Hybridisation of brownboost classifier and glow-worm swarm based optimal sensor placement for water leakage detection

. Water Distribution System distributes the water to customer with the better quality and pressure. Due to the leakage, the sufficient amount of water is not delivered to the consumer. Many researchers introduced the techniques for detecting the water leakage in distribution system. But, the water leakage detection accuracy was not improved and time consumption was also not reduced. To improve the water leakage detection performance, Enhanced BrownBoost Classifier 10 based Glowworm Swarm Optimization (EBBC-GWO) Method is introduced. Enhanced BrownBoost Classifier model considers k-Nearest Neighbor (k-NN) classifier as weak classifier. It combines all k-NN classifier to construct strong classifier for classifying normal data or abnormal data with higher classification accuracy. After classification, optimization process is executed where every solution corresponds to the glowworm (i.e., abnormal pressure data node) in search space. Glowworm updates its location to the glowworm in dynamic decision space and optimal one is selected for water leakage 15 detection. By this way, water leakage detection accuracy is improved with lesser false positive rate. Experimental results demonstrated that EBBC-GWO method is higher in case of classification accuracy, false positive rate, classification time and water leakage detection accuracy than the state-of-the-art method.


Introduction
Water Management System is efficient through the development of automated system for leakage detection in water network.Water leaks in Water Distribution Network (WDN) leads to economic losses for final consumer in huge quantities of fluid transportation.Therefore, finding leaks is important one in water network.In different WDN, losses because of leaks are calculated to 30% of extracted water.Water distribution is installed with underground pipes.The underground pipeline monitoring is difficult than open space water pipelines.Leaks in pipes are due to many several factors like pipe's age, improper fitting and disasters (Sadeghioon et al. 2018).Leak management in pipe networks are developed in (Puust et al. 2010) through leakage assessment methods.A solution is needed to identify and to determine leak in the water network.
When the leaks are detected early, corrective actions are carried out to avoid wastage of natural resources and economical losses (Perfido et al. 2017).
A fast water leakage detection system with adaptive design was introduced in (Kang et al. 2018) that combined onedimensional convolutional neural network and support vector machine (1D-CNN-SVM) model.An actual water pipeline network was constructed in graph network and leakage occurred at virtual points on graph.However, the water leakage detection accuracy was not improved.In order to improve the water detection accuracy, a multi-stage graph partitioning approach was introduced in (Rajeswaran et al. 2018) to find the off-line flow measurements were required for reducing cost.
But, the classification performance was not carried out for water leak detection.
To improve the classification performance, a new statistical framework was introduced in (Fagiani et al. 2016) for leakage identification.The framework comprises three sections for extraction and selection of features at leakage detection.
But, the classification accuracy was not improved using new statistical framework.Leakage detection and localization on water transportation was introduced in (Kayaalp et al. 2017) to recognize and find leaks on water pipelines positions with pressure data.Three multi-label classification techniques like RA-kELd, BRkNN, and BR with SVM were employed for leak detection.The false positive rate was not minimized in this system.Therefore, a new leak detection technique was presented in (Sadeghioon et al. 2018) for water distribution pipelines as it buried pressurized fluid flow pipe.The detection technique was dependent on pressure sensor connected with temperature difference.The pipe failure detection accuracy was improved with no false positives.But, the time taken for water leakage detection was not reduced using anomaly detection algorithm.
To solve the above issue, nominal variable was developed in (Dawidowicz 2017) with artificial neural network.The developed scheme also addresses the issues with pressure and division of water distribution system into the pressure zones.
The classification was carried out depending on neural network variables explaining particular parameters that change the water distribution system design.However, the calculation of accuracy was not efficient.To address this issue, a new model was introduced in (Hajibandeh 2018) for leakage amount and location detection.A nodal pressures and demands were adjusted through multi-objective ant colony based optimization model.But, leakage detection time consumption was not reduced using ant colony based optimization model.
To analyze the time consumption, an algorithm structure was constructed in (Kumar et al. 2017) with modularity of wavelet and neural network that join the capability of wavelet transform through examining leakage signals and classification ability of artificial neural networks.The study authenticated that time domain was not apparent to features concerning noisy leak signals and significance of selection of mother wavelet in water distribution pipes.Yet another time consumption issue was addressed by Leak detection technique in (Martini et al. 2015) through vibration monitoring methods.
The long-term objective was attained for automatic detection of leaks.The leak detection accuracy was not improved for automatic detection.To overcome this limitation, the designed approach was introduced in (Sousa et al. 2015) depending on the steady-state modeling through monitoring tank flow and pressure at strategic nodes.The selection of pressure monitoring nodes was performed with graph theory ideas in WDN.But, the classification process was not carried out in designed approach.
