INTELLIGENT HOME: EMPOWERING SMART HOME WITH
MACHINE LEARNING FOR USER ACTION PREDICTION
Ayad
A. Saleem a,*, Masoud M. Hassan b, Ismael A. Ali b
a Technical College of Petroleum and
Mineral Sciences, Duhok Polytechnic University, Zakho, Kurdistan Region, Iraq -
ayad.abdulrahman@dpu.edu.krd
b CCNP Research Lab, Department of
Computer Science, Faculty of Science, Zakho, Kurdistan Region, Iraq - (masoud.hassan, ismael.ali)@uoz.edu.krd
ABSTRACT:
Smart home is
an emerging technology that is transforming the way people live and interact
with their homes. These homes are equipped with various devices and
technologies that allow the homeowner to control, monitor, and automate various
aspects of their home. This can include lighting, heating and cooling, security
systems, and appliances. However, to enhance the efficiency of these homes,
machine learning algorithms can be utilized to analyze the data generated from
the home environment and adapt to user behaviors. This paper proposes a smart
home system empowered by machine learning algorithms for enhanced user behavior
prediction and automation. The proposed system is composed of three modes,
including manual, automatic, and intelligent, with the objectives of maximizing
security, minimizing human effort, reducing power consumption, and facilitating
user interaction. The manual mode offers control and monitoring capabilities
through a web-based user interface, accessible from anywhere and at any time.
The automatic mode provides security alerts and appliances control to minimize
human intervention. Additionally, the intelligent mode employs machine learning
classification algorithms, such as decision tree, K-nearest neighbors, and
multi-layer perceptron, to track and predict user actions, thereby reducing user
intervention and providing additional comfort to homeowners. Experiments
conducted employing the three classifiers resulted in accuracies of 97.4%,
97.22%, and 97.36%, respectively. The proposed smart home system can
potentially enhance the quality of life for homeowners while reducing energy
consumption and increasing security.
KEYWORDS: Smart home, Machine Learning,
Raspberry Pi, Decision Tree, K-Nearest Neighbors, Multi-Layer Perceptron, ANN,
User Behavior.
1.
Introduction
The
remarkable advancement in technology has facilitated the ability to establish a
connection between any device and the Internet, thus giving rise to the notion
of the Internet of Things (IoT). The IoT refers to a network of
internet-connected devices and objects that range from simple household
appliances to complex machinery in various settings (Saleem et al., 2022). These devices collect and share
data, which enables them to function seamlessly and offer insights into
different aspects of human life (Ibrahim et al., 2022; Taiwo et al.,
2022). The integration of
IoT has led to enhanced efficiency, convenience, and optimization in both
residential and commercial settings, resulting in increased revenue and
improved customer service for businesses. The application areas for the IoT
encompass various fields such as healthcare, smart homes, smart cities,
industrial automation, and transportation. Among these, and since the smart
home pertains more closely to individuals' daily lives, the smart home has
garnered significant interest from both the industrial and academic communities
(Jabbar et al., 2018).
The use of
smart homes is growing rapidly and is expected to continue to grow in the
future as technology continues to advance and become more accessible. The
development of smart homes is being driven by the increasing demand for home
automation and the growing need for energy efficiency. However, there are
challenges associated with the implementation of smart homes, including cost
and security concerns (Jabbar et al., 2019). Nevertheless, the potential benefits
of smart homes are substantial and are likely to result in significant positive
impacts on the quality of life for homeowners. The goal of smart homes is to
increase comfort, convenience, safety, and efficiency while reducing energy
consumption. The provision of a user interface (UI) for home control and
monitoring is crucial, with a preference for web-based applications accessible
at any time and from anywhere via the internet. Such an interface should be
user-friendly and compatible with all major operating system platforms,
including Android, Windows, and iOS, to facilitate the way the user interacts
with their smart homes. In addition to enhancing user comfort, home control
systems can also promote energy conservation by enabling automatic control. To
ensure maximum security, the system should also include real-time notifications
and alarms.
The current
era of the Fourth Industrial Revolution (Industry 4.0 or 4IR) has resulted in
an abundance of digital data, including IoT data, business data, health data,
mobile data, social media data, and cybersecurity data, among others. The
effective analysis of these data and the creation of related automated and
smart applications require a solid understanding of Artificial Intelligence
(AI), particularly Machine Learning (ML). Within ML, there are several
different algorithms such as supervised, unsupervised, semi-supervised, and
reinforcement learning. Of particular significance is deep learning, a subset
of ML that possesses the capability to analyze vast amounts of data
intelligently (Sarker, 2021a). In order to increase the
intelligence of the smart home, ML methods can be utilized. ML algorithms are
increasingly being integrated into smart homes to enhance automation and
improve the user experience. These algorithms allow for real-time analysis of
large amounts of data and personalization based on user behavior and
environmental factors. Smart homes powered by ML are a new and rapidly evolving
field, combining the benefits of smart homes with the power of AI. ML algorithms
can be used to analyze data generated by smart home systems, such as energy
usage patterns, occupancy information, and sensor readings, and hence make
predictions about future behavior. ML models have proven to be an effective
tool in enabling smart home automation to achieve a multitude of objectives.
These include detecting and recognizing objects, human activities, and faces,
as well as controlling household appliances intelligently, optimizing energy
consumption, monitoring homes, and enhancing safety and security measures (Taiwo et al., 2022). Smart home employs several ML classification
algorithms in its applications, including but not limited to K-Nearest
Neighbors (KNN), Decision Tree (DT), Support Vector Machine (SVM), Naive Bayes,
and Random Forest. In addition, deep learning (DL) models are also utilized in
smart home applications (Saleem et al., 2022). In this work, ML classification
model is employed to adapt user behavior and learn from performed actions.
This paper
aims to design and implement an intelligent home system empowered by ML methods
for enhanced user behavior prediction and automation. The proposed system aims
to reduce user effort, power consumption, and human intervention, while
increasing security. It includes a web-based UI accessible from any operating
system platform to control various home appliances such as lights, TV, air
conditioners, window, door, curtain, with the capability to be controlled
manually by the user through a friendly UI. The proposed system is also capable
of performing automatic tasks without user intervention, such as turning on/off
lights in order to reduce power consumption. Furthermore, it provides real-time
voice and text message alerts through the UI and email for detecting intruders
and fire inside the home. Additionally, the system incorporates the power of
artificial intelligence, especially ML algorithms, to increase the system
intelligence. This work makes use of various sensors and user actions to
predict user preferences, particularly regarding controlling the curtain using
a classification model. Additionally, this work applies different classifiers
to identify the best model to be implemented in the proposed system. The system
provides a responsive UI to monitor and control the home environment. The
proposed system offers a smart, secure, and user-friendly solution for
intelligent home automation.
The rest of
this paper is structured as follows. Section 2 presents a review of related
work to enable a comparison with the proposed system. Section 3 provides a
background on the fundamentals of smart home, ML, and the various tools
utilized in this study. Section 4 details the proposed method, including system
architecture and design considerations. Section 5 presents the results obtained
from applying various ML algorithms to the created dataset, along with a
discussion of their performance. Finally, Section 6 provides the conclusion,
highlights the contributions of this work, and outlines potential directions
for future research.
2.
RELATED WORKS
In this section, numerous state-of-the-art smart home systems are
reviewed to provide a comparison to the proposed system.
Various smart
home systems were proposed for controlling home appliances and monitoring the
home environment without making use of the power of ML algorithms. For
instance, (Okorie et al., 2020) proposed a smart, cost-effective
system for controlling home appliances and monitoring the status of diverse
sensors using smart Android phones, in order to help the elderly and people
with disabilities to live their lives in the easiest way. The system consists
of an Arduino connected to sensors (light and temperature sensors), appliances
(TV, fan, light), and a Bluetooth module to provide communication between the
Arduino and the smartphone. This system can be controlled by any android-based
device using the "arduDroid" application, which is capable of sending
controlling commands to the Arduino, such as turning on/off the TV, fan, and
light. In addition, it receives information from Arduino related to the value
of temperature and light intensity sensors. The system is easy to use and
implement, and costly effective. However, this system is controlled only via
Android devices.
In another
study, home control and monitoring system based on the web application UI was
presented (Iqbal et al., 2018). Their system comprises three main
functionalities: door control, fan and light control, and water pump control.
In the door control functionality, a motion sensor, a camera, and a
Light-Emitting Diode (LED) are installed in front of the door. As soon as any
motion is detected, the RPi will receive a signal, and then it will turn the
light on automatically in case it is night time. In addition, the camera will
take a picture and then store it in the database for security purposes. Thus,
the user can open the door via a UI. Furthermore, in the fan and light
controlling functionality, the user can turn on/off lights and the exhaust fan.
A humidity and temperature sensor are installed to read the temperature and
humidity status and show them on the web page for the user to decide whenever
to turn on the exhaust fan. Finally, in the water pump controlling
functionality, two water level sensors are installed on the top and bottom of
the water tank, which are connected to the Arduino to send the signals through
ZigBee to the RPi to decide when to turn on the water pump automatically. The
data from the sensors' and actuators' status is temporarily stored on the MySQL
database in RPi and then backed up on the cloud server to be ready for a user
history report or even a third-party service provider. As an advantage of the
system, an online UI is used to make the system capable of being accessed
anywhere and anytime, and the authors took into consideration sensors' status
before controlling the actuators. Although the system has multiple features, no
ML method is used.
