Pipe explosion event positioning method based on pipe network hydraulic state fusion data set

Abstract

The application discloses a pipe burst event positioning method based on pipe network hydraulic state fusion data set, first, collecting and classifying pipe network monitoring data, and constructing a water supply pipe network hydraulic model. Secondly, generating pipe burst simulation pressure data, and fusing the pipe burst simulation pressure data with the pipe network monitoring pressure data through pressure drop as a medium to form a pipe network hydraulic state fusion data set. On this basis, an EVO-LSTM-based pipe burst event identification and positioning model is constructed. Finally, inputting real-time monitoring data into the model can obtain pipe network working condition judgment and regional positioning result under pipe burst working condition. The application combines pipe network hydraulic characteristics and user long-term water use rules through the hydraulic state fusion data set, expands the training set samples, and improves the sample quality. Meanwhile, the EVO optimization algorithm is used to optimize the hyperparameters of the LSTM model, providing a more accurate and more practical choice for pipe burst event identification and positioning.

Description

Technical Field

[0001] This invention relates to a method for locating pipe burst events, and more particularly to a method for locating pipe burst events based on a fusion dataset of hydraulic conditions in a pipe network. Background Technology

[0002] Water supply networks are an indispensable infrastructure of modern cities. However, water supply systems have a long lifespan, and long-term operation can easily lead to pipe aging and bursts, resulting in high maintenance costs. Traditionally, burst identification and location are done manually, relying on citizen reports and manual inspections to screen and locate burst points throughout the network. However, due to the vast area of ​​water supply networks, manual screening is inefficient. With the development of computer and Internet of Things (IoT) technologies, the cost of network monitoring equipment has decreased and its accuracy has increased, making it possible to deploy monitoring points over a wide area and locate burst points by analyzing monitoring data. Among various monitoring data, pressure monitoring data, as the most readily available data within the network, has been widely used in burst location analysis.

[0003] Currently, various methods for locating pipe burst points using pressure data have emerged both domestically and internationally. These methods can be mainly divided into two categories: simulation data analysis and monitoring data analysis.

[0004] Simulation data analysis involves filtering historical data within a certain operating period and constructing a hydraulic model of the pipeline network based on this data and the network's hydraulic characteristics. This model then simulates and generates a large number of pressure data samples related to pipe bursts. Mathematical methods are used to analyze these data samples to pinpoint the exact location of the burst event. While this method provides a large number of simulated burst data samples, the limitations of the hydraulic model mean it cannot reflect the impact of fluctuations in residential water consumption on network pressure during long-term operation, resulting in poor prediction accuracy in practical applications.

[0005] The monitoring data analysis method directly uses all historical pressure data collected from pipeline monitoring points for analysis. The preprocessed pressure data is used to train mathematical methods such as machine learning or deep learning models, and then the mathematical model is used to identify and locate pipe burst events. However, since pipe burst accidents are low-probability events, the number of pipe burst accident samples in the monitoring data is relatively small. Overfitting often occurs when training the mathematical model, resulting in poor prediction performance. Summary of the Invention

[0006] In view of the above-mentioned prior art, the present invention provides a method for locating pipe burst events based on a fusion dataset of pipe network hydraulic status. This method integrates a large amount of simulated pipe burst accident data with pipe network hydraulic status characteristics with pipe network monitoring data with long-term water usage fluctuation information of users. The resulting fusion dataset of pipe network hydraulic status is used to train an artificial intelligence model, thereby generating a more accurate mathematical model and achieving more precise pipe burst identification and location.

[0007] To address the problems of low prediction accuracy and poor adaptability to changes in user water usage patterns in existing pipe burst location methods, this invention proposes a pipe burst event location method based on a fusion dataset of pipe network hydraulic status, comprising the following steps:

[0008] Step 1: Collect water supply network monitoring data, including user water consumption monitoring data, water plant pump station monitoring data, and monitoring point pressure monitoring data. Based on the daily water consumption, filter out the low, medium, and high daily network monitoring data for each quarter, and classify and organize them to obtain a network monitoring data sample set; collect water supply network GIS data; import the water supply network monitoring data and water supply network GIS data into EPANET software to build a water supply network hydraulic model;

[0009] Step 2: Generate pipe burst simulation data and, using pressure drop as a medium, overlay the pipe burst simulation data with actual monitoring data to construct a fusion dataset of pipe network hydraulic status. This includes: based on the thin-walled large-orifice outflow principle, using EPANET software to add ejector coefficients to model nodes to simulate pipe burst events; using the dynamic link library of EPANET software to traverse the pipe burst events of all nodes to generate a large batch of pipe burst event simulation data samples, and classifying them with pipe burst quarter s', water flow rate q', pipe burst time t', pipe burst degree n, and pipe burst node number j as labels; subtracting the simulated pressure data of pipe burst events from the simulated pressure data of monitoring points without ejector coefficients under normal operating conditions to obtain the corresponding pressure drop data sample set; and fusing the pipe network monitoring data sample set with the pressure drop data sample set to obtain the final fusion dataset of pipe network hydraulic status that combines user water flow time series patterns and pipe burst characteristics.

