Content uploaded by Ulaş Yurtsever
Author content
All content in this area was uploaded by Ulaş Yurtsever on Apr 10, 2024
Content may be subject to copyright.
Alexandria Engineering Journal 96 (2024) 112–123
Available online 8 April 2024
1110-0168/© 2024 The Author(s). Published by Elsevier BV on behalf of Faculty of Engineering, Alexandria University This is an open access article under the CC
BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
Enhancing electrical panel anomaly detection for predictive maintenance
with machine learning and IoT
Muhammed Fatih Peks
¸en
a
,
*
, Ulas
¸ Yurtsever
b
,
c
, Yılmaz Uyaro˘
glu
d
a
Vocational School of Adapazarı, Sakarya University, Sakarya 54050, Turkey
b
Department of Computer Engineering, Sakarya University, Sakarya 54050, Turkey
c
Interdisciplinary Articial Intelligence Lab., Research Development and Application Center (SARGEM), Sakarya University, Sakarya 54050, Turkey
d
Department of Electrical and Electronics Engineering, Sakarya University, Sakarya 54050, Turkey
ARTICLE INFO
Keywords:
Articial intelligence
Machine learning
IoT
Fire
Predictive maintenance
ABSTRACT
This study aims to detect electrical panel res using the Internet of Things (IoT) framework and Machine
Learning (ML) algorithms. Within the scope of the study, an experimental process was carried out using Arduino
and Raspberry Pi platforms to collect essential data such as gas, temperature, and humidity. In this experimental
work conducted on 3478 data points to detect electrical panel res, Decision Tree (DT), Gaussian Naive Bayes
(GNB), Gaussian Process Classier (GPC), and Support Vector Machine (SVM) algorithms were used, and their
performances were evaluated. For each algorithm, the best hyperparameters were selected using the 5 k-fold
cross-validation method, and different models trained with these hyperparameters were examined. The ndings
indicate that the GPC algorithm has higher accuracy values than others, and its performance is consistent with a
high potential for generalization. The GPC algorithm has stood out from the others with its high-performance
values. For the GPC algorithm, the accuracy, precision, recall, F1 score, and Area Under the Curve (AUC)
metrics values were 99.56%, 0.978, 0.989, 0.983, and 0.99, respectively. Additionally, Receiver Operating
Characteristic (ROC) analysis has shown that GPC, DT, and SVM algorithms effectively distinguish positive and
negative classes.
1. Introduction
An examination of reports published in previous years regarding re
data exhibits that a signicant proportion of incidents are seen to be
caused by electricity. In the Malaysian Fire and Rescue Department
activity report, it was determined that approximately 61% of the res
that occurred in 2020 (4256 incidents) and approximately 60% of the
res that occurred in 2021 (4558 incidents) were caused by electricity.
In addition, it was determined that approximately 13% of electrical res
were caused by arcing, 15% by short circuits, 5% by overvoltage and
overload, and 28% by heating [29]. Similarly, in the report shared by
the Istanbul Metropolitan Municipality Fire Department, it was stated
that 23.5% (29,200 incidents) of the res that occurred between 2015
and 2019 were caused by electricity [20].
With the development of Industry 4.0 in today’s world, there has
been a notable increase in the use of full and semi-autonomous systems,
particularly in large industrial facilities such as automotive and food
factories. This rise in automation technology also corresponds to an
increased demand for electrical energy. Consequently, electrical panels
may operate under more severe conditions than their initial installation
parameters, potentially leading to re hazards within the electrical
distribution panel. It is commonly observed that malfunctions stemming
from electrical systems can result in signicant damage and res.
Various factors contribute to these hazards, including loss of contact,
short circuits, and overheating of electrical devices [62]. Additionally,
loose connections of conductors within the panel and insulation dete-
rioration due to connector overheating are known re risk factors [19].
Moreover, electrical distribution faults causing arcing present a signi-
cant re risk, posing dangers to both property and human life [18,36]. If
ignition within the panel goes undetected in its early stages, it can
escalate rapidly and develop into a well-amed re. Such res occurring
in the workplace can lead to partial delays or complete halts in pro-
duction activities.
At this point, relevant regulations and standards mandate the annual
maintenance and inspection of all electrical equipment, electrical
panels, and cables in terms of both re safety and equipment integrity
* Corresponding author.
E-mail address: mpeksen@sakarya.edu.tr (M.F. Peks
¸en).
Contents lists available at ScienceDirect
Alexandria Engineering Journal
journal homepage: www.elsevier.com/locate/aej
https://doi.org/10.1016/j.aej.2024.03.106
Received 26 October 2023; Received in revised form 8 February 2024; Accepted 31 March 2024
Alexandria Engineering Journal 96 (2024) 112–123
113
[34]. However, shortcomings and issues in inspections and mainte-
nance, improper use of equipment after maintenance, and even post-
ponement of maintenance due to a busy production schedule can cause
res in electrical panels.
Predictive maintenance, one of the types of maintenance, is a tech-
nique based on monitoring certain parameters of operating equipment
and periodically analyzing them, allowing for the early detection of
faults [32]. This method contributes to monitoring conditions without
interrupting production and obtaining measures for re safety. In this
way, it is possible to prevent situations such as loss of life, injury, and
material damage that may be caused by res. Among the predictive
maintenance methods, there are frequency, insulation, short-circuit
measurements, circuit breaker (thermal-magnetic), transformer, earth-
ing inspections, leakage current, fuse, cable tests, and infrared in-
spections [40,43]. However, if such inspections are not performed by
qualied and experienced personnel in the eld, the results obtained can
be misinterpreted [52]. Therefore, scientists research fast and reliable
technologies that can detect res at the initial stage through various
studies [50].
1.1. Literature review
Upon examining the eld literature, no studies on predictive main-
tenance based on ML for detecting re-based anomalies in electrical
panels have been found. However, there are re-based predictive and
preventive studies in different elds that utilize articial intelligence
(AI). In a study on forest re prediction, data were collected from the
forest area using smoke detection sensors, and temperature and hu-
midity sensors were processed using deep learning methods. The results
were then sent as warnings to the relevant department [3]. In another
study for predicting building res, a deep learning model was utilized to
determine a re from smoke images. Simulation data were trained with
Convolutional Neural Network (CNN), and attempts were made to pre-
dict temporary re scenarios using smoke images obtained from outside
[55]. Additionally, in a different study, a reinforced logistic regression
model was selected to predict potential re control points for pre-re
planning and operational re management. It was noted that the ML
model created accurately classied res without considering weather
conditions [35].
In examining AI studies within the electricity domain, a focused
investigation aimed at optimizing electricity consumption costs
explored the impact of design parameter changes on a building’s elec-
tricity consumption and construction materials cost, employing articial
neural networks and genetic algorithms [4]. Additionally, research has
been held out on the optimization of electricity load distribution in
stadiums using AI and smart building design technology [57]. Regarding
the utilization of smart electrical systems, studies have delved into the
operation of articial intelligence-based electrical devices. One such
study involved the development of an intelligent battery-based electrical
system employing articial neural networks and demand-based load
planning [28]. In another machine learning-based investigation
focusing on re frequency analysis, the impact of climate change on
wildres was analyzed, and predictions were made regarding the in-
crease in the number of future res [58]. Furthermore, a study con-
cerning the classication of re accelerators in arson cases yielded
effective results utilizing the adaptive synthetic sampling method and
CNN model, based on the examination of approximately 4000 cases
[37].
An examination of studies about electrical panel re demonstrates
that sensitivity analysis of input variables has been studied to determine
re damage thresholds on the electrical panels. In one study, a re dy-
namics simulator was used to create dimensionless geometry-based
damage thresholds for different electrical panels, and sensitivity ana-
lyses were calculated [21]. Another study focused on conducting a re
risk assessment to develop compliance strategies based on re protec-
tion requirements for backup electrical panels in a nuclear power plant.