To address the classification process, a new method was introduced in (Seyoum et al. 2017) for household leakage detection through sound signal recordings.The designed approach comprised recording of sound signals created through water fixtures and appliances.The recordings were employed to identify abnormal situation (i.e., leak).But, the classification accuracy was not improved.To solve the above said problem, a new methodology was introduced in (Steffelbauer and Fuchs-Hanusch 2016) with uncertainties of different sources in optimal sensor placement issues for leak localization on potential pressure measurement points.The issues were addressed for diverse sensors and uncertainties.Yet another sensor placement optimization issues were addressed in (Christodoulou et al. 2013) in urban water distribution networks through entropy-based approach for feasible waterloss incident detection.The designed method was employed for longitudinal for acoustic, pressure or flow sensors on pipe segments.But, the optimization was not carried out in effective manner because it failed to consider the real-time operating pressure and flow data.
To solve the above issue, a new sensor placement approach was introduced in (Casillas et al. 2013) for leak location in Water Distribution Networks (WDNs).The sensor placement issues were addressed through solving integer optimization problem.However, sensor placement approach was not effective in detecting the water leak.Therefore, Water leakage detection techniques were introduced and issues were highlighted in (Adedeji et al. 2017).It accomplished the effort and development in leakage detection technologies.It was efficient detection of background kind leakages.But, time consumption was not reduced using water leakage detection techniques.To address this problem, a leakage detection method was introduced in (Pérez et al. 2009) depending on discrepancies among pressure measurements and estimations attained from calibrated water distribution network.Every sensor in network identified discrepancy in pressure because of leakage based on its location.
Yet another leak detection method was introduced in (Zhang and Wang 2011) depending on Bayesian theory and Fisher's law for water distribution systems.A hydraulic model was linked with parameters of leaks.The randomness of parameter values was computed through the probability density function and Bayesian theory.The false positive rate was not lessened using leak detection method.In order to reduce the false positive rate, a factorized distribution assigned the inflow demand across the consumption nodes with individual billing data and transmitted the demand across consumption nodes in (Moors et al. 2018).The automatic leak localization method with demand distribution models was developed to compute the leak localization performance.But, the performance of leak identification was not carried out in effective manner because it varied based on the timing and duration of the measurement.To solve this, the location of leak was determined in (Wachla et al. 2015) by group of neuro-fuzzy classifiers.The main aim of the method was to increase the accuracy of leak detection and location.The designed classifier only predicts the leak in the one subarea of water distribution network.In order to overcome this problem, a novel method called robust leakage detection was introduced in (Kim et al. 2015) using pressure data.However, the leakage prediction with minimum time was not ensured.
To solve the above issue, A Sampling Design (SD) method was introduced in (Gamboa-Medina and Reis 2017) for localization and quantification of pressure sensors in WDS for leak detection.Sampling design is the selection of test locations in water distribution system for data collection.SD addressed four criteria, namely improvement of leak sensitivity and sensitivity constancy and reduction of information redundancy and sensor count.Though, the iterative increase in the number of sensing devices/locations on a network, the SD was unable to be performed well.However, the classification time was not reduced by using SD method.To address this issue, a new technique was designed in (Islam et al. 2011) to identify and to diagnose the leakage in WDS.A fuzzy-based algorithm incorporated many uncertainties into many WDS parameters like roughness, nodal demands and water reservoir levels.But, the water leakage detection accuracy was not improved by fuzzy-based algorithm.
The certain problems are identified from the existing methods are lesser classification accuracy, higher classification time consumption, lesser water leakage detection accuracy, higher false positive rate, and so on.In order to overcome these problems, proposed EBBC-GWO method is developed with the objective of increasing classification accuracy with minimum time while detecting water leakage.
Enhanced BrownBoost Classifier model in EBBC-GWO takes k-NN classifier as weak and categorize training samples by majority vote of neighbor.Then the strong classifier is constructed by combining all weak classifiers.This in turns, classification accuracy is increased (i.e., 20% than the state-of-the-art-method) with minimum time.
In Glowworm Swarm Optimization model is used to find the optimal node to place the sensor for detecting water leakage.This is done by computing objective function of each glowworm.
The paper is structured as follows; Section 2 explains EBBC-GWO Method.Section 3 discusses the simulation setting.Section 4 presents result analysis of proposed EBBC-GWO method with various parameters.In section 5, the conclusion of the work is given.

Methodology
In order to perform the leak detection, EBBC-GWO Method is introduced.The major contribution of the paper is described as follows.EBBC-GWO method comprises two models namely, Enhanced BrownBoost Classifier model and Glowworm Swarm Optimization model.Enhanced BrownBoost Classifier model uses k-Nearest Neighbor classifier as a weak.k-NN classifier is a non-parametric method to classify training samples by majority vote of neighbor with object being assigned to the class.After that, Brownboost combines the weak learner to form the strong classifier.By this way, the data are classified as normal data or abnormal data with higher classification accuracy.In Glowworm Swarm Optimization model, every solution corresponds to the glowworm in the search space.Every glowworm has the objective function for solving the optimization problem.Every glowworm sets the objective function at current location into luciferin value.Each glowworm functions in probabilistic manner to find the neighbor with higher luciferin value and travels toward it.Glowworm updates its location to the glowworm in dynamic decision space and optimal one is selected for water leakage detection.

System Model
A water distribution system comprises the set of graph branches ( ) where denotes set of edges (i.e., pipes) and denotes the set of vertices (i.e., nodes) explaining about the pipe connections and endings in (R. Sarrate et al. 2014).A water distribution system is a collection of pipes (i.e., links) connected to the nodes (junctions, tanks and reservoirs).The water flows are computed through hydraulic balancing, and through solving the equations at every node and links.EPANET software works out the network hydraulic equations automatically (Rossman 2000).By using the graph model, EPANET software obtains the pressure and flow rate of every node and links from hydraulic simulation.EPANET software tool is simulated at without leak condition and pressure data readings are collected.The collected readings are taken as the threshold pressure data.

Enhanced BrownBoost Classifier based Glowworm Swarm Optimization (EBBC-GWO) Method
The key aim of EBBC-GWO method is to place the minimal number of sensors at optimal nodes for detecting the leakages.EBBC-GWO method comprises two models, namely Enhanced BrownBoost Classifier Model and Glowworm Swarm Optimization Model for detecting the leakage in water distribution system.In Enhanced BrownBoost Classifier Model, the pressure data obtained from the EPANET tool is taken as an input.Then, the pressure data is classified as normal pressure data or abnormal pressure data with higher classification accuracy.The pressure flow is normal (i.e., maintained by a constant pressure) in the collection pressure data it is classified or identified as normal pressure data.Whereas, the pressure flow is abnormal (not a constant value) in the collection of pressure data is identified as abnormal pressure data.The abnormal pressure occurs when the pipe break, water leakage or fault in the pipeline is identified.After classification, abnormal pressure data is considered and the sensor is placed at the optimal nodes using Glowworm Swarm Optimization for detecting the leakage.The architectural diagram of EBBC-GWO method is illustrated in figure 1.
From figure 1, the architectural diagram of EBBC-GWO method is described.Initially, the WDN is constructed by using EPANET tool and reading is taken at without leak condition.After that, the pressure data are collected from the nodes in WDS.The similarity between the pressure data and threshold pressure data are calculated.The classification process is carried out as normal pressure data or abnormal pressure data.Then, abnormal pressure data nodes are distributed and taken as glowworms.The objective function of each pressure data is estimated.After that, the lesser brighter glowworm moves towards the higher brighter glowworm and then updates its position.Finally, the optimal glowworms are taken for sensor placement to detect the water leakage.The brief explanation of Enhanced BrownBoost Classifier Model and GSO Model are given next subsections.

Enhanced BrownBoost Classifier Model
Initially, large number of pressure data is collected from nodes.After performing the data collection, classification process is carried out by using Enhanced BrownBoost Classifier (EBBC) in (Freund 2000) to find out the node is normal node or leakage possibility node.Enhanced BrownBoost classifier is a boosting method with two result outcomes.The traditional classification technique failed to classify the pressure data with more accuracy and minimal time for leakage detection.In order to increase the accuracy, Enhanced BrownBoost classifier is used which combines weak hypotheses (i.e.weak learner) to constructs the final strong classifier.EBBC-GWO method employs k-NN classifier as weak learner.k-NN classifier is non-parametric process to classify the training samples (i.e.pressure data of node) through majority neighbor votes for allocating to the class.In EBBC-GWO method, enhanced Brownboost classifier combines the weak learner to form the strong classifier.The flow process diagram of enhanced brownboost classifier is explained in figure 2.
From ( 10), represents the normal pressure data and symbolizes the abnormal pressure data.
The algorithmic process of enhanced Brown boost classifier is described as follows.

