Intelligent early warning system and method for leakage risk of large-diameter pipeline

Abstract

The invention discloses a large-diameter pipeline leakage risk intelligent early warning system and method, and relates to the technical field of pipeline risk early warning. Upstream and downstream operation data of a pipeline are collected, a boundary self-adaptive model is constructed, a multi-dimensional feature map of internal state change of the pipeline is constructed, and a dynamic response mode of the multi-dimensional feature map under the upstream and downstream change condition is analyzed; and comparing the current multi-dimensional feature map dynamic response with the predicted multi-dimensional feature map dynamic response, analyzing whether an abnormal dynamic state exists or not based on a comparison result, and generating an intelligent early warning prompt and a recommendation action if the abnormal state exists. The early warning system not only effectively solves the core problems of high false alarm rate, high missing report rate, difficulty in early leakage identification, inaccurate early warning response and the like in the prior art, but also greatly improves the accuracy, the real-time performance and the intelligent level of the large-diameter pipeline leakage risk early warning system through multi-dimensional, multi-modal and multi-factor intelligent fusion analysis; and a powerful technical guarantee is provided for safe operation of a pipe network.

Description

Technical Field

[0001] This invention relates to the field of pipeline risk early warning technology, specifically to an intelligent early warning system and method for leakage risks in large-diameter pipelines. Background Technology

[0002] With the continuous acceleration of urbanization and the rapid development of industrial infrastructure, various pipeline systems (such as urban water supply pipelines, industrial raw material transportation pipelines, and long-distance oil and gas pipelines) are widely laid underground or on the ground. In particular, large-diameter pipelines, due to their large flow rate, wide coverage, and high operating pressure, will not only cause serious waste of resources and economic losses in the event of a leakage accident, but may also cause environmental pollution, ground subsidence, and even safety accidents such as casualties.

[0003] The existing technology has the following shortcomings:

[0004] 1. In terms of monitoring the internal condition of pipelines, analysis is often conducted using a single data dimension. There is a lack of multi-dimensional and multi-modal perception and joint analysis of the complex dynamic processes inside the pipeline. It is difficult to fully reflect the impact of leakage or pipeline defects on the internal fluid state. This limitation makes it difficult to accurately identify early minor leaks and results in a delayed early warning response.

[0005] 2. Traditional leakage early warning systems often lack a comprehensive multi-factor assessment of abnormal events and rely on a single indicator to determine leakage risk. This results in a lack of rich hierarchical classification and spatial positioning of early warning information, making it difficult to provide maintenance personnel with specific handling suggestions and optimized response paths.

[0006] 3. Traditional pipeline leakage detection methods mostly rely on data from a single type of sensor or make anomaly judgments based on empirical thresholds. They are difficult to cope with the complex and ever-changing pipeline operating environment and the dynamic fluctuations in operating conditions. This often results in the early warning system being insufficiently sensitive to changes in normal operating conditions, which can easily lead to a large number of false alarms or missed alarms, affecting the response efficiency of operation and maintenance personnel and the pipeline safety assurance capability.

[0007] Based on this, the present invention proposes an intelligent early warning system and method for leakage risks in large-diameter pipelines. It not only effectively solves the core problems of high false alarm rate, high false alarm rate, difficulty in early leakage identification and inaccurate early warning response in the existing technology, but also greatly improves the accuracy, real-time performance and intelligence level of the early warning system for leakage risks in large-diameter pipelines through intelligent fusion analysis of multiple dimensions, multiple modes and multiple factors, providing a strong technical guarantee for the safe operation of pipeline networks. Summary of the Invention

[0008] The purpose of this invention is to provide an intelligent early warning system and method for leakage risks in large-diameter pipelines, in order to address the shortcomings of the prior art.

[0009] To achieve the above objectives, the present invention provides the following technical solution: an intelligent early warning method for leakage risk in large-diameter pipelines, the early warning method comprising the following steps:

[0010] Real-time acquisition of upstream and downstream pipeline operation data, and construction of boundary adaptive models;

[0011] Construct a multidimensional feature map of changes in the internal state of the pipeline, and analyze the dynamic response mode of the multidimensional feature map under upstream and downstream changes.

[0012] The current dynamic response of the multidimensional feature map is compared with the predicted dynamic response of the multidimensional feature map. Based on the comparison results, it is analyzed whether there are abnormal dynamics. If there are abnormalities, intelligent early warning prompts and recommended actions are generated.

[0013] In a preferred embodiment, the current dynamic response of the multidimensional feature map is compared with the predicted dynamic response of the multidimensional feature map, and the analysis based on the comparison results is performed to determine whether there is abnormal dynamics, including the following steps:

[0014] Real values ​​are collected synchronously during operation. Compared with the predicted value Compare and calculate the node prediction error. ;

[0015] The anomaly degree of the overall graph structure is calculated by weighting and aggregating the difference indicators of all nodes. ;

[0016] when hour, This is the dynamic anomaly threshold, used to determine if a dynamic anomaly response has occurred.

[0017] In a preferred embodiment, the difference indices of all nodes are weighted and aggregated to calculate the degree of anomaly in the overall graph structure. : ,in, Indicates the total number of nodes. Indicates the first The weight of each node, This represents the difference in node features.

[0018] In a preferred embodiment, real values ​​are collected synchronously during operation. Compared with the predicted value Compare and calculate the node prediction error. : In the formula, This represents the predicted value of parameter i at time t+1. This represents the true value of parameter i at time t+1.