[0010] Step 3: Construct the EVO-LSTM algorithm framework and train it to obtain the EVO-LSTM-based pipe burst event identification and localization model. The EVO-LSTM algorithm framework is obtained by coupling the energy valley optimization EVO algorithm on the basis of the Long Short-Term Memory Network (LSTM) deep learning model. It includes an input layer, an LSTM layer, a fully connected layer, a softmax layer, and a classification layer. The energy valley optimization EVO algorithm is used to optimize the hyperparameters in the LSTM layer. The pipe burst identification model and the pipe burst localization model share the EVO-LSTM algorithm framework and constitute the EVO-LSTM-based pipe burst event identification and localization model.

[0011] The input to the pipe burst identification model is pressure monitoring data from monitoring points obtained in real time. The output of the model is a binary classification label for pipe burst condition and normal condition, where the pipe burst condition label is 1 and the normal condition label is 2. The model is used to determine the pipe burst status in the pipeline network. In the pipe burst location model, the softmax layer loss function is set as the cross-entropy between the predicted classification probability and the actual classification probability, as expressed below:

[0012]

[0013]

[0014] In the formula, J represents the number of pipe burst nodes in the pipe burst event; The probability of the actual classification; To predict the probability of classification; This represents the score corresponding to the i-th classification result output after the input data is processed by the LSTM layer and the fully connected layer;

[0015] The pipe burst location model is activated only when the pipe burst identification model outputs the label corresponding to the pipe burst condition. The input of the pipe burst location model is the pressure monitoring data of the monitoring points obtained from real-time monitoring, and the output of the pipe burst location model is the unique calorific value classification label of the top 50 pipe burst points with the model prediction score. According to the different requirements of the training set for pipe burst identification and pipe burst location work, the feature matrix composed of the pipeline hydraulic state fusion dataset in steps 2-3) and the monitoring point pressure monitoring dataset in the monitoring data sample set in step 1 is labeled and then jointly input into the pipe burst identification model. The feature matrix composed of the pipeline hydraulic state fusion dataset is labeled and then separately input into the pipe burst location model. This is called the pipe burst event identification and location model based on EVO-LSTM.

[0016] Step 4: Regional positioning of the pipe burst point prediction results: The pressure monitoring data of the monitoring points obtained in real time are used as the input of the pipe burst event identification and positioning model based on EVO-LSTM. According to the output of the pipe burst positioning model, the TOP50 node set for pipe burst prediction is determined. Finally, the circular positioning area is calculated based on the position coordinates of the nodes in the TOP-50 node set and their output probabilities in the softmax layer of the pipe burst positioning model. The center of the positioning area is the weighted center of the output probabilities, and the radius is the weighted average distance from each node to the center of the circle.

[0017] Compared with the prior art, the beneficial effects of the present invention are:

[0018] 1. This invention utilizes the hydraulic and topological characteristics of the pipe network to generate a large number of pipe burst data samples of different categories using only a small amount of water consumption monitoring data during periods of low, medium and high water consumption. This solves the problem of insufficient actual pipe burst data samples in traditional deep learning training, and makes the trained deep learning model more accurate in recognition.

[0019] 2. This invention employs a data fusion method, using pressure drop as a medium to combine actual monitoring data with simulated data of pipe burst events in the pipeline network. The resulting fused dataset possesses both the hydraulic state characteristics of the pipeline network and information on the long-term water usage fluctuation patterns of water users. This makes the training data of the deep learning model more closely resemble the actual operating conditions of the pipeline network, and maintains the same stable accuracy in pipeline network scenarios under different operating conditions.

[0020] 3. This invention uses the EVO-LSTM deep learning algorithm framework to identify pipe burst events, making full use of the time series characteristics of pipeline pressure data and the advantages of the LSTM model in identifying time series data features, thereby further improving the accuracy of pipe burst event identification. At the same time, the EVO intelligent optimization algorithm is used to optimize the hyperparameters of the LSTM model, reducing the impact of manually setting hyperparameters on pipe burst identification and further improving the accuracy of model prediction. Attached Figure Description

[0021] Figure 1 This is a flowchart of the pipe burst event localization method of the present invention;

[0022] Figure 2 This is a general framework diagram of the pipe burst event localization method of the present invention;

[0023] Figure 3 This is a schematic diagram of the hydraulic state fusion dataset of the present invention. Figure 1 ;

[0024] Figure 4 This is a schematic diagram of the hydraulic state fusion dataset of the present invention. Figure 2 ;

[0025] Figure 5 This is a schematic diagram of the structure of the pipe burst identification and localization model. Detailed Implementation