Output: Increases classification accuracy and reduces classification time
Step 1: Begin Step 2: For each pressure data Step 3: Initialize Step 4: Calculate Step 5:

Find weak classifier
Step 6: Compute value of and Step 7: Update the margins Step 8: Update remaining time Step 9: if ( ) then Step 10: Pressure data is normal Step 11: else Step 12: Pressure data is abnormal Step 13: End if Step 14: End for Step 15: End Algorithm 1: Enhanced Brown boost Classifier Algorithm Algorithm 1 describes the enhanced Brownboost classifier with kNN base classifier for classifying the pressure data with higher classification accuracy and minimum time.The collection of pressure data is correctly classified as normal and 5 abnormal pressure data is called as classification accuracy.The amount of time taken to perform classification process is known as classification time.Initially, the remaining time is initialized for performing the classification.Subsequently, the weak classifier is determined when output of weak learner is greater than zero.kNN classifier is combined for providing the better classification results.k-NN classifier performs classification on votes of neighbor with object allocated to the class.
After that, enhanced brownboost classifier combines the results of all kNN classifiers.The strong Brownboost classifier 10 produces the results +1 or -1 where +1 represents that the pressure data is normal and -1 represents the pressure data is abnormal.Brownboost classifier includes the ability to learn a collection of pressure data into a fixed level of accuracy thus classifies the data with classification accuracy than the state-of-the-art methods.After classification of pressure data, the abnormal pressure data are taken for performing the optimization process to place the sensor.

Glowworm Swarm Optimization (GSO) Model
The GSO is a kind of new swarm intelligence optimization method based on the behaviour of glowworms by (Krishnan and and Ghose 2006).The behaviour pattern of glowworms varies with the intensity of luciferin emission and glow with different intensities.The agents in GSO are taken as the glowworms (i.e, abnormal pressure data nodes) that send the luminescent called luciferin.Every glowworm uses luciferin to transmit the information of the current location to their neighbors.Glowworms are attracted to brighter luciferin neighbors.The movement depends on the local information and selective neighbor interactions enable the swarm to partition into the disjoint subgroups that satisfies the objective function.
In GSO algorithm, every solution corresponds to glowworm in search space and every glowworm has objective function (i.e., higher flow rate) for addressing the optimization problem (i.e., sensor placement on optimal node for leakage detection).Objective function is function that aims to maximize the flow rate or minimize the error rate.Every glowworm sets the objective function at current location into luciferin value.The glowworms are based on the variable neighbourhood to detect the neighbors and compute the movements.Each glowworm employs the probabilistic way to choose the neighbor with higher luciferin value and moves toward it.Each glowworm has diverse decision space with luciferin values enhanced than itself and distance inside decision radius.Glowworm renews the location to the glowworm in decision space probability.GSO has two main objectives: The agents glow at intensities proportional to objective function being optimized.It is taken that the glowworms of brighter intensities attracts the glowworms with lower intensity.
GSO algorithm comprises decision range where distant glowworms is reduced when it has adequate neighbors count or range beyond perception.
Every glowworm comprises the local-decision domain bound by radial sensor range.In local decision domain, every glowworm detects their neighbor and gets attracted by brighter glow of additional glowworms in neighbourhood set.
Local-decision domain size is changeable and gets affected by its neighbors.If the neighbor has lesser density, local-decision domain increases for identifying the neighbors.Otherwise, local-decision domain gets reduced.Lastly, the movement of glowworms resulted in the majority gathering to many optima.Figure 3   The GSO algorithm comprises four steps, namely deployment of glowworms phase, luciferin-update phase, movement phase and local-decision domain update phase for finding the optimal glowworm.

Deployment of glowworms phase:
In this phase, the glowworms are distributed in whole objective space.Every glowworm comprises the equal quantity of luciferin and sensor range.

Luciferin-update phase:
The luciferin update is depending on the function value at glowworm position and all glowworms initiates the similar luciferin value during initial iteration.The values get varied consistent with the function values at current positions.
During the luciferin update phase, every glowworm inserts to the previous luciferin level.A luciferin amount depends on the measured value of sensed profile.A fraction of luciferin value is subtracted to replicate the decay in the luciferin with time.
From ( 12), the objective function (i.e., higher flow rate) of each glow warm is calculated.

Movement-phase
During the movement-phase, each glowworm takes the decision by probabilistic method for shifting to their neighbor with higher luciferin value.They are attracted to the neighbors which glow brighter.For every glowworm , probability of moving to the neighbor is formulated, ; From ( 13) and ( 14), denotes the time index.signify the Euclidian distance between glowworms and at time .represent the luciferin level with glowworm at time .denotes the variable local-decision range linked with glowworm at time .is the radial range of luciferin sensor.Each glowworm updates their location by, From ( 15), denotes the step size.symbolizes the Euclidean Operator of glowworms.

