Urban pipe network anomaly identification and positioning system based on edge computing

Abstract

The application discloses an urban pipe network anomaly identification and positioning system based on edge computing and relates to the technical field of urban pipe network monitoring. The system comprises a data acquisition module, a data synchronization module, a feature extraction module, an anomaly identification module and a positioning calculation module. The data acquisition module acquires pressure wave, acoustic emission and flow signals collected by edge computing devices deployed on pipe network nodes. The data synchronization module performs time window alignment on multi-source sensing data. The feature extraction module inputs the synchronized data stream into a graph signal processing model, extracts spatial propagation features and time evolution features and constructs a feature vector. The anomaly identification module identifies pipe network anomalies based on a sparse Bayesian learning classifier and generates a trigger signal. The positioning calculation module calculates abnormal event position coordinates based on time difference of arrival and pipe network topology in response to the trigger signal. The application realizes low-delay identification and high-precision positioning of pipe network anomalies on the edge side.

Description

Technical Field

[0001] This invention relates to the field of urban pipeline network monitoring and intelligent operation and maintenance technology, specifically to an urban pipeline network anomaly identification and location system based on edge computing. Background Technology

[0002] In the field of urban pipeline network operation monitoring, traditional methods rely on transmitting large amounts of sensor data back to cloud servers for centralized processing and analysis. These methods typically employ simple threshold comparisons or frequency domain analysis based on Fourier transforms to detect changes in parameters such as pressure and flow, and utilize cross-correlation methods or single time-of-arrival (TOA) models to estimate the location of anomalies. However, urban pipeline networks are complex, with a topological distribution. Signals from abnormal events such as leaks or blockages undergo attenuation, reflection, and multipath interference during propagation, exhibiting strong nonlinearity and coupling in their spatial propagation patterns and temporal evolution. Existing centralized processing methods face bottlenecks in data transmission bandwidth and long processing delays, making it difficult to meet the demands of real-time monitoring and rapid response. Furthermore, methods based on simple thresholds or single frequency domain analysis cannot effectively extract the spatial propagation characteristics of signals within the pipeline network topology, easily misjudging normal operational disturbances as anomalies, and lacking sufficient sensitivity to identify weak leak signals. In addition, traditional TOA location methods do not fully consider the pipeline routing and node connectivity constraints in the actual pipeline network topology, often resulting in solutions to the location equations that deviate from the actual pipeline locations, leading to significant errors in the location results. Therefore, a technical solution is needed that can identify and accurately locate abnormal events in the pipeline network in real time on edge computing devices, in order to solve the problems of high data processing latency, incomplete extraction of abnormal features, and the influence of topology on positioning accuracy in existing technologies. Summary of the Invention

[0003] The present invention aims to provide an urban pipeline network anomaly identification and location system based on edge computing, so as to solve the problems of high latency, low anomaly identification sensitivity and location results deviating from pipeline topology caused by centralized data processing in the prior art.

[0004] The objective of this invention can be achieved through the following technical solutions: An edge computing-based urban pipeline network anomaly identification and location system includes a data acquisition module, a data synchronization module, a feature extraction module, an anomaly identification module, and a location calculation module. The data acquisition module acquires multi-source sensor data collected by edge computing devices deployed at urban pipeline network nodes. This multi-source sensor data includes pressure wave signals, acoustic emission signals, and flow signals. The data synchronization module performs time window alignment on the multi-source sensor data to obtain a synchronized sensor data stream. Specifically, the timestamp sequence of each sensing channel, including pressure sensing, acoustic emission sensing, and flow sensing, is read from the local cache of the edge computing device. The earliest timestamp in the timestamp sequence is selected as the alignment reference point, and starting from this reference point, corresponding raw data segments are extracted from each sensing channel according to a preset time window length. The extracted raw data segments of each sensing channel are packaged according to the sensing channel identifier to generate a synchronized sensing data stream. When a missing data is detected in a sensing channel within a time window, the complete data of the adjacent time window is retrieved from the historical data cache of that sensing channel and inserted into the missing data position to obtain complete time window data. The complete data is a linear prediction value based on the previous time window data of the same sensing channel or a spatial interpolation value based on the same time window data of adjacent nodes. This synchronization mechanism effectively solves the time asynchrony problem caused by clock drift or data packet loss of multi-source sensors, ensures the consistency of input data for subsequent feature extraction and positioning calculation, and improves the accuracy of anomaly identification and positioning.

[0005] The feature extraction module inputs the synchronous sensing data stream into a graph signal processing model pre-deployed on an edge computing device to extract spatial propagation and temporal evolution features. A pipeline network graph structure is established based on the physical connections of the urban pipeline network, with nodes corresponding to pipeline monitoring points and edges corresponding to pipeline connections. The synchronous sensing data stream is mapped onto each node to form a graph signal sequence. A graph Fourier transform is performed on the graph signal sequence to convert it from the node domain to the spectral domain, obtaining a spectral domain signal. A set of graph wavelet basis functions is designed in the spectral domain, and the spectral domain signal is decomposed into multiple scales using these functions to extract spectral coefficients at different scales as spatial propagation features. The graph wavelet basis functions use hot kernel graph wavelets, and the propagation radius of the graph signal in different local areas is controlled by the scale parameter. For the Laplace matrix of the pipeline network graph structure... eigenvalues and its corresponding eigenvectors Defined in the scale parameter Wavelet coefficients of the heat kernel diagram for:

[0006] in: The total number of nodes. For node indexing, For scale parameters, For the heat core filter function, The Fourier coefficients are plotted on the graph. For the feature vector at the node The component at that location. By selecting... Different scale parameters The wavelet coefficients of each node at various scales are calculated to form a multi-scale spatial propagation feature matrix. Simultaneously, an autoregressive moving average model is performed on the graph signal sequence along the time axis to estimate the time-varying parameters of each node as temporal evolution features. Specifically, for the first node in the network diagram structure... The node, in the , will be in the . The signal values ​​for each time window are denoted as follows: Construct an autoregressive moving average model:

[0007] in: Let the order be the autoregressive order. The moving average order is... and These are the autoregressive coefficient and the moving average coefficient, respectively. For nodes In the past Signal values ​​within a time window, For nodes In the past Noise term for each time window, In order to be in Remove the white noise residuals; solve for the parameter set of each node using the maximum likelihood estimation method. As time-varying parameters, the time-varying parameters of all nodes together constitute the time evolution characteristics. This modeling method can capture the autocorrelation and noise effects of pipeline network signals in the time dimension, and enhance the system's sensitivity to slowly evolving anomalies (such as gradual leakage).

