to Figure 3 The flowchart of the output remaining lifetime prediction value shows that the system performs single-step forward time prediction on 1000 particles based on the probability transfer model, obtains the predicted thickness value of each particle in the next year, and simultaneously retrieves the failure probability increase risk transmission feature with a value of 0.35 from the spatiotemporal graph attention network.

[0086] Furthermore, the importance weight of each particle is corrected using the risk transmission feature. Specifically, the Euclidean distance between each predicted particle and the 0.12 mm thinning of the tube wall in the real-time monitored holographic state is calculated, and the negative exponential function value of this distance is multiplied and corrected with the risk transmission feature 0.35 to obtain the updated normalized weight value of each particle.

[0087] Specifically, the system performs a resampling operation using the corrected weights. It randomly selects operators to retain the top 80% of the particles by weight and splits and replicates the high-weight particles to make up the total number of 1000 particles. This process removes low-confidence particles with weights below 0.01 and outputs the posterior particle distribution that reflects the true state at the current moment.

[0088] The resampling update of particle weights specifically includes: calculating the Kolb-Leibler divergence of the particle set before and after resampling; triggering a regularization perturbation mechanism when the divergence value is lower than a preset particle depletion threshold; smoothing the surviving high-weight particles using a Gaussian kernel function to generate a new particle set with tiny random displacements in the state space until the particle diversity index after resampling recovers to a preset confidence interval, preventing premature convergence of the lifetime prediction trajectory due to particle degradation; Specifically, when using the particle filter algorithm for long-term (e.g., the next five years) lifetime prediction, the standard resampling algorithm can lead to a high concentration of particles on a few high-weight samples (particle depletion), resulting in an overly narrow confidence interval for the predicted remaining lifetime and giving incorrect "false high confidence." The system calculates the number of effective particles after each resampling step. If the number of effective particles is less than 20% of the total number of particles, the percentage of effective particles will be calculated in real time when using the particle filter to perform lifetime evolution simulation for the next 20 years. Setting it to 20% is to achieve a balance between "concentration of high-weight samples" and "potential failure path exploration capability." Once the number of effective particles falls below 1 / 5 of the total, particle depletion is determined to have occurred, and a regularization perturbation mechanism based on the Gaussian kernel function is triggered to forcibly widen the particle distribution. This numerical setting not only refers to the general principle in statistics to prevent the model from falling into "false high confidence" but also ensures that the remaining lifetime prediction results can truly reflect future uncertainties. If impoverishment occurs, a regularization process is initiated, using a Gaussian kernel function with a bandwidth of 0.1 to jitter the particle positions. This operation is equivalent to superimposing a smoothing window on the posterior probability density function, forcibly widening the distribution range of particles and preserving more exploration capabilities for low-probability extreme failure events. This effectively solves the problem of overly optimistic or collapsed prediction results caused by "sample impoverishment" in long-term prediction of particle filtering algorithms. The deviation between the inclusion probability of the remaining lifetime prediction results and the actual failure rate is reduced from 18% to 4.5%, more realistically reflecting future uncertainties.

[0089] Furthermore, using the obtained posterior particle distribution as the initial state, the probability transfer model is used to perform multi-step recursive iterative calculations with a step size of 1 year over the next 20 years. In each iteration, the particle distribution at the previous moment is used as the input to the state transition probability matrix, and the predicted distribution at the next moment is output.

[0090] Specifically, by recording the evolution paths of 1,000 particles in 20 iterations, a trajectory cluster describing the future wall degradation trend of the pipeline node is generated in the spatiotemporal dimension. This trajectory cluster is represented as a numerical matrix of 1,000 rows × 20 columns.

[0091] Furthermore, the safety failure threshold of the gas pipeline network is set to 40% of the initial wall thickness, i.e., 3.2 mm. The system traverses each element in the trajectory cluster matrix and identifies the specific year in which each trajectory reaches or falls below 3.2 mm. In this invention, "continuous time integration" refers to performing multi-step discrete recursion on the probability transfer model with a preset step size on the time axis, and calculating the relationship between the state probability distribution and the failure threshold at discrete time points. Thus, by numerically integrating the discrete probability distribution of failure time, the mathematical expectation of the remaining life is obtained. The time step size (such as month, quarter, year) can also be adjusted according to business needs to achieve approximate integration of the continuous time evolution process while ensuring numerical convergence.

