ORE Open Research Exeter

TITLE
Literature review of data analytics for leak detection in water distribution networks: A focus on pressure
and flow smart sensors

AUTHORS
Wan, X; Kuhanestani, PK; Farmani, R; et al.

JOURNAL
Journal of Water Resources Planning and Management

DEPOSITED IN ORE
10 June 2022

This version available at

[Link]

COPYRIGHT AND REUSE

Open Research Exeter makes this work available in accordance with publisher policies.

A NOTE ON VERSIONS
The version presented here may differ from the published version. If citing, you are advised to consult the published version for pagination, volume/issue and date of
publication
Manuscript Click here to access/download;Manuscript;literature review -
[Link]

1 Literature Review of Data Analytics for Leak Detection in Water

2 Distribution Networks: A Focus on Pressure and Flow Smart Sensors

3 Xi Wan1, Parisa Khorsandi Kuhanestani2, Raziyeh Farmani3, Edward Keedwell4

1
4 Ph.D. Student, Centre for Water Systems, College of Engineering, Mathematics and Physical Science,

5 University of Exeter, Harrison Building, North Park Rd., Exeter, Devon, EX4 4QF, United Kingdom

6 (corresponding author). Email: xw355@[Link]

2
7 Ph.D. Student, Department of Marine and Fluvial Systems, University of Twente, 7522 NB Enschede,

8 Netherlands. Email: [Link]@[Link]

3
9 Centre for Water Systems, College of Engineering, Mathematics and Physical Science, University of Exeter,

10 Harrison Building, North Park Rd., Exeter, Devon, EX4 4QF, United Kingdom. Email:

11 [Link]@[Link]
4
12 College of Engineering, Mathematics and Physical Science, University of Exeter, Exeter EX4 4QF, United

13 Kingdom. Email: [Link]@[Link]

14 ABSTRACT

15 Leakage detection is one of the important aspects of water distribution management. Water

16 companies are exploring alternative approaches to detect leaks in a timely manner with high accuracy

17 to reduce water losses and minimise environmental and economic consequences. In this article, a

18 literature review is presented to develop a step-by-step analytic framework for the leakage detection

19 process based on flow and pressure data collected from water distribution networks. The main steps of

20 the data analytic for leakage detection are: setting up the goals, data collection, preparing the gathered

21 data, analysing the prepared data, and method evaluation. The issues of concern for each step of the

22 proposed leakage detection framework are analysed and discussed. The smart sensor-based leakage

23 detection methods can be categorised as data-driven methods and model-based methods. Data-driven

24 methods can be further categorised as statistical process control-based methods, prediction-

25 classification methods, and clustering methods. Hydraulic model-based methods can be further

26 categorised as calibration-based methods, sensitivity analysis, and classifier-based methods. The

27 advantages and disadvantages of each method are discussed, and suggestions for future research are

1
28 provided. This review represents a new perspective on the subject from five aspects: 1) most of the

29 leakage detection methods are focused on burst detection, and different types of leakages should be

30 considered in future research; 2) it is important to consider data uncertainties, and more robust real-time

31 leakage detection methods should be developed; 3) it is important to consider hydraulic model

32 uncertainties; 4) unrealistic assumptions should be addressed in future research; 5) spatial relations

33 between sensors could provide more information and should be considered.

34 INTRODUCTION

35 Leakage is the loss of water from the supply network through uncontrolled actions. In addition to

36 water loss, there are other negative consequences caused by leakage (Colombo et al. 2009; Farah and

37 Shahrour 2017), such as: 1) potential risks to public health (Romano et al. 2011) due to the entrance of

38 contaminants from the environment into the pipes with negative pressure and changes in water quality

39 (Xu et al. 2014), 2) environmental issues due to energy used by pumps to deliver water to compensate

40 for the pressure drop, and chemicals used to treat the water in the treatment plant, and 3) leaked water

41 that ends up in surface water with a potential negative impact on living organisms due to high chlorine

42 concentration. All of these lead to great socio-economic losses (Colombo et al. 2009; Farah and

43 Shahrour 2017; Romano et al. 2011; Wu et al. 2010a), and thus, timely detection and localisation of

44 leakage events in Water Distribution Networks (WDNs) have received considerable attention for more

45 than two decades (Zaman et al. 2020).

46 Leakage detection and localisation are the processes to identify leakage in the WDN and specify the

47 leakage location. Timely detection of leaks could decrease the amount of water loss from the system. It

48 can also have other benefits (Bohorquez et al. 2020; Puust et al. 2010). For example, it can (i) reduce

49 environmental impacts by reducing water losses, (ii) allow planned interruption of supply and therefore

50 reduce the impact on customers and (iii) reduce financial costs by reducing the level of pumping and

51 financial losses linked to the amount of water lost. After detecting a leak, the location of the leakage

52 should be determined, so that repair can take place. The determination of leakage location has three

53 different phases. The first phase is to “localise”, i.e., limit the location of a leak to a specific district.

2
54 The second phase is to “locate” the pipes in a district area where leaks occur. The third phase is to

55 “pinpoint”, i.e., limit the leakage location to an area with a small radius of 2-3 feet (Qahtani et al. 2020).

56 Almost all leakage detection methods can be broadly categorised into hardware-based methods and

57 software-based methods (Ismail et al. 2019). Hardware-based methods, also called direct methods

58 (Zaman et al. 2020) or passive methods (Chan et al. 2018), usually rely on hardware devices to detect

59 leak events. Based on the principles that apply to the hardware devices, hardware-based methods can

60 be further divided into acoustic techniques and non-acoustic techniques (see Fig. 1). The detailed

61 information about the hardware-based methods can be found in Ismail et al. (2019) and Chan et al.

62 (2018). Although the accuracy of the hardware-based methods is increasingly high, they are costly,

63 time-consuming and labour-intensive, as expensive equipment and professional staff are needed.

64 Furthermore, the results could be influenced by pipe materials (e.g. acoustic methods), soil types and

65 conditions (e.g. infrared thermography, ground penetrating radar), and other factors depending on the

66 equipment. Therefore, hardware-based methods are mostly used in the third phase of leakage detection

67 and localisation, i.e. “pinpoint”.

68 Software-based methods, also called indirect methods (Zaman et al. 2020) or active methods (Chan

69 et al. 2018), could detect leaks by inference from internal pipeline parameters (such as pipe flow and

70 pressure data) rather than detect leak-related information (e.g. leak noise, infrared radiation) directly.

71 Software-based methods use computer software systems to monitor hydraulic parameters to detect

72 possible leaks continuously. Unlike hardware-based methods that try to pinpoint the leakage location

73 accurately, software-based methods aim to detect leakage and limit the area to a specific district. Based

74 on the hydraulic state of the pipeline system, the software-based methods can be classified as transient-

75 state methods and non-transient methods.

76 Colombo et al. (2009) and Abdulshaheed et al. (2017) provided a comprehensive review of the

77 transient-state leakage detection method. Transient analysis is based on the idea that any changes (e.g.

78 blockage, leakage) in the pipe’s physical structure will alter the flow and pressure response. To

79 adequately capture the transients at all time scales, the system requires many measurement points with

80 high sampling frequency, which results in a costly, labour-intensive process. Moreover, this technique

81 often relies on complex transient simulation models and is mainly applied to a single pipeline to predict

3
 82 its features (Bohorquez et al. 2020; Keramat et al. 2019). This type of analysis is computationally

83 expensive and unsuitable for real-time monitoring of large urban areas.

84 Non-transient methods, which is the focus of the present study, can be further classified as hydraulic

85 model-based methods and data-driven methods according to whether a hydraulic model is used.

86 Compared with transient-state methods, non-transient methods could use monitoring data with a much

87 lower sampling rate (e.g. 5 minutes, 15 minutes), which is readily available. Non-transient methods

88 provide a promising solution for the long-term monitoring of large-scale WDNs. When a leak occurs,

89 it will change the hydraulic behaviour in the WDS, and flow and pressure readings also will change.

90 The flow will increase by additional demand, which results in larger head loss, and leads to different

91 pressures within the network. Leakage detection is based on the difference between the predicted

92 hydraulic parameter in the absence of leakage provided by hydraulic model or data-driven methods and

93 the field observations collected by sensors.

94 Therefore, real-time monitoring of changes in water distribution networks to detect leaks is one of

95 the most promising methods. The monitoring is often performed by installing pressure sensors in

96 different parts of the system or installing flow sensors in the transmission mains. With the rapid

97 development of the internet of things and big data technologies in recent years, smart and intelligent

98 water systems could be more connected and operated with more data in real-time to achieve maximum

99 efficiency and effectiveness. The sensors create big data, and by applying appropriate data analytics to

100 them, valuable information could be obtained and make the detection and localisation of leakages

101 possible. However, the focus on big data is relatively new in the water industry, and most of the decision

102 making is done either independently or with limited use of available data. Therefore, this paper focuses

103 on smart sensor-based leakage detection systems that use time-series data from pressure and flow

104 sensors to detect leakage in distribution networks.

105 The literature review on existing methods allows researchers to form reasoned, logical and

106 confirmed arguments (Denyer and Tranfield 2006). Several literature review articles have been

107 published on leakage management with the main focus on detection. Puust et al. (2010) presented a

108 review of leakage management methods and classified these methods into three groups: leakage

109 assessment methods, leakage detection methods and leakage control models. The authors concluded

4
110 that future works should focus on the real-time models for pipe networks. Gupta and Kulat (2018)

111 carried out a similar review to Puust et al. (2010) and highlighted that more effort is needed in online

112 monitoring and online leakage detection. El-Zahab et al. (2016) reviewed leakage detection methods

113 focusing on: 1) classification of leak detection phases (identification, localisation, and pinpointing

114 leaks), 2) sensor installation type (static and dynamic leak detection systems). Wu and Liu (2017)

115 reviewed data-driven methods using data from Supervisory Control and Data Acquisition (SCADA)

116 systems but only for burst detection. They categorised the methodologies into three groups:

117 classification method, prediction method, and statistical method. Hu et al. (2021a) reviewed model-

118 based and data-driven approaches for leakage detection and location from the aspect of methodology.

119 Hu et al. (2021a) provide a good overview of methodologies that have been developed for leakage

120 detection, but in this paper, a more comprehensive step-by-step analysis of the process for leakage

121 detection is provided, including the data pre-processing techniques, the types of case study, the size of

122 leaks that are possible to be detected, etc. In order to thoroughly discuss those topics, this paper provide

123 a step-by-step analytic framework for real-time leakage detection process based on big data gathered

124 from pressure and flow sensors, from the perspective of water resources planning and management.

125 The aim of this review is to clarify the state of knowledge, identify research gaps, and form a

126 consensus on the subject. A five-step framework has been developed to analyse and compare a suite of

127 leakage detection and localisation methods for low-frequency (compared with data for transient analysis)

128 pressure and flow data measured by in situ sensors (Fig. 2). It is no doubt that goal identification is the

129 prerequisite of data analysis. After installing sensors in the distribution network, the first step is to

130 collect data for the following analysis. Identifying the characteristics of the collected data is a crucial

131 stage for data analytics-based studies, and suitable data analytic methods could be chosen accordingly.

132 By applying the appropriate methodology, the information will gain more value and be used in decision

133 making. Based on the data collected from the SCADA system and the analysis of the characteristics of

134 collected data, different methods have been applied to detect and localise leakage events. Each step will

135 be analysed step by step in the following research.