[0008] Feature vectors for pipeline network anomalies are constructed based on spatial propagation and temporal evolution characteristics. Each spectral coefficient in the spatial propagation characteristics is multiplied by its corresponding graph wavelet scaling factor to obtain a weighted spectral feature vector. The time-varying parameters of each node in the temporal evolution characteristics are arranged according to node number to obtain a time-varying parameter sequence. A Kronecker product is then performed on the weighted spectral feature vector and the time-varying parameter sequence to obtain a fused feature matrix. ,in The dimension of the weighted spectral eigenvector. The length of the time-varying parameter sequence. Perform higher-order singular value decomposition in tensor decomposition on the fused feature matrix: ,in It is a left singular matrix. It is a right singular matrix. It is a diagonal matrix whose diagonal elements are singular values. Before extraction The largest singular value and its corresponding left singular vector The product of each singular value and its corresponding left singular vector is taken as the principal component factor, i.e., the first singular value. The principal component factors are The principal component factors are arranged in descending order of their singular values, and the top few are selected to form the feature vector of the pipeline network anomaly event. This fusion and dimensionality reduction method effectively combines multi-scale information in the spatial domain with time-varying parameters in the temporal domain, eliminates redundant noise, extracts the most discriminative features, and significantly improves the robustness of anomaly classification.

[0009] The anomaly detection module inputs feature vectors into a sparse Bayesian learning-based classifier to identify the presence of pipeline network anomalies. The sparse Bayesian classifier is pre-trained using samples from both normal and abnormal pipeline network states, storing the weights of the relevant vector machine for each training sample. The sparse Bayesian classifier calculates the posterior probability of a feature vector belonging to the anomaly category and compares this probability with a preset anomaly decision threshold. When the posterior probability is greater than or equal to the threshold, an anomaly is detected, an anomaly trigger signal is generated, and the anomaly confidence level corresponding to the feature vector is recorded. When the posterior probability is less than the threshold, no anomaly is detected, and the feature vector and its corresponding posterior probability are stored in the incremental learning cache of the edge computing device for subsequent online updates of the classifier. Once the data in the incremental learning cache reaches a preset threshold, incremental retraining of the sparse Bayesian classifier is triggered, updating the relevant vector machine weights. This classifier uses a sparse Bayesian framework to automatically select features, retaining only a few relevant vectors to achieve high classification accuracy, significantly reducing the storage and computational overhead of edge devices; at the same time, the incremental learning capability enables the system to adapt to distribution drift caused by pipeline aging, environmental changes, etc., and maintain long-term stable anomaly detection performance.

[0010] In response to an anomaly trigger signal, the positioning calculation module invokes a positioning engine based on the time-of-arrival (TOA) method and the pipeline network topology to calculate the location coordinates of the anomaly. It extracts the arrival times of the anomalous wavefronts detected by each sensor node from the synchronous sensor data stream. These anomalous wavefronts include pressure surge wavefronts and acoustic emission wavefronts. A set of TOA equations is established based on the spatial coordinates of each sensor node and the arrival times of the anomalous wavefronts. Each equation represents a hyperbolic relationship between the TOA between two different sensor nodes and the location of the anomaly. Boundary conditions are applied to the TOA equations in conjunction with pipeline routing constraints and node connectivity constraints in the pipeline network topology to obtain the constrained positioning equations. Specifically, let the location coordinates of the anomaly be... , No. The spatial coordinates of each sensor node are The arrival time of the detected abnormal wavefront is The original arrival time difference equation is: ,in This represents the propagation speed of the anomalous wave in the pipe medium. Apply pipe alignment constraints: [Set the position coordinates...] The constraint is on the set consisting of all pipe segments in the network topology. Above, that is ,in This represents the total number of pipeline segments. For the first A set of points for a pipeline segment; apply node connectivity constraints: location of abnormal events. It must be located on at least one shortest path connecting two sensor nodes, and this shortest path is determined by the edge weights (pipe lengths) in the pipeline network structure. Introducing these two constraints into the time difference of arrival equation yields the constrained positioning equation: ,in To detect the total number of sensor nodes in the abnormal wavefront, The arrival time of the reference node corresponding to the earliest arrival time. Let be the spatial coordinates of the reference node. The constrained positioning equation is solved using a weighted least squares iterative algorithm. The initial iterative values ​​are provided by the coordinates of the pipe intersection closest to the earliest arriving node of the anomalous wavefront in the pipe network topology. Specifically, let the earliest arriving sensor node of the anomalous wavefront be numbered as... Its spatial coordinates are Extract the coordinate set of all pipe intersections from the pipeline network topology. Calculate the Euclidean distance between the earliest arriving node and each intersection point. The coordinates of the intersection point corresponding to the minimum distance are selected as the initial iteration value. ,in If multiple intersection points have the same distance to the earliest arriving node and all of them are at their minimum distances, then the average time difference error between these intersection points and other nodes that detected the wavefront is calculated, and the intersection point with the smallest error is selected as the initial iteration value. This positioning method effectively eliminates non-physical solutions caused by pipe bends and branches by introducing pipe topology constraints, significantly reducing the search space. Combined with weighted least squares iteration and reasonable initial values, it improves the positioning convergence speed and accuracy, and is particularly suitable for complex urban pipe network scenarios.

