Research On The Optimal Placement Method Of Pressure Sensors Based On The Leakage Model Of Pipe Network

Water is a precious natural resource that we depend on for our survival and to maintain the normal operation of human production and life.In every city,there exists a complex network to distribute these water resources,which we call Water Distribution Network(WDN).However,the global water wastage is exacerbated by leakage problems due to a number of reasons,making some methods of anomaly detection and leakage location necessary.In general,anomaly detection and leakage location are collected by placing sensors in the WDN,so it is necessary to study the anomaly detection and leakage location by placing pressure sensors in the WDN.This thesis focuses on an approach to pressure sensor placement under a leakage model in a water supply distribution network.In real life,it is impractical to arbitrarily destroy pipes in urban WDNs to collect pressure data,and therefore a dataset for supervised learning cannot be formed.That is,the placement of pressure sensors in the process of solving anomaly detection and leak localization problems is performed under unsupervised conditions.In addition,the WDN itself has a network topology,and combining it with the graphical analysis approach for solving the pressure sensor placement problem has great advantages and research space.Therefore,this thesis conducts a lot of research from these two perspectives to further optimize the scheme of sensor placement in the WDN.In order to make full use of this important feature of the graph structure of WDNs to better solve the problem of optimal sensor placement for leak detection and localization techniques,this thesis proposes an improved Graph Convolution Network(IGCN)algorithm,which converts the problem of optimal sensor placement under unsupervised conditions by the algorithm converts the node selection problem into the maximum correlation problem between nodes by converting the sensor optimization placement problem under unsupervised conditions into the node classification problem under semi-supervised conditions.By using the K-Means Clustering(KMC)algorithm,the residual sum of squares of each sample node to the center node of each cluster is calculated,and the minimum residual sum of squares is selected as the pseudo-label node,and only one pseudo-label node exists in each class of sample nodes.After training the semi-supervised graph convolutional network by combining graphical analysis with network topology,the WDN is divided into several regions.Next,the most representative node in each region is selected by the maximum correlation of the nodes in that region,and that node location is used to monitor the whole water supply distribution network by placing sensors.The experimental results show that the IGCN method proposed in this thesis is 1.07 times more accurate than the fully linear Dense Net-based method,and also 1.09 times more accurate than the traditional semi-supervised method.Similarly,it also outperforms the comparison method in terms of average topological distance and mean square error index,which in turn demonstrates the effectiveness and superiority of the IGCN method proposed in this thesis.In the study of pressure sensor placement in WDNs,two important metrics are often overlooked: 1.the time required for sensors in the network to detect and locate leaks;and 2.the coverage of the entire network in terms of monitoring leak anomalies by the sensors placed in the WDN.However,these two indicators have an important impact on the sensor placement in the WDN.Therefore,an improved Non-dominated Sorting Genetic Algorithm-II(INSGA-II)algorithm is proposed to minimize two objective functions,i.e.,the probability of sensor placement detection failure and the detection time in the WDN.The results show that by optimizing the two objectives,coverage and time,the scheme of placing sensors can be optimized and thus the accuracy of leak detection and localization is improved compared to pursuing monitoring coverage or total monitoring time alone,which means that a single objective function cannot be considered alone to accomplish this task.