[0019] In a preferred embodiment, analyzing the dynamic response pattern of a multidimensional feature map under upstream / downstream changes includes the following steps:

[0020] Each graph structure The input graph neural network is expressed as:

[0021] ,in, This represents the adjacency matrix after adding self-joins. Represents the node degree matrix, Indicates the first Layer node representation, Represents the weight matrix. Indicates the activation function;

[0022] For each time step Perform the process separately, outputting a sequence of node embedding vectors to represent the graph embedding at each time step. As a time series input, it is fed into the LSTM network to model the evolution trend of node states over time.

[0023] In a preferred embodiment, constructing a multidimensional feature map of the internal state changes of the pipeline includes the following steps:

[0024] Collect operational status information inside the pipeline and divide the pipeline into several continuous monitoring unit sections;

[0025] The feature representation of each node is a vector. Formal representation, constructing graph structure , This represents a set of nodes, where each node represents a physical segment. Represents the set of edges;

[0026] In each sampling time window Generate a graph structure Obtain the time series graph structure based on the node matrix and edge matrix. T represents the duration window length.

[0027] In a preferred embodiment, the operating status information inside the pipeline is collected, including instantaneous pressure, pressure change gradient, sound wave propagation delay, echo characteristics, attenuation signal, strain distribution, temperature gradient, and vibration mode.

[0028] The pipeline is divided into several continuous monitoring unit sections, with each unit section serving as a node. The node-bound features include the average pressure value, pressure gradient change rate, echo delay, signal-to-noise ratio decrease, strain peak value, and vibration anomaly frequency within the current time window of the unit.

[0029] The feature representation of each node is a vector. Formal representation: ,in, This represents the average pressure. This represents the rate of change of pressure. Indicates the amplitude or attenuation coefficient of the reflected sound wave. Sound wave propagation delay, Indicates the abnormal strain value of the optical fiber. This indicates the number of high-frequency vibrations in the optical fiber.

[0030] In a preferred embodiment, real-time acquisition of upstream and downstream pipeline operating data and construction of a boundary adaptive model include the following steps:

[0031] Collect upstream pipeline operation data, including inlet pressure, water flow rate, and temperature fluctuations; collect downstream pipeline operation data, including outlet pressure, water flow rate, and temperature fluctuations.

[0032] The collected historical operation data is organized to construct a structured historical operation data sample set. The historical operation data is classified and labeled based on the dimensions of operation period, seasonal changes, and water source fluctuations, and historical operation trajectories under multiple operation scenarios are established.

[0033] Fluctuation characteristics of operational indicators in historical operational data are extracted and analyzed, and the mean, variance, coefficient of variation, and rate of change of each parameter are dynamically calculated using a sliding time window;

[0034] A boundary adaptive model is established based on historical fluctuation characteristics using the K-Means clustering algorithm.

[0035] In a preferred embodiment, a boundary adaptive model is established based on historical fluctuation characteristics using the K-Means clustering algorithm, including the following steps:

[0036] Construct the feature vector for each historical data point: , The mean, For variance, The coefficient of variation is 1. The rate of change;

[0037] The training set is constructed by collecting feature vectors from multiple time points: ,in, The number of feature samples;

[0038] Historical feature samples are divided into Each running state category is used to optimize the objective function. In the formula, Indicates the first Each sample feature vector Indicates the first Cluster categories, Indicates the first The feature center of the class Represents the Euclidean distance between vectors;

[0039] For each category Calculate the upper and lower boundaries: In the formula, Indicates the first The sample at the th Values ​​in each feature dimension Indicates the first The class of Maximum dimension Indicates the first The class of Minimum dimension.

[0040] The intelligent early warning system for leakage risks in large-diameter pipelines includes a model building module, a response prediction module, and an anomaly analysis module.

[0041] Model building module: Real-time collection of upstream and downstream pipeline operation data using sensors, and construction of boundary adaptive models based on historical operation data;

[0042] Response prediction module: Constructs a multi-dimensional feature map of changes in the internal state of the pipeline using technologies such as pressure sensors, acoustic detection, and fiber optic monitoring. Uses graph neural network + LSTM algorithm to analyze upstream and downstream changes and predict the dynamic response mode of the multi-dimensional feature map.

[0043] Anomaly Analysis Module: Compares the current dynamic response of the multidimensional feature map with the predicted dynamic response of the multidimensional feature map. Based on the comparison results, it analyzes whether there are abnormal dynamics. If anomalies are found, it generates intelligent early warning prompts and recommended actions by combining the anomaly level, spatial location and environmental information.

[0044] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0045] This invention collects real-time upstream and downstream pipeline operating data and constructs a boundary adaptive model to generate a multi-dimensional feature map of pipeline internal state changes. It analyzes the dynamic response patterns of this multi-dimensional feature map under upstream and downstream variations, compares the current dynamic response with the predicted dynamic response, and analyzes the presence of abnormal dynamics based on the comparison results. If an anomaly is found, it generates intelligent early warning prompts and recommended actions. This early warning system not only effectively solves the core problems of existing technologies, such as high false alarm rates, high false negative rates, difficulty in identifying early leaks, and inaccurate early warning responses, but also greatly improves the accuracy, real-time performance, and intelligence level of large-diameter pipeline leakage risk early warning systems through intelligent fusion analysis of multiple dimensions, modalities, and factors, providing strong technical support for the safe operation of pipeline networks. Attached Figure Description

[0046] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0047] Figure 1 This is a flowchart of the method of the present invention.

[0048] Figure 2 This is a system architecture diagram of the present invention. Detailed Implementation

[0049] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, 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.