[0092] Specifically, the time points when all 1000 trajectories reached the 3.2 mm threshold were statistically analyzed, and a failure time probability distribution interval was constructed with the year as the horizontal axis and the occurrence frequency as the vertical axis. The mathematical expectation value of this distribution interval was calculated using an integral operator.

[0093] Furthermore, by calculating the difference between the expected value and the current year, the final output is a predicted remaining life of 8.5 years for the target pipe section, and this predicted value, along with the failure probability distribution curve, is sent to the display terminal of the management node.

[0094] Specifically, the system automatically triggers the maintenance level determination logic based on the predicted value of 8.5 years. If the predicted value is less than 5 years, an emergency inspection work order associated with the asset node is automatically generated in the management node of the pipeline knowledge graph, completing the closed-loop processing from data collection to life prediction to maintenance decision-making.

[0095] The value output by the prediction module is extracted as the remaining lifespan prediction value of 8.5 years. An index mapping relationship is established between this prediction value and the corresponding pipeline node identifier RQ-GD105 in the holographic state set. By constructing a structured data object containing node ID, predicted lifespan value, predicted timestamp and current failure probability of 0.05, the logical binding between the prediction result and the physical entity is realized. Furthermore, a consistent hash operation is performed on the network node identifier using a preset hash algorithm. The calculated hash value is mapped to a pre-defined non-relational database cluster containing 64 storage shards to determine the target storage partition index of the predicted data object. Specifically, the consistent hash operation obtains the binary encoding sequence of the identifier string and inputs it into the hash function to generate a 32-bit unsigned integer. Then, a modulo operation is performed between this integer and the total number of shards, 64. The lifetime prediction data of node RQ-GD105 is accurately routed to the 12th storage shard based on the operation result, thereby achieving load balancing and efficient addressing of large-scale network prediction data at the physical storage level.

[0096] Furthermore, the mapped structured data object is written in JSON format to the document-oriented storage partition of the non-relational database, and a global index based on a B-tree structure is established for the pipeline node identifier field during the writing process; Specifically, the B-tree structure maintains a three-layer multi-branch tree storage architecture, stores the character features of the node identifier as search keys in the leaf nodes, and points to the physical offset address in the storage partition through pointer logic, thereby ensuring that the system can quickly retrieve the remaining life data of a specific pipe segment within 10 milliseconds when performing pipeline risk retrieval.

[0097] Furthermore, the persistent write operation of the database is performed. By starting the write-ahead log mechanism, the lifetime prediction data to be written is first recorded in the non-volatile transaction log buffer. Specifically, a forced flush operation is triggered every 512KB of data accumulated in the transaction log buffer or every 100 milliseconds, synchronously writing the log data into the persistent sector of the magnetic media storage. Subsequently, the database engine writes the prediction data into the data file of the non-relational database according to the log records, ensuring that in the event of a sudden power outage or system crash, the lifetime prediction value and the node mapping relationship can be fully recovered by replaying the log.

[0098] A dynamic probability transfer model was established to couple the natural decay law of materials with external environmental disturbances. Soil corrosivity and ground dynamic load were introduced as random disturbance terms into the pipe wall degradation equation, realizing a mathematical expression of the evolution law of the health state of the pipeline network under the combined action of physical mechanisms and environmental stress, thus improving the environmental adaptability of the model. The risk transmission characteristics calculated in real time were integrated into the state recursion process using a particle filtering algorithm. The particle weights were dynamically corrected through likelihood matching and resampling mechanisms, which effectively solved the prediction deviation caused by parameter drift in the theoretical degradation model in actual operation, and ensured that the state probability distribution could track the real physical state of the pipeline network in real time.

