A water supply network leakage monitoring method and device based on a deep learning model

Abstract

The application discloses a kind of based on deep learning model's water supply pipe network leakage monitoring method, comprising the following steps: by hydraulic model construction normal operating condition pressure monitoring data and abnormal pressure monitoring data;According to the geographical information of water supply pipe network and each pressure monitoring data constructs continuous pressure spatial distribution chart, and pressure spatial distribution chart and label are formed into data set;Predictive network is constructed;Data set is used to train the predictive network, to obtain the leakage monitoring model for monitoring leakage event;Monitoring data and geographical information are input into leakage monitoring model, to output the result whether there is leakage event in pipe network.The application also provides a kind of water supply pipe network leakage monitoring device.The method of the application can effectively improve the performance and generalization of leakage monitoring model.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of urban water supply pipe network, and particularly relates to a water supply pipe network leakage monitoring method and device based on a deep learning model. BACKGROUND

[0002] Water leakage in water distribution networks (WDNs) is an important issue in water resource management. Severe leakage incidents can cause a large amount of water and energy resources to be wasted, and can also trigger secondary disasters, such as the invasion of pipe network pollutants and pathogens, ground subsidence, and the like. However, the pipe network is usually buried underground, and has a large service area, so that small leaks are difficult to detect in a short period of time, making the identification of abnormal events complex.

[0003] Water supply pipe network burst monitoring methods are mainly divided into two categories based on hardware and software. Professional equipment such as sound sticks, noise recorders, thermal infrared, ground penetrating radar, and the like can achieve high-precision identification of leakage events, but are expensive, slow, time-consuming, and labor-intensive, greatly limiting their application potential in large-scale urban water supply pipe network leakage monitoring.

[0004] Data-driven leakage detection methods using machine learning techniques have been widely studied, but these methods are mainly good at detecting severe leaks (such as pipe bursts) or identifying abnormal events in relatively simple WDNs. In order to effectively overcome the limitations of machine learning, a new method must be developed that has the following characteristics:

[0005] (1) the ability to analyze complex features;

[0006] (2) robustness sufficient to withstand uncertain factors such as changes in water usage and monitoring noise;

[0007] (3) scalability to adapt to large detection data sets;

[0008] (4) adaptability to WDNs of different sizes and configurations;

[0009] (5) autonomous learning ability to reduce dependence on human design.

[0010] Patent document CN116951334A discloses a method for constructing a water supply pipe leakage identification model based on deep learning, which obtains Mel frequency cepstral coefficients as leakage identification features by analyzing leakage sound signals, and a deep learning model learns these features to achieve the function of leakage identification. This method can dynamically identify leakage events, but the collection of leakage signals can be disturbed by environmental noise, which can affect the clarity of the sound signals and the recognition ability of the deep learning model.

[0011] Patent document CN117722611A discloses a water supply pipeline leakage positioning method, comprising the following steps: obtaining geographic information data of a water supply pipe network, and establishing a topological structure of the water supply pipe network according to the geographic information data; simulating node parameters of the water supply pipe network by using a pre-constructed hydraulic model to obtain simulated values of the node parameters; comparing the simulated values with measured values to obtain difference values of each node parameter; evaluating the leakage rate of the water supply pipe network according to the difference values to obtain a leakage rate evaluation result; determining the water flow direction in the leakage pipe section according to the leakage rate evaluation result, and tracing the leakage position in combination with the score of the pipe section. SUMMARY

[0012] The purpose of the present application is to provide a water supply pipe network leakage monitoring method and device based on a deep learning model, which can effectively improve the performance and generalization of the leakage monitoring model.

[0013] In order to achieve the first purpose of the present application, the following technical solution is provided: a water supply pipe network leakage monitoring method based on a deep learning model, comprising the following steps:

[0014] Generate the node water demand and pressure values of the pipe network under normal working conditions by a pre-constructed hydraulic model to construct normal working condition pressure monitoring data;

[0015] Perform pipe burst testing on each pipe by using the hydraulic model to construct abnormal pressure monitoring data under leakage events;

[0016] Obtain geographic information of the water supply pipe network, and couple the geographic information with the normal working condition pressure monitoring data and the abnormal pressure monitoring data respectively, and convert the discrete data obtained by coupling into continuous pressure spatial distribution maps by using the Kriging interpolation method, label the pressure spatial distribution maps as abnormal and normal, and form a data set by combining the pressure spatial distribution maps and the labels;

[0017] Construct a prediction network based on a deep learning framework, the prediction network comprising an image generation module, a feature extraction module, a data processing module, and a prediction module, the image generation module generating corresponding pressure spatial distribution maps according to input data, the feature extraction module being used to extract image features of the pressure spatial distribution maps, the data processing module being used to perform dimension reduction processing on the extracted image features to obtain low-dimensional data features, and the prediction module being used to perform prediction according to the input low-dimensional data features to output prediction results;

[0018] Train the prediction network by using the data set to obtain a leakage monitoring model for monitoring leakage events;

[0019] Input the monitoring data and the geographic information into the leakage monitoring model to output the result of whether there is a leakage event in the pipe network.