[0050] Example 1: Please refer to Figure 1 As shown in this embodiment, the intelligent early warning method for leakage risk in large-diameter pipelines includes the following steps:

[0051] By utilizing sensors to collect real-time upstream and downstream pipeline operating data and constructing a boundary adaptive model based on historical operating data, the fluctuation characteristics and reasonable range of upstream and downstream operating data are identified. A multi-dimensional feature map of pipeline internal state changes is constructed using technologies such as pressure sensors, acoustic detection, and fiber optic monitoring. A graph neural network (GNN) + LSTM algorithm is used to analyze the dynamic response pattern of the multi-dimensional feature map under upstream and downstream changes. The current dynamic response of the multi-dimensional feature map is compared with the predicted dynamic response of the multi-dimensional feature map. Based on the comparison results, it is analyzed whether there are abnormal dynamics. If an anomaly is found, intelligent early warning prompts and recommended actions are generated by combining the anomaly level, spatial location, and environmental information.

[0052] This application collects real-time upstream and downstream pipeline operating data and constructs a boundary adaptive model to build a multi-dimensional feature map of pipeline internal state changes. It analyzes the dynamic response pattern of the multi-dimensional feature map under upstream and downstream changes, compares the current dynamic response of the multi-dimensional feature map with the predicted dynamic response, and analyzes the presence of abnormal dynamics based on the comparison results. If an anomaly is found, it generates intelligent early warning prompts and recommended actions. This early warning system not only effectively solves the core problems of high false alarm rate, high false alarm rate, difficulty in identifying early leakage, and inaccurate early warning response in existing technologies, but also greatly improves the accuracy, real-time performance, and intelligence level of large-diameter pipeline leakage risk early warning systems through intelligent fusion analysis of multiple dimensions, modalities, and factors, providing strong technical support for the safe operation of pipeline networks.

[0053] The aforementioned intelligent early warning method for leakage risks in large-diameter pipelines achieves accurate identification and intelligent early warning of pipeline leakage risks by integrating sensor data acquisition, boundary adaptive model construction, multi-dimensional feature map dynamic analysis, and advanced graph neural networks and temporal deep learning algorithms. Compared with existing technologies, it solves several key defects, specifically in the following aspects:

[0054] First, traditional pipeline leak detection methods mostly rely on data from a single type of sensor or on anomaly judgment based on empirical thresholds, making it difficult to cope with the complex and volatile pipeline operating environment and dynamic fluctuations in operating conditions. This often results in insufficient sensitivity of early warning systems to changes in normal operating conditions, leading to a large number of false alarms or missed alarms, affecting the response efficiency of maintenance personnel and the pipeline network's safety assurance capabilities. This technical solution, by constructing a boundary adaptive model based on historical operating data, can accurately capture the reasonable fluctuation range of upstream and downstream operating data, achieving dynamic modeling of adaptive fluctuations in operating conditions. This significantly reduces the false alarm rate caused by natural changes in operating conditions and improves the reliability of early warnings.

[0055] Secondly, existing technologies for monitoring the internal condition of pipelines mostly employ single-dimensional data analysis, lacking multi-dimensional and multi-modal perception and joint analysis of the complex dynamic processes within the pipeline. This makes it difficult to comprehensively reflect the impact of leaks or pipeline defects on the internal fluid state. This limitation makes it difficult to accurately identify early-stage minor leaks, resulting in delayed early warning responses. This early warning method utilizes multiple technologies such as pressure sensing, acoustic detection, and fiber optic monitoring to construct a multi-dimensional feature map of the pipeline's internal state. Combined with the deep temporal modeling capabilities of graph neural networks and long short-term memory networks, it effectively mines the spatiotemporal correlation features contained in pipeline operation data, achieving accurate prediction and analysis of the dynamic processes within the pipeline. This significantly improves the sensitivity and identification capability for micro-leaks and abnormal dynamics.

[0056] Furthermore, traditional leakage early warning systems typically lack comprehensive multi-factor assessments of abnormal events, often relying on single indicators to determine leakage risk. This results in early warning information lacking rich hierarchical classification and spatial location, making it difficult to provide maintenance personnel with specific handling suggestions and optimized response paths. The solution presented in this paper, through intelligent fusion analysis combining anomaly level, spatial location, and environmental information, can comprehensively assess anomaly dynamics, generate tiered early warning prompts and targeted recommended action guidance, significantly improving the practical value of early warnings and the ability to support maintenance decision-making.

[0057] In summary, this technical solution not only effectively solves the core problems of high false alarm rate, large false alarm rate, difficulty in early leakage identification, and inaccurate early warning response in existing technologies, but also greatly improves the accuracy, real-time performance, and intelligence level of the large-diameter pipeline leakage risk early warning system through intelligent fusion analysis of multiple dimensions, multiple modes, and multiple factors, providing a strong technical guarantee for the safe operation of pipeline networks.

[0058] Example 2: Real-time acquisition of upstream and downstream pipeline operating data using sensors, and construction of a boundary adaptive model based on historical operating data to identify the fluctuation characteristics and reasonable range of variation of upstream and downstream operating data, including the following steps:

[0059] In the first stage of the early warning method, high-precision sensor systems need to be deployed at key upstream and downstream nodes of large-diameter pipelines to achieve real-time perception of operational status. Specifically, pressure sensors, flow meters, temperature sensors, and other auxiliary equipment, such as valve status recorders and pump station start-stop monitoring devices, should be installed upstream and downstream, respectively. These sensors can continuously collect key data, including inlet and outlet pressure, water flow rate, and temperature fluctuations, around the clock, to construct a multi-dimensional operational status dataset. Since the operational status of the pipeline system varies significantly under different time periods, loads, and environmental conditions, relying solely on status data at a single moment can easily lead to erroneous judgments. Therefore, the acquisition system must ensure time synchronization, spatial correspondence, and data integrity to provide a high-quality data foundation for subsequent model building.