136 GOAL IDENTIFICATION (STEP 0)

5
137 Identifying the needs and goals of the end-user is the first step in the leakage detection and

138 localisation framework. Setting goals has a direct impact on the selection of the most suitable detection

139 methods. In most proposed methods, leakage identification and localisation are the two goals of leakage

140 detection based on data analytics. Tables 1 and 2 provide the summary of steps 0 and 1 of data-driven

141 and hydraulic model based leakage detection methods, respectively. Data-driven methods model the

142 system behaviour based on historical data mining, and model-based methods use a well-calibrated

143 hydraulic model to represent the current state of a network. Therefore, for leakage identification,

144 hydraulic model-based methods are preferred when the amount of historical data is limited, and data-

145 driven methods could be more efficient and provide more accurate predictions when a long-term

146 monitoring dataset is available. For leakage localisation, model-based methods are more preferred since

147 the hydraulic model can provide more topological information of the network.

148 Leakage identification usually ends with binary results (i.e. alarm on or off) that represent whether

149 or not a leakage is happening in the system. In order to provide more information for the operator,

150 Mounce et al. (2007, 2010) have provided fuzzy values and probability values between 0 to 1 to

151 represent the likelihood of a leakage event. Ye and Fenner (2011, 2014) defined the burst size as the

152 difference between the predicted value and the observed value. The determined burst size may not be

153 very accurate due to prediction errors, measurement errors, unpredictable random consumers’

154 behaviour, etc. The results of leakage localisation are presented as a leak map to compare the predicted

155 leak area and the actual leak location. Visualisation of the results helps the decision-makers to gain

156 knowledge and make informed decisions.

157 DATASET GENERATION (STEP 1)

158 One of the essential stages of data analytics-based studies is data collection and data characteristics.

159 Data specifications should be based on the needs of the study and the goals that have been set. Most of

160 the studies have applied their method to a sample case to validate or explain the presented methodology

161 of leakage detection. Those algorithms are usually tested on three kinds of datasets:

6
162  A synthetic dataset generated by a hydraulic model (such as EPANET, WNTR etc.) that usually

163 made simplifications to the network’s condition and customers’ behaviour (such as following

164 a very regular pattern or assuming that consumers’ behaviour are known in advance);

165  Engineered tests operated in real water distribution systems by opening fire hydrants that

166 usually simulated burst events with sudden water loss and short duration.

167  Historical monitoring data that contain real leakage events.

168 The artificially synthesised dataset allows researchers more flexibility to adjust parameters and

169 model different types of leakage events. However, it is difficult to model the high uncertainties in a

170 real-life network. The simplifications of simulations can generate a clean and easy-to-learn dataset that

171 is beneficial for evaluating the tested algorithms and leaves a question of the suitability of the algorithms

172 for real-life networks. Therefore, these methods should be validated in engineered test datasets or

173 historical datasets. As shown in Table 1, engineered tests that can introduce artificial leakage to the real

174 system are widely-used when researchers evaluate their methods. It should be noted that the engineered

175 tests used in the current studies only simulated burst events but not incipient leakage. However, incipient

176 leakage can cause more water loss than bursts due to a longer awareness time. Furthermore, incipient

177 leakages may last for days, or even weeks, unlike burst events that lasted for a few hours simulated in

178 the engineered tests. Therefore, it is crucial to generate an early warning for incipient leakage before it

179 reaches its maximum level. Historical monitoring datasets that contain real leakage events can reflect

180 the efficacy of detection methods in real-life scenarios. However, it is hard to know the exact starting

181 time of leakages in real life, and the information can only be inferred from the maintenance work or

182 customer contacts, which brings difficulty for method evaluation. Within this context, the Battle of the

183 Leakage Detection and Isolation Methods (BattleDIM) (Vrachminis et al. 2020) provides a hydraulic

184 model called L-Town that contains two years of real-life demand data. The datasets of BattleDIM

185 provide two years of monitoring dataset, including flow data, pressure data, and demand data. Different

186 types of leakages are modelled in the system, including background leakages, gradual leakages, and

187 bursts. These datasets could be a good option to evaluate the performance of leakage detection and

188 localisation methods.

7
189 The commonly used data are flow and pressure monitoring data, which are critical hydraulic

190 parameters that change with any alteration in the distribution system. In some cases, the flow values are

191 reported as the average value, while pressure values are reported instantaneously (Mounce et al. 2012).

192 Therefore, the data shows smoothed flow values and missing some changes in flow between time

193 intervals (Farley et al. 2013). Compared with flow time series, pressure time series have more variation

194 in their profile than flow time series (Romano et al. 2011). Furthermore, flow instruments are usually

195 installed at inlets, and they are sensitive to downstream changes, while pressure values are affected by

196 head loss and pressure changes upstream and around the instrument position (Geiger 2005). Therefore,

197 in the experiment conducted by Ye and Fenner (2011), pressure-based detection seems less sensitive to

198 a burst event than flow-based detection, especially when a pressure sensor is remote from the burst

199 location. It can also be observed from Tables 1 and 2 that most of the burst identification methods

200 preferred to use flow data. However, the installation of flow sensors is more expensive than pressure

201 sensors (Romano et al. 2011). Hence, usually, there are fewer flow data available than pressure data. It

202 has been suggested that the pressure data can be used to provide additional information as a way of

203 confirming the flow-based detection results (Ye and Fenner 2011). Therefore, pressure data are

204 preferred when the goal is leakage localisation.

205 DATA PREPARATION (STEP 2)

206 Raw monitoring data may contain lots of noises, missing data, or data from faulty loggers. Therefore,

207 it is necessary to perform data pre-processing before data is analysed by data analytics. Furthermore, it

208 is difficult to construct a proper function to model the pattern of the raw monitoring data because it is

209 highly nonlinear. In addition, the variation over different weekdays makes the flow and the pressure

210 pattern more complicated to model. While some papers have tried to model water usage patterns directly,

211 most papers adopted pre-processing techniques before leakage detection. Moreover, different methods

212 have different requirements that need to be prepared before analysing. Tables 3 and 4 summarise steps

213 2-4 of each leakage detection method.

214 Data-Driven Methods

215 As shown in Table 3, the most frequently used data preparation procedures for data-driven methods are:

8
216 1. Data correction: Data collected by the sensors is a type of big data, and due to its real-time

217 nature, sometimes incomplete/incorrect data may exist in the time series, due to missing data,

218 data from faulty loggers, erroneous timestamps etc. To ensure a continuous data stream, in some

219 studies, the missing data are replaced by an alternative value that is calculated through a

220 statistical process such as a filter interpolation (Mounce et al. 2002; Mounce and Machell 2006;

221 Romano et al. 2014). Furthermore, statistical tests could be applied to the time series to ensure

222 that an adequate amount of good quality data is available for the analysis (Romano et al. 2014).

223 2. Data de-noising: Recorded pressure and flow data are usually accompanied by noise. The

224 presence of this noise may cause some small leaks to be undetected or cause false alarms.

225 Therefore, in some studies, this noise is removed in the pre-processing stage. Misiunas (2006)

226 used an adaptive recursive least squares filter, and Romano et al. (2014) used discrete wavelet

227 transform to remove noise from data.

228 3. Data selection: In most studies, a range of normal data (data without leakage events) is needed

229 to train machine learning models or used as a reference library so that the normal behaviour of

230 the distribution network can be accurately represented. However, data collected from real

231 WDNs usually contains both leak and non-leak events. Therefore, data selection is needed to

232 ensure the performance of the detection method. For example, Palau et al. (2012) used an

233 iterative procedure to eliminate outliers during Principle Component Analysis (PCA) model

234 construction. Wu et al. (2020) used an abnormal subsequence searching (ASS) algorithm to

235 search and remove the abnormal subsequences in the library.

236 4. Data reformatting: Different techniques require different formats for the data, such as

237 normalisation, label information assigning, time-series restructure etc. When the analytic

238 method is sensitive to the numerical ranges of the variables, such as PCA (Palau et al. 2012),

239 mean centring and scaling processes are needed to normalise the data to ensure that the data

240 falls within the same range. Label information such as time of day, day of the week could

241 provide more information to the machine learning model and improve the prediction accuracy.

242 Furthermore, given a sequence of numbers for a time series dataset, data needs to be pre-

243 processed (e.g. a tapped delay line format) to prepare for neural network presentation.

9
244 Hydraulic Model-Based Methods

245 As shown in Table 4, the most frequently used data preparation procedures for hydraulic model-based

246 methods are:

247 1. Model calibration: For hydraulic model-based methods, hydraulic model calibration is

248 unavoidable for hydraulic model-based methods. The aim of the model calibration is to develop

249 the best values for the unknown model parameters, so that the hydraulic model could reasonably

250 represent the performance of the WDN. A poorly calibrated model could result in significant

251 errors in leak detection. A reliable hydraulic model requires structural and hydraulic data for

252 calibration and validation (Giorgio Bort et al. 2014). Due to the fact that it is extremely difficult

253 to accurately obtain the roughness and diameter of every pipe and water demand at each node,

254 these parameters should be properly calibrated and validated before the hydraulic model can be

255 used. Evolutionary methods and least-squares are the most frequently used methods for model

256 calibration (Sanz et al. 2016).

257 2. Data generation: With the availability of a well-calibrated hydraulic model, various leakage

258 scenarios (including the non-leakage scenario) under different boundary conditions can be

259 modelled (Soldevila et al. 2019). By simulate different leakages with different locations and

260 different scales, a hydraulic model could provide numerous training example for leakage area

261 classification. The leakage can be modelled as an emitter flow that represented as a function of

262 the pressure at the junction node, given as

263 𝑄𝑖 (𝑡) = 𝑘𝑖 [𝑃𝑖 (𝑡)𝛼 ] (1)

264 where 𝑄𝑖 is the leak flow at node i at time t, 𝑘𝑖 is the emitter coefficient at node i, 𝑃𝑖 (𝑡) is the

265 nodal pressure at node i at time t, α represent the emitter pressure exponent.

266 Historical leak-free monitoring measurements (such as pressures, flows, reservoir conditions,

267 etc.) must be provided to the hydraulic model as boundary conditions. The network behaviour

268 can be described by steady-state models concatenated in an extended period simulation (EPS)

269 (Perez et al. 2014). In addition, pressure-driven analysis (PDA) (Wagner et al. 1988) in

270 EPANET provides a more realistic representation of the pressure-leakage relationship.

10
271 3. Zone division: Usually, a small number of monitoring devices are equipped in WDNs. Thus, it

272 is hard to locate the exact location of leakage with limited monitoring devices since leakage

273 that happened at neighbouring pipes might have a very similar influence on available devices.

274 Therefore, in order to locate the leakage in a small possible area within a district meter area

275 (DMA), one solution is to divide the DMA into small zones. Zhang et al. (2016) used k-means

276 clustering algorithm to divide WDNs into k leakage zones. Wu et al. (2022) used fuzzy c-means

277 (FCM) to cluster pipes and place the sensors. Romero et al. (2022) adopted a method called

278 graph agglomerative clustering (GAC) to cluster the network based on its topology.

279 DATA ANALYTIC METHODS (STEP 3)

280 Various techniques have been explored to mining the monitoring data and to provide effective

281 solutions for leakage detection. As mentioned before, leakage detection methods can be categorised as

282 data-driven methods and hydraulic model-based methods.