[0026] The design concept of this invention, a method for locating pipe burst events based on a fusion dataset of pipe network hydraulic conditions, is as follows: By establishing a fusion dataset of hydraulic conditions, the hydraulic characteristics of the pipe network and the long-term water usage patterns of users are combined, thereby expanding the number of training samples while improving the quality of sample information. Simultaneously, the EVO optimization algorithm is used to optimize the structure of the traditional LSTM model, providing a more accurate and practical option for the identification and location of pipe burst events. The main steps include: First, collecting and classifying pipe network monitoring data samples to construct a water supply network hydraulic model; second, generating pipe burst simulation data and fusing it with pipe network monitoring data using pressure drop as a medium to establish a fusion dataset of pipe network hydraulic conditions; then, constructing a deep learning model for pipe burst event identification and location based on the EVO-LSTM algorithm framework; finally, by inputting real-time pressure monitoring data from monitoring points into the model, the presence or absence of a pipe burst and the TOP-50 node set of pipe burst events can be obtained, and the area location can be performed based on the node set results.

[0027] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but the following embodiments are by no means intended to limit the present invention.

[0028] The present invention proposes a method for locating pipe burst events based on a fusion dataset of pipe network hydraulic conditions. The main process is as follows: Figure 1 As shown, the overall framework is as follows Figure 2 As shown, the specific content is as follows:

[0029] Step 1: Collect pipeline monitoring data, and classify and organize the pipeline monitoring data sample set;

[0030] Network monitoring data includes user water consumption data, water plant pump station information, and pressure data from monitoring points in the water supply network. Due to the large volume of collected network monitoring data, this method further categorizes the operation and maintenance data to more accurately describe its characteristics and facilitate data selection.

[0031] 1) Integrate various types of pipeline monitoring data into a pipeline monitoring data sample set with daily samples;

[0032] 2) To describe the seasonal variations in user water usage patterns, based on the climate characteristics of the area where the water supply network is located, the collected daily data samples are divided into four seasonal sample sets, denoted as S. s (The subscript 's' represents the corresponding season);

[0033] 3) Calculate the total daily water consumption data leaving the water plant in the seasonal sample set, and further calculate the average value of the daily total water consumption data based on statistical principles. and mean square error σ 总 ;

[0034] 4) Select As the dividing threshold between high-day, average-day, and low-day water consumption sample sets, the seasonal sample set is further divided into three sub-sample sets with low, medium, and high water consumption levels, denoted as S. sq (The subscript q represents the corresponding water consumption level).

[0035] Step 2: Import the pipeline network GIS data and pipeline network monitoring data sample set into EPANET software to build a hydraulic model of the water supply network, which includes the following steps.

[0036] 1) Collect GIS information of water supply network, import the geographical structure information of water supply network into EPANET software, and establish the topology of water supply network.

[0037] 2) Select representative daily sample data from 12 sub-sample sets representing low, medium and high water usage levels across four quarters in the pipeline monitoring data sample set and import them into the hydraulic model. Use the flow-driven method to initially establish the pipeline hydraulic model for the 12 representative days.

[0038] The specific selection principles for the representative daily sample data are as follows: Based on the segmented sample sets established in the first step, sample data with the median daily total water consumption of users in each segmented sample set are selected as the segmented sample set S. sq Representative daily sample data S' sq Import it into the hydraulic model of the water supply network as the driving data of the model.

[0039] The representative daily sample data specifically includes daily user water consumption monitoring data, water plant pump station monitoring data, and monitoring point pressure monitoring data.

[0040] 3) Select the pressure data information of the corresponding water supply network monitoring point to verify the preliminary hydraulic model and further improve the simulation accuracy of the hydraulic model.

[0041] The third step involves generating pipe burst simulation data and, using pressure drop as a medium, fusing the pipe burst simulation data with pipeline network monitoring data to establish a pipeline network hydraulic state fusion dataset. This specifically includes the following steps:

[0042] 1) Simulation of pipe burst event

[0043] After establishing a relatively accurate hydraulic model of the water supply network, the method of this invention utilizes the dynamic link library of the EPANET software to traverse the burst events of all nodes and simulates the burst events by adding jet coefficients to the model nodes. The specific node jet coefficients... The calculation method for the corresponding injection flow rate is shown in the following formula:

[0044]

[0045]

[0046]

[0047]

[0048] In the formula: C c is the orifice contraction coefficient, which can be calculated using empirical formulas in pipe rupture events; n is the ratio of the pipe cross-sectional area to the area of ​​the ruptured orifice, used to represent the degree of pipe rupture, and the value of n ranges from 0 to 1.0; C q It is the flow coefficient, which represents the reduction factor of the actual flow rate relative to the ideal flow rate under unobstructed conditions; is the drag coefficient, with a value of 0.06; A is the pipe cross-sectional area; g is the gravitational acceleration; H is the pressure head at the node when the pipe burst occurs, obtained by simulation calculation using EPANET software; It is a stress index with a value of 0.5;