[0060] Next, the system needs to systematically organize the historical data collected over a long period to construct a structured historical operational data sample set. This stage involves more than just simple data aggregation; it also includes preprocessing processes such as data cleaning, outlier removal, and missing value completion. Based on this, the data should be further classified and labeled according to dimensions such as operating time, seasonal changes, and water source fluctuations to establish historical operational trajectories under multiple typical operating scenarios. Through this diverse operational data, the system can understand the various fluctuation patterns and interference factors experienced by the pipeline in actual operation, thereby avoiding over-reliance on a single scenario in subsequent modeling processes, which could lead to insufficient model generalization ability.

[0061] After processing historical data, the early warning system will extract and analyze the fluctuation characteristics of key operational indicators from this data. For parameters such as pressure, flow rate, and temperature, the system will use a sliding time window to dynamically calculate the statistical characteristics of each indicator, such as mean, variance, coefficient of variation, and rate of change. The goal of this stage is to quantify the "reasonable fluctuation behavior" of each parameter under normal operation through mathematical means, thereby providing a controllable boundary range for the next step of model construction. If an indicator maintains a certain fluctuation pattern or characteristic range in most scenarios, then these patterns will be incorporated into the boundary model as "criteria" for stable operating conditions.

[0062] To capture the dynamic changes of operational metrics (such as pressure, flow rate, and temperature) over time, a sliding time window method is first used to extract four core feature metrics: mean, variance, coefficient of variation, and rate of change.

[0063] Mean calculation formula: In the formula, Indicates the first A value of a certain indicator collected at a specific time. Indicates the length of the sliding window. Indicated by time The mean is the average value within the window at the end. The formula for calculating the mean is to calculate the average level of the operating indicators within the time window, reflecting the basic operating status of the current working conditions.

[0064] Variance calculation formula: In the formula, This represents the variance within the current sliding window. The variance calculation formula measures the degree of fluctuation of the indicator value within the time window, reflecting whether there are drastic changes.

[0065] The formula for calculating the coefficient of variation is as follows: In the formula, Represents the coefficient of variation. It represents the standard deviation (square root of variance). The formula for calculating the coefficient of variation, which represents the mean, reflects relative volatility. It eliminates the influence of different dimensions or mean values ​​and helps to identify unsteady operating conditions.

[0066] Formula for calculating the rate of change: In the formula, Indicates the rate of change. This represents the current value compared to the previous value. The sampling time interval is indicated, and the formula for calculating the rate of change reflects the abrupt trend or gradient change of a certain indicator in the time dimension.

[0067] After feature extraction, the system builds a boundary adaptive model based on these historical fluctuation features. This model is constructed using clustering algorithms (such as K-Means and DBSCAN) for training, learning the upper and lower boundary ranges of parameter fluctuations under various operating conditions. The key advantage of this type of model lies in its "dynamic adaptability," meaning it can automatically identify fluctuation patterns under different conditions and generate corresponding upper and lower floating boundaries based on the operating scenario. Unlike traditional methods based on manually set fixed thresholds, the boundary adaptive model can automatically adjust the judgment criteria according to the input data, enabling the model to maintain good robustness and adaptability even under complex operating conditions.

[0068] After extracting the aforementioned fluctuation characteristics, a feature vector is constructed for each historical data point. The system collects feature vectors from multiple time points (or multiple running cycles) to form a complete training set: ,in, The number of feature samples. A core step of the K-Means algorithm is to divide historical feature samples into... Given several operating state categories (e.g., stable, slightly fluctuating, moderately fluctuating, severely fluctuating, etc.), optimize the objective function (minimize the K-Means algorithm): Minimize the Euclidean distance between each sample and its class center. Where: Indicates the first Each sample feature vector Indicates the first Cluster categories, Indicates the first The feature center of the class This represents the Euclidean distance between vectors.

[0069] For each category Calculate the upper and lower boundaries of the characteristic values ​​of each indicator: The boundary values ​​of each feature dimension in each cluster category are extracted to form an adaptive discriminative region. In the formula, Indicates the first The sample at the th Values ​​in each feature dimension Indicates the first The class of Maximum dimension Indicates the first The class of Minimum dimensionality: For real-time data within the current sliding window, calculate its feature vector. Determine whether it falls within the boundary range of a certain cluster:

[0070] Otherwise, it is marked as a potential abnormal fluctuation and proceeds to the next step of the dynamic response prediction and early warning process.

[0071] This method extracts dynamic statistical features through a sliding window, effectively capturing the temporal variation patterns of operational indicators. It then utilizes the K-Means algorithm to cluster and model historical operating conditions, achieving a boundary judgment mechanism with self-learning capabilities. This overcomes the problem of fixed thresholds in traditional methods being unsuitable for complex operating states, providing a precise, dynamic, and stable input basis for subsequent intelligent leak early warning.

[0072] Finally, during actual system operation, the real-time collected upstream and downstream data will be continuously matched and compared with the boundary adaptive model. If a certain indicator value is consistently within the reasonable fluctuation range set by the model, the operating state is considered normal under the current environment, and no alarm needs to be triggered. However, once one or more key parameters continuously deviate from the boundary range, the system will immediately mark the state as "potentially abnormal" and use this deviation as the starting point for triggering subsequent multi-dimensional feature map dynamic modeling and risk analysis. The judgment here considers not only instantaneous deviations but also continuous deviations, trend deviations, and other factors, thereby achieving more sensitive and intelligent dynamic risk identification.

[0073] Through the synergistic effect of the above-mentioned steps, this early warning method effectively overcomes the problem of misjudgment that traditional systems are prone to when faced with fluctuations in operating conditions. The boundary adaptive model can not only dynamically identify the reasonable operating range, but also has the ability to accurately model in multiple scenarios, laying a solid data and algorithm foundation for the entire intelligent leakage risk assessment process, and significantly improving the early warning accuracy and response robustness of the system in actual operation.