Utilizing a probability risk assessment method alongside deterministic
re protection standards, the study assessed the impact of electrical
panels on re risk in nuclear power plants [23]. In a different study,
multi-sensor re detection using fuzzy logic in a low-voltage electrical
panel was explored. This study compared both conventional and
modular fuzzy logic methods for processing data from temperature,
smoke, ame, and carbon monoxide sensors, revealing that the modular
fuzzy logic method yielded more realistic results [44]. Another experi-
mental study focused on the reduced-scale re spread between adja-
cently placed electrical panels. It provided signicant insights into the
conditions of cables located at the top of electrical panels coated with
non-ammable (metallic) and ammable (poly methyl) materials [61].
However, there is currently no study on machine learning-based pre-
dictive maintenance for detecting anomalies in electrical panels in the
literature. Future research in this domain holds promise for enhancing
our understanding of the re risks associated with electrical panels and
contributing to their mitigation.
As depicted in Table 1, various studies have investigated machine
learning and its sub-methods concerning re and re safety across
different elds. However, it is notable that specic examinations of
anomalies in electrical panels for predictive maintenance and re safety
have been lacking in the literature. Therefore, this study stands as the
pioneering endeavor aiming to ll this gap and contribute substantially
to the existing body of knowledge.
1.2. Signicance of the research
In this study, a machine learning-based model has been proposed for
the predictive maintenance of re-related anomalies in electrical panels.
The proposed machine learning-based automatic monitoring system will
enable the detection of anomalies in electrical panels without the need
for expert personnel. To achieve this, thermal imaging cameras, tem-
perature, humidity, and gas-detecting sensors, along with a module for
measuring current and voltage, were integrated into the electrical panel
for experiments. In the experiments, a cable between the contactor and
the capacitor inside the panel was selected, and multiple changes were
made using different cable cross-sections. Additionally, various modi-
cations were made to the insulation of the cables, such as cuts, crushes,
reducing the number of cable strands, and using weak cables. A total of
24 experiments were conducted. Using the collected data on tempera-
ture, humidity, current, and voltage, a machine learning-based predic-
tive maintenance model was developed.
Similar studies have been conducted using Supervisory Control and
Data Acquisition (SCADA) systems. However, SCADA systems primarily
monitor machine systems used in production processes and do not
collect data specically about electrical panels. Feature engineering
plays a crucial role in collecting all systemic data, determining required
features through feature selection, and detecting anomalies [53,63].
Through the experiments, machine learning-based anomaly detection
was studied alongside feature selection. Upon examining the literature,
it was found that no study within this scope had been conducted yet.
Implementing this system within the panel allows for faster monitoring
of conditions and the early detection of any anomalies without human
intervention. This is particularly advantageous in facilities with
numerous electrical panels, as it saves time for maintenance teams,
enables the proper execution of maintenance planning processes, and
improves performance criteria by employing sufcient man/days. This
system alleviates the workload of maintenance teams there is no need
for real information or data from previous years. Additionally, since the
panels are monitored 24/7, potential re incidents are prevented before
they occur. Consequently, potential workplace shutdowns due to elec-
trical panel malfunctions or occupational accidents caused by panel
ignition or re can be averted. However, the system also has some dis-
advantages in detecting certain anomalies. Among these, it was
observed that the sensor that reads the temperature inside the panel is
affected by the external temperature, which may lead to false anomaly
M.F. Peks¸en et al.
Alexandria Engineering Journal 96 (2024) 112–123
114
Table 1
Comparing this study with relevant research.
Research Research Origin Parameters Methodology Performance
Wang et al. [55] Predicting transient building
re
Heat release rate, opening size, and fuel type Deep learning Accuracy of relative error 20%
Baghoolizadeh
et al. [4]
Annual electricity
consumption in the residential
building
Insulation thickness, phase change materials
thickness, melting temperature, phase change
temperature, and insulation conductivity
Articial neural networks (ANN). ANN Correlation coefcient (R
2
)
0.76–0.86
Xu et al. [58] Fire frequency analysis and
prediction of re using ML
with a case study.
Historical data, daily maximum and minimum
temperature, precipitation, and wind force.
Random forest (RF), SVM and
Polynomial regression (PR)
R
2
RF (T
max
) 0.95,
R
2
SVM (T
max
) 0.51, R
2
PR (T
max
)
0.87.
Park et al. [37] Arson, re accelerant
classication using ML with a
case study.
Fire accelerant, gasoline, diesel, candle, solvent
…
CNN, RF, SVM. RF accuracy 0.88, weighted
average precision 0.88,
recall 0.88,
F1-score 0.88, AUC 0.98,
SVM accuracy 0.88, weighted
average precision 0.89,
recall 0.88,
F1-score 0.88, AUC 0.98
CNN accuracy 0.92, weighted
average precision 0.92,
recall 0.92,
F1-score 092, AUC 0.99
Kim et al. [21] Sensitivity analysis for re
damage at the electrical
panels
Separation distance Image processing and re dynamic
simulator
-
Lee et al. [23] Fire risk assessment at the
redundant electrical panels.
Heat release rate, ignition, hot surface. Probabilistic risk assessment (PRA),
re dynamics tools (FDT)
-
Sahid & Alaydrus
[44]
Fire Detection in Low Voltage
Electrical Panel
Smoke, carbon monoxide, ame, and
temperature.
Modular Fuzzy Logic -
Zavaleta [61] Experimental study re
spread between electrical
cabinets
Separation distance, overhead electric cable
trays, heat ux
Test measurements -
Hong et al. [17] Mine tunnel re smoke way
prediction.
Tunnel height, width, ambient temperature,
ventilation velocity, heat release rate, and
inclination angle,
SVM, RF, classication, and
regression tree (CART), and ANN
SVM accuracy 98.67%, precision
0.99, recall 0.97, F1 score 0,98.
RF accuracy 98.67%, precision
0.99, recall 0.97, F1 score 0,98.
CART accuracy 95.67%,
precision 0.91, recall 0.96, F1
score 0,94.
ANN accuracy 97%, precision
0.98, recall 0.93, F1 score 0,95.
Sun et al. [51] Coal industry re safety
assessment.
Coal temperature, Smoke concentration, Carbon
monoxide concentration
SVM, Naive Bayes Classier (NBC), SVM accuracy 90.82%. recall
0.94, precision 0.9, F1 score
0.92,
NBC accuracy 73.67%, recall
0.78, precision 0.85, F1 score
0.81,
Chen et al. [8] Early-stage re detection. Temperature, Carbon monoxide, smoke, SVM, RF, K-means SVM accuracy 89.50%, precision
0.88, recall 0.88, F1 score 0.88.
RF accuracy 87.50%, precision
0.86, recall 0.87, F1 score 0.86.
K-means accuracy 50%,
precision 0.49, recall 0.48, F1
score 0.48.
Zulfauzi et al.
[65]
Anomaly detection at the
photovoltaic plant
Temperature and current K-Means for clustering, Long-short
term memory (LSTM), ANN
LSTM relative error 4.316%
ANN relative error 4.363%.
Mellit et al. [31] Fault diagnosis of
photovoltaic arrays using
embedded system
Current, voltage, incident solar irradiance, cell
temperature
ANN with eXtreme Gradient
Boosting (XGBoost), Light Gradient
Boosting Machine (LightGBM)
XGBoost accuracy 96.80%,
precision 0.99, recall 0.97, F1
score 0.98,
LightGBM accuracy-, precision
0.97, recall 0.96, F1 score 0.97,
This study Predictive maintenance, re
safety
panel ambient temperature, panel humidity,
CO, CO2, voltage, current, cable temperature,
outdoor temperature, and outdoor humidity
DT, GPC, SVM, GNB For the S1 class,
DT accuracy 98.99%, precision
0.966, recall 0.956, F1 score
0.961, and AUC 0.97.
GNB accuracy 90.51%, precision
0.603, recall 0.802, F1 score
0.688, and AUC 0.95.
GPC accuracy 99.56%, precision
0.978, recall 0.989, F1 score
0.983, and AUC 0.99.
SVM accuracy 98.41%, precision
0.934, recall 0.945, F1 score
0.939, and AUC 0.99.
M.F. Peks¸en et al.
