A method for detecting and locating multi-pipe leakage of a water supply network
By collecting and processing time-series pressure data from water supply networks, and utilizing leakage inference calculation models and state inversion algorithms, an inversion equation set is constructed. This solves the accuracy problem of multi-pipe leakage detection and location in existing technologies, enabling the quantification and type differentiation of leakage, and generating detailed detection and location reports.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SHANDONG GREEN SOURCE WATER SAVING TECHNOLOGY RESEARCH INSTITUTE CO LTD
- Filing Date
- 2026-04-03
- Publication Date
- 2026-06-09
AI Technical Summary
Existing water supply network leakage detection and location technologies cannot accurately identify leakage in multiple pipelines, cannot quantify the probability of leakage, and are difficult to accurately detect and locate leakage in scenarios with concurrent leakage in multiple pipelines. Furthermore, they are prone to missed detections and false diagnoses.
Pressure time-series data from monitoring points in the water supply network are collected to generate standardized pressure fluctuation sequences. A pre-trained leakage inference calculation model is used to score the likelihood of leakage and classify its type. A transient flow state inversion equation set is constructed by combining the water supply network topology and pipeline hydraulic model. The changes in the internal hydraulic state of the network are reversed through the state inversion algorithm to identify abrupt changes in pressure gradient and locate specific pipe sections.
It enables accurate detection and location of leaks in multiple pipes of the water supply network, quantifies the probability of leak occurrence, distinguishes different types of leaks, filters out noise interference, and generates detailed leak detection and location reports, adapting to complex working conditions with concurrent leaks in multiple pipes.
Smart Images
Figure CN121980409B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of water supply network operation and maintenance testing technology, and in particular to a method for detecting and locating leaks in multiple pipes of a water supply network. Background Technology
[0002] Current methods for detecting and locating leaks in water supply networks mostly employ single-point pressure monitoring combined with simple threshold judgments. Some technologies rely on conventional hydraulic models for leak analysis, which can only provide preliminary identification of leaks in a single pipe. They cannot quantify the probability of a leak occurring, nor can they classify leaks of different causes and forms. When there are abnormal fluctuations in network pressure, these technologies tend to misclassify non-leakage interference signals as leak events, failing to accurately extract the spatiotemporal information corresponding to valid leak events. In scenarios where multiple pipes experience simultaneous pressure anomalies, they are prone to both missed and false positives.
[0003] Current methods for locating leaks in pipe networks primarily employ forward hydraulic simulation, which can only roughly define the leak area based on pressure changes at monitoring points. It cannot incorporate inverse calculations using transient flow conditions, making it difficult to capture abrupt changes in pressure gradients within pipe segments. Existing technologies cannot use standardized pressure fluctuation sequences as boundary conditions to construct inversion equations to deduce changes in the hydraulic state of the pipe network. Furthermore, they cannot combine leak classification information derived from the model with pipe segment location results, failing to generate complete multi-pipe leak detection and location information and making them ill-suited for complex pipe network operation and maintenance scenarios involving concurrent leaks in multiple pipes. Summary of the Invention
[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a method for detecting and locating leaks in multiple pipes of a water supply network.
[0005] To achieve the above objectives, the present invention adopts the following technical solution: a method for detecting and locating leakage in multiple pipes of a water supply network, comprising:
[0006] Collect and process pressure time-series data from selected monitoring points in the water supply network to generate standardized pressure fluctuation sequences;
[0007] Based on a pre-trained leakage inference calculation model, pressure time series data of water supply network are analyzed to obtain leakage suspicion score and leakage type classification;
[0008] Based on the suspected leakage score, suspected leakage events that exceed the set threshold are filtered out, and the monitoring point location and timestamp of their occurrence are obtained;
[0009] Using the topology model of the water supply network, the hydraulic model of the pipeline, and the occurrence time of the suspected leakage event, a set of inversion equations for the transient flow state of the network is constructed.
[0010] The state inversion algorithm is used to solve the transient flow state inversion equation set of the pipeline network. The state inversion algorithm uses the standardized pressure fluctuation sequence measured by the pressure sensor as the boundary condition to reverse the hydraulic state changes of each pipe segment inside the pipeline network within the time window of the suspected leakage event.
[0011] Based on the hydraulic state changes derived from the state inversion algorithm, the pipe network connection points where the pressure gradient changes abruptly are identified, and the specific pipe segments associated with the pressure gradient abrupt change are located.
[0012] By combining the leakage type classification output by the leakage inference calculation model with the specific pipe segment located by the state inversion algorithm, a multi-pipe leakage detection and location report containing leakage location, leakage type and occurrence time is generated.
[0013] As a further aspect of the present invention, the step of collecting and processing pressure time-series data from selected monitoring points in the water supply network to generate a standardized pressure fluctuation sequence includes:
[0014] Pressure time-series data of selected monitoring points in the water supply network are collected, and the pressure time-series data is continuously acquired by pressure sensors deployed at key nodes of the network.
[0015] The collected pressure time-series data undergoes signal preprocessing to generate a standardized pressure fluctuation sequence, specifically including:
[0016] Wavelet threshold denoising is applied to the pressure time series data to separate and filter out high-frequency noise components;
[0017] A polynomial fitting method was used on the denoised pressure data to extract the long-term trend term caused by changes in water use patterns, resulting in a pressure residual sequence after removing the trend term.
[0018] The pressure residual sequence is resampled at fixed time intervals to form the standardized pressure fluctuation sequence;
[0019] The standardized pressure fluctuation sequence is normalized to eliminate the magnitude difference between different monitoring points caused by different reference pressures.
[0020] As a further aspect of the present invention, pressure time-series data of the water supply network are analyzed based on a pre-trained leakage inference calculation model to obtain leakage suspicion scores and leakage type classifications, including:
[0021] Based on the standardized pressure fluctuation sequence, abnormal pressure events contained therein are identified by a pattern recognition algorithm. These abnormal pressure events are characterized by an abnormal sudden drop or continuous decline in pressure values.