[0074] A multi-dimensional feature map of internal state changes in pipelines is constructed using technologies such as pressure sensors, acoustic wave detection, and fiber optic monitoring, including the following steps:

[0075] The system deploys pressure sensors, acoustic detection devices, and distributed fiber optic sensing systems at key nodes of the pipeline to continuously collect operational status information inside the pipeline, including: instantaneous pressure and pressure change gradients acquired by pressure sensors; acoustic propagation delay, echo characteristics, and attenuation signals acquired by acoustic sensors; and strain distribution, temperature gradient, and vibration modes acquired by fiber optic sensors.

[0076] Based on the physical structure of the pipeline, the pipeline is divided into several continuous monitoring unit sections (e.g., each unit is 5 meters long), and each unit serves as a node in the diagram. The features bound to a node include the average pressure value and pressure gradient change rate within the current time window of the unit; the echo delay and signal-to-noise ratio decrease in the unit's acoustic signal; and the strain peak value and vibration anomaly frequency in the corresponding fiber optic sensing channel of the unit. Ultimately, the features of each node can be expressed as a vector form. ,in, This represents the average pressure. This represents the rate of change of pressure. Indicates the amplitude or attenuation coefficient of the reflected sound wave. Sound wave propagation delay, Indicates the abnormal strain value of the optical fiber. This indicates the number of high-frequency vibrations in the optical fiber.

[0077] Based on the physical continuity of the pipeline and the sensor deployment density, a graph structure is constructed. , This represents a set of nodes, where each node represents a physical segment. This represents a set of edges, defining the spatial connections between nodes. Connection weights can be defined based on the physical distance between nodes or the correlation of sound wave or pressure propagation. For example, a higher connection weight is assigned if two nodes are close together or their signal fluctuations are highly correlated. Feature similarity is calculated using indicators such as the Pearson correlation coefficient and used as an attribute of the edge.

[0078] In each sampling time window Each of these will generate a graph structure. Its node matrix ( Indicates the number of nodes. (representing the feature dimension of each node), the edge matrix is... As time progresses, the system acquires a series of time series graph structures. T represents the duration window length. This set of time series diagrams constitutes a "multidimensional feature map sequence," which comprehensively describes the internal physical response and abnormal evolution trajectory of the pipeline under different operating conditions.

[0079] To provide a basis for subsequent model training and comparative analysis, based on manual inspections or historical event data, several sections of the graph structure with leakage or structural anomalies are marked, forming a labeled training set. Anomalies may manifest as: a sudden increase in strain at a node that rapidly spreads to adjacent nodes; a sudden and drastic drop in pressure gradient in a section; or acoustic delay values ​​on certain edges exceeding the normal propagation range.

[0080] Through the above steps, the system transforms the complex and ever-changing internal operating status information of pipelines into a well-structured, feature-rich, and time-continuous multidimensional graphical representation, providing structured and semantic input for graph neural networks and time-series prediction models. Compared to traditional methods that rely solely on single-point data or linear analysis, this multidimensional feature map scheme significantly improves the accuracy of anomaly identification, the ability to track local changes, and the overall structural perception capability, laying the foundation for subsequent intelligent early warning.

[0081] Since different sensors have different data sampling frequencies, physical principles, and data delays, the first step is to perform unified time reference calibration and time window alignment on all sensor data to ensure the temporal consistency and spatial registration accuracy of subsequent analysis.

[0082] Combining Graph Neural Networks (GNNs) with Long Short-Term Memory Networks (LSTMs) to analyze the dynamic response of large-diameter pipelines under changing upstream and downstream operating conditions is a key step in achieving multidimensional feature map prediction and anomaly identification. The GNN + LSTM algorithm is used to analyze and predict the dynamic response patterns of multidimensional feature maps under varying upstream and downstream conditions, including the following steps:

[0083] In the previous stage, a multidimensional feature map sequence that varies over time was constructed, namely... Each diagram ,in: This represents the adjacency matrix between nodes, describing the connection relationships and edge weights of the pipeline structure. Let be a node matrix, representing the time interval of each node. The graph contains multidimensional state information (such as pressure, strain, and sound wave reflection characteristics). This sequence of graphs not only reflects the spatial structure but also preserves continuous temporal information, facilitating dynamic graph modeling and trend prediction.

[0084] To extract the spatial dependency features between different nodes in each time-matrix, firstly, each... Input is a graph neural network, such as a graph convolutional network (GCN) or a graph attention network (GAT). Taking GCN as an example, its core calculation formula is as follows: ,in, This represents the adjacency matrix after adding self-joins. Represents the node degree matrix, Indicates the first Layer node representation (initial) , This represents the learnable weight matrix. This represents the activation function (such as ReLU), and the process is performed at each time step. Perform each step separately, and output a new sequence of node embedding vectors. Each This represents the features after aggregation in the spatial dimension.

[0085] Embed the graph at each time step in the previous step to represent As a time series input, it is fed into the LSTM network to model the evolution trend of node states over time. The core mechanism of LSTM is as follows (taking a specific point as an example): ,in, Indicates the input at the current time (i.e.) (embedding vector of a node in the data). The hidden state represents the output, and the predicted trend represents the trend. These represent the forget gate, input gate, and output gate, respectively. This indicates element-wise multiplication. The output of LSTM is a future... The predicted value of the node dynamic features at any given time, i.e., the multidimensional feature map for predicting the future. .

