applied

sciences
Article
Prediction of Water Leakage in Pipeline Networks Using Graph
Convolutional Network Method
Ersin Şahin 1, * and Hüseyin Yüce 2

1 Computer Programming, Beykoz Vocational School, Beykoz University, 34820 Istanbul, Turkey
2 Mechatronics Engineering, Faculty of Technology, Marmara University, 34722 Istanbul, Turkey;
[email protected]
* Correspondence: [email protected]

Featured Application: Considering the theoretical contribution of the study to science, the use of
graphs in monitoring leaks in pipelines and the application of graph-based machine learning for
detection represent a novel approach in the literature. The datasets generated in this study will
be made available to other scientists, serving as a foundation for further research and offering
various benefits. When assessing the impact of this work on social life, it becomes crucial to
utilize water resources effectively and efficiently due to increasing demand resulting from both
global warming and urbanization. Ensuring the sustainability of our world heavily relies on
this aspect.

Abstract: This study aims to predict leaks in water-carrying pipelines by monitoring pressure drops.
Timely detection of leaks is crucial for prompt intervention and repair efforts. In this research, we
represent the network structure of pipelines using graph representations. Consequently, we propose
a machine learning model called Graph Convolutional Neural Network (GCN) that leverages graph-
type data structures for leak prediction. Conventional machine learning models often overlook the
dependencies between nodes and edges in graph structures, which are critical in complex systems
like pipelines. GCN offers an advantage in capturing the intricate relationships among connections in
pipelines. To assess the predictive performance of our proposed GCN model, we compare it against
the Support Vector Machine (SVM) model, a widely used traditional machine learning approach.
Citation: Şahin, E.; Yüce, H. In this study, we conducted experimental studies to collect the required pressure and flow data to
Prediction of Water Leakage in train the GCN and SVM models. The obtained results were visualized and analyzed to evaluate
Pipeline Networks Using Graph their respective performances. The GCN model achieved a performance rate of 94%, while the SVM
Convolutional Network Method. model achieved 87%. These results demonstrated the potential of the GCN model in accurately
Appl. Sci. 2023, 13, 7427. https://
detecting water leaks in pipeline systems. The findings hold significant implications for water
doi.org/10.3390/app13137427
resource management and environmental protection. The knowledge acquired from this study can
Academic Editor: José António serve as a foundation for predicting leaks in pipelines that transport gas and oil.
Correia
Keywords: graph convolutional network; graph machine learning; leakage detection
Received: 26 May 2023
Revised: 15 June 2023
Accepted: 19 June 2023
Published: 22 June 2023
1. Introduction
Due to the development of the industry, the need to transport water resources to
cities through pipelines has arisen during urbanization, both from dams and underground
Copyright: © 2023 by the authors. sources. The demand for water resources is increasing due to population growth, while
Licensee MDPI, Basel, Switzerland.
water resources are decreasing due to global warming. Given the increasing demand and
This article is an open access article
reduced resources, it is essential to transport water resources without loss. Open or closed
distributed under the terms and
pipeline networks are used to transport and distribute water resources. Failures in pipeline
conditions of the Creative Commons
networks are common and significant and a global problem. Failures in pipeline networks
Attribution (CC BY) license (https://
may occur in two ways: leaks and blockages. This study focuses on the detection of water
creativecommons.org/licenses/by/
4.0/).
leaks in pipeline networks. Pipeline leaks that arise from pressure or temperature changes,

Appl. Sci. 2023, 13, 7427. https://doi.org/10.3390/app13137427 https://www.mdpi.com/journal/applsci

Appl. Sci. 2023, 13, 7427 2 of 16

corrosion, wear and tear, and third-party damages [1] may lead to critical problems in
pipeline transportation systems. Accurately detecting water leaks in pipeline networks is
essential for avoiding the economic impacts of losses and protecting the environment [2].
Therefore, it is crucial to detect and repair pipeline network failures rapidly. Timely
and accurate intervention in pipeline failures is a critical decision-making problem. Two
main approaches are used to detect failures in pipeline networks: physical inspection
and mathematical model simulation. Physical methods are costly due to the production
stoppage, although they accurately identify the location and size of the failure. On the
other hand, the mathematical approach theoretically detects leaks and is, therefore, much
less expensive [3]. Various methods are used to detect leaks in pipeline networks. Some of
these methods include acoustic methods [4,5], negative pressure wave measurement [6,7],
measurement of input and output values of pressure and fluid velocity [8–10], vibration
analysis, distributed fiber optic sensing, infrared cameras, and lidar systems [11]. Many
machine learning algorithms based on convolutional neural networks and artificial neural
networks are used for leakage detection and classification [12]. Many studies have been
conducted to detect and localize pipeline system leaks [13,14]. A growing body of literature
on using machine learning applications for leak detection in pipelines [15,16]. However,
studies on machine learning methods for leak detection in pipelines are still limited and
have remained under-researched [17].
A literature review reveals that studies have been conducted on detecting leaks in
pipelines using machine learning methods, which can be categorized into five main ap-
proaches. Machine learning models employed in the studies in the literature are as follows:
1. For the negative pressure wave (NPW) method, a wireless sensor network-based
machine learning (WML) algorithm was utilized [18].
2. The acoustic-based model, including support vector machines (SVM) [19], neural
networks [20], artificial neural networks (ANNs), as well as SVM [21], and the combi-
nation of SVM with Relevance Vector Machine (RVM) [22], was employed for leak
detection methods.
3. The thermal infrared (IR) camera-based method utilized the immune neural net-
work [23] and convolutional neural network (CNN) approaches [24] for leak detection.
Some studies specifically focused on the application of these methods in petroleum
pipelines.
4. Pressure-monitoring methods employed the sparrow search algorithm (SPSA) and
CNN [25] for leak detection in petroleum pipelines, while a deep-learning method
called Deeppipe was also used [26]
5. The fluid transient waves method utilized artificial neural networks (ANNs) [27] and
a machine learning (ML)-based framework [28].
This academic literature review highlights the utilization of various machine learning
algorithms to detect leaks in pipelines, which demonstrates the successful application of
different methods and encourages further research to achieve improved results in leak
detection techniques. Using machine learning algorithms, accuracy rates for pipeline fault
diagnosis have been reported to range from 78.51% to 99%. Representing many complex
systems encountered in real life theoretically can be challenging. These complex systems
can be described using graph structures.
Many structures, such as airline, banking, social, medical, and supply chain structures,
can be expressed using graph structures. It is feasible to depict intricate systems through
graphs, such as distribution networks with multiple pipeline convergence or divergence
points. In the literature, pipeline networks have been represented using graphs, and the
spatial correlations of the pipeline network have been captured using the GCN model [29].
In this study, pipelines carrying water have been represented using graphs. The pressure
values at certain points on the pipeline were visualized over time by modeling and observ-
ing pressure drops. This study used the graph-based machine learning algorithm GCN to
detect pipeline leakage situations.
Appl. Sci. 2023, 13, 7427 3 of 16

