Detection And Source Identification Of Contamination In Water Distribution Networks Based On Online Sensor Data

Water distribution network(WDN)is one of the most critical infrastructures of our city.Any contaminants existing in the WDN will spread quickly in the pipes and may seriously endanger the health of many residents.Therefore,a reliable water contamination monitoring system is necessary for a WDN.To this end,this dissertation focuses on investigating the problems of contamination detection and contamination source identification in WDNs based on the online sensor data collected from WDNs.Accordingly,the dissertation includes the following two parts:First,it investigates the problem of contamination detection,i.e.,how to determine if there is any contamination event occurring at a specific monitoring station in a WDN based on the sensor data collected from that station.In this dissertation,the contamination detection problem is modelled as a binary classification problem,i.e.,whether a contamination event occurs or not.However,since contamination events rarely occur in real life WDNs,the data set used for training a classifier is often class-imbalanced.To deal with this issue,we propose to adopt a data balancing method by which the two categories of data(with and without contamination)are sampled at a fixed ratio for model training.This method can ensure the relative balance of the two categories of data,thereby avoiding the bias of the classifier.In addition,considering the importance of temporal information for contamination detection,a sequence-to-point learning method is used to extract the temporal features from the original sensor data.Combining these two methods,a LSTM classifier is constructed and trained for detecting contamination in WDNs.By carrying out a set of experiments using a public data set,we show that the proposed contamination detection method outperforms the state-of-art detection methods by 28% in terms of F1 score.Second,it investigates the problem of contamination source identification(CSI),i.e.,how to identify the source node,starting time,duration and strength of a contamination event based on the sensor data collected from all the monitoring stations of a WDN.With the help of the simulation tool EPANET,the occurrence of various contamination events in a WDN can be simulated,and the problem of CSI is modelled as an optimisation problem.In order to solve this optimisation problem,a two-stage CSI framework based on deep learning and evolutionary algorithms is proposed.In the first stage,a deep learning model is used to locate the contamination source node,returning the top K candidate nodes in terms of their probability.In the second stage,by taking the candidate nodes obtained from the first stage as the prior knowledge,an evolutionary algorithm is used to evolve two populations of individuals in parallel to speed up the searching process of determining all the contamination source information including contamination source node,starting time,duration and strength.By carrying out a set of experiments,the effectiveness of the proposed framework is evaluated.