The authors
in (Pavithra & Balakrishnan, 2015) proposed a smart home system for the
purpose of controlling home appliances and detecting fire. The system is
designed to operate through the use of a central controller, which is based on
the Raspberry Pi (RPi) platform. Specifically, the system employs relays to
switch on and off lights when an infrared (IR) sensor detects an object, and a
Passive Infrared (PIR) sensor to switch on the fan when motion is detected.
Additionally, the system integrates a fire detection sensor that, when
activated, triggers a camera to capture an image of the fire in order to send it
along with an alarm message to the user's phone. The phone subsequently makes
an automatic call to the closest fire station. The control and monitoring of
the system's lights and fans can be managed manually through a web page UI,
which is accessible from any operating system platform. It is worth noting that
the system's architecture does not involve the use of ML algorithms.
Consequently, the system does not have the capability to learn or adapt to
changing environmental conditions. While the system's design is commendable in
its utilization of a central controller for home automation and fire detection,
it does not exhibit intelligent behavior as ML algorithms would provide.
The authors
in (Gota et al., 2020) designed and implemented a home
automation system by controlling lighting, doors, and windows, as well as
monitoring temperature and humidity inside and outside the home. The process of
controlling and monitoring was done through a web page to make the system
usable on different operating systems. In order to control the windows and
doors inside the house, a servo motor was used, which can rotate at an angle of
0 to 180 degrees. Another type of motor called a stepper motor was used to
control the opening and closing of the garage door, as this type has the
ability to rotate in two directions and at multiple angles. The motors stop
when they reach a certain point that is sensed by the magnetic sensor. For
manual or automatic ventilation, the temperature inside and outside the house
was read along with the humidity reading. Thus, when a certain temperature is
reached, the ventilation system is turned on automatically or manually via the
web page.
On the other
hand, several smart homes systems have been proposed that incorporate ML
algorithm for controlling home appliances and enhancing home security. For
example, (Abbas & Abdullah, 2021) proposed an innovative approach
towards tracking and predicting user behavior using ML algorithms. The system
comprises two essential components, a RPi and a NodeMCU ESP32, which work in
tandem to perform the intended function. The television is connected to the
ESP32 via a relay, while the pushbutton serves as a switch to activate or
deactivate the TV. The RPi records and archives the device's status and user
behavior, which are then subjected to a classification process. The decision to
turn the television on or off is derived through the use of DT algorithm. This
model is adaptable, allowing for updates to be made with newly collected daily
data. It is noteworthy that the system does not include a UI and does not
utilize any wireless communication technology.
In another
study by (Mehmood et al., 2019), the authors devised a deep
learning-based system for the detection of individuals. Their system utilized
Amazon Web Service to facilitate remote monitoring by the user. The system
comprises a camera module, connected to a RPi, which captures live video and
employs the Single Shot MultiBox Detector (SSD) algorithm trained with the
Microsoft Common Objects in Context (COCO) dataset for video analysis. Upon
detection of a person, notifications are sent to the user either through Short
Message Service (SMS) or email. Additionally, a control message is transmitted
to the actuating device, a Node MicroController Unit (NodeMCU) ESP8266, to turn
on or off a LED. The system's accuracy is impacted by decreasing light intensity,
increasing distance, and larger frame size, resulting in a reduction from
95-100 to 80-85 percentage.
(Crisnapati et al., 2016) proposed a smart home monitoring and
control system that utilizes fuzzy logic to decrease energy consumption. The
system features an HD camera that is integrated with a motion detection system,
a motion sensor, and RFID for security. In this system, the user can manage
recorded videos and captured photos, control functions (turning lights on and
off), and monitor the home through a user-friendly web interface. Furthermore,
the system's artificial intelligence capabilities allow it to automatically
control temperature and lighting to promote energy efficiency.
The authors
in (Paredes-Valverde et al., 2020) presented a sophisticated system
designed for power consumption monitoring and management, called IntelliHome.
The system utilizes data obtained from the usage of IoT devices such as
electrical appliances, sensors, and smart switches, to provide energy-saving
recommendations to the user based on their behavior and preferences. Upon
accepting the recommendation, the system implements control over home appliances.
The analysis and processing of the data are handled by the Home Energy
Consumption Monitor (HECM) module, which utilizes the Holt-Winters-RNN
algorithm. Although the system provides a UI, it is only compatible with
Android devices.
In (Peng et al., 2019), the authors propose an intelligent
home-control system with the utilization of a convolutional neural network
(CNN) to recognize and classify human gestures (human point attitude). The
system makes use of an Arduino as a central controller for controlling the
window, air conditioner, and LED. The gestures are captured via a Kinect sensor
that is connected to a personal computer, which processes the obtained data and
sends resulting information via Zigbee wireless technology to the Arduino to
execute controlling commands. Finally, the obtained results were excellent for
all gesture classifications. The system does not have a user interface.
The authors in (Raju et al., 2021) presented a smart
home system that enables monitoring of the home, control of appliances such as
lights and fans, and prediction of power conservation. This system is built
using a Raspberry Pi, various sensors (including motion, temperature, light,
and sound), and actuators. Users can control their appliances via a mobile
application that sends commands to the Raspberry Pi, which then sends control
commands to the relay module to turn the appliances on or off. The system can
also automatically control appliances, and users can view the energy
consumption of each device via the user interface. To predict energy
consumption, the system employs several machine learning algorithms, including
decision tree regression, KNN, support vector regression, and random forest
regression. The PIR sensor is used to control both the light and the fan, and
Bluetooth communication is used to connect the mobile phone with the Raspberry
Pi for wireless communication. While the system has a user interface, it cannot
be accessed via the internet. Notably, the Decision Tree algorithm was found to
provide the highest accuracy among all applied algorithms.
3.
BACKGROUND
This section provides a brief overview of smart home systems,
and machine learning methods, along with the various tools and devices utilized
in this work.
3.1
Smart Homes
As defined by (Marikyan et al., 2019),
a smart home “is a residence equipped with smart technologies aimed at
providing tailored services for users”. A "smart home" is referred to
as a home automation system that is designed with controlling, monitoring, and
sensing functionalities such as surveillance, ventilation, lighting, and
conditioning systems. These smart systems consist of various essential
components, including actuators and sensors connected wired or wirelessly to a
central controller (Saleem et al., 2022).
This controller receives data from sensors, and sends controlling commands to
the actuators. Controlling commands are provided either by the user via a UI or
automatically based on some pre-programmed conditions. Furthermore, householders
can monitor the houses via a graphical UI using a tablet, computer, or
smartphone (Saleem et al., 2022).
Smart home as one of the most common IoT applications (Nikou, 2019), provides access to the
components remotely anytime and anywhere via any smart devices (Choi et al., 2021). Smart homes offer
diverse services to the homeowner, including:
· Remote monitoring: monitor the environment inside and around
the home from anywhere and at any time. For instance, monitoring the devices'
status (on/off), and sensors' status (readings).
· Remote controlling: controlling the home appliances anytime
and anywhere via the UI.
· Reducing human effort: the user can control home appliances
by their phones without any physical movement. For example, switching on the
air conditioner, TV, and lights, as well as opening the door.
· Reducing power consumption: for example, turning the lights
off during the daytime. In addition, turning the lights on when motion is
detected somewhere inside the home would lead to saving energy.
· Security maximizing: integrating a security option in the
smart home is essential to prevent a home from being stolen and to provide a
real-time alert. For example, installing a surveillant camera in addition to
setting up a motion detector to alert the householder by sending an alarming message
via email or phone calls, as well as activating buzzers.
3.2
Raspberry Pi
Raspberry Pi (RPi) is a small-sized highly performance single
board computer (SBC) that has all standard computer components including
processor, RAM, I/O units, and GPU, incorporated in a single board (Jolles, 2021; Pajankar, 2021).
In addition, as compared to the available SBCs in the market, the RPi is one of
the most common SBCs and the best-selling computers in the world (Pajankar, 2021). Unlike
traditional computers, the main disadvantage of the RPi is that the hardware
component cannot be upgraded (Pajankar, 2021). Like the other
SBCs and due to their suitable size and performance, the RPi is essentially
employed in embedded systems, especially for robotics and IoT applications. The
first model of RPi was released in February 2012 and was developed by the RPi
foundation in the United Kingdom (Jolles, 2021; Kurniawan, 2019).
Among all RPi model, the newest model is the RPi 4 model B which comes with the
following specifications: 64-bit quad-core Cortex- A72 1.5GHz Broadcom BCM2711
processor. LPDDR4 3200 SDRAM comes with 1GB, 2GB, 4GB, or 8GB of RAM. Dual-band
2.4GHz and 5GHz wireless networking IEEE 802.11ac, gigabit ethernet, and
Bluetooth 5, two USB 3.0 and two USB 2.0 ports. An array of 40-pin headers
(described in Figure 2), 28-pins out of these
40 are General Purpose Input/Output (GPIO) used to connect sensors and
actuators for controlling and monitoring purposes. Two micro-HDMI ports support
up to 4k video streaming. Serial interface port to connect RPi camera, microSD
card slot, and it operates with 5V 3A DC input power, (See Figure 1). Although RPi has its
official operating system called Raspberry Pi OS (Raspbian OS, previously) (Jolles, 2021), it can run with other
OS including android, windows 10 IoT, and ubuntu OS family.
Figure 1: Raspberry Pi 4 model B on board components (Pajankar, 2021).