[0020] The present application fuses geographic information with pressure monitoring data to generate a continuous pressure spatial distribution map, and then uses image features of the pressure spatial distribution map for subsequent leakage event prediction tasks.

[0021] Specifically, the construction process of the normal working condition pressure monitoring data is as follows: in view of the problem of leakage events existing in real pipe network historical data, the empirical mode decomposition method is used to obtain the historical data of the pipe network, the empirical mode decomposition method is used to extract the independent metering partition water mode in the pipe network, and the inverse Fourier transform is used to reconstruct the regional water consumption change in the region where the pipe network is located within one year, the independent metering partition water mode includes a seasonal trend for extracting in a year and a week mode for extracting; according to all node water consumption modes and node basic water demand in the reconstructed regional water consumption change, the WNTR simulation is used to obtain the pressure monitoring value under the normal working condition.

[0022] Specifically, the acquisition process of the abnormal pressure monitoring data is as follows:

[0023] All water supply pipes are regarded as potential leakage pipe sections, each potential leakage pipe is divided into N equal parts, and N different degrees of leakage events may occur in each pipe;

[0024] A water consumption node is added at the midpoint of the pipe section as a leakage hole, the leakage pipe diameter is set, the leakage pipe diameter is gradually increased to complete pipe burst, and the pressure value of each monitoring node of each leakage event is output through hydraulic model calculation.

[0025] Specifically, the calculation expression of the monitoring node pressure value is as follows: When , in the formula, is the leakage flow rate; is the flow coefficient, and the default value of the turbulent state is 0.75; is the area of the leakage hole; is the index, and the default value is 0.5; is the gravitational acceleration; is the pressure water head.

[0026] Specifically, the pipe section diameter is gradually increased in a proportional manner until the pipe section is completely burst, and the larger the pipe diameter, the greater the leakage degree increment, which is more close to the actual situation compared with setting a fixed leakage flow rate increment.

[0027] Specifically, the specific construction process of the pressure spatial distribution map is as follows:

[0028] The obtained pressure monitoring data and the geographic information of the water supply pipe network are coupled and input into the Kriging interpolation model, the non-sample positions are interpolated, and the discrete pressure monitoring data is converted into a continuous pressure spatial distribution image.

[0029] Specifically, the prediction network is built by using a VGG16 architecture, the VGG16 architecture is composed of a feature extraction layer and a classification layer, and the parameter layer is increased to 16 layers.

[0030] Specifically, the feature extraction module includes a plurality of 3*3 convolution filters.

[0031] Specifically, the prediction module uses the prediction results of three consecutive time steps to output the final prediction result, if the prediction results of three consecutive time steps are all abnormal, the final prediction result is output as abnormal, and if there is no prediction result of three consecutive time steps that is all abnormal, the final prediction result is output as normal.

[0032] In order to realize the second object of the application, the technical scheme is provided as follows: a water supply network leakage monitoring device is realized by the above-mentioned water supply network leakage monitoring method based on a deep learning model.

[0033] Compared with the prior art, the beneficial effects of the present application are:

[0034] A complete water supply network leakage monitoring model is established by using deep learning and data imaging technology, and through advanced image recognition technology, the deep features of the image are fully mined, and a precise and efficient method for city pipe explosion monitoring is provided. BRIEF DESCRIPTION OF DRAWINGS

[0035] Figure 1 The city water supply network topology provided for this embodiment;

[0036] Figure 2 The flowchart of the water supply network leakage monitoring method based on the deep learning model provided for this embodiment;

[0037] Figure 3 The pipe network pressure distribution gray scale diagram under different operating conditions provided for this embodiment;

[0038] Figure 4 The training and verification loss diagram of the VGG16 model provided for this embodiment;

[0039] Figure 5 The ROC-AUC curve of the model evaluation provided for this embodiment. DETAILED DESCRIPTION

[0040] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0041] like Figure 1 The diagram shown is a topology map of an urban water supply network in a certain area provided in this embodiment. The basic data of the network includes 2 reservoirs, 2 water tanks, 1 water pump, 782 water usage nodes, and 905 pipe segments.