[0099] The vulnerability identification module extracts real-time physical attribute data of the target asset node from the holographic status set through the data bus. Specifically, it includes real-time measured pipe wall thickness of 7.5 mm, pipe yield strength of 350 MPa, and operating pressure of 0.42 MPa. These data are then used as basic observation variables and input into the endogenous health assessment operator to calculate the endogenous health index that reflects the bearing capacity of the pipe section itself.

[0100] Furthermore, a pre-built multilayer perceptron model is used to perform the quantitative calculation of the endogenous health index. The model architecture specifically includes an input layer with 4 neurons, a hidden layer with 32 neurons using the ReLU activation function, and an output layer with a single neuron.

[0101] Specifically, the input layer receives the normalized residual wall thickness (0.93), pressure load ratio (0.26), yield strength utilization rate (0.32), and normalized service duration (0.4). The hidden layer performs nonlinear weight mapping on the features, and the output layer uses the Sigmoid activation function to generate a value between 0 and 1. For example, the intrinsic health index of the pipe segment asset node is calculated to be 0.85, indicating that the node is in a good operating condition without considering external influences.

[0102] Furthermore, based on the aggregation weights dynamically allocated by the spatiotemporal graph attention network, and combined with the failure probability increase of 0.35 and the propagation path attenuation coefficient of 0.68 in the risk transmission characteristics, the weighted impact value of neighborhood risk on the central node is quantified to generate the exogenous stress index.

[0103] Specifically, by extracting the aggregate weight values ​​of the three neighboring nodes that have topological connections with the central node, which are 0.45, 0.30 and 0.25 respectively, and performing a multiplication and accumulation operation on the real-time failure probability of each neighboring node and the corresponding aggregate weight, the calculated exogenous stress index is 0.72. This value characterizes the intensity of potential risk compression on the central node caused by the high-risk state of neighboring nodes through the topological structure and environmental coupling edges.

[0104] Furthermore, a risk comparison discriminant function is constructed to screen the vulnerability of all pipeline network nodes. This function sets a health threshold of 0.8, which is derived from the statistical analysis of a large number of stable operating pipeline samples, serving as a baseline for distinguishing whether a node is in a safe operating range and possesses sufficient structural integrity. By extracting features from a large number of pipeline samples in a stable operating state, the cluster centers of samples that maintain structural integrity and do not show an accelerated degradation trend are marked at the 0.8 position to distinguish whether a node is in a safe and healthy range. The warning threshold of 0.7 serves as the classification boundary, derived from... Cascaded failure simulation and engineering risk tolerance experience characterize the critical point of potential risk compression intensity of the central node caused by the neighboring risk through the topology. By simulating the transmission of neighboring risks of different intensities, the critical stress intensity that leads to stress instability or accelerated corrosion of the central node is identified and used as a classification warning line. The value of 0.7 is set to ensure that sufficient inspection and maintenance response time can be reserved before the neighboring risk causes irreversible damage. The calculated endogenous health index of 0.85 and exogenous stress index of 0.72 are logically compared with the corresponding thresholds.

[0105] Specifically, the screening steps for the risk comparison discriminant function are as follows: Step 1: Extraction and Normalization of Endogenous Health Features. Utilizing the physical attribute data from the holographic state set, extract real-time monitoring values ​​of the target asset node, including pipe wall thickness (7.5 mm), pipe yield strength (350 MPa), and operating pressure (0.42 MPa). Transform these raw data into normalized feature vectors, specifically including pipe wall thickness residual rate (0.93), pressure load ratio (0.26), yield strength utilization rate (0.32), and normalized service time value (0.40). Step Two: Quantifying the Endogenous Health Index. The aforementioned feature vector is input into the endogenous health assessment operator, which employs a multilayer perceptron model architecture. Through nonlinear weight mapping of the hidden layers and processing by the activation function of the output layer, a value reflecting the pipe segment's own bearing capacity is generated. In this embodiment, the calculated endogenous health index is 0.85. Step 3: Generate the exogenous stress index. Based on the neighborhood aggregation weights dynamically allocated by the spatiotemporal graph attention network (e.g., 0.45, 0.3, 0.25), combined with the failure probability increase (0.35) and propagation path attenuation coefficient (0.68) in the risk transmission characteristics, the real-time failure probability of each neighboring node is multiplied and accumulated with the corresponding aggregation weight to quantify the weighted impact of neighborhood risk on the central node. The final output exogenous stress index is 0.72.