Figure 2: Raspberry Pi 40-pins header description (Raspberry Pi Documentation, n.d.).
As illustrated in Figure 2, the RPi consists of 40 pins
defined as follows:
· Two power pins provide 5V, and two power pins provide 3.3V.
· 8 unconfigurable ground pins provide 0V.
· The other 28 pins called GPIO pins, which can be used either
as output and can be set from 3.3V (high) to 0V (low), or as input which can
read also from 3.3V to 0V.
Machine Learning (ML) is a branch of AI that consist of a set
of algorithms and techniques that enable computer systems to learn and make
predictions or decisions to learn from available data based on previous
experiences (Bertolini et al., 2021)
without being explicitly programmed. ML methods are used in diverse areas
including object detection, text and speech interpretation, classification and
pattern recognition (Singh et al., 2020). Prior to the
implementation of the algorithm on a specific problem, the algorithm is trained
on a dataset (available data) to result in the most accurate model. The dataset
contains a number of columns called attributes or features, in addition to the
output variable (in case of supervised learning). Furthermore, the dataset
should be split into two sets, one used for training the model, called the
training subset. In contrast, the other called the testing subset utilized for
testing the model accuracy (Bertolini et al., 2021).
Moreover, the dataset features can be continuous, binary, or categorical (Singh et al., 2020), and the output
variable could be continuous or categorical (Bertolini et al., 2021).
The types of available data help to choose the best-fit algorithm for the
available case.
ML consists of several algorithms, each of which is intended
to solve specific kinds of problems, such as classification (in case of
categorical output), regression (in case of continuous output), and clustering.
There are three main types of ML: supervised learning, unsupervised learning,
and reinforcement learning. Classification and regression, are methods of
supervised learning, in which the data are labelled (consisting of input and
output) (Bertolini et al., 2021; Singh et al., 2020). While the clustering is an unsupervised learning (output
variables are missing) (Bertolini et al., 2021; Singh et al., 2020). The classification (predictive learning) comprises various
algorithms including Decision Tree (DT), Naive Bayes (NB), Random Forest (RF),
Artificial Neural Network (ANN), K-Nearest Neighbors (KNN), Logistic Regression
(LR), and many other algorithms that utilized for prediction (Bertolini et al., 2021; Singh et al., 2020). While unsupervised learning, which is also called
descriptive learning, analyzes the given dataset and intends not to predict the
missing output, but to discover the hidden pattern behind the given data (Bertolini et al., 2021).
For example, in clustering, the data are divided into several groups, each of
which contains data that are similar to each other, but differ from the other
groups (Bertolini et al., 2021).
The most common clustering algorithm is K-means. Semi-supervised learning is
the combination of both supervised and unsupervised learning which can handle
labelled and unlabelled data. It is useful when the dataset contains a small
number of labelled data and a large number of unlabelled ones (Sarker, 2021b; van Engelen & Hoos, 2020). Further, it is located between supervised and unsupervised
learning. As it aims to enhance the performance of one of the previous
techniques by making use of data corresponding to the other. For example, in
case of handling classification problems, a large amount of unlabelled samples
are utilized to improve the classification process. While in clustering,
labelled observations can be used to improve the clustering process as well (van Engelen & Hoos, 2020).
A fraud detection and machine translation considered application of
semi-supervised method (Sarker, 2021b).
ML has a wide range of applications in various industries,
such as finance, healthcare, retail, and many others. It is used to solve a
variety of problems, such as predictive analytics, image recognition, natural
language processing, robotics, spam filtering, face recognition, handwriting
and speech recognition, DNA classification, and computer games (Singh et al., 2020). In the proposed
work, supervised ML classification is used in the smart home system to predict
the user action regarding to opening and closing the curtain. In this work,
several classifiers, including DT, KNN, and MLP, are applied to the Curtain
dataset to predict user action. These three classifiers were selected to
perform the comparison among the tree-based algorithm, the distance-based
algorithm, and the optimization-based algorithm in order to recognize the impact.
3.4
The Decision Tree (DT)
The Decision Tree (DT) is a tree-like graphical
representation classifier used for supervised learning in both regression and
classification tasks (Géron, 2019; Mukherjee et al., 2019).
It consists of nodes, with decision nodes representing attributes and leaf
nodes representing class labels, that are connected via arrows, namely directed
edges. Iterative Dichotomiser 3 (ID3) and C4.5 are the most commonly used
algorithms for constructing DTs (Mienye et al., 2019).
ID3 is used only for categorical data, while C4.5 can be used for both
numerical and categorical data (Mienye et al., 2019).
To construct a DT, impurity measures (such as Entropy or Gini) and information
gain are utilized for each feature, whereas the most informative attribute with
the maximum information gain is selected as the root node (Mienye et al., 2019).
This process is repeated to determine the best-fit attribute for each node
until all attributes are included in the tree. The resulting DT is then
translated into rules comprising if-then statements (Mukherjee et al., 2019; Patel & Prajapati, 2018). Decision trees have several advantages, including their
interpretability, simplicity, and ability to handle both categorical and
numerical features.
3.5
K-Nearest Neighbors (KNN)
K-Nearest Neighbors (KNN) algorithm is another supervised
learning algorithm, which is considered the simplest ML algorithm as it
requires only storing the training dataset to build the model (Gao & Li, 2020) KNN is utilized
for solving both classification problems for categorical label, as well as
regression problems for numerical label (Jung, 2022). The key parameter of
the KNN algorithm is the number of neighbors that will be considered (Jung, 2022). In this paper, KNN is
employed for classification of user behaviors. In order to predict the class of
the new data example, KNN intends to determine the closest data examples out of
all the examples in the training dataset (Li et al., 2023). The simplest
version is to set K to 1. Thus, it means considering only one nearest neighbor
in order to figure out which class the new data point belongs to. In general,
for a binary classification, it is preferable to set k greater than two; in
this case, voting is utilized to determine the correct label. This implies that
the correct label is the class that has the most frequent neighbors to the new
data point (Mahmoodzadeh et al., 2020).
KNN makes use of distance metrics to calculate the distance between the new
point and all other training set points (Kubat & An, 2021).
Indeed, the three most commonly utilized distance metrics are Euclidean,
Manhattan (also known as city block distance), and Minkowski (Gao & Li, 2020). The Minkowski is
a generalization of both Euclidean (when 
= 2) and Manhattan
distance metric (when = 1) (Gao & Li, 2020).
3.6
Artificial Neural Network
(ANN)
Artificial Neural Network (ANN) is a complex system that
operates similarly to the human brain and nervous system which can provide
self-learning (Desai & Shah, 2021).
ANN is one of the most well-known ML techniques used for the learning process
and provides satisfying results for complex problems that cannot be easily
interpreted (Desai & Shah, 2021).
Furthermore, ANN can solve various problems, including regression,
classification, and clustering (Heidari et al., 2020).
The learning process can be performed via several iterations called epochs (Desai & Shah, 2021),
each iteration consists of two major phases, including the feed-forward and
backpropagation processes. ANNs have various forms to represent, such as
single-layer perceptron (SLP), MLP, and deep learning (Heidari et al., 2020).
SLP consists of only two layers: the input layer and the output layer. It
cannot perform well with patterns that are not linearly separable. For this
reason and in order to alleviate the drawbacks of SLP, MLP was developed.
Unlike SLP, this kind has more than two layers, including an input layer,
hidden layers, and an output layer. The input layer comprises a number of
neurons representing the number of features in the dataset, while the output
layer is composed of only one neuron (Car et al., 2020). The key benefits
of MLP are that it performs well with the availability of noise, provides high
learning accuracy, and handles non-linear separable data (Heidari et al., 2020).
The performance of MLP is highly influenced by various factors, such as wights
vector and learning technique.
MLP is one of the most common FeedForward Neural Networks
(FFNNs) algorithms that offers a high degree of reliability based on the
layered structure of the neural network. The layers in FFNNs consist of nodes
called neurons. Each of which is fully connected by connection links to all
neurons of the next layer (Desai & Shah, 2021; Heidari et al., 2020). These edges are associated with a real number called
weights. Each node can carry out two kinds of functions (Heidari et al., 2020);
summation function, which is calculated by taking the summation of the
production of input values and weights, with adding bias weight. Thus, the
result of the summation function is passed through an activation function to
determine whether the next node will be activated or not (Heidari et al., 2020).
Several activation functions can be applied, including the most common ones
such as Identity (the values stay the same), ReLU (only positive values will be
accepted, and the negative ones will be replaced with zero), Sigmoid (the
output value will always be in the range 0 to 1), and Tanh (mapping the value
of the summation function into the range from -1 to 1). The working mechanism
of MLP is first started from the input layer, which is called forward
propagation, represented by setting initial weights for the edges (Heidari et al., 2020),
the output of the neurons in one layer is considered to be the input of the
next layer neurons, after adding weight values to them (Desai & Shah, 2021).
This process represents one iteration. For repeating iteration, the process is
passed through the backpropagation technique, which operates in reverse from
the output layer, ending in the input layer. This technique provides weight
updating to enhance accuracy and it is repeated until the desired results are
obtained that have the minimum loss (Desai & Shah, 2021; Heidari et al., 2020). This algorithm is used in this study to predict the user
action behaviors.
3.7
Cross Validation
Cross-validation is a statistical technique used to
evaluate the performance of a ML model by testing it on multiple subsets of the
available data. The basic idea behind cross-validation is to split the
available data into two sets: a training set, which is used to fit the model,
and a testing set, which is used to evaluate its performance. This technique is
used to evaluate the model, and to avoid overfitting and underfitting problems.