[0022] Extract the waveform features of each of the abnormal pressure events;
[0023] The extracted waveform features are input into a pre-trained leakage inference calculation model, which outputs a leakage suspicion score and a leakage type classification, specifically including:
[0024] The leakage inference calculation model is based on a gradient boosting decision tree architecture, and its input layer receives a feature vector composed of the waveform features.
[0025] The hidden layer of the leakage inference calculation model splits and judges the input feature vector through a multi-level decision tree, and extracts nonlinear feature combinations related to leakage layer by layer.
[0026] In the output layer of the leakage inference calculation model, one branch outputs the leakage suspicion score, which is a continuous value between zero and one.
[0027] In the output layer of the leakage inference computation model, another branch outputs the classification probability of the leakage type.
[0028] As a further aspect of the present invention, the method of constructing a set of inversion equations for transient flow state of the water supply network using the topological model of the water supply network, the hydraulic model of the pipeline, and the occurrence time of the suspected leakage event includes:
[0029] The topology model defines the connection relationships between all pipes, nodes, water sources and monitoring points in the pipeline network;
[0030] The pipeline hydraulic model includes the pipeline's length, diameter, material, friction coefficient, and valve status hydraulic properties.
[0031] The occurrence time of the suspected leakage event is taken as the starting time of the state inversion, and the data of the standardized pressure fluctuation sequence before and after this time are taken as known observations;
[0032] Based on the equations of mass conservation and momentum conservation, a set of equations is established with the flow rate of all unknown nodes in the pipeline network and the pressure drop of the pipeline section as state variables. The coefficient matrix of the set of equations is determined by the topology model and the pipeline hydraulic model.
[0033] Substituting the known observations as boundary conditions into the system of equations forms a closed system of equations for the transient flow state inversion of the pipeline network.
[0034] As a further aspect of the present invention, the state inversion algorithm is used to solve the transient flow state inversion equation set of the pipeline network. The state inversion algorithm uses the standardized pressure fluctuation sequence measured by pressure sensors as boundary conditions to reverse-engineer the hydraulic state changes of each pipe segment within the pipeline network within the time window of the suspected leakage event, including:
[0035] The state inversion algorithm adopts the adjoint method, which constructs the objective function by minimizing the difference between the pipeline node pressure calculated by the model and the observed pressure in the standardized pressure fluctuation sequence.
[0036] By solving the gradient of the objective function with respect to the state variables, the unknown state variables in the transient flow state inversion equation set of the pipeline network are iteratively adjusted.
[0037] In each iteration, the pressure distribution of the entire pipeline network is recalculated using the adjusted state variables and compared with the observed values to update the gradient direction;
[0038] When the value of the objective function converges to within the preset tolerance or reaches the maximum number of iterations, the iteration stops. At this point, the solution of the state variable is the hydraulic state change of each pipe segment inside the pipeline network obtained by inversion.
[0039] As a further aspect of the present invention, based on the hydraulic state changes derived from the state inversion algorithm, the pipe network connection points where pressure gradients abruptly change are identified, and the specific pipe segments associated with the pressure gradient abrupt changes are located, including:
[0040] Based on the hydraulic state solution finally obtained by the state inversion algorithm, the pressure gradient distribution along the pipe in the network is calculated.
[0041] In the pressure gradient distribution, search for spatial locations where the pressure gradient value exceeds the threshold of the normal fluctuation range;
[0042] The searched spatial location points are mapped back to the topology model of the water supply network to determine the network connection point where the pressure gradient change occurs.
[0043] In the topology model, all pipe segments directly connected to the network connection point are retrieved, and the specific pipe segment where leakage occurs is determined by combining the direction of the pressure gradient.
[0044] As a further aspect of the present invention, the process of combining the leakage type classification output by the leakage inference calculation model with the specific pipe segment located by the state inversion algorithm to generate a multi-pipe leakage detection and location report containing leakage location, leakage type, and occurrence time includes:
[0045] The specific pipe segment number, start node, and end node information located by the state inversion algorithm are used as leakage location information;
[0046] The leakage type information is classified into the suspected leakage events associated with specific pipe sections output by the leakage inference calculation model.
[0047] The time of occurrence of the suspected leakage event is used as the leakage occurrence time information;
[0048] The leakage location information, the leakage type information, and the leakage occurrence time information are associated and integrated;
[0049] According to the preset report template, the integrated information is filled in, and the pressure gradient distribution map used for location analysis is attached to form the final multi-pipe leakage detection and location report.
[0050] As a further aspect of the present invention, the construction steps of the leakage inference calculation model include:
[0051] Acquire pressure time-series data of the water supply network under normal operating conditions and various known leakage events during historical periods to form an initial training dataset;
[0052] For each known leakage event in the initial training dataset, the signal preprocessing step is performed to generate the corresponding standardized pressure fluctuation sequence.
[0053] From the standardized pressure fluctuation sequence, the abnormal pressure events and their corresponding waveform features are extracted based on the labeled time windows of known leakage events;
[0054] The extracted waveform features are associated with the corresponding real labels of the leakage events. The real labels include leakage existence labels and leakage type labels, which constitute feature-label sample pairs for model training.
[0055] The gradient boosting decision tree algorithm is adopted, taking the waveform features in the feature-label sample pairs as input and the corresponding leakage existence label and leakage type label as supervision targets, and iteratively optimizing the model parameters during training.
[0056] During model training, the feature-label sample pairs are divided into a training subset and a validation subset. The training subset is used to update the model parameters, and the validation subset is used to evaluate the model's predictive performance on leakage suspicion score and leakage type classification, so as to prevent model overfitting.
[0057] When the model's prediction performance on the validation subset meets the preset accuracy and recall metrics, training stops, and the trained leakage inference calculation model is obtained.
[0058] As a further aspect of the present invention, the establishment of a set of equations based on the mass and momentum conservation equations, with the flow rate at all unknown nodes in the pipeline network and the pressure drop in the pipeline section as state variables, includes:
[0059] For each node in the pipeline topology model, a node flow balance equation is established based on the mass conservation equation.