[0042] To better monitor leakage events in this area, through methods such as Figure 2 The method for monitoring leakage in water supply networks based on deep learning models, as shown, involves the construction of a monitoring model. The specific process is as follows:

[0043] Generate normal operating condition pressure monitoring data: Fourier series is used to simulate the weekly and seasonal variation patterns of regional water use, and the water demand and pressure values ​​of the pipeline nodes under normal operating conditions over a period of one year are generated using the WNTR (Water Network Tool for Resilience, a Python package) hydraulic model.

[0044] Simulate different levels of leakage events in potentially leaking pipelines and collect pressure data from monitoring points: First, the diameter of each potentially leaking pipeline is evenly divided into 20 parts to simulate 20 different levels of leakage events that may occur in each pipeline. Next, the leakage diameter is set. d l The pressure is gradually increased until a complete pipe burst is simulated. These different operating conditions are input into the hydraulic model to calculate the pressure monitoring data under various leakage events. The parameter information for the leakage event simulation is shown in Table 1.

[0045] .

[0046] Wherein, *DN represents the nominal diameter of the leaking pipe. Furthermore, all water supply pipes are considered as potentially leaking pipe sections, and the diameter of each potentially leaking pipe is divided into 20 equal parts, corresponding to 20 different degrees of leakage events that may occur in each pipe.

[0047] Add a water node as a leakage hole at the midpoint of the pipe segment, and set the leakage pipe diameter , gradually increasing to complete the pipe burst.

[0048] All root pipe segments in the pipe network, a possible leakage event.

[0049] After hydraulic model calculation, the monitoring node pressure value of each leakage event is output, and its expression is as follows: the mathematical expression is: , when , in the formula, is the leakage flow rate; is the flow coefficient, and the default value of the turbulent state is 0.75; is the area of the leakage hole; is the index, and the default value is 0.5; is the gravitational acceleration ( ); is the pressure head (m).

[0050] Then the pressure monitoring data and the geographic information of the water supply network are coupled, the discrete pressure monitoring data is converted into a continuous pressure spatial distribution grayscale image by using the Kriging interpolation method, and then the pixel value is scaled to the interval of 0-255, and adjusted to a rectangular image of 224x224 pixels.

[0051] According to the average value and standard deviation of the normal image set, the image is normalized by Z-score, and the image preprocessing result is shown in Figure 3 , wherein Figure 3 (a) in Figure 3 (b) in Figure 3 , the scatter points in the grayscale image at the third moment represent all water nodes, and the pipe segment pointed by the arrow is the leakage pipe segment.

[0052] Further, the pressure spatial distribution image is generated, and the specific steps are as follows:

[0053] The obtained pressure monitoring data under the no-leakage condition and the geographic information of the water supply network are coupled, input into the Kriging interpolation model, and interpolated for non-sample positions, so as to convert the discrete pressure monitoring data into a continuous pressure spatial distribution image.

[0054] The interpolation algorithm adopted is the basic estimation formula of weighted average of sampling data. Considering that the spatial variability does not increase linearly with distance, the semi-variance function gamma ( hThe similarity between spatial data points is quantified using a method called spherical, exponential, and Gaussian semivariance models. The predictive performance of these models was compared by calculating the mean absolute error (MAE), and the Gaussian model was ultimately chosen. Its mathematical expression is as follows: ; ; ; In the formula, For sample points The estimated value, for The observed values, For the weight of the data, The interval is The number of point pairs, To ensure that the predicted values ​​are unbiased Lagrange multipliers, This is the value of the reciprocal of the variogram.

[0055] Based on the aforementioned data imaging technique, all pressure monitoring data within a single time step are input into the Kriging model to generate a grayscale image of the pressure spatial distribution. The image pixel values ​​are scaled to the range of 0-255, and the size is adjusted to a rectangular image of 224×224 pixels. The mean and standard deviation of the normal image set are calculated, and then the image set is normalized according to the Z-score, the expression of which is as follows: In the formula, For the pixel values ​​of each image, and The mean and standard deviation of all normal image pixel values.

[0056] A deep learning-based image loss monitoring model was established: Considering the imbalance of data in the real historical database, this technique randomly selected 5500 normal images and 480 images with image loss. These images were allocated in a 6:2:2 ratio for the training set, validation set, and test set. First, the training set was input into the VGG16 model for training. This model transforms high-dimensional data into a low-dimensional space through hierarchical mapping, thereby extracting image features in depth. Subsequently, the classifier automatically classified the images into two types: normal and abnormal.

[0057] like Figure 4 As shown, the model calculates the loss function by comparing the difference between the predicted labels and the true labels in the validation set, and then dynamically adjusts the model parameters. Once the model converges, the relevant weights and biases are saved and applied to the test set for image recognition. If images from three consecutive time points in the test set are marked as abnormal, a leakage event is determined to exist in the pipeline network.