In general, cross-validation can be applied with two common methods, including
hold-out and 
-Fold (Mahmoodzadeh et al., 2020; Tatsat et al., 2020). The hold-out method divides the original dataset into
training and testing samples, while the -Fold method divides the data into
several folds (subsets) based on the value of , each of which has the same size. One
fold is considered as the testing set and the other 
folds are the training set (Mahmoodzadeh et al., 2020). Therefore, the testing set is evaluated using evaluation metrics (such
as accuracy and F1-score) for each fold independently to gain results. Consequently, the final
result is obtained by calculating the average of the obtained 
(Tatsat et al., 2020). Logically, in the hold-out method, the size percentage of the training
set must be large as compared with the testing set size. Figure 3 demonstrates
the idea behind K-Fold cross-validation (Tatsat et al., 2020).
Figure 3: K-Fold cross validation when K=5.
The strength of the -Fold technique is that each data point
in the original dataset is considered to be testing data once, and as training
data -1 times, leading to a decrease in
bias. Because of this, it was implemented in this study in the experiments
conducted with -Fold values equal to 3, 5, and 10.
3.8
Evaluation Metrics
In order to evaluate the performance of the
classification model, different evaluation metrics are used. Basically, these
metrics rely on the confusion matrix, which is a table used to evaluate the
performance of a classification algorithm. It is a summary of prediction
results on a binary classification problem. The confusion matrix comprises four
parts including True Positive (TP), True Negative (TN), False Positive (FP),
and False Negative (FN). Thus, this matrix shows the relationship between the
actual class and the predicted one (Klok & Nazarathy, 2021). Furthermore, when the classifier predicts the class that matches the
actual one, it means it is a TP and TN case. Moreover, as soon as the predicted
value does not match the actual value, then it is the case of either FP or FN.
Positive classes refer to the class with the values 1, True, or Yes; while the
Negative classes represent the values 0, False, or No, and so on. The most
common measures obtained from confusion matrix are: Accuracy, Precision,
Recall, and F1-score.
The confusion matrix is demonstrated in Table 1, and the
evaluation metrics are defined as in the following equations:
Another evaluation metric was employed called the Area
Under the Curve (AUC) which represents the area that falls under Receiver
Operating Characteristics (ROC). AUC is a widely used performance metric in ML,
which measures the ability of a binary classification model to distinguish
between positive and negative classes by computing the area under the ROC
curve. It is a robust and intuitive metric that provides a single scalar value
summarizing the overall performance of the model, which is insensitive to the
imbalance of class distributions.
Table 1: Confusion matrix for binary classification problem.
|
|
Actual Class
|
|
Positive (1)
|
Negative (0)
|
|
Predicted Class
|
Positive (1)
|
TP
|
FP
|
|
Negative (0)
|
FN
|
TN
|
4.
Proposed Smart Home System
This section discusses the structure of the proposed
smart home system, including hardware components (appliances, sensor, and
devices), and system workflow (with and without ML).
In order to build the proposed system, various devices
and sensors were utilized and connected together to collect and exchange data
with each other. Figure 5 demonstrates the home architecture of the proposed
system and its connected sensors, and appliances. Figure 4 provides a top view
of the proposed system prototype. The objectives of the utilized devices and
sensors in the proposed system are illustrated clearly in Table 2. In addition,
Table 3 shows each sensor and device and their signal type, provided voltage
(VCC: Voltage Common Collector), pin mode, and allocated RPi pin number.
Figure 4: Top view of the prototype.
Table 2. The objective and the position of
utilized devices and sensors.
|
Device
|
Count
|
Place
|
Objective
|
|
RPi
|
1
|
Controlling box
|
To connect all devices, send commands, and receive and store data. As well as
implementing ML model. Considered local server.
|
|
PIR
|
2
|
Bathroom
|
Sensing human existence to turn on the bathroom light
|
|
Living room
|
Sensing human existence to notify the householders about intruders
when “nobody home” mode is on. As well as to turn off the tv after a period
of no movement found.
|
|
Light Dependent Resistor (LDR)
|
1
|
Outside in front of the main door
|
Installed outside to calculate the real-time light intensity.
Useful especially for ML mode
|
|
Digital Temperature and Humidity (DHT22)
|
2
|
Bedroom
|
Sensing the temperature and humidity inside the room
|
|
outside
|
Sensing the temperature and humidity outside
|
|
Gas sensor (MQ-2)
|
1
|
Kitchen
|
To measure the intensity of smoke and gas.
|
|
Raindrops
|
1
|
Outside
|
Detecting if it is rainy or not.
|
|
Flame
|
1
|
Kitchen
|
Detecting if there is a flame inside the kitchen
|
|
Analog-to-Digital Converter System (ADS 1115/1015)
|
1
|
Controlling box
|
Converts analog signals received from sensors into digital
signals.
|
|
Stepper motor
|
1
|
Living room
|
To open and close the living room curtain in both modes: user
commands and intelligent (ML) mode
|
|
Servo motor
|
2
|
Living room
|
Controlling window
|
|
Main door
|
Controlling door
|
Figure 5: Home architecture and connected sensors, and appliances (red lines
indicates wired RPi output, blue lines indicate wired RPi input, and the green
lines refers to wireless communication).
Table 3. Home overall utilized sensors, actuators,
and appliances associated with allocated pin number and provided voltage.
|
Sensers / Appliances
|
Place
|
Allocated pin number
|
Pin mode
|
Signal type
|
VCC
|
|
PIR
|
Bathroom
|
6
|
Input
|
Digital
|
5v
|
|
Living room
|
19
|
Input
|
Digital
|
5v
|
|
DHT22
|
Living room
|
26
|
Input
|
Analog
|
3.3v
|
|
Outside
|
23
|
Input
|
Analog
|
5v
|
|
ADS 1115/1015
|
Between analog sensors and RPi GPIO
|
I2C protocol (SCL, SDA)
|
Input
|
Analog to Digital converter (ADC)
A0, A1, A2, A3
|
3.3v
|
|
LDR
|
Outside
|
A0
|
Input
|
Analog
|
5v
|
|
MQ-2
|
Kitchen
|
A1
|
Input
|
Analog
|
5v
|
|
Raindrops HL-83
|
Outside
|
A2
|
Input
|
Analog
|
3.3v
|
|
Flame
|
Kitchen
|
24
|
Input
|
Digital
|
5v
|
|
Stepper
|
Living room curtain
|
10, 16, 20, 21
|
Output
|
Digital
|
5v
|
|
Servo
|
Bedroom window
|
13
|
Output
|
Analog
|
5v
|
|
Main-door (outdoor)
|
12
|
Output
|
Analog
|
5v
|
|
Exhausted fan
|
Kitchen
|
25
|
Output
|
Digital
|
12v external adapter
|
|
LED bulb
|
Kitchen
|
7
|
Output
|
Digital
|
12v external adapter
|
|
Bedroom
|
4
|
Output
|
Digital
|
12v external adapter
|
|
Bathroom
|
17
|
Output
|
Digital
|
12v external adapter
|
|
Outside
|
8
|
Output
|
Digital
|
12v external adapter
|
|
Living room
|
18
|
Output
|
Digital
|
12v external adapter
|
|
TV
|
Living room
|
27
|
Output
|
Digital
|
12v external adapter
|
|
Air conditioner (AC)
|
Living room
|
22
|
Output
|
Digital
|
12v external adapter
|
|
Bedroom
|
5
|
Output
|
Digital
|
12v external adapter
|
The proposed work intends to create a dataset, referred
to as "Curtain dataset" that captures user behavior regarding the
opening and closing of curtain inside the house. Although the dataset is
intended to store real-time data received from the proposed smart home system,
the first version of the Curtain dataset was generated using logical simulation
data representing as much as possible real data. The dataset represents a
classification-based problem that records the data based on specific times and
includes six input features and a binary classification output (user behavior).
It comprises 8030 samples collected over one year, with some features were
encoded, such as in "time", "weekday", and
"out_light", while other features remain in raw data format.
Table 4 provides a detailed
description of the Curtain dataset, including the encoding methodology used for
each feature. Indeed, before generating the semi-real simulation data,
experiments were conducted on the sensors utilized to identify the range and
factors that influence their values. For example, the “out_light” feature,
which represents the value of the LDR sensor, is influenced by the date (month)
and time, particularly sunrise and sunset times (noted in (Sunrise and Sunset Zakho Dahuk Iraq, n.d.)) for each month.
Figure 7 demonstrates the correlation between the
dataset features and the output label, while Figure 6 displays the frequency
distribution of the available classes. Some features in the dataset were
generated randomly, whereas this randomness was controlled based on the
established range and experiments performed. The Curtain dataset is updated daily
every 30 minutes with actual real-time sensors data to provide accurate
prediction for the model. Table 5 provides a sample of the Curtain dataset,
consisting of a few observations.