[0086] After performing error statistics on the entire map, an anomaly scoring matrix is ​​constructed. If the error in a certain region (node ​​or submap) is consistently large or shows a sudden increase in variation, the system determines that the region has experienced an unexpected dynamic response. This may indicate a sudden pressure drop triggered in the early stages of leakage; a change in vibration mode due to pipeline structural fatigue; or abnormal local sound wave reflection caused by environmental disturbances. This anomaly score, combined with the anomaly level (minor, significant, severe), spatial location, and environmental context information, generates intelligent early warning prompts and recommended response actions (such as inspection scheduling, pressure regulation, and suspension of operation).

[0087] The following example illustrates how this intelligent early warning method for leakage risks in large-diameter pipelines based on GNN+LSTM works, and clearly demonstrates its key analysis process and technical advantages.

[0088] In a city's water supply system, there exists a large-diameter main pipe (1.2 meters in diameter) used to transport purified water from the water treatment plant to the city's core area. Typically, residential water consumption is lower at night, resulting in more stable pressure at the upstream pumping station and a relatively stable flow rate downstream. The system has deployed the following monitoring devices:

[0089] Pressure sensor: Used to monitor water pressure at each critical node along the route;

[0090] Acoustic sensor: used to identify acoustic signals of water flow disturbances and possible leaks;

[0091] Fiber optic monitoring system: used to sense minute strain changes; the system has accumulated 3 years of historical data.

[0092] The system treats each monitoring section as a node, and establishes a graph structure between nodes based on their actual pipeline connections. Each node's characteristics include: current pressure value; acoustic wave reflection characteristic parameters (amplitude and frequency changes); and fiber optic strain value. These node attributes, which constitute the graph, are updated every 5 minutes, forming a continuous time series.

[0093] Through the Generative Neural Network (GNN), the model learned how downstream node pressure depends on upstream pumping stations and how pipe bends affect acoustic wave propagation. This spatial relationship exhibits different patterns under different operating conditions, and GNN modeling can identify these structural changes. The system uses a multi-dimensional feature map sequence (12 frames) from the past hour to predict the normal trend of feature changes at each node within the next 5 minutes. For example, the system "knows" that if upstream pressure decreases slowly, downstream pressure will also decrease slowly, and acoustic wave characteristics will not change drastically.

[0094] During a certain period late at night, the system predicted a pressure value of 3.2 MPa for a certain node, but the actual collected value suddenly dropped to 2.6 MPa, accompanied by high-frequency acoustic disturbances and increased fiber optic strain. The system determined that: spatially, this error was limited to that node and two neighboring nodes; temporally, the error occurred suddenly rather than evolving slowly; and there were no corresponding flow regulation actions upstream or downstream. Combining model comparison and empirical thresholds, the system judged that "this is likely an abnormal dynamic caused by leakage."

[0095] The system automatically generates the following warning results:

[0096] Warning level: Medium;

[0097] Anomaly type: Suspected localized leakage;

[0098] Location: 340 meters downstream of XX section of XX Road, XX District;

[0099] Response Recommendation: Dispatch inspection personnel with acoustic positioning devices to confirm the situation; it is recommended to switch to a bypass branch to reduce risk.

[0100] Technical effect description:

[0101] Compared to the traditional method of relying solely on single-point pressure difference to determine whether there is a leak, this method achieves the following: comprehensive analysis of multiple sensor channels to improve the accuracy of anomaly identification; consideration of fluctuations in upstream and downstream conditions to avoid false alarms and missed alarms; "comparative judgment" combined with predictive models, which can issue early warnings 5 ​​to 10 minutes in advance before anomalies cause widespread impact; and automatic location of abnormal areas, saving manual investigation time.

[0102] The current dynamic response of the multidimensional feature map is compared with the predicted dynamic response of the multidimensional feature map. Based on the comparison results, the analysis is performed to determine whether there are any abnormal dynamics, including the following steps:

[0103] During system operation, real-time data is collected synchronously. , and the prediction Compare and calculate the node prediction error. : In the formula, This represents the predicted value of parameter i at time t+1. This represents the true value of parameter i at time t+1. This is the formula for calculating Euclidean distance.

[0104] The anomaly degree of the overall graph structure is calculated by weighting and aggregating the difference indicators of all nodes. : ,in: Indicates the total number of nodes; Indicates the first The weight of each node (e.g., important pipe sections can be assigned a larger weight), This value represents the difference in node features. It serves as a core quantitative indicator for measuring the degree of abnormality in the current overall state.

[0105] The system sets dynamic anomaly thresholds to determine whether anomalies exist. These thresholds are based on historical data or statistical models. .when When this occurs, the system determines it to be a "dynamic abnormal response". Thresholds can be set using the following methods: ,in: This represents the mean of the historical error distribution. The standard deviation of the historical error distribution is represented by... This indicates the sensitivity adjustment coefficient (e.g., between 1.5 and 2).

[0106] The background for including the pipeline is as follows:

[0107] Pipeline type: Large-diameter main water supply pipeline, 1 meter in diameter, 6 kilometers in total length;

[0108] Monitoring equipment deployment: A set of sensor nodes is deployed every 500 meters, for a total of 12 nodes;

[0109] Each node group includes: a pressure sensor, an acoustic interference sensor, and a fiber optic micro-strain sensor.

[0110] An intelligent early warning platform deployed with graph neural networks and LSTM.

[0111] Model prediction and actual comparison process

[0112] During normal operation:

[0113] LSTM prediction results (at a certain time point):

[0114] Node number Pressure prediction (MPa) Acoustic wave characteristic prediction (Hz) Strain prediction (με) 5 3.20 1200 30 6 3.18 1195 32 7 3.15 1192 31

[0115] Actual collected values:

[0116] Node number Actual pressure (MPa) Actual sound wave characteristics (Hz) Actual strain (με) 5 3.18 1198 31 6 3.16 1197 33 7 3.12 1194 32

[0117] The model differed very little from reality and did not trigger any abnormal warnings.