Alexandria Engineering Journal 96 (2024) 112–123
115
detection. To prevent this, the outdoor temperature was measured using
a different sensor, and the situations were closely monitored. This
approach resulted in excessive data processing and a minor loss of time.
In addition, instead of measuring the ambient temperature, the use of an
expensive thermal camera to read the focal temperature yielded more
efcient results in anomaly detection, albeit increasing system costs.
The suggested system does have some limitations. The rst of these is
that not every panel size is suitable for installing the system. The thermal
camera requires a wide angle to measure the focal temperature inside
the panel. To ensure a wide angle, the measurement must be made from
a distance. The second limitation is the cost of the system. Installing an
anomaly detection system on all panels will entail additional expenses
for the business.
During the creation of this model, feature selection processes were
applied to choose the most suitable features among the dataset’s char-
acteristics. The dataset was divided into test and training sets using the
k-fold cross-validation method, and DT, GNB, GPC, and SVM algorithms
were applied. These algorithms and applied methods have been dis-
cussed in detail in the subsequent sections of the study. This work is
considered as an important step in terms of the safety and maintenance
of electrical panels. It is anticipated that the results will play a crucial
role in industrial applications.
The implications of this research are signicant in the context of
predictive maintenance and re safety. It facilitates the accurate and
early detection of anomalies in electrical panels, thereby preventing
res within these panels. Thus, re safety is enhanced in both industrial
and residential environments. The integration of IoT and machine
learning in this eld will lead to more reliable and efcient monitoring
systems. Additionally, this study makes a valuable contribution to the
domain of machine learning-based predictive maintenance and exem-
plies the practical application of technology in a critical area of re
safety.
Below is a summary of the general structure of the article. The second
section briey describes the experimental design carried out using
Arduino and Raspberry Pi, the data collection process, data set creation,
data preprocessing steps, feature selection, the extraction of thermal
images of the electrical panel cables, and commonly used algorithms and
techniques for automatically classifying the electrical panels into normal
and re stages. The third section covers experimental studies, ndings,
and discussions regarding ML classiers and evaluation metrics such as
accuracy, recall, precision, and F1 score. Finally, the results are pre-
sented in the fourth section.
2. Materials and methods
2.1. Experiment design
For this study, an electrical panel with dimensions of 50 cm x 70 cm x
20 cm was equipped with a ex glass cover at the front to allow for
observation and was set up as shown in Fig. 1. Within the panel, a 25
Ampere contactor, a 50 Ampere fuse, a terminal block, and capacitors
with a capacity of a 5kvar (Kilovolt Ampere Reactive) and 2.5kvar were
installed. Additionally, the cable between the contactor, highlighted in
red in Fig. 1, and the capacitor was selected as the experimental
connection point. This cable was used in initial experiments with the
5kvar capacitor and operating cables. Subsequently, experiments were
conducted using the same cable variety with a 2.5kvar capacitor, thus
providing variety in the experiments.
A Tibcon brand 5kvar and 2.5kvar capacitor were used in the elec-
trical panel to provide variety in experiments. With these capacitors,
experiments were completed using different cables, as shown in Fig. 2.
The original cable shown in Fig. 2(a) is black, with green-tipped ends,
and was covered in Polyvinyl chloride (PVC) on the outside and contains
multiple wire bundles wrapped around each other; its cross-section is
4.8 mm. The blue PVC-coated cables shown in Fig. 2(b to e) each have a
cross-section of 3 mm. The black PVC-coated cables shown in Fig. 2(f to
i) each also have a cross-section of 3 mm. The white PVC-coated cables
shown in Fig. 2(j to l) each have a cross-section of 2 mm and were
connected at congurations of three layers, two layers, and a single
layer. The gray cable shown in Fig. 2(m) has a cross-section of 1.5 mm.
In the experiments with the 5kvar capacitor, cables shown in Fig. 2(a to
e, j to l) were used, while in experiments with the 2.5kvar capacitor,
cables shown in Fig. 2(a, f to m) were used.
Fig. 1. Electrical panel setup for experimental purposes.
Fig. 2. Cables used in the panel (a) original black cable, b) blue cable, c) blue
cable with reduced strands, d) blue cable with crush (crushed area within the
red region), e) blue cable with cuts in the insulation, f) black cable, g) black
cable with reduced strands, h) black cable with crush (crushed area within the
red region), i) black cable with cuts in the insulation, j) 3-layer white cable, k)
2-layer white cable, l) single-layer white cable, m) thin cable.
M.F. Peks¸en et al.
Alexandria Engineering Journal 96 (2024) 112–123
116
The procedural owchart for the experimental cables connected to
the electrical panel is shown in Fig. 3. These procedures were applied for
the 5kvar and 2.5kvar capacitors and for all different cable types of
experiments.
During the experiment, an Arduino Mega 2560 R3 microcontroller
and a Raspberry Pi 4 minicomputer were used for data collection. The
Arduino was xed onto the electrical panel, and using cables extended
into the panel, measurements were taken for ambient temperature and
humidity, as well as the toxic gas carbon monoxide affecting air quality,
and the simple asphyxiant gas carbon dioxide inside the panel. The
ambient temperature and humidity were measured using the DHT22
sensor, while gas measurements were performed using the MQ135 and
MQ4 sensors. Additionally, a current clamp was attached to the con-
tactor cable inside the panel. The current and voltage values were ob-
tained using the PZEM-004 T sensor. For wireless communication of the
Arduino, the ESP-8266 module was used. The Raspberry Pi was placed
next to the electrical panel, and through cables extended into the panel,
the MLX90640 infrared thermal camera was mounted inside the panel.
Also, with the Raspberry, external ambient temperature and humidity
were measured using another DHT22 sensor. The Raspberry Pi was
connected to the internet via a wired connection on its ethernet card.
The system block diagram is shown in Fig. 4.
A software program for the Arduino microcontroller was developed
in the C programming language. This program collected data from
sensors and transmitted it to the cloud system. Additionally, an audio
and visual alarm system was added to Arduino in case data beyond the
expected range was received. The detailed working diagram of this
system is shown in Fig. 5.
When the Arduino microcontroller is rst started, the system waits
for 15 seconds to warm up the sensors. Afterward, the wireless internet
connection protocol is initiated through ESP8266. If the connection is
successful, communication is established with the private channel ID
address on the cloud system, hosted on thingspeak.com. After successful
communication is established, the Arduino gathers data from the sensors
every 5 seconds, including temperature and humidity from the DHT22
sensor, ambient air quality values for carbon monoxide and carbon di-
oxide from the MQ135 sensor, and smoke and hydrocarbon gas values
from the MQ4 sensor, as well as voltage and current values from the
PZEM-004 T module. These values are then saved to the thingspeak.com
cloud system. If any of the measured values exceed the predetermined
thresholds, the Arduino microcontroller triggers both an audible and
visual warning system.
The software for the Raspberry minicomputer was developed using
the Python programming language. It gathered data from the sensors
and communicated with the cloud system. Additionally, an audible and
visual warning and alarm system were added to the Raspberry in case
data outside the expected range was received. The detailed operational
diagram of this system is shown in Fig. 6.
When the Raspberry Pi minicomputer is rst powered on, it retrieves
data from the DHT22 sensor and the MLX90640 infrared thermal camera
module. It then initiates a wired internet connection protocol via its
ethernet card. If the connection is successful, it communicates with the
private channel ID address created on the cloud system thingspeak.com.
Upon successful communication establishment, every 5 seconds, the
Raspberry Pi sends the temperature and humidity values from the
DHT22 sensor and the focal temperature value on the cable subjected to
a test with the MLX90640 infrared thermal camera to the thingspeak.
com cloud system. If the measured values exceed a specied limit, both
an audible and visual alarm system are activated. The difference be-
tween the Raspberry Pi minicomputer and the Arduino microprocessor
is that the Raspberry Pi can capture images using the MLX90640
infrared thermal camera. It saves the images both in its own memory and
backs them up to a Google Drive account via a protocol created with the
Google Drive account every 5 seconds.