Table 4: Curtain dataset description.
|
Input/output
|
Column name
|
Data type
|
Description and range
|
|
Input feature
|
date
|
Integer
|
Represent the month from 1 to 12
|
|
time
|
Integer
|
Represents the time of 24 hours, represented as follows:
20-4:29=0, 4:30-4:59=1, 5:00-5:29=2, 5:30-5:59=3, 6:00-6:29=4,
6:30-6:59=5, 7:00-7:29=6, 7:30-7:59=7, 8:00-8:29=8, 8:30-8:59=9,
9:00-9:29=10, 9:30-9:59=11, 10:00-10:29=12, 10:30-10:59=13, 11:00-16:29=14,
16:30-16:59=15, 17:00-17:29=16, 17:30-17:59=17, 18:00-18:29=18, 18:30-18:59=19,
19:00-19:29=20, 19:30-19:59=21.
|
|
weekday
|
Binary
|
Represents either 0 (weekend: Friday and Saturday) or 1 (working
days: Sunday to Thursday)
|
|
motion
|
Binary
|
Generated randomly. Represents the current value of the motion
(PIR) sensor
0 = no motion, 1 = motion detected.
|
|
out_light
|
Integer
|
Generated randomly taking into account the time and date feature.
Represents the outside light intensity which is calculated based on the value
that passes from the resistance (LDR sensor) to the RPi GPIO pin. The lower
the value received from RPi, the higher the light intensity. The values are
converted to percentages to make them easy to understand.
|
|
smoke
|
Integer
|
Generated randomly based on the tests conducted. Represents the
data of the gas/smoke sensor. The normal rang when no gas or smoke detected
is from 350 to 390. The cases of greater than 400 means smoke or gas is
detected.
|
|
Classification output
|
user action
|
Binary
|
Represents the user behavior.
0 = closing the curtain, 1= opening the curtain.
|
Table 5: Curtain Dataset Sample.
|
date
|
time
|
weekday
|
motion
|
out_
light
|
smoke
|
user
action
|
|
7
|
6
|
0
|
0
|
82
|
362
|
0
|
|
7
|
7
|
0
|
0
|
83
|
382
|
0
|
|
7
|
8
|
0
|
0
|
84
|
352
|
0
|
|
7
|
9
|
0
|
1
|
84
|
359
|
0
|
|
7
|
10
|
0
|
0
|
85
|
373
|
0
|
|
7
|
11
|
0
|
0
|
86
|
355
|
1
|
|
7
|
12
|
0
|
0
|
87
|
354
|
1
|
|
7
|
13
|
0
|
1
|
88
|
367
|
1
|
|
7
|
14
|
0
|
0
|
89
|
386
|
1
|
|
7
|
15
|
0
|
1
|
88
|
361
|
1
|
|
7
|
16
|
0
|
0
|
84
|
364
|
1
|
|
7
|
17
|
0
|
1
|
79
|
366
|
1
|
|
7
|
18
|
0
|
1
|
65
|
358
|
1
|
|
7
|
19
|
0
|
0
|
60
|
387
|
1
|
|
7
|
20
|
0
|
0
|
40
|
387
|
0
|
|
7
|
21
|
0
|
1
|
0
|
380
|
0
|
|
8
|
0
|
1
|
1
|
0
|
373
|
0
|
|
8
|
1
|
1
|
0
|
0
|
374
|
0
|
|
8
|
2
|
1
|
0
|
38
|
389
|
0
|
|
8
|
3
|
1
|
1
|
70
|
355
|
0
|
|
8
|
4
|
1
|
1
|
78
|
351
|
0
|
|
8
|
5
|
1
|
1
|
82
|
364
|
0
|
|
8
|
6
|
1
|
1
|
82
|
385
|
1
|
|
8
|
7
|
1
|
0
|
83
|
387
|
1
|
|
8
|
8
|
1
|
0
|
84
|
372
|
1
|
Figure 6: Curtain dataset class distribution.
Figure 7: Curtain dataset column correlations.
4.2
System Workflow
The proposed system is designed to provide various
functionalities that enable the creation of a fully automated home system with
enhanced security and comfort, while also reducing power consumption and human
effort. For this, the system consists of three modes, namely manual, automatic,
and intelligent (ML) modes. In the manual mode, the user has full control over
the appliances and can operate them through user intervention. In constant, the
automatic mode allows the system to control the appliances based on predefined
conditions using sensor values. Finally, in the intelligent mode, the system
employs ML algorithms to perform actions based on the output of the trained
model.
By integrating these three modes, the system provides a
robust and flexible solution that can adapt to the user's preferences and
environment conditions. This approach allows for the creation of a more
efficient and effective home system that reduces energy consumption and
increases comfort and security. In the following subsections, each mode is
described in detail to provide a better understanding of the proposed system's
capabilities and functionalities.
4.3
Manual Mode:
In this mode, the user sends control commands to the
actuators using a web-based UI to control the following appliances:
· Opening and closing the living room
curtain using a stepper motor.
· Turning on/off lights in the bedroom,
living room, and kitchen via the relay module.
· Switching on/off the living room and
bedroom air conditioners (represented as a fan in the prototype) via relay
module.
· Switching on/off the TV using relay
module.
· Turning on/off the kitchen exhaust fan
using relay module.
· Controlling the main door and bedroom
window by utilizing servo motors.
4.4
Automatic Mode:
This mode provides automatic control based on predefined
conditions that rely on sensors’ values. The following actions demonstrate the
mode:
· Based on the outside darkness obtained
from LDR, turning on or off outside light.
· Based on human existence, turning on
and off the bathroom light. Thus, the light will turn on as soon as motion is
detected via the PIR sensor, and it will turn off after 2 minutes, starting
from the time when no motion is detected.
· Turning on the exhaust fan when
abnormal smoke/gas is detected. In addition to sending an email, and viewing
text and voice messages on the UI.
· For security reasons, the system
contains an option called "Nobody at home". The home residents can
turn on this option when they leave home to tell the system to send an alert
via email and also show a text and voice notification on the GUI when the PIR
(installed in front of the main door) senses any motion. Thus, detecting the
intruder.
· Automatically close windows when rain
is detected via raindrop sensor.
· Turning off the TV automatically after
10 minutes starting from the moment when no motion is detected.
· Sending email and viewing text and
voice alerts as soon as a fire is detected.
· The system automatically stores the
values of the mentioned sensors in the Curtain dataset every 30 minutes based
on the specific pattern.
· Show text and voice alert if the
outside temperature is very hot or very cold.
· Show text and voice alerts if the
main-door stays open for a while.
4.5
Intelligent Mode:
The proposed system is incorporated with an intelligent
mode which relies on using ML technique.
This mode can be activated by the user via
the UI, allowing the system to predict user preferences for automatically
opening or shutting the curtain. In this mode, the prediction is performed
using a DT classifier. The reason behind using this predictor is discussed in
the next section. The process of storing data into Curtain dataset was
presented in section 1.1. Incorporating this part into the proposed work resulted in reduced user
effort and energy consumption, as the user will no longer need to turn on the
light when the outside sunlight enters the home after opening the curtain. This
also demonstrates the potential of ML to make home environments smarter, more
comfortable, and more feasible.
Figure 8 depicts the workflow of the proposed system,
highlighting the use of ML in predicting user preferences and automating the
opening and closing of the curtain. The decision tree classifier was chosen for
its ability to handle both numerical and categorical data, making it suitable
for the data collected in the "Curtain" dataset.
Figure 8 illustrates the functioning of the intelligent
mode in detail. This mode is designed to leverage the RPi to collect and store
data from sensors it in the Curtain dataset, along with the corresponding date
(month, and weekday) and time after encoding some data and converting some
others into a more comprehensible format. Prior to fitting classification
models, the essential process of feature selection and the classification label
determination is carried out. Moreover, feature scaling techniques such as
MinMax Scaler (for the KNN classifier) and Standard Scaler (for the MLP classifier)
are applied, with the exception of the DT classifier, since its performance not
affected by feature scaling. The dataset is subsequently split into training
and testing sets utilizing the K-Fold cross-validation method. The training
sets are used for the fitting process to obtain the trained model, which is
then used to predict the output of the testing set comprising previously unseen
samples. Several models are trained to determine the optimal model and
classifier for the intelligent mode. The selected model is then utilized by the
RPi to periodically check the sensors' data, along with the date and time,
every 10 minutes to predict the final decision of opening or closing the curtain.
5.
Results and Discussion
This section represents the results of various
experiments carried out the “Curtain” dataset using K-Fold cross-validation
(with a specified random state parameter of 42) to predict user behavior
regarding the opening and closing the curtain. The experiments involved
applying three different classification algorithms: decision tree (DT),
K-Nearest Neighbor (KNN), and Multilayer Perceptron (MLP) artificial neural
network (ANN). The objective was to evaluate the performance of each algorithm
and determine the best model to be implemented in the proposed system.
Initially, the experiments were conducted independently
for each algorithm, with diverse results obtained to assess the accuracy of the
classifiers. The results of each experiment were compared to determine the
optimal parameter values that would enhance the accuracy of the classifier.
Thereafter, the highest accuracies for each algorithm were compared against
each other to select the best overall model for the proposed system.
All experiments were carried out using the same
computer, which was equipped with Windows 10 64-bit operating system, x64-based
processor, Intel(R) Core (TM) i7-4700MQ CPU @2.40GHz, and 8.00 GB of RAM. To
facilitate the implementation of the experiments, popular Python programming
language packages such as Scikit-Learn, Pandas, and NumPy were used.
Furthermore, Table 6 outlines the parameters and their respective values
employed for the DT, KNN, and ANN classifiers in each conducted experiment.