283 Data-Driven Methods

284 Currently, data-driven methods are mainly used for leakage identification (especially for burst

285 detection) instead of leakage localisation. With a large amount of historical data, the pattern can be

286 analysed by statistical methods or learnt automatically by machine learning models. If the

287 characteristics of new data are substantially different from historical data, it can be inferred that an

288 abnormal event occurred in the distribution system. For example, a sudden pressure drop and flow

289 increase are the most frequently used criteria for burst events. Fig. 3 shows the flowchart of data analytic

290 steps for data-driven leakage detection methods, and Table 3 summarises the techniques used in each

291 method. Based on the principle of the data analysis techniques, data-driven methods can be further

292 categorised into three categories: statistical process control (SPC)-based methods, prediction-

293 classification methods, and clustering-based methods.

294 SPC-based methods

295 Statistical process monitoring charts, also called control charts, with a set of control limits, are used

296 to display and detect the unusual variability in the data. SPC methods are the most intuitive and simple

297 but powerful methods used to monitor the unusual behaviour of a process. These charts contain three

11
298 characteristics: a target representing the mean value for the in-control process, upper control limit, and

299 lower control limit used to determine the in-control limits. The control limits can be set by calculating

300 the statistical characteristics of the historical data. Data that is outside these thresholds are assumed as

301 invalid or abnormal. For example, the well-known “3-sigma” method is a Shewhart-type method, which

302 means that the data three standard deviations from the mean are considered under normal conditions.

303 The most common SPC methods include univariate methods such as Shewhart chart (Loureiro et al.

304 2016), Western Electrical Company (WEC) rules (Ahn and Jung 2019; Jung et al. 2015), cumulative

305 sum (CUSUM) control chart (Misiunas et al. 2006), exponentially weighted moving average (EWMA)

306 (Jung et al. 2015), and multivariate methods that consider the correlation between data from multiple

307 sensors, such as Hotelling 𝑇 2 (Palau et al. 2012), multivariate EWMA (Jung et al. 2015), and

308 multivariate CUSUM (Jung et al. 2015). Shewhart-type approaches provide effective detection of large

309 faults, while CUSUM and EWMA are more sensitive in detecting small changes but do not guarantee

310 to detect large faults (Harrou et al. 2020). Some of the SPC methods can be viewed as a simpler version

311 of prediction model-based methods discussed in the next category. For example, Shewhart uses the

312 mean of historical value as the predicted value for the next data point, and EWMA uses an exponentially

313 weighted moving average value.

314 It should be noted that some assumptions that underlie the quality control process are: 1. Data comes

315 from a single statistical distribution; 2. The data distribution is a normal (Gaussian) distribution; 3. The

316 errors are uncorrelated over time. It is clear that none of these assumptions holds true in the raw

317 monitoring data. Therefore, some researchers (Jung et al. 2015; Loureiro et al. 2016) assumed that the

318 data at the same time every day comes from the same distribution and reformat the data before applying

319 the SPC methods. The Minimum Night Flow (MNF) analysis proposed by Farah and Shahrour (2017)

320 also implies this idea, in which the minimum value of a day is checked based on the calculation of

321 moving average and moving standard deviation. Loureiro et al. (2016) improved SPC methods by using

322 a quantile-based approach instead of a sample mean to consider the asymmetric behaviour of flow data.

323 Besides, instead of considering each data point separately, Palau et al. (2012) divided a day of

324 monitoring data into different time periods (such as morning, afternoon, and night) and used PCA to

325 compress the data and to reduce the unnecessary variability. Then, Hotelling 𝑇 2 and distance to model

12
326 can be calculated to determine the outliers. However, this method takes a relatively long time to make

327 the decision, which may not be conceived as very effective for burst detection.

328 Prediction-classification methods

329 Prediction-classification methods are the most common approaches in the literature. While SPC-

330 based methods try to construct boundaries directly for the monitoring data points based on unrealistic

331 assumptions, prediction-classification methods construct more complicated and accurate models that

332 can represent the expected behaviour of the distribution system in a healthy state. The prediction model

333 is trained with the collected historical data and then gives predicted values. If there is no leakage in the

334 future, there will be a reasonable match between predicted and measure values. Therefore, anomalies

335 representing leakage events can be detected by analysing the residuals between the observed value and

336 the predicted value (see Fig. 3).

337 Various approaches have been explored to detect leakage events in WDNs. Ye and Fenner (2014)

338 have proposed the weighted least squares with the expectation-maximisation algorithm for burst

339 detection. They (Ye and Fenner 2011) also explored the application of Kalman Filter (KF) for burst

340 detection and achieved very good results. However, one shortcoming of these two approaches is that

341 they considered data within the same day separately. If the sampling rate is 15 minutes, it needs to build

342 672 independent models to eliminate diurnal patterns. If complex patterns are considered, such as

343 weekly patterns, more models need to be built. Thus, building a large number of filters cannot radically

344 solve the problem of equal-state assumption.

345 Machine learning techniques can learn from data without relying on rules-based programming.

346 Mounce et al. (2002, 2010, 2011) have used support vector regression (SVR) and artificial neural

347 network (ANN) for leakage detection on real data from a water distribution system. Romano et al. (2014)

348 provided an online system for leakage detection using ANN combined with the Bayesian inference

349 system (BIS). However, a standard ANN does not share features across different steps of time series.

350 In contrast, the recurrent neural network (RNN) is widely recognised as a suitable method to deal with

351 sequential data due to its ability to connect previous information to the present task. Therefore, Wang

352 et al. (2020) used the long short-term memory (LSTM) network - a special kind of RNN - for flow

353 prediction and the detection of burst events.

13
354 Besides the methods that have been explored in water leakage detection (see Table 3), various kinds

355 of regression or prediction models can be applied in the prediction stage (Han et al. 2019), as long as

356 the models are capable of time series modelling. Prediction-classification detection methods highly rely

357 on the accuracy of the prediction model (Wu and Liu 2017). However, time-series predictors generally

358 have no inbuilt mechanism for subsequent classification. Thus, additional classification methods, such

359 as control charts, have been used for the final alarm raising or decision making. However, setting the

360 threshold for event detection is not a trivial issue. The determination of thresholds often depends on

361 experience, which greatly influences the detection effect (Wang et al. 2020). Additionally, most of these

362 prediction-classification methods need to be updated regularly to adapt to time changes (Mounce et al.

363 2010).

364 Clustering-based methods

365 Clustering-based methods are based on comparing time series subsequences or their representations,

366 using a reference of normality, without the need for fitting a prediction model. Clustering analysis is

367 used to create clusters by grouping points or subsequences that are similar to each other and separating

368 dissimilar points or subsequences into different clusters. Abnormal subsequences are those that are

369 dissimilar to normal subsequences, and they can be determined based on the distance to the centroid of

370 the cluster of normal sequences belongs.

371 Wu et al. (2016) used cosine distance to calculate the dissimilarity between vectors combined with

372 the information from different sensors. Aksela et al. (2009) proposed a method based on the self-

373 organising map (SOM) to detect leakage by finding similarities between flow data from other weeks,

374 facilitated by a leak function that describes the relationship between the confidence in the existence of

375 a leak and the distance between flow meters and leakage locations. Wu et al. (2020) proposed a shape

376 similarity-based (SSB) method that detected bursts by analysing the shape of flow time series data

377 within the same period from different days. Huang et al. (2018) applied dynamic time warping (DTW)

378 to study the similarity of daily water demand and found the most unusual daily pattern.

379 The construction of the reference library is the most critical step of clustering-based methods. In

380 order to consider various uncertainties in the data, such as weekday patterns, weekend patterns, holiday

381 patterns, etc., a large amount of historical data may be needed. Furthermore, those discord discovery

14
382 techniques (e.g. Wu et al. (2020), Huang et al. (2018)) require the users to specify the length of the

383 leakage event in advance, which in many cases may not be known and could only be determined by

384 experience. It should be noted that most of the methods are designed for burst detection. Most of the

385 clustering-based methods are based on the assumption that leakage events will cause unusual shapes in

386 the data. However, the shape of the pattern of incipient leakage may stay the same during the beginning

387 stage but grow in the long term. Thus, the ability of clustering methods for incipient leakage detection

388 is still a question.

389 Hydraulic Model-Based Methods

390 Hydraulic model-based approaches rely on a hydraulic model of a network. The accuracy of these

391 models depends on their calibration, and a prerequisite of accurate leakage localisation is a well-

392 calibrated hydraulic model. By comparing the simulations generated by the well-calibrated hydraulic

393 model and the data collected from pressure and flow sensors, the leakage can be detected, and the most

394 probable area of the network can be found. Currently, hydraulic model-based methods are mainly used

395 for leakage localisation. In most studies, the leakage detection is based on the results that a leakage

396 event has already been known to exist in the network. Based on the principles used to detect leakage,

397 hydraulic model-based methods can be further categorised as calibration-based methods, sensitivity

398 analysis-based methods, and classification-based methods. Details of these methods can be found in

399 Table 4. The general steps of hydraulic model-based leakage detection methods have been described in

400 Fig. 4.

401 Calibration-based methods

402 Leakage detection based on model calibration is defined as an inverse problem of parameter

403 identification of the hydraulic model. The leakage detection is initiated by obtaining the well-calibrated

404 hydraulic model and the field measurements. Model calibration is used to minimise the discrepancies

405 between the observed flow and pressure values and the values simulated at junctions in the hydraulic

406 model affected by possible leaks. After selecting the optimisation criteria and the optimisation objective,

407 an optimiser will be used to seek the best solution from all possible solutions automatically. Possible

408 solutions are usually represented as a number of leakage nodes with positive emitter coefficients (Wu

15
409 et al. 2010a). Table 4 summarises the objective function and the optimisation method that have been

410 explored in each study.

411 Misiunas (2006) searched for all the locations, and the node with the smallest objective value (i.e.

412 the sum of difference squares) is declared to be the burst position. This method is based on the leakage

413 identification procedure, the amount of water loss needs to be estimated, and the demand value is

414 assigned uniformly to all nodes. However, the leakage demands should be pressure-dependent, which

415 is more in line with reality. Therefore, Wu et al. (2010b) developed a pressure-dependent leakage

416 detection (PDLD) method that uses pressure-dependent emitter flow at a junction to represent a leakage

417 event. A Genetic algorithm (GA) was then used to search for the optimal solution. Then, Wu et al.

418 (2010b) provided an application report of the PDLD method to two water systems and proved its

419 effectiveness compared with acoustic leak loggers.

420 The number of decision variables is directly related to the number of candidate leak locations in a

421 DMA, which means that the optimisation algorithm needs to solve a nonlinear inverse problem with

422 thousands of decision variables for a medium-sized system. When there are a large number of decision

423 variables because multiple combinations of decision variables may generate equally fit solutions and

424 result in inaccurate localisation results (Sophocleous 2019), which greatly limits the applicability of

425 calibration-based methods. Although the decision variables have been reduced by specifying the

426 maximum number of possible leaks within a system, this information heavily relies on engineering

427 judgment. Therefore, Sophocleous et al. (2018, 2019) introduced a search space reduction stage before

428 leak localisation to reduce the search area of the optimisation. A real case from the UK is investigated

429 by this method and proved its effectiveness. However, these methods have only been demonstrated in

430 single-leak cases because of the combinatorial complexity and a large number of decision variables. In

431 order to take into account multiple leakage scenarios, Berglund et al. (2017) proved that multiple leaks

432 could be modelled as a linear combination of single-leak scenarios under certain limitations (e.g. leak

433 coefficients, leak number). Based on linear programming (LP) and mixed-integer linear programming

434 (MILP), a linear combination of leaks can be determined to approximate the observed pressure values.