Local-decision range update rule
When glowworms are depending on the local information to decide movements, it is considered that the number of peaks captured is strong function of radial sensor range.The radius of every glowworm dynamic decision space is utilizes the current radius of dynamic decision space and links with the radial range of luciferin sensor.The update rule of every glowworm dynamic decision space radius is given by, ( 16) From ( 16), denotes the constant parameter.symbolizes explicit threshold parameter.From this updated rule, the best glowworm is identified.The algorithmic process of Glowworm Swarm Optimization is given as follows,

\\ Glowworm Swarm Optimization Algorithm
Input: Abnormal pressure data nodes Output: Optimizes the node for sensor placement Step 1: Begin Step 2: For each abnormal pressure data nodes Step 3: Calculate the objective function Step 4: Find probability of movement by comparing objective function value with its neighbor Step 5: if ( ) then Step 6: Movement takes place to neighbouring position Step 7: else Step 8: No movement takes place Step 9: End if Step 10: Updates position of node Step 11: End for Step 12: Selects optimal node for sensor placement Step 13: End Algorithm 2: Glowworm Swarm Optimization Algorithm Algorithm 2 explains the Glowworm Swarm Optimization to select the optimal one with minimum false positive rate and maximum water leakage detection accuracy.False positive rate is measured as the leak which is incorrectly detected through the sensor placed nodes.Whereas, water leakage detection accuracy is the correct detection of leak by placing sensor at optimal nodes.The agents in GSO are the abnormal pressure data nodes.Every glowworm estimates their objective function and sends the information of current location to their neighbors based on objective function.Lesser objective function nodes get attracted by brighter objective function neighbors.Every glowworm takes the decision through probabilistic mechanism for moving to its neighbor with higher objective function.After that, the position of nodes gets updated and the optimal node is selected for the sensor placement to detect the leakage in WDS.By this way, the water leakage detection accuracy is improved.

Simulation Settings
EBBC-GWO Method is implemented in MATLAB Simulink with 3.4 GHz Intel Core i3 processor, 4GB RAM, and windows 7 platform for water leak detection.The nodes pressure and demand flow in water pipeline networks are collected at without leak condition from extended-period hydraulic simulation using EPANET software.EBBC-GWO method used Barcelona water distribution network.From these source 4645 km of pipeline is considered.EBBC-GWO Method is designed for detecting the water leakage in WDS and compared with existing onedimensional convolutional neural network and support vector machine (1D-CNN-SVM) model (Kang et al. 2018) and Multi-Stage Graph Partitioning Approach (Rajeswaran et al. 2018).The efficiency of EBBC-GWO method is evaluated along with the metrics such as classification accuracy, classification time, water leakage detection accuracy and false positive rate.

Classification Accuracy (CA)
Classification accuracy is the one of the important parameter to evaluate the performance of proposed work.It is measured in terms of percentage (%).Classification accuracy is mathematically formulated by, ( 17)

Sample calculation:
 1D-CNN-SVM model: the number of pressure data correctly classified is 30 from simulation and the total number of pressure data is 50.After that, classification accuracy is computed as,  Multi-Stage Graph Partitioning: the number of pressure data correctly classified is 39 using simulation and the total number of pressure data is 50.After that classification accuracy is calculated as,  Proposed EBBC-GWO Method: the number of pressure data correctly classified is 45 using simulation and the total number of pressure data is 50.After that, the classification accuracy is obtained as, When the classification accuracy is higher, method is efficient.(Rajeswaran et al. 2018).
The pressure data is taken from DMA in Barcelona water distribution network.In order to analyze the classification performance, the pressure data is gradually increased.While considering number of pressure data is 500, the classification accuracy is improved up to 88%, 93% and 98% using 1D-CNN-SVM model, Multi-Stage Graph Partitioning Approach and EBBC-GWO Method respectively.From the table, it clear that the classification accuracy using EBBC-GWO Method is higher than other existing techniques (Kang et al. 2018), (Rajeswaran et al. 2018).
Classification process is performed to identify the data as normal data or abnormal data to further processing.For performing the classification, enhanced brownboost classifier is used in our work.In this classifier, k-nearest neighbor classifier is taken as base classifier.The enhanced brownboost classifier combines the output of all classifier to produce strong classifier with higher classification accuracy.In our work, the number of pressure data is taken as training samples.
After that, k-nearest neighbor classifier classifies the pressure data as normal data or abnormal data through finding the similarity between the pressure data and threshold pressure data.After obtaining the results from all weak classifier, enhanced brownboost classifier boosts the results through combining all kNN classifiers to produce the strong classifier output.By this way, the classification of pressure data is carried out with higher accuracy.
Let us consider the number of instances as 10.Every instance the number of pressure data gets varied.When number of pressure data is taken as 300 for simulation work, EBBC-GWO Method gets 96% of classification accuracy while