[0118] Abnormal runtime (2:30 AM):

[0119] At this moment, the system detected an abnormal change:

[0120] Forecast values ​​(based on normal trends):

[0121] Node number Pressure prediction (MPa) Sound wave prediction (Hz) Strain prediction (με) 6 3.18 1195 32

[0122] Actual value:

[0123] Node number Actual pressure (MPa) Actual sound wave frequency (Hz) Actual strain (με) 6 2.72 1380 89

[0124] Difference calculation:

[0125] Pressure error = |2.72–3.18| = 0.46 MPa;

[0126] Acoustic error = |1380–1195| = ​​185 Hz;

[0127] Strain error = |89–32| = 57με;

[0128] The system automatically combines these difference vectors to form the difference vector of node 6. And calculate using a function: Meanwhile, analysis of upstream and downstream nodes (5 and 7) showed no significant differences, indicating that the anomaly may be limited to the vicinity of node 6.

[0129] If an anomaly is detected, an intelligent early warning and recommended actions are generated by combining the anomaly level, spatial location, and environmental information, including the following steps:

[0130] Anomaly node identification: Preliminary localization is performed based on nodes whose anomaly vectors exceed a threshold;

[0131] Spatial correlation analysis: By combining the edge connections in the graph structure, determine whether it is an isolated anomaly or resonates with neighboring nodes (such as a leak point causing synchronous mutations in multiple surrounding nodes).

[0132] GIS geographic mapping: Maps abnormal nodes to the actual geographic coordinate system or electronic map, and outputs the precise location (e.g., "50 meters east of the intersection of ×× Road and ×× Street").

[0133] Step 3: Environmental Information Fusion

[0134] Environmental data integration: The system connects with auxiliary information such as meteorological data, geological information, and construction plan databases;

[0135] Impact factor analysis:

[0136] Rainy weather + low-lying areas → high probability of water seepage and accumulation;

[0137] Construction area → External disturbance may cause leakage;

[0138] Pipe material + age information → aging sections are more prone to damage;

[0139] Weighted scoring: The anomaly level is adjusted based on the weighting of the impact factors to improve the accuracy of early warning.

[0140] A structured early warning message is generated based on the comprehensive evaluation results, containing the following content:

[0141] Warning level: High / Medium / Low;

[0142] Spatial location: accurate to the pipe section number or geographic coordinates;

[0143] Anomaly types: such as "sudden leakage", "accumulation of abnormal strain", "abnormal acoustic interference", etc.

[0144] Impact range estimation: such as "may affect water supply pressure within 2 kilometers downstream";

[0145] Confidence score: A score based on a fusion of model output and environment, such as "92.7%";

[0146] Presentation methods: via visual platform pop-ups, mobile app push notifications, SMS notifications to on-duty personnel, etc.

[0147] Based on the warning level and fault type, the system provides tiered response suggestions;

[0148] Warning level:

[0149] Low, recommended action: increase the sampling frequency of this segment of sensor data and continue observation.

[0150] Recommended action: Dispatch inspection personnel to conduct secondary verification using an acoustic wave meter or fiber optic camera.

[0151] High, recommended action: Automatically activate the bypass water supply switching logic, close the valve in this section, and issue a temporary water supply suspension notice for the area.

[0152] Example 3: Please refer to Figure 2 As shown in the figure, the intelligent early warning system for leakage risk of large-diameter pipelines described in this embodiment includes a model building module, a response prediction module, and an anomaly analysis module;

[0153] Model building module: Real-time collection of upstream and downstream pipeline operation data by sensors, and construction of boundary adaptive model based on historical operation data to identify the fluctuation characteristics and reasonable range of change of upstream and downstream operation data. The boundary adaptive model is sent to the response prediction module, and the upstream and downstream pipeline operation data is sent to the response prediction module.

[0154] Response prediction module: Constructs a multi-dimensional feature map of changes in the internal state of the pipeline using technologies such as pressure sensors, acoustic detection, and fiber optic monitoring. Using graph neural network (GNN) + LSTM algorithm, it analyzes the dynamic response mode of the multi-dimensional feature map under upstream and downstream changes and sends the prediction results to the anomaly analysis module.

[0155] Anomaly Analysis Module: Compares the current dynamic response of the multidimensional feature map with the predicted dynamic response of the multidimensional feature map. Based on the comparison results, it analyzes whether there are abnormal dynamics. If anomalies are found, it generates intelligent early warning prompts and recommended actions by combining the anomaly level, spatial location and environmental information.

[0156] It should be understood that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. A and B can be singular or plural. Additionally, the character " / " in this article generally indicates an "or" relationship between the preceding and following related objects, but it can also represent an "and / or" relationship. Please refer to the context for a more accurate understanding.

[0157] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0158] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0159] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An intelligent early warning method for leakage risk in large-diameter pipelines, characterized by: The early warning method includes the following steps: Real-time acquisition of upstream and downstream pipeline operation data, and construction of boundary adaptive models; Construct a multidimensional feature map of changes in the internal state of the pipeline, and analyze the dynamic response mode of the multidimensional feature map under upstream and downstream changes. The current dynamic response of the multidimensional feature map is compared with the predicted dynamic response of the multidimensional feature map. Based on the comparison results, it is analyzed whether there are abnormal dynamics. If there are abnormalities, intelligent early warning prompts and recommended actions are generated.