435 The advantages of calibration-based methods are: 1) the leak position can be accurately determined;

436 2) the leak amount can also be determined. However, one major drawback of these methods is that they

16
437 are computationally demanding and have limited real-time applicability (Berglund et al. 2017). Sanz et

438 al. (2016) proposed an online leakage detection method based on model calibration. The leak can be

439 determined based on the demand difference between the hydraulic model being calibrated before and

440 after. However, this method is computationally demanding. The calibration-based methods are mostly

441 applied to small networks since the effectivity of the optimisation algorithm can be greatly limited.

442 Sensitivity analysis-based methods

443 Sensitivity analysis-based methods detect leaks based on the comparison of modelled data versus

444 observed data. The quantification of the difference between the actual pressure measurements with the

445 predictions predicted by the hydraulic model is called pressure residual. After the model has been

446 calibrated using the historical data, the pressure residual set for ns sensors can be obtained by calculating

447 the theoretical pressure difference between the non-leak scenario, 𝑝̂ 0 ∈ ℝ𝑛𝑠 , and all potential leak

448 scenarios, 𝑝 ∈ ℝ𝑛𝑠 , simulated at each junction in turn:

449 𝑟 = 𝑝 − 𝑝̂0 (2)

450 Based on evaluating the theoretical effect of all potential leaks fi of all monitored nodes, pi, the

451 sensitivity matrix S can be determined as:

𝜕𝑝1 𝜕𝑝1
⋯
𝜕𝑓1 𝜕𝑓𝑛𝑝
452 S= ⋮ ⋱ ⋮ (3)
𝜕𝑝𝑛𝑠 𝜕𝑝𝑛𝑠
⋯
[ 𝜕𝑓1 𝜕𝑓𝑛𝑝 ]

453 where 𝑛𝑠 is the number of sensors, 𝑛𝑝 is the number of potential leaks (network nodes). However, it is

454 extremely difficult to calculate the sensitivity matrix S analytically for a real network since the WDS is

455 a nonlinear system without an explicit solution (Perez et al. 2014). Therefore, many ways have been

456 proposed to approximate the sensitivity matrix. For example, Giorgio Bort et al. (2014) and Okeya et

457 al. (2015) estimated the burst flow before leakage localisation, and the estimated burst flow was

458 simulated in turn at each node to obtain the sensitivity matrix. Perez et al. (2011, 2014) introduced the

459 same leakage in each node and recorded the pressure increment to approximate the sensitivity matrix.

460 To construct a more robust sensitivity matrix, Farley et al. (2013) used the chi-squared value of pressure

461 increment. Furthermore, to consider the uncertainties that exist in real-life, Pérez et al. (2011) applied

17
462 a threshold to the sensitivity matrix so that only the strong relations between leaks and pressure sensors

463 could be considered.

464 The possible leak area can be determined by ranking the sensitivity of sensors to leaks and compared

465 with the observed pressure residual. Giorgio Bort et al. (2014) performed PCA analysis on the

466 sensitivity matrix to rank the measurement nodes according to the most important feature. Perez et al.

467 (2014) and Steffelbauer et al. (2022) located the most probable leak nodes by identifying the largest

468 correlation values between the observed pressure residual and sensitivity matrix. Theoretically, the

469 residuals should be zero under non-leakage scenarios. However, due to the existence of measurement

470 errors, calibration errors, random customer behaviour etc., the residual will not remain zero even under

471 healthy conditions. Furthermore, the sensitivity of pressure sensors to different leak scenarios is hard to

472 quantify using a constant value, which reduces the accuracy of leak localisation.

473 Classification-based methods

474 With the development of machine learning techniques, the classification of leakage scenarios can

475 automatically be trained by a classifier. In the first stage, the pressure map or pressure residual map of

476 each leakage scenario can be generated by the hydraulic model and used as training data. In the second

477 stage, the training data will be processed and fed to train a classifier. After a burst is detected, observed

478 pressure values will be processed, and the trained classifier could be used to determine the leak area.

479 Zhou et al. (2019a) used the fully-linear DenseNet (FL-DenseNet) to extract features in pressure

480 patterns for burst localisation. Javadiha et al. (2019) used a convolutional neural network (CNN) to

481 learn the different pressure residual maps. Since the number of pressure sensors in a system is limited,

482 some node leaks may present a similar leak signature and can be indistinguishable. Therefore, Soldevila

483 et al. (2016) used a node grouping procedure prior to the k-Nearest Neighbour (kNN) classifier training.

484 Zhang et al. (2016) used K-means clustering to divide the network into different k zones based on the

485 pressure residual matrix generated by the hydraulic model. Then, the leakage events were represented

486 by adding a random leakage demand to the junction selected by the Monte-Carlo method in the

487 hydraulic model. Training samples are generated by the hydraulic model and used to train the M-SVM

488 model. Romero et al. (2022) used an image coding procedure called Gramian angular field (GAF) to

489 transform pressure vectors into images, and the task of leakage localisation has been transformed into

18
490 image classification. A set of deep neural networks (DNNs) are organised hierarchically to obtain a

491 classification tree to localise the leakage area. Zhang et al. (2022) used FCM to divide network into

492 different zones and combined the extreme gradient boosting (XGBoost) to identify the leakage zone.

493 The results showed superior performance than the back-propagation neural network (BPNN). More

494 information can be found in Table 4.

495 With the benefit of a hydraulic model, a large number of training data can be generated and provided

496 for classifier training. Once the classifier is well-trained, the results can be generated efficiently.

497 However, a well-calibrated hydraulic model is hard to be maintained to reflect the real-time condition

498 of the network. Any changes in the network will cause inaccurate estimation, such as the addition or

499 elimination of any element (pipes, valves, tanks, etc.), the changes of pipe roughness coefficient and

500 the changes of pipe diameter caused by increasing pipe ages. Furthermore, consumers’ demand is hard

501 to determine and difficult to adjust its real-time variation. Currently, hydraulic model-based methods

502 have not reached the maturity of real-time monitoring for WDNs.

503 PERFORMANCE EVALUATION (STEP 4)

504 Method evaluation has a critical role in method development, and different goals need different

505 metrics. Each method should be evaluated before being applied to real life. In this section, model

506 validation and method performance with real-life data will be analysed.

507 Leakage Identification

508 For leakage identification, the process usually ends up with binary classification. Essentially, each

509 data point or a data subsequence needs to be labelled as an anomaly or not. If leakage happens and data

510 points during that time period are identified as anomalies, this case is a true positive. If there is no

511 leakage happening and the detection results showed negative all the time (meaning the system is

512 healthy), then it’s called a true negative. However, there are cases the detection method can fail. If the

513 system is healthy, but an alarm is rising, this case is a false positive. If leakage happens but the detection

514 results show negative, this case is a false negative. Therefore, as one of the most comprehensive ways,

515 confusion matrices (Alla and Adari 2019) have been widely used to evaluate leakage identification

516 methods' performances. True Positive Rate (TPR) and False Positive Rate (FPR) are the two most

19
517 commonly used criteria for the evaluation of leakage detection methods. This is because failure events

518 rarely occur in real-life scenarios, resulting in considerable parts of the observed data being labelled as

519 normal, and a few parts of it are labelled as abnormal. As such, other indicators such as F1 score,

520 Receiver Operating Curve (ROC) or area under the curve (AUC) can be used to evaluate such biases in

521 data. For prediction-classification methods, in the prediction stage, additional evaluation metrics such

522 as rooted mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error

523 (MAPE) should be used to measure and quantify the prediction error. It is important to emphasize that

524 the prediction error will never be zero because no model can perfectly predict the future. From Table 5,

525 it could be observed that the current methodologies have achieved very high accuracy, but it should be

526 noted that most of them focused on burst events. Furthermore, only a few of them are applied in real

527 life.

528 Furthermore, the false alarm presents a serious issue. Some results could have false positives every

529 day (Jung et al. 2015, Jung and Lansey 2015, Xu et al. 2020), it is clearly impossible to raise alarms at

530 every false positive point. It is well-known that one abnormal data point cannot solely represent the

531 leakage event, because it has a high probability that it is generated by the data noise or the random

532 behaviour of water consumption, while continuous disruptive data is more suitable to indicate the

533 occurrence of a leakage event. For example, Mounce et al. (2007) combined ANN with Fuzzy Inference

534 System (FIS). Mounce et al. (2011) applied a time window for the detection results, and an alarm will

535 be raised only if enough anomalies occur within a moving event window. Romano et al. (2014) used

536 the BIS to generate probabilities for burst events. However, there is no consensus on the issue of how

537 to represent a leakage event, and currently, the rules that have been used to raise the alarm are

538 determined intuitively by researchers.

539 Detection time (DT) or average detection time (ADT) describes the time duration between the start

540 time of leakage and the time when a method successfully raised the alarm. It is important to raise the

541 alarm as early as possible. In a real-life dataset, the DT could be difficult to be determined, and it could

542 only be inferred from the customers’ contacts or maintenance history. Based on the summary provided

543 in Table 5, it could be observed that prediction-classification methods could receive the quickest

544 response time for burst detection. It is hard to draw a conclusion about which method is the best method

20
545 for leakage detection since each method has their own advantages and disadvantages. Also, the

546 performance of a leakage detection method may depend on the time of occurrence and the magnitude

547 and types of leakage. Moreover, different evaluation criteria are used in those studies, and thus it

548 becomes more difficult to compare the performance.

549 Leakage Localisation

550 Currently, there is no consensus about the evaluation of leakage localisation methods. Most methods

551 show their results using graphical representations of the probable leak nodes or areas and the true leak

552 locations. The visualisation of the leak map could provide an intuitive view of the accuracy of leak

553 localisation. However, quantification metrics are needed so that different methods can be compared

554 with each other. Graphical distance to real leak and pipeline distance to real leak are the two

555 quantification metrics that have been used in literature, and currently, the accuracy could only achieve

556 200 m (see Table 5). In addition, for classification-based method, the classification accuracy for leakage

557 zone localisation have been used for method evaluation.

558 Leakage localisation is important to reduce disruption to customers and traffic by identifying the

559 leak’s location as close as possible. However, it is clear that accuracy and effectiveness have a great

560 potential for improvement. Furthermore, it is important to evaluate the localisation methods and assess

561 their capability in real-life datasets. The accuracy of leakage localisation is affected by several factors

562 such as: (1) the size and types of leakage; (2) location of the leakage; (3) calibration of the model; (4)

563 number and location of the sensors. Therefore, it is obvious that more comprehensive evaluation criteria

564 for leakage localisation are needed. Considering the burst events, Qi et al. (2018) proposed a

565 methodology to investigate the capacity of pressure-based burst detection using several quantitative

566 metrics: (1) undetectable nodes, represent the effectiveness of a detection method and provide

567 information of the need for additional sensors; (2) undetectable demands at those undetectable nodes to

568 assess the potential capacity of a method; (3) detection dimension that indicates the correlation between

569 nodes and the entire pressure sensor distribution; (4) spatial partition that investigate the influence of

570 each sensor considering the distance of the sensor to the leakage; (5) detectable threshold that represents

571 the minimum detectable burst flow. It is important to assess these quantitative metrics for a localisation

572 method in the future research, especially for a system equipped with a large number of sensors.

21
573 CHALLENGES AND LIMITATIONS

574 Although the advantages and disadvantages of some methods have been discussed in step 3 and step

575 4, there are some other issues in leakage detection methods that still need to be addressed.