Classification Time
In order to analyse the performance of EBBC-GWO Method, classification time metric is choosen for WDN while classifying the data as normal data and abnormal data.It is the difference of starting and ending time of classification.It is measured in terms of milliseconds (ms).It is formulated as After that classification time is calculated as,  Proposed EBBC-GWO Method: the starting time of classification is 0 and the ending time of classification is 36.
After that, the classification time is obtained as, When the classification time is lesser, method is more efficient.Table3 describes the value of classification time which is obtained from the simulation output.The classification time of three different techniques implemented in Matlab simulation with number of pressure data ranging from 50-500.The performance of classification time using EBBC-GWO Method is compared with the existing 1D-CNN-SVM model (Kang et al. 2018) and Multi-Stage Graph Partitioning Approach (Rajeswaran et al. 2018).From the table, it clear that the classification time using EBBC-GWO Method is lesser than other existing techniques (Kang et al., 2018), (Rajeswaran et al. 2018).Classification process categorizes the training data as the normal data or abnormal data for future use.An enhanced brownboost classifier is employed for the classification process in our research work.In this classifier, training samples are given for all k-nearest neighbor classifier at the same time to reduce the time consumption for classification.After that, the output from all the weak classifier is combined to produce the strong classifier.In our work, the number of pressure data is taken as input for all weak classifiers at the same time.After that, k-nearest neighbor classifier categorizes the pressure data as normal or abnormal data through discovering the extended jaccard similarity between the pressure data and threshold pressure data.After obtaining the results from all weak classifier, enhanced brownboost classifier improves the results through combining all kNN classifiers to form the strong classifier.By this way, the classification of pressure data is carried out with minimal amount of time.
Let us take the number of instances as 10.For every instance, the number of pressure data gets changed.When number of pressure data is taken as 200, EBBC-GWO Method classifies the pressure data in 52ms while 1D-CNN-SVM model (Kang et al. 2018) and Multi-Stage Graph Partitioning Approach (Rajeswaran et al. 2018) classifies the data in 79ms and 62ms respectively.The convergence time of optimal sensor placement method using EBBC-GWO Method is 1.3ms.The classification time consumption of EBBC-GWO Method is decreased by 41% and 20% compared to existing 1D-CNN-SVM model (Kang et al. 2018) and Multi-Stage Graph Partitioning Approach (Rajeswaran et al. 2018) respectively.

Water Leakage Detection Accuracy (WLDA)
Next, Water leakage detection accuracy metric is employed to the water leakage through placing the sensor at optimal nodes.It is measured in terms of percentage (%).It is measured by,

Sample calculation:
 1D-CNN-SVM model: the number of sensor nodes that detects leak accurately is 35 and the number of sensor placement nodes is 50.After that, water leakage detection accuracy is computed as,  Multi-Stage Graph Partitioning: the number of pressure data correctly classified is 39 and the total number of sensor placement nodes is 50.After that water leakage detection accuracy is calculated as,  Proposed EBBC-GWO Method: the number of pressure data correctly classified is 43 and the total number of sensor placement nodes is 50.After that, the water leakage detection accuracy is obtained as, When the water leakage detection accuracy is higher, the method is efficient.(Kang et al. 2018), (Rajeswaran et al. 2018).
Classification process classifies the training data as the normal data or abnormal data.After performing the classification process, optimization process is carried out to identify the optimal node.An enhanced brownboost classifier is employed for the classification process and glowworm swarm optimization is employed for choosing the optimal one in our research work.In this classifier, the number of pressure data is taken as input.K-nearest neighbor classifier classifies the pressure data as normal or abnormal data through finding the extended jaccard similarity between pressure data and threshold pressure data.Then, enhanced brownboost classifier combines all kNN classifiers output to form the strong classifier.After that, the abnormal pressure data node is taken and objective function is calculated for all data.Based on the brightness movement takes place and position gets updated.Finally, the optimal node is selected based on the brightness and water leakage gets detected.By this way, the water leakage gets detected with higher accuracy.
Let us take the number of instances as 10.For every instance, the number of sensor placement nodes gets changed.
When number of sensor placement data is taken as 400, EBBC-GWO Method obtains the 97% of water leakage detection accuracy while 1D-CNN-SVM model (Kang et al. 2018)

Sample calculation:
 1D-CNN-SVM model: the number of sensor nodes that incorrectly detects leak is 15 and the number of sensor placement nodes is 50.After that, false positive rate is calculated as,  Multi-Stage Graph Partitioning: the number of sensor nodes that incorrectly detects leak is 11 and the total 5 number of pressure data is 50.After that false positive rate is calculated as,  Proposed EBBC-GWO Method: the number of sensor nodes that incorrectly detects leak is 7 and the total number of pressure data is 50.After that, the false positive rate is obtained as, When the false positive rate is lesser, the method is said to be more efficient.(Kang et al. 2018), (Rajeswaran et al. 2018).Theses value are calculated from the results of simulation.
The localization of sensor is done by identifying optimal nodes in WDN.At first, pressure data is collected from the nodes.Then the classification process is carried out to classify the data as the normal data or abnormal data for performing the optimization process.An optimization process is carried out to identify the data as optimal one.An enhanced brownboost classifier is classifies the data and glowworm swarm optimization optimizes the data to detect the water leakage in our research work.Enhanced brownboost classifier combines all kNN classifiers results to form the strong classifier output.
Then, the abnormal pressure data is distributed and taken as glowworms.Besides, objective function (luciferin value) for each glowworm is computed where lesser brighter glowworm attracted by higher brighter glowworm and then updates its position.This helps to identify the optimal node for sensor localization.Finally, water leakage gets detected by placing the sensor at that particular node.By this way, the false positive rate gets reduced in our research work.
The key aim of EBBC-GWO method is to place the minimal number of sensors at optimal nodes for detecting the leakages.
For conducting the simulation experiment, number of instances taken are 10.

Leakage size
Lastly, leakage size is computed for analyzing the performance of water leakage detection.Leakage size is measured the amount water wasted through the time.Here, time is considered as 1 second.Leakage size is measured in terms of liters per second (lps). (21)

Sample calculation:
 1D-CNN-SVM model: the amount of water leakage is 15 lps and the time is considered as 1 sec.After that, leakage size is calculated as, lps  Multi-Stage Graph Partitioning: the amount of water leakage is 13 lps and the time is considered as 1 sec.After that, leakage size is calculated as, lps  Proposed EBBC-GWO Method: the amount of water leakage is 11 lps and the time is considered as 1 sec.After that, leakage size is calculated as, lps data.The number of pressure data is varied from 50-500 for performing simulation.The performance of leakage size is using proposed EBBC-GWO Method is compared with the existing 1D-CNN-SVM model (Kang et al. 2018) and Multi-Stage Graph Partitioning Approach (Rajeswaran et al. 2018).From the comparison, it is clear that the leakage size of EBBC-GWO Method is minimum than state-of-the-art methods (Kang et al. 2018), (Rajeswaran et al. 2018).
Enhanced BrownBoost Classifier model is employed to classify the pressure data is normal or abnormal through considering the k-Nearest Neighbor classifier as base classifier.This in turns, brownboost combines the weak learner to make the strong classifier.After that, optimization model is employed to find the optimal node for finding the sensor placement.From that, the leak in the water network is identified early using EBBC-GWO Method.For conducting the simulation experiment, 10 number of instances are considered.For each instance, the amount of water leakage is varied.
While considering number of pressure data is 250, the amount of water wasted in proposed EBBC-GWO Method is 11lps.

Conclusion
An EBBC-GWO Method is introduced for leakage detection in water distribution system.Enhanced BrownBoost Classifier model taken k-NN classifier as weak classifier and classifies training samples through majority vote of neighbor.After that, Brownboost classifier combines all k-NN classifier output to build the strong classifier.By this way, data are classified as the normal data or abnormal.After classification, glowworm swarm optimization process is carried out where every solution corresponds to abnormal pressure data node in search space.Every glowworm has objective function for addressing the optimization issue.Every glowworm employs probabilistic way to choose neighbor with higher objective function value and moves toward it.Glowworm updates its location to the glowworm in dynamic decision space and optimal one is selected for senor placement to detect the leakage.By this way, water leakage detection accuracy is improved with lesser false positive rate.The performance of EBBC-GWO method is carried out with the parameters such as classification accuracy, false positive rate, classification time, leakage size and water leakage detection accuracy.The results analysis of EBBC-GWO method improves classification accuracy and water leakage detection accuracy with minimum false positive rate and classification time than the state-of-art methods.

Figure 2
Figure 2 describes the enhanced BrownBoost Classification process with weak k-NN classifier for efficient classification.In enhanced BrownBoost classifier, k-NN base classifier is trained repeatedly with Pressure data.Enhanced BrownBoost classifier has n number of training data where represents the set of pressure data and denotes the final classifier output results.Let the threshold range of the pressure data be (i.e., ).In k-NN classifier, initially k number of classes is initialized for classification.After initialization, the similarity between the training pressure data and threshold pressure range are to be computed in kNN classifier by using Extended Jaccard Similarity Measure.It is computed by, explains the directed graph among set of eight glowworms depending on relative luciferin levels and accessibility of local information.
Every glowworm varies luciferin value consistent with the objective function value of current location.The luciferin update rule is given by, ), denotes the luciferin decay constant.symbolizes luciferin enhancement constant.symbolizes luciferin value.represent the objective function value at glowworm location at time t.From that, the objective function of water flow rate is calculated as follows,


1D-CNN-SVM model: the starting time of classification is 0 and the ending time of classification is 72.After that, classification time is calculated as,  Multi-Stage Graph Partitioning: the starting time of classification is 0 and the ending time of classification is 52.

Figure
Figure 6 Convergence and convergence speed of EBBC-GWO

Table 1 Values of parameters that are kept fixed in all the simulations
It consists of 883 nodes, 927 pipes which distributes water to 639 consumers.The pipe diameters varied from 70 to 400 mm.The pressure at night flow changes between 29.31m and 43.46 m.Flow and pressure are identified at inflow and outflow point.DMA includes 311 nodes with demand (RM type), 60l nodes without demand (EC type), 48 hydrant nodes without demand (HI type), 14 dummy valve nodes without demand (VT type) and 448 dummy nodes without demand (XX type) (R. Sarrate et al. 2014).The water pipeline has two inflow inputs (i.e., reservoir nodes).Leak detection is based on the damage (leaks) in different locations of piping network includes the liquid outflow at leak location that changes the flow features (pressure heads, flow rates, acoustics signals, etc.) at monitoring locations of piping network.It is imagined that leaks occur at XX type nodes in which 448 potential leaks are detected.Leaks occur at any node or pipe.Flow rate are employed for leak detection and gathering pressure data.In simulation, five leaks are identified.The size of the leakage was 142 meter, 0.99meter, 0.99meter, 12.31meter.The simulation settings of the constants in the optimization algorithm such as beta, gamma, rho values are fixed in all simulations is shown in Table1.

Table 2
describes classification accuracy of three different methods implemented in Matlab simulation by taking diverse number of pressure data ranging from 50-500.The values in the table 1 are obtained from the results of simulation using EPANET software.The simulation result of classification accuracy using EBBC-GWO Method is compared with existing 1D-CNN-SVM model (Kang et al. 2018) and Multi-Stage Graph Partitioning Approach

Table 4
(Rajeswaran et al. 2018)ge detection accuracy of three different techniques with respect to number of sensor placement nodes ranging from 50-500.The performance of water leakage detection accuracy using EBBC-GWO Method is compared with the existing 1D-CNN-SVM model(Kang et al. 2018) and Multi-Stage Graph Partitioning Approach(Rajeswaran et al. 2018).Table values are listed from the output results of simulation.From the table, it is clear that the water leakage detection accuracy of EBBC-GWO Method is higher than other existing techniques (Rajeswaran et al. 2018)titioningApproach(Rajeswaran et al. 2018) produces 81% and 91% of water leakage detection accuracy respectively.The water leakage detection accuracy of EBBC-GWO Method is increased by 21% and 9% compared to existing 1D-CNN-SVM model(Kang et al. 2018) and Multi-Stage Graph Partitioning Approach(Rajeswaran et al. 2018)respectively.In addition, the convergence and convergence speed of the EBBC-GWO Method is shown in below.It is moves toward the same point where it is meet.

Table 5 : Tabulation for False Positive Rate 10 Number of Sensor Placement Nodes (Number) False Positive Rate (%) 1D-CNN-SVM model Multi-Stage Graph
(Rajeswaran et al. 2018)nce of false positive rate using EBBC-GWO Method is compared with the existing 1D-CNN-SVM model(Kang et al. 2018) and Multi-Stage Graph Partitioning Approach(Rajeswaran et al. 2018).From the table, it is clear that the false positive rate of EBBC-GWO Method is lesser than other existing techniques (Kang et al. 2018e, the number of sensor placement nodes gets changed.When number of sensor placement data is taken as 350, EBBC-GWO Method attains the 3% of false positive rate while 1D-CNN-SVM model(Kang et al. 2018) and Multi-Stage Graph Partitioning Approach (Rajeswaran et al. 2018)produces 18% and 11% of false positive rate respectively.The false positive rate of EBBC-GWO Method is reduced by 69% and 54% compared to existing 1D-CNN-SVM model(Kang et al. 2018) and Multi-Stage Graph Partitioning Approach(Rajeswaran et al. 2018)respectively.

Table 6
illustrates the leakage size of proposed and existing methods with respect to different number of pressure