Table 6: Classification algorithms and their
utilized parameters.
|
Classifier
|
Parameter
|
Utilized values
|
Description
|
|
DT
|
criterion
|
‘gini’, and ‘entropy’
|
Measurement function to calculate the quality of the feature.
|
|
splitter
|
‘best’, and ‘random’
|
A strategy used to split the tree in each node.
|
|
max_depth
|
Integer number from 3 to 12
|
The depth of the tree. Represents the number of levels in the
tree.
|
|
KNN
|
n_neibors
|
Integer number from 1 to 10
|
The number of nearest neighbors helps in deciding which class the
new data point belongs to.
|
|
p
|
1, 2, and 4
|
Representing distance metric. It is the power value for the
Minkowski distance metric that determines the utilized metrics among
Manhattan, Euclidean, or Minkowski.
|
|
MLP
|
hidden_layer_sizes
|
Integer number from 6 to 10
|
Number of hidden layers (Only one hidden layer utilized in this
study) with the number of neurons in each layer.
|
|
activation
|
‘identity’, ‘logistic’, ‘tanh’, and ‘relu’
|
It is the activation function which calculate the importance of
the neurons to decide if the neuron is active or not.
|
|
solver
|
‘sgd,’ and ’adam’
|
Optimizer to use for weights update.
|
|
learning_rate
|
‘Constant’
|
This parameter is uset to determine if the initial learning rate
will be changed or be constant. in this study, ‘constant’ is used. used only
when solver=’sgd’.
|
|
learning_rate_init
|
0.1, 0.01, 0.001
|
This is the actual learning rate to use as the initial value of
the learning rate for weight updates. It controls the step-size in updating
the weights. Only used when solver=’adam’ or ‘sgd’.
|
|
max_iter
|
700
|
Representing epoch, the max number of times that weights will be
updated for each data instance.
|
Figure 8: Workflow of the proposed system incorporating with ML.
5.1
Experiment #1: DT Classifier
Performance
In the
context of the experiments conducted in this study, the DT classifier was
utilized to determine the optimal model with the best parameter values. Several
tests were performed on different models, and the accuracy scores obtained from
the DT models were demonstrated using K-Fold cross-validation with 3, 5, and 10
folds. The results were summarized in Figure 9. The best results of the three
K-Fold values were compared and presented in Table 7 and Figure 10. The
findings indicate that the highest accuracy was achieved when the best splitter
was performed on the tree. Moreover, the other parameters, including criterion
and max_depth, were set to entropy and 7, respectively. The results
demonstrated that the model with 10 folds provided the best accuracy score of
97.4% and F1-score of 97.16%, outperforming the other values of 3 and 5 folds.
Therefore, the accuracy was found to increase with the increase in the number
of folds used in the experiment. It should be noted that no normalization
techniques were applied to the dataset during the pre-processing phase. This
decision was made since DT is not a distance-based algorithm, and therefore,
normalization does not significantly impact its performance.
The experiments
performed in this study involved testing the DT classifier with different
values for several parameters, as illustrated in Figure 9. The findings
indicate that when the splitter parameter was set to "best", the
criterion "gini" produced higher accuracy scores than the criterion
"entropy" when the maximum depth of the tree was less than or equal
to six. However, when the maximum depth was greater than six, the criterion
"entropy" performed better. On the other hand, when the random
splitter was employed, the results of both criteria ("entropy" and
"gini") exhibited an irregular pattern and did not follow a clear
trend. These findings highlight the importance of selecting appropriate
parameter values to achieve the best performance of the DT classifier.
Table 7: DT results for the best models selected
based on different K-Fold value (3, 5, and 10).
|
K-Fold
|
Criterion
|
Max_depth
|
Splitter
|
Accuracy
|
F1-score
|
|
3
|
Entropy
|
7
|
Best
|
97.17
|
96.93
|
|
5
|
Entropy
|
7
|
Best
|
97.30
|
97.05
|
|
10
|
Entropy
|
7
|
Best
|
97.40
|
97.16
|
To validate
the performance of the best models obtained from each K-Fold value (3, 5, and
10), the ROC curves were plotted and AUCs were calculated, as shown in Figure
11. The AUC values presented in each figure correspond to the AUC result for
each independent test split (fold). For example, the graph illustrating the
results of the selected best model for K-Fold when assigned to 3 displays three
AUC values. Similarly, the graph with five AUC values represents the model for
K-Fold = 5, and so on. Furthermore, to ensure the reliability of obtained
results, the training accuracy against the testing accuracy for each fold were
compared, as shown in Figure 12. These analyses provide insights into the
generalization performance of the models and demonstrate how well they can
perform on new, unseen data. The results suggest that the models trained with
the best parameters achieve high accuracy scores and can effectively predict
user behavior regarding the opening and closing of curtains.
Figure 9: Results of DT classifier when: (a) K-Fold=3. (b) K-Fold=5. (c)
K-Fold=10.
|
|
|
|
(a)
|
(b)
|
Figure 10: DT performance with different K-Folds. (a) Accuracy score. (b)
F1-score.
Figure 11: ROC curve with AUC results for DT classifier. (a) K-Fold=3. (b)
K-Fold=5. (c) K-Fold=10.
Figure 12: Model performance for DT classifier (Train vs Test accuracy). (a)
K-Fold=3. (b) K-Fold=5. (c) K-Fold=10.
The trained models exhibit high reliability, as
evidenced by their ability to achieve a high true positive rate of over 98%, as
shown in Figure 11. Furthermore, Figure 12 indicates that the train and test
accuracies were consistently high and that their values were very similar for
each split, implying that there was no evidence of overfitting or underfitting
in the trained models. This indicates that the models can accurately generalize
to new, unseen data, which is essential for their practical implementation in
real-world scenarios. Overall, these findings demonstrate the efficacy of the
selected classifiers and parameter values in predicting user behavior regarding
the opening and closing of curtains.
5.2
Experiment #2: KNN Classifier
Performance
This experiment investigates the performance of the KNN
algorithm as a classifier to predict the user's action. Given that KNN is a
distance-based algorithm, feature scaling techniques, particularly
Normalization, can have a significant influence on its performance. To address
this issue, the “MinMax Scaler” normalization method was employed, which has
shown to yield better results compared to the “Standard Scaler” standardization
method. To ensure a fair comparison among the utilized classifiers, the K-Fold
cross-validation was used with the same values of 3, 5, and 10 folds, as
previously done in Experiment #1. For each of these three K-Fold values, this
experiment conducted tests with the KNN classifier and reported the parameter
values and results obtained for each model with and without normalization. The
results, as shown in Figure 13, indicate that the performance of the KNN
classifier significantly improves with the use of normalization. This
improvement is particularly evident in cases where the number of folds is
higher (i.e., 5 and 10 folds). Therefore, the use of the "MinMax
Scaler" normalization technique was recommended when applying KNN for user
action prediction.
Figure 13: KNN performance results when: (a) K-Fold=3. (b) K-Fold=5. (c)
K-Fold=10.
Figure 13 provides a visual representation of the
parameters used in these experiments and their impact on the results. these
findings suggest that a value of "p" equal to one, which corresponds
to the Manhattan distance metric, yields the highest accuracies in most cases,
irrespective of the number of neighbors used. Additionally, the obtained
results indicate that the accuracy of the KNN classifier improves significantly
when using normalized data compared to raw data. Moreover, the Euclidean
distance metric outperforms the Minkowski metric in the conducted experiments.
These findings highlight the importance of carefully selecting the appropriate
distance metric for achieving optimal performance when using KNN as a
classifier.
The experimental results presented in the three
subfigures of Figure 13 are summarized in Table 8 and visualized in Figure 14,
where the best model that achieved the highest accuracy among all tests is
selected. According to Table 8, the highest accuracy is achieved using a
10-fold test with four neighbors and the Manhattan distance metric, where the
value of "p" is set to one. On the other hand, the best model in
terms of the F1-score evaluation metric is obtained using five neighbors, the
Manhattan distance metric, and a five-fold test. To further validate the
selected best models, this experiment presents the ROC curves in Figure 15 and
plot the training accuracy against the testing accuracy in Figure 16, using the
same procedures as in Experiment #1. These analyses demonstrate that the
selected models provide high accuracy and robustness, thus confirming their
suitability for user action prediction using the KNN classifier.
Table 8: KNN results for the best models selected
based on different K-Fold values (3, 5, and 10).
|
K-Fold
|
n_neighbors
|
P
(Distance metric)
|
Accuracy
|
F1-score
|
|
3
|
4
|
1
|
97.02
|
96.76
|
|
5
|
5
|
1
|
97.20
|
96.97
|
|
10
|
4
|
1
|
97.22
|
96.96
|
|
|
|
|
(a)
|
(b)
|
Figure 14: KNN performance with different K-Folds. (a) Accuracy score (b)
F1-score.
Figure 15: ROC curve with AUC results for KNN classifier. (a) K-Fold=3. (b)
K-Fold=5. (c) K-Fold=10.
Figure 16: Model performance for KNN classifier (Train vs Test accuracy). (a)
K-Fold=3. (b) K-Fold=5. (c) K-Fold=10.
The ROC curves displayed in Figure 15 indicate that all
models achieved a high percentage of the true positive rate, which suggests
that they are highly reliable for user action prediction. The high AUC rates
observed in the ROC curves further confirm the robustness and accuracy of the
selected models. Furthermore, Figure 16 presents the plot of training accuracy
against testing accuracy for each split, which provides insights into the
performance of the selected models in terms of overfitting and underfitting. It
can be observed that the models were well-trained, as they do not exhibit
significant overfitting or underfitting, and the train and test accuracies in
each split are very close to each other. This suggests that the selected models
generalize well to unseen data, which is an essential property for any machine
learning model to be effective in real-world applications.