[0049] The calculated nodal injector coefficient By importing the hydraulic model of the water supply network into the EPANET software, pipe burst events of different rupture degrees (n) can be simulated. The pressure data of all nodes in the network after all simulated pipe burst events are obtained, forming a pipe burst event simulation data sample set. The pressure data of the corresponding nodes of the network monitoring points are selected, and labels are made according to the conditions for generating the pressure data: pipe burst quarter (s'), water flow rate (q'), pipe burst time (t'), pipe burst degree (n), and pipe burst node number (j), forming a pipe burst event simulation pressure dataset of the monitoring points.

[0050] The simulated pressure data of pipe burst events at monitoring points under normal operating conditions (without the ejector coefficient) is subtracted from the simulated pressure data of pipe burst events at monitoring points with the ejector coefficient added, resulting in the corresponding pressure drop data sample set. The pipeline monitoring data sample set and the pressure drop data sample set are then fused to obtain the final pipeline hydraulic state fusion dataset, which combines user water usage time-series patterns and pipe burst characteristics. The construction process of the hydraulic state fusion dataset is as follows: Figure 3 and Figure 4 As shown.

[0051] 2) Classification of pipe burst incidents

[0052] To facilitate rapid retrieval of pipe burst events in subsequent data fusion work, each pipe burst event is distinguished by five dimensions of information: burst quarter s', water flow rate q', burst time t', burst severity n, and burst node number j, forming a subdivided sample set S of pipe burst events. s’q’t’nj .

[0053] Wherein, the bursting quarter s' is the quarter of the subdivided sample set corresponding to the representative daily data used by the hydraulic model to generate the simulation data; q' is the water consumption level of the subdivided sample set corresponding to the representative daily data used by the hydraulic model to generate the simulation data; bursting time t' is the time when the simulated bursting event occurs; bursting node number j is the number of the node where the bursting event occurs; and bursting degree n is the degree of pipe rupture, used to test the model's ability to identify bursting events of different degrees in the pipe network.

[0054] 3) Establish a pressure drop sample set

[0055] Subtracting the simulated pressure data of the burst pipe condition from the simulated pressure data of the normal operating condition without the injector factor yields the corresponding pressure drop data sample set LS. s’q’t’nj .

[0056] The fourth step is to construct a fusion data sample set of pipe burst events.

[0057] The monitoring data will be further subdivided into sample sets S based on the different degrees of pipe rupture. sq Subtract the pressure drop data samples LS of the pipe burst events at different levels n and pipe burst node j for each quarter s and water consumption level q at time t. s’t’nj The final fusion data sample MS of the pipe burst event, which combines the time sequence pattern of user water use and the characteristics of pipe bursts, was obtained. stnj .

[0058] The fifth step is to construct the feature matrices and corresponding labels for the training and testing sets of the burst pipe recognition model and the burst pipe localization model, respectively.

[0059] 1) Construct the input feature matrix and corresponding output label of the burst pipe identification model.

[0060] The input feature matrix comprises two parts: the first is the data feature of the pipe burst condition, which consists of a feature vector composed of the pressure values ​​at all pressure monitoring points at each moment in the pipeline hydraulic state fusion dataset; the second is the data feature of the normal condition, which consists of a feature vector composed of the pressure values ​​collected at each monitoring point in the pipeline monitoring data sample set at each moment. Output labels are set for the pipe burst condition and the normal condition, respectively, where 1 represents the pipe burst condition and 0 represents the normal condition.

[0061] 2) Construct the input feature matrix and corresponding output labels for the burst pipe localization model.

[0062] The input feature matrix is ​​derived from the fusion sample set MS of pipe burst events. stnj The output label is the classification code corresponding to the burst pipe node number j, which is composed of the feature vector of pressure values ​​at all pressure monitoring points at each time moment. The classification code uses the unique thermal code Y. j express.

[0063] One-hot encoding, also known as single-bit encoding, uses an N-bit state register to encode N states, with each state having its own independent register bit. In this invention, one-hot encoding is used to classify pipe burst events occurring at different nodes. For the pipe burst event fusion sample set MS... stnj A total of J pipe burst events occurred at a total of J nodes. These events were independently distinguished using a J-dimensional vector. For the fused sample set MS... stnj The burst pipe event occurring at node j corresponds to the one-hot encoding Y. j The j-th dimension has a value of 1, and the values ​​of the other dimensions are zero, that is: Y j = (0, 0, … ,y j , … , 0, 0) T y j = 1.

[0064] 3) Divide the input and output matrices for the training and test sets. To enable the model to adapt to the randomness of the data and identify the cyclical characteristics of user water usage, the feature vectors in the input and output matrices are randomly divided into the training and test sets in a 7:3 ratio, with each day as a unit (if there are T feature vectors at different times in a day, they are randomly allocated in units of T times).

