Water supply network pressure measuring point optimization method based on structured depth clustering

Abstract

The invention provides a water supply pipe network pressure measuring point optimization method based on structured deep clustering, and relates to the technical field of pipe network pressure monitoring, the method obtains high-dimensional feature representation of original data through an SDCNs model, removes redundant information in the original data, then fuses high-dimensional features and structural features in the data, and finally optimizes the pressure measuring points of the water supply pipe network. The SDCNs learn the structural features of the data while learning the high-dimensional features of the data, so that the SDCNs can obtain a good clustering effect on the pressure monitoring point arrangement problem, initial populations of a traditional GA algorithm and a PSO algorithm are optimized based on the good clustering effect, a better pressure monitoring point arrangement scheme is obtained, and the pressure monitoring point arrangement efficiency is improved. The SDCNs can be effectively applied to the arrangement scheme of the pressure monitoring points of the water supply pipe network, and the space positions and the number of the monitoring points are optimized, so that the leakage monitoring capability of the pipe network is improved, and the balance of economy and effectiveness of a monitoring system is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of pipe network pressure monitoring, and in particular to a method for optimizing water supply pipe network pressure measuring points based on structured deep clustering. Background Art

[0002] Leakage in the water supply pipeline network will have a serious impact on people's normal life and production, and increase the cost of water supply. Pipeline pressure gauges are an important means of monitoring leaking pipes. In the design of water supply systems, pressure monitoring points are usually arranged based on experience, and are generally selected at locations such as the node with the largest flow, the most unfavorable water supply point, the water supply boundary line, and the end of the pipeline. However, if the pressure monitoring points of the pipeline network are set based on experience, on the one hand, too many points need to be set or the monitoring range is limited. On the other hand, when the pipeline network has problems such as leakage and burst pipes, they cannot be discovered and repaired in time. In order to improve the leakage monitoring capabilities of the pipeline network, reduce the pipeline pressure test points, and increase economic benefits, it is particularly important to optimize the location and number of monitoring points.

[0003] The differences in the optimization selection of early pressure monitoring points are mainly reflected in the differences in goals and constraints. Either the different goals, such as faster detection of pipe bursts and a larger monitoring range, are used as the primary conditions for selection, or the different hydraulic characteristics of nodes, such as pipeline node pressure, node flow, and the relationship between nodes, are used as constraints. The goals and constraints are combined and used as the basis for selecting pressure monitoring points. These methods have the following disadvantages in the layout of pressure monitoring points: (1) Locality. The layout of pressure monitoring points is often concentrated in a specific area, resulting in insufficient monitoring of overall pressure changes and difficulty in achieving comprehensive coverage. (2) The data generated by a large number of monitoring points needs to be centrally processed and analyzed. The amount of data is huge, and the data processing method may be complex and easy to introduce errors. (3) Insufficient intelligence. Many monitoring systems lack intelligent features and cannot perform self-diagnosis and fault warning, making it difficult to achieve real-time and efficient monitoring. Summary of the invention

[0004] The purpose of the present invention is to use a monitoring point optimization method based on autoencoder dimensionality reduction and graph convolutional network (GCN) reasoning, by clustering the monitoring points into different clusters, to achieve comprehensive monitoring of pipeline leakage with as few monitoring points as possible, avoid redundant layout of pressure monitoring points and overlapping monitoring ranges, and improve the accuracy and comprehensiveness of monitoring results.

[0005] To achieve the above object, the present invention provides a water supply network pressure measurement point optimization method based on structured deep clustering, the method comprising: Data processing: Process the original pipe network pressure data collected by the sensor to obtain the training set and test set; Model construction, building the SDCN model based on autoencoder dimensionality reduction and GCN inference; Pressure measurement point optimization, use the SDCN model to cluster the training set, use the GA algorithm and PSO algorithm to optimize the final clustering results of the SDCN model to obtain the optimal layout plan of the pressure monitoring points.

[0006] Furthermore, the data processing specifically includes the following operations: S101, obtaining the pipe network pressure data of each pipeline leakage of all simulated nodes, performing value processing on the data value of each original pipe network pressure data within the target time period, obtaining the data points of each sensor in the target time period, and forming an original data set; S102, checking the original data set to determine whether there are missing values ​​and abnormal values; S103, using a structured deep clustering network to perform clustering processing on the original data set, generating a corresponding KNN graph, and obtaining graph structure features of the original data set; S104, standardize the original data set.

[0007] Furthermore, the original data set is standardized by mapping the data points in the original data set to between (0, 1), and the calculation formula is as follows:

[0008]

[0009] In the above formula, is the data to be normalized, that is, the data points in the original data set, is a row vector consisting of the minimum values ​​in each column of the original data set, is a row vector consisting of the maximum value in each column of the original data set, max is the maximum value of the interval to be mapped, and min is the minimum value of the interval to be mapped. is the result of standardization. is the result of normalization.

[0010] Furthermore, the SDCN model includes three parts. The first part is composed of an autoencoder, which uses a DNN model; the second part is composed of a GCN module; the third part is a dual self-supervision module, and the input of the dual self-supervision module is the data of the hidden layer of the autoencoder and the output distribution Z of the GCN module.

[0011] Furthermore, the autoencoder includes an encoder, a hidden layer and a decoder.

[0012] Furthermore, the GCN module includes three graph convolutional layers, each of which receives high-dimensional features of data output by the encoder, hidden layer and decoder respectively.

[0013] Furthermore, the dual self-supervision module designs a soft task distribution Q according to the data of the hidden layer of the autoencoder, uses t distribution as a measure of the similarity between the data representation and the cluster center vector, and optimizes the effect of unsupervised clustering. At the same time, distribution P is designed based on distribution Q, and high-confidence assignments are learned to optimize the data representation. KL divergence is used as the clustering loss, and distribution P is used to simultaneously supervise distribution Q and distribution Z, and their KL divergence and the reconstruction loss of the autoencoder are calculated as the loss of the network.

[0014] Furthermore, Softmax is added as an activation function after the three parts of the SDCN model, and Adamax optimizer is used in the forward operation of the SDCN model.

[0015] Furthermore, the pressure measurement point optimization specifically includes the following operations: S201, using the training set to train the SDCN model, and using the cluster center obtained by the training model as the initial population; S202, maximizing the number of nodes monitored by the pressure monitoring point is taken as the objective function, and the expression of the objective function is:

[0016] Among them, x is the layout location of the pressure monitoring point, y is the leakage point data, f() is the leakage detection function, and z is the number of leakage points detected in the pipeline network; S203, constructing constraint conditions, wherein the constraint conditions include pressure correlation constraint, shortest path constraint, economic cost constraint and sensitivity constraint; S204, using the GA algorithm and the POS algorithm to perform optimization training based on the objective function and the constraint conditions to obtain the optimal arrangement plan of the pressure monitoring points.

[0017] Compared with the prior art, the present invention has the following beneficial effects: The water supply network pressure measuring point optimization method based on structured deep clustering provided by the present invention obtains high-dimensional feature representation of original data through the SDCN model, removes redundant information in the original data, and then fuses the high-dimensional features and structural features in the data. The SDCN model learns the structural features of the data while learning the high-dimensional features of the data, so that the SDCN model can obtain a good clustering effect on the pressure monitoring point layout problem. Based on the good clustering effect, the initial population of the traditional GA algorithm and the PSO algorithm is optimized to obtain a better pressure monitoring point layout plan. The SDCN model can be effectively applied to the layout plan of the water supply network pressure monitoring points, optimize the spatial position and number of monitoring points, thereby improving the leakage monitoring capability of the pipeline network and achieving a balance between the economy and effectiveness of the monitoring system. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only preferred embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0019] Figure 1 It is a schematic diagram of the overall process of the water supply network pressure measurement point optimization method based on structured deep clustering provided in an embodiment of the present invention.

[0020] Figure 2 It is a schematic diagram of the structure of the autoencoder model provided by an embodiment of the present invention.

[0021] Figure 3 It is a structural schematic diagram of the graph convolutional neural network model provided by an embodiment of the present invention.

[0022] Figure 4 This is a schematic diagram of the confusion matrix of the predicted target value provided by the embodiment of the present invention. Figure 1 .

[0023] Figure 5 This is a schematic diagram of the confusion matrix of the predicted target value provided by the embodiment of the present invention. Figure 2 . DETAILED DESCRIPTION

[0024] The principles and features of the present invention are described below in conjunction with the accompanying drawings. The enumerated embodiments are only used to explain the present invention and are not used to limit the scope of the present invention.

[0025] Reference Figure 1 The embodiment of the present invention provides a water supply network pressure measurement point optimization method based on structured deep clustering, the method includes three parts: data processing, model construction, and pressure measurement point optimization: Data processing: Process the original pipe network pressure data collected by the sensor to obtain the training set and test set; Model construction, building the SDCN model based on autoencoder dimensionality reduction and GCN inference; Pressure measurement point optimization, use the SDCN model to cluster the training set, use the GA algorithm and PSO algorithm to optimize the final clustering results of the SDCN model to obtain the optimal layout plan of the pressure monitoring points.

[0026] As a possible implementation, the data processing specifically includes the following operations: S101, obtaining the pipe network pressure data of each pipeline leakage of all simulated nodes, performing value processing on the data value of each original pipe network pressure data within a target time period, obtaining the data points of each sensor in the target time period, and forming an original data set.

[0027] S102: Check the original data set to determine whether there are missing values ​​and outliers.

[0028] S103. Use a structured deep clustering network to cluster the original data set, generate a corresponding KNN graph, obtain the graph structure features of the original data set, generate a corresponding adjacency matrix, and prepare for subsequent data dimensionality reduction and clustering.

[0029] S104, standardize the original data set.

[0030] In this implementation, the original data is standardized, specifically mapping the data points in the original data set to between (0, 1). Exemplarily, the method "MinMaxScaler().fit_transform()" in the "sklearn" library can be called to map the data points of the indicator to between (0, 1), and the calculation formula is as follows:

[0031]

[0032] In the above formula, is the data to be normalized, that is, the data points in the original data set, is a row vector consisting of the minimum values ​​in each column of the original data set, is a row vector consisting of the maximum value in each column of the original data set, max is the maximum value of the interval to be mapped, and min is the minimum value of the interval to be mapped. is the result of standardization. is the result of normalization.

[0033] As a possible implementation, the SDCN model includes three parts.

[0034] The first part consists of an autoencoder. Figure 2, the autoencoder includes an encoder, a hidden layer and a decoder. It is an unsupervised method and does not require labeled data. Therefore, this implementation uses the autoencoder to obtain high-dimensional feature representation of the original data and remove redundant information in the original data. In this implementation, the autoencoder adopts a DNN (Deep Neural Network) model, the encoder and the decoder are both a Dense module, each Dense module contains three hidden fully connected layers, and uses the ReLU activation function. The ReLU activation function can introduce nonlinear transformations, which can enable the network to learn complex patterns and relationships of the data while alleviating the gradient vanishing problem. The autoencoder adjusts the weight of the entire branch through the gradient descent method to adapt to the input data, so that it can better learn the complex patterns and nonlinear relationships in the data.

[0035] For example, the number of neurons corresponding to the three fully connected layers of the encoder is 500, 500, and 2000 respectively. The number of neurons corresponding to the hidden layer is 10. The number of neurons corresponding to the three fully connected layers of the decoder is 2000, 500, and 500 respectively. The activation function of the hidden layer is Linear, and the main function of the Linear activation function is to adjust the input dimension of the linear layer.

[0036] The second part consists of a GCN (graph convolutional neural network) module. Graph convolutional neural networks are used to process graph structured data and can learn the structural features of the data. Figure 3 , the GCN module includes three graph convolution layers, each of which receives high-dimensional features of the data output by the encoder, hidden layer, and decoder. This implementation passes each layer of representation learned by the autoencoder to the corresponding graph convolution layer of the GCN module through a transfer operator, that is, each graph convolution layer receives high-dimensional features of the data output by the encoder, hidden layer, and decoder. Through the transfer operator, the GCN module improves the representation specific to the autoencoder into a high-order graph regularization constraint, and the autoencoder helps alleviate the over-smoothing problem in the GCN module.

[0037] The third part is a dual self-supervision module, which designs a soft task distribution Q based on the data of the hidden layer of the autoencoder, and uses Student's t-distribution as a measure of the similarity between the data representation and the cluster center vector to optimize the effect of unsupervised clustering. At the same time, distribution P is designed based on distribution Q, and data representation is optimized by learning high-confidence assignments. KL divergence is used as the clustering loss to improve the cohesion of the network. Distribution P is used to supervise distribution Q and distribution Z at the same time, and its KL divergence and the reconstruction loss of the autoencoder are calculated as the network loss.

[0038] The loss function is set as the weighted sum of the reconstruction loss of the autoencoder and the KL divergence of the distribution P supervised distribution Q and distribution Z. Through this mechanism, the SDCN model can directly focus two different goals, namely clustering goals and classification goals, in one loss function. In practice, after training until the maximum epoch, the SDCN model will get a stable result, and then the label can be set to the sample, and the soft assignment in the distribution Z is selected as the final clustering result. Table 1 shows a possible hyperparameter setting for the SDCN model.

[0039] Table 1 SDCN model hyperparameter settings

[0040] This implementation connects the autoencoder with the GCN module in the neural network architecture. However, the autoencoder and GCN modules are not designed for deep clustering. Basically, the autoencoder is mainly used for data representation learning, which belongs to the unsupervised learning scenario. The traditional GCN belongs to the semi-supervised learning scenario. Neither of these two methods can be directly applied to clustering problems. This implementation unifies the autoencoder and GCN modules in one framework through dual self-supervised modules, and effectively performs end-to-end clustering training on the autoencoder and GCN modules.

[0041] As another possible implementation, Softmax is added as an activation function after the three parts of SDCN, and the number of output neurons is set to 4. The Softmax function is mostly used in the output layer of multi-classification tasks, converting the raw data output by the entire model into a set of probability distributions, converting the output value of each category into a set of probability values, and making the sum of all probabilities equal to 1. The advantage of this is that the model can better explain the confidence of each category, and its mathematical formula can be expressed as:

[0042] In the above formula, represents the Softmax function, xi Represented as the first i elements, and k Expressed as the length of the vector.

[0043] At the same time, the Adamax optimizer is used in the forward operation of the SDCN model. The Adamax optimizer is a variant of the Adam optimizer. Compared with other optimizers, the Adamax optimizer uses a simpler calculation method when calculating gradient updates, which reduces the demand for memory and computing resources, improves the stability of gradient improvement, and is more applicable to the network.

[0044] As another possible implementation, the pressure measurement point optimization specifically includes the following operations: S201. Train the SDCN model using the training set, and use the clustering centers obtained from training the model as the initial population.

[0045] S202. Take the maximization of the number of nodes monitored by the pressure monitoring points as the objective function, and the expression of the objective function is:

[0046] where x is the layout position of the pressure monitoring points, y is the data of the leakage points, f() is the leakage detection function, and z is the number of detected pipeline network leakage points. Exemplarily, f() can be trained using a traditional two-layer bp network.

[0047] S203. Construct constraint conditions, and the constraint conditions include pressure correlation constraint, shortest path constraint, economic cost constraint, and sensitivity constraint.

[0048] Exemplarily, the pressure correlation constraint can be expressed as: P_pressure < g(x), where g() is a function for calculating the number of points related to the pressure at the monitoring points, and P_pressure represents the pressure correlation threshold.

[0049] The shortest path constraint can be expressed as P_distance < h(x), where h() is a function for calculating the number of nodes passed by the shortest path between the monitoring points, and P_distance represents the shortest path requirement.

[0050] The economic cost constraint can be expressed as P_cost > cost(x), where cost() is a function for calculating the number of nodes with the layout cost of the monitoring points less than the threshold, and P_cost represents the cost budget value.

[0051] The sensitivity constraint can be expressed as P_sensitivity < s(x), where s() is a function for calculating the number of nodes with the sensitivity of the monitoring points greater than the other nodes, and P_sensitivity represents the sensitivity threshold of the pressure monitoring points.

[0052] S204. Use the Genetic Algorithm (GA) and Particle Swarm Optimization (PSO) to perform optimization training based on the objective function and the constraint conditions to obtain the optimal layout plan for the pressure monitoring points.

[0053] This embodiment uses a test set to evaluate the accuracy of the SDCN model, the loss value, and the accuracy of the confusion matrix. It includes the evaluation of two target values, namely, the binary classification of leak detection prediction based on structured deep clustering network optimization - that is, the binary classification of leak detection prediction based on the method provided in this embodiment, and the binary classification of leak detection prediction based on traditional GA algorithm and PSO algorithm optimization. In the end, the accuracy of leak detection prediction based on structured deep clustering network optimization reached 92.3%; the accuracy of leak detection prediction optimized by traditional GA algorithm and PSO algorithm reached 88.3%. Figure 4 and Figure 5 The confusion matrix for the predicted target values ​​is shown.

[0054] The accuracy of model classification is one of the important criteria for evaluating model performance. A high accuracy rate means that the model correctly identifies most of the sample data, indicating that the model has good classification and generalization capabilities. The expression is:

[0055] Among them TP Indicates the number of samples correctly predicted by the model as positive. TN Indicates the number of samples that the model correctly predicts as negative classes, FP Indicates the number of samples that the model incorrectly predicts as positive. FN Indicates the number of samples that the model incorrectly predicts as negative.

[0056] Accuracy It can be predicted as the accuracy in the positive sample, and its formula is: 。

[0057] In order to illustrate the effectiveness of the method provided in this embodiment, we use bp network and CNN network as the generation of f(), and compare the GA algorithm and PSO algorithm combined with the SDCN model with the traditional GA algorithm and PSO algorithm. The comparison results are shown in Table 2. The comparison results show that compared with other models, the fault diagnosis accuracy of the SDCN model is improved under both target conditions after optimization.

[0058] Table 2 Comparison results

[0059] The method provided in this embodiment can not only improve the accuracy of monitoring water supply network leakage by optimizing the initial population of the traditional GA algorithm and the PSO algorithm, but also use fewer initial populations for iterative optimization, which can greatly reduce the time required for the convergence of the optimization objective function and ensure the accuracy of detecting water supply network leakage, so as to obtain better optimization solutions in less time.

Claims

1. A water supply network pressure measurement point optimization method based on structured deep clustering, characterized in that: The method comprises: Data processing: Process the original pipe network pressure data collected by the sensor to obtain the training set and test set; Model construction, building the SDCN model based on autoencoder dimensionality reduction and GCN inference; Pressure measurement point optimization, use the SDCN model to cluster the training set, use the final clustering results of the SDCN model to optimize the GA algorithm and PSO algorithm to obtain the optimal layout plan of the pressure monitoring points.

2. According to claim 1, a water supply network pressure measurement point optimization method based on structured deep clustering is characterized in that: The data processing specifically includes the following operations: S101, obtaining the pipe network pressure data of each pipeline leakage of all simulated nodes, performing value processing on the data value of each original pipe network pressure data within the target time period, obtaining the data points of each sensor in the target time period, and forming an original data set; S102, checking the original data set to determine whether there are missing values ​​and abnormal values; S103, using a structured deep clustering network to perform clustering processing on the original data set, generating a corresponding KNN graph, and obtaining graph structure features of the original data set; S104, standardize the original data set.

3. The water supply network pressure measurement point optimization method based on structured deep clustering according to claim 2 is characterized in that: The standardization of the original data set is specifically: mapping the data points in the original data set to between (0, 1), and the calculation formula is as follows: In the above formula, is the data to be normalized, that is, the data points in the original data set, is a row vector consisting of the minimum values ​​in each column of the original data set, is a row vector consisting of the maximum value in each column of the original data set, max is the maximum value of the interval to be mapped, and min is the minimum value of the interval to be mapped. is the result of standardization. is the result of normalization.

4. The water supply network pressure measurement point optimization method based on structured deep clustering according to claim 2 is characterized in that: The SDCN model includes three parts. The first part is composed of an autoencoder, which uses a DNN model; the second part is composed of a GCN module; the third part is a dual self-supervision module, and the input of the dual self-supervision module is the data of the hidden layer of the autoencoder and the output distribution Z of the GCN module.

5. The method for optimizing water supply network pressure measurement points based on structured deep clustering according to claim 4, characterized in that: The autoencoder includes an encoder, a hidden layer and a decoder.

6. The method for optimizing water supply network pressure measurement points based on structured deep clustering according to claim 5, characterized in that: The GCN module includes three graph convolutional layers, each of which receives high-dimensional features of data output by the encoder, hidden layer and decoder respectively.

7. The method for optimizing water supply network pressure measurement points based on structured deep clustering according to claim 5, characterized in that: The dual self-supervision module designs a soft task distribution Q according to the data of the hidden layer of the autoencoder, uses t distribution as a measure of the similarity between the data representation and the cluster center vector, and optimizes the effect of unsupervised clustering. At the same time, distribution P is designed based on distribution Q, and high confidence assignments are learned to optimize data representation. KL divergence is used as clustering loss, and distribution P is used to simultaneously supervise distribution Q and distribution Z, and their KL divergence and the reconstruction loss of the autoencoder are calculated as the loss of the network.

8. The method for optimizing water supply network pressure measurement points based on structured deep clustering according to claim 4, characterized in that: Softmax is added as an activation function after the three parts of the SDCN model, and Adamax optimizer is used in the forward operation of the SDCN model.

9. The method for optimizing water supply network pressure measurement points based on structured deep clustering according to claim 1, characterized in that: The pressure measurement point optimization specifically includes the following operations: S201, using the training set to train the SDCN model, and using the cluster center obtained by the training model as the initial population; S202, maximizing the number of nodes monitored by the pressure monitoring point is taken as the objective function, and the expression of the objective function is: Among them, x is the layout location of the pressure monitoring point, y is the leakage point data, f() is the leakage detection function, and z is the number of leakage points detected in the pipeline network; S203, constructing constraint conditions, wherein the constraint conditions include pressure correlation constraint, shortest path constraint, economic cost constraint and sensitivity constraint; S204, using the GA algorithm and the POS algorithm to perform optimization training based on the objective function and the constraint conditions to obtain the optimal arrangement plan of the pressure monitoring points.

Images