576 Different Types of Leakage

577 Leaks could happen in all WDNs. Depending on the size of leakage, they can also be categorised as:

578 1) Background leaks (small flow rate, invisible), 2) Unreported leaks (moderate flow rate, invisible), 3)

579 Reported leaks/bursts (high flow rate, visible above ground). The burst events can be easily detected

580 due to a large amount of water loss, but it will have a negative impact on customer satisfaction and may

581 also cause contamination intrusion (Wu and Liu 2017). Compared with bursts, background and

582 unreported leaks can accumulate into greater water loss due to a longer time to awareness. Pre-

583 detection/detection of background and unreported leaks is challenging since the magnitude of leaks is

584 small. Jung and Lansey (2015) seem reached an accurate detection result for small magnitude burst

585 events, but the results were generated using synthetic data. Furthermore, the proposed method used

586 nearly 2,000 days of normal data for statistics calculation, and this kind of information usually is

587 unavailable in real life.

588 As shown in Table 1 and Table 2, most studies, especially data-driven methods, are focused on burst

589 detection and did not evaluate the detection ability for gradual leakage events that developed from

590 incipient leakages to burst events. Unlike bursts that can cause variation in a relatively short period (a

591 few hours) and can be characterised by sudden flow increase and sudden pressure drop, gradual leakages

592 will not generate noticeable deviation in the beginning stage and can be more challenging to detect. In

593 addition, gradual leakage can cause more damage if they remain undetected. However, to the best of

594 the authors’ knowledge, there is no literature currently that have addressed this issue, and an early

595 warning system for gradual leakage events could be a topic in future research. The dataset created by

596 BattleDIM (Vrachminis et al. 2020) modelled different types of leakage (including burst, gradual

597 leakage and background leakage) to evaluate the performance of competitors’ methods. The automatic

598 meter readings (AMRs) provide valuable information for accurate demand calibration, and most of

599 those methods are developed based on the well-calibrated hydraulic model (Steffelbauer et al. 2022,

600 Marzola et al. 2022). However, it should be noted that most water companies do not equip smart water

22
601 meters in real-life. Moreover, due to the complexity of BattleDIM, Marzola et al. (2022) used

602 engineering judgement and visual inspection instead of automated detection.

603 Data Uncertainties

604 In general, WDNs display the same daily water demand pattern, an increase in the early morning

605 and late afternoon during weekdays and a slightly different pattern during weekends. Leakage acts as a

606 demand in the network and affects pressure and flow values, but it does not follow the consumption

607 patterns. The uncertainties within the data have posed great difficulties for leak detection. Most methods

608 did not consider the demand variation caused by weather or the demand variation caused by population

609 increases. The model needs to be retrained regularly to adapt to the changing situation, which may be

610 time-consuming. Moreover, how to differentiate the variation caused by leakage and by weather,

611 holiday behaviour, human activities etc., is still an issue that needs to be addressed.

612 How to handle the spurious outliers is also a question that needs to be considered during leakage

613 detection. The current statistical information can be twisted by the incoming outliers and make the

614 baseline that represents the normal behaviour inaccurate. In order to reduce the influence caused by

615 outliers, Ye and Fenner (2014) used the expectation maximum algorithm to assign different weights to

616 the data to reduce the influence of spurious points. Wang et al. (2020) designed a feedback loop to

617 replace the detected outliers with a more appropriate value. Therefore, a feedback control system that

618 can combine the information of online detection results and model updating is needed. It is especially

619 important for real-time leak detection methods to involve the automatic adjustment of parameters during

620 failed conditions or develop a method that is robust to outliers.

621 Hydraulic Model Uncertainties

622 Hydraulic model-based methods are influenced by uncertainties in both the model and the

623 measurements. In the context of the hydraulic model, several simplifications in modelling will cause an

624 unrealistic representation of the WDN, such as: 1) pipes that are considered not essential will be

625 removed since it is computationally impractical to model all pipes of a large WDN; 2) water demands

626 are aggregated at junctions during the modelling process, but in real life, water usage happened along

627 pipes; 3) WDN input parameters contain uncertainties, such as pipe roughness, emitter coefficient. In

628 addition, the accuracy of the hydraulic model could be influenced by uncertainty in nodal demands,

23
629 measurements errors, etc. Thus, the residual between actual measurements and model output can be

630 different from zero, even in the absence of leaks. It is important to consider these uncertainties and

631 reduce the impact of these uncertainties so that a robust leak localisation method can be developed.

632 Cugueró-Escofet et al. (2015) studied the effect of demand uncertainty on the ability of localisation

633 methods, and Blesa and Pérez (2018) proposed a method of modelling the effect of these uncertainties

634 on model-based localisation methods. The results suggested that future works should consider the

635 uncertainty in the nominal value of the leak, inflow, and sensor measurements.

636 Unrealistic Assumptions

637 Most of the data-driven methods have a very important hidden assumption is that the historical data

638 used for model training doesn’t contain any leakage events and is under normal operation. This is

639 because the accurate parameter estimation of the health situation of a WDN is needed to distinguish the

640 abnormal event. Romano et al. (2014) used SPC for historical data selection to ensure the quality of

641 training data. The accuracy of SPC is limited, and many normal data could be deleted even there are no

642 leakage events. Another frequently used assumption is that the system operation stage is assumed stable

643 during detection because it changes the behaviour of monitoring data. For example, if there is a pump

644 station located downstream of a DMA that pumps water to a water tank, the pumping station will also

645 cause sudden water outflow in the system, and the impact of the sudden pumping flow on the data

646 pattern will be the same as a burst event. Future research should consider how to detect burst or leakage

647 events even under the changing operation situation so that methods could be more robust in real-life

648 applications.

649 Most hydraulic model-based methods use some unrealistic assumptions for real-world deployment

650 such as: 1) assuming no measurement error, modelling error, calibration error, etc. 2) assuming that

651 customer behaviour doesn’t change; 3) modelling leakage events as aggregated demands at junctions

652 or nodes, but in reality, the majority of leaks happen on pipes; 4) assuming that there are no uncertainties

653 in pipe roughness coefficient, which may vary depending on the pipe materials, age, or encrusted

654 materials on the pipe walls. Therefore, hydraulic model-based methods are targeted at finding leaks that

655 occur after calibration. For leaks that have not been correctly identified and located, pipeline roughness

656 values are often misadjusted to compensate for the head loss caused by those unidentified leaks (Wu et

24
657 al. 2010a). Currently, model-based methods still have room for improvement and have not reached the

658 maturity for mainstream adoption (Sophocleous et al. 2019). Thus, the question of how to develop a

659 robust method to overcome these assumptions is not easily solved.

660 Spatial Relation of Sensors

661 Currently, very few studies (Wu et al. 2018a; b) have considered the data from multiple sensors at

662 the same time, but developing a model for every sensor and the decision are made separately. Mounce

663 et al. (2003) proposed a data fusion technology to fuse the information from different sensors and

664 consider the cascading effect between each DMA. Multivariate methods which have been widely

665 applied in fault detection of smart-grid (Zhou et al. 2019b) have a bright future in leakage detection.

666 Compared with the univariate methods, multivariate analysis-based methods can consider the

667 correlations and, therefore, should provide a more efficient detection performance (Ni et al. 2020).

668 From Table 1, it can be observed that the application of data-driven methods is mainly focused on

669 leakage identification. However, it is worth exploring the application of data-driven methods for

670 leakage localisation. Since the reference behaviour can be described by a prediction model or statistical

671 characteristics, the influence of leakage to each sensor can be estimated by the deviance between leak

672 data and reference value. The influence of leakage is relative to the distance between leakage and the

673 sensor, and by quantifying the deviance value, the possible leakage area can be determined. Following

674 this idea, Wu et al. (2018) identified the approximate location information by calculating the

675 abnormality degree of each pressure sensor. Soldevila et al. (2019) have proposed data-driven leakage

676 localisation methods by comparing the pressure map estimated by Kriging spatial interpolation, and

677 Bayesian reasoning is applied to consider the temporal evolution to improve the accuracy of leakage

678 localisation.

679 CONCLUSIONS

680 Leakage affects the majority of water utilities in developed and developing countries. Many efforts

681 have been made to reduce leakage in water distribution systems. Leakage management has five aspects:

682 prevention, assessment, control, detection, and localization and repair. In this article, the systematic

683 review of leakage detection methods provides a compendium of information on existing technologies,

25
684 their main implementation steps and issues of concern for each step, their suitability for different cases

685 studies and their advantages and disadvantages.

686 A framework was developed to assess these methods. The framework has the following main steps:

687 1) definition of the main goal of the study, 2) data collection, 3) pre-processing of data, 4) data analysis

688 and 5) method evaluation. Leakage detection methods based on pressure and flow data were categorized

689 into hydraulic model-based and data-driven approaches. For leakage identification, data-driven methods

690 can take advantage of a large amount of monitoring data and explore more valuable information. For

691 leakage localisation, hydraulic model-based methods could take into account pressure data from

692 multiple sensors and provide more accurate localisation. For each method, the advantages and

693 disadvantages are provided in this paper.

694 In the future, researchers need to consider the limitations of current methodologies. Firstly, almost

695 all papers focus on burst events, but it is crucial to develop an early warning system to detect gradual

696 leakage before it causes obvious disruption and causes more water loss. Secondly, the uncertainties in

697 both monitoring data and hydraulic models have impeded the application of leakage detection methods,

698 and methods that can be robust to these uncertainties are needed in the future. Furthermore, it is

699 important for leak detection methods to involve the automatic adjustment of parameters to achieve

700 better real-time performance. Thirdly, it is important to realise the assumptions that have been made

701 when developing the method, and real-life scenarios are the final goal that makes fewer assumptions.

702 Fourthly, an information interaction system that can consider the information from multiple sensors is

703 needed so that decision making could be made comprehensively instead of independently. Finally, a

704 more comprehensive leakage detection evaluation method needs to be developed.

705 DATA AVAILABILITY STATEMENTS

706 No data, models, or code were generated or used during the study (e.g. opinion or dateless paper).

707 ACKNOWLEDGEMENT

708 The first author is funded by the China Scholarship Council (No. 202006370080), and the work is

709 supported by a Royal Academy of Engineering Industrial Fellowship to resource Raziyeh Farmani’s

710 involvement (IF\192057).

26
711 REFERENCES

712 Abdulshaheed, A., Mustapha, F., and Ghavamian, A. (2017). “A Pressure-Based Method for

713 Monitoring Leaks in a Pipe Distribution System: A Review.” Renewable and Sustainable Energy

714 Reviews, 69, 902–911.

715 Ahn, J., and Jung, D. (2019). “Hybrid Statistical Process Control Method for Water Distribution Pipe

716 Burst Detection.” Journal of Water Resources Planning and Management, 145(9), 06019008.

717 Aksela, K., Aksela, M., and Vahala, R. (2009). “Leakage Detection in a Real Distribution Network

718 Using a SOM.” Urban Water Journal, 6(4), 279–289.

719 Alla, S., and Adari, S. K. (2019). Beginning Anomaly Detection Using Python-Based Deep Learning.

720 New Jersey: Apress.

721 Bakker, M., Vreeburg, J. H. G., Van De Roer, M., and Rietveld, L. C. (2014). “Heuristic Burst Detection

722 Method Using Flow and Pressure Measurements.” Journal of Hydroinformatics, 16(5), 1194–

723 1209.