[0065] Step 6: Construct the Energy Valley Algorithm-Optimized Long Short-Term Memory Network (EVO-LSTM) algorithm framework.

[0066] In this invention, the Energy Valley Intelligent Optimization Algorithm is used to optimize three hyperparameters in the Long Short-Term Memory Network model: the number of iterations, the initial learning rate, and the number of neurons in the LSTM hidden layer. This reduces the negative impact of manually selected hyperparameters on the classification performance of the LSTM model, thereby constructing the EVO-LSTM algorithm framework.

[0067] The EVO-LSTM algorithm framework consists of an input layer, an LSTM layer, a fully connected layer, a softmax layer, and a classification layer. The energy valley optimization algorithm is used to optimize the hyperparameters in the LSTM layer.

[0068] In the EVO-LSTM framework for burst pipe localization, the softmax layer loss function is set as the binary cross entropy (BCE) between the predicted and actual classification probabilities, as shown in the following expression:

[0069]

[0070] In the formula, J represents the number of pipe burst nodes in the pipe burst event; The probability of the actual classification; The expression for predicting the probability of classification is:

[0071]

[0072] In the formula, This represents the probability corresponding to the i-th classification result; This represents the score corresponding to the i-th classification result output after the input data is processed by the LSTM layer and the fully connected layer.

[0073] Among them, the Energy Valley Optimizer (EVO) is a novel metaheuristic algorithm proposed by Mahdi Azizi et al., which has few artificial constraints and fast convergence speed. Its design concept is based on the physical principles of stability and different particle decay modes. The specific optimization process is as follows:

[0074] 1) Perform algorithm initialization. Based on the physical principle of particle decay, this algorithm assumes the search space of the solution is the universe, and selects candidate solutions ( Assuming the search involves particles with different stability levels within the cosmic space (i.e., the search space for solutions), the specific formula is as follows:

[0075]

[0076]

[0077] In the formula, n represents the total number of particles (candidate solutions) in the universe (the search space of solutions); d is the dimension of the optimization problem. It is the initial spatial position corresponding to the j-th dimension variable of the i-th candidate solution; This represents the upper bound of the j-th dimension variable of the i-th candidate solution; represents the lower bound of the j-th dimension variable of the i-th candidate solution; rand is a random number uniformly distributed in the range [0, 1]. In the method of this invention, each dimension variable of each candidate solution particle corresponds to each hyperparameter to be optimized in the LSTM model, and the upper and lower bounds of each dimension variable of the candidate solution particles of the energy valley intelligent optimization algorithm are set based on this.

[0078] 2) Calculate the objective function. To describe the difference between neutron-rich and neutron-poor particles, the concept of the enrichment boundary (EB) is introduced. The value of the objective function corresponding to each solution particle is calculated as the neutron enrichment level (NEL) of that particle, and the enrichment boundary (EB) of the particle is calculated using the following formula:

[0079]

[0080] In the formula, Let be the neutron enrichment level (NEL) of the i-th particle. In the method of this invention, are the mean absolute errors between the output of the softmax layer of the identification and localization model and the true value, respectively, and EB is the enrichment boundary of particles in the universe.

[0081] 3) Calculate the particle's stability level (SL), the specific formula is as follows:

[0082]

[0083] In the formula, represents the stability level of the i-th particle. The higher the stability level value, the lower the stability of the particle. BS and WS represent the particles with the best and worst stability levels in the universe, respectively. Their values ​​are equal to the minimum and maximum values ​​of the loss function among the particles found so far.

[0084] In the main search loop of EVO, if the neutron enrichment level of the particle is higher than the neutron enrichment boundary ( If the value is >EB), it indicates that the particle has a high stability level (SL), and according to relevant physical principles, the particle will undergo alpha, beta, and gamma decay. The following update scheme will be adopted for this type of candidate solution:

[0085] First, a random number is generated between [0, 1] to simulate the stability bound (SB) in the universe. This is based on the stability level of the particles. Below the stability limit SB ( If the particle is considered more stable, it will undergo two generations of updates: alpha decay and gamma decay, generating two new candidate solution particles with a relatively small spatial span.

[0086] a) Alpha decay update. In this case, an alpha decay update will occur first. The program first generates a random integer Alpha Index I in the range [1, d] as the number of alpha rays generated during the decay process, and generates a random integer array Alpha Index II of dimension Alpha Index I to determine the variable dimension number of the candidate particle ray replacement. The variable values ​​in the candidate solution corresponding to the numbers in the integer array Alpha Index II are then replaced with the current best solution ( The variable values ​​of the corresponding dimension in the equation are used as new candidate solution particles generated in this update, thus completing an alpha decay update. Its mathematical expression is as follows:

[0087]

[0088] In the formula, These are candidate solution particles newly generated after alpha decay. It is the solution vector corresponding to the current candidate solution particle; It is the solution vector with the best stability, that is, the lowest loss function value NEL; It is the dimension that emits alpha rays, that is, the dimension that is updated and replaced.