2.2. Dataset and data processing
In this study, after all experiments were completed, a dataset con-
sisting of 3478 samples was obtained by synchronizing the timings of the
Arduino and the Raspberry Pi. The dataset included data from two
different sources. The rst source contained data retrieved from the
Arduino, including panel ambient temperature, ambient humidity,
carbon monoxide levels in the panel, carbon dioxide levels, and voltage
and current values from the busbar cable. The second source contained
data retrieved from the Raspberry Pi, including cable focal temperature,
external ambient temperature, and ambient humidity levels. The
sampled data obtained with the Arduino and the Raspberry Pi is pre-
sented in Table 2.
In the experiments, it was observed that the temperature in the cable
initially increased rapidly, and when it started to reach the regime point,
the temperature increase rate decreased, but the increasing trend
continued. Ignition, re, or explosion occurred in seven of the experi-
ments. Pictures showing the condition of the cables after these seven
tests are given in Fig. 7.
Infrared thermal camera images of the results obtained in the 5kvar
white cable experiment are given in Fig. 8 and No Infrared (NoIR)
camera images placed inside the panel are given in Fig. 9.
As a result of the experiments, the data obtained from Arduino and
Fig. 3. Experimental procedure diagram.
Fig. 4. Block diagram schematic.
M.F. Peks¸en et al.
Alexandria Engineering Journal 96 (2024) 112–123
117
Raspberry were evaluated according to the Heat Release Rate (HRR)
graph shown in Fig. 9, and based on this, the results (outputs) were
labeled.
According to the evaluation in Fig. 9, (N) represents normal condi-
tions, Stage1 (S1) represents pre-ignition conditions like smoldering,
and Stage2 (S2) represents ignition-ashover interval conditions like
well ame development. Stage3 (S3) represents fully re developed
interval conditions, and Stage4 (S4) represents post-decay conditions.
Based on this information, from the experiments conducted, a total of
3478 data points were obtained, with 3025 data points for N and 453
data points for S1. Since enough data for S2, S3, and S4 weren’t
collected, the AI model was trained for binary classication using only
the data, including N and S1 classes.
2.3. Feature selection algorithms
Feature selection is a method that aids in obtaining a subset from the
original feature set based on specic criteria, selecting relevant features
of the dataset, and allowing compression of the data processing scale
[7]. In this method, all features are ranked based on specic criteria.
Then, the features with the highest scores are selected. Three methods
operate in the feature selection model depending on the search strate-
gies. These are classied as lter methods, wrapper methods, and
embedded methods [33]. The lter method is based on the general
characteristics of the data and evaluates features without involving any
learning algorithm. In contrast, the wrapper method requires a
pre-determined learning algorithm and uses performance as an evalua-
tion criterion for selecting features [64]. Embedded methods are similar
to wrapper methods but are a hybrid of wrapper and lter methods, and
Fig. 5. Signal processing diagram of the Arduino Mega 2560 R3 re detection system.
Fig. 6. Raspberry Pi 4 re detection system signal processing diagram.
Table 2
Arduino Mega 2560 R3 and Raspberry Pi 4 sample data.
Arduino Mega 2560 R3 Raspberry Pi 4
Panel Ambient Temp. (
o
C) Panel Humidity CO CO
2
Voltage (Volt) Current (Amper) Cable Temp. (
o
C) Outdoor Temp. (
o
C) Outdoor Humidity
22,8 41,6 3,1 402,39 228,5 11,13 24,2 22 31
26,7 33,4 2,52 402,06 229,6 11,27 26,7 22 28
28,3 30,5 2,76 402,19 228,6 11,19 27,6 22 27
28,8 29,4 2,18 401,85 229 11,24 28,1 22 26
29,4 28,6 1,97 401,72 227,2 11,15 28,6 24 26
Fig. 7. a) 5kvar white cable, b) 2,5 kvar 5 wires missing white cable, c) 2,5
kvar 8 wires missing white cable, d) 2,5 kvar 6 wires missing white cable, e)
thin cable, f) 2,5 kvar 6 wires missing white cable g) 2,5 kvar 7 wires missing
white cable.
M.F. Peks¸en et al.
Alexandria Engineering Journal 96 (2024) 112–123
118
they are more powerful than wrapper methods. This method uses clas-
sication algorithms with built-in capabilities to select features [27].
The application of feature selection algorithms has several advan-
tages, especially in reducing storage requirements and computational
resources, lowering computational costs, enhancing the interpretability
of predictive models, and selecting relevant features from the dataset
[45].
2.3.1. Recursive feature elimination
The Recursive Feature Elimination (RFE) algorithm is one of the
widely used feature selection algorithms in ML and data analysis due to
its ease of use, congurability, ability to predict target variables, effec-
tiveness in removing weak features, and selecting features in training
datasets [47]. This technique aims to eliminate features through itera-
tion. Its goal is to identify the most important features in a dataset by
repetitively removing less relevant ones. In situations with a large
number of features, such as those dening re conditions in training
data, there is a potential risk of overtting for a ML method due to high
dimensionality. This situation can be mitigated by applying the RFE
method. When examining re data, it may not be the case that all fea-
tures have a positive impact on predictions. RFE is an efcient method
that recursively reduces model complexity by removing irrelevant and
unnecessary features based on ranking criteria [25]. It is a wrapper
feature selection algorithm that also internally uses lter feature selec-
tion, and it can easily be integrated with different ML algorithms [15]. In
RFE, prediction models are initially trained for each feature by assigning
weights based on the original features. Then, features with the smallest
absolute weights are removed, simplifying the feature set. This process
continues until the number of remaining features reaches a sufcient
level, usually when classication accuracy reaches an acceptable level
[26,48].
An electrical panel anomaly dataset containing more than one
feature was used in this study. This dataset includes panel ambient
temperature, panel humidity, CO, CO2, voltage, current, cable temper-
ature, outdoor temperature, and outdoor humidity features. The RFE
algorithm was applied to systematically eliminate the least essential
features from the nine features in this dataset. This process was per-
formed iteratively, evaluating the model’s performance at each step.
Throughout this process, the RFE algorithm identied the four most
important features-panel humidity, CO, CO2, and cable temperature—as
the most important indicators of the desired results, in order of impor-
tance. This approach ensured that the nal model was efcient and
effective while maintaining high prediction accuracy with a simplied
feature set. This method is particularly effective and helpful in high-
dimensional data scenarios where feature selection is crucial for
model performance.
2.4. Machine learning model
ML is a subeld of computer science that belongs to the AI segment
and is focused on discovering hidden patterns in complex datasets [46].
Essentially, ML involves computers learning (or being trained) from data
obtained from repeated samples, guiding them to perform tasks intelli-
gently beyond traditional computation [10].
ML is categorized into four main groups: supervised learning (SL),
unsupervised learning (USL), semi-supervised learning (SSL), and rein-
forcement learning (RL). SL is the most commonly used ML technique,
and SL algorithms are designed to recognize the model from training
data and predict new input pairs or previously unseen observations [1].
In USL, only unlabeled data is provided to the model. Various selection
approaches, such as clustering, anomaly detection, autoencoders, and
more, are based on assumptions [24]. SSL is a hybrid of SL and USL
models. In other words, it is an ML method that utilizes both labeled and
unlabeled data to learn mappings between examples and desired outputs
[22]. RL is an ML method where an agent learns to complete a specic
task successfully through a sequence of actions with cumulative feed-
back. Unlike SL, it doesn’t require labeled example-output data and
Fig. 8. Sample camera and infrared images from the electrical panel experiment (a) Detected temperature and thermal image at the beginning of the experiment, (b)
Thermal image showing intensication in purple color according to the heating level after a certain period, (c) Increase in purple color of the cable during the
combustion stage, (d) Thermal image depicting the state where electrical current is cut off from the cable, (e) Image obtained from the NoIR camera placed inside the
panel at the beginning of the experiment, (f) Corresponding image to (b), (g) Image of smoke emission during the combustion phase, (h) Image after the panel’s
power is urgently cut off and intervention in the re.
Fig. 9. Different phases in the development of a compartment re [2].
M.F. Peks¸en et al.
Alexandria Engineering Journal 96 (2024) 112–123
119
directly assesses incorrect or insufcient predictions or actions [22].