Pipeline networks are often not limited to a single pipeline; significantly, in systems
like water distribution networks, there are structures where multiple pipelines converge
or diverge. Representing these complex structures with graph data structures provides
advantages compared to traditional data structures. Conventional machine learning models
(SVM, ANNs) often overlook the inherent dependencies among nodes and edges in pipeline
networks. However, it is crucial to acknowledge that a leakage occurrence in one pipeline
can have cascading effects on interconnected pipelines. In light of this, this study proposes
an approach that leverages graph data structures to effectively represent pipeline networks,
accompanied by the utilization of graph convolutional networks (GCN) as a graph-based
machine learning model. By incorporating GCN, the model can effectively capture and
contain the interdependencies between nodes and edges, thereby enhancing the accuracy
and predictive capabilities specifically tailored to the domain of pipeline networks. The
studies on GCN in the literature generally focus on object recognition and classification.
The GCN algorithm has achieved accuracy rates of 75% in urban pipeline networks [30],
94% in image processing [31], 88% in customer product recommendation [32], and 86% in
fault detection on steam turbines [33].
This study aims to predict leaks by monitoring pressure drops in water-carrying
pipelines, with a specific focus on accurately detecting pipeline leaks. In this research,
we represent the network structure of pipelines using graph representations. The graph
convolutional neural network (GCN) model is utilized. GCN offer an advantage in cap-
turing the intricate relationships among connections in pipelines. The data required for
the GCN algorithm were collected from a test set using experimental methods, wherein a
leakage scenario was deliberately created. Two datasets, edge and node, were specifically
constructed for this purpose. The performance of the GCN algorithm in leak detection was
compared with existing studies in the literature. The primary objective was to evaluate the
performance of GCN in graph-based machine learning and perform a comparative analysis
with traditional machine learning methods, explicitly employing the Support Vector Ma-
chine (SVM) algorithm as a reference. Given the escalating demand for water resources and
the concurrent decline in availability, preventing leaks in pipeline transportation systems
has become paramount. Detecting leaks facilitates prompt intervention, maintenance, and
repair, minimizing economic losses and mitigating environmental impact. The findings
of this study highlight the efficacy of the GCN algorithm in water leakage detection and
underscore the potential of graph-based machine learning in tackling intricate problems.

2. Materials and Methods

This study utilized the graph-based machine learning algorithms GCN and SVM.
These algorithms were implemented using StellarGraph, Pandas, Numpy, Sklearn, Ten-
sorflow, IPython, matplotlib, and Pyvis libraries based on Python. The performance
evaluations of the GCN and SVM algorithms were conducted using a confusion matrix,
accuracy, precision, recall, and f1 scoring methods. The GCN algorithms were implemented
using early-stopping methods to prevent overfitting.

2.1. Graph Convolutional Networks

Graph Convolutional Networks (GCN) is a current artificial neural network research
topic. The GCN model is derived from graph theory and convolution theorems to apply
machine learning to data represented by graphs. In general, a node is represented by com-
bining its attributes with the attributes of its neighbors [34]. GCN learns the representation
of a node by propagating neighbor information based on the graph structure. A GCN is
a multilayer neural network that operates directly on a graph and induces embedding
vectors of nodes based on the properties of their neighborhoods. Formally, consider a graph
G = (V, E) [35], where V and E are sets of nodes and edges, respectively [36].
Appl. Sci. 2023, 13, 7427 4 of 16

The GCN algorithm was proposed by Kipf et al. [37], and its working principle for a
single layer is expressed in Equation (1) below.
 1 1

e−2 A e − 2 X W = σ Â X W


Z=σ D eD (1)

e = A+I
A (2)
The expression Ae given in Equation (2) is the normalized adjacency matrix with self-
loops. The degree of the normalized adjacency matrix is denoted by e Dii in Equation (3).

e ii =
D ∑j Ae ij (3)

“σ” in Equation (1) represents the activation function. In this study, tanh and Softmax
functions were used as activation functions. “W” is the trainable weight matrix. “X”
represents the input features of nodes and edges in the graph structure (pressure and
flow rate values from sensors). “Z” is the output of the GCN layer containing first-degree
neighborhood information of all nodes. If k layers are used in the GCN algorithm, the
output of Z contains k-degree neighborhood (spatial) information. Therefore, the hidden-
layer data of GCN can provide more preliminary information for model training, enabling
trained hidden-layer neurons to have a more profound feature expression ability [38]. The
difference between GCN and classical machine learning algorithms lies in the ability of
GCN to directly operate on graph data structures and effectively utilize neighborhood
information to model the graph structure. With these capabilities, GCN can better capture
patterns and relationships in graph data, enabling more accurate predictions.
The application stages of the Graph Convolutional Networks (GCN) algorithm can be
defined in 11 steps:
1. Initialize the initial node feature matrix Z0 = X.
2. Determine the number of GCN layers (K) (16, 16).
3. Create weight matrices (Wk ) for each GCN layer with random initial values.
4. Define an activation function (σ = tanh).
5. Perform the graph convolution operation: Zk+1 = σ(D−1 ×A × Zk × Wk ).
6. Apply additional steps for each GCN layer, such as dropout or normalization.
7. Create the output layer: Zout = softmax(ZK × WK+1 ).
8. Define the loss function (L), typically cross-entropy.
9. Choose an optimization algorithm (Adam optimizer).
10. Update the weights using the training data: Wk , WK+1 = optimizer. minimize(L).
11. Make predictions on new data using the trained model: Zout_test = softmax(ZK ×
WK+1 ).
Zk represents the matrix representing the node features; Wk represents the weight ma-
trix, A represents the adjacency matrix representing the graph structure and D represents
the degree matrix. The algorithm updates the node features by utilizing the neighbor-
hood information from the graph structure and uses the softmax activation function for
classification. During the training stage, the weights are updated using an optimization
algorithm, and during the testing stage, the trained model makes predictions on new data.
These steps describe the implementation stages of the Graph Convolutional Networks
(GCN) algorithm, including initializing the node features, determining the number of
layers, creating weight matrices, applying the graph convolution operation, defining the
output layer, defining the loss function, choosing an optimization algorithm, updating the
weights during training, and making predictions on new data during testing. Table 1 below
gives the parameters of the GCN model used in this study.
Appl. Sci. 2023, 13, 7427 5 of 16

Table 1. Parameters of the GCN model.