[0011] After calculating the location coordinates of the abnormal event, the system further correlates these coordinates with the pressure wave attenuation amplitude in the synchronous sensor data stream to obtain the equivalent intensity factor of the abnormal event. Based on this equivalent intensity factor, the system calls upon a pre-established anomaly type knowledge base to classify and identify the abnormal event type, including leakage, blockage, cavitation, or external impact. The system then packages the location coordinates, event type, and corresponding anomaly confidence level of the trigger signal into an anomaly event report, which is uploaded to the cloud management platform and simultaneously stored in the local storage of the edge computing device for updating the local anomaly event database. Historical anomaly event records are retrieved from the local storage of edge computing devices. Anomaly reports are spatiotemporally correlated and matched with these historical records to determine if the current anomaly is a recurring event. If it is determined to be a recurring event, the evolution trend parameters of the previous anomaly are extracted from the historical records and combined with the equivalent intensity factor of the current anomaly to update the severity score. The sampling frequency of the edge computing devices' sensor data is dynamically adjusted based on the updated severity score. When the severity score exceeds a preset warning threshold, the sampling frequency is increased to the maximum sampling rate, and adjacent edge computing devices are configured into collaborative monitoring mode. In collaborative monitoring mode, adjacent edge computing devices exchange their extracted arrival times and morphological characteristics of anomalous wavefronts via direct device communication. This mechanism achieves closed-loop management of anomalies. Through intensity and type identification, recurrence determination, and dynamic severity assessment, it adaptively increases the monitoring density in key areas, avoiding resource waste and enhancing the response capability to sudden severe anomalies.

[0012] Before acquiring multi-source sensor data, the system performs time synchronization calibration and sensor self-test on each edge computing device. The time synchronization calibration employs a dual verification mechanism using both the Network Time Protocol (NTP) and the Global Navigation Satellite System (GNSS) time source to ensure that the time deviation between each edge computing device is less than a preset time deviation threshold, providing a high-precision time reference for subsequent positioning calculations. The sensor self-test process includes zero drift detection, sensitivity detection, and noise floor detection, and generates a sensor health score based on the self-test results. When the sensor health score of an edge computing device falls below a preset health lower limit, the edge computing device is marked as an untrusted node, and its data is excluded from subsequent anomaly identification and positioning calculations, effectively preventing false alarms or missed alarms caused by sensor failures. After identifying an anomaly in the pipeline network, the system also retrieves historical synchronous sensor data streams from the local storage of the edge computing device within a predetermined time window prior to the anomaly as precursor data. This data undergoes recursive quantitative analysis to calculate three recursive features: recursion rate, determinism, and laminarity. Correlation analysis is then performed between these recursive features and the feature vector of the current anomaly to obtain an evolution stage identifier. Based on this stage identifier, corresponding emergency control strategies are matched from a pre-set contingency plan database. These strategies include valve shut-off commands, pump speed adjustment commands, and pressure relief commands. The matched emergency control strategies are then issued as commands to the corresponding pipeline network actuators. This early warning and response linkage mechanism allows the system to trace precursor features, determine the anomaly's development stage, and automatically execute optimal control measures as early as possible, minimizing pipeline network operational risks and economic losses.

[0013] The beneficial effects of this invention are: This study employs a graph signal processing model to extract spatial propagation and temporal evolution features from synchronous sensor data streams. The scheme establishes a network graph structure based on the physical connections of the urban pipeline network, mapping the synchronous sensor data stream into a graph signal sequence. A graph Fourier transform is then performed on the graph signal sequence to convert it to the spectral domain. In the spectral domain, multi-scale decomposition using graph wavelet basis functions extracts spectral coefficients as spatial propagation features. Simultaneously, an autoregressive moving average model is performed along the time axis of the graph signal sequence to estimate the time-varying parameters of each node as temporal evolution features. Compared to conventional methods using single sensor signals or simple Fourier transforms, this scheme can capture the propagation patterns of anomalous signals within the pipeline network topology, distinguish between local disturbances and global propagation events, enhance the ability to identify weak anomalous signals, and improve the robustness of anomaly identification by fusing spatial and temporal features. A positioning engine combining time-of-arrival (TOA) localization with the pipeline network topology is used to calculate the location coordinates of anomalous events. This scheme extracts the arrival times of anomalous wavefronts detected by each sensor node from the synchronous sensor data stream. It then applies boundary conditions to the time difference of arrival (TDOA) equations based on pipeline routing constraints and node connectivity constraints within the pipeline network topology, resulting in constrained location equations. These equations are then solved using a weighted least squares iterative algorithm. Compared to traditional free-space TDOA location algorithms, this scheme utilizes topological constraints to limit the location solution to the actual pipeline segment, preventing deviations from the pipeline and improving consistency between the location results and the actual pipeline network space. Furthermore, the addition of boundary conditions reduces the number of invalid solutions, enabling the iterative algorithm to converge to a physically feasible location more quickly. Attached Figure Description

[0014] The invention will now be further described with reference to the accompanying drawings.

[0015] Figure 1 This is a flowchart of the urban pipeline network anomaly identification and location system based on edge computing described in this invention; Figure 2 This is a flowchart of time window alignment and missing data completion for multi-source sensor data; Figure 3 This is a flowchart for constructing feature vectors of pipeline network anomaly events. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] See Figure 1This invention provides an urban pipeline network anomaly identification and location system based on edge computing. The system specifically includes: a data acquisition module for acquiring multi-source sensor data collected by edge computing devices deployed at urban pipeline network nodes, the multi-source sensor data including pressure wave signals, acoustic emission signals, and flow signals; a data synchronization module for performing time window alignment on the multi-source sensor data to obtain a synchronized sensor data stream; a feature extraction module for inputting the synchronized sensor data stream into a graph signal processing model pre-deployed on the edge computing device, extracting spatial propagation features and temporal evolution features from the synchronized sensor data stream through the graph signal processing model, and constructing a feature vector for pipeline network anomaly events based on the spatial propagation features and the temporal evolution features; an anomaly identification module for inputting the feature vector into a sparse Bayesian learning-based classifier to identify whether a pipeline network anomaly event exists, and generating an anomaly trigger signal when an anomaly event is identified; and a location calculation module for responding to the anomaly trigger signal by calling a location engine based on time difference of arrival (TDOA) and pipeline network topology to calculate the location coordinates of the anomaly event.

[0018] Example 1: In a specific implementation, a time window alignment operation is performed on the multi-source sensor data to obtain a synchronized sensor data stream. (See also...) Figure 2 The timestamp sequence of each sensor channel, including a pressure sensor channel, an acoustic emission sensor channel, and a flow sensor channel, is read from the local cache of the edge computing device. The timestamp sequence records the time information of data acquisition for each sensor channel, and the length of the timestamp sequence for each sensor channel is related to the data acquisition frequency of the corresponding channel.

[0019] In some embodiments, the earliest timestamp in the timestamp sequence is selected as the alignment reference point. Starting from the alignment reference point, an alignment is performed according to a preset time window length. Extract the corresponding raw data segment from each sensing channel. The preset time window length... This is a fixed value, set according to the time resolution of urban pipeline monitoring and the duration of abnormal events. For example, it can be set to 100 milliseconds. This value can cover the normal propagation delay range of pressure waves, acoustic emission signals, and flow signals in the pipeline. The interception operation starts from the alignment reference point and intercepts a length of [length missing]. Data from all sampling points within the time range is used to obtain the raw data fragments for each sensing channel.

[0020] In practice, the original data segments from each sensor channel are packaged according to the sensor channel identifier to generate the synchronous sensing data stream. The sensor channel identifier is a unique number for each sensor channel, for example, the pressure sensor channel is numbered "P01", the acoustic emission sensor channel is numbered "AE01", and the flow sensor channel is numbered "F01". The packaged synchronous sensing data stream is a data structure that contains each sensor channel identifier and its corresponding original data segment. The data segments of all sensor channels in this data stream are aligned in time to the same starting point of the time window.

[0021] In some embodiments, when a data gap is detected in a sensing channel within the time window, complete data from the adjacent time window is retrieved from the historical data cache of that sensing channel, and the complete data is inserted into the missing data position to obtain the complete time window data. The data gap detection method is as follows: for each sensing channel, calculate the data within the time window... If the actual number of sampling points collected is less than the expected number of sampling points calculated according to the sampling frequency of the channel, then it is determined that there is missing data. The location of the missing data is determined by the start and end times of the missing time period.

[0022] Optionally, the supplementary data is a linear prediction based on the previous time window data of the same sensing channel. Specifically, for sensing channels with missing data, the data sequence of the previous complete time window is extracted from the historical data cache of that channel. A linear prediction model is used to extrapolate the data values ​​corresponding to the missing positions within the previous time window, generating predicted values ​​for the missing positions. The linear prediction model employs first-order linear regression, assuming the missing time... Compared to the previous window closing time The time difference is Then the predicted value ,in The average rate of change of data within the previous time window is calculated by dividing the difference between adjacent sampling points within the previous time window by the sampling interval.

[0023] Optionally, the completed data is a spatial interpolation value based on the window data of adjacent nodes at the same time. Specifically, when a certain sensing node (denoted as node...)... When data is missing, identify the nodes from the pipeline topology. The set of sensor nodes that are physically adjacent. The sensor data values ​​of these adjacent nodes at corresponding times within the same time window are extracted, and the node is calculated using the inverse distance weighted spatial interpolation method. In the absence of time completion value The inverse distance weighted formula is:

[0024] in: Adjacent nodes At any moment The sensor data values, As weight, For nodes With neighboring nodes The weight of the pipe length between them is inversely proportional to the square of the distance. The number of nodes. If adjacent nodes also have missing data at the same time, only adjacent nodes with complete data are used for calculation; if all adjacent nodes have missing data, the linear prediction method described above is used to generate complete data.

[0025] Example 2: In a specific implementation, the synchronous sensing data stream is input to a graph signal processing model pre-deployed on an edge computing device. The graph signal processing model extracts the spatial propagation and temporal evolution features of the synchronous sensing data stream. Based on the physical connections of the urban pipeline network, a pipeline network diagram structure is established. Nodes in the pipeline network diagram structure correspond to pipeline monitoring points, and edges correspond to pipeline connections. Pipeline monitoring points are the locations of sensor nodes where edge computing devices are deployed. Pipeline connections are determined by the actual direction and connectivity of the pipelines. Each node is assigned a unique number, and each edge is assigned a pipeline length weight.

[0026] In practical implementation, the synchronous sensing data stream is mapped to each node of the pipeline network structure, forming a graph signal sequence. This graph signal sequence is a collection of multi-channel signals that vary over time. Within each time window, each node corresponds to a multi-source sensing data value (one dimension of pressure wave signal, acoustic emission signal, or flow signal). The data from all nodes form a graph signal vector, denoted as... ,in This represents the total number of nodes in the pipeline network structure. The time window index is used. A graph Fourier transform is performed on the graph signal sequence to convert it from the node domain to the spectral domain, obtaining the spectral domain signal. The graph Fourier transform is defined as: calculating the Laplace matrix of the network diagram structure. Laplace matrix ,in It is a degree matrix, and the diagonal elements are the degree of each node (the number of edges connected to that node). Given an adjacency matrix, if the node and nodes If there is an edge connecting them, then Otherwise, it is 0. For the Laplace matrix... Eigenvalues ​​are obtained by performing eigenvalue decomposition. and the corresponding feature vectors The eigenvalues ​​are arranged in ascending order. The Fourier transform of the graph signal Projecting onto the space spanned by the eigenvectors yields the graphical Fourier coefficients. ,in For nodes In the time window The signal value, For the first Each feature vector at the node The component at that location, For the first Fourier coefficients. The spectral domain signal is the set of all Fourier coefficients across all time windows. .