[0060] For each section of the pipeline in the hydraulic model of the pipeline network, the pipeline pressure drop equation is established based on the momentum conservation equation;
[0061] The nodal flow balance equations established for all nodes and the pipeline pressure drop equations established for all pipe segments are combined to form a set of nonlinear equations with the flow rates of all unknown nodes and the pressure drops of all unknown pipe segments in the pipeline network as state variables.
[0062] The coefficient matrix of the nonlinear equation system is jointly determined by the node-pipe connection relationship defined by the topology model and the pipe attribute parameters defined by the pipe hydraulic model.
[0063] As a further aspect of the present invention, the unknown state variables in the transient flow state inversion equation set of the pipeline network are iteratively adjusted by solving the gradient of the objective function with respect to the state variables, including:
[0064] The objective function is defined as the squared L2 of the difference between the pressure vector at the pipeline node calculated by the model and the observed pressure vector in the standardized pressure fluctuation sequence measured in reality.
[0065] Calculate the gradient vector of the objective function with respect to all unknown state variables in the system of inversion equations for the transient flow state of the pipeline network;
[0066] Based on the calculated gradient vector, determine the search direction that causes the objective function to decrease;
[0067] Along the search direction, the unknown state variables are updated with a preset step size to obtain a new set of state variable estimates;
[0068] Substitute the updated state variable estimates into the pipeline transient flow state inversion equation set, recalculate the pressure distribution of the entire pipeline, and then calculate the new model-calculated pressure vector.
[0069] Based on the new model, the pressure vector and the observed pressure vector are calculated, and the value of the objective function is recalculated.
[0070] Determine whether the recalculated value of the objective function is less than the preset convergence tolerance, or whether the number of iterations has reached the preset maximum number of iterations;
[0071] If the stopping condition is not met, the difference between the pressure vector and the observed pressure vector is calculated based on the new model, the gradient vector of the objective function with respect to the updated state variables is recalculated, and the steps of determining the search direction, updating the state variables, recalculating the pressure distribution and objective function, and determining whether the stopping condition is met are repeated.
[0072] If the stopping condition is met, the iteration process is terminated, and the estimated value of the state variable obtained from the last update is taken as the final solution of the state inversion algorithm, that is, the hydraulic state changes of each pipe segment inside the pipeline network within the time window of the suspected leakage event obtained by inversion.
[0073] Compared with the prior art, the advantages and positive effects of the present invention are as follows:
[0074] Pressure time-series data collected and standardized from monitoring points in the water supply network are analyzed using a pre-trained leakage inference calculation model. The model outputs leakage suspicion scores and leakage type classification results. Based on set thresholds, high-susceptibility leakage events are filtered out. The location and timestamp of the monitoring point corresponding to the event are obtained simultaneously. This allows for direct quantification of the probability of leakage occurrence corresponding to abnormal network pressure, differentiation of leakage types with different characteristics, filtering out invalid abnormal signals caused by pressure fluctuation noise, accurately locating the spatial location and occurrence time of leakage events to be analyzed in depth, and distinguishing the leakage attributes and credibility of different pressure anomalies in the network.
[0075] By combining the water supply network topology model, pipeline hydraulic model, and the occurrence time of suspected leakage events, a set of transient flow state inversion equations for the pipeline network is constructed. Standardized pressure fluctuation sequences measured by pressure sensors are used as boundary conditions. The state inversion algorithm is used to reverse-engineer the hydraulic state changes of each pipe segment within the corresponding time window. This can identify abrupt changes in pressure gradients at pipeline connection points, determine the specific pipe segment associated with the abrupt change, and combine the leakage type classification results with the specific pipe segment information obtained from the location to form a multi-pipeline leakage detection and location report covering the leakage location, leakage type, and occurrence time. This report is suitable for complex operating conditions with multiple pipelines experiencing concurrent leakage, clearly presenting the pipe segment, type, and occurrence time information corresponding to each leakage, and comprehensively presenting the core detection results of multiple pipeline leakage. Attached Figure Description
[0076] Figure 1 This is a flowchart of a method for detecting and locating leakage in multiple pipes of a water supply network, as described in this invention.
[0077] Figure 2 A flowchart for constructing the inversion equation set of transient flow state in pipeline network;
[0078] Figure 3 This is a topology model diagram of the water supply network;
[0079] Figure 4 Convergence curve of the water supply network leakage detection status inversion algorithm;
[0080] Figure 5 A flowchart for identifying abrupt changes in pressure gradients and locating specific pipe segments. Detailed Implementation
[0081] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.
[0082] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.
[0083] See Figure 1 Pressure time-series data were collected from selected monitoring points in the water supply network and processed to generate standardized pressure fluctuation sequences. A pre-trained leakage inference calculation model was used to analyze the pressure time-series data of the water supply network. This model can output a leakage suspicion score and classify leakage types. Based on the calculated leakage suspicion score, suspected leakage events with scores higher than a set threshold were screened, and the monitoring point locations and specific timestamps of these events were recorded. Combining the topology model of the water supply network, the pipeline hydraulic model, and the occurrence time of the identified suspected leakage events, an inversion equation set describing the transient flow state of the network was constructed. The state inversion algorithm was used to solve this equation set. This algorithm uses the standardized pressure fluctuation sequence obtained from pressure sensors as boundary conditions to inversely deduce the changes in the hydraulic state of each pipe segment within the network within the time window of the suspected leakage event. Based on the hydraulic state change results derived by the state inversion algorithm, the connection points in the network where pressure gradients change abruptly, and then the specific pipe segments associated with these pressure gradient changes are located. By combining the leakage type classification results output by the integrated leakage inference calculation model with the specific pipe segment information located by the state inversion algorithm, a detailed report is generated. This report includes the location of the leakage, its type, and the time of occurrence, thus completing the detection and location of leakage in multiple pipelines.