[0106] The corresponding real-time failure probabilities in the embodiment (0.80, 0.68, 0.62) are multiplied and accumulated as follows: 0.8×0.45+0.68×0.3+0.62×0.25=0.719, which is rounded to the nearest whole number to generate the exogenous stress index of 0.72. Step 4: Perform risk comparison and discrimination screening. Construct a risk comparison and discrimination function, introduce a health threshold (0.8) and a warning threshold (0.7) as classification boundaries, and logically compare the calculated endogenous health index (0.85) and exogenous stress index (0.72) with the corresponding thresholds respectively. Step 5: Marking Vulnerable Nodes and Generating Instructions. Nodes that meet the criteria of having an endogenous health index higher than the health threshold and an exogenous stress index higher than the warning threshold are selected. Such nodes (e.g., RQ-GD105) are marked as vulnerable nodes, and inspection and maintenance instructions for these nodes are automatically generated based on the marking results.

[0107] Specifically, the system identified that the node's endogenous health index was higher than the health threshold, indicating that its physical condition was good, but its exogenous stress index was higher than the warning threshold, indicating that it was highly susceptible to the risk of cascading failures in the surrounding area. Therefore, the node was marked as a vulnerable node, and its corresponding asset identifier RQ-GD105 was extracted.

[0108] Furthermore, the vulnerability identification module triggers a response mechanism based on the marking results, retrieves the associated geographic coordinate data 116.39 and 39.91 from the management node, and generates an inspection and maintenance instruction for the vulnerable node by combining the material grade information stored in the construction traceability edge.

[0109] Specifically, the inspection and maintenance instruction was sent to the mobile terminal in the form of a JSON data packet. The instruction included specific operational suggestions such as shortening the inspection cycle to 48 hours and focusing on checking the insulation potential of the anti-corrosion layer, thus completing a complete business loop from status monitoring to identification of potential risk nodes and maintenance decision-making.

[0110] Starting with the corrected posterior particle distribution, multi-step recursive iteration and numerical integration are performed to generate a trajectory cluster describing the pipe wall degradation trend within the future time window. By statistically analyzing the probability distribution range of the trajectory reaching the failure threshold, a leap from single numerical prediction to probability range prediction is achieved, providing a more valuable fault tolerance space.

[0111] This embodiment achieves a closed-loop technology for underground gas pipeline networks by nesting spatiotemporal graph attention networks and dynamic probability transfer models, from heterogeneous data completion and semantic topology correction to quantification of failure risk transmission and probabilistic prediction of remaining life. Under the premise of ensuring that the model calculation logic conforms to the mechanical and electrochemical mechanisms of gas engineering, it provides a data-supported quantitative basis for precise maintenance decisions of pipeline facilities.

[0112] Example 2 This embodiment is applied to the scenario of predicting the lifespan of gas pipeline networks in industrial parks of megacities, which are driven by both high-frequency vibration interference from subway operation and uneven settlement of soft soil layers.

[0113] The acquisition module first obtains the original vibration acceleration time series with a sampling frequency of 1000Hz through a piezoelectric accelerometer deployed on the surface of the tube, and then inputs it into a signal feature extraction model based on a one-dimensional convolutional neural network architecture. Specifically, the signal feature extraction model consists of an input layer capable of accommodating 1000 sampling points, three consecutive convolutional layers, a dropout layer with a dropout rate of 0.2, and a fully connected layer. The first convolutional layer has 32 convolutional kernels of size 5 and stride 1, the second convolutional layer has 64 convolutional kernels of size 3 and stride 1, and the third convolutional layer has 128 convolutional kernels of size 3. Each convolutional layer uses the ReLU activation function for nonlinear feature mapping.