GCN Model Parameters

Layer sizes (16, 16)
Epochs 70
Optimization method Adam (lr = 0.01)
Loss function Categorical cross-entropy
Activation functions Tanh, Softmax
Drop out 0.4%
Model train/rest rate 0.24%
Verbose 2

2.2. Support Vector Machines

Support Vector Machines (SVM) are an efficient alternative to supervised classifica-
tion [39]. The foundation of SVM was developed by Vapnik [40] and has gained acceptance
due to its various attractive features and promising performance. The formulation em-
bodies the structural risk minimization principle and is superior to conventional machine
learning methods’ traditional empirical risk minimization principle employed by conven-
tional machine learning methods. The SVM was initially developed to solve classification
problems [41]. During the training process, the SVM model was trained with the param-
eters test size 0.25 and random state 48. Table 2 below gives the parameters of the SVM
model used in this study.

Table 2. Parameters of the SVM model.

SVM Model Parameters

C 1.0
Break ties False
Cache size 200
Class weight None
Coef0 0.0
Decision function shape ovr
Degree 3
Gamma scale
Kernel RBF

2.3. Experimental Setup

This study conducted an experimental investigation to obtain the necessary datasets for
the proposed GCN and SVM algorithms, aiming to predict pipeline leaks. The experimental
design and implementation phase consisted of four main steps, which are provided below.
• Research Design
• Selection of Data Collection Methods
• Data Collection Process
• Data Recording
During the research design phase, the required data to be collected (such as pressure
and flow rate) were determined. The experimental dataset was prepared to be ready for
data collection following the created leak scenario, and the sensors were calibrated. The
limitations and constraints of this research were identified.
The limitations of this experimental research are as follows:
• The system is assumed to be represented by eight pressure sensors and two flow
meters.
• It is assumed that the experimental setup and measurement instruments are unaffected
by external factors, such as heat, vibration, noise, and light.
• It is assumed that the characteristics of the region where the experimental setup is
located, such as altitude, do not affect the precision of the measurements.
Appl. Sci. 2023, 13, 7427 6 of 16

• It is assumed that the PVC pipe type used in the experimental setup represents other
pipe types used in transmission lines.
The limitations of the experimental research are as follows:
• The findings are limited to the data set obtained from the experimental setup.
• The representation of liquid fluids is limited to water, used as the fluid in the experi-
mental setup.
• The pressure measurements are limited to the eight pressure sensors, two flow meters,
and PVC pipes used in the experimental setup.
An experimental setup was designed to create data sets required to operate graph-
based machine learning algorithms. A leakage scenario, which is detailed in the titled
scenario, was made to represent possible leakage situations. Pressure and flow variables
were measured over time through pressure sensors (MPS 500) and flowmeters using the
experimental setup shown in Figure 1 for data collection.

Figure 1. Experimental setup for pressure and flow measurement in the pipeline.

The data collection method and data recording processes were planned. During the
selection of data collection methods, the nature and volume of the data were considered to
determine how the data would be collected. In the data collection process, considerable
attention was paid to collecting the data accurately and consistently. A 12-channel data
acquisition card was used for recording the data. The data acquisition card records ten data
points per second. The recorded data were transferred to electronic spreadsheets, and the
accuracy and integrity of the data were checked. Any instances of incorrect or missing data
were identified and corrected.
For graph-based machine learning algorithms to work, the obtained data must be
represented by nodes and edges that connect these nodes. In the experimental dataset,
the pressure sensors represent the nodes, and the connections between these sensors
represent the edges. A total of eight pressure sensors were used in the dataset. Since the
Appl. Sci. 2023, 13, 7427 7 of 16

measurement intervals of the variables used in the dataset (such as flow rate and pressure)
differed, the dataset was standardized using the Python programming language and the
Sklearn and StandardScaler libraries. Additionally, the dataset was normalized using the
MinMaxScaler library, as artificial intelligence models perform more efficiently with values
ranging between 0 and 1. The experimental setup includes faucets and valves at various
locations for simulating leakage scenarios. Table 3 presents three columns: edge weight,
source (node), and destination (node). The table contains a total of 855 rows representing
the states of leakage (s3a), risk (s3o), and normal (s3n). The expression ‘s3n’ indicates the
absence of leakage or fault, ‘s3o’ indicates the presence of a leakage risk, and ‘s3a’ indicates
the presence of leakage or fault. Edge weights are assigned as 3 for leakage scenarios,
2 for risky scenarios, and 1 for normal scenarios. The expressions A0, A3, . . . , A8 in Table 3
represent the pressure sensors in the experimental dataset, which are the graph nodes.
The expressions _1, _2, _15 in Table 3 represent the values measured at different times
corresponding to the respective nodes.

Table 3. Edge list.

Source Target Edge Weight

s3n_A0_1 s3n_A3_1 1
s3n_A3_2 s3n_A8_2 1
s3o_A8_3 s3o_A9_3 2
s3o_A7_4 s3o_A9_4 2
s3a_A8_15 s3a_A9_15 3
s3a_A7_15 s3a_A8_15 3

Table 4 details 360 rows representing the nodes. The node list has 339 columns
representing the attributes of each node, such as pressure and flow rate. The node names
are encoded as follows: s3n for the case of no leakage, s3o for the risky situation, s3a for the
case of leakage, and A0_time for the zeroth node. For the leakage scenario, the node of the
A0 sensor belonging to the normal state at the first second is represented as s3n_A0_1.

Table 4. Node list.