724 Bakker, M., Vreeburg, J. H. G., van Schagen, K. M., and Rietveld, L. C. (2013). “A Fully Adaptive

725 Forecasting Model for Short-Term Drinking Water Demand.” Environmental Modelling and

726 Software, 48, 141–151.

727 Berglund, A., Areti, V. S., Brill, D., and Mahinthakumar, G. (Kumar). (2017). “Successive Linear

728 Approximation Methods for Leak Detection in Water Distribution Systems.” Journal of Water

729 Resources Planning and Management, 143(8), 04017042.

730 Blesa, J., and Pérez, R. (2018). “Modelling Uncertainty for Leak Localization in Water Networks.”

731 IFAC-PapersOnLine, 51(24), 730–735.

732 Bohorquez, J., Alexander, B., Simpson, A. R., and Lambert, M. F. (2020). “Leak Detection and

733 Topology Identification in Pipelines Using Fluid Transients and Artificial Neural Networks.”

734 Journal of Water Resources Planning and Management, 146(6), 1–11.

735 Chan, T. K., Chin, C. S., Member, S., and Zhong, X. (2018). “Review of Current Technologies and

736 Proposed Intelligent Methodologies for Water Distributed Network Leakage Detection.” IEEE

737 Access, IEEE, 6, 78846–78867.

27
738 Colombo, A. F., Lee, P., and Karney, B. W. (2009). “A Selective Literature Review of Transient-Based

739 Leak Detection Methods.” Journal of Hydro-Environment Research, 2(4), 212–227.

740 Cugueró-Escofet, P., Blesa, J., Pérez, R., Cugueró-Escofet, M., and Sanz, G. (2015). “Assessment of a

741 Leak Localization Algorithm in Water Networks Under Demand Uncertainty.” IFAC-

742 PapersOnLine, 28(21), 226–231.

743 Denyer, D., and Tranfield, D. (2006). “Using Qualitative Research Synthesis to Build an Actionable

744 Knowledge base.” Management decision.

745 El-Zahab, S., Asaad, A., Mohammed Abdelkader, E., and Zayed, T. (2016). “Collective Thinking

746 Approach for Improving Leak Detection Systems.” Smart Water, 2(1), 1–10.

747 Farah, E., and Shahrour, I. (2017). “Leakage Detection Using Smart Water System: Combination of

748 Water Balance and Automated Minimum Night Flow.” Water Resources Management, 31(15),

749 4821–4833.

750 Farley, B., Mounce, S. R., and Boxall, J. B. (2013). “Development and Field Validation of a Burst

751 Localization Methodology.” Journal of Water Resources Planning and Management, 139(6),

752 604–613.

753 Geiger, I. G. (2005). Principles of Leak Detection. KROHNE oil and gas 399.

754 Giorgio Bort, C. M., Righetti, M., and Bertola, P. (2014). “Methodology for Leakage Isolation Using

755 Pressure Sensitivity and Correlation Analysis in Water Distribution Systems.” Procedia

756 Engineering, 89, 1561–1568.

757 Gupta, A., and Kulat, K. D. (2018). “A Selective Literature Review on Leak Management Techniques

758 for Water Distribution System.” Water Resources Management, 32(10), 3247–3269.

759 Han, Z., Zhao, J., Leung, H., Ma, K. F., and Wang, W. (2019). “A Review of Deep Learning Models

760 for Time Series Prediction. ” IEEE Sensors Journal, 21(6), 7833–7848.

761 Harrou, F., Sun, Y., S. Hering, A., Madakyaru, M., and Dairi, A. (2020). Statistical Process Monitoring

762 using Advanced Data-Driven and Deep Learning Approaches. Elsevier.

763 Hu, Z., Chen, B., Chen, W., Tan, D., and Shen, D. (2021a). “Review of model-based and data-driven

764 approaches for leak detection and location in water distribution systems.” Water Supply, 21(7),

765 3282–3306.

28
766 Hu, X., Han, Y., Yu, B., Geng, Z., and Fan, J. (2021b). “Novel Leakage Detection and Water Loss

767 Management of Urban Water Supply Network Using Multiscale Neural Networks.” Journal of

768 Cleaner Production, 278, 123611.

769 Huang, P., Zhu, N., Hou, D., Chen, J., Xiao, Y., Yu, J., Zhang, G., and Zhang, H. (2018). “Real-Time

770 Burst Detection in District Metering Areas in Water Distribution System Based on Patterns of

771 Water Demand with Supervised Learning.” Water (Switzerland), 10(12), 1–16.

772 Ismail, M. I. M., Dziyauddin, R. A., Salleh, N. A. A., Muhammad-Sukki, F., Bani, N. A., Izhar, M. A.

773 M., and Latiff, L. A. (2019). “A Review of Vibration Detection Methods Using Accelerometer

774 Sensors for Water Pipeline Leakage.” IEEE Access, 7, 51965--51981.

775 Javadiha, M., Blesa, J., Soldevila, A., and Puig, V. (2019). “Leak Localization in Water Distribution

776 Networks Using Deep Learning.” 2019 6th International Conference on Control, Decision and

777 Information Technologies, CoDIT 2019, 1426–1431.

778 Jung, D., Kang, D., Liu, J., and Lansey, K. (2015). “Improving the Rapidity of Responses to Pipe Burst

779 in Water Distribution Systems: A Comparison of Statistical Process Control Methods.” Journal

780 of Hydroinformatics, 17(2), 307–328.

781 Jung, D., and Lansey, K. (2015). “Water Distribution System Burst Detection Using a Nonlinear

782 Kalman Filter.” Journal of Water Resources Planning and Management, 141(5), 04014070.

783 Kang, D., and Lansey, K. (2014). “Novel Approach to Detecting Pipe Bursts in Water Distribution

784 Networks.” Journal of Water Resources Planning and Management, 140(1), 121–127.

785 Keramat, A., Wang, X., Louati, M., Meniconi, S., Brunone, B., and Ghidaoui, M. S. (2019). “Objective

786 Functions for Transient-Based Pipeline Leakage Detection in a Noisy Environment: Least Square

787 and Matched-Filter.” Journal of Water Resources Planning and Management, 145(10), 04019042.

788 Loureiro, D., Amado, C., Martins, A., Vitorino, D., Mamade, A., and Coelho, S. T. (2016). “Water

789 Distribution Systems Flow Monitoring and Anomalous Event Detection: A Practical Approach.”

790 Urban Water Journal, Taylor & Francis, 13(3), 242–252.

791 Marzola, I., Mazzoni, F., Alvisi, S., and Franchini, M. (2022). “Leakage Detection and Localization in

792 a Water Distribution Network through Comparison of Observed and Simulated Pressure Data.”

793 Journal of Water Resources Planning and Management, 148(1), 1–11.

29
794 Misiunas, D., Vitkovský, J., Olsson, G., Lambert, M., and Simpson, A. (2006). “Failure Monitoring in

795 Water Distribution Networks.” Water Science and Technology, 53(4–5), 503–511.

796 Mounce, S. R., Boxall, J. B., and Machell, J. (2007). “An Artificial Neural Network/Fuzzy Logic

797 System for DMA Flow Meter Data Analysis Providing Burst Identification and Size Estimation.”

798 Proceedings of the Combined International Conference of Computing and Control for the Water

799 Industry, CCWI2007 and Sustainable Urban Water Management, SUWM2007, (January), 313–

800 320.

801 Mounce, S. R., Boxall, J. B., and Machell, J. (2010). “Development and Verification of an Online

802 Artificial Intelligence System for Detection of Bursts and Other Abnormal Flows.” Journal of

803 Water Resources Planning and Management, 136(3), 309–318.

804 Mounce, S. R., Day, A. J., Wood, A. S., Khan, A., Widdop, P. D., and Machell, J. (2002). “A Neural

805 Network Approach to Burst Detection.” Water Science and Technology, 45(4–5), 237–246.

806 Mounce, S. R., Khan, A., Wood, A. S., Day, A. J., Widdop, P. D., and Machell, J. (2003). “Sensor-

807 Fusion of Hydraulic Data for Burst Detection and Location in a Treated Water Distribution

808 System.” Information Fusion, 4(3), 217–229.

809 Mounce, S. R., and Machell, J. (2006). “Burst Detection Using Hydraulic Data from Water Distribution

810 Systems with Artificial Neural Networks.” Urban Water Journal, 3(1), 21–31.

811 Mounce, S. R., Mounce, R. B., and Boxall, J. B. (2011). “Novelty Detection for Time Series Data

812 Analysis in Water Distribution Systems Using Support Vector Machines.” Journal of

813 Hydroinformatics, 13(4), 672–686.

814 Mounce, S. R., Mounce, R. B., and Boxall, J. B (2012). “Identifying Sampling Interval for Event

815 Detection in Water Distribution Networks.” Journal of Water Resources Planning and

816 Management, 138(2), 187–191.

817 Ni, F. T., Zhang, J., and Noori, M. N. (2020). “Deep Learning for Data Anomaly Detection and Data

818 Compression of a Long-Span Suspension Bridge.” Computer-Aided Civil and Infrastructure

819 Engineering, 35(7), 685–700.

820 Okeya, I., Hutton, C., and Kapelan, Z. (2015). “Locating Pipe Bursts in a District Metered Area via

821 Online Hydraulic Modelling.” Procedia Engineering, 119(1), 101–110.

30
822 Palau, C. V., Arregui, F. J., and Carlos, M. (2012). “Burst Detection in Water Networks Using Principal

823 Component Analysis.” Journal of Water Resources Planning and Management, 138(1), 47–54.

824 Pérez, R., Puig, V., Pascual, J., Quevedo, J., Landeros, E., and Peralta, A. (2011). “Methodology for

825 Leakage Isolation Using Pressure Sensitivity Analysis in Water Distribution Networks.” Control

826 Engineering Practice, 19(10), 1157–1167.

827 Perez, R., Sanz, G., Puig, V., Quevedo, J., Cuguero Escofet, M. A., Nejjari, F., Meseguer, J., Cembrano,

828 G., Mirats Tur, J. M., and Sarrate, R. (2014). “Leak Localization in Water Networks: A Model-

829 Based Methodology Using Pressure Sensors Applied to a Real Network in Barcelona.” IEEE

830 Control Systems, 34(4), 24–36.

831 Porwal, S., Akbar, S. A., and Jain, S. C. (2017). “Leakage Detection and Prediction of Location in a

832 Smart Water Grid using SVM Classification.” International Conference on Energy,

833 Communication, Data Analytics and Soft Computing, IEEE, 3288–3292.

834 Puust, R., Kapelan, Z., Savic, D. A., and Koppel, T. (2010). “A Review of Methods for Leakage

835 Management in Pipe Networks.” Urban Water Journal, 7(1), 25–45.

836 Qahtani, T. Al, Yaakob, M. S., Yidris, N., and Sulaiman, S. (2020). “A Review on Water Leakage

837 Detection Method in the Water Distribution Network.” Journal of Advanced Research in Fluid

838 Mechanics and Thermal Sciences, 68(2), 152–163.

839 Qi, Z., Zheng, F., Guo, D., Maier, H. R., Zhang, T., Yu, T., and Shao, Y. (2018). “Better Understanding

840 of the Capacity of Pressure Sensor Systems to Detect Pipe Burst within Water Distribution

841 Networks.” Journal of Water Resources Planning and Management, 144(7), 04018035.

842 Romano, M., Kapelan, Z., and Savić, D. A. (2011). “Real-Time Leak Detection in Water Distribution

843 Systems.” Water Distribution Systems Analysis 2010, American Society of Civil Engineers,

844 Reston, VA, 1074–1082.

845 Romano, M., Kapelan, Z., and Savić, D. A. (2014). “Automated Detection of Pipe Bursts and Other

846 Events in Water Distribution Systems.” Journal of Water Resources Planning and Management,

847 140(4), 457–467.