[0027] In specific implementation, a set of graph wavelet basis functions is designed in the spectral domain. These functions are used to perform multi-scale decomposition of the spectral domain signal, extracting spectral coefficients at different scales. These spectral coefficients are then used as the spatial propagation features. The graph wavelet basis functions employ heat kernel graph wavelets, and the propagation radius of the graph signal in different local regions is controlled by scale parameters. For the Laplace matrix of the network diagram structure... eigenvalues and its corresponding eigenvectors Defined in the scale parameter The wavelet coefficients of the heat kernel diagram are:

[0028] in: This represents the total number of nodes in the pipeline network diagram structure. For node indexing; This is a scaling parameter; a larger value indicates a wider range of signal propagation on the graph. This is a thermonuclear filter function used to control the attenuation level of different frequency components; The signal sequence in the figure is at the 1st Graph Fourier coefficients on each eigenvector; For the first Each feature vector at the node The component at the location; For nodes In scale The wavelet coefficients are selected. Different scale parameters The wavelet coefficients of each node at each scale are calculated to form a multi-scale spatial propagation feature matrix. The scale parameters... The set of values ​​is set based on the average path length and signal propagation speed of the pipeline network topology, for example... The wavelet takes eight values ​​from 0.1 to 10 at logarithmic intervals, allowing it to cover different scales from local node neighborhoods to the global graph. The dimension of the multi-scale spatial propagation feature matrix is... Each element is a node. In scale wavelet coefficients .

[0029] In specific implementation, an autoregressive moving average model is performed on the graph signal sequence along the time axis to estimate the time-varying parameters of each node, and these time-varying parameters are used as the time evolution features. For the first node in the pipeline network diagram structure... The node, in the , will be in the . The signal values ​​for each time window are denoted as follows: Construct the following autoregressive moving average model:

[0030] in: The autoregressive order represents the dependency length of the historical signal values; is the moving average order, and represents the dependency length of the historical noise term; For the first The node of the first Autoregressive coefficient of order; For the first The node of the first The moving average coefficient; This is the white noise residual term at the current moment, with a mean of zero; For nodes In the past Signal values ​​within a time window; For nodes In the past Noise terms within each time window. The autoregressive order. and moving average order Pre-set based on the time correlation analysis of pipeline network signals, for example Setting it to 3 indicates that the current signal value depends on the historical signal values ​​of the previous 3 time windows; Setting it to 2 indicates that the current noise depends on the historical noise terms from the previous two time windows. The parameter set for each node is solved using the maximum likelihood estimation method. The parameter set is used as the node. The time-varying parameters of all nodes collectively constitute the time evolution characteristics. The maximum likelihood estimation method employs Kalman filtering iterative solutions or uses the BIC (Bayesian Information Criterion) to select the optimal order. In specific implementations, for each node, past... A time window (e.g.) Parameter estimation is performed on the historical data sequence to obtain the parameters for each node. Each parameter value. The time-varying parameters of all nodes are arranged according to their node numbers, forming a... The matrix is ​​the time evolution feature.

[0031] Example 3: See Figure 3 In specific implementation, a feature vector of pipeline network anomalies is constructed based on the spatial propagation characteristics and the temporal evolution characteristics. Each spectral coefficient in the spatial propagation characteristics is multiplied by its corresponding graph wavelet scaling factor to obtain a weighted spectral feature vector. The spectral coefficients are elements in the multi-scale spatial propagation feature matrix obtained through heat kernel graph wavelet multi-scale decomposition, and the graph wavelet scaling factor is a pre-set scaling parameter. The reciprocal of the value or the weighting factor calculated based on the scale parameter; in specific implementation, each node is weighted according to the scale. wavelet coefficients Multiply by the corresponding graph wavelet scaling factor ,in The maximum value among all scale parameters is used for normalization. After weighting, the weighted spectral feature vector is obtained, whose dimension is equal to the total number of nodes. Multiply by scale quantity , recorded as All weighted values ​​are arranged into a one-dimensional vector according to the node number and scale number, which is the weighted spectral feature vector.

[0032] In specific implementation, the time-varying parameters of each node in the time evolution feature are arranged according to the node number to obtain a time-varying parameter sequence. The time evolution feature is a set of time-varying parameters for all nodes, and each node... The time-varying parameters include autoregressive coefficients. and moving average coefficient ,common One parameter. Numbered from 1 to... The order of each node The parameters are arranged sequentially to form a string of length. A one-dimensional sequence, denoted as the time-varying parameter sequence, has a length denoted as . .

[0033] In a specific implementation, the weighted spectral feature vector and the time-varying parameter sequence are subjected to a Kronecker product operation to obtain a fused feature matrix. Let the weighted spectral feature vector be... The time-varying parameter sequence is The Kronecker product operation is defined as: fusing the feature matrices ,in Represents the Kronecker product, the result of the operation. The dimension is Each element in the matrix ,in The first eigenvector of the weighted spectral eigenvector One element, The first of the time-varying parameter sequences Each element.