[0114] Furthermore, the signal feature extraction model uses the sliding convolution operation of each layer's convolution kernel on the time axis to extract high-frequency texture features from the original signal, compresses the feature dimension from 1000 dimensions to 128 dimensions through the max pooling layer, and uses the feature gain parameters obtained after 500 iterations of training by the Adam optimizer to identify the subway operating environment noise mode in the 40Hz to 60Hz frequency band in the fully connected layer. Specifically, the algorithm logic defines the extracted environmental noise mode as the interference component. By subtracting the preset standard environmental reference vector from the feature vector output by the fully connected layer, the structural feature residual reflecting the micro-fatigue damage of the pipe wall is obtained. This residual is output in the form of a 128-dimensional dense vector and is transmitted in real time to the dynamic attribute field of the corresponding asset entity in the pipeline network knowledge graph.

[0115] Furthermore, the prediction module synchronously retrieves the average annual settlement rate of the soft soil layer of 3.5 mm recorded in the geographic information system for this area, and combines it with the structural feature residuals obtained in the aforementioned steps, which together serve as the input of the external disturbance term for the dynamic probability transfer model. Specifically, the probability transfer model nests an evolution trend predictor based on a bidirectional long short-term memory network. The predictor contains two hidden layers, each with 256 recurrent units. By receiving the fatigue damage residual sequence and settlement displacement sequence of the past 30 consecutive calendar days, it uses a gating mechanism to calculate the composite contribution weight of settlement load and vibration load to the pipe wall crack propagation.

[0116] Furthermore, the evolution trend predictor outputs a cumulative fatigue damage probability distribution map for the next 12 months, and uses the probability density value in the high-risk range of the distribution map as the resampling bias of the particle filter algorithm to guide the recursive evolution direction of 1000 random particles on the time axis. Specifically, the system statistically analyzes the time points when 1000 particles reach the material's fatigue limit threshold, which is set to 40% of the material's original fatigue strength (approximately 140 MPa for L360 steel). This threshold is calibrated based on the material fatigue damage accumulation theory and long-term service experimental data, representing the physical critical point at which the pipe evolves from micro-cracks to macro-damage under cyclic loading. It is used to determine whether the particle evolution trajectory reaches the boundary of structural failure. The system calculates the mathematical expectation of its failure probability distribution, thereby outputting a predicted remaining life of 6.2 years for the target pipe section under the coupled environment of complex dynamic loads and geological disasters.

[0117] Specifically, the implementation steps for failure time point statistics are as follows: Step 1: Generate initial particle set: The prediction module uses a dynamic probability transfer model to generate a particle set containing 1000 random particles. Each particle stores a specific floating-point value to represent a potential distribution of the current fatigue damage state of the pipeline node. Step 2: Multi-step recursive iterative deduction: Using the current posterior particle distribution as the initial condition, multi-step recursive iterative calculations are performed on the future time axis with a step size of 1 year. Each iteration takes the external environmental load (such as the subway vibration residual and soft soil layer settlement displacement extracted by a one-dimensional convolutional neural network) as a random disturbance term as input, and outputs the predicted distribution for the next time step. Step 3: Construct the evolution trajectory cluster: Record the evolution path of these 1000 particles in the continuous iteration process, and generate a 1000-row × 20-column numerical matrix in the spatiotemporal dimension, which is the trajectory cluster describing the future fatigue damage trend of the pipe segment; Step 4: Identify the contact time: The system traverses each element in the trajectory cluster matrix to identify the specific year in which each particle trajectory first reaches or falls below the aforementioned material fatigue limit threshold. Step 5: Numerical Integration and Lifetime Output: Statistically analyze the time points when all 1000 trajectories reach the threshold, construct a failure time probability distribution interval with year as the horizontal axis and frequency of occurrence as the vertical axis, and calculate the expected value of this distribution using an integral operator. The final output is a predicted remaining lifetime of 6.2 years for the target pipe section.

[0118] This embodiment achieves accurate decoupling of weak damage signals in complex urban environments and simulation of evolution trajectories under multi-source load coupling by deeply nesting one-dimensional convolutional neural networks and bidirectional long short-term memory networks. It solves the technical problem that conventional prediction models are prone to prediction drift under dynamic external disturbances, and greatly improves the confidence of gas pipeline facility life prediction under special and complex working conditions.