The ML model enhances AI applications in many areas, including
automotive, nance, military, and healthcare services. The primary
reason for this is that it offers advantages in terms of predictive accuracy
compared to "classical" statistical models [13]. In facilities where pro-
duction operations are intensive, ML methods are frequently used. It is
especially one of the effective methods for critical and preventive
maintenance. With regular and real-time monitoring of the working data
of electrical panels, res in panels and personal injuries or fatalities
caused by panel res can also be effectively prevented. Therefore, using
the current new technology, electrical panels can be monitored and
utilized as predictive maintenance work. In this study, DT, GNB, SVM,
and GPC algorithms were selected from supervised ML methods for re
classication in electrical panels.
2.4.1. Decision tree
ML is a signicant model frequently used in statistics and data
mining. It is an intuitive SL algorithm commonly employed for classi-
cation and regression problems [39]. DT, known as classication trees,
is a popular method due to its ability to represent different consider-
ations for each outcome in a owchart, reecting human intuition [30].
DT is visually represented by a graphic and often drawn upside down
[60]. The rst step in developing a DT is selecting the attribute that will
be placed at the root node, from which recursive branches will grow. At
each node, the examples are progressively "split" into branches based on
dened criteria [9]. This process proceeds in a tree-like fashion,
repeating the process until it reaches a leaf node. When a leaf node is
reached, the tree predicts the outcome [49]. In this study, DT was
selected due to its low computational cost, ability to deal with missing
values, and simple processing of mixed data. This method has provided
good results, especially in completing missing data in some cases.
2.4.2. Gaussian naive bayes
GNB, one of the simplest classication algorithms, calculates the
probability of an event based on previous knowledge, describing it in
terms of any feature according to Bayes’ probability and statistical rule
[6]. This theorem calculates the probability of an event occurring based
on experience according to Eq. (1) [14].
P(cx) = P(xc) ∗ P(c)
P(x)(1)
Here, P(cx) represents the probability value of x belonging to class c,
P(xc) represents the actual probability value of x for class c in the
training data, P(c) represents the actual probability value of x in the
training data, and P(x) represents the actual probability value calculated
from the training data [14]. Because of the method’s ease of use, high
efciency, good performance in working with large data, and trans-
parency of the ndings, GNB was used. This method was used to obtain
accurate results with high efciency due to the confusion occurring in
the 3478 samples used in the system.
2.4.3. Gaussian process classier
GPC was rst introduced by Christopher K. I. Williams in 1998 in his
work titled “Computation with Innite Neural Networks” [56]. GPC is
considered a non-parametric probabilistic classication model, and its
main feature is its ability to use sample data directly without needing
estimates of characteristic parameters in the theoretical distribution
[42]. Prediction probabilities are calculated according to Eq. (2).
p(yn|xn,X,y) = ∫p(yn|f(xn))⋅p(f(xn)|X,y)df (xn)(2)
Here, p(y
n
| x
n
,X,y) represents the conditional probability of class
label y
n
given input x
n
. x
n
represents the input for which a prediction will
be made, y
n
represents the class label for the input, f(x
n
) represents the
estimated function for input x
n
, X represents the dataset, and y repre-
sents the class labels of the dataset.
GPC is used in conjunction with other Gaussian Process-based
methods such as Gaussian Process Regression (GPR) or Gaussian Pro-
cess Machine Learning (GPML) and is especially applied in pattern
recognition and classication problems. It has been successfully applied
in areas such as image classication, fault diagnosis, text categorization,
product quality prediction, spam email ltering, facial expression
recognition, and document classication [38,54]. For this study, GPC
was selected because of its ability to enhance prediction accuracy by
minimizing uncertainty in forecasts. As will be seen in the following
sections, the best results were obtained using GPC to determine per-
formance criteria.
2.4.4. Support vector machine
SVM, developed by Vladimir Vapnik in the 1990 s, is based on sta-
tistical learning theory [5]. It is a classication and regression method
that can be used for both linear and non-linear problems. SVM’s ability
to perform well with a small number of features, its robustness against
model errors, and its efcient computation compared to other ML
methods make it a preferred choice [12]. It is becoming increasingly
popular, particularly in elds such as articial intelligence, data mining,
pattern recognition, and optimization problems, and is considered a
powerful toolset [41].
SVM is used to establish the hyperplane that best separates the two
classes with the maximum margin. It employs linear separation mech-
anisms in two-dimensional space, planar mechanisms in three-
dimensional space, and hyperplane mechanisms in multi-dimensional
space [16]. The data points that are closest to the optimal separating
hyperplane and have a larger margin width are the most challenging to
classify [11]. Therefore, SVM utilizes a training phase to identify the
optimal separation hyperplane [59]. This method was particularly used
due to its strong generalization ability and adaptability to nonlinear
examples. Because the 3478 samples obtained were at nonlinear
conditions.
2.4.5. Performance evaluation
In machine learning, performance evaluation metrics are frequently
used to assess a model’s predictive capabilities and performance. These
metrics are essential for evaluating the effectiveness of a model in
classication tasks. The four main metrics used to evaluate a classi-
cation model are accuracy, precision, recall, and F1 score. These metrics
are derived from the confusion matrix, a table (Table 3) used to evaluate
the performance of a classication model on a dataset where the actual
values are known. These values are calculated based on the confusion
matrix values True Positives (TP), False Positives (FP), False Negatives
(FN), and True Negatives (TN).
Normal (N) conditions represent pre-ignition conditions, and Stage1
(S1) represents ignition-pre-ashover conditions such as smoldering. In
this study, for the N and S1 conditions, TP represents the model that
correctly predicts an anomaly. That is, the model’s prediction is positive,
and the actual condition is indeed an anomaly. FP refers to cases in
which the model incorrectly predicts an anomaly. This means that the
model’s prediction is positive, but the actual condition is normal.
Conversely, FN refers to situations in which the model cannot detect an
anomaly. This means that the model’s prediction is negative, yet the
actual condition is anomalous. TN refers to situations in which the
model correctly identies the absence of an anomaly.
Accuracy is the most common evaluation measurement for classi-
Table 3
The confusion matrix for two-class classication.
Actual
Positive Negative
Predicted Positive TP FP
Negative FN TN
M.F. Peks¸en et al.
Alexandria Engineering Journal 96 (2024) 112–123
120
cation models, and it measures the ratio of correct predictions—both
true positives and true negatives—made by the model to all projections
made. It is calculated as the following Eq. (3). Despite accuracy being a
prevalent metric, its interpretation within imbalanced datasets warrants
caution. It can be misleading in the presence of class imbalance. The
prevalence of one class signicantly outweighing another may lead to a
perception of high performance.
Accuracy =TP +TN
TP +TN +FP +FN (3)
Precision quanties the proportion of positive identications that are
actually correct. At the same time, precision is a measurement that
represents the percentage of TP predictions made by the model among
all positive predictions. In other words, the rate of precision is the per-
centage of total anomalies reported out of the true anomalies. High
precision indicates a low rate of false positives, which is crucial in ap-
plications where the cost of a false positive is high. This measurement
exhibits the precision and reproducibility of the measurements and is
calculated using Eq. (4).
Precision =TP
TP +FP (4)
Recall is a measure of how many of the positive cases in the dataset
are correctly identied by the model. In other words, the rate of recall is
the percentage of actual anomalies that are reported among the total
number of existing anomalies. High recall is important in scenarios
where missing a positive instance is costly. For example, misclassica-
tion of abnormal conditions as normal in electrical panels can lead not
only to high costs due to production stoppages but also to situations that
can cause loss of life. It is calculated using Eq. (5).
Recall =TP
TP +FN (5)
The F1 score is the harmonic mean of the precision and recall values.
It is particularly useful for imbalanced datasets, and its value ranges
from 0 to 1, where a higher value indicates better performance. In
addition, the F1 score serves as a commonly used metric to evaluate
model precision, offering the benet of incorporating both precision and
recall. Finally, the high F1 score indicates that the model minimizes FN
while making correct positive predictions and is calculated using Eq. (6).