Nodes Feature-1 ... Feature-339 System State

s1n_A0_1 1 ... 1 Normal
s1n_A2_1 0 ... 0 Normal
s1o_A2_1 0 ... 1 Risky
s1a_A2_1 0 ... 0 Leakage

3. Scenario
A scenario was created to represent leakage fault conditions in pipeline systems. Based
on this scenario, the data collection process was carried out using the experimental dataset
described above. Water was used as the fluid in the experimental dataset. Figure 2 shows
the graph and the location of the leakage for the leakage scenario. The graph nodes (A0, A2,
A3, A4, A6, A7, A8, A9) represent the pressure-measurement points in the experimental
dataset. Pressure measurements were taken at eight different points. In the leakage scenario,
the experimental dataset included four pipelines shown in green in Figure 2. Two different
pipe diameters, 1 mm and 3 mm, were used in the transmission lines for the leakage
scenario.
 Appl. Sci. 2023, 13, x FOR PEER REVIEW 8 of 17

Appl. Sci. 2023, 13, 7427 8 of 16

Figure 2. Two different pipe diameters, 1 mm and 3 mm, were used in the transmission
lines for the leakage scenario.

A2 A7

A3 A8

Leakage Point

A6
A0 A9
A4

Figure2.2.AAscenario
Figure scenariodiagram
diagram was created
created for
formeasuring
measuringthe
thepipeline’s
pipeline’spressure
pressure and
and flow
flow rate.
rate.

The
Theconditions
conditionsof ofthe
the created scenario are
created scenario areas
asfollows:
follows:
1.1. The
Thescenario
scenario was
was conducted
conducted at atroom
roomtemperature.
temperature.
2.2. Before
Beforedata
datacollection,
collection,the
theairairininthethe
pipes was
pipes purged
was purgedafter operating
after operatingthethe
experimental
experi-
mental setup.
setup.
3.3. Thesystem
The system was
was allowed
allowed toto reach
reachaastablestablestate,
state,and
andthe
thepipes
pipeswere
were completely
completelyfilled
filled
withwater
with water before
before data collection.
collection.
4.4. Thepressure
The pressure sensors
sensors were
were calibrated.
calibrated.
5.5. Considering the structure of water
Considering the structure water distribution
distributionnetworks,
networks,pipes
pipeswith
with different diam-
different diame-
eters and heights were selected for the scenario. The location of the leakage
ters and heights were selected for the scenario. The location of the leakage point point waswas
chosenso
chosen so that
that measurements
measurements could couldbe betaken
takenboth
bothbefore
beforeand after
and thethe
after leakage point
leakage point
using sensors.
using sensors.
The graph structure depicted in Figure 3 represents the pressure sensors and PVC
water pipes in our experimental setup. The pressure sensors are the graph nodes, while
the water pipes carry the edge attributes. The graph is directed, as there is a flow direction
within the pipelines.
 Appl. Sci. 2023, 13, x FOR PEER REVIEW 9 of 17

The graph structure depicted in Figure 3 represents the pressure sensors and PVC
Appl. Sci. 2023, 13, 7427 water pipes in our experimental setup. The pressure sensors are the graph nodes, 9while of 16
the water pipes carry the edge attributes. The graph is directed, as there is a flow direction
within the pipelines.

Figure3.3.Graph
Figure Graph representation
representation of
of the
thescenario
scenariocreated
createdfor
forpressure
pressureand
andflow
flowmeasurement
measurementin
inthe
the
pipeline.
pipeline.

The
Theexpressions
expressionsA0,
A0,A2,
A2,A3,
A3,A4,
A4,A6, A7,
A6, A8,
A7, A8,andandA9A9
depicted in Figure
depicted 3 represent
in Figure the
3 represent
pressure sensors
the pressure present
sensors in the in
present experimental dataset dataset
the experimental given ingiven
Figurein1.Figure
When 1. theWhen
pipeline
the
network in the experimental dataset is represented as a graph structure, the
pipeline network in the experimental dataset is represented as a graph structure, thepoints A0, A2,
A3, A4,A0,
points A6,A2,
A7,A3,
A8,A4,
and A9A7,
A6, serve
A8,asandtheA9
nodes
serveofasthe
thegraph
nodesstructure.
of the graphThese node points
structure. These
are recorded
node points by
aremeasuring
recorded bythemeasuring
pressure values over time
the pressure to create
values a dataset.
over time to create a dataset.

4.4.Results
Results
Data
Datacollected
collectedwith
witheight different
eight differentpressure sensors
pressure sensorsandand
flowflow
meters on the
meters on experimen-
the experi-
tal
mental setup were labeled for normal, risky, and leakage scenarios for a 10-s interval
setup were labeled for normal, risky, and leakage scenarios for a 10-s time and
time interval
sampled in Table 5. The samples of risk and leakage are significantly different
and sampled in Table 5. The samples of risk and leakage are significantly different from from most
other
most samples regarding
other samples characteristics
regarding [42]. The
characteristics data
[42]. Theindata
Tablein5Table
are visualized as a graph
5 are visualized as a
in Figures 4 and 5.
graph in Figures 4 and 5.
Each color seen in Figures 4 and 5 represents an individual pressure sensor within
the experimental dataset, and their corresponding mappings are as follows: A0 (blue), A4
(red), A6 (orange), A9 (yellow), A3 (green), A2 (violet), A7 (brown), A8 (turquoise).
Using graphs, Figure 4 displays the experiment data (Table 5) representing normal,
leak, and risky scenarios.
The central nodes represented in black show the time information. The colored
nodes connected to the black nodes by edges, A0, A2, . . . and A9 represent the pressure
information obtained from the sensors. The radii of the nodes vary according to the
measured pressure values. In leak scenarios, i.e., in the first, seventh, and eighth central
nodes, the node sizes are very small, as the pressure values are relatively lower than
the normal scenarios. In risky scenarios, i.e., in the fourth and sixth central nodes, it is
observed that the nodes shrink compared to normal scenarios. Figure 5 below illustrates
the clustering of similar nodes (based on node and edge characteristics) using the Python
NetworkX library for the values provided in Table 5, where the nodes are clustered closely
together.
 Appl. Sci. 2023, 13, 7427 10 of 16

Table 5. (0–9) second experiment set data labeled as Risky, Normal, and Leakage (flow rate, pressure
[mA]).