5.3
Experiment #3: MLP Classifier
Performance
Another experiment employing MLP classification
algorithm was conducted over “Curtain” dataset, which yielded obtaining high
accuracies. Similar to the previous experiments, K-Fold cross-validation was
used to ensure a fair comparison among the different models. Therefore,
different results were obtained from several models represented in Figure 17.
Unlike the KNN which employed MinMax Scaler as the feature scaler, the MLP
classifier makes use of Standard Scaler standardization. Thus, this scaler led
to better results and helped the model to converge faster. Indeed, it can be
observed that the MLP provided very low results when the feature scaling
technique was not applied.
The results indicate that the best-performing MLP model
achieved an accuracy score of 97.36% and an F1-score of 97.13% when K-Fold was
set to 10, and the following parameters were assigned: "logistic"
activation function, one hidden layer with six nodes, a learning rate of 0.1,
and "sgd" solver (optimizer). Additionally, the max_iter parameter,
which determines the number of epochs necessary to train the model, was set to
700. The model was observed to converge entirely at this value after testing
various max_iter values, ranging from 100 to 700 in increments of 100. The loss
function used was "cross_entropy," which is the only option available
in the MLP classifier in the Sklearn package.
These findings are summarized in Table 9, which
highlights the optimal model for each K-Fold value. Additionally, Figure 18
illustrates the performance of the best model configurations across all K-Fold
values.
Figure 17: MLP performance with different K-Folds. (a) K-Fold=3. (b) K-Fold=5.
(c) K-Fold=10.
Figure 17 (a) demonstrates that when the solver=”adam”
the best results obtained are when the value of the learning rate equals to
0.01, and “tanh” as an activation function. While for the solver “sgd”, the
highest results obtained with the learning rate 0.1 and the activation function
is assigned to “logistic”. From this figure, it can be observed that the solver
“adam” provides higher results than the “sgd”.
It can be seen in Figure 17 (b) that the best results
obtained were with activation function “logistic” as well as when the value of
the learning rate assigned to 0.1. Furthermore, the learning rate with the
value 0.001 performs the lowest results in the most cases as compared with the
other learning rate values (0.1, and 0.01).
As visualized in Figure 17 (c), the activation function
“logistic” performs the highest results among all other activation function.
Moreover, the learning rate 0.01 outperform the other learning rate when the
“adam” optimizer is used. While the learning rate of 0.1 outperforms the other
learning rate utilized in this experiment when associated with “sgd” solver. It
can be observed that from the recently three mentioned figures, the lowest
results gained when the “identity” activation function is utilized.
Table 9: MLP results for the best models selected
based on different K-Fold values (3, 5, and 10)
|
K-Fold
|
Activation
function
|
Hidden
layer sizes
|
Learning
rate
|
Solver
|
Accur-acy
|
F1-score
|
|
3
|
tanh
|
6
|
0.01
|
adam
|
97.35
|
97.11
|
|
5
|
tanh
|
9
|
0.1
|
sgd
|
97.33
|
97.10
|
|
10
|
logistic
|
6
|
0.1
|
sgd
|
97.36
|
97.13
|
|
|
|
|
(a)
|
(b)
|
Figure 18: Performance with different K-Folds. (a) Accuracy score. (b) F1-score.
Figure 18, illustrates that the highest results obtained
when the number of folds assigned to 10. In the next figures, the validations
of the best MLP models are demonstrated, including those with 3, 5, and 10 as
K-Fold values. In Figure 19, the ROC curves associated with AUC rates were
presented, while Figure 20 displays the testing against training accuracy results.
Figure 19: ROC curve with AUC results for MLP classifier. (a) K-Fold=3. (b)
K-Fold=5. (c) K-Fold=10.
Figure 20: Model performance for MLP classifier (Train vs Test accuracy). (a)
K-Fold=3. (b) K-Fold=5. (c) K-Fold=10.
The performance of the models was assessed by analyzing
true positive rates across different test splits, as shown in Figure 19. The
results indicate that the models were highly reliable, achieving true positive
rates of more than 99% for each test split. Furthermore, Figure 20 illustrates
that the models were appropriately trained, with neither overfitting nor
underfitting observed. The train and test accuracies were found to be extremely
close to one another across all splits, providing further evidence of the
models' robustness and generalizability. These findings suggest that the models
developed in this study can be considered highly reliable and adequately
trained for use in real-world scenarios.
Table 10: A comparison among the best results for
each classifier with their training time.
|
K-Fold
|
Classifier
|
Accuracy
|
F1-score
|
Average training time in milliseconds
|
|
10
|
DT
|
97.40
|
97.16
|
5.45
|
|
10
|
KNN
|
97.22
|
96.96
|
11.25
|
|
10
|
MLP
|
97.36
|
97.13
|
3157.6 (4.51 per epoch)
|
Upon analyzing the results of the three
experiments conducted on the "Curtain" dataset, it can be concluded
that the DT classifier demonstrated superior performance compared to the other
classifiers in terms of accuracy and training time (as illustrated in Table 10).
This can be attributed to the fact that the "Curtain" dataset is
composed solely of numerical values, which are well-suited for decision
tree-based algorithms.
To obtain the training times for the
classifiers' best models, the average time in milliseconds was calculated by
averaging the results of four different executions of the same model. As a
result, the average training times for the DT, KNN, and MLP classifiers were
found to be 5.45, 11.25, and 3157.6 milliseconds (4.51 per epoch),
respectively. Based on these findings, it is recommended that the DT model be
implemented in the system due to its superior performance in terms of accuracy
and training time.
5.4
Comparison of the proposed
system with the state-of-the-art systems
Table 11 presents a comparative analysis of
the proposed system against several reviewed systems in the literature.
Table
11: Summary of proposed system
versus reviewed works, where MLu: Machine Learning is utilized, RPi: Raspberry
Pi used, oRPi: only the RPi utilized as a central controller, DPS: Dataset
Provided by the system, Adp: the system is adaptive to the user behavior, SP:
security purpose in the system, AM: does any appliances controlled
automatically, MM: does any appliances controlled manually, MF: Muli-Featured,
WUI: Web-based User Interface.
|
Ref.
|
MLu
|
RPi
|
oRPi
|
DPS
|
Adp
|
SP
|
AM
|
MM
|
MF
|
WUI
|
|
(Okorie et al., 2020)
|
û
|
û
|
û
|
û
|
û
|
û
|
û
|
ü
|
ü
|
û
|
|
(Iqbal et al., 2018)
|
û
|
ü
|
û
|
û
|
û
|
ü
|
ü
|
ü
|
ü
|
ü
|
|
(Pavithra & Balakrishnan, 2015)
|
û
|
ü
|
ü
|
û
|
û
|
ü
|
ü
|
ü
|
ü
|
ü
|
|
(Gota et al., 2020)
|
û
|
û
|
û
|
û
|
û
|
û
|
ü
|
ü
|
ü
|
ü
|
|
(Abbas & Abdullah, 2021)
|
ü
|
ü
|
û
|
ü
|
ü
|
û
|
ü
|
ü
|
û
|
û
|
|
(Mehmood et al., 2019)
|
ü
|
ü
|
û
|
û
|
û
|
ü
|
ü
|
û
|
û
|
ü
|
|
(Crisnapati et al., 2016)
|
ü
|
ü
|
û
|
û
|
û
|
ü
|
ü
|
ü
|
ü
|
ü
|
|
(Paredes-Valverde et al., 2020)
|
ü
|
û
|
û
|
ü
|
û
|
û
|
û
|
ü
|
ü
|
û
|
|
(Peng et al., 2019)
|
ü
|
û
|
û
|
û
|
û
|
û
|
û
|
ü
|
û
|
û
|
|
(Raju et al., 2021)
|
ü
|
ü
|
ü
|
û
|
û
|
û
|
ü
|
ü
|
ü
|
û
|
|
Proposed system
|
ü
|
ü
|
ü
|
ü
|
ü
|
ü
|
ü
|
ü
|
ü
|
ü
|
The proposed system outperforms the reviewed systems in
various aspects. Firstly, the proposed system excels having a web-based user
interface (UI) that enables local and internet-based system access surpassing
the systems reviewed in (Abbas & Abdullah, 2021; Okorie et al., 2020;
Paredes-Valverde et al., 2020; Peng et al., 2019; Raju et al., 2021). Secondly, the proposed system stands out for its
ability to use only a RPi as a central controller, outperforming all reviewed
systems except for the ones in (Pavithra & Balakrishnan, 2015; Peng et al., 2019). Thirdly, this system has a unique ability to adapt to
user behavior by leveraging the dataset it creates, a feature missing in all
reviewed systems except for the one in (Abbas & Abdullah, 2021) which does not take into consideration the environmental conditions.
Additionally, the proposed system outperforms the reviewed system in terms of using
ML techniques to enhance home intelligence, surpassing the systems reviewed in (Gota et al., 2020; Iqbal et al., 2018; Okorie et al.,
2020; Pavithra & Balakrishnan, 2015). Moreover, this system provides a variety of ways for users to interact
with the system, including features related to security, energy-saving, and
more, outperforming all reviewed works. Furthermore, this system provides
real-time voice and text alerts to users via the UI and email to improve home
security and safety. Finally, the proposed system provides a fully automated
home, utilizing several sensors and appliances, and enhances home intelligence
by applying automatic decision-making on behalf of the user.