[0058] To improve the performance of deep learning in identifying abnormal events, the performance of the model was compared under various hyperparameter settings, and the final determined hyperparameter information is shown in Table 2.

[0059] .

[0060] When different degrees and different locations of leakage events occur in the pipe network, the monitoring performance of the VGG16 model and other deep learning models is compared as shown in Tables 3 and Figure 5 The TPR (True Positive Rate) is the true positive rate, representing the probability of correctly identifying a burst pipe; the FPR (False Positive Rate) is the false positive rate, representing the probability of incorrectly identifying a burst pipe; the AUC (Area Under Curve) is an evaluation index of the binary classification model, and the closer the AUC is to 1, the higher the model performance.

[0061] .

[0062] Further, the establishment of the image set: all normal working condition pressure images and leakage working condition images are collected to form a complete image set. In order to simulate the characteristics of data imbalance in the real historical data set (the number of normal data is much larger than that of leakage data), a large number of normal working condition images and a small number of leakage working condition images are randomly selected from the image set, and the selected images are divided into a training set, a validation set and a test set according to a certain proportion, which are used for training and verification of the deep learning model.

[0063] Construction of the leakage monitoring model: the invention uses a convolutional neural network to realize the function of extracting deep image information, and the VGG16 architecture is composed of a feature extraction layer and a classification layer, and the parameter layer is increased to 16 layers. The core of the feature extraction layer is a plurality of 3x3 convolution filters, and the classification layer is composed of three fully connected layers, and finally the image category is input through the sigmoid function.

[0064] This model maps high-dimensional pressure images to low-dimensional space in layers, and has the ability to deeply analyze image features.

[0065] Determination of judgment criteria: single-time abnormal images are considered to significantly increase the false positive rate of the algorithm if there is a leakage event in the pipe network, and continuous-time abnormal labels can more reasonably indicate the occurrence of leakage. The invention sets that if the image labels of three consecutive time steps are all determined as abnormal, it is considered that there is a leakage in the WDN.

[0066] Evaluation of model performance: in order to comprehensively evaluate the robustness of the model, the weight and bias information with the best performance on the validation set is loaded onto the test set for calculation. Four indicators are used to evaluate the performance of the model: true positive rate (TPR), false positive rate (FPR), F1 score and receiver operating characteristic (ROC) curve and area under the curve (ROC-AUC), whose expressions are as follows: ; ; ; ; wherein TP (True Positive) represents a leak in the pipe network that is correctly detected; FN (False Negative) represents a leak in the pipe network that is not detected; TN (True Negative) represents no leak occurred and is correctly not detected as occurring; and FP (False Positive) represents a false detection of a leak, although no leak actually occurred.

[0067] The embodiment also provides a water supply network leakage monitoring device, which is realized by the water supply network leakage monitoring method based on a deep learning model provided in the above embodiment.

[0068] As known from the above, the technology fuses the geographic information and pressure monitoring data of the urban water supply network, and uses the Kriging interpolation algorithm to encode the discrete monitoring data into a continuous pressure distribution map, thereby enriching the monitoring information and providing comprehensive and reliable data support;

[0069] By developing an abnormal image recognition model based on a convolutional neural network, and inputting the pressure spatial distribution image and its derived label into the model for training, the technology enables the model to autonomously learn the deep features of the image, overcomes the limitations of traditional machine learning, such as the need for manual input of features and insufficient knowledge learning, thereby significantly improving the monitoring performance;

[0070] The automated water supply network leakage monitoring technology based on data imaging and deep learning models makes full use of advanced data processing and deep learning algorithms, effectively improves the accuracy and efficiency of leakage monitoring, and provides strong technical support for the sustainable operation of the water supply system. In addition, the technology not only has good robustness and transferability, but also shows potential for application in large-scale water supply networks, marking its important application value in the water supply industry.

[0071] In addition, the terms "upper", "lower", "inner", "outer", "front", "back" are only used for description purposes, and cannot be understood as indicating or implying relative importance. Unless otherwise specified, the relative steps, numerical expressions and numerical values of the components and steps set forth in these embodiments do not limit the scope of the present application.

[0072] Of course, the above only describes specific embodiments of the present application, and is not intended to limit the scope of the present application. Any equivalent changes or modifications made to the structure, features and principles described in the scope of the present application are included in the scope of the present application.

[0073] Finally, it should be noted that the above-described embodiments are merely specific embodiments of the present application, which are used to illustrate the technical solutions of the present application, but not to limit the same. The protection scope of the present application is not limited thereto. Although the present application has been described in detail with reference to the foregoing embodiments, it should be understood by those skilled in the art that any person skilled in the art can still modify or easily think of changes to the technical solutions recorded in the foregoing embodiments, or make equivalent replacements to some of the technical features, within the technical scope disclosed by the present application. The modifications, changes or replacements do not cause the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for monitoring water leakage in a water distribution network based on a deep learning model, characterized in that, The method comprises the following steps: generating the node water demand and pressure value of the pipe network under normal working conditions through a pre-built hydraulic model to construct normal working condition pressure monitoring data; testing each pipe through the hydraulic model to construct abnormal pressure monitoring data under leakage events; obtaining the geographic information of the water supply pipe network, coupling the geographic information with the normal working condition pressure monitoring data and the abnormal pressure monitoring data respectively, and converting the discrete data obtained by coupling into continuous pressure spatial distribution maps by using the Kriging interpolation method, and labeling the pressure spatial distribution maps and labels to form a data set; constructing a prediction network based on a deep learning framework, the prediction network comprising an image generation module, a feature extraction module, a data processing module and a prediction module, the image generation module generating corresponding pressure spatial distribution maps according to input data, the feature extraction module being used for extracting image features of the pressure spatial distribution maps, the data processing module being used for dimension reduction processing of the extracted image features to obtain low-dimensional data features, and the prediction module being used for prediction according to input low-dimensional data features to output prediction results; training the prediction network by using the data set to obtain a leakage monitoring model for monitoring leakage events; inputting monitoring data and geographic information into the leakage monitoring model to output results of whether there is a leakage event in the pipe network. 2.The water distribution network leakage monitoring method based on a deep learning model according to claim 1, wherein, The construction process of the normal working condition pressure monitoring data is as follows: In view of the problem that there are leakage events in the real pipe network historical data, the empirical mode decomposition method is used to obtain the historical data of the pipe network, the independent metering subarea water use mode in the pipe network is extracted by using the empirical mode decomposition method, and the regional water use change in the region where the pipe network is located within one year is reconstructed by inverse Fourier transform, the independent metering subarea water use mode comprising a seasonal trend for extracting in a year and a week mode; the pressure monitoring values under normal working conditions are obtained by using WNTR simulation according to all node water use modes and node basic water demand in the reconstructed regional water use change. 3.The water distribution network leakage monitoring method based on a deep learning model according to claim 1, characterized in that, The acquisition process of the abnormal pressure monitoring data is as follows: all water supply pipes are regarded as potential leakage pipes, each potential leakage pipe is divided into N equal parts, and N different degrees of leakage events may occur in each pipe; a water use node is added at the midpoint of the pipe section as a leakage hole, the leakage pipe diameter is set, the leakage pipe diameter is gradually increased to complete pipe explosion, and the monitoring node pressure values of each leakage event are output by hydraulic model calculation. 4.The water distribution network leakage monitoring method based on a deep learning model according to claim 3, characterized in that, The calculation expression of the monitoring node pressure value is as follows: When , in the formula, is the leakage flow rate; is the flow coefficient, and the default value of the turbulent state is 0.75; is the area of the leakage hole; is the index, and the default value is 0.5; is the gravitational acceleration; is the pressure water head. 5.The water distribution network leakage monitoring method based on a deep learning model according to claim 3, wherein, The pipe section diameter is increased proportionally to gradually increase the pipe section until complete pipe explosion. 6.The water distribution network leakage monitoring method based on a deep learning model according to claim 1, wherein, The specific construction process of the pressure spatial distribution map is as follows: the obtained pressure monitoring data and the geographic information of the water supply pipe network are coupled, input into the Kriging interpolation model, interpolated for non-sample positions, and the discrete pressure monitoring data is converted into continuous pressure spatial distribution images. 7.The deep learning model based water distribution network leakage monitoring method of claim 1, wherein, The prediction network is constructed by using the VGG16 architecture. 8.The water distribution network leakage monitoring method based on a deep learning model according to claim 1, wherein, The feature extraction module comprises a plurality of 3x3 convolution filters. 9.The water distribution network leakage monitoring method based on a deep learning model according to claim 1, wherein, The prediction module adopts the prediction results of three continuous time steps to output a final prediction result, and if the prediction results of three continuous time steps are all abnormal, the final prediction result is output as abnormal, and if there is no prediction result of three continuous time steps that is all abnormal, the final prediction result is output as normal.

10. A water distribution network leakage monitoring apparatus, characterised in that, The water supply network leakage monitoring method based on the deep learning model is realized by the method according to any one of claims 1-9.

Images