[0089] b) Gamma Decay Update. After the first generation of alpha decay update, a gamma decay update will occur. The program first generates a random integer Gamma Index I in the range [1, d] to represent the number of gamma rays generated during the decay process, and generates a random integer array Gamma Index II of dimension Gamma Index I to determine the variable dimension number for ray replacement of the candidate particle. The variable value in the candidate solution corresponding to the number in the integer array Gamma Index II is replaced with the mean vector of the neighboring solution particles ( The variable values ​​of the corresponding dimension in the equation are used as new candidate solution particles generated in this update, thus completing a gamma decay update. Its mathematical expression is as follows:

[0090]

[0091] In the formula, These are candidate solution particles newly generated after gamma decay; It is the solution vector corresponding to the current candidate solution particle; It is the mean vector of a random number of solution particles near a candidate particle in Euclidean space; It is the dimension that emits gamma rays, that is, the dimension in which updates and replacements are performed.

[0092] However, if the stability level of the particles Above the stability limit SB ( If the particle has a low stability level, it will undergo two generations of different types of beta decay, causing the solution particle to perform a large-scale controlled motion towards the candidate particle with the best stability level and the particle center when it is being updated.

[0093] The mathematical expression for the first-generation beta decay process is as follows:

[0094]

[0095]

[0096] In the formula, and These are the first-generation update position and the current position of the i-th candidate solution particle in the search space, respectively; It is the position vector of the particle's center; It represents the stability level of the i-th particle; and These are two random numbers between [0, 1] that determine the amount of motion of the particle.

[0097] The mathematical expression for the second-generation beta decay process is as follows:

[0098]

[0099] In the formula, and These are the first-generation update position and the current position of the i-th candidate solution particle in the search space, respectively; It is the solution vector with the best stability, that is, the lowest loss function value NEL; It is the mean vector of a random number of solution particles near a candidate particle in Euclidean space; and These are two random numbers between [0, 1] that determine the amount of motion of the particle.

[0100] In the main search loop of EVO, if the neutron enrichment level of the particles is below the neutron enrichment boundary ( If the particle's stability is less than EB, it indicates that the particle is relatively stable. In this case, the particle tends to move towards the stable band by capturing electrons or emitting positrons. The mathematical expression for this process is as follows:

[0101]

[0102] In the formula, and These are the first-generation update position and the current position of the i-th candidate solution particle in the search space, respectively; r is a random number between [0, 1], which determines the particle's motion.

[0103] In the EVO-LSTM burst tube localization model, the objective function of the EVO intelligent optimization algorithm—the neutron enrichment level (NEL)—is set as the mean absolute error (MAE) of the test set loss function values. The formula for calculating the objective function is as follows:

[0104]

[0105] In the formula, The value of the objective function of the model trained with the j-th set of hyperparameters is represented by m; m represents the number of samples in the test set. This represents the loss function value when the model predicts the i-th test sample.

[0106] Step 7: Based on the EVO-LSTM algorithm framework, train the burst pipe recognition and localization model to generate the burst pipe event localization area.

[0107] 1) Initialize the basic parameters of each structural layer in the EVO-LSTM model. Since user water usage follows a daily cycle, the number of neurons in the input layer and fully connected layer is set to the number of time points (t) during which data is collected per day. The loss function for the softmax layer is the cross-entropy (BCE) between the predicted and actual classification probabilities. The optimizer uses Adaptive Moment Estimation (Adam), and the maximum number of iterations is set to 20.

[0108] 2) Initialize the energy valley optimization algorithm in the EVO-LSTM model. In this invention, the energy valley algorithm is used to optimize the number of iterations, initial learning rate, and number of hidden layer neurons in the LSTM model. The threshold ranges for the above three hyperparameters are set as shown in the table below. The dimension of the optimization variable is set to 3, the initial number of candidate particles for the energy valley algorithm is 20, and the maximum number of iterations for the optimization algorithm is 50.

[0109] threshold 10-50 0.01-0.15 10-500

[0110] 3) The energy valley optimization algorithm is used to optimize the burst pipe identification and localization model to obtain a pre-trained burst pipe identification and localization model.

[0111] Among them, the structure of the burst pipe identification and localization model is as follows: Figure 5 As shown, it includes two parts: an identification model and a localization model. Both the identification model and the localization model are based on the EVO-LSTM algorithm architecture. The difference is that the input of the identification model is normal working condition data and pipe bursting condition data, and the output is the model score of normal working condition (0) and pipe bursting condition (1). The classification result is the maximum value of the model score. The input of the localization model is only pipe bursting condition data, and the output is the model score of the node where the pipe bursting event occurs. The top 50 nodes with the highest model scores are used to form the TOP-50 node set of the pipe bursting event.