[0119] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A life prediction system for underground pipeline facilities based on multi-parameter time series analysis, characterized in that, include: The acquisition module collects and completes high-frequency time-series data from the flow and pressure monitoring system, discrete records from manual inspections, and static vector data from the geographic information system to generate a holographic status set of the pipeline network. The knowledge graph construction module collects historical maintenance logs, geological exploration reports, and construction records. It uses natural language processing to extract entities and their relationships, including pipe sections, valves, soil types, and construction parties, to construct a pipeline network knowledge graph. It also performs logical verification and implicit relationship deduction to establish a pipeline network topology. Network construction module: Maps the holographic state set to the corresponding nodes of the pipeline topology network, constructs a spatiotemporal graph attention network, dynamically allocates the aggregation weight of each node through the attention mechanism, and calculates the risk transmission characteristics of a single point failure to adjacent nodes; Prediction module: Constructs a probability transfer model that includes environmental load, corrosion mechanism and material properties; using the risk transmission characteristics as external observation variables, and based on the probability transfer model, uses a particle filter algorithm to recursively update the state probability distribution of each node on the time axis, performs continuous time integration on the pipeline state, and outputs the remaining life prediction value.

2. The underground pipeline facility life prediction system based on multi-parameter time series analysis according to claim 1, characterized in that, The acquisition module generates a holographic state set of the pipeline network, specifically including: By aligning the time granularity of the high-frequency time-series data, discrete records, and static vector data through a multi-scale sliding window, heterogeneous data is mapped into a multimodal array. The tensor filling algorithm is used to mine the correlation between the spatial topology of the pipeline network and the similarity of the time period, inferring and filling the numerical missing values ​​in the multimodal array, and generating the holographic state set.

3. The underground pipeline facility life prediction system based on multi-parameter time series analysis according to claim 1, characterized in that, The knowledge graph construction module for building the pipeline network specifically includes: Named entity recognition algorithms are used to extract three types of node entities from the historical maintenance logs, geological exploration reports, and construction records: asset nodes including pipe sections and valves, environmental nodes including soil type and groundwater level, and management nodes including construction parties and maintenance records. Directed edges connecting these node entities are constructed based on engineering semantic logic. These directed edges include at least: topological connectivity edges representing the direction of medium flow and physical connections, environmental coupling edges representing the corrosive effect of environmental factors on pipe materials, and construction traceability edges representing construction quality and responsibility attribution. The asset nodes, environmental nodes, and management nodes are instantiated and connected through these topological connectivity edges, environmental coupling edges, and construction traceability edges to generate the pipeline network knowledge graph.

4. The underground pipeline facility life prediction system based on multi-parameter time series analysis according to claim 1, characterized in that, The process of performing logical verification and implicit relationship derivation to establish the pipeline network topology specifically includes a three-layer verification system: Pressure rating system logic verification: Based on the gas pipeline network pressure rating standard, the connection relationship of pipeline nodes is traversed to identify and block abnormal topologies in which high and low pressure pipeline sections are directly interconnected without pressure regulating facilities, ensuring that physical connections comply with pressure reduction and rating system rules; Electrocorrosion connectivity derivation: For spatial discontinuities in geographic information system data, cluster analysis is performed using the spatial continuity characteristics of cathodic protection potential gradient to derive physically separated but electrochemically conductive implicit connections, thus completing the current loop logic used for corrosion rate calculation. Hydraulic topology flow direction correction: Based on the law of conservation of flow at nodes and pipe diameter constraints, the contradiction between the medium flow direction and the pipe diameter gradient is detected, reverse flow or loop deadlock paths are corrected, and a pipe network topology with complete hydraulic and electrical logic is generated.