[0034] In specific implementation, tensor decomposition is performed on the fused feature matrix to extract principal component factors. These principal component factors are then sorted from largest to smallest contribution, and the top few principal component factors are selected to form the feature vector of the pipeline network anomaly event. Specifically, performing tensor decomposition on the fused feature matrix to extract principal component factors includes: denoting the fused feature matrix as... ,in The dimension of the weighted spectral eigenvector. Given the length of the time-varying parameter sequence, perform the following higher-order singular value decomposition in tensor decomposition:

[0035] in: It is a left singular matrix, and its column vectors are left singular vectors; It is a right singular matrix, and its column vectors are right singular vectors; It is a diagonal matrix whose diagonal elements are singular values. ; for Transpose. Before extraction. The largest singular value and its corresponding left singular vector The product of each singular value and its corresponding left singular vector is taken as the principal component factor, i.e., the first singular value. The principal component factors are ,in The principal component factors are arranged in descending order of their singular values, and the first few are selected to form the feature vector of the pipeline network anomaly event. The value of is determined based on the cumulative contribution of singular values, and the formula for calculating the cumulative contribution is: In specific implementation, select The smallest integer that makes the cumulative contribution reach 95%. For example, if the cumulative contribution of the first 10 singular values ​​reaches 95%, then... The selected first The principal component factors are concatenated into a one-dimensional vector with dimension 1. (Because each principal component factor) yes (dimensional vector), but in order to construct the final feature vector of pipeline anomalies, each principal component factor is treated as a feature component, that is, the feature vector is composed of... indivual The principal component factors of the dimension are cascaded together, and the total dimension is . Alternatively, reduce the dimensionality of each principal component factor (e.g., retain only the first few components). (element) and then concatenate, but according to the original description of Example 3, "selecting the first few principal component factors to form the feature vector of the pipeline network abnormal event" means directly selecting the first few elements. Each of the principal component factors (each factor being a vector in itself) forms part of the feature vector, therefore the final feature vector dimension is... In practical implementation, the principal component factors can also be flattened into a one-dimensional vector.

[0036] Example 4: In a specific implementation, the feature vector is input into a classifier based on sparse Bayesian learning to identify whether there are any abnormal events in the pipeline network. When an abnormal event is identified, an abnormality trigger signal is generated. The sparse Bayesian classifier is pre-trained using samples of normal and abnormal pipeline network states. The sparse Bayesian classifier stores the relevant vector machine weights corresponding to each training sample.

[0037] In its implementation, the sparse Bayesian classifier adopts a Relevance Vector Machine (RVM) architecture, comprising an input layer, a kernel function layer, and an output layer. The input layer receives feature vectors with a dimension of [missing information]. ,in The length of the feature vector for the pipeline network anomaly event is given. The kernel layer uses a radial basis function (RBF), with the kernel center composed of relevance vectors from the training samples. The output layer outputs the posterior probability of belonging to the anomaly category, calculated as follows: for the input feature vector... Output value ,in The number of related vectors, For the first 1 related vector, For radial basis kernel functions, For the first The weights corresponding to each relevant vector This is the bias term. The posterior probability is converted using the sigmoid function. Mapping to the interval [0,1] yields the posterior probability. .

[0038] In practice, the training process of the sparse Bayesian classifier is completed offline. The training sample set includes data collected from historical operation data of the urban pipeline network. Each feature vector contains samples of the pipeline network under normal and abnormal conditions, with each sample accompanied by a category label. 0 indicates normal, and 1 indicates abnormal. During training, a sparse Bayesian learning algorithm is used to iteratively solve for the relevance vector machine weights. and hyperparameters hyperparameters The sparsity of the weights is controlled. After training, the sparse Bayesian classifier stores each relevance vector. and their corresponding weights .

[0039] In a specific implementation, the posterior probability that the feature vector belongs to the anomaly category is calculated using the sparse Bayesian classifier, and this posterior probability is compared with a preset anomaly judgment threshold. The preset anomaly judgment threshold is denoted as... The value is a real number between 0 and 1, determined based on the balance curve of the false alarm rate and false negative rate of the classifier on the validation set. For example, set... An anomaly is determined when the posterior probability is greater than or equal to 0.85. This value ensures that the false alarm rate is less than 1% and the false negative rate is less than 5% (based on offline verification results).

[0040] In practice, when the posterior probability is greater than or equal to the anomaly judgment threshold, an anomaly event in the pipeline network is determined to exist, and the anomaly trigger signal is generated. Simultaneously, the anomaly confidence level corresponding to the feature vector is recorded. The anomaly confidence level is the numerical value of the posterior probability, stored as a floating-point number in the memory of the edge computing device.

[0041] In practical implementation, when the posterior probability is less than the anomaly judgment threshold, it is determined that there is no pipeline anomaly event, and the feature vector and its corresponding posterior probability are stored in the incremental learning cache of the edge computing device for subsequent online updates of the classifier. The incremental learning cache is a first-in-first-out queue in the local storage of the edge computing device, with a capacity set to... 1 record, for example When the amount of data stored in the cache reaches a preset threshold, incremental retraining of the sparse Bayesian classifier is triggered, updating the weights of the relevant vector machines.

[0042] In specific implementation, the preset quantity threshold is denoted as... The value is set to 500. Incremental retraining is triggered when the number of data entries in the incremental learning cache reaches 500. Incremental retraining uses an online learning-based update algorithm, taking 500 new samples (feature vectors and their corresponding posterior probabilities) from the cache as training data, while retaining the main relevance vectors from the original relevance vector machine model. The hyperparameters and weights are iteratively optimized and adjusted to obtain the updated relevance vector machine weights. More: In specific implementation, in response to the abnormal trigger signal, a positioning engine based on the time-of-arrival (TOA) method and pipeline topology is invoked to calculate the location coordinates of the abnormal event. In response to the abnormal trigger signal, the arrival times of the abnormal wavefronts detected by each sensor node are extracted from the synchronous sensor data stream. These abnormal wavefronts include pressure surge wavefronts and acoustic emission wavefronts. For each sensor node, wavefront detection is performed on the pressure wave signal and acoustic emission signal separately in the synchronous sensor data stream. The arrival time of the wavefront is located using a combination of thresholding and short-time zero-crossing rate: a threshold value for the pressure wave signal is set. Harmony emission signal energy threshold When the signal amplitude exceeds Or signal energy exceeds The time marker is used as the arrival time of the wavefront, and noise interference is eliminated with the help of short-time zero-crossing rate. The extracted wavefront... The arrival times of the anomalous wavefronts of each sensing node are denoted as follows: The spatial coordinates of the sensing node are as follows: ,in and The coordinates (in meters) are pre-calibrated using the Global Navigation Satellite System. A node set is formed for all sensing nodes that detected the anomalous wavefront. ,in To detect the total number of sensor nodes in the abnormal wavefront, and .

[0043] In practical implementation, a set of arrival time difference equations is established based on the spatial coordinates of each sensing node and the arrival time of the anomalous wavefront. Let the location coordinates of the anomalous event be... , No. The spatial coordinates of each sensor node are The arrival time of the detected abnormal wavefront is The original arrival time difference equation is:

[0044] in: This refers to the propagation speed of the abnormal wave in the pipeline medium, measured in meters per second. This propagation speed is predetermined based on the pipeline material and the type of medium. For example, for pressure waves in a cast iron water supply pipeline, the propagation speed is... meters per second; For the first The arrival time of the abnormal wavefront of each sensing node; Let represent the Euclidean norm, i.e., the straight-line distance between two points. The sensor node corresponding to the earliest arrival time is selected as the reference node; its node number is 1, and its arrival time is . For ,get There are several time-of-arrival equations, each representing a hyperbolic relationship between the time-of-arrival difference between two different sensing nodes and the location of the abnormal event.

[0045] In practical implementation, by combining the pipeline routing constraints and node connectivity constraints in the pipeline network topology, boundary conditions are applied to the time difference of arrival equations to obtain the constrained positioning equations. Applying pipeline routing constraints: The position coordinates... The constraint is on the set consisting of all pipe segments in the network topology. Above, that is ,in This represents the total number of pipeline segments. For the first A set of points for each pipeline segment. Each pipeline segment is defined by the coordinates of its start and end points, and the points on the pipeline segment satisfy a linear parametric equation. ,in and The first The coordinates of the start and end points of the pipeline segment, parameters. Apply node connectivity constraints: Location of abnormal events It must be located on at least one shortest path connecting two sensor nodes, which is determined by the edge weights (pipe lengths) in the network graph structure. The shortest path is calculated on the network graph structure using Dijkstra's algorithm, where the edge weights of the graph structure are the pipe lengths between adjacent nodes.

[0046] In practical implementation, the above two constraints are introduced into the arrival time difference equation to form the constrained positioning equation:

[0047] in: To detect the total number of sensor nodes in the abnormal wavefront, The arrival time of the reference node corresponding to the earliest arrival time. Let be the spatial coordinates of the reference node. The objective of this optimization problem is to minimize the sum of squares of the residuals of the arrival time difference equations for all sensing nodes, while simultaneously finding the solution... It must be located on any pipe segment. The coordinates of the anomalous event location that satisfy the topological constraints are obtained by solving this minimization problem.

[0048] In practice, a weighted least squares iterative algorithm is used to solve the constrained positioning equation to obtain the location coordinates of the abnormal event. The weighted least squares iterative algorithm is as follows: first, an initial iteration value is set. Then in the In the next iteration, the current residual vector is calculated. , of which Each component is Construct the weight matrix The weight , For the first The standard deviation of the measurement error at the arrival time of each sensing node is set according to the sensor performance parameters, for example... Seconds. Calculate the Jacobian matrix. , its first row pair The partial derivatives are Update iteration values: After each iteration, Project onto the nearest pipe segment: Traverse all pipe segments ,calculate Find the minimum Euclidean distance to each line segment, select the line segment with the minimum distance, and set... Replace with the coordinates of the nearest point on the line segment. Repeat the iteration until... Less than the convergence threshold (For example (meters), at this time That is, the coordinates of the location where the abnormal event occurred.

[0049] In practical implementation, the initial iteration values ​​of the weighted least squares iterative algorithm are provided by the coordinates of the pipe intersection point closest to the earliest arriving node of the anomalous wavefront in the pipeline network topology. Let the earliest arriving sensor node of the anomalous wavefront be numbered as follows: Its spatial coordinates are Extract the coordinate set of all pipe intersections from the pipeline network topology. ,in This represents the total number of intersections. A pipe intersection is defined as a node where two or more pipe segments intersect. Calculate the earliest reached node. Euclidean distance to each intersection:

[0050] The coordinates of the intersection point corresponding to the minimum distance are selected as the initial iteration value. :

[0051] If multiple intersection points have the same minimum distance to the earliest arriving node, then the average time difference of arrival error between these intersection points and other nodes that detected the wavefront is calculated, and the intersection point with the smallest error is selected as the initial iteration value. The average time difference of arrival error is calculated as follows: for each candidate intersection point with the smallest distance... Using this as a hypothetical location of the anomalous event, the predicted time difference of arrival is calculated based on the time difference of arrival equation, and then compared with the actual time difference of arrival. By comparison, the sum of squared errors is obtained. ,choose The smallest candidate intersection point is used as the initial iteration value.

[0052] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. An urban pipeline network anomaly identification and location system based on edge computing, characterized in that, include: The data acquisition module is used to acquire multi-source sensor data collected by edge computing devices deployed at urban pipeline network nodes. The multi-source sensor data includes pressure wave signals, acoustic emission signals, and flow signals. The data synchronization module is used to perform time window alignment operation on the multi-source sensor data to obtain a synchronized sensor data stream; The feature extraction module is used to input the synchronous sensing data stream into a graph signal processing model pre-deployed on an edge computing device, extract the spatial propagation features and temporal evolution features in the synchronous sensing data stream through the graph signal processing model, and construct the feature vector of pipeline abnormal events based on the spatial propagation features and the temporal evolution features. An anomaly detection module is used to input the feature vector into a classifier based on sparse Bayesian learning to identify whether there are any abnormal events in the pipeline network, and to generate an anomaly trigger signal when an abnormal event in the pipeline network is detected. The positioning calculation module is used to respond to the abnormal trigger signal by calling the positioning engine based on the time difference of arrival and the pipeline topology to calculate the location coordinates of the abnormal event.