[0112] 4) Real-time early warning of monitoring data: Input the real-time monitoring data of a single monitoring point into the pre-trained pipe burst identification and positioning model. When the identification model identifies a pipe burst event, the positioning model starts to further locate the pipe burst node and outputs the TOP-50 node set of the pipe burst event.

[0113] 5) Generate the location area of ​​the burst pipe event based on the TOP-50 node set of burst pipe events.

[0114] The location area for the burst pipe event is a circular region. The center of this region is the weighted average coordinate obtained by multiplying the coordinates of the TOP-50 nodes by the output probabilities of their corresponding softmax layers. The radius of the region is the weighted average radius obtained by multiplying the distance from each node to the center of the region by the output probabilities of its corresponding softmax layers. The calculation formula is as follows:

[0115]

[0116]

[0117]

[0118] In the formula, The x-coordinate of the center of the region is used to locate the area. Let x be the x-coordinate of the i-th node; The output probability of the softmax layer for the i-th node; The ordinate of the center of the region is used for location; Let be the y-coordinate of the i-th node. The radius of the positioning area; The distance between the i-th node and the center of the circle;

[0119] Although the present invention has been described above in conjunction with the accompanying drawings, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many improvements and changes under the guidance of the present invention without departing from the spirit of the present invention, and these improvements and changes are all within the protection scope of the present invention.

Claims

1. A method for locating pipe burst events based on a fusion dataset of pipe network hydraulic states, characterized in that, Includes the following steps: Step 1: Collect water supply network monitoring data, including user water consumption monitoring data, water plant pump station monitoring data, and monitoring point pressure monitoring data. Based on the daily water consumption, filter out the low, medium, and high daily network monitoring data for each quarter, and classify and organize them to obtain a network monitoring data sample set; collect water supply network GIS data; import the water supply network monitoring data and water supply network GIS data into EPANET software to build a water supply network hydraulic model; Step 2: Generate pipe burst simulation data, and using pressure drop as a medium, overlay the pipe burst simulation data with actual monitoring data to construct a fusion dataset of the pipe network hydraulic state, including: 2-1) Based on the principle of thin-walled large-orifice outflow, EPANET software was used to add model node ejector coefficients to simulate pipe burst events. The dynamic link library of EPANET software was used to traverse the pipe burst events of all nodes to generate a large batch of pipe burst event simulation data samples, and the burst events were classified with the burst quarter s', water flow rate q', burst time t', burst degree n and burst node number j as labels. 2-2) Subtract the simulated pressure data of the monitoring point under normal operating conditions calculated by EPANET software without adding the injector coefficient from the simulated pressure data of the monitoring point after the pipe burst event, and calculate the pressure drop value of the monitoring point after different pipe burst events to obtain the corresponding pressure drop data sample set. 2-3) Subtract the pressure monitoring data of the monitoring points in the monitoring data sample set in step 1 from the pressure drop value of the corresponding monitoring point calculated in step 2-2) to obtain the pressure data of the burst pipe event after hydraulic state fusion, and construct a pipeline hydraulic state fusion dataset that combines user water use time sequence pattern and burst pipe characteristics. Step 3: Construct the EVO-LSTM algorithm framework and train it to obtain the EVO-LSTM-based model for identifying and locating burst pipe events. The EVO-LSTM algorithm framework is obtained by coupling the energy valley optimization EVO algorithm on the basis of the Long Short-Term Memory Network LSTM deep learning model; the EVO-LSTM-based pipe bursting event identification and localization model includes a pipe bursting identification model and a pipe bursting localization model; the pipe bursting identification model and the pipe bursting localization model share the EVO-LSTM algorithm framework. The EVO-LSTM algorithm framework includes an input layer, an LSTM layer, a fully connected layer, a softmax layer, and a classification layer, and the energy valley optimization EVO algorithm is used to optimize the hyperparameters in the LSTM layer. The input to the pipe burst identification model is the pressure monitoring data of the monitoring points obtained in real time. The output of the pipe burst identification model is a binary label for pipe burst condition and normal condition, where the pipe burst condition label is 1 and the normal condition label is 2. The pipe burst identification model is used to determine the pipe burst status in the pipeline network. In the burst pipe localization model, the loss function of the softmax layer is set as the cross-entropy between the predicted classification probability and the actual classification probability, as expressed below: In the formula, J represents the number of pipe burst nodes in the pipe burst event; The probability of the actual classification; To predict the probability of classification; This represents the score corresponding to the i-th classification result output after the input data is processed by the LSTM layer and the fully connected layer; The pipe burst location model is activated only when the pipe burst identification model outputs the label corresponding to the pipe burst condition. The input of the pipe burst location model is the pressure monitoring data of the monitoring points obtained from real-time monitoring, and the output of the pipe burst location model is the unique calorific value classification label of the top 50 pipe burst points with the model prediction score. According to the different requirements of the training set for pipe burst identification and pipe burst location work, the feature matrix composed of the pipeline hydraulic state fusion dataset in steps 2-3) and the monitoring point pressure monitoring dataset in the monitoring data sample set in step 1 is labeled and then jointly input into the pipe burst identification model. The feature matrix composed of the pipeline hydraulic state fusion dataset is labeled and then separately input into the pipe burst location model. This is called the pipe burst event identification and location model based on EVO-LSTM. Step 4: Locate the burst pipe location based on the predicted burst pipe location results: The pressure monitoring data of the monitoring points obtained in real time is used as the input of the pipe bursting event identification and localization model based on EVO-LSTM. According to the output of the pipe bursting localization model, the TOP50 node set for pipe bursting prediction is determined. Finally, the circular localization area is calculated based on the position coordinates of the nodes in the TOP-50 node set and their output probabilities in the softmax layer of the pipe bursting localization model. The center of the localization area is the weighted center of the output probabilities, and the radius is the weighted average distance from each node to the center of the circle.