848 Romero, L., Blesa, J., Puig, V., and Cembrano, G. (2022). “Clustering-Learning Approach to the

849 Localization of Leaks in Water Distribution Networks.” Journal of Water Resources Planning and

31
850 Management, 148(5), 1–11.

851 Sanz, G., Pérez, R., Kapelan, Z., and Savic, D. (2016). “Leak Detection and Localization through

852 Demand Components Calibration.” Journal of Water Resources Planning and Management,

853 142(2), 04015057.

854 Soldevila, A., Blesa, J., Fernandez-Canti, R. M., Tornil-Sin, S., and Puig, V. (2019). “Data-Driven

855 Approach for Leak Localization in Water Distribution Networks Using Pressure Sensors and

856 Spatial interpolation.” Water (Switzerland), 11(7).

857 Soldevila, A., Blesa, J., Tornil-Sin, S., Duviella, E., Fernandez-Canti, R. M., and Puig, V. (2016). “Leak

858 Localization in Water Distribution Networks Using a Mixed Model-Based/Data-Driven

859 Approach.” Control Engineering Practice, 55, 162–173.

860 Sophocleous, S. (2019). “Development of the Next Generation of Water Distribution Network

861 Modelling Tools Using Inverse Methods.”

862 Sophocleous, S., Savić, D., and Kapelan, Z. (2019). “Leak Localization in a Real Water Distribution

863 Network Based on Search-Space Reduction.” Journal of Water Resources Planning and

864 Management, 145(7), 04019024.

865 Sophocleous, S., Savic, D., Kapelan, Z., Gilbert, C., and Sage, P. (2018). “Leak Detection and

866 Localization Based on Search Space Reduction and Hydraulic Modelling.” WDSA / CCWI Joint

867 Conference Proceedings.

868 Steffelbauer, D. B., Deuerlein, J., Gilbert, D., Abraham, E., and Piller, O. (2022). “Pressure-Leak

869 Duality for Leak Detection and Localization in Water Distribution Systems.” Journal of Water

870 Resources Planning and Management, 148(3), 1–13.

871 Vrachimis, S. G., Eliades, D. G., Taormina, R., Ostfeld, A., Kapelan, Z., Liu, S., Kyriakou, M., Pavlou,

872 P., Qiu, M., and Polycarpou, M. M. (2020). “BattLeDIM : Battle of the Leakage Detection and

873 Isolation Methods. ” 2nd International CCWI/WDSA Joint Conference.

874 Wagner, J.M., Shamir, U., and Marks, D.H. (1988). “Water Distribution Reliability: Simulation

875 Methods. ” Journal of Water Resources Planning and Management, 114(3), 276–294.

876 Wang, X., Guo, G., Liu, S., Wu, Y., Xu, X., and Smith, K. (2020). “Burst Detection in District Metering

877 Areas Using Deep Learning Method.” Journal of Water Resources Planning and Management,

32
878 146(6), 1–12.

879 Wu, Y., and Liu, S. (2017). “A Review of Data-Driven Approaches for Burst Detection in Water

880 Distribution Systems.” Urban Water Journal, Taylor & Francis, 14(9), 972–983.

881 Wu, Y., Liu, S., and Asce, A. M. (2020). “Burst Detection by Analyzing Shape Similarity of Time

882 Series Subsequences in District Metering Areas.” Journal of Water Resources Planning and

883 Management, 146(1), 1–12.

884 Wu, Y., Liu, S., Smith, K., and Wang, X. (2018a). “Using Correlation between Data from Multiple

885 Monitoring Sensors to Detect Bursts in Water Distribution Systems.” Journal of Water Resources

886 Planning and Management, 144(2), 04017084.

887 Wu, Y., Liu, S., and Wang, X. (2018b). “Distance-Based Burst Detection Using Multiple Pressure

888 Sensors in District Metering Areas.” Journal of Water Resources Planning and Management,

889 144(11), 06018009.

890 Wu, Y., Liu, S., Wu, X., Liu, Y., and Guan, Y. (2016). “Burst Detection in District Metering Areas

891 Using a Data Driven Clustering Algorithm.” Water Research, 100, 28–37.

892 Wu, Z. Y., Sage, P., and Turtle, D. (2010a). “Pressure-Dependent Leak Detection Model and Its

893 Application to a District Water System.” Journal of Water Resources Planning and Management,

894 136(1), 116–128.

895 Wu, Z. Y., Burrows, R., Moorcroft, J., Croxton, N., and Limanond, S. (2010b). “Pressure-Dependent

896 Leakage Detection Method Compared with Conventional Techniques.” Water Distribution

897 Systems Analysis 2010, 1083–1092.

898 Wu, J., Ma, D., and Wang, W. (2022). “Leakage Identification in Water Distribution Networks Based

899 on XGBoost Algorithm.” Journal of Water Resources Planning and Management, 148(3), 1–13.

900 Xu, Q., Liu, R., Chen, Q., and Li, R. (2014). “Review on Water Leakage Control in Distribution

901 Networks and the Associated Environmental Benefits.” Journal of Environmental Sciences, The

902 Research Centre for Eco-Environmental Sciences, Chinese Academy of Sciences, 26(5), 955–961.

903 Xu, Z., Ying, Z., Li, Y., He, B., and Chen, Y. (2020). “Pressure Prediction and Abnormal Working

904 Conditions Detection of Water Supply Network Based on LSTM.” Water Science and Technology:

905 Water Supply, 20(3), 963–974.

33
906 Ye, G., and Fenner, R. A. (2011). “Kalman Filtering of Hydraulic Measurements for Burst Detection in

907 Water Distribution Systems.” Journal of Pipeline Systems Engineering and Practice, 2(1), 14–22.

908 Ye, G., and Fenner, R. A. (2014). “Weighted Least Squares with Expectation-Maximization Algorithm

909 for Burst Detection in U.K. Water Distribution Systems.” Journal of Water Resources Planning

910 and Management, 140(4), 417–424.

911 Zaman, D., Tiwari, M. K., Gupta, A. K., and Sen, D. (2020). “A Review of Leakage Detection Strategies

912 for Pressurised Pipeline in Steady-State.” Engineering Failure Analysis, 109, 104264.

913 Zhang, Q., Wu, Z. Y., Zhao, M., Qi, J., Huang, Y., and Zhao, H. (2016). “Leakage Zone Identification

914 in Large-Scale Water Distribution Systems Using Multiclass Support Vector Machines.” Journal

915 of Water Resources Planning and Management, 142(11), 04016042.

916 Zhou, X., Tang, Z., Xu, W., Meng, F., Chu, X., Xin, K., and Fu, G. (2019a). “Deep Learning Identifies

917 Accurate Burst Locations in Water Distribution Networks.” Water Research, 166, 115058.

918 Zhou, Y., Arghandeh, R., Zou, H., and Spanos, C. J. (2019b). “Nonparametric Event Detection in

919 Multiple Time Series for Power Distribution Networks.” IEEE Transactions on Industrial

920 Electronics, 66(2), 1619–1628.

34
Table 1. Summary of steps 0-1 of data-driven leakage detection studies Click here to access/download;Table;Table [Link]

Table 1. Summary of steps 0-1 of data-driven leakage detection studies

Data Collection (Step 1)

Category Reference Goal (Step 0) Leak
Case Study Data Type Frequency Leak Size
Type
SPC-based Misiunas et al. 2006 Identification/ 1 simulated dataset Flow 1 min Burst 5-10 L/s
methods Localisation
Palau et al. 2012 Identification 1 historical dataset Flow 5 min Burst Around 5% of the average flow
Jung et al. 2015 Identification 2 simulated datasets, 1 Flow, Pressure 5 min Burst 0.13%-0.72% of the average flow
historical dataset
Loureiro et al. 2016 Identification 1 historical dataset Flow 15 min Burst --
Ahn and Jung 2019 Identification 1 simulated dataset Flow 5 min Burst 0.1%-3.3% of the mean total
demand
Prediction- Mounce et al. 2002 Identification 1 engineered test dataset Flow, Pressure 15 min Burst --
classification Mounce et al. 2007 Identification 1 engineered test dataset Flow 15 min Burst 5-7 L/s
methods Mounce et al. 2010 Identification 1 historical dataset Flow 15 min Burst/Le 9%-32% of the average demand
ak
Romano et al. 2011, Identification 1 engineered test dataset Flow 15 min Burst 5%-16% of the average inflow
2014
Mounce et al. 2011 Identification 1 engineered test dataset, Flow, Pressure 15 min Burst 6%-12% of the average daily
1 historical dataset maximum flow
Ye and Fenner 2011, Identification 1 engineered test dataset, Flow, Pressure 15 min Burst/Le 1-5 L/s
2014 1 historical dataset ak
Bakker et al. 2014 Identification 1 historical dataset Flow, Pressure 5 min Burst 150 m3/h for largest area, 7 m3/h
for smallest area
Jung and Lansey 2015 Identification 1 simulated dataset Flow, Pressure 5 min Burst 0.3%–7.0% of the mean total
demand
Wang et al. 2020 Identification 1 engineered test dataset Flow 5 min Burst 2.8%-14% of the average inflow
Xu et al. 2020 Identification 1 engineered test dataset Flow, Pressure 5 min Burst --
Clustering- Aksela et al. 2009 Identification 1 historical dataset Flow 1 hour Leak --
based methods Wu et al. 2016 Identification 1 engineered test dataset Flow 5 min Burst 13.3%-23.1% of current inflow
Wu et al. 2018a Identification 1 engineered test dataset Pressure 5 min Burst 13.3%-23.1% of current inflow
Huang et al. 2018 Identification 1 engineered test dataset Flow 15 min Burst 10%-20% of the average inflow
Wu et al. 2020b Identification 1 simulated dataset, 1 Flow 5 min, 15 Burst 6%–12% of the average inflow
engineered test dataset min
-- represents that the information did not provide in the paper
Table 2. Summary of steps 0-1 of hydraulic model-based leakage detection studies Click here to access/download;Table;table 2 - [Link]

Table 2. Summary of steps 0-1 of hydraulic model-based leakage detection studies

Data Collection (Step 1)