2. The intelligent early warning method for leakage risk in large-diameter pipelines according to claim 1, characterized in that: The current dynamic response of the multidimensional feature map is compared with the predicted dynamic response of the multidimensional feature map. Based on the comparison results, the analysis is performed to determine whether there are any abnormal dynamics, including the following steps: Real values ​​are collected synchronously during operation. Compared with the predicted value Compare and calculate the node prediction error. ; The anomaly degree of the overall graph structure is calculated by weighting and aggregating the difference indicators of all nodes. ; when hour, This is the dynamic anomaly threshold, used to determine if a dynamic anomaly response has occurred.

3. The intelligent early warning method for leakage risk in large-diameter pipelines according to claim 2, characterized in that: The anomaly degree of the overall graph structure is calculated by weighting and aggregating the difference indicators of all nodes. : ,in, Indicates the total number of nodes. Indicates the first The weight of each node, This represents the difference in node features.

4. The intelligent early warning method for leakage risk in large-diameter pipelines according to claim 3, characterized in that: Real values ​​are collected synchronously during operation. Compared with the predicted value Compare and calculate the node prediction error. : In the formula, This represents the predicted value of parameter i at time t+1. This represents the true value of parameter i at time t+1.

5. The intelligent early warning method for leakage risk in large-diameter pipelines according to claim 4, characterized in that: Analyzing the dynamic response patterns of multidimensional feature maps under upstream and downstream changes includes the following steps: Each graph structure The input graph neural network is expressed as: ,in, This represents the adjacency matrix after adding self-joins. Represents the node degree matrix, Indicates the first Layer node representation, Represents the weight matrix. Indicates the activation function; For each time step Perform the process separately, outputting a sequence of node embedding vectors to represent the graph embedding at each time step. As a time series input, it is fed into the LSTM network to model the evolution trend of node states over time.

6. The intelligent early warning method for leakage risk in large-diameter pipelines according to claim 5, characterized in that: Constructing a multidimensional feature map of the internal state changes of the pipeline includes the following steps: Collect operational status information inside the pipeline and divide the pipeline into several continuous monitoring unit sections; The feature representation of each node is a vector. Formal representation, constructing graph structure , This represents a set of nodes, where each node represents a physical segment. Represents the set of edges; In each sampling time window Generate a graph structure Obtain the time series graph structure based on the node matrix and edge matrix. T represents the duration window length.

7. The intelligent early warning method for leakage risk in large-diameter pipelines according to claim 6, characterized in that: Collect operational status information inside the pipeline, including instantaneous pressure, pressure change gradient, sound wave propagation delay, echo characteristics, attenuation signal, strain distribution, temperature gradient, and vibration mode; The pipeline is divided into several continuous monitoring unit sections, with each unit section serving as a node. The node-bound features include the average pressure value, pressure gradient change rate, echo delay, signal-to-noise ratio decrease, strain peak value, and vibration anomaly frequency within the current time window of the unit. The feature representation of each node is a vector. Formal representation: ,in, This represents the average pressure. This represents the rate of change of pressure. Indicates the amplitude or attenuation coefficient of the reflected sound wave. Sound wave propagation delay, Indicates the abnormal strain value of the optical fiber. This indicates the number of high-frequency vibrations in the optical fiber.

8. The intelligent early warning method for leakage risk in large-diameter pipelines according to claim 7, characterized in that: Real-time acquisition of upstream and downstream pipeline operation data, and construction of a boundary adaptive model, including the following steps: Collect upstream pipeline operation data, including inlet pressure, water flow rate, and temperature fluctuations; collect downstream pipeline operation data, including outlet pressure, water flow rate, and temperature fluctuations. The collected historical operation data is organized to construct a structured historical operation data sample set. The historical operation data is classified and labeled based on the dimensions of operation period, seasonal changes, and water source fluctuations, and historical operation trajectories under multiple operation scenarios are established. Fluctuation characteristics of operational indicators in historical operational data are extracted and analyzed, and the mean, variance, coefficient of variation, and rate of change of each parameter are dynamically calculated using a sliding time window; A boundary adaptive model is established based on historical fluctuation characteristics using the K-Means clustering algorithm.

9. The intelligent early warning method for leakage risk in large-diameter pipelines according to claim 8, characterized in that: The boundary adaptive model is established based on historical fluctuation characteristics using the K-Means clustering algorithm, including the following steps: Construct the feature vector for each historical data point: , The mean, For variance, The coefficient of variation is 1. The rate of change; The training set is constructed by collecting feature vectors from multiple time points: ,in, The number of feature samples; Historical feature samples are divided into Each running state category is used to optimize the objective function. In the formula, Indicates the first Each sample feature vector Indicates the first Cluster categories, Indicates the first The feature center of the class Represents the Euclidean distance between vectors; For each category Calculate the upper and lower boundaries: In the formula, Indicates the first The sample at the th Values ​​in each feature dimension Indicates the first The class of Maximum dimension Indicates the first The class of Minimum dimension.

10. An intelligent early warning system for leakage risks in large-diameter pipelines, used to implement the early warning method described in any one of claims 1-9, characterized in that: It includes a model building module, a response prediction module, and an anomaly analysis module; Model building module: Real-time collection of upstream and downstream pipeline operation data using sensors, and construction of boundary adaptive models based on historical operation data; Response prediction module: Constructs a multi-dimensional feature map of changes in the internal state of the pipeline using technologies such as pressure sensors, acoustic detection, and fiber optic monitoring. Uses graph neural network + LSTM algorithm to analyze upstream and downstream changes and predict the dynamic response mode of the multi-dimensional feature map. Anomaly Analysis Module: Compares the current dynamic response of the multidimensional feature map with the predicted dynamic response of the multidimensional feature map. Based on the comparison results, it analyzes whether there are abnormal dynamics. If anomalies are found, it generates intelligent early warning prompts and recommended actions by combining the anomaly level, spatial location and environmental information.

Images