2. The method for locating a burst pipe event according to claim 1, characterized in that, In step 1, the low, medium, and high daily water consumption data for each quarter are selected based on the daily water consumption. After classification and organization, a sample set of pipeline monitoring data is obtained, which includes: The pipeline monitoring data sample set is based on daily data. According to the climate characteristics of the area where the water supply network is located, the collected daily data samples are divided into four seasonal sample sets, denoted as Ss, where the subscript s represents the corresponding season (s=1,2,3,4). The total daily water consumption data leaving the water plant is calculated from the seasonal sample sets, and the average daily total water consumption data is calculated based on statistical principles. and mean square error σ 总 ; Select As the dividing threshold between high-day, average-day, and low-day water consumption sample sets, the seasonal sample set is divided into three sub-sample sets of low, medium, and high water consumption levels, denoted as S. sq The subscript q represents the corresponding water consumption level, where q = low, medium, and high.

3. The method for locating a burst pipe event according to claim 2, characterized in that, In step 1, the water supply network monitoring data and water supply network GIS data are imported into the EPANET software to construct a water supply network hydraulic model. The basic data used is a total of 12 subdivided sample sets S representing low, medium, and high water usage levels across four seasons. sq The representative daily sample data for each of the subdivided sample sets S; the representative daily sample data for each subdivided sample set S sq The sample data is the median daily total water consumption of the users; the representative daily sample data includes the representative daily user water consumption data, water plant pump station output water volume and water pressure information.

4. The method for locating a burst pipe event according to claim 1, characterized in that, In step 2-1), the node injector coefficient The calculation method for the corresponding injection flow rate Q is as follows: In the formula: C c is the orifice contraction coefficient, which can be calculated using empirical formulas in pipe rupture events; n is the ratio of the pipe cross-sectional area to the area of ​​the ruptured orifice, used to represent the degree of pipe rupture, and the value of n ranges from 0 to 1.0; C q It is the flow coefficient, which represents the reduction factor of the actual flow rate relative to the ideal flow rate under unobstructed conditions; is the drag coefficient, with a value of 0.06; A is the pipe cross-sectional area; g is the gravitational acceleration; H is the pressure head at the node when the pipe burst occurs, obtained by simulation calculation using EPANET software; It is a stress index with a value of 0.5; The calculated nodal injector coefficient By importing the hydraulic model of the water supply network into the EPANET software, pipe burst events of different rupture degrees (n) can be simulated. This yields pressure data for all nodes in the network after each simulated pipe burst event, forming a sample set of pipe burst event simulation data. Pressure data for nodes corresponding to network monitoring points are then selected, and labels are created based on the conditions for generating the pressure data: pipe burst quarter (s'), water flow rate (q'), pipe burst time (t'), pipe burst degree (n), and pipe burst node number (j). This forms a simulated pressure dataset of pipe burst events for the monitoring points.

5. The method for locating a burst pipe event according to claim 1, characterized in that, In step 4, the pipe burst event location area is a circular area. The center of this area is the weighted average coordinate obtained by summing the coordinates of the TOP-50 nodes multiplied by their output probabilities in the softmax layer of the pipe burst event location model. The radius of the area is the weighted average radius obtained by multiplying the distance from each node to the center of the area by their output probabilities in the softmax layer of the pipe burst event location model. The calculation formula is as follows: In the formula, The x-coordinate of the center of the region is used to locate the area. Let x be the x-coordinate of the i-th node; The output probability of the softmax layer for the i-th node; The ordinate of the center of the region is used for location; Let be the y-coordinate of the i-th node. The radius of the positioning area; Let be the distance between the i-th node and the center of the circle.

Images