[0084] In one embodiment of the present invention, pressure time-series data from selected monitoring points in a water supply network are collected and processed to generate a standardized pressure fluctuation sequence. The pressure time-series data is continuously acquired by pressure sensors deployed at key nodes in the network. The sensors record water pressure values at specific locations within the network at a fixed sampling frequency, forming a raw data sequence corresponding to time and pressure values. In a specific implementation, the collected pressure time-series data undergoes signal preprocessing, including noise reduction, trend term removal, and data resampling. Wavelet thresholding is applied to the pressure time-series data. A suitable wavelet basis function and decomposition level are selected to perform multi-resolution analysis on the raw pressure time-series data, separating high-frequency detail coefficients characterizing noise. By setting a threshold, wavelet coefficients with amplitudes below the threshold are set to zero. Then, wavelet reconstruction is performed on the processed coefficients to separate and filter out high-frequency noise components, obtaining pre-denoised pressure data.
[0085] In some embodiments, a polynomial fitting method is used to remove the long-term trend term caused by changes in water usage patterns from the denoised pressure data. A polynomial curve matching the trend of the denoised pressure data is fitted using the least squares method. This curve reflects the periodic or gradual variation of water load. The values of the corresponding points on this polynomial curve are subtracted point by point from the original denoised pressure data to obtain the pressure residual sequence after removing the trend term. The pressure residual sequence mainly contains pressure transient signals caused by sudden leakage events and residual random fluctuations. In some embodiments, the pressure residual sequence is resampled at fixed time intervals. Linear interpolation or spline interpolation methods are used to map the non-uniform or asynchronous pressure residual sequence data points onto a unified Gaussian time axis, ensuring that the data from all monitoring points have a uniform time resolution and forming a standardized pressure fluctuation sequence that is strictly aligned at time points.
[0086] Optionally, the standardized pressure fluctuation series is normalized to eliminate magnitude differences. Normalization is performed independently for each standardized pressure fluctuation series generated at each monitoring point. The arithmetic mean and standard deviation of all data points in the standardized pressure fluctuation series for a single monitoring point are calculated. By subtracting the mean from each data point in the series and then dividing by the standard deviation, the original pressure fluctuation values are converted into standardized values with a mean of zero and a standard deviation of one. The formula used for normalization can be understood as follows:
[0087]
[0088] in: Representing the The standardized pressure fluctuation sequence of the monitoring points, after normalization processing, is the first... A number, The corresponding number in the standardized pressure fluctuation sequence at that monitoring point The original pressure fluctuation value, This represents the arithmetic mean of all values in the entire standardized pressure fluctuation sequence at that monitoring point. This represents the standard deviation of all values in the standardized pressure fluctuation series for that monitoring point. It can be understood that, after the above processing, the overall data offset and amplitude scale differences caused by variations in installation location and reference pressure at different monitoring points are eliminated. The pressure fluctuation data from all monitoring points are transformed to the same statistical scale, facilitating subsequent unified analysis and comparison.
[0089] In one embodiment of the present invention, a pre-trained leakage inference calculation model is used to analyze the pressure time-series data of the water supply network to obtain a leakage suspicion score and leakage type classification. Based on a standardized pressure fluctuation sequence, an abnormal pressure event is identified using a pattern recognition algorithm. Abnormal pressure events are characterized by an abnormal sudden drop or continuous decrease in pressure value. The pattern recognition algorithm traverses the standardized pressure fluctuation sequence through a sliding time window, calculates the statistical characteristics of the pressure data within the window, and compares them with historical normal fluctuation intervals to mark abnormal segments that deviate from the normal pattern. Waveform features of each abnormal pressure event are extracted. These features include the slope of the pressure drop, the depth of the pressure trough, the shape of the pressure recovery, and the duration of the event. The slope of the pressure drop is obtained by calculating the ratio of the pressure change to the time change between the starting point of the abnormal pressure event and the lowest pressure point. The depth of the pressure trough refers to the magnitude of the pressure drop relative to the steady-state pressure baseline value before the event. The shape of the pressure recovery is described by the curve shape parameters quantifying the process of pressure recovering from the trough to a stable state. The duration of the event refers to the time elapsed from the onset of the abnormal pressure drop to the recovery to a stable state.
[0090] In some embodiments, the extracted waveform features are input into a pre-trained leakage inference computation model, which outputs a leakage suspicion score and a leakage type classification. The leakage inference computation model is based on a gradient boosting decision tree architecture. Its input layer receives a feature vector composed of waveform features, which is a one-dimensional array consisting of multiple waveform feature values such as the slope of pressure drop, the depth of pressure trough, the morphological parameters of pressure recovery, and the duration of the event, arranged in a fixed order. The hidden layer of the leakage inference computation model splits and judges the input feature vector through multi-level decision trees. Each decision tree sets internal node splitting conditions based on different components and values of the feature vector, dividing the input samples into different leaf nodes. The multi-level decision trees are combined in an additive model, extracting nonlinear feature combinations related to leakage layer by layer. In the output layer of the leakage inference computation model, one branch outputs a leakage suspicion score, which is a continuous value between zero and one. This score is obtained by weighted summation of the output values of all decision trees and mapping through a sigmoid function. In the output layer of the leakage inference computation model, another branch outputs the classification probability of the leakage type. The leakage types include perforation leakage, burst pipe leakage, and interface leakage. For multi-classification tasks, another branch of the gradient boosting decision tree architecture will calculate a probability value for each leakage type, and the sum of the probabilities of all types is one.
[0091] In some embodiments, the construction steps of the leakage inference calculation model include acquiring pressure time-series data of the water supply network under normal operating conditions and various known leakage events over historical periods to form an initial training dataset. The historical data covers network pressure change scenarios under different seasons, time periods, and leakage types. For the pressure time-series data corresponding to each known leakage event in the initial training dataset, a signal preprocessing step is performed to generate a corresponding standardized pressure fluctuation sequence. The process and parameter settings of the signal preprocessing step must be consistent with the real-time analysis process. From the standardized pressure fluctuation sequence, based on the labeled time window of the known leakage events, abnormal pressure events and their corresponding waveform features are extracted. The waveform features include the slope of the pressure drop, the depth of the pressure trough, the shape of the pressure recovery, and the duration of the event. The labeled time window of the known leakage events is accurately calibrated manually or by an automated system based on maintenance records or confirmed alarm information. The extracted waveform features are associated with the real labels of the corresponding leakage events. The real labels include leakage existence labels and leakage type labels, which constitute feature-label sample pairs for model training. The leakage existence label is used to indicate whether the waveform feature originates from a real leakage event, and the leakage type label is used to indicate which type of leakage event it belongs to: perforation leakage, pipe burst leakage, or interface leakage.