5. The underground pipeline facility life prediction system based on multi-parameter time series analysis according to claim 1, characterized in that, The network construction module calculates risk transmission characteristics specifically including: Based on the node identifiers of the pipeline network knowledge graph, the multimodal array in the holographic state set is mapped one by one to the asset nodes, environment nodes, and management nodes of the pipeline network topology to form a spatiotemporal embedding vector. A spatiotemporal graph attention network is constructed, in which the time series features of each node are used as input. The spatiotemporal dependencies are captured through a multi-head attention layer, and the neighborhood aggregation weight is dynamically calculated. The weight allocation considers the topological distance between nodes, the environmental coupling strength, and the similarity of historical failures. A single-point failure scenario is simulated, and the failure signal is propagated using the aggregation weight. The risk transmission characteristics of adjacent nodes are quantified. The risk transmission characteristics include the failure probability increase, the propagation path attenuation coefficient, and the cascading failure threshold.

6. The underground pipeline facility life prediction system based on multi-parameter time series analysis according to claim 1, characterized in that, The construction of the probabilistic transfer model, which includes environmental load, corrosion mechanism, and material properties, specifically includes: A state transition equation for pipe wall degradation based on material properties and service life is established, and a power-law function is used to simulate the natural decay process of pipe wall thickness over time. An external disturbance state equation based on soil corrosivity and ground dynamic load is established to quantify the nonlinear accelerating effect of environmental factors on the pipe wall degradation rate. The external disturbance state equation is coupled to the pipe wall degradation state transition equation as a random disturbance term to form a probabilistic transfer model describing the evolution of the health status of pipeline nodes over time.

7. The underground pipeline facility life prediction system based on multi-parameter time series analysis according to claim 1, characterized in that, The prediction module outputs the remaining lifetime prediction value, specifically including: A set of random particles representing the current health status of pipeline nodes is generated. A single-step forward prediction is performed based on the probability transfer model. The particle weights are corrected using the risk transmission characteristics, and resampling is performed to obtain a posterior particle distribution reflecting the true state at the current moment, thus completing the real-time update of the state probability distribution. Using the posterior particle distribution as initial conditions, multi-step recursive iterations are performed on the time axis using the probability transfer model. A trajectory cluster describing the pipe wall degradation trend of each particle within a future time window is generated through numerical integration. A pipeline failure threshold is set, and the time points when each trajectory in the trajectory cluster reaches the failure threshold are statistically analyzed. The probability distribution interval of the time points is calculated, and the expected value of the interval is determined as the remaining lifetime prediction value. An index mapping relationship is established between the remaining lifetime prediction value and the corresponding pipeline node identifier, and the mapped data is written to the storage partition of a non-relational database for persistent storage.

8. The underground pipeline facility life prediction system based on multi-parameter time series analysis according to claim 1, characterized in that, The system also includes a vulnerability identification module, which specifically performs the following steps: using the physical attribute data in the holographic state set to calculate the endogenous health index of the node; based on the risk transmission characteristics and combined with the dynamically allocated aggregation weights, quantifying the weighted impact value of neighborhood risks on the central node, and generating an exogenous stress index; constructing a risk comparison discrimination function to screen out nodes that meet the conditions of the endogenous health index being higher than the health threshold and the exogenous stress index being higher than the warning threshold, marking them as vulnerable nodes, and generating inspection and maintenance instructions for vulnerable nodes.

9. A method for predicting the lifespan of underground pipeline facilities based on multi-parameter time series analysis, characterized in that, include: Collect and supplement high-frequency time-series data from flow and pressure monitoring systems, discrete records from manual inspections, and static vector data from geographic information systems to generate a holographic status set of the pipeline network; Historical maintenance logs, geological exploration reports, and construction records are collected. Natural language processing is used to extract entities and their interrelationships, including pipe sections, valves, soil types, and construction parties, to construct a pipeline network knowledge graph. Logical verification and implicit relationship deduction are performed to establish a pipeline network topology. The holographic state set is mapped to the corresponding node of the pipeline network topology to construct a spatiotemporal graph attention network. The aggregation weight of each node is dynamically allocated through the attention mechanism to calculate the risk transmission characteristics of a single point failure to adjacent nodes. A probability transfer model incorporating environmental load, corrosion mechanism, and material properties is constructed. Using the risk transmission characteristics as external observation variables, and based on the probability transfer model, the state probability distribution of each node is recursively updated on the time axis using a particle filter algorithm. The pipeline state is then continuously integrated over time to output the predicted remaining life value.

Images