F1score =2∗Precision ∗Recall
Precision +Recall (6)
Additionally, a comparative analysis was conducted with metrics
such as AUC for ROC to provide a well-rounded evaluation of model
effectiveness.
3. Experimental results and discussion
Within the scope of this study, a data set consisting of 3478 samples
was used. For each algorithm, models were created with a k-fold of 5 and
using different hyperparameters. At this stage, the best generalizing
model was designed, considering the overall accuracy of the model, the
F1 score, and other critical performance metrics. For GPC and GNB al-
gorithms, a single parameter was used. For DT and SVM algorithms,
different hyperparameter combinations were used. For the SVM algo-
rithm, models with different hyperparameter combinations were
created with C parameter was [0.1, 0.5, 1, 5, 50, 100], gamma parameter
was [0.01, 0.1, 1, ’auto’, ’scale’], kernel parameter was [’linear’, ’rbf’,
’sigmoid’, ’poly’] and degree parameter [1, 3, 5, 7, 10, 12]. For the GNB
algorithm, the priors parameter was designed as [None], and the var_-
smoothing parameter was 1e-9. For the GPC algorithm, the kernel
parameter was designed as [1**2 * RBF (length_scale=1)]. For the DT
algorithm, models with different hyperparameter combinations were
created with max_depth parameter was [None, 2, 5, 8, 10, 12], max_-
features parameter was [None, ’auto’, ’sqrt’, ’log2’, 1, 2, 5], criterion
parameter was [’squared_error’, ’friedman_mse’, ’absolute_error’,
’poisson’], min_samples_split parameter [2, 5, 10] and min_samples_leaf
parameter [1, 2, 4]. For each algorithm, the best hyperparameter com-
binations were selected using the Python RandomizeSearchCV library.
The 5-fold cross-validation method was applied for each algorithm
using different hyperparameter combinations, and the model accuracy
values obtained were averaged. The best hyperparameters were selected
by comparing these accuracy values. The variation of the model accu-
racy values obtained is shown in a boxplot plot for each algorithm, as
shown in Fig. 10. Since the GPC and GNB algorithms have only one
parameter combination, only one model accuracy is shown in the graph.
For the SVM algorithm, 135 combinations of hyperparameters were
tested, and the lowest model accuracy was 0.74. For the DT algorithm,
300 combinations of hyperparameters were also tested; the most insuf-
cient model accuracy was 0.92. All variability is illustrated in Fig. 10.
As a result of this training, the best hyperparameters for each algo-
rithm were determined, and these hyperparameters are given in Table 4.
The best hyperparameters and the distribution of accuracy values
from model trainings using 5-fold cross validation were represented
using a boxplot graph and are shown in Fig. 11. The model was retrained
according to the best hyperparameters, and the results are shown in
Table 5.
The boxplot graph was created using training accuracy values in
different data sections to visually represent the distribution and general
consistency of the performance of each algorithm in different data sec-
tions. Upon examining Fig. 11, it is evident that DT and GPC algorithms
exhibit higher accuracy values than the other algorithms. Additionally,
the box plots of the training accuracy values of these algorithms indicate
consistent training results and a high generalization potential. These
ndings were used for model selection and hyperparameter tuning.
According to the best hyperparameters selected in Table 3, the al-
gorithms were tested on the data set, 80% of which (2782 data) was
divided into training and 20% (696 data) into the test data set. The
confusion matrices obtained for each algorithm are given in Fig. 12.
Within the scope of the study, all operations regarding the dataset and
ML were carried out using the Python programming language and a
single desktop computer with Ubuntu 18.04 64-bit operating system,
64 GB RAM, Intel i7–9800X CPU, and two NVIDIA RTX A5000 GPUs.
Python v3.10.12 programming language, sckit-image v0.22.0, sckit-
learn v1.3.0, numpy v1.24.4, pandas v2.0.3, and matplotlib v3.7.2 li-
braries were used for application and machine learning models
development.
The results presented in Fig. 12 were used to objectively assess the
performance of different classication algorithms and the effectiveness
Fig. 10. Distribution of model accuracy values obtained by 5 K-fold cross-
validation of different hyperparameters.
M.F. Peks¸en et al.
Alexandria Engineering Journal 96 (2024) 112–123
121
of the model, aiming to evaluate the classication accuracy, precision,
recall, and F1 score of each algorithm. According to the results, the GPC
and DT algorithms had the highest success. Also, it was observed that
GPC and DT algorithms completed the classication task more suc-
cessfully. On the other hand, the SVM algorithm exhibited lower per-
formance than the DT algorithm, while the lowest success rate was
associated with the GNB algorithm.
The results of the ROC analysis based on the AUC values for the four
different classication algorithms (DT, GNB, GPC, and SVM) are pre-
sented in Fig. 13. Upon examining the gure, high AUC values were
observed for GPC, SVM, and DT algorithms. These high AUC values
indicate that GPC, SVM, and DT algorithms have a good effect on dis-
tinguishing positive and negative classes.
As a result of modeling and analysis performed on the dataset of
electrical panel re experiments with classications N and S1, the
calculated F1 Score, Precision, Recall, Accuracy, and AUC-ROC Value
results are presented in Table 5. Overall, the GPC algorithm has the
highest performance results for both N and S1 classes, whereas the GNB
algorithm has the lowest performance results.
Furthermore, performance metric results for F1 Score, Precision, and
Recall for both classes (N and S1) have been presented as bar charts on
an algorithm basis in Fig. 14.
A general evaluation on an algorithm basis according to the perfor-
mance results in Table 5, Fig. 13, and Fig. 14, exhibits that the DT model
achieves a high F1 score for both classes. In particular, an F1 of 0.994 for
the ‘N’ class and high precision and recall values for the ‘S1’ class
indicate that the model successfully classies these classes. The accuracy
value of 98.99% and the AUC value of 0.97 reect the overall success of
this DT model. It is observed that although the GNB model has high
sensitivity for the ‘N’ class, it has a low F1 score and AUC value for the
‘S1’ class, indicating that the GNB model has more difculty dis-
tinguishing this class. The GPC model has high F1 scores, precision, and
recall values for both classes, indicating the successful classication of
both classes. The 99.56% accuracy value and 0.99 AUC value has
revealed that this model has the highest performance. The SVM model
has a high F1 score for the ‘N’ class and high precision values for the ‘S1’
Table 4
The best hyperparameters selected for algorithms.
Algorithms Best Hyperparameters
SVM {probability: True, kernel: ’rbf’, gamma: ’scale’, C: 100}
GPC {kernel: 1**2 * RBF (length_scale=1)}
GNB {priors: None, var_smoothing: 1e-9}
DT {min_samples_split: 2, min_samples_leaf: 1, max_features: ’None’,
max_depth: None, criterion: ’entropy’}
Fig. 11. Distribution of training model accuracy values obtained using 5 K-Fold
cross validation.
Table 5
Comparative performance analysis of the methods used for N and S1 classes.
Model Class F1
Score
Precision Recall Accuracy
(%)
AUC-ROC
Value
DT N
S1
0.994
0.961
0.993
0.966
0.995
0.956
98.99 0.97
GNB N
S1
0.944
0.688
0.968
0.603
0.920
0.802
90.51 0.95
GPC N
S1
0.997
0.983
0.998
0.978
0.996
0.989
99.56 0.99
SVM N
S1
0.990
0.939
0.991
0.934
0.990
0.945
98.41 0.99
Fig. 12. Confusion Matrix for a) DT b) GNB c) GPC d) SVM on the N class and S1 class.
Fig. 13. Comparative ROC curve and AUC analysis of DT, GNB, GPC, and
SVM algorithms.
M.F. Peks¸en et al.
Alexandria Engineering Journal 96 (2024) 112–123
122
class. However, the recall value for the ‘S1’ class is slightly lower,
indicating that the model distinguishes this class slightly less effectively.