[0092] It is understandable that the gradient boosting decision tree algorithm is used, taking waveform features from feature-label sample pairs as input and the corresponding leakage existence label and leakage type label as supervision targets, to iteratively optimize and train the model parameters. The gradient boosting decision tree algorithm constructs a series of decision trees sequentially, with the learning objective of each new tree being to fit the residual between the prediction result of the previous tree and the true label, thereby gradually reducing prediction error. During model training, feature-label sample pairs are divided into training and validation subsets. The training subset is used to update the model parameters, and the validation subset is used to evaluate the model's predictive performance on leakage suspicion score and leakage type classification to prevent overfitting. It is understandable that training stops when the model's predictive performance on the validation subset meets the preset accuracy and recall metrics, resulting in a trained leakage inference calculation model. The model outputs a leakage suspicion score. It can be represented as an input feature vector Functions:
[0093]
[0094] in: The score represents the degree of suspected leakage. Represents the sigmoid activation function. This represents the total number of decision trees in a gradient boosting decision tree model. Representing the Decision Tree Weighting coefficients in an additive model This represents a feature vector composed of waveform features. Representing the Decision trees for feature vectors The predicted output value.
[0095] In one embodiment of the present invention, a set of inversion equations for the transient flow state of a water supply network is constructed using a topological model of the water supply network, a hydraulic model of the pipeline, and the occurrence time of suspected leakage events. The topological model defines the connection relationships between all pipes, nodes, water sources, and monitoring points in the network. The topological model can be mathematically expressed using adjacency matrices or incidence matrices from graph theory, where the rows and columns of the matrix correspond to nodes in the network, and the matrix element values represent the pipe connection relationships and water flow direction. The hydraulic model of the pipeline includes the pipe's length, diameter, material, friction coefficient, and valve state hydraulic properties. The hydraulic model stores the physical properties and hydraulic parameters of each pipe segment in the form of a structured data table, providing fundamental parameters for calculating the pipe flow resistance. (See also...) Figure 2The occurrence time of the suspected leakage event is taken as the starting time of the state inversion. The data of the standardized pressure fluctuation sequence before and after this time are taken as the known observations. The observations refer to the set of standardized pressure fluctuation sequence data points corresponding to all monitoring points where pressure sensors are deployed within the time window of the suspected leakage event.
[0096] In some embodiments, based on the mass and momentum conservation equations, a set of equations is established with the flow rates of all unknown nodes in the pipe network and the pressure drop of the pipe segments as state variables. For each node in the pipe network topology model, a node flow balance equation is established based on the mass conservation equation. The node flow balance equation states that the total flow rate flowing into the node is equal to the total flow rate flowing out of the node. For any intermediate node in the pipe network, the sum of all pipe flows flowing into that node must be equal to the sum of all pipe flows flowing out of that node. For each pipe segment in the pipe network hydraulic model, a pipe pressure drop equation is established based on the momentum conservation equation. The pipe pressure drop equation states that the pressure difference between the two ends of the pipe is equal to the sum of the pipe friction resistance and the local resistance, where the resistance is a function of the pipe flow rate, pipe length, diameter, and material friction coefficient. The pipe pressure drop equation is usually described using hydraulic calculation formulas such as the Darcy-Weisbach formula or the Hayson-Williams formula. The nodal flow balance equations established for all nodes are combined with the pipeline pressure drop equations established for all pipe segments to form a set of nonlinear equations with the flow rates of all unknown nodes and the pressure drops of all unknown pipe segments in the pipeline network as state variables. The total number of state variables is equal to the sum of the number of unknown nodes and the number of unknown pipe segments in the pipeline network.
[0097] In some embodiments, the coefficient matrix of the nonlinear equation system is jointly determined by the node-pipe connection relationships defined by the topology model and the pipe attribute parameters defined by the pipe hydraulic model. The node-pipe connection relationships determine the coefficients of the flow variable in the mass conservation equation, while the pipe attribute parameters determine the specific form and value of the drag term in the momentum conservation equation. It can be understood that substituting known observations as boundary conditions into the equation system constitutes a closed system of equations for the transient flow state inversion in a pipe network. The observations, as boundary conditions, provide known pressure values for some nodes (monitoring point locations), thereby reducing the number of unknowns in the equations and making the equation system mathematically solvable. The system of equations for the transient flow state inversion in a pipe network can be compactly represented as solving for the state vector. The form makes ,in It is a vector of state variables containing the flow rates of all unknown nodes and the pressure drop of the pipe section. It is an operator of a nonlinear system of equations composed of the mass conservation equation and the momentum conservation equation. This represents a vector of known pressure observations at the monitoring point location, extracted from a standardized pressure fluctuation sequence.
[0098] See Figure 3This is a topology model of a water supply network. This diagram visualizes the network topology and leakage node location, clearly showing the distribution of network nodes and leakage locations. Blue dots represent normally operating nodes, key monitoring points or connection points in the water supply network, connected by black pipelines to form the complete network topology. Red dots represent leakage nodes, located by a state inversion algorithm, with coordinates approximately (82, 46). The leakage node is located in the middle-right region of the network, connecting multiple pipes, and is a critical hydraulic convergence point. Leakage will significantly affect the pressure distribution in the surrounding area. The network exhibits a "multi-branch + local ring" structure. The area where the leakage node is located has a high node density and complex hydraulic interactions, consistent with the actual layout characteristics of a water supply network. Labeled samples are provided for the leakage inference model to optimize the model's accuracy in identifying leakage at different locations.