Category Reference Goal (Step 0)
Case Study Data Type Frequency Leak Type Leak Size
Calibration-based Misiunas et al. 2006 Identification, 1 simulated dataset Pressure 1 min Burst 5-10 L/s
methods Localisation
Wu et al. 2010b Localisation 1 engineered test dataset, Pressure 30 min Leak --
1 historical dataset
Sanz et al. 2016 Identification, 1 simulated dataset Flow, Pressure 10 min Burst 2.5%-13% of total
Localisation consumption
Berglund et al. 2017 Localisation 3 simulated datasets Pressure 1 hour Leak Less than 0.5% of the
total inflow
Sophocleous et al. 2018, Localisation 1 simulated dataset, 1 Flow, Pressure 15 min Burst 5%-50% of the inlet
2019 historical dataset flow
Sensitivity analysis- Farley et al. 2013 Localisation 1 engineered test dataset Flow, Pressure 15 min Burst --
based methods Kang and Lansey 2014 Identification, 1 simulated dataset Flow, Pressure 1 hour Burst Emitter coefficient of
Localisation 0.1
Perez et al. 2014 Localisation 1 engineered test dataset Flow, Pressure 10 min Leak About 5.6 L/s
Okeya et al. 2015 Localisation 1 simulated dataset Flow, Pressure 15 min Burst 5%-50% of the
average demand
Giorgio Bort et al. 2014 Localisation 1 simulated dataset Pressure -- Leak --
Steffelbauer et al. 2022 Identification, 1 simulated dataset Flow, Pressure, 5 min Burst and 5-30 m3/h
Localisation AMR leak
Classification-based Soldevila et al. 2016 Localisation 3 simulated datasets Flow, Pressure 1 hour Leak 0.84%-2.51% of the
methods total demand
Zhang et al. 2016 Localisation 2 simulated datasets Pressure -- Leak Around 3% of the
average demand
Porwal et al. 2017 Localisation 1 simulated dataset Flow, Pressure 30 min Leak Emitter coefficient
of 0.005-0.1
Zhou et al. 2019a Localisation 2 simulated datasets Pressure 15 min Burst Intensity coefficient
of 10%-30%
Hu et al. 2021b Localisation 1 engineered test dataset Flow, Pressure -- Leak 10-38 L/s
Romero et al. 2022 Localisation 1 historical dataset Pressure 2 min Leak 1.15 L/s
Wu et al. 2022 Localisation 2 simulated datasets Pressure -- Leak 0-25 L/s
-- represents that the information did not provide in the paper
Table 3. Summary of steps 2-4 of data-driven leakage detection Click here to access/download;Table;table [Link]
studies

Table 3. Summary of steps 2-4 of data-driven leakage detection studies

SPC-Based Methods
Evaluation metrics
Reference Data Preparation (Step 2) Data Analytics (Step 3)
(Step 4)
Misiunas et al.
RLS (Denoising) CUSUM DT
2006
Palau et al. Mean centering and Detection
Hotelling 𝑇 2 , DMOD
2012 scaling, PCA effectiveness
Jung et al. WEC, CUSUM, EWMA, M-CUSUM, M-
Normalisation ADT, TPR, NF
2015 EWMA, Hotelling T2
Loureiro et al. Moving average, data
Modified Shewhart chart TPR, FPR
2016 correction, normalisation
Ahn and Jung
Normalisation Hybrid method of WEC and CUSUM ADT, TPR, FPR
2019
Prediction-Classification Methods
Data Analytics (Step 3) Evaluation metrics
Reference Data Preparation (Step 2)
Prediction model Classification (Step 4)
Mounce et al. Data correction, MDN A classification --
2002 normalisation, reformatting module
Mounce et al. Data correction, ANN FIS --
2007, 2010 normalisation, reformatting
Romano et al. SPC (Data selection), data ANN SPC, BIS DT, AUC
2011, 2014 correction, WT (de-
noising)
Mounce et al. Reformatting SVR Binomial event --
2011 discriminator
Ye and Fenner Reformatting KF A user-defined --
2011 threshold
Ye and Fenner Reformatting Weighted least A user-defined --
2014 squares threshold
Bakker et al. -- Adaptive forecasting A user-defined DT, TPR, FPR,
2014 model (Bakker et al. threshold AUC
2013)
Jung and -- KF, NKF CUSUM, Hotelling T 2 ADT, TPR, FPR
Lansey 2015
Wang et al. Reformatting LSTM Multithreshold DT, TPR, FPR
2020 classification based on
time-varying z-score
Xu et al. 2020 Linear interpolation (data A parallel LSTM A user defined value DT, NF
correction), WT (data de- tandem deep neural
noising), normalisation network
Clustering-Based Methods
Data Analysis (Step 3) Evaluation metrics
Reference Data Preparation (step 2)
Similarity measure Clustering (Step 4)
Aksela et al. -- A leak function SOM
2009
Wu et al. 2016 Reformatting Euclidean distance Clustering algorithm TPR, FPR
Wu et al. Data selection, Cosine distance Clustering algorithm TPR, FPR
2018a reformatting, normalisation
Huang et al. Data selection DTW Random forest TPR, FPR
2018
Wu et al. Normalisation, data Increase-state ASS algorithm TPR, NF, FPR
2020b selection, reformatting distance
-- represents that the information did not provide in the paper
Table 4. Summary of steps 2-4 of hydraulic model-based Click here to access/download;Table;table 4 - [Link]
leakage detection studies

Table 4. Summary of steps 2-4 of hydraulic model-based leakage detection studies

Calibration-Based Methods
Data Analytics (Step 3)
Data Preparation Evaluation
Reference Optimisation
(Step 2) Objective Function metrics (Step 4)
Algorithm
Misiunas et al. De-noising (RLS), Minimising the sum of Trial-and- Comparison of
2006 model calibration difference squares error candidate node
and predict node
Wu et al. Model calibration 1. Minimise the sum of Genetic Geographic
2010b difference squares Algorithm distance to leak
2. Minimise the sum of absolute (GA)
differences
3. Minimise the maximum
absolute difference
Sanz et al. Model calibration Minimising the error in pressure Least squares Graphical distance
2016 and flow measurements to leak, Pipe
distance to leak
Berglund et al. Model calibration Minimising the sum of absolute LP, MILP Comparison of
2017 pressure difference candidate node
and predict node
Sophocleous et Model calibration 1. single-leak: minimise the Search space Geographic
al. 2018, 2019 weighted sum of squared flow reduction, distance to leak
2. n leak: minimise the weighted GA
sum of squared differences for
both pressure and flow
Sensitivity Analysis-Based Methods
Data Data Analytics (Step 3) Evaluation
Reference Preparation
Sensitivity Analysis Decision Making metrics (Step 4)
(Step 2)
Farley et al. Model calibration Jacobian sensitivity GA --
2013 matrix
Giorgio Bort et Model calibration Sensitivity matrix PCA, Least squares --
al. 2014
Kang and Model calibration Binarised sensitivity Statistical analysis --
Lansey 2014 matrix
Perez et al. Reformatting, Sensitivity matrix Biggest correlation Visualization of
2014 Model calibration values leak map
Okeya et al. Model calibration Binarised matrix Trial-and-error Visualization of
2015 leak map
Steffelbauer et Model calibration Jacobian sensitivity Highest pearson Geographic
al. 2022 based on a so- matrix correlation sum distance to leak
called dual
approach
Classification-Based Methods
Data Preparation Data Analytics (Step 3) Evaluation metrics (Step
Reference
(Step 2) Training Data Classifier 4)
Soldevila et al. Model calibration, data Pressure residuals kNN Graphical distance to
2016 generation, node grouping leak
Zhang et al. Model calibration, zone Pressure residuals M-SVM Visualization of leak
2016 division (k-means), data map, classification
generation accuracy
Porwal et al. Model calibration, data Leakage and non- SVM Classification accuracy
2017 generation leakage scenario
Zhou et al. Model calibration, data Leakage and non- FL- Visualization of leak
2019a generation leakage scenario DenseNet map
Hu et al. Model calibration Pressure and flow DBSCAN- Comparison of candidate
2021b residuals MFCN node and predict node
Romero et al. Data generation, image Pressure data DNN Pipe distance to leak
2022 encoding (GAF), zone
division (GAC)
Wu et al. 2022 Data generation, zone Pressure residuals XGBoost Classification accuracy
division (FCM)
-- represents that the information did not provide in the paper
Table 5. Summary of performance evaluation of each method Click here to access/download;Table;table 5 - [Link]

Table 5. Summary of performance evaluation of each method

Performance
Results References Categories
evaluation criteria
Leak identification
Detection time Within 5 minutes Misiunas et al. 2006 SPC-based
(DT) or average Around 2-12 hours Jung et al. 2015 SPC-based
detection time 1 hour 45 minutes average Prediction-classification
(ADT) In most cases, within 15 Romano et al. 2011, 2014 Prediction-classification
minutes
Less than 30 minutes Bakker et al. 2014 Prediction-classification
Around 1 – 5 hours Jung and Lansey 2015 Prediction-classification
Within 10 minutes Wang et al. 2020 Prediction-classification
Within 10 minutes Xu et al. 2020 Prediction-classification
Around 3-10 hours Sanz et al. 2016 Calibration-based
True positive rate Around 55% - 78% Jung et al. 2015 SPC-based
(TPR) Around 80% - 93% Loureiro et al. 2016 SPC-based
Or detection In the beset model 81% Ahn and Jung 2019 SPC-based
probability (DP) In the best case, 90% Bakker et al. 2014 Prediction-classification
In the best model, 98% Jung and Lansey 2015 Prediction-classification
In the best model, 100% Wang et al. 2020 Prediction-classification
In the best model, 71.43% Wu et al. 2016 Clustering-based
In the best model, 100% Huang et al. 2018 Clustering-based
In the best model, 90% Wu et al. 2020b Clustering-based
False Positive rate Around 10% - 16% Loureiro et al. 2016 SPC-based
(FPR) Around 0 - 1% per day Ahn and Jung 2019 SPC-based
In the best cases, 2.1% Bakker et al. 2014 Prediction-classification
Around 0 – 1% per day Jung and Lansey 2015 Prediction-classification
In the best model, 0.41% Wang et al. 2020 Prediction-classification
Around 0.4% – 0.8% Wu et al. 2016 Clustering-based
In the best model, 0% Huang et al. 2018 Clustering-based
Around 5% - 10% Wu et al. 2020b Clustering-based
Number of false 1-18 false positives per day Jung and Lansey 2015 Prediction-classification
positives (NF) 2 false positives per day Xu et al. 2020 Prediction-classification
50-100 false positives Wu et al. 2020b Clustering-based
AUC 0.88 Romano et al. 2011, 2014 Prediction-classification
0.972 for larger bursts, 0.535 Bakker et al. 2014 Prediction-classification
for all bursts
Leak localisation
Geographic In most cases, around 200 m Sanz et al. 2016 Calibration-based
distance to leak Within an area of 100 m radius Wu et al. 2010b Calibration-based
In most cases, within 200 m Sophocleous et al. 2018, Calibration-based
2019
Around 200 m Soldevila et al. 2016 Classification-based
Within 500 m Porwal et al. 2017 Classification-based
Pipe distance to In most cases, around 200-400 Sanz et al. 2016 Calibration-based
leak m
In most cases, within 250 m Steffelbauer et al. 2022 Sensitivity analysis-
based
Leakage scope is narrowed to Zhang et al. 2016 Classification-based
the pipelines near the predicted
leakage nodes
442 m Romero et al. 2022 Classification-based
Classification Around 80.78% - 99.25% Zhang et al. 2016 Classification-based
accuracy Around 40% - 90% Porwal et al. 2017 Classification-based
Around 85 % - 100% Zhou et al. 2019 Classification-based
Around 67.2% - 90.4% Wu et al. 2022 Classification-based
Fig. 1. Classification of leakage management and leakage detection methods Click here to access/download;Figure;Fig. 1. .tiff
Fig. 2. The framework of data analytics for leakage detection Click here to access/download;Figure;Fig. 2. .tiff
Fig. 3. Data analytic steps for data-driven leakage detection methods Click here to access/download;Figure;Fig. 3. .tiff
Fig. 4. Data analytic steps for hydraulic model-based leakage detection methods Click here to access/download;Figure;Fig. 4. .tiff
Figure Caption List

List of Figures

Fig. 1. Classification of leakage management and leakage detection methods

Fig. 2. The framework of data analytics for leakage detection

Fig. 3. Data analytic steps for data-driven leakage detection methods

Fig. 4. Data analytic steps for hydraulic model-based leakage detection methods