4. Summary and conclusion
This signicant work aims to substantially reduce the risk of re in
electrical panels through anomaly detection and improve the predictive
maintenance of electrical panels. By combining IoT frameworks and ML
algorithms to detect early signs of re in electrical enclosures, this study
presents a novel approach to monitoring electrical enclosure health. At
the heart of this approach is a complex system of Arduino and Raspberry
Pi platforms designed to collect sensor data to gather information such
as gas, temperature, and humidity levels. This system has generated a
signicant dataset to test the effectiveness of various machine learning
algorithms, including DT, GNB, GPC, and SVM, in identifying patterns
indicative of potential electrical panel res.
In this study, four different ML algorithms, namely DT, GPC, SVM,
and GNB, were comparatively analyzed, focusing on their application in
the early detection of electrical panel res. After the analysis and model
training process, the best hyperparameters were determined for each
algorithm. After extensive testing and evaluation, these algorithms were
compared on the basis of performance metrics such as accuracy, preci-
sion, recall, F1 score, and AUC value.
The results indicate that GPC, DT, and SVM algorithms exhibit high
accuracy and generalization capabilities for early anomaly detection in
electrical panels. In particular, the GPC algorithm demonstrates the
highest model performance, while the GNB algorithm exhibits the
lowest generalization and model performance. The ROC analysis results
reveal that GPC, SVM, and DT algorithms can effectively distinguish
between positive and negative classes. These models have different ad-
vantages and weaknesses. The preferred model may vary depending on
the hardware, speed, application requirements, and performance
criteria. Moreover, the results demonstrate the applicability of these
models in predictive maintenance frameworks aimed at improving re
safety protocols in electrical panels. Additionally, they provide valuable
insights into integrating IoT and ML technologies into preventive
maintenance strategies. Furthermore, this research will serve as an
important resource for improving feature and model selection, as well as
hyperparameter tuning, for future electrical panel re detection and
similar classication tasks.
Conict of interest
∘All authors have participated in (a) conception and design, or anal-
ysis and interpretation of the data; (b) drafting the article or revising
it critically for important intellectual content; and (c) approval of the
nal version.
∘ This manuscript has not been submitted to, nor is under review at,
another journal or other publishing venue.
∘ The authors have no afliation with any organization with a direct or
indirect nancial interest in the subject matter discussed in the
manuscript.
References
[1] N. Al-Azzam, I. Shatnawi, Comparing supervised and semi-supervised machine
learning models on diagnosing breast cancer, Ann. Med. Surg. 62 (2021) 53–64.
[2] A.A.S. Alari, H.N. Phylaktou, G.E. Andrews, What kills people in a re? heat or
smoke? 9th SSC Proc. (2016).
[3] J. Ananthi, N. Sengottaiyan, S. Anbukaruppusamy, K. Upreti, A.K. Dubey, Forest
re prediction using IoT and deep learning, Int. J. Adv. Technol. Eng. Explor. 9
(87) (2022) 246–256.
[4] M. Baghoolizadeh, S.A.H.H. Dehkordi, M. Rostamzadeh-Renani, R. Rostamzadeh-
Renani, N.K. Azarkhavarani, D. Toghraie, Optimization of annual electricity
consumption costs and the costs of insulation and phase change materials in the
residential building using articial neural network and genetic algorithm methods,
J. Energy Storage 62 (2023) 106916.
[5] S.M. Basha, D.S. Rajput, Survey on evaluating the performance of machine learning
algorithms: past contributions and future roadmap. In Deep Learning and Parallel
Computing Environment for Bioengineering Systems, Academic Press, 2019,
pp. 153–164.
[6] C.M. Bishop, N.M. Nasrabadi, Pattern recognition and machine learning, Vol. 4,
springer, New York, 2006, p. 738.
[7] J. Cai, J. Luo, S. Wang, S. Yang, Feature selection in machine learning: A new
perspective, Neurocomputing 300 (2018) 70–79.
[8] S. Chen, J. Ren, Y. Yan, M. Sun, F. Hu, H. Zhao, Multi-sourced sensing and support
vector machine classication for effective detection of re hazard in early stage,
Comput. Electr. Eng. 101 (2022) 108046.
[9] D. Colledani, P. Anselmi, E. Robusto, Machine learning-decision tree classiers in
psychiatric assessment: An application to the diagnosis of major depressive
disorder, Psychiatry Res. 322 (2023) 115127.
[10] I. El Naqa, M.J. Murphy, What Is Machine Learning? in: I. El Naqa, R. Li,
M. Murphy (Eds.), Machine Learning in Radiation Oncology Springer, Cham, 2015
https://doi.org/10.1007/978-3-319-18305-3_1.
[11] M.C. Frunza, Support Vector Machines. In Solving Modern Crime in Financial
Markets, Academic Press, 2016, pp. 205–215.
[12] R. Gholami, N. Fakhari, Support vector machine: principles, parameters, and
applications. In Handbook of neural computation, Academic Press, 2017,
pp. 515–535.
[13] P. Giudici, A. Gramegna, E. Rafnetti, Machine Learning Classication Model
Comparison, Socio-Econ. Plan. Sci. (2023) 101560.
[14] T. Goswami, Machine learning behind classication tasks in various engineering
and science domains, Cogn. Inform., Comput. Model., Cogn. Sci. (2020) 339–356.
[15] I. Guyon, J. Weston, S. Barnhill, V. Vapnik, Gene selection for cancer classication
using support vector machines, Mach. Learn. 46 (2002) 389–422.
[16] A. Güran, M. Uysal, ¨
O. Do˘
grus¨
oz, Effects of Support Vector Machines Parameter
Optimization on Sentiment Analysis, Dokuz Eylül Univ. Fac. Eng. J. Sci. Eng. 16
(48) (2014) 86–93.
[17] Y. Hong, J. Kang, C. Fu, Rapid prediction of mine tunnel re smoke movement with
machine learning and supercomputing techniques, Fire Saf. J. 127 (2022) 103492.
[18] A.N. Huda, S. Taib, M.S. Jadin, D. Ishak, A semi-automatic approach for
thermographic inspection of electrical installations within buildings, Energy Build.
55 (2012) 585–591.
[19] A.S.N. Huda, S. Taib, K. Ghazali, M.S. Jadin, A new thermographic NDT for
condition monitoring of electrical components using ANN with condence level
analysis, ISA Trans. 53 (3) (2014) 717–724.
Fig. 14. Performance metric graphs of algorithms by class a) N class b) S1 class.
M.F. Peks¸en et al.
Alexandria Engineering Journal 96 (2024) 112–123
123
[20] ˙
Istanbul Büyüks
¸ehir Belediyesi ˙
Itfaiye Daire Bas
¸kanlı˘
gı (2020). ˙
Istatistikler 2020.
http://itfaiye.ibb.gov.tr/img/_152291222020_.pdf.
[21] B. Kim, J. Lee, W.G. Shin, Sensitivity analysis of input variables to establish re
damage thresholds for redundant electrical panels, Nucl. Eng. Technol. 54 (1)
(2022) 84–96.
[22] A. Kumar, L. Bi, J. Kim, D.D. Feng, Machine learning in medical imaging, Biomed.
Inf. Technol. (2020) 167–196.
[23] J. Lee, F. Joglar, U. Farradj, A. Ratchford, Fire risk assessment to develop a
compliance strategy for sample redundant electrical panels at a nuclear power
plant based on deterministic re protection requirements, Prog. Nucl. Energy 128
(2020) 103467.
[24] M. Levy, F. Naiser, Machine learning at the edge. In Softwre Engineering for
Embedded Systems, Newnes, 2019, pp. 549–601.
[25] L. Li, W.K. Ching, Z.P. Liu, Robust biomarker screening from gene expression data
by stable machine learning-recursive feature elimination methods, Comput. Biol.
Chem. 100 (2022) 107747.
[26] W. Lian, G. Nie, B. Jia, D. Shi, Q. Fan, Y. Liang, An intrusion detection method
based on decision tree-recursive feature elimination in ensemble learning, Math.
Probl. Eng. 2020 (2020) 1–15.
[27] H. Liu, M. Zhou, Q. Liu, An embedded feature selection method for imbalanced
data classication, IEEE/CAA J. Autom. Sin. 6 (3) (2019) 703–715.
[28] Z. Liu, Y. Gao, B. Liu, An articial intelligence-based electric multiple units using a
smart power grid system, Energy Rep. 8 (2022) 13376–13388.
[29] Malaysia Fire and Rescue Department (2021). Annual report 2021. https://www.
bomba.gov.my/wp-content/uploads/2023/01/Laporan-Tahun-2021_compressed.
pdf.
[30] P. Mccaffrey, Introduction to machine learning: Support vector machines, tree-
based models, clustering, and explainability, Introd. Healthc. Inform. (2020)
211–225.
[31] A. Mellit, M. Benghanem, S. Kalogirou, A.M. Pavan, An embedded system for
remote monitoring and fault diagnosis of photovoltaic arrays using machine
learning and the internet of things, Renew. Energy 208 (2023) 399–408.
[32] M.A. Mendes, L.G.R. Tonini, P.R. Muniz, C.B. Donadel, Thermographic analysis of
parallelly cables: A method to avoid misdiagnosis, Appl. Therm. Eng. 104 (2016)
231–236.
[33] J. Miao, L. Niu, A survey on feature selection, Procedia Comput. Sci. 91 (2016)
919–926.
[34] NFPA 70B, Recommended Practice for Electrical Equipment Maintenance, 2019
edition, Section 11, https://link.nfpa.org/free-access/publications/70B/2019.
[35] C.D. O’Connor, D.E. Calkin, M.P. Thompson, An empirical machine learning
method for predicting potential re control locations for pre-re planning and
operational re management, Int. J. Wildland re 26 (7) (2017) 587–597.
[36] G. Parise, E. Hesla, C.S. Mardegan, L. Parise, E.B. Capaccini, Measures to minimize
series faults in electrical cords and extension cords, IEEE Trans. Ind. Appl. 55 (5)
(2019) 4551–4556.
[37] C. Park, J.B. Lee, W. Park, D.K. Lee, Fire accelerant classication from GC–MS data
of suspected arson cases using machine–learning models. Forensic Sci. Int. (2023)
111646.
[38] D. Petschke, T.E. Staab, A supervised machine learning approach using naive
Gaussian Bayes classication for shape-sensitive detector pulse discrimination in
positron annihilation lifetime spectroscopy (PALS), Nucl. Instrum. Methods Phys.
Res. Sect. A: Accel., Spectrometers, Detect. Assoc. Equip. 947 (2019) 162742.
[39] V.C. Pezoulas, T.P. Exarchos, D.I. Fotiadis, Machine learning and data analytics,
Med. Data Shar., Harmon. Anal. (2020) 227–309.
[40] P. Pierleoni, L. Palma, A. Belli, L. Sabbatini, Using plastic injection moulding
machine process parameters for predictive maintenance purposes. In 2020.
International Conference on Intelligent Engineering and Management (ICIEM),
IEEE, 2020, June, pp. 115–120.
[41] D.A. Pisner, D.M. Schnyer, Support vector machine. In Machine learning, Academic
Press, 2020, pp. 101–121.
[42] Z. Qin, W. Li, X. Xiong, Estimating wind speed probability distribution using kernel
density method, Electr. Power Syst. Res. 81 (12) (2011) 2139–2146.
[43] L.C. Ribeiro, E.L. Bonaldi, L.E.L. De Oliveira, L.B. da Silva, C.P. Salomon, W.
C. Santana, G. Lambert-Torres, Equipment for predictive maintenance in
hydrogenerators, AASRI Procedia 7 (2014) 75–80.
[44] Sahid, D., & Alaydrus, M. (2020, September). Multi sensor re detection in low
voltage electrical panel using modular fuzzy logic. In 2020 2nd International
Conference on Broadband Communications, Wireless Sensors and Powering
(BCWSP) (pp. 31-35). IEEE.
[45] Santiago Egea, Albert Rego Manez, Belen Carro, Antonio Sanchez-Esgunevillas,
Jaime Lloret, Intelligent IoT trafc classication using novel search strategy for
fast-based-correlation feature selection in industrial environment, IEEE Internet
Things J. vol. 5 (3) (JUNE 2018).
[46] G. Sarailidis, T. Wagener, F. Pianosi, Integrating scientic knowledge into machine
learning using interactive decision trees, Comput. Geosci. 170 (2023) 105248.
[47] E.M. Senan, M.H. Al-Adhaileh, F.W. Alsaade, T.H. Aldhyani, A.A. Alqarni,
N. Alsharif, M.Y. Alzahrani, Diagnosis of chronic kidney disease using effective
classication algorithms and recursive feature elimination techniques, J. Healthc.
Eng. 2021 (2021).
[48] N.V. Sharma, N.S. Yadav, An optimal intrusion detection system using recursive
feature elimination and ensemble of classiers, Microprocess. Microsyst. 85 (2021)
104293.
[49] G. Shobha, S. Rangaswamy, Chapter 8-Machine Learning Handbook of Statistics,
Elsevier, 2018.
[50] A. Sol´
orzano, J. Eichmann, L. Fern´
andez, B. Ziems, J.M. Jim´
enez-Soto, S. Marco,
J. Fonollosa, Early re detection based on gas sensor arrays: multivariate
calibration and validation, Sens. Actuators B: Chem. 352 (2022) 130961.
[51] S. Sun, D. Gura, B. Dong, Fire safety assessment models based on machine learning
methods for the coal industry, Chemom. Intell. Lab. Syst. 231 (2022) 104693.
[52] S. Taib, M.S. Jadin, S. Kabir, Thermal imaging for qualitative-based measurements
of thermal anomalies in electrical components. 2011 Saudi International
Electronics, Communications and Photonics Conference (SIECPC), IEEE, 2011,
April, pp. 1–6.
[53] M. Wadinger, M. Kvasnica, Adaptable and interpretable framework for anomaly
detection in SCADA-based industrial systems, Expert Syst. Appl. (2024) 123200.
[54] J. Wang, C. Zhao, A probabilistic framework with concurrent analytics of Gaussian
process regression and classication for multivariate control performance
assessment, J. Process Control 101 (2021) 78–92.
[55] Z. Wang, T. Zhang, X. Wu, X. Huang, Predicting transient building re based on
external smoke images and deep learning, J. Build. Eng. 47 (2022) 103823.
[56] C.K. Williams, Computation with innite neural networks, Neural Comput. 10 (5)
(1998) 1203–1216.
[57] H. Xu, Coordinated distribution of stadium electric load unloading based on
articial intelligence, Energy Rep. 8 (2022) 10594–10603.
[58] Z. Xu, D. Liu, L. Yan, Temperature-based re frequency analysis using machine
learning: a case of Changsha, China, Clim. Risk Manag. 31 (2021) 100276.
[59] I. Yahyaoui, N. Yumus¸ ak, Machine learning techniques for data classication.
Advances in renewable energies and power technologies, Elsevier, 2018,
pp. 441–450.
[60] X.-S. Yang, Data mining techniques. Introduction to Algorithms for Data Mining
and Machine Learning, 109–128. 2019.
[61] P. Zavaleta, Experimental study at reduced-scale of re spread between electrical
cabinets located opposite each other, Fire Saf. J. 122 (2021) 103319.
[62] J. Zhang, L. Huang, T. Chen, G. Su, Simulation based analysis of electrical re risks
caused by poor electric contact between plug and receptacle, Fire Saf. J. 126
(2021) 103434.
[63] B. Zhao, W. Zhang, Y. Zhang, C. Zhang, C. Zhang, J. Zhang, Research on the
remaining useful life prediction method for lithium-ion batteries by fusion of
feature engineering and deep learning, Appl. Energy 358 (2024) 122325.
[64] Z. Zhao, F. Morstatter, S. Sharma, S. Alelyani, A. Anand, H. Liu, Advancing feature
selection research, ASU Feature Sel. Repos. (2010) 1–28.
[65] I.A. Zulfauzi, N.Y. Dahlan, H. Sintuya, W. Setthapun, Anomaly detection using K-
means and long-short term memory for predictive maintenance of large-scale solar
(LSS) photovoltaic plant, Energy Rep. 9 (2023) 154–158.
M.F. Peks¸en et al.