[0099] In one embodiment of the present invention, a state inversion algorithm is used to solve the transient flow state inversion equations of a pipeline network. The state inversion algorithm uses standardized pressure fluctuation sequences measured by pressure sensors as boundary conditions to inversely extrapolate the hydraulic state changes of each pipe segment within the pipeline network during the time window of a suspected leakage event. The state inversion algorithm employs the adjoint method, constructing an objective function by minimizing the difference between the network node pressure calculated by the model and the observed pressure in the standardized pressure fluctuation sequence. The adjoint method incorporates state constraints (i.e., the transient flow state inversion equations of the pipeline network) into the optimization process by introducing Lagrange multipliers (adjoint variables), which significantly improves the computational efficiency of the objective function gradient. By solving the gradient of the objective function relative to the state variables, the unknown state variables in the transient flow state inversion equations of the pipeline network are iteratively adjusted. The state variables include the flow rate at each unknown node and the pressure drop of each pipe segment. The objective function is defined as the squared L2 of the difference between the network node pressure vector calculated by the model and the observed pressure vector in the standardized pressure fluctuation sequence. The objective function value directly quantifies the overall deviation between the model simulation results and the field measured data.
[0100] In some embodiments, the gradient vector of the objective function with respect to all unknown state variables in the transient flow state inversion equations of the pipeline network is calculated. The gradient vector is a vector with the same dimension as the state variables, and each component represents the rate of change of the objective function value relative to the corresponding state variable. When calculating the gradient vector, the residual between the observed pressure and the model-calculated pressure is backpropagated using the adjoint equation to efficiently obtain the gradient information corresponding to all state variables. Based on the calculated gradient vector, a search direction that reduces the objective function is determined. The search direction is usually chosen as the negative gradient direction or a conjugate direction determined by optimization algorithms such as the conjugate gradient method. Along the search direction, the unknown state variables are updated with a preset step size to obtain a new set of state variable estimates. The step size can be determined using a line search strategy to ensure that the objective function value decreases sufficiently in each iteration. The updated state variable estimates are substituted into the transient flow state inversion equations of the pipeline network to recalculate the pressure distribution of the entire pipeline network, and then a new model-calculated pressure vector is calculated. This step requires calling the hydraulic simulation engine to solve for the pipeline pressure field that satisfies mass and momentum conservation under the new state variable estimates.
[0101] In some embodiments, the pressure vector and observed pressure vector are calculated based on the new model, and the objective function is recalculated. It is determined whether the recalculated objective function value is less than a preset convergence tolerance, or whether the number of iterations has reached a preset maximum number of iterations. The convergence tolerance is a positive number close to zero, used to determine whether the optimization process has achieved satisfactory accuracy. If the stopping condition is not met, the difference between the pressure vector and observed pressure vector is calculated based on the new model, the gradient vector of the objective function relative to the updated state variables is recalculated, and the steps of determining the search direction, updating the state variables, recalculating the pressure distribution and objective function, and determining whether the stopping condition is met are repeated. If the stopping condition is met, the iteration process is terminated, and the estimated value of the state variables obtained from the last update is taken as the final solution of the state inversion algorithm, i.e., the hydraulic state changes of each pipe segment within the pipeline network within the suspected leakage event time window. Refer to Table 1 for the iterative update process and key parameters of the state inversion algorithm.
[0102] Table 1: Example of iterative process of state inversion algorithm
[0103]
[0104] See Figure 4This is a convergence curve of the water supply network leakage detection state inversion algorithm, showing the iterative optimization process of the algorithm. The core is the trend of the objective function value changing with the number of iterations. The convergence tolerance threshold (20) is used when the objective function value is lower than this value, indicating that the algorithm has converged. The objective function iteration decline curve visually shows the gradual reduction of the deviation in the filled area. The initial objective function value is 1250.4, and the deviation between the model-calculated pressure and the measured pressure is extremely large, indicating that the algorithm is in the initial state. After 5 iterations, the objective function value drops sharply to 15.7, falling below the convergence tolerance threshold of 20 for the first time. The algorithm reaches the preset accuracy requirement, and the iteration process terminates. The convergence after 5 iterations indicates that the algorithm has high computational efficiency and stability in the water supply network leakage detection scenario.
[0105] In one embodiment of the present invention, see [reference] Figure 5 Based on the hydraulic state changes derived from the state inversion algorithm, the pipe network connection points where pressure gradients abruptly change are identified, and the specific pipe segments associated with these pressure gradient abrupt changes are located. Based on the final hydraulic state solution obtained from the state inversion algorithm, the pressure gradient distribution along the pipe network is calculated. The hydraulic state solution includes the spatial distribution of pressure at all nodes and flow rates in the pipe segments. The pressure gradient is obtained by calculating the ratio of the pressure difference between the two ends of the pipe to the pipe length. Within the pressure gradient distribution, spatial locations where the pressure gradient value exceeds the normal fluctuation range threshold are searched. The normal fluctuation range threshold is obtained by statistically analyzing the pressure gradient data of each pipe during the historical period of normal operation without leakage, and is set as the historical average pressure gradient plus a certain multiple of the historical pressure gradient standard deviation. The searched spatial locations are mapped back to the topology model of the water supply network to determine the pipe network connection points where pressure gradient abrupt changes occur. Pipe network connection points refer to nodes where pipes intersect or where pipes connect to equipment.
[0106] In some embodiments, within the topology model, all pipe segments directly connected to the network connection point are retrieved. Combining this with the direction of the pressure gradient, the specific pipe segment where leakage occurred is determined. The direction of the pressure gradient indicates the axis of the most severe pressure drop. By comparing the magnitude and direction of the pressure gradients of each connected pipe segment at the connection point, the pipe most likely to experience leakage can be inferred. Combining the leakage type classification output by the leakage inference calculation model with the specific pipe segment located by the state inversion algorithm, a multi-pipe leakage detection and location report is generated, including the leakage location, leakage type, and occurrence time. The number, start node, and end node information of the specific pipe segment located by the state inversion algorithm are used as the leakage location information. The specific pipe segment number is a unique identifier assigned to that pipe in the pipeline hydraulic model. The leakage type classification of suspected leakage events associated with the specific pipe segment, output by the leakage inference calculation model, is used as the leakage type information. The leakage type classification is the classification result output by the model after judging the waveform characteristics of abnormal pressure events. The occurrence time of the suspected leakage event is used as the leakage occurrence time information. The occurrence time is the start timestamp of the abnormal pressure event identified from the standardized pressure fluctuation sequence.