6.
Conclusion and Future Works
Smart homes using Machine Learning (ML) offer
significant benefits over traditional smart homes by incorporating the power of
artificial intelligence. By leveraging ML algorithms, smart homes can
automatically adjust settings and operations based on learned patterns and
preferences, providing a more personalized and efficient experience for
residents. One of the key applications of ML in smart homes is in the area of
home automation. For example, ML algorithms can be used to control lighting,
heating, cooling, and security systems based on the presence and behavior of
residents.
This study proposes a smart home system that aims to
enhance security, comfort, energy efficiency, and intelligence. The system
comprises three primary modes, including the manual mode, automatic mode, and
intelligent mode. The manual mode allows the user to control different
appliances and monitor the home environment through a web-based application.
The automatic mode, on the other hand, reduces human intervention and maximizes
security. This mode provides real-time alerts through the UI in both text and
voice form and also sends text alerts via email in the event of an intruder or
fire detection. Additionally, this mode controls appliances automatically based
on sensor readings to minimize energy consumption. The intelligent mode employs
ML algorithms to predict and track user behavior patterns related to opening
and closing curtains. Several experiments were conducted to compare the
performance of various classifiers (DT, KNN, and MLP) applied to the
"Curtain" dataset. The results revealed that DT provided the highest
accuracy (97.4%) with the shortest average training time (5.45 milliseconds)
among all the implemented classifiers, making it the best choice for
implementation in the proposed smart home system.
For the future goal it has been aimed to enable the user
to interact with the system using voice commands. Additionally, it has been
planned to incorporate image processing capabilities into the system to
automatically control the main door based on the image of the authorized
person. Furthermore, for the intelligent mode, it has been aimed to create
another dataset for user action for predicting user preferences using
regression-based models. In addition, applying various feature selection
techniques on the "Curtain" dataset in order to evaluate their impact
on the performance of the utilized models.
References
Abbas, A. F., & Abdullah, M. Z. (2021). Design and
Implementation of Tracking a user’s Behavior in a Smart Home. IOP Conference
Series: Materials Science and Engineering, 1094(1), 012008.
https://doi.org/10.1088/1757-899x/1094/1/012008
Bertolini, M., Mezzogori, D., Neroni, M., & Zammori,
F. (2021). Machine Learning for industrial applications: A comprehensive
literature review. Expert Systems with Applications, 175(February),
114820. https://doi.org/10.1016/j.eswa.2021.114820
Car, Z., Baressi Šegota, S., Anđelić, N.,
Lorencin, I., & Mrzljak, V. (2020). Modeling the Spread of COVID-19
Infection Using a Multilayer Perceptron. Computational and Mathematical
Methods in Medicine, 2020. https://doi.org/10.1155/2020/5714714
Choi, W., Kim, J., Lee, S. E., & Park, E. (2021).
Smart home and internet of things: A bibliometic study. Journal of Cleaner
Production, 301, 126908.
https://doi.org/10.1016/j.jclepro.2021.126908
Crisnapati, P. N., Wardana, I. N. K., & Aryanto, I.
K. A. A. (2016). Rudas: Energy and sensor devices management system in home
automation. Proceedings - 2016 IEEE Region 10 Symposium, TENSYMP 2016,
184–187. https://doi.org/10.1109/TENCONSpring.2016.7519401
Desai, M., & Shah, M. (2021). An anatomization on
breast cancer detection and diagnosis employing multi-layer perceptron neural
network (MLP) and Convolutional neural network (CNN). Clinical EHealth, 4,
1–11. https://doi.org/10.1016/j.ceh.2020.11.002
Gao, X., & Li, G. (2020). A KNN Model Based on
Manhattan Distance to Identify the SNARE Proteins. IEEE Access, 8,
112922–112931. https://doi.org/10.1109/ACCESS.2020.3003086
Géron, A. (2019). Hands-on Machine Learning with
Scikit-Learn, Keras, and TensorFlow. In O’Reilly Media, Inc (Second
Edi). O’Reilly Media, Inc.
Gota, D. I., Puscasiu, A., Fanca, A., Miclea, L., &
Valean, H. (2020). Smart home automation system using Arduino microcontrollers.
2020 22nd IEEE International Conference on Automation, Quality and Testing,
Robotics - THETA, AQTR 2020 - Proceedings.
https://doi.org/10.1109/AQTR49680.2020.9129989
Heidari, A. A., Faris, H., Mirjalili, S., Aljarah, I.,
& Mafarja, M. (2020). Ant lion optimizer: Theory, literature review, and
application in multi-layer perceptron neural networks. Studies in
Computational Intelligence, 811, 23–46. https://doi.org/10.1007/978-3-030-12127-3_3
Ibrahim, A. K., Hassan, M. M., & Ali, I. A. (2022).
Smart Homes for Disabled People: A Review Study. Science Journal of
University of Zakho, 10(4), 213–221.
https://doi.org/10.25271/sjuoz.2022.10.4.1038
Iqbal, A., Ullah, F., Anwar, H., Kwak, K. S., Imran, M.,
Jamal, W., & Rahman, A. ur. (2018). Interoperable Internet-of-Things
platform for smart home system using Web-of-Objects and cloud. Sustainable
Cities and Society, 38, 636–646.
https://doi.org/10.1016/j.scs.2018.01.044
Jabbar, W. A., Alsibai, M. H., Amran, N. S. S., &
Mahayadin, S. K. (2018). Design and Implementation of IoT-Based Automation
System for Smart Home. 2018 International Symposium on Networks, Computers
and Communications, ISNCC 2018, November 2018, 1–6. https://doi.org/10.1109/ISNCC.2018.8531006
Jabbar, W. A., Kian, T. K., Ramli, R. M., Zubir, S. N.,
Zamrizaman, N. S. M., Balfaqih, M., Shepelev, V., & Alharbi, S. (2019).
Design and Fabrication of Smart Home with Internet of Things Enabled Automation
System. IEEE Access, 7, 144059–144074.
https://doi.org/10.1109/ACCESS.2019.2942846
Jolles, J. W. (2021). Broad-scale applications of the
Raspberry Pi: A review and guide for biologists. Methods in Ecology and
Evolution, 12(9), 1562–1579. https://doi.org/10.1111/2041-210X.13652
Jung, A. (2022). Machine Learning: Foundations,
Methodologies, and Applications. Springer Nature.
Klok, H., & Nazarathy, Y. (2021). Statistics with
Julia: Fundamentals for Data Science, Machine Learning and Artificial
Intelligence. Springer Nature. https://statisticswithjulia.org
Kubat, M., & An. (2021). An Introduction to Machine
Learning. In Springer Nature (Third Edit). Springer Nature.
https://doi.org/10.1002/9781119720492.ch7
Kurniawan, A. (2019). Introduction to Raspberry Pi. In Raspbian
OS Programming with the Raspberry Pi (pp. 1–25). Apress, Berkeley, CA.
https://doi.org/https://doi.org/10.1007/978-1-4842-4212-4_1
Li, J., Gao, F., Lin, S., Guo, M., Li, Y., Liu, H., Qin,
S., & Wen, Q. (2023). Quantum K-fold cross-validation for nearest neighbor
classification algorithm. Physica A: Statistical Mechanics and Its
Applications, 611, 128435.
https://doi.org/10.1016/j.physa.2022.128435
Mahmoodzadeh, A., Mohammadi, M., Daraei, A., Farid Hama
Ali, H., Kameran Al-Salihi, N., & Mohammed Dler Omer, R. (2020).
Forecasting maximum surface settlement caused by urban tunneling. Automation
in Construction, 120(July), 103375.
https://doi.org/10.1016/j.autcon.2020.103375
Marikyan, D., Papagiannidis, S., & Alamanos, E.
(2019). A systematic review of the smart home literature: A user perspective. Technological
Forecasting and Social Change, 138(November 2017), 139–154.
https://doi.org/10.1016/j.techfore.2018.08.015
Mehmood, F., Ullah, I., Ahmad, S., & Kim, D. H.
(2019). Object detection mechanism based on deep learning algorithm using
embedded IoT devices for smart home appliances control in CoT. Journal of
Ambient Intelligence and Humanized Computing, 0123456789.
https://doi.org/10.1007/s12652-019-01272-8
Mienye, I. D., Sun, Y., & Wang, Z. (2019).
Prediction performance of improved decision tree-based algorithms: A review. Procedia
Manufacturing, 35, 698–703.
https://doi.org/10.1016/j.promfg.2019.06.011
Mukherjee, A., Mondal, S., Chaki, N., & Khatua, S.
(2019). Naive bayes and decision tree classifier for streaming data using
hbase. In Advances in Intelligent Systems and Computing (Vol. 897).
Springer Singapore. https://doi.org/10.1007/978-981-13-3250-0_8
Nikou, S. (2019). Factors driving the adoption of smart
home technology: An empirical assessment. Telematics and Informatics, 45(September),
101283. https://doi.org/10.1016/j.tele.2019.101283
Okorie, P. U., Ibraim, A. A., & Auwal, D. (2020).
Design and Implementation of an Arduino Based Smart Home. HORA 2020 - 2nd
International Congress on Human-Computer Interaction, Optimization and Robotic
Applications, Proceedings, October 2012.
https://doi.org/10.1109/HORA49412.2020.9152922
Pajankar, A. (2021). Introduction to Raspberry Pi. In Practical
Linux with Raspberry Pi OS. Apress, Berