A municipal pipeline network monitoring system based on the Internet of Things

Abstract

This invention relates to the field of data processing technology, specifically to an Internet of Things (IoT)-based municipal pipeline network monitoring system. The system includes: a multi-source sensing mapping module that receives fluid pressure values ​​and maps them to node feature vectors; a topology association construction module that constructs a dynamic directed graph based on a physical architecture and calculates the transfer function matrix to eliminate redundant connections; a feature fingerprint indexing module that extracts high-dimensional graph features and compresses them into state fingerprint codes to construct a historical fault feature library; and an anomaly retrieval and judgment module that uses the Hamming distance algorithm to retrieve historical fault records and pinpoint potential leak areas. In this invention, by constructing a dynamic directed graph based on a physical architecture, discrete monitoring data is transformed into structured features with spatial attributes. State fingerprint coding enables efficient retrieval and comparison of massive heterogeneous data, accurately solving the problems of delayed fault location and high false alarm rates caused by the lack of spatial topology association in traditional technologies.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a municipal pipeline network monitoring system based on the Internet of Things. Background Technology

[0002] Data processing technology involves systematic engineering for collecting, cleaning, storing, retrieving, and analyzing massive, heterogeneous information with complex spatiotemporal attributes. Its core function lies in extracting features and logically associating raw data through algorithmic models, thereby uncovering the hidden patterns and value behind the data. It is widely used in smart cities, industrial monitoring, and logistics networks, aiming to provide decision-making systems with accurate quantitative data and efficient query services. Traditional municipal pipeline monitoring systems utilize basic sensing devices such as pressure gauges and flow meters deployed at key nodes in the pipeline network to collect data on the fluid state inside the pipeline. The collected analog signals are converted into digital signals and transmitted to a central server via a wireless network. The server uses a relational database to store these time-series arranged values ​​and issues alarms for exceeding limits at individual nodes based on preset static threshold ranges.

[0003] Existing technologies typically only discretize and linearly record single-dimensional physical quantity values ​​collected from various dispersed monitoring points, and store them in a simple time series. The data streams of each monitoring node lack deep logical connections based on the actual spatial topology of the pipeline network. This makes it impossible for the system to accurately establish a dynamic mapping relationship between monitoring data fluctuations and specific fault spatial locations when facing the risk of nonlinear cascading faults hidden in complex pipeline networks. Furthermore, this unstructured data stacking based on a flat database leads to extremely low efficiency in retrieval and verification, making it difficult to quickly locate fault evolution patterns with similar characteristics in massive historical records. This severely restricts the real-time response speed and comprehensive judgment accuracy for sudden pipeline network anomalies. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing an Internet of Things-based municipal pipeline network monitoring system.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a municipal pipeline network monitoring system based on the Internet of Things, comprising: The multi-source sensing mapping module is used to receive fluid pressure and instantaneous flow values ​​in real time from distributed IoT acquisition terminals through a dedicated interface, and to spatially map the fluid pressure and instantaneous flow values ​​according to the geographic spatial coordinates of the pipeline network to generate node feature vectors. The topology association construction module is used to cascade the node feature vectors based on the real physical pipeline architecture, construct a dynamic directed graph, calculate the topology transfer function matrix of the dynamic directed graph, quantify the energy decay gradient, and remove redundant connection edges in the dynamic directed graph based on the energy decay gradient. The feature fingerprint index module is used to perform graph convolution operation on the dynamic directed graph after removing redundant connection edges, extract high-dimensional graph features, compress the high-dimensional graph features into binary state fingerprint codes using the locality sensitive hashing algorithm, and store the state fingerprint codes into the inverted index database to construct a historical fault feature library. The anomaly retrieval and analysis module is used to collect current transient hydraulic characteristics, generate real-time fingerprint codes, calculate the Hamming distance between the real-time fingerprint codes and the historical fault feature database, retrieve historical fault records, and perform reverse inference based on the historical fault records to infer potential leakage areas.

[0006] As a further aspect of the present invention, the specific function of the multi-source sensing mapping module is as follows: The terminal data parsing submodule acquires the fluid pressure value and the instantaneous flow rate value, identifies and removes abnormal abrupt measurement values ​​that exceed the preset sensor range limit, applies an adaptive Kalman filter algorithm to smooth the nonlinear acquisition sequence on the time series, and combines an autoregressive moving average model to eliminate high-frequency water hammer fluctuation interference signals caused by pump station start-up and shutdown, and establishes a clean fluid dataset. The spatial attribute binding submodule obtains the pure fluid dataset and the geospatial coordinates of the pipeline network, parses the globally unique hardware identification serial number in the data packet header, retrieves the matching latitude and longitude positioning anchor points in the static deployment mapping form, calls the Gauss-Kruger projection transformation algorithm to convert the spherical physical coordinates into standard grid points in the plane rectangular coordinate system, and performs multidimensional tensor splicing between the standard grid points and the pure fluid dataset to generate a spatial fluid attribute matrix. The feature dimensionality reduction mapping submodule obtains the spatial fluid attribute matrix, applies a nonlinear kernel principal component analysis algorithm to calculate the variance contribution ratio parameters of fluid physical quantities in different topological dimensions, selects core principal component orthogonal basis vectors that meet the cumulative contribution rate according to a preset information entropy retention threshold criterion, and uses the core principal component orthogonal basis vectors to perform linear projection dimensionality reduction processing on the original high-dimensional attribute tensor to generate the node feature vector.

[0007] As a further aspect of the present invention, the specific functional implementation of the topology association construction module is as follows: The physical graph cascaded submodule obtains the node feature vectors, retrieves the real three-dimensional wiring drawings of the urban underground pipe gallery as the topological connection constraint boundary, uses the Delaunay triangulation algorithm to perform spatial connectivity analysis on nodes distributed at different geographical elevations, establishes a directed weighted edge structure that reflects the upstream and downstream directions of real fluid transportation, and constructs the dynamic directed graph. The transfer function solution submodule extracts the node connectivity and pipe material roughness parameters from the dynamic directed graph, constructs a set of partial differential equations for pressure head loss between adjacent nodes based on Bernoulli's law of conservation of energy, solves the set of equations to obtain the flow transmission damping coefficient matrix between different node pairs, and calculates the topology transfer function matrix. The attenuation gradient optimization submodule obtains the topology transfer function matrix, calculates the hydraulic energy attenuation gradient of all connected paths in the graph from the source water plant node to the end user node, sets a lower limit of the connectivity threshold for determining dead water areas or closed-loop short-circuit redundant pipe sections at the end of the pipeline network, compares the energy attenuation gradient with the lower limit of the connectivity threshold, and removes redundant connecting edges in the dynamic directed graph.

[0008] As a further aspect of the present invention, the specific function of the feature fingerprint index module is as follows: The graph feature extraction submodule obtains the dynamic directed graph after removing the redundant connection edges, constructs a multi-level Chebyshev polynomial graph convolutional neural network architecture, and uses a learnable frequency domain convolutional filter to perform multi-scale spatial feature aggregation calculation on the hydraulic state features within the local neighborhood of the node to extract the high-dimensional graph features. The Local Sensitive Hash Submodule acquires the high-dimensional graph features, performs multiplication-addition dimensionality reduction operations on the continuous floating-point high-dimensional feature vector using a pre-trained set of random projected Gaussian matrices, combines the step sign function to quantize and map the dimensionality-reduced continuous numerical values ​​into discrete binary string sequences, performs bit XOR verification to enhance anti-collision performance, and generates the state fingerprint code. The inverted index construction submodule obtains the status fingerprint code, uses the status fingerprint code as the key primary key search term of the data form, obtains historically verified pipeline fault load data, and uses a hash chain list collision resolution strategy to persistently store the correspondence between the status fingerprint code and the pipeline fault load data to the disk array to construct the historical fault feature library. The pipeline fault load data includes the time, geographical coordinates, and leakage severity parameters of the pipeline rupture or leakage event.

[0009] As a further aspect of the present invention, the specific function of the anomaly retrieval and judgment module is as follows: The real-time encoding generation submodule collects the real-time sensor time series of the micro-hydraulic fluctuation state of the pipeline network within the current time window, and synchronously calls the map feature extraction and local sensitive hash algorithm pipeline to perform end-to-end feature compression and dimensionality reduction encoding processing on the instantaneous data to generate the real-time fingerprint code; Hamming distance comparison submodule extracts the real-time fingerprint code and the benchmark fingerprint library in the historical fault feature library, performs the sum of the number of different bits in the bitwise XOR logic operator at the underlying hardware level, and obtains the Hamming distance metric value that represents the similarity between the current state and the historical fault mode, and matches the historical fault record. The leakage inversion and location submodule acquires the historical fault records, extracts the coordinates of the leakage center source point and its radiation diffusion attenuation range parameters recorded in the historical fault records, performs Bayesian posterior probability update and inference by combining the real-time pressure drop slope curve of the current pipeline node, and uses Gaussian mixture clustering algorithm to delineate the geographical polygon boundary with the highest probability density to infer the potential leakage area.

[0010] As a further aspect of the present invention, the mapping process of the node feature vector specifically includes: Obtain the spatial fluid attribute matrix, calculate the mean square error of the basic flow rate in the adjacent connected regions of the target node in the matrix, apply the singular value decomposition algorithm to orthogonally diagonalize the local covariance matrix, extract the eigenvector direction corresponding to the largest eigenvalue as the principal axis direction of the fluid disturbance, and construct the node reference vector. Obtain the node reference vector, calculate the Mahalanobis spatial distance between adjacent nodes to quantify the spatial heterogeneity distribution difference of hydraulic state, and nonlinearly map the multidimensional spatial distance into similarity weight allocation coefficients in the interval of zero to one based on the Gaussian radial basis kernel function to establish the adjacency mapping weight matrix. The adjacency mapping weight matrix is ​​obtained, and the tensor Kronecker product multiplication rule is used to perform cross-dimensional cross-fusion calculation between the adjacency mapping weight matrix and the spatial fluid attribute matrix. A linear rectified activation function with leakage correction is introduced to perform nonlinear amplitude limiting operation on the fusion calculation result to generate the node feature vector.

[0011] As a further aspect of the present invention, the calculation process of the topological transfer function matrix specifically includes: The roughness coefficient of the inner wall of each physical pipe in the dynamic directed graph, the physical inner diameter of the pipe, and the fluid motion viscosity coefficient at the current temperature are obtained. The friction resistance coefficient under the fully turbulent state is solved by Newton's iterative method based on the Kolblock empirical formula, and the hydraulic friction loss matrix is ​​established. Extract the hydraulic friction loss matrix, statistically analyze the local geometric abrupt change angle parameters and pipe diameter scaling factor at the bifurcation node in the flow spectrum structure, calculate the local vortex loss distribution ratio caused by the rapid separation of fluid streamlines, and linearly weight and superimpose the friction resistance coefficient along the friction and the local vortex loss distribution ratio to calculate the comprehensive resistance coefficient. The comprehensive resistance impedance coefficient is obtained, and a network admittance matrix architecture characterizing the energy input and output balance state inside the system is constructed. Based on Kirchhoff's current conservation law, the Jacobian matrix derivative operation is performed on the flow continuity equation of each intersection node inside the pipeline network to extract the transfer response relationship and calculate the topology transfer function matrix.

[0012] As a further aspect of the present invention, the quantization process of the energy decay gradient specifically includes: Extract the topology transfer function matrix, calculate the cumulative drop in pressure head along the pipeline series path from the source water plant to the target node, measure the average kinetic energy loss ratio per unit distance in the direction of water flow within the pipeline network, and establish a basic energy attenuation profile. The spatial parameters of the physical elevation change rate of the auxiliary pipeline facilities around the target node and the dynamic opening ratio coefficient of the control valve are collected. An exponential attenuation compensation mechanism based on historical operating conditions is introduced to perform multi-dimensional local calibration and dynamic correction adjustment on the basic energy attenuation profile and calculate the attenuation gradient compensation vector. Obtain the attenuation gradient compensation vector, and calculate the energy attenuation gradient using the following mathematical formula: ; in, Representing the The energy decay gradient values ​​of each target node This represents the total physical length of the pipeline from the water source to the target node. This represents the inherent natural damping attenuation coefficient of the pipeline material. The variable representing the spatial distance from the starting point, Represents distance from the starting point The instantaneous pressure head drop gradient function at the point, This represents the total number of auxiliary pipeline facilities surrounding the target node. The traversal index variable representing the auxiliary pipeline facilities. Representing the Factors affecting the valve opening adjustment of ancillary facilities Representing the The projection weight values ​​of the decay gradient compensation vector corresponding to each ancillary facility in the local dimension.

[0013] As a further aspect of the present invention, the extraction process of the high-dimensional graph features specifically includes: The preprocessed dynamic directed graph is obtained, the corresponding strictly normalized Laplacian eigensymmetric transformation matrix is ​​extracted, a large-scale eigenvalue spectral decomposition numerical operation is performed on the Laplacian matrix, the frequency domain orthogonal basis sequence vector group reflecting the global topological structure of the graph is obtained, and the graph convolution smoothing operator is constructed. The graph convolution smoothing operator is obtained, and a polynomial approximate frequency domain filter combination structure with Chebyshev truncation expansion property is constructed based on the preset multi-level neighbor perception receptive field physical range. Random high-frequency hydraulic noise is filtered by dynamically adjusting the convergence weight coefficient of the polynomial, while retaining the low-frequency core signal components, and a graph convolution kernel matrix is ​​generated. Obtain the graph convolution kernel matrix, and calculate the high-dimensional graph features according to the following feature space mapping model: ; in, The high-dimensional graph feature tensor matrix representing the output. This represents the nonlinear hyperbolic tangent activation response function. This represents the maximum truncation order of the Chebyshev polynomial. The index variable represents the order of the polynomial. Representing the Learnable network parameters corresponding to order polynomials, This represents the normalized Laplacian matrix whose eigenvalues ​​are restricted to the interval between -1 and +1 after scaling. Represents the first type of Chebyshev polynomial operation unit. The input is a dynamic directed graph node feature observation data matrix. This represents the weight matrix for cross-layer feature projection transformation in a deep neural network.

[0014] As a further aspect of the present invention, the deduction process of the potential leakage area specifically includes: The historical fault records are obtained, and the seasonal ambient temperature change parameters at the time of the fault and the load benchmark fluctuation index during the peak operation period of the pipeline network are extracted from the historical fault records. The nonlinear error distribution compensation and correction calculation is performed by comparing the current real-time micro pipeline network flow load demand curve with the nonlinear error distribution curve, and a fault spatial source matrix is ​​established. The fault spatial source matrix is ​​obtained, and the Markov Monte Carlo random sampling approximation algorithm is called to generate virtual leak detection particle sources in batches within the limited geographical grid coverage space of the pipeline network. The posterior confidence probability of the virtual leak detection particle sources generating hydraulic state variation phenomena that are consistent with the current terminal observation data is calculated through forward hydraulic simulation, and a fault source probability density heat map is constructed. Extract the geometric set of connected pixel blocks whose probability density distribution values ​​exceed the preset extreme value warning threshold from the probability density heatmap of the fault source. Use the image morphology centroid calculation rule to solve for the coordinates of the geometric center anchor point of the geometric set of connected pixel blocks as a theoretical estimate of the leakage source. Use the theoretical estimate of the leakage source to radiate outward and dynamically adaptively expand the safety buffer isolation distance to deduce the potential leakage area.

[0015] Compared with the prior art, the advantages and positive effects of the present invention are as follows: In this invention, by constructing a dynamic directed graph based on the physical architecture of the pipeline network, the originally discrete and isolated sensor monitoring data is transformed into structured feature nodes with clear spatial attributes and logical relationships. The energy attenuation gradient between nodes is accurately quantified using the transfer function matrix, thereby reconstructing the operating state of the pipeline network at the logical level. By introducing a locality-sensitive hashing algorithm to generate binary state fingerprint codes, efficient retrieval and comparison of massive heterogeneous data is achieved. It can accurately match historical fault evolution paths that are highly consistent with the current abnormal fluctuations within milliseconds, thereby effectively solving the problem of fault location delay caused by the lack of spatial topological association. Attached Figure Description

[0016] Figure 1 This is a diagram showing the overall structure of the system of the present invention; Figure 2 This is a structural diagram of the multi-source sensing mapping module of the present invention; Figure 3 This is a structural diagram of the topology association construction module of the present invention; Figure 4 This is a structural diagram of the feature fingerprint index module of the present invention; Figure 5 This is a structural diagram of the anomaly retrieval and judgment module of the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the software-based technical solution is described in detail below with reference to system architecture diagrams and embodiments. It should be understood that the specific embodiments described herein are only for explaining the technical solutions of this invention and do not constitute a limitation on the scope of protection.

[0018] In the description of this invention, the system architecture relationships or data processing flows indicated by terms such as "layer," "module," "interface," "data flow," "client," and "server" are all defined based on the architecture diagram or flowchart corresponding to the embodiments. This way of describing is only used to clearly illustrate the logical relationships between the elements in the technical solution, and not to limit the physical deployment form. The term "multiple" includes two or more technical units, including but not limited to multiple data nodes, processing threads, service instances, or functional components and other scalable elements. The specific number is determined according to the actual business scenario and needs to be specifically specified.

[0019] Please see Figure 1 and Figure 2 This invention provides a technical solution: a municipal pipeline network monitoring system based on the Internet of Things, comprising: The multi-source sensing mapping module is used to receive fluid pressure and instantaneous flow values ​​in real time from distributed IoT acquisition terminals through a dedicated interface, and to spatially map the fluid pressure and instantaneous flow values ​​according to the geographic coordinates of the pipeline network to generate node feature vectors. The specific functions of the multi-source sensing mapping module are as follows: The terminal data parsing submodule acquires fluid pressure and instantaneous flow values, identifies and removes abnormal and sudden measurement values ​​that exceed the preset sensor range limit, applies an adaptive Kalman filter algorithm to smooth the nonlinear acquisition sequence on the time series, and combines an autoregressive moving average model to eliminate high-frequency water hammer fluctuation interference signals caused by pump station start-up and shutdown, and establishes a clean fluid dataset. The spatial attribute binding submodule obtains the pure fluid dataset and the geospatial coordinates of the pipeline network, parses the globally unique hardware identification serial number in the data packet header, retrieves the matching latitude and longitude positioning anchor points in the static deployment mapping form, calls the Gauss-Kruger projection transformation algorithm to convert the spherical physical coordinates into standard grid points in the plane rectangular coordinate system, and performs multidimensional tensor splicing between the standard grid points and the pure fluid dataset to generate a spatial fluid attribute matrix. The feature dimensionality reduction mapping submodule obtains the spatial fluid attribute matrix, applies the nonlinear kernel principal component analysis algorithm to calculate the variance contribution ratio of fluid physical quantities in different topological dimensions, selects core principal component orthogonal basis vectors that meet the cumulative contribution rate according to the preset information entropy retention threshold criterion, and uses the core principal component orthogonal basis vectors to perform linear projection dimensionality reduction on the original high-dimensional attribute tensor to generate node feature vectors. The mapping process of node feature vectors specifically includes: Obtain the spatial fluid property matrix, calculate the mean square error of the basic flow rate in the adjacent connected regions of the target node in the matrix, apply the singular value decomposition algorithm to orthogonally diagonalize the local covariance matrix, extract the eigenvector direction corresponding to the largest eigenvalue as the principal axis direction of the fluid disturbance, and construct the node reference vector. Obtain the node reference vector, calculate the Mahalanobis spatial distance between adjacent nodes to quantify the spatial heterogeneity distribution difference of hydraulic state, and nonlinearly map the multidimensional spatial distance into similarity weight allocation coefficients in the interval of zero to one based on the Gaussian radial basis kernel function to establish the adjacency mapping weight matrix. The adjacency mapping weight matrix is ​​obtained, and the tensor Kronecker product multiplication rule is used to perform cross-dimensional cross-fusion calculation between the adjacency mapping weight matrix and the spatial fluid attribute matrix. A linear rectified activation function with leakage correction is introduced to perform nonlinear amplitude limiting operation on the fusion calculation result to generate node feature vectors.

[0020] The multi-source sensing mapping module acquires the analog electrical signal sequences transmitted back by 150 high-frequency piezoresistive sensors and 200 electromagnetic flowmeters distributed in geographic space nodes. The analog electrical signal sequences are discretized into 16-bit floating-point arrays containing timestamps through an analog-to-digital converter to form fluid pressure values ​​and instantaneous flow values. The terminal data parsing submodule obtains fluid pressure and instantaneous flow values, reads the sensor factory calibration parameter form, retrieves the absolute range upper limit of 1.6MPa and lower limit of 0MPa for the piezoresistive sensor, and retrieves the instantaneous flow range upper limit of 5000 cubic meters per hour and lower limit of 0 cubic meters per hour for the electromagnetic flowmeter. The terminal data parsing submodule performs a sliding window scan calculation on the input floating-point array, setting the window width to 50 sampling points and the sliding step size to 10 sampling points, and identifies and removes abnormal abrupt measurement values ​​that exceed the above range limit.

[0021] The terminal data parsing submodule applies an adaptive Kalman filter algorithm to smooth and filter the nonlinear acquisition sequence on the time series, constructs a mapping matrix between a 4-dimensional state vector and a 2-dimensional observation vector, and dynamically calculates the ratio parameter between the process noise covariance matrix and the observation noise covariance matrix. When the fluctuation of the sequence variance exceeds the baseline threshold of 15% within 20 consecutive sampling periods, the terminal data parsing submodule automatically increases the weight of the diagonal elements of the process noise covariance matrix by 0.5 times to enhance the filter's tracking response sensitivity to sudden hydraulic jumps. The terminal data parsing submodule combines an autoregressive moving average model to eliminate high-frequency water hammer fluctuation interference signals caused by pump station start-up and shutdown. Based on the Akaike Information Criterion, the autoregressive order is dynamically determined to be 3 and the moving average order is 2. The maximum likelihood estimation method is used to solve the weight parameters of the 5 coefficient terms in the model. The mechanical vibration white noise interference with a frequency higher than 50Hz is filtered out through differential operation to generate a pure fluid dataset with a signal-to-noise ratio improved to more than 45dB. The spatial attribute binding submodule obtains the pure fluid dataset and the geospatial coordinates of the pipeline network, parses the data packet header encapsulated in the pure fluid dataset using the TCP / IP protocol, and extracts the globally unique hardware identification serial number composed of 48-bit binary encoding. The spatial attribute binding submodule uses a binary search tree data structure to retrieve and match the corresponding latitude and longitude positioning anchor points in the static deployment mapping form. It calls the Gauss-Kruger projection transformation algorithm to convert the spherical physical coordinates based on the WGS84 ellipsoid model into standard grid points in the plane rectangular coordinate system, sets the central meridian to 114 degrees, and adds an eastward pseudo-offset constant of 500,000 to eliminate negative coordinate values. The spatial attribute binding submodule performs cross-dimensional splicing operations on standard grid points and fluid pressure, flow rate, temperature and other physical field variables in the pure fluid dataset to construct a multidimensional tensor with 300 rows and 8 columns, and generate a spatial fluid attribute matrix. The feature dimensionality reduction mapping submodule obtains the spatial fluid property matrix, applies the nonlinear kernel principal component analysis algorithm to calculate the variance contribution ratio of fluid physical quantities in different topological dimensions, constructs a Gaussian radial basis kernel matrix with dimensions of 300 by 300, and sets the scale hyperparameter of the kernel function to 0.5. The feature dimensionality reduction mapping submodule performs eigenvalue decomposition on the kernel matrix, arranges the eigenvalue spectrum sequence in descending order of numerical value, and selects the top 12 core principal component orthogonal basis vectors that meet the cumulative contribution rate based on the preset information entropy retention threshold of 85%. The feature dimensionality reduction mapping submodule uses 12 core principal component orthogonal basis vectors to perform inner product projection transformation on the original 8-dimensional high-dimensional attribute tensor, compressing the feature space dimension to 3-dimensional and generating node feature vectors. The feature dimensionality reduction mapping submodule obtains the spatial fluid attribute matrix, extracts instantaneous flow data slices of three adjacent connected regions of the target node within the physical topology connection range in the matrix, and calculates the basic flow mean square error value of 100 consecutive sampling periods in each region. The feature dimensionality reduction mapping submodule extracts the local covariance matrix and applies the singular value decomposition algorithm to perform orthogonal diagonalization processing. The matrix is ​​decomposed into the product form of the left singular vector matrix, the diagonal singular value matrix and the right singular vector matrix. The left singular vector direction corresponding to the maximum singular value is extracted as the main axis direction of the fluid disturbance characterizing the intensity of local flow field turbulence. A node reference vector consisting of 3 single-precision floating-point numbers is constructed. The feature dimensionality reduction mapping submodule obtains the node reference vector, calculates the Mahalanobis spatial distance between adjacent nodes, and calls the covariance matrix of the joint distribution of the head loss rate along the route and the node elevation difference to quantify the spatial heterogeneity distribution difference of the hydraulic state. The feature dimensionality reduction mapping submodule uses the Gaussian radial basis kernel function to nonlinearly map the calculated multidimensional spatial distance into similarity weight allocation coefficients in the interval of 0 to 1, and establishes a dense symmetric adjacency mapping weight matrix with a dimension of 300 by 300. The feature dimensionality reduction mapping submodule obtains the adjacency mapping weight matrix. It uses the tensor Kronecker product multiplication rule to perform cross-dimensional cross-fusion calculation between the adjacency mapping weight matrix and the spatial fluid attribute matrix containing 300 node data. This expands the original feature tensor to generate a high-order cross-feature mapping space with a dimension of 90,000. The feature dimensionality reduction mapping submodule introduces a linear rectified activation function with leakage correction to perform nonlinear amplitude limiting operation on the fused calculation results. The leakage correction coefficient hyperparameter of the negative half axis is set to 0.01, preserving the weak negative wave characteristics of hydraulic disturbance, and finally generating node feature vectors reduced to 128 dimensions. Experimental data show that when the information entropy retention threshold is set to 85% and the kernel function scale hyperparameter is selected as 0.5, the feature mapping latency of the multi-source sensing mapping module when processing 10,000 concurrent data nodes is reduced to less than 12 milliseconds, which is 22.5% higher than that of the linear principal component analysis method in terms of the integrity of feature retention of small flow leakage.

[0022] The aforementioned information entropy retention threshold refers to the lower limit of the percentage of the total variance contained in the selected principal components relative to the total variance of the original data during the feature dimensionality reduction process. This threshold is used to ensure that the core hydraulic fluctuation information of the original fluid physics field is still retained after dimensionality reduction.

[0023] Please see Figure 1 and Figure 3 The topology association construction module is used to cascade and connect node feature vectors based on the real physical pipeline architecture, construct a dynamic directed graph, calculate the topology transfer function matrix of the dynamic directed graph, quantify the energy decay gradient, and remove redundant connection edges in the dynamic directed graph based on the energy decay gradient. The specific functionalities of the topology association construction module are as follows: The physical graph cascaded submodule obtains node feature vectors, retrieves real 3D wiring drawings of urban underground pipe corridors as topological connection constraint boundaries, uses Delaunay triangulation algorithm to perform spatial connectivity analysis on nodes distributed at different geographical elevations, establishes a directed weighted edge structure that reflects the upstream and downstream directions of real fluid transport, and constructs a dynamic directed graph. The transfer function solution submodule extracts the node connectivity and pipe material roughness parameters from the dynamic directed graph. Based on Bernoulli's law of conservation of energy in fluid dynamics, it constructs a set of partial differential equations for pressure head loss between adjacent nodes. Solving the set of equations yields the flow conduction damping coefficient matrix between different node pairs, and the topology transfer function matrix is ​​calculated. The attenuation gradient optimization submodule obtains the topology transfer function matrix, calculates the hydraulic energy attenuation gradient of all connected paths in the graph from the source water plant node to the end user node, sets the lower limit of the connectivity threshold for determining dead water areas or closed-loop short-circuit redundant pipe sections at the end of the pipe network, compares the energy attenuation gradient with the lower limit of the connectivity threshold, and removes redundant connecting edges in the dynamic directed graph. The calculation process of the topological transfer function matrix specifically includes: The roughness coefficient of the inner wall of each physical pipe in the dynamic directed graph, the physical inner diameter of the pipe, and the fluid motion viscosity coefficient at the current temperature are obtained. The friction resistance coefficient under the fully turbulent state is solved by Newton's iterative method based on the Kolblock empirical formula, and the hydraulic friction loss matrix is ​​established. Extract the hydraulic friction loss matrix, statistically analyze the local geometric abrupt change angle parameters and pipe diameter scaling factor at the bifurcation node in the flow spectrum structure, calculate the local vortex loss distribution ratio caused by the rapid separation of fluid streamlines, and linearly weight and superimpose the friction resistance coefficient along the line with the local vortex loss distribution ratio to calculate the comprehensive resistance coefficient. Obtain the comprehensive resistance coefficient, construct the network admittance matrix architecture to characterize the energy input and output balance state inside the system, perform Jacobian matrix differentiation on the flow continuity equation of each intersection node inside the pipeline based on Kirchhoff's current conservation law, extract the transfer response relationship, and calculate the topology transfer function matrix. The quantization process of the energy decay gradient specifically includes: Extract the topology transfer function matrix, calculate the cumulative pressure head drop along the pipeline series path from the source water plant to the target node, measure the average kinetic energy loss ratio per unit distance in the direction of water flow within the pipeline network, and establish a basic energy attenuation profile. The spatial parameters of the physical elevation change rate of the auxiliary pipeline facilities around the target node and the dynamic opening ratio coefficient of the control valve are collected. An exponential attenuation compensation mechanism based on historical operating conditions is introduced to perform multi-dimensional local calibration and dynamic correction adjustment on the basic energy attenuation profile and calculate the attenuation gradient compensation vector. Obtain the decay gradient compensation vector and calculate the energy decay gradient using the following mathematical formula: ; in, Representing the The energy decay gradient values ​​of each target node This represents the total physical length of the pipeline from the water source to the target node. This represents the inherent natural damping attenuation coefficient of the pipeline material. The variable representing the spatial distance from the starting point, Represents distance from the starting point The instantaneous pressure head drop gradient function at the point, This represents the total number of auxiliary pipeline facilities surrounding the target node. The traversal index variable representing the auxiliary pipeline facilities. Representing the Factors affecting the valve opening adjustment of ancillary facilities Representing the The projection weight values ​​of the decay gradient compensation vector corresponding to each ancillary facility in the local dimension.

[0024] The physical map cascaded submodule obtains 128-dimensional node feature vectors, retrieves the vector data package of real 3D wiring drawings of urban underground utility tunnels stored in the geographic information system database through a dedicated fiber optic transmission network, and extracts the shape control points containing elevation attributes and pipe diameter dimensions in the drawings as topological connection constraint boundaries. The physical map cascade submodule reads 1,500 discrete spatial coordinate nodes distributed at different geographical elevations from 0 meters to -15 meters. It uses the Delaunay triangulation algorithm to perform spatial connectivity analysis on the nodes distributed at different geographical elevations, constructs a non-overlapping tetrahedral mesh to divide the spatial set, and removes non-connected inner edges that cross physical obstacles. Based on the natural gravity-guided characteristic of water flow potential energy flowing from high-elevation nodes to low-elevation nodes, the cascaded submodule of the physical graph assigns directional arrows and length weights to each partitioned connected edge, establishing a directed weighted edge structure that reflects the upstream and downstream directions of real fluid transport. The feature vectors of 1500 nodes are attached to the vertex domain of the graph, constructing a dynamic directed graph containing 1500 nodes and 4200 connected edges. The transfer function solving submodule extracts the node connectivity matrix and pipe material roughness parameters of each physical connection edge in the dynamic directed graph. Based on Bernoulli's law of conservation of energy in fluid dynamics, it integrates the node elevation difference, flow velocity head difference and friction loss terms to construct a 1500-dimensional partial differential equation system between adjacent nodes. The transfer function solution submodule calls the Jacobi over-relaxation iterative numerical solution algorithm, sets the maximum number of iterations to 500, and the lower limit of the convergence tolerance threshold to 0.0001. It solves the equation system to obtain the flow conduction damping coefficient matrix, which characterizes the rate of change in flow response caused by small pressure fluctuations at the node, and completes the calculation operation of the topology transfer function matrix. The attenuation gradient optimization submodule obtains the topology transfer function matrix containing 4200 connections, calls the generalized Dijkstra shortest path search algorithm, and calculates the hydraulic energy attenuation gradient of all connected paths in the graph from the main pressurization node of the source water plant to the water distribution nodes of 1500 independent users at the end. The attenuation gradient optimization submodule sets the lower limit of the connectivity threshold for determining dead water zones or closed-loop short-circuit redundant pipe sections at the end of the pipe network to 0.15, extracts the set of energy attenuation gradient values, and compares the energy attenuation gradient of each local micro-pipe section with the lower limit of the connectivity threshold of 0.15 using floating-point values. When the energy attenuation gradient value of a certain pipe segment is less than 0.15, the attenuation gradient optimization submodule determines that the connection edge has a very low effective energy contribution to the global hydraulic transport, and performs a graph pruning operation to remove the corresponding redundant connection edges in the dynamic directed graph. The transfer function solver module obtains the inner wall roughness coefficient, physical inner diameter of each physical pipe in the dynamic directed graph, and the fluid motion viscosity coefficient of 1.004 × 10⁻⁶ square meters per second at the current ambient temperature of 20 degrees Celsius. The transfer function solving submodule uses the Kolblock empirical formula and Newton's iteration method to solve the friction resistance coefficient under the fully turbulent state with a Reynolds number greater than 4000. The upper limit of the error control for the iteration approximation is set to 0.00001, and a hydraulic friction loss matrix with a dimension of 1500 by 1500 is established. The transfer function solving submodule extracts the hydraulic friction loss matrix, and statistically analyzes the local geometric abrupt change angle parameters at 450 water flow bifurcation intersection nodes in the dynamic directed graph structure, as well as the pipe diameter scaling factor at different main and branch pipe interfaces. The transfer function solving submodule calculates the local vortex shedding distribution ratio caused by the rapid separation of fluid streamlines at the sudden expansion or contraction section based on the Buda experience correlation formula. A linear weighting coefficient of 0.7 is used to assign the friction drag coefficient along the line, and a weighting coefficient of 0.3 is used to assign the local vortex shedding distribution ratio. A weighted superposition synthesis operation is performed to calculate the comprehensive drag resistance coefficient. The transfer function solving submodule obtains the comprehensive resistance impedance coefficient, converts the pipe section water resistance into an equivalent hydraulic admittance parameter by taking the reciprocal, and constructs a 1500th order sparse network admittance matrix architecture that characterizes the balance state between water pressure energy input and flow output inside the system. The transfer function solution module performs Jacobian matrix differentiation on the flow continuity balance equation of each intersection node in the pipeline network based on Kirchhoff's current conservation law, extracts the partial derivative transfer response relationship caused by the branch flow fluctuation caused by the node pressure micro-variable, and calculates the analytical solution of the topology transfer function matrix using the Gaussian elimination algorithm. The attenuation gradient optimization submodule extracts the topology transfer function matrix, calculates the cumulative drop in pressure head along the serial pipeline path from the source water plant to the target node along 30 independent water conveyance trunk lines in the map, quantifies the average kinetic energy loss ratio per unit distance in the water flow direction within the pipeline network, and establishes a one-dimensional basic energy attenuation profile containing the distance coordinate system and the drop in head. The attenuation gradient optimization submodule collects the spatial parameters of the physical elevation change rate of 8 auxiliary pipeline facilities within a 50-meter radius around the target node, and reads the percentage parameters of the dynamic opening ratio coefficient of the control valve in the SCADA monitoring system. The attenuation gradient optimization submodule introduces an exponential attenuation compensation function model based on fitting historical operating conditions over the past 5 years. Based on the daily cross temperature difference of ambient temperature and the impact load parameters of water usage during morning and evening peak flow rates, the module performs local curvature calibration and dynamic correction adjustment on the basic energy attenuation profile to generate an attenuation gradient compensation vector with a dimension of 1 x 8. The attenuation gradient optimization submodule obtains the attenuation gradient compensation vector containing the above 8 compensation components, and substitutes the various measured environmental input parameters into the energy attenuation gradient calculation formula to perform numerical integration. In the specific implementation of this formula, the total physical extension length of the pipeline is obtained by directly calling the actual pipeline laying trajectory mapping and accumulating the length in the geographic information system. For a certain trunk pipeline node in this embodiment, the specific value extracted is 1500 meters. The inherent natural damping attenuation coefficient of the pipeline material is extracted based on the average value of the physical aging test report of the material fatigue tensile laboratory of the city water affairs bureau over a 5-year period. In this embodiment, the natural attenuation damping coefficient of the ductile iron pipe material is taken as 0.002. The instantaneous pressure head drop gradient function is generated by fitting a cubic spline curve function using high-frequency sampling points of a nodal piezoresistive sensor array; The total number of auxiliary pipeline facilities around the target node is 3. The opening adjustment influence factors of the 3 valves are extracted as 0.8, 0.6 and 0.4 respectively, and the projection weights of the corresponding attenuation gradient compensation vectors in the local dimension are extracted as 0.5, 0.3 and 0.2 respectively. The decay gradient optimization submodule substitutes the specific numerical set mentioned above into the formula: By performing definite integral and summation operations from 0 to 1500 meters, the comprehensive energy decay gradient of the target node was calculated to be 0.45. The calculated value of 0.45 indicates that the energy consumption of the local resistance of the current pipe section is within the safe range of the normal design flow state. The relationship between its value and the lower limit of the connectivity threshold of 0.15 proves that the path is a high-frequency effective water conveyance link. The attenuation gradient optimization submodule will retain the directed graph connection edge based on this value and refuse to execute the pruning and removal command. Experimental data show that when the lower limit of the connectivity threshold is set to 0.15, the redundant node removal rate of the dynamic directed graph reaches 24.6%. Compared with the original topology without pruning optimization, the matrix inversion operation time in the subsequent graph feature extraction stage is reduced by 31.8%.

[0025] The aforementioned comprehensive energy attenuation gradient refers to the quantitative distribution index of head loss per unit distance caused by pipe wall friction, local resistance of pipe fittings, and elevation changes during fluid transmission in the pipeline network.

[0026] Table 1: Comparison of Optimized Overall Impedance Coefficient ; As shown in Table 1, the specific value of the comprehensive impedance coefficient is obtained by extracting the friction coefficient and vortex shedding ratio of different pipe materials and using a fixed weighting ratio of 0.7 and 0.3.

[0027] Please see Figure 1 and Figure 4 The feature fingerprint index module is used to perform graph convolution operation on the dynamic directed graph after removing redundant connection edges, extract high-dimensional graph features, compress the high-dimensional graph features into binary state fingerprint codes using the locality sensitive hashing algorithm, and store the state fingerprint codes into the inverted index database to build a historical fault feature library. The specific functions of the feature fingerprint index module are as follows: The graph feature extraction submodule obtains the dynamic directed graph after removing redundant connecting edges, constructs a multi-level Chebyshev polynomial graph convolutional neural network architecture, and uses a learnable frequency domain convolutional filter to perform multi-scale spatial feature aggregation calculation on the hydraulic state features within the local neighborhood of the node to extract high-dimensional graph features. The Local Sensitive Hash (LSH) submodule acquires high-dimensional graph features, uses a pre-trained set of random projected Gaussian matrices to perform multiplication-addition dimensionality reduction operations on continuous floating-point high-dimensional feature vectors, combines a step sign function to quantize and map the dimensionality-reduced continuous numerical values ​​into discrete binary string sequences, performs bit XOR verification to enhance anti-collision performance, and generates state fingerprint encoding. The inverted index construction submodule obtains the status fingerprint code, uses the status fingerprint code as the key primary key search term in the data form, obtains historically verified pipeline fault load data, and uses a hash chain list collision resolution strategy to persistently store the correspondence between the status fingerprint code and the pipeline fault load data to the disk array to build a historical fault feature library. Pipeline fault load data includes the time, geographical coordinates, and leakage severity parameters of the pipeline rupture or leakage event. The process of extracting features from high-dimensional graphs specifically includes: The preprocessed dynamic directed graph is obtained, the corresponding strictly normalized Laplacian eigensymmetric transformation matrix is ​​extracted, large-scale eigenvalue spectral decomposition numerical operation is performed on the Laplacian matrix, the frequency domain orthogonal basis sequence vector group reflecting the global topological structure of the graph is obtained, and the graph convolution smoothing operator is constructed. Obtain the graph convolution smoothing operator, construct a polynomial approximate frequency domain filter combination structure with Chebyshev truncation expansion property based on the preset multi-level neighbor perception receptive field physical range, filter random high-frequency hydraulic noise by dynamically adjusting the convergence weight coefficients of the polynomial, and retain the low-frequency core signal components to generate the graph convolution kernel matrix. Obtain the graph convolution kernel matrix, and calculate the high-dimensional graph features based on the following feature space mapping model: ; in, The high-dimensional graph feature tensor matrix representing the output. This represents the nonlinear hyperbolic tangent activation response function. This represents the maximum truncation order of the Chebyshev polynomial. The index variable represents the order of the polynomial. Representing the Learnable network parameters corresponding to order polynomials, This represents the normalized Laplacian matrix whose eigenvalues ​​are restricted to the interval between -1 and +1 after scaling. Represents the first type of Chebyshev polynomial operation unit. The input is a dynamic directed graph node feature observation data matrix. This represents the weight matrix for cross-layer feature projection transformation in a deep neural network.

[0028] The graph feature extraction submodule obtains a dynamic directed graph containing 1131 valid nodes and 2500 connecting edges after removing redundant connecting edges, and constructs a multi-level Chebyshev polynomial graph convolutional neural network architecture with 3 hidden layers. The graph feature extraction submodule initializes the hidden layer nodes of the graph convolutional neural network with dimensions of 256, 128, and 64 respectively, and uses a weight matrix that randomly follows a normal distribution to construct a learnable frequency domain convolutional filter. The graph feature extraction submodule inputs the smoothed node pressure and flow feature tensors into the input layer. Through the frequency domain polynomial expansion of the Laplacian matrix, it performs 3-hop neighbor hydraulic state feature information aggregation within the local neighborhood of the node in the first and second hidden layers. In the third hidden layer, it applies the nonlinear hyperbolic tangent activation function to perform multi-scale spatial feature aggregation calculation, generating a high-dimensional graph feature tensor matrix with dimensions of 1131 x 64. The Local Sensitive Hash submodule obtains high-dimensional graph features with a dimension of 1131 x 64, and calls a Gaussian distribution pseudo-random number generator to generate 10 sets of random projected Gaussian matrices with a dimension of 64 x 72. The Locality Sensitive Hash (LSH) submodule performs dense matrix multiplication and addition dimensionality reduction operations on the continuous floating-point high-dimensional feature vector array and the set of randomly projected Gaussian matrices, mapping the 64-dimensional features to a 72-dimensional low-dimensional Euclidean space. The locality-sensitive hashing submodule, combined with the step sign function, sets a hard decision threshold of 0 for the dimensionality-reduced numerical results. It truncates and quantizes floating-point results greater than or equal to 0 into discrete values ​​of 1, and truncates and quantizes floating-point results less than 0 into discrete values ​​of 0, generating a binary string sequence containing 72 bits. The Local Sensitive Hash Submodule performs an XOR check operation on the 72 bits in groups of 8 bits each, generating a 9-bit parity check bit appended to the end of the string to enhance the anti-collision performance of the bit string when faced with sensor white noise disturbance, and merges them to generate an 81-bit state fingerprint code. The inverted index building submodule obtains the 81-bit status fingerprint code, starts the high-concurrency non-relational database write process of the disk array, and uses the 81-bit status fingerprint code as the key primary key search term of the data form; The inverted index building submodule sends a data request query packet to the historical maintenance record server of the city water group through the Open Database Interconnection Protocol to obtain pipeline fault load data that has been manually verified and confirmed in the past 10 years. The pipeline fault load data specifically includes the time of the pipeline rupture or leakage event recorded in ISO standard time format, centimeter-level geographic coordinates recorded by high-precision differential GPS, and leakage degree parameters divided into 1 to 10 hazard levels based on the percentage drop in water pressure. The inverted index construction submodule uses a separate linked list hash collision resolution strategy, dynamically allocates heap memory space to build a balanced binary tree structure to store redundant data items that collide with hash slots, and persists the one-to-many correspondence between the state fingerprint code and the pipeline fault load data into the file system of the independent disk redundant array, and builds a historical fault feature library that can hold more than 500,000 data records. The graph feature extraction submodule obtains the preprocessed dynamic directed graph topology connection adjacency matrix containing 1131 nodes, and calculates the sum of the elements in each row of the adjacency matrix to generate the angle matrix. The graph feature extraction submodule performs a symmetric multiplication operation by multiplying the negative square root matrix of the degree matrix by the left and right sides of the adjacency matrix, thereby extracting the strictly normalized Laplace feature symmetric transformation matrix. The graph feature extraction submodule calls the Lanzos iterative algorithm to perform eigenvalue spectral decomposition numerical operations on a large-scale normalized Laplacian matrix, sets the maximum iterative convergence residual to 0.000001, and obtains a frequency domain orthogonal basis sequence vector group containing 1131 orthogonal unit vectors that reflects the global topology of the graph, thus completing the construction of the basic matrix structure for the graph convolution smoothing operator. The graph feature extraction submodule obtains the graph convolution smoothing operator. Based on the preset multi-level neighbor perception receptive field physical range, the maximum truncation order of the Chebyshev polynomial is forcibly set to 3, and a polynomial approximate frequency domain filter combination structure with Chebyshev truncation expansion property is constructed. The graph feature extraction submodule dynamically adjusts the learnable network parameters corresponding to polynomials of order 0, 1, 2, and 3 based on the gradient descent direction using the backpropagation algorithm. It suppresses high-frequency feature components with Laplacian eigenvalues ​​greater than 1.5 to filter out random high-frequency hydraulic noise, enhances low-frequency core signal components with eigenvalues ​​less than 0.5, and generates a graph convolution kernel matrix that participates in the final graph convolution multiplication operation. The graph feature extraction submodule obtains the graph convolution kernel matrix and substitutes the input feature matrix containing real-time sampled traffic data from node sensors into the high-dimensional graph feature mapping space formula to perform large-scale parallel tensor operations. In the specific execution flow of this formula, the maximum truncation order of the parameter is pre-determined through optimization on the validation set using a 10-fold cross-validation algorithm, and in this embodiment, it is set to 3. The normalized Laplacian matrix, after scaling, is obtained by forcibly reducing the maximum singular value of the original matrix using a normalization formula, and its internal eigenvalue matrix is ​​strictly limited to the range of -1 to +1. The polynomial order traversal index variable is sequentially increased from 0 to 3. The input dynamic directed graph node feature observation data matrix contains pressure and flow dual-channel information of 1131 nodes. The graph feature extraction submodule performs a point-by-point matrix multiplication of the learnable network parameters with preset weights and the normalized matrix expanded by the polynomial operation unit. The product is then multiplied with the input feature data matrix and the weight matrix of the deep neural network cross-layer feature projection transformation initialized from a Gaussian distribution, and a three-dimensional tensor multiplication is performed. The sum of the matrix products is input into a nonlinear hyperbolic tangent activation response function. The hyperbolic tangent function is used to compress and map the output value to the output interval from -1 to +1, and a high-dimensional graph feature tensor matrix with a dimension of 1131 x 64 is calculated. The numerical results of this operation show that the originally isolated low-dimensional time-series hydraulic sensor data has been integrated with the spatial correlation features within the three-hop physical distance range of the entire pipeline network topology. The high-dimensional graph feature tensor matrix will be directly used as the data input source for the local sensitive hash submodule to generate the dimensionality-reduced fingerprint encoding. Experimental data show that when the Chebyshev polynomial truncation order is set to 3 and combined with the hyperbolic tangent activation function, the extracted high-dimensional graph features maintain a cosine similarity of over 92.1% on the spatial feature vector set in the test environment with a 3% random packet loss rate of the sensor. This improves the noise resistance stability by 14.7% compared with the traditional spatial domain graph attention network architecture.

[0029] The aforementioned high-frequency characteristic components refer to the eigenvector projections of the Laplace matrix corresponding to larger eigenvalues, which typically characterize rapid and localized hydraulic data fluctuations in the pipeline network caused by mechanical vibration or sensor electrical noise.

[0030] Table 2 Fingerprint Hash Collision Rate Test Table: ; Referring to Table 2, after introducing the setting parameter of 81 hash bits, the system's hash collision rate dropped to an extremely low level of 0.004, proving the efficient distinguishing ability of using 81-bit encoding combination in matching massive historical fault features.

[0031] Please see Figure 1 and Figure 5 The anomaly retrieval and analysis module is used to collect current transient hydraulic characteristics, generate real-time fingerprint codes, calculate the Hamming distance between the real-time fingerprint codes and the historical fault feature database, retrieve historical fault records, and perform reverse inference based on historical fault records to infer potential leakage areas. The specific functions of the anomaly retrieval and analysis module are as follows: The real-time encoding generation submodule collects real-time sensor time series of the micro-hydraulic fluctuation state of the pipeline network within the current time window, and synchronously calls the map feature extraction and local sensitive hash algorithm pipeline to perform end-to-end feature compression and dimensionality reduction encoding processing on the instantaneous data to generate real-time fingerprint encoding. The Hamming distance matching submodule extracts the real-time fingerprint code and the benchmark fingerprint library in the historical fault feature library, performs the bitwise XOR logic operation at the underlying hardware level, sums the number of different bits in the operator, and obtains the Hamming distance metric value that represents the similarity between the current state and the historical fault mode, and matches the historical fault records. The leakage inversion and localization submodule acquires historical fault records, extracts the coordinates of the leakage center source point and its radiation diffusion attenuation range parameters recorded in the historical fault records, combines the real-time pressure drop slope curve of the current pipeline node to perform Bayesian posterior probability update and inference, and uses Gaussian mixture clustering algorithm to delineate the geographical polygon boundary with the highest probability density to infer the potential leakage area. The process of extrapolating potential leak areas specifically includes: Historical fault records are obtained, and the seasonal ambient temperature change parameters at the time of the fault and the load benchmark fluctuation index during the peak operation period of the pipeline network are extracted from the historical fault records. The nonlinear error distribution compensation and correction calculation is performed by comparing the current real-time micro pipeline network flow load demand curve with the nonlinear error distribution curve, and a fault spatial source matrix is ​​established. Obtain the fault spatial source matrix, call the Markov Monte Carlo random sampling approximation algorithm to generate virtual leak detection particle sources in batches within the limited geographical grid coverage space of the pipeline network, calculate the posterior confidence probability of the virtual leak detection particle sources generating hydraulic state variation phenomena that are consistent with the current terminal observation data through forward hydraulic simulation, and construct the fault source probability density heat map. Extract the geometric set of connected pixel blocks whose probability density distribution values ​​exceed the preset extreme value warning threshold from the probability density heatmap of fault source tracing. Use the image morphology centroid calculation rule to solve for the coordinates of the geometric center anchor point of the geometric set of connected pixel blocks as a theoretical estimate of the leakage source. Use the theoretically estimated leakage source to radiate outward and dynamically adaptively expand the safety buffer isolation distance to infer the potential leakage area.

[0032] The real-time encoding generation submodule continuously collects the real-time sensor time series of the micro hydraulic fluctuation state of the pipeline network, which contains 300 sampling points, within the current 5-minute time window through the industrial Ethernet interface at a sampling interval of 1 second, and performs zero-mean and unit variance normalization data preprocessing. The real-time encoding generation submodule uses the real-time feature tensor in the memory bus as a dynamic parameter call pointer, synchronously calls the dynamic link library operation operators in the graph feature extraction and locality sensitive hashing algorithm pipeline, and uses the streaming multiprocessor of the graphics processor to perform 1024 threads of concurrent end-to-end feature compression and hash dimensionality reduction encoding processing, compressing and converting floating-point data within 45 milliseconds to generate a real-time fingerprint code consisting of 81 discrete binary bits. Hamming distance matching submodule extracts 81-bit real-time fingerprint code and puts it into the highest level register of the central processing unit. It then loads the historical fault feature library containing 500,000 historical baseline codes into the memory buffer pool at high speed through memory-mapped file technology. The Hamming distance matching submodule calls the advanced vector extension instruction set in the underlying x86 hardware architecture and uses a 256-bit wide single instruction multiple data register to perform wide bit-by-bit XOR logic gate operations on the real-time encoding and the benchmark fingerprint database simultaneously. The Hamming distance comparison submodule uses the built-in instruction to calculate the number of 1s in a binary number, and the sum of the number of 1 bits in the output result of the hardware-level system operator to obtain an integer value of Hamming distance metric that represents the similarity between the current real-time hydraulic microstate and various historical fault modes. The Hamming distance comparison submodule determines historical records with Hamming distance values ​​less than or equal to 15 bits of error boundary as highly suspected fault matching objects, and retrieves the top 5 highly overlapping historical fault records from 500,000 records. The leakage inversion and localization submodule obtains the above 5 historical fault records, parses the form attribute segments, extracts the latitude and longitude coordinates of the leakage center source point based on the WGS84 coordinate system recorded in the historical fault records, and the parameters of the outward radiation diffusion attenuation within a radius of 500 meters drawn based on the historical water pressure drop contour lines. The leak inversion and location submodule extracts discrete data sets of the real-time pressure drop slope of the pipeline node calculated every minute over the past 15 minutes by real-time sensors, and inputs them into a pre-set Bayesian inference network. The leakage inversion and localization submodule uses the historical leakage frequency as the prior probability distribution constant, and combines it with the likelihood function formed by the current real-time descent slope curve to perform Bayesian posterior probability update equation derivation and calculate the current posterior confidence of the existence of a leakage point for each geographic coordinate grid point. The leakage inversion and localization submodule inputs grid coordinates with a posterior confidence greater than 0.6 into the Gaussian mixture clustering algorithm model, sets the upper limit hyperparameter of the number of Gaussian component centers in the algorithm to 3, executes the maximum expectation iteration algorithm to solve the model covariance parameter, delineates the core geographic polygon closed loop boundary with the highest probability density integral area on the map projection plane, and outputs the spatial polygon vertex array set of the potential leakage area. The leak inversion and location submodule obtains five highly similar historical fault records, extracts the seasonal average environmental temperature change parameter in degrees Celsius on the day the fault occurred, and the water flow load benchmark fluctuation variance index of the pipeline network during the peak operating periods at 8 am and 6 pm. The leakage inversion and location submodule collects the current real-time transient curve of the flow load demand output by the total flow meter of the micro-pipeline network, inputs the absolute difference sequence between the two into the nonlinear error distribution compensation and correction formula, performs compensation scaling on the historically extrapolated pressure distribution deviation value, and establishes a 3D fault space source matrix including the node number axis, the spatial elevation axis and the corrected hydraulic residual axis. The leak inversion and localization submodule obtains the fault spatial source matrix, sets the grid division granularity to 5 meters by 5 meters, and calls the Markov Monte Carlo random sampling approximation algorithm to generate 10,000 virtual leak detection particle sources with independent states in a batch within a 10 square kilometer spatial area covered by the limited pipeline network geographic grid. The leakage inversion and localization submodule calls the one-dimensional unsteady flow hydrodynamic equations to perform forward hydraulic simulation for each virtual leakage particle source. It compares the hydraulic state variation phenomenon output by the simulation with the root mean square error of the current actual terminal observation data matrix, and converts the error into the distribution of the probability of occurrence of the inverse posterior confidence. Based on the coordinates and probability values, it constructs a fault source probability density heat map containing 10,000 pixels. The leakage inversion and localization submodule extracts the probability density heat map of the fault source and uses a bilinear interpolation algorithm to smooth the pixel array of the heat map. Threshold cutting is performed on the probability density distribution values ​​of all pixels. The lower limit of the extreme value warning threshold is set to 0.85. All background noise points with a probability lower than this value are removed, and multiple independent connected domain pixel block geometric sets are retained. The leakage inversion and localization submodule uses the morphological centroid calculation rule based on image grayscale weight to solve the geometric center anchor point plane coordinates of the geometric set of connected domain pixel blocks as a theoretical estimate of the leakage source. The theoretically estimated leakage source radiates outward to dynamically adaptive safety buffer isolation distance, which is calculated based on the water flow diffusion speed multiplied by a 5-minute buffer time and set to 50 meters. Using this circumferential expansion boundary, a high-confidence inferred potential leakage area polygon control coordinate point array is finally output. Experimental data show that after applying an extreme value early warning threshold of 0.85 and introducing a high-resolution extrapolation framework with a Markov Monte Carlo particle source scale of 10,000, the absolute average error of the model in locating the physical inversion of the simulated micro-crack leakage source is reduced to within 12.5 meters. Compared with the basic clustering algorithm without seasonal load correction compensation, its false alarm rate is reduced by 41.2% during the peak water use period of the pipeline network.

[0033] The aforementioned virtual leak detection particle source refers to the simulated coordinate point entity generated in the Monte Carlo random sampling algorithm. Each entity carries specific hydraulic dissipation parameters and is used to simulate the initial excitation source state when a real pipeline network ruptures and leaks in the digital twin space.

[0034] Table 3 Leakage Location Accuracy Data: ; Table 3 lists the experimental test data results under different parameter configurations. The data items prove that when the particle source size is increased to 10,000 and the warning threshold is set to a safety cutoff high of 0.85, the false alarm risk rate of the system is significantly suppressed, which verifies the improvement of the spatial geometric accuracy of the inversion addressing logic.

[0035] The above embodiments illustrate preferred embodiments of the present invention. Any equivalent adjustments to the technical solution based on software engineering methods are within the scope of protection, including but not limited to: implementing algorithm logic using different programming languages, refactoring functional modules into services, adjusting data interaction protocols, and optimizing resource scheduling strategies. Any implementation scheme derived from reasonable modifications to the data processing flow, service call chain, or system architecture layer without departing from the core technology of the present invention should be considered within the protection scope defined by the technical solution of the present invention.

Claims

1. A municipal pipeline network monitoring system based on the Internet of Things, characterized in that, The system includes: The multi-source sensing mapping module is used to receive fluid pressure and instantaneous flow values ​​in real time from distributed IoT acquisition terminals through a dedicated interface, and to spatially map the fluid pressure and instantaneous flow values ​​according to the geographic coordinates of the pipeline network to generate node feature vectors. The topology association construction module is used to cascade the node feature vectors based on the real physical pipeline architecture, construct a dynamic directed graph, calculate the topology transfer function matrix of the dynamic directed graph, quantify the energy decay gradient, and remove redundant connection edges in the dynamic directed graph based on the energy decay gradient. The feature fingerprint index module is used to perform graph convolution operation on the dynamic directed graph after removing redundant connection edges, extract high-dimensional graph features, compress the high-dimensional graph features into binary state fingerprint codes using the locality sensitive hashing algorithm, and store the state fingerprint codes into the inverted index database to build a historical fault feature library. The anomaly retrieval and analysis module is used to collect current transient hydraulic characteristics, generate real-time fingerprint codes, calculate the Hamming distance between the real-time fingerprint codes and the historical fault feature database, retrieve historical fault records, and perform reverse inference based on the historical fault records to infer potential leakage areas.

2. The municipal pipeline network monitoring system based on the Internet of Things according to claim 1, characterized in that, The specific functions of the multi-source sensing mapping module are as follows: The terminal data parsing submodule acquires the fluid pressure value and the instantaneous flow rate value, identifies and removes abnormal abrupt measurement values ​​that exceed the preset sensor range limit, applies an adaptive Kalman filter algorithm to smooth the nonlinear acquisition sequence on the time series, and combines an autoregressive moving average model to eliminate high-frequency water hammer fluctuation interference signals caused by pump station start-up and shutdown, and establishes a clean fluid dataset. The spatial attribute binding submodule obtains the pure fluid dataset and the geospatial coordinates of the pipeline network, parses the globally unique hardware identification serial number in the data packet header, retrieves the matching latitude and longitude positioning anchor points in the static deployment mapping form, calls the Gauss-Kruger projection transformation algorithm to convert the spherical physical coordinates into standard grid points in the plane rectangular coordinate system, and performs multidimensional tensor splicing between the standard grid points and the pure fluid dataset to generate a spatial fluid attribute matrix. The feature dimensionality reduction mapping submodule obtains the spatial fluid attribute matrix, applies a nonlinear kernel principal component analysis algorithm to calculate the variance contribution ratio parameters of fluid physical quantities in different topological dimensions, selects core principal component orthogonal basis vectors that meet the cumulative contribution rate according to a preset information entropy retention threshold criterion, and uses the core principal component orthogonal basis vectors to perform linear projection dimensionality reduction processing on the original high-dimensional attribute tensor to generate the node feature vector.

3. The municipal pipeline network monitoring system based on the Internet of Things according to claim 1, characterized in that, The specific functional implementation of the topology association construction module is as follows: The physical graph cascaded submodule obtains the node feature vectors, retrieves the real three-dimensional wiring drawings of the urban underground pipe gallery as the topological connection constraint boundary, uses the Delaunay triangulation algorithm to perform spatial connectivity analysis on nodes distributed at different geographical elevations, establishes a directed weighted edge structure that reflects the upstream and downstream directions of real fluid transportation, and constructs the dynamic directed graph. The transfer function solution submodule extracts the node connectivity and pipe material roughness parameters from the dynamic directed graph, constructs a set of partial differential equations for pressure head loss between adjacent nodes based on Bernoulli's law of conservation of energy, solves the set of equations to obtain the flow transmission damping coefficient matrix between different node pairs, and calculates the topology transfer function matrix. The attenuation gradient optimization submodule obtains the topology transfer function matrix, calculates the hydraulic energy attenuation gradient of all connected paths in the graph from the source water plant node to the end user node, sets a lower limit of the connectivity threshold for determining dead water areas or closed-loop short-circuit redundant pipe sections at the end of the pipeline network, compares the energy attenuation gradient with the lower limit of the connectivity threshold, and removes redundant connecting edges in the dynamic directed graph.

4. The municipal pipeline network monitoring system based on the Internet of Things according to claim 1, characterized in that, The specific function of the feature fingerprint index module is as follows: The graph feature extraction submodule obtains the dynamic directed graph after removing the redundant connection edges, constructs a multi-level Chebyshev polynomial graph convolutional neural network architecture, and uses a learnable frequency domain convolutional filter to perform multi-scale spatial feature aggregation calculation on the hydraulic state features within the local neighborhood of the node to extract the high-dimensional graph features. The Local Sensitive Hash Submodule acquires the high-dimensional graph features, performs multiplication-addition dimensionality reduction operations on the continuous floating-point high-dimensional feature vector using a pre-trained set of random projected Gaussian matrices, combines the step sign function to quantize and map the dimensionality-reduced continuous numerical values ​​into discrete binary string sequences, performs bit XOR verification to enhance anti-collision performance, and generates the state fingerprint code. The inverted index construction submodule obtains the status fingerprint code, uses the status fingerprint code as the key primary key search term of the data form, obtains historically verified pipeline fault load data, and uses a hash chain list collision resolution strategy to persistently store the correspondence between the status fingerprint code and the pipeline fault load data to the disk array to construct the historical fault feature library. The pipeline fault load data includes the time, geographical coordinates, and leakage severity parameters of the pipeline rupture or leakage event.

5. The municipal pipeline network monitoring system based on the Internet of Things according to claim 1, characterized in that, The specific functions of the anomaly retrieval and judgment module are as follows: The real-time encoding generation submodule collects the real-time sensor time series of the micro-hydraulic fluctuation state of the pipeline network within the current time window, and synchronously calls the map feature extraction and local sensitive hash algorithm pipeline to perform end-to-end feature compression and dimensionality reduction encoding processing on the instantaneous data to generate the real-time fingerprint code; Hamming distance comparison submodule extracts the real-time fingerprint code and the benchmark fingerprint library in the historical fault feature library, performs the sum of the number of different bits in the bitwise XOR logic operator at the underlying hardware level, and obtains the Hamming distance metric value that represents the similarity between the current state and the historical fault mode, and matches the historical fault record. The leakage inversion and location submodule acquires the historical fault records, extracts the coordinates of the leakage center source point and its radiation diffusion attenuation range parameters recorded in the historical fault records, performs Bayesian posterior probability update and inference by combining the real-time pressure drop slope curve of the current pipeline node, and uses Gaussian mixture clustering algorithm to delineate the geographical polygon boundary with the highest probability density to infer the potential leakage area.

6. The municipal pipeline network monitoring system based on the Internet of Things according to claim 2, characterized in that, The process of generating the node feature vector specifically includes: Obtain the spatial fluid attribute matrix, calculate the mean square error of the basic flow rate in the adjacent connected regions of the target node in the matrix, apply the singular value decomposition algorithm to orthogonally diagonalize the local covariance matrix, extract the eigenvector direction corresponding to the largest eigenvalue as the principal axis direction of the fluid disturbance, and construct the node reference vector. Obtain the node reference vector, calculate the Mahalanobis spatial distance between adjacent nodes to quantify the spatial heterogeneity distribution difference of hydraulic state, and nonlinearly map the multidimensional spatial distance into similarity weight allocation coefficients in the interval of zero to one based on the Gaussian radial basis kernel function to establish the adjacency mapping weight matrix. The adjacency mapping weight matrix is ​​obtained, and the tensor Kronecker product multiplication rule is used to perform cross-dimensional cross-fusion calculation between the adjacency mapping weight matrix and the spatial fluid attribute matrix. A linear rectified activation function with leakage correction is introduced to perform nonlinear amplitude limiting operation on the fusion calculation result to generate the node feature vector.

7. The municipal pipeline network monitoring system based on the Internet of Things according to claim 3, characterized in that, The process of calculating the topology transfer function matrix specifically includes: The roughness coefficient of the inner wall of each physical pipe in the dynamic directed graph, the physical inner diameter of the pipe, and the fluid motion viscosity coefficient at the current temperature are obtained. The friction resistance coefficient under the fully turbulent state is solved by Newton's iterative method based on the Kolblock empirical formula, and the hydraulic friction loss matrix is ​​established. Extract the hydraulic friction loss matrix, statistically analyze the local geometric abrupt change angle parameters and pipe diameter scaling factor at the bifurcation node in the flow spectrum structure, calculate the local vortex loss distribution ratio caused by the rapid separation of fluid streamlines, and linearly weight and superimpose the friction resistance coefficient along the friction and the local vortex loss distribution ratio to calculate the comprehensive resistance coefficient. The comprehensive resistance impedance coefficient is obtained, and a network admittance matrix architecture characterizing the energy input and output balance state inside the system is constructed. Based on Kirchhoff's current conservation law, the Jacobian matrix derivative operation is performed on the flow continuity equation of each intersection node inside the pipeline network to extract the transfer response relationship and calculate the topology transfer function matrix.

8. The municipal pipeline network monitoring system based on the Internet of Things according to claim 3, characterized in that, The process of quantifying the energy decay gradient specifically includes: Extract the topology transfer function matrix, calculate the cumulative drop in pressure head along the pipeline series path from the source water plant to the target node, measure the average kinetic energy loss ratio per unit distance in the direction of water flow within the pipeline network, and establish a basic energy attenuation profile. The spatial parameters of the physical elevation change rate of the auxiliary pipeline facilities around the target node and the dynamic opening ratio coefficient of the control valve are collected. An exponential attenuation compensation mechanism based on historical operating conditions is introduced to perform multi-dimensional local calibration and dynamic correction adjustment on the basic energy attenuation profile and calculate the attenuation gradient compensation vector. Obtain the attenuation gradient compensation vector, and calculate the energy attenuation gradient using the following mathematical formula: ； in, Representing the The energy decay gradient values ​​of each target node This represents the total physical length of the pipeline from the water source to the target node. This represents the inherent natural damping attenuation coefficient of the pipeline material. The variable representing the spatial distance from the starting point, Represents distance from the starting point The instantaneous pressure head drop gradient function at the point, This represents the total number of auxiliary pipeline facilities surrounding the target node. The traversal index variable representing the auxiliary pipeline facilities. Representing the Factors affecting valve opening adjustment of ancillary facilities Representing the The projection weight values ​​of the decay gradient compensation vector corresponding to each ancillary facility in the local dimension.

9. The municipal pipeline network monitoring system based on the Internet of Things according to claim 4, characterized in that, The process of extracting the high-dimensional graph features specifically includes: The preprocessed dynamic directed graph is obtained, the corresponding strictly normalized Laplacian eigensymmetric transformation matrix is ​​extracted, a large-scale eigenvalue spectral decomposition numerical operation is performed on the Laplacian matrix, the frequency domain orthogonal basis sequence vector group reflecting the global topological structure of the graph is obtained, and the graph convolution smoothing operator is constructed. The graph convolution smoothing operator is obtained, and a polynomial approximate frequency domain filter combination structure with Chebyshev truncation expansion property is constructed based on the preset multi-level neighbor perception receptive field physical range. Random high-frequency hydraulic noise is filtered by dynamically adjusting the convergence weight coefficient of the polynomial, while retaining the low-frequency core signal components, and a graph convolution kernel matrix is ​​generated. Obtain the graph convolution kernel matrix, and calculate the high-dimensional graph features according to the following feature space mapping model: ； in, The high-dimensional graph feature tensor matrix representing the output. This represents the nonlinear hyperbolic tangent activation response function. This represents the maximum truncation order of the Chebyshev polynomial. The index variable represents the order of the polynomial. Representing the Learnable network parameters corresponding to order polynomials, This represents the normalized Laplacian matrix whose eigenvalues ​​are restricted to the interval between -1 and +1 after scaling. Represents the first type of Chebyshev polynomial operation unit. The input is a dynamic directed graph node feature observation data matrix. This represents the weight matrix for cross-layer feature projection transformation in a deep neural network.

10. The municipal pipeline network monitoring system based on the Internet of Things according to claim 5, characterized in that, The process of extrapolating the potential leakage area specifically includes: The historical fault records are obtained, and the seasonal ambient temperature change parameters at the time of the fault and the load benchmark fluctuation index during the peak operation period of the pipeline network are extracted from the historical fault records. The nonlinear error distribution compensation and correction calculation is performed by comparing the current real-time micro pipeline network flow load demand curve with the nonlinear error distribution curve, and a fault spatial source matrix is ​​established. The fault spatial source matrix is ​​obtained, and the Markov Monte Carlo random sampling approximation algorithm is called to generate virtual leak detection particle sources in batches within the limited geographical grid coverage space of the pipeline network. The posterior confidence probability of the virtual leak detection particle sources generating hydraulic state variation phenomena that are consistent with the current terminal observation data is calculated through forward hydraulic simulation, and a fault source probability density heat map is constructed. Extract the geometric set of connected pixel blocks whose probability density distribution values ​​exceed the preset extreme value warning threshold from the probability density heatmap of the fault source. Use the image morphology centroid calculation rule to solve for the coordinates of the geometric center anchor point of the geometric set of connected pixel blocks as a theoretical estimate of the leakage source. Use the theoretical estimate of the leakage source to radiate outward and dynamically adaptively expand the safety buffer isolation distance to deduce the potential leakage area.

Images