[0107] In some embodiments, leakage location information, leakage type information, and leakage occurrence time information are associated and integrated. This association is based on the fact that these information originates from the output results of different stages in the analysis process of the same suspected leakage event. Following a preset report template, the integrated information is filled in, and a pressure gradient distribution map for location analysis is attached to form the final multi-pipe leakage detection and location report. The report template predefines the presentation position and format of fields such as leakage location, leakage type, and occurrence time in a structured table or document format. The pressure gradient distribution map uses a pipeline network map as a base map, and uses line segments of different colors or widths to overlay the spatial distribution of pressure gradients at corresponding pipeline locations. The map highlights locations where the pressure gradient exceeds a threshold.
[0108] It is understandable that the identification of pressure gradient abrupt change points can be achieved by comparing the relative change in the pressure gradient value at the abrupt change point with the background pressure gradient value. The calculation formula is:
[0109]
[0110] in: This represents the relative change in the pressure gradient. This represents the abnormal pressure gradient value identified within the suspected leakage time window. This represents the long-term average pressure gradient of the pipe section during its historical normal operation. This represents the standard deviation of the pressure gradient during the historical normal operation of this pipe section. The larger the absolute value, the more significant the pressure gradient change at that point, and the more likely it is to be the location of a leak. It is understandable that a multi-pipe leak detection and location report can systematically present the spatial location, type, and occurrence time of multiple potential leak points, providing clear maintenance objectives and decision-making basis for pipeline maintenance.
[0111] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.
Claims
1. A method for detecting and locating leaks in multiple pipes of a water supply network, characterized in that, include: Collect and process pressure time-series data from selected monitoring points in the water supply network to generate standardized pressure fluctuation sequences; Based on a pre-trained leakage inference calculation model, pressure time series data of water supply network are analyzed to obtain leakage suspicion score and leakage type classification; Based on the suspected leakage score, suspected leakage events that exceed the set threshold are filtered out, and the monitoring point location and timestamp of their occurrence are obtained; Using the topology model of the water supply network, the hydraulic model of the pipeline, and the occurrence time of the suspected leakage event, a set of inversion equations for the transient flow state of the network is constructed. The state inversion algorithm is used to solve the transient flow state inversion equations of the pipeline network. This algorithm uses the standardized pressure fluctuation sequence measured by pressure sensors as boundary conditions to inversely extrapolate the hydraulic state changes of each pipe segment within the suspected leakage event time window, including: The state inversion algorithm adopts the adjoint method, which constructs the objective function by minimizing the difference between the pipeline node pressure calculated by the model and the observed pressure in the standardized pressure fluctuation sequence. By solving the gradient of the objective function with respect to the state variables, the unknown state variables in the transient flow state inversion equation set of the pipeline network are iteratively adjusted. In each iteration, the pressure distribution of the entire pipeline network is recalculated using the adjusted state variables and compared with the observed values to update the gradient direction; When the value of the objective function converges to within the preset tolerance, or when the maximum number of iterations is reached, the iteration stops. At this time, the solution of the state variable is the hydraulic state change of each pipe segment inside the pipeline network obtained by inversion. Based on the hydraulic state changes derived from the state inversion algorithm, the pipe network connection points where the pressure gradient changes abruptly are identified, and the specific pipe segments associated with the pressure gradient abrupt change are located. By combining the leakage type classification output by the leakage inference calculation model with the specific pipe segment located by the state inversion algorithm, a multi-pipe leakage detection and location report containing leakage location, leakage type and occurrence time is generated.
2. The method for detecting and locating leakage in multiple pipes of a water supply network as described in claim 1, characterized in that, The process of collecting and processing pressure time-series data from selected monitoring points in the water supply network to generate a standardized pressure fluctuation sequence includes: Pressure time-series data of selected monitoring points in the water supply network are collected, and the pressure time-series data is continuously acquired by pressure sensors deployed at key nodes of the network. The collected pressure time-series data undergoes signal preprocessing to generate a standardized pressure fluctuation sequence, specifically including: Wavelet threshold denoising is applied to the pressure time series data to separate and filter out high-frequency noise components; A polynomial fitting method was used on the denoised pressure data to extract the long-term trend term caused by changes in water use patterns, resulting in a pressure residual sequence after removing the trend term. The pressure residual sequence is resampled at fixed time intervals to form the standardized pressure fluctuation sequence; The standardized pressure fluctuation sequence is normalized to eliminate the magnitude difference between different monitoring points caused by different reference pressures.
3. The method for detecting and locating leakage in multiple pipes of a water supply network as described in claim 2, characterized in that, Based on a pre-trained leakage inference calculation model, pressure time-series data of the water supply network are analyzed to obtain leakage suspicion scores and leakage type classifications, including: Based on the standardized pressure fluctuation sequence, abnormal pressure events contained therein are identified by a pattern recognition algorithm. These abnormal pressure events are characterized by an abnormal sudden drop or continuous decline in pressure values. Extract the waveform features of each of the abnormal pressure events; The extracted waveform features are input into a pre-trained leakage inference calculation model, which outputs a leakage suspicion score and a leakage type classification, specifically including: The leakage inference calculation model is based on a gradient boosting decision tree architecture, and its input layer receives a feature vector composed of the waveform features. The hidden layer of the leakage inference calculation model splits and judges the input feature vector through a multi-level decision tree, and extracts nonlinear feature combinations related to leakage layer by layer. In the output layer of the leakage inference calculation model, one branch outputs the leakage suspicion score, which is a continuous value between zero and one. In the output layer of the leakage inference computation model, another branch outputs the classification probability of the leakage type.