Time A0 A2 A3 A4 A6 A7 A8 A9 Flow 1 Flow 2 System State

0 8.5852 7.7438 8.0905 7.9858 8.1753 9.9996 7.0367 9.9996 0.6433 0.6253 Normal
1 2.4679 2.7862 2.7090 2.7395 2.6995 4.5739 3.1051 4.9981 0.6365 0.6253 Leakage
2 8.6276 7.7267 8.0307 7.9724 8.1848 9.9996 7.0370 9.9996 0.6552 0.6298 Normal
3 8.6093 7.7267 8.0447 7.9504 8.1417 9.9996 6.9976 9.9996 0.6799 0.6179 Normal
4 5.0442 4.8980 4.9563 4.9505 4.9908 7.0755 4.7476 7.6266 0.6015 0.6021 Risky
5 8.5580 7.6980 7.9895 7.9409 8.1707 9.9996 7.0004 9.9996 0.6689 0.6253 Normal
6 5.0234 4.9099 4.9774 4.9642 5.0503 7.0767 4.7598 7.6596 0.6021 0.6024 Risky
7 2.5045 2.8204 2.7096 2.7218 2.7200 4.5855 3.1057 4.9801 0.6436 0.6253 Leakage
8 2.4459 2.8164 2.7236 2.7355 2.7081 4.5532 3.0987 4.9758 0.6668 0.6295 Leakage
Appl. Sci. 2023, 13, x FOR PEER REVIEW 10 of 17
9 8.5006 7.6895 8.0026 7.9452 8.1314 9.9996 6.9760 9.9996 0.6536 0.6195 Normal

Figure 4. Graphical
Figure representation
4. Graphical overover
representation timetime
of the
ofdata provided
the data in Table
provided 5. 5.
in Table

As shown in Figure 5, the sensor values obtained under normal conditions are clus-
tered in a different area than those obtained during leaks and risky situations. The nodes in
Figure 5 are numbered 1, 2, 3, etc. For each period in Table 5, eight sensor or node values
are generated. For time 0, the node numbers are 1, 2, 3, . . . , 8, and for time 1, the nodes
continue from 9, 10, 11, . . . , 16. The leakage situations in Table 5 are given at times 1, 7,
and 8. It can be observed that the sizes of the nodes corresponding to these periods, which
fall between nodes 9 and 16, 57 and 64, and 65 and 72, have shrunk to a size that is too
small to be observed. The parameters related to the performance of the GCN algorithm are
determined as an accuracy score of 0.9420, a loss value of 0.0980, and an f1 score of 0.94.
The summary of the classification of scenarios according to the GCN algorithm is given in
Table 6.
Appl. Sci. 2023, 13, 7427 11 of 16

Figure 4. Graphical representation over time of the data provided in Table 5.

Figure Figure 5. A clustered

5. A clustered graphical
graphical representation of
representation ofthe
thepressure values
pressure of theof
values sensors in Table 5,
the sensors ingrouped
Table 5,
according to their similarities.
grouped according to their similarities.

Table 6. Summary of classification of scenarios according to the GCN algorithm.

Each color seen in Figures 4 and 5 represents an individual pressure sensor within
the experimental dataset, and
System Status their corresponding
Precision Recall mappingsf1-Score
are as follows: A0 (blue), A4
Sample
(red), A6 (orange),
Normal A9 (yellow), A3
1.00 (green), A2 (violet),
0.82 A7 (brown),
1.00 A8 (turquoise).
120
Risky 1.00 1.00 0.94 120
Leakage 0.90 0.89 0.95 120
Accuracy 0.94 360
Macro Average 0.97 0.96 0.96 360
Weighted
0.97 0.96 0.96 360
Average

For the leakage situations in Table 6, the f1-score is detected as 95%, the detection
of risky situations with 94% accuracy, and the detection of normal situations with 100%
accuracy. The system’s average leakage, risk, and normal status are detected with a 94.20%
accuracy rate. The real-time monitoring of leakage situations in pipeline networks and the
detection of leakage situations occur at the moment of leakage. This situation causes a waste
of time and water resources during the intervention phase of leakages. It is crucial to detect
leakage situations in pipeline networks at an early stage. The system status being labeled
as risky indicates the need for intervention in case of leaks. The prediction of a possible
significant leakage situation is made in advance. Small interventions or maintenance and
repair can be addressed at this stage before the failure escalates. The graph of the accuracy
and loss values of the GCN algorithm is given in Figure 6.
 waste of time and water resources during the intervention phase of leakages. It is crucial
to detect leakage situations in pipeline networks at an early stage. The system status being
labeled as risky indicates the need for intervention in case of leaks. The prediction of a
possible significant leakage situation is made in advance. Small interventions or mainte-
Appl. Sci. 2023, 13, 7427 12 of 16
nance and repair can be addressed at this stage before the failure escalates. The graph of
the accuracy and loss values of the GCN algorithm is given in Figure 6.

Figure 6. Accuracy and loss graph for the GCN algorithm.

The classification
The classificationofofthe
thepredicted
predicted values
values of the
of the GCN GCN algorithm
algorithm forsystem’s
for the the system’s
risky,
leakage and normal states is shown in Figure 7, which was created using the the
risky, leakage and normal states is shown in Figure 7, which was created using t-SNE
t-SNE (t-
(t-Distributed
Distributed Stochastic
Stochastic Neighbor
Neighbor Embedding)
Embedding) technique,
technique, which
which is is employed
employed for
for visual-
visualiz-
Appl. Sci. 2023, 13, x FOR PEER REVIEW 13 of 17
izing
ing high-dimensional
high-dimensional data
data byby assigning
assigning each
each datadata point
point a location
a location on on a two-
a two- or three-
or three-di-
dimensional
mensional map
map [43].
[43].

Figure
Figure 7.
7. Classification
Classification of
of the
the system’s
system’s risky, leakage, and
risky, leakage, and normal
normal states
states with
with GCN.
GCN.

The X 1 and X 2 axes in Figure 7 represent the coordinates of the two-dimensional

space generated by the t-SNE algorithm. The positions of the points on the graph are de-
termined based on the similarities among the data points. The nodes labeled 1 and 2
within the circles in Figure 7 show the confusion of the algorithm in predicting leakage
and risky situations, indicating that the nodes belonging to these situations have similar
Appl. Sci. 2023, 13, 7427 Figure 7. Classification of the system’s risky, leakage, and normal states with GCN.13 of 16

The X1 and X2 axes in Figure 7 represent the coordinates of the two-dimensional spac
The X1 by
generated andthe
X2 t-SNE
axes in algorithm.
Figure 7 represent the coordinates
The positions of the two-dimensional
of the points on the graph arespace
determined
generated by the t-SNE algorithm. The positions of the points on the graph are determined
based on the similarities among the data points. The nodes labeled 1 and 2 within th
based on the similarities among the data points. The nodes labeled 1 and 2 within the circles
circles in Figure 7 show the confusion of the algorithm in predicting leakage and risky
in Figure 7 show the confusion of the algorithm in predicting leakage and risky situations,
situations, indicating that the nodes belonging to these situations have simila
indicating that the nodes belonging to these situations have similar characteristics. The
characteristics.
confusion The confusion
matrix comparing matrix
the actual andcomparing the actual
predicted values and algorithm
of the GCN predictedfor
values
the of th
GCN algorithm for the system’s risky, leakage, and normal
system’s risky, leakage, and normal states is shown in Figure 8. states is shown in Figure 8.

Figure8.8.Confusion
Figure Confusionmatrix according
matrix to GCN
according results.
to GCN results.
The matrix compares the actual states with the predicted states obtained from applying
Thealgorithm
the GCN matrix compares
in Figure the actual
8. The statesshowed
findings with the predicted
that states
the system obtained
confused from apply
the risky
ing the GCN algorithm in Figure 8. The findings showed that the
situations with the leakage situations. The performance of the Graph Convolutional system confused th
risky situations
Network with the leakage
(GCN) algorithm, used forsituations. The performance
machine learning of the Graph
on graph-structured Convolutiona
data, was
Networkby
evaluated (GCN) algorithm,
employing usedmachine
a classical for machine learning
learning on graph-structured
algorithm, data, was eval
Support Vector Machine
(SVM),
uated for
by predicting
employing leakages in pipeline
a classical systems
machine using algorithm,
learning the same scenario
SupportandVector
dataset.Machin
The parameters
(SVM), related to
for predicting the performance
leakages in pipelineof the SVM algorithm
systems using theare
samedetermined
scenarioasand
an dataset
accuracy score of 0.8666, a loss value of 0.1333, and an f1 score of 0.8667. The summary of
The parameters related to the performance of the SVM algorithm are determined as an
the classification of scenarios according to the SVM algorithm is given in Table 7.
accuracy score of 0.8666, a loss value of 0.1333, and an f1 score of 0.8667. The summary o
the classification
Table of scenariosofaccording
7. Summary of classification to the SVM
scenarios according algorithm
to the SVM is given in Table 7.
algorithm.

System
Table Status
7. Summary Precision
of classification Recallaccording to
of scenarios f1-Score Sample
the SVM algorithm.
Normal 1.00 1.00 1.00 25
System
Risky Status 0.87 Precision 0.77 Recall 0.82 f1-Score 35 Sample
Normal
Leakage 0.76 1.00 0.87 1.00 0.81 1.00 30 25
Accuracy 0.87 90
Risky 0.87 0.77 0.82 35
Macro Average 0.88 0.87 0.87 90

For the leakage situations in Table 7, the f1-score is detected as 81%, the detection
of risky situations with 82% accuracy, and the detection of normal situations with 100%
accuracy. The system’s average leakage, risk, and normal status are detected with an 87.66%
accuracy rate. The SVM algorithm, which solely relies on node features and does not
utilize edge features, has fewer samples than the GCN algorithm. The confusion matrix
comparing the actual and predicted values of the SVM algorithm for the system’s risky,
leakage, and normal states is shown in Figure 9.
 risky situations with 82% accuracy, and the detection of normal situations with 100% ac
curacy. The system’s average leakage, risk, and normal status are detected with an 87.66%
accuracy rate. The SVM algorithm, which solely relies on node features and does not uti
lize edge features, has fewer samples than the GCN algorithm. The confusion matrix
Appl. Sci. 2023, 13, 7427 comparing the actual and predicted values of the SVM algorithm for the system’s
14 of 16 risky
leakage, and normal states is shown in Figure 9.

Figure9.9.Confusion
Figure Confusionmatrix
matrixaccording to SVM
according results.
to SVM results.
The matrix compares the actual states with the predicted states obtained from applying
The matrix compares the actual states with the predicted states obtained from apply
the SVM algorithm in Figure 9. The findings showed that the system confused the risky
ing the SVM
situations algorithm
with the in Figure 9. The findings showed that the system confused the
leakage situations.
risky situations with the leakage situations.
5. Conclusions
5. Conclusions
When the literature was reviewed, it was found that machine learning algorithms
have been
When utilized for fault diagnosis
the literature in pipeline
was reviewed, systems,
it was foundyielding accuracy rates
that machine ranging
learning algorithm
from 78.51% to 99%. However, it is worth noting that non-graph-based machine learning
have been utilized for fault diagnosis in pipeline systems, yielding accuracy rates ranging
models, such as SVM, ANNs, RVM, CNN, SPSA, and WML, have been predominantly
from 78.51% to 99%. However, it is worth noting that non-graph-based machine learning
employed instead of graph-based machine learning models for predicting and detecting
models,
leaks such as SVM, ANNs, RVM, CNN, SPSA, and WML, have been predominantly
in pipelines.
employed instead
In this study, theofGCN
graph-based machine
model achieved learningaccuracy
a detection models of for95%
predicting and detecting
for leak condi-
leaks94%
tions, in pipelines.
for risky situations, and 100% for normal situations. The SVM (Support Vector
Machine) model
In this was used
study, as a reference
the GCN to determine
model achieved the GCN model’s
a detection accuracy effectiveness.
of 95% forTheleak condi
SVM model achieved a detection accuracy of 81% for leaks, 82% for risky
tions, 94% for risky situations, and 100% for normal situations. The SVM (Support conditions, and Vecto
100% for normal
Machine) model conditions.
was usedOverall, when comparing
as a reference the performance
to determine of the two
the GCN model’s models
effectiveness. Th
(considering leaks, risky conditions, and normal conditions), SVM achieved an accuracy of
SVM model achieved a detection accuracy of 81% for leaks, 82% for risky conditions, and
87%, while GCN achieved 94% accuracy.
100% for normal conditions. Overall, when comparing the performance of the two model
Based on the findings, the following conclusions have been drawn:
(considering leaks, risky conditions, and normal conditions), SVM achieved an accuracy
1. The results reveal important implications for water resource management, reduction
of 87%, while GCN achieved 94% accuracy.
of water losses and sustainable living.
2. Based
The GCNonmodel
the findings, the following
demonstrated conclusions
high accuracy have been
and performance drawn: pipeline
in predicting
1. water leaks, including
The results leaks andimplications
reveal important risky conditions.
for The
waterresults highlight
resource the effective-
management, reduction
ness of the graph-based machine learning
of water losses and sustainable living. model. The high accuracy of the GCN
2. model
The GCN in detecting risky conditionshigh
model demonstrated is crucial for early
accuracy andleak detection and
performance preventive pipeline
in predicting
maintenance planning. Situations labeled as risky indicate the need to act to prevent
water leaks, including leaks and risky conditions. The results highlight the effective
leaks. Thus, predicting a potential major leak can be made in advance. At this stage,
ness of the graph-based machine learning model. The high accuracy of the GCN
it is aimed to solve the problem before it grows by performing small interventions or
model in detecting
maintenance risky conditions is crucial for early leak detection and preventive
and repair.
3. maintenance
Both GCN and SVMplanning.
modelsSituations
exhibited labeled as risky
equally good indicatein
performance the need tonormal
detecting act to preven
conditions.
4. The GCN model outperformed the SVM model when comparing the prediction
results for leaks and risky conditions. The GCN algorithm’s ability to analyze edge
and spatial relationships made a difference in leak detection.
The results indicate the potential of graph-based machine learning methods in solving
complex problems such as detecting water leaks in pipelines. Based on the experience
gained from this study, it is aimed to replicate this study with other graph-based machine
Appl. Sci. 2023, 13, 7427 15 of 16

learning algorithms, such as GraphSAGE, HinSAGE, RGCN GAT, SGC, PPNP, APPNP,
and Cluster-GCN, given the detection of leak locations, and to predict leak and blockage
situations in pipeline systems with gaseous fluids in further studies.

Author Contributions: Conceptualization, E.Ş.; formal analysis, H.Y.; investigation, E.Ş.; methodol-
ogy, E.Ş.; project administration, H.Y.; writing—original draft, E.Ş.; writing—review and editing, H.Y.
and E.Ş. All authors have read and agreed to the published version of the manuscript.
Funding: This study was supported by a scientific research project numbered “FDK-2023-10459” by
Marmara University within the scope of the doctoral thesis.
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Data Availability Statement: The data that support the findings of this study are available upon
reasonable request from the corresponding author.
Acknowledgments: The experimental stages of this study were conducted using the experimental
setup developed for this study at the Fault Diagnosis Laboratory of Marmara University, Faculty of
Technology, Mechatronics Engineering Department in Turkey. I am indebted to Marmara University
for providing the facilities to conduct the present study.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Wang, W.; Sun, H.; Guo, J.; Lao, L.; Wu, S.; Zhang, J. Experimental study on water pipeline leak using in pipe acoustic signal
analysis and artificial neural network prediction. Measurement 2021, 186, 110094. [CrossRef]
2. Anfinsen, H.; Aamo, O.M. Leak detection, size estimation and localization in branched pipe flows. Automatica 2022, 140, 110213.
[CrossRef]
3. Shivananju, B.N.; Kiran, M.; Nithin, S.P.; Vidya, M.J.; Hegde, G.M.; Asokan, S. Real-time monitoring of petroleum leakage
detection using etched fiber Bragg grating. In International Conference on Optics in Precision Engineering and Nanotechnology
(icOPEN2013); SPIE: Bellingham, WA, USA, 2013; Volume 8769.
4. Zeng, W.; Do, N.; Lambert, M.; Gong, J.; Cazzolato, B.; Stephens, M. Linear phase detector for detecting multiple leaks in water
pipes. Appl. Acoust. 2023, 202, 109152. [CrossRef]
5. Datta, S.; Sarkar, S. A review on different pipeline fault detection methods. J. Loss Prev. Process Ind. 2016, 41, 97–106. [CrossRef]
6. Liang, L.; Feng, K.; Xu, G.; Zhu, Z.; Zhou, X. Pipeline leakage test based on FBG pressure sensor. In IOP Conference Series: Earth
and Environmental Science; IOP Publishing: Bristol, UK, 2018; Volume 170.
7. Yu, D.; Zhang, L.; Wei, L.; Wang, Z. Projective Noise Reduction Algorithm for Negative Pressure Wave Signal Processing. In
Proceedings of the 2008 7th International Pipeline Conference, Calgary, AB, Canada, 29 September–3 October 2008; Volume 48579.
8. Bai, L.; Yue, Q.; Li, H. Sub-sea Pipelines Leak Detection and Location Based on Fluid Transient and FDI. In Proceedings of the
Fourteenth International Offshore and Polar Engineering Conference, Toulon, France, 23–28 May 2004; OnePetro: Richardson, TX,
USA, 2004.
9. Ling, K.; He, J. A new method for leak detection in gas pipelines. Oil Gas Facil. 2015, 4, 97–106. [CrossRef]
10. Reddy, R.S.; Payal, G.; Karkulali, P.; Himanshu, M.; Ukil, A.; Dauwels, J. Pressure and flow variation in gas distribution pipeline
for leak detection. In Proceedings of the 2016 IEEE International Conference on Industrial Technology (ICIT), Taipei, Taiwan,
14–17 March 2016.
11. Korlapati, N.V.S.; Khan, F.; Noor, Q.; Mirza, S.; Vaddiraju, S. Review and analysis of pipeline leak detection methods. J. Pipeline
Sci. Eng. 2022, 2, 100074. [CrossRef]
12. Goni, M.F.; Nahiduzzaman; Anower, M.; Rahman, M.; Islam, M.; Ahsan, M.; Haider, J.; Shahjalal, M. Fast and Accurate Fault
Detection and Classification in Transmission Lines using Extreme Learning Machine. E-Prime-Adv. Electr. Eng. Electron. Energy
2023, 3, 100107.
13. Mandal, S.K.; Chan, F.T.; Tiwari, M. Leak detection of pipeline: An integrated approach of rough set theory and artificial bee
colony trained SVM. Expert Syst. Appl. 2012, 39, 3071–3080. [CrossRef]
14. Ayati, A.H.; Haghighi, A.; Ghafouri, H.R. Machine Learning–Assisted Model for Leak Detection in Water Distribution Networks
Using Hydraulic Transient Flows. J. Water Resour. Plan. Manag. 2022, 148, 04021104. [CrossRef]
15. Eastvedt, D.; Naterer, G.; Duan, X. Detection of faults in subsea pipelines by flow monitoring with regression supervised machine
learning. Process Saf. Environ. Prot. 2022, 161, 409–420. [CrossRef]
16. Ahn, B.; Kim, J.; Choi, B. Artificial intelligence-based machine learning considering flow and temperature of the pipeline for leak
early detection using acoustic emission. Eng. Fract. Mech. 2019, 210, 381–392. [CrossRef]
Appl. Sci. 2023, 13, 7427 16 of 16

17. Miao, X.; Zhao, H.; Xiang, Z. Leakage detection in the natural gas pipeline based on unsupervised learning and stress perception.
Process. Saf. Environ. Prot. 2023, 170, 76–88. [CrossRef]
18. Rashid, S.; Akram, U.; Khan, S.A. WML: Wireless sensor network-based machine learning for leakage detection and size
estimation. Procedia Comput. Sci. 2015, 63, 171–176. [CrossRef]
19. Xu, T.; Zeng, Z.; Huang, X.; Li, J.; Feng, H. Pipeline leak detection based on variational mode decomposition and support vector
machine using an interior spherical detector. Process. Saf. Environ. Prot. 2021, 153, 167–177. [CrossRef]
20. Noseda, F.; Marnet, L.R.; Carlim, C.; Costa, L.R.; de Moura, N., Jr.; Calôba, L.P.; Soares, S.D.; Clarke, T.; Jacques, R.C. A Neural
Network System for Fault Prediction in Pipelines by Acoustic Emission Techniques. Res. Nondestruct. Eval. 2021, 32, 132–146.
[CrossRef]
21. Baronti, L.; Castellani, M.; Hefft, D.; Alberini, F. Neural network identification of water pipe blockage from smart embedded
passive acoustic measurements. Can. J. Chem. Eng. 2022, 100, 521–539. [CrossRef]
22. Xiao, R.; Li, J. Evaluation of acoustic techniques for leak detection in a complex low-pressure gas pipeline network. Eng. Fail.
Anal. 2023, 143, 106897. [CrossRef]
23. Banjara, N.K.; Sasmal, S.; Voggu, S. Machine learning supported acoustic emission technique for leakage detection in pipelines.
Int. J. Press. Vessel. Pip. 2020, 188, 104243. [CrossRef]
24. Yu, X.; Tian, X. A fault detection algorithm for pipeline insulation layer based on immune neural network. Int. J. Press. Vessel. Pip.
2022, 196, 104611. [CrossRef]
25. De Kerf, T.; Gladines, J.; Sels, S.; Vanlanduit, S. Oil spill detection using machine learning and infrared images. Remote Sens. 2020,
12, 4090. [CrossRef]
26. Li, Q.; Shi, Y.; Lin, R.; Qiao, W.; Ba, W. A novel oil pipeline leakage detection method based on the sparrow search algorithm and
CNN. Measurement 2022, 204, 112122. [CrossRef]
27. Zheng, J.; Wang, C.; Liang, Y.; Liao, Q.; Li, Z.; Wang, B. Deeppipe: A deep-learning method for anomaly detection of multi-product
pipelines. Energy 2022, 259, 125025. [CrossRef]
28. Bohorquez, J.; Alexander, B.; Simpson, A.R.; Lambert, M.F. Leak detection and topology identification in pipelines using fluid
transients and artificial neural Networks. J. Water Resour. Plan. Manag. 2020, 146, 04020040. [CrossRef]
29. Asghari, V.; Kazemi, M.H.; Duan, H.-F.; Hsu, S.-C.; Keramat, A. Machine learning modeling for spectral transient-based leak
detection. Autom. Constr. 2023, 146, 104686. [CrossRef]
30. Liao, Z.; Liu, M.; Du, B.; Zhou, H.; Li, L. A temporal and spatial prediction method for urban pipeline network based on deep
learning. Phys. A Stat. Mech. Its Appl. 2022, 608, 128299. [CrossRef]
31. Edwards, M.; Xie, X. Graph-based convolutional neural network. arXiv 2016, arXiv:1609.08965.
32. Shafqat, W.; Byun, Y.-C. Incorporating similarity measures to optimize graph Convolutional neural network for a product
recommendation. Appl. Sci. 2021, 11, 1366. [CrossRef]
33. Zhang, Y.-J.; Hu, L.-S. Fault Propagation Inference Based on a Graph Neural Network for Steam Turbine Systems. Energies 2021,
14, 309. [CrossRef]
34. Wu, Z.; Pan, S.; Chen, F.; Long, G.; Zhang, C.; Philip, S.Y. A comprehensive survey on graph neural networks. IEEE Trans. Neural
Netw. Learn. Syst. 2020, 32, 4–24. [CrossRef]
35. Toulouse, M.; Dai, H.K.; Le, T.G. Distributed load-balancing for account-based sharded blockchains. Int. J. Web Inf. Syst, 2022;
ahead of print.
36. Yao, L.; Mao, C.; Luo, Y. Graph Convolutional Networks for Text Classification. In Proceedings of the AAAI Conference on
Artificial Intelligence, Honolulu, HI, USA, 27 January 27–1 February 2019; Volume 33, pp. 7370–7377.
37. Kipf, T.N.; Welling, M. Semi-supervised classification with graph convolutional networks. arXiv 2016, arXiv:1609.02907.
38. Tong, H.; Qiu, R.C.; Zhang, D.; Yang, H.; Ding, Q.; Shi, X. Detection and classification of transmission line transient faults based
on graph convolutional neural Network. CSEE J. Power Energy Syst. 2021, 7, 456–471.
39. Valero-Carreras, D.; Alcaraz, J.; Landete, M. Comparing two SVM models through different metrics based on the confusion
matrix. Comput. Oper. Res. 2023, 152, 106131. [CrossRef]
40. Vapnik, V. The Nature of Statistical Learning Theory; Springer Science & Business Media: Berlin/Heidelberg, Germany, 1999.
41. Roy, A.; Chakraborty, S. Support vector machine in structural reliability analysis: A review. Reliab. Eng. Syst. Saf. 2023, 233, 109126.
[CrossRef]
42. Gao, H.; Qiu, B.; Barroso, R.J.D.; Hussain, W.; Xu, Y.; Wang, X. TSMAE: A novel anomaly detection approach for the internet of
things time series data using memory-augmented autoencoder. IEEE Trans. Netw. Sci. Eng. 2022. [CrossRef]
43. Van der Maaten, L.; Hinton, G. Visualizing data using t-SNE. J. Mach. Learn. Res. 2008, 9, 2579–2605.

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.