2. The urban pipeline network anomaly identification and location system based on edge computing according to claim 1, characterized in that, Perform a time window alignment operation on the multi-source sensor data to obtain a synchronized sensor data stream, including: The timestamp sequence of each sensing channel, including a pressure sensing channel, an acoustic emission sensing channel, and a flow sensing channel, is read from the local cache of the edge computing device. The earliest timestamp in the timestamp sequence is selected as the alignment reference point, and the corresponding original data segment is extracted from each sensing channel according to the preset time window length, starting from the alignment reference point. The original data segments of each sensor channel are packaged according to the sensor channel identifier to generate the synchronous sensor data stream. When a missing data is detected in a certain sensing channel within the time window, the complete data of the adjacent time window is retrieved from the historical data cache of that sensing channel, and the complete data is inserted into the missing data position to obtain the complete time window data.

3. The urban pipeline network anomaly identification and location system based on edge computing according to claim 2, characterized in that, The completed data is either a linear prediction value based on the previous window data of the same sensing channel or a spatial interpolation value based on the window data of adjacent nodes at the same time.

4. The urban pipeline network anomaly identification and location system based on edge computing according to claim 1, characterized in that, The synchronous sensing data stream is input into a graph signal processing model pre-deployed on an edge computing device. The graph signal processing model extracts the spatial propagation features and temporal evolution features from the synchronous sensing data stream, including: Based on the physical connection relationship of the urban pipeline network, a pipeline network diagram structure is established. The nodes in the pipeline network diagram structure correspond to pipeline monitoring points, and the edges in the pipeline network diagram structure correspond to pipeline connection relationships. The synchronous sensing data stream is mapped onto each node of the network diagram structure to form a graph signal sequence; Perform a graph Fourier transform on the graph signal sequence to convert the graph signal sequence from the node domain to the spectral domain, and obtain a spectral domain signal; In the spectral domain, a set of graph wavelet basis functions are designed. The spectral domain signal is decomposed into multiple scales using the graph wavelet basis functions to extract spectral coefficients at different scales. The spectral coefficients are then used as the spatial propagation features. Simultaneously, an autoregressive moving average model is performed on the graph signal sequence along the time axis to estimate the time-varying parameters of each node, and the time-varying parameters are used as the time evolution features.

5. The urban pipeline network anomaly identification and location system based on edge computing according to claim 4, characterized in that, The graph wavelet basis function adopts the thermal kernel graph wavelet, and the propagation radius of the graph signal in different local areas is controlled by the scale parameter.

6. The urban pipeline network anomaly identification and location system based on edge computing according to claim 4, characterized in that, Based on the spatial propagation characteristics and the temporal evolution characteristics, a feature vector for pipeline network anomaly events is constructed, including: Multiply each spectral coefficient in the spatial propagation feature by the corresponding graph wavelet scaling factor to obtain the weighted spectral feature vector; Arrange the time-varying parameters of each node in the time evolution feature according to the node number to obtain the time-varying parameter sequence; Perform a Kronecker product operation on the weighted spectral feature vector and the time-varying parameter sequence to obtain a fused feature matrix; Tensor decomposition is performed on the fused feature matrix to extract principal component factors. The principal component factors are sorted from largest to smallest according to their contribution, and the top few principal component factors are selected to form the feature vector of the pipeline network abnormal event.

7. The urban pipeline network anomaly identification and location system based on edge computing according to claim 1, characterized in that, The feature vector is input into a sparse Bayesian learning-based classifier to identify whether there are pipeline network anomalies. Upon detection of an anomaly, an anomaly trigger signal is generated, including: The feature vector is used as the input of a sparse Bayesian classifier, which is pre-trained using samples of normal and abnormal states of the pipeline network. The sparse Bayesian classifier stores the weights of the relevant vector machine corresponding to each training sample. The posterior probability of the feature vector belonging to the anomaly category is calculated by the sparse Bayes classifier, and the posterior probability is compared with a preset anomaly judgment threshold. When the posterior probability is greater than or equal to the anomaly judgment threshold, it is determined that there is a pipeline anomaly event, and the anomaly trigger signal is generated. At the same time, the anomaly confidence level corresponding to the feature vector is recorded. When the posterior probability is less than the anomaly judgment threshold, it is determined that there is no pipeline anomaly event, and the feature vector and its corresponding posterior probability are stored in the incremental learning cache of the edge computing device for subsequent online updates of the classifier.

8. The urban pipeline network anomaly identification and location system based on edge computing according to claim 7, characterized in that, Once the data in the incremental learning cache reaches a preset threshold, the incremental retraining of the sparse Bayesian classifier is triggered, and the weights of the relevant vector machines are updated.

9. The urban pipeline network anomaly identification and location system based on edge computing according to claim 7, characterized in that, In response to the aforementioned abnormal trigger signal, a positioning engine based on the time difference of arrival (TDOA) and the pipeline topology is invoked to calculate the location coordinates of the abnormal event, including: In response to the abnormal trigger signal, the arrival time of the abnormal wavefront detected by each sensing node is extracted from the synchronous sensing data stream. The abnormal wavefront includes a pressure change wavefront and an acoustic emission wavefront. Based on the spatial coordinates of each sensor node and the arrival time of the abnormal wavefront, a set of arrival time difference equations is established. Each equation in the set of arrival time difference equations represents the hyperbolic relationship between the arrival time difference between two different sensor nodes and the location of the abnormal event. By combining the pipeline routing constraints and node connectivity constraints in the pipeline network topology, boundary conditions are applied to the time difference of arrival equations to obtain the constrained positioning equations. The location equation after the constraints are solved by using a weighted least squares iterative algorithm to obtain the location coordinates of the abnormal event.

10. The urban pipeline network anomaly identification and location system based on edge computing according to claim 9, characterized in that, The initial iteration value of the weighted least squares iterative algorithm is provided by the coordinates of the pipe intersection point closest to the earliest arriving node of the anomalous wavefront in the pipeline network topology.

Images