4. The method for detecting and locating leakage in multiple pipes of a water supply network as described in claim 3, characterized in that, The method utilizes the topological model of the water supply network, the hydraulic model of the pipeline, and the occurrence time of the suspected leakage event to construct a set of inversion equations for the transient flow state of the network, including: The topology model defines the connection relationships between all pipes, nodes, water sources and monitoring points in the pipeline network; The pipeline hydraulic model includes the pipeline's length, diameter, material, friction coefficient, and valve status hydraulic properties. The occurrence time of the suspected leakage event is taken as the starting time of the state inversion, and the data of the standardized pressure fluctuation sequence before and after this time are taken as known observations; Based on the equations of mass conservation and momentum conservation, a set of equations is established with the flow rate of all unknown nodes in the pipeline network and the pressure drop of the pipeline section as state variables. The coefficient matrix of the set of equations is determined by the topology model and the pipeline hydraulic model. Substituting the known observations as boundary conditions into the system of equations forms a closed system of equations for the transient flow state inversion of the pipeline network.
5. The method for detecting and locating leakage in multiple pipes of a water supply network as described in claim 4, characterized in that, Based on the hydraulic state changes derived from the state inversion algorithm, the pipe network connection points where pressure gradients abruptly change are identified, and the specific pipe segments associated with these pressure gradient abrupt changes are located, including: Based on the hydraulic state solution finally obtained by the state inversion algorithm, the pressure gradient distribution along the pipe in the network is calculated. In the pressure gradient distribution, search for spatial locations where the pressure gradient value exceeds the threshold of the normal fluctuation range; The searched spatial location points are mapped back to the topology model of the water supply network to determine the network connection point where the pressure gradient change occurs. In the topology model, all pipe segments directly connected to the network connection point are retrieved, and the specific pipe segment where leakage occurs is determined by combining the direction of the pressure gradient.
6. The method for detecting and locating leakage in multiple pipes of a water supply network as described in claim 5, characterized in that, The process combines the leakage type classification output by the leakage inference calculation model with the specific pipe segment located by the state inversion algorithm to generate a multi-pipe leakage detection and location report containing leakage location, leakage type, and occurrence time, including: The specific pipe segment number, start node, and end node information located by the state inversion algorithm are used as leakage location information; The leakage type classification of the suspected leakage events associated with specific pipe sections, output by the leakage inference calculation model, is used as leakage type information. The time of occurrence of the suspected leakage event is used as the leakage occurrence time information; The leakage location information, the leakage type information, and the leakage occurrence time information are associated and integrated; According to the preset report template, the integrated information is filled in, and the pressure gradient distribution map used for location analysis is attached to form the final multi-pipe leakage detection and location report.
7. The method for detecting and locating leakage in multiple pipes of a water supply network as described in claim 6, characterized in that, The steps for constructing the leakage inference calculation model include: Acquire pressure time-series data of the water supply network under normal operating conditions and various known leakage events during historical periods to form an initial training dataset; For each known leakage event in the initial training dataset, the signal preprocessing step is performed to generate the corresponding standardized pressure fluctuation sequence. From the standardized pressure fluctuation sequence, the abnormal pressure events and their corresponding waveform features are extracted based on the labeled time windows of known leakage events; The extracted waveform features are associated with the corresponding real labels of the leakage events. The real labels include leakage existence labels and leakage type labels, which constitute feature-label sample pairs for model training. The gradient boosting decision tree algorithm is adopted, taking the waveform features in the feature-label sample pairs as input and the corresponding leakage existence label and leakage type label as supervision targets, and iteratively optimizing the model parameters during training. During model training, the feature-label sample pairs are divided into a training subset and a validation subset. The training subset is used to update the model parameters, and the validation subset is used to evaluate the model's predictive performance on leakage suspicion score and leakage type classification, so as to prevent model overfitting. When the model's prediction performance on the validation subset meets the preset accuracy and recall metrics, training stops, and the trained leakage inference computation model is obtained.
8. The method for detecting and locating leakage in multiple pipes of a water supply network as described in claim 7, characterized in that, The system of equations, based on the equations of mass and momentum conservation, is established with the flow rate at all unknown nodes in the pipeline network and the pressure drop in the pipeline section as state variables, including: For each node in the pipeline topology model, a node flow balance equation is established based on the mass conservation equation. For each section of the pipeline in the hydraulic model of the pipeline network, the pipeline pressure drop equation is established based on the momentum conservation equation; The nodal flow balance equations established for all nodes and the pipeline pressure drop equations established for all pipe segments are combined to form a set of nonlinear equations with the flow rates of all unknown nodes and the pressure drops of all unknown pipe segments in the pipeline network as state variables. The coefficient matrix of the nonlinear equation system is jointly determined by the node-pipe connection relationship defined by the topology model and the pipe attribute parameters defined by the pipe hydraulic model.
9. The method for detecting and locating leakage in multiple pipes of a water supply network as described in claim 8, characterized in that, By solving for the gradient of the objective function with respect to the state variables, the unknown state variables in the transient flow state inversion equations of the pipeline network are iteratively adjusted, including: The objective function is defined as the squared L2 of the difference between the pressure vector at the pipeline node calculated by the model and the observed pressure vector in the standardized pressure fluctuation sequence measured in reality. Calculate the gradient vector of the objective function with respect to all unknown state variables in the system of inversion equations for the transient flow state of the pipeline network; Based on the calculated gradient vector, determine the search direction that causes the objective function to decrease; Along the search direction, the unknown state variables are updated with a preset step size to obtain a new set of state variable estimates; Substitute the updated state variable estimates into the pipeline transient flow state inversion equation set, recalculate the pressure distribution of the entire pipeline, and then calculate the new model-calculated pressure vector. Based on the new model, the pressure vector and the observed pressure vector are calculated, and the value of the objective function is recalculated. Determine whether the recalculated value of the objective function is less than the preset convergence tolerance, or whether the number of iterations has reached the preset maximum number of iterations; If the stopping condition is not met, the difference between the pressure vector and the observed pressure vector is calculated based on the new model, the gradient vector of the objective function with respect to the updated state variables is recalculated, and the steps of determining the search direction, updating the state variables, recalculating the pressure distribution and objective function, and determining whether the stopping condition is met are repeated. If the stopping condition is met, the iteration process is terminated, and the estimated value of the state variable obtained from the last update is taken as the final solution of the state inversion algorithm, that is, the hydraulic state changes of each pipe segment inside the pipeline network within the time window of the suspected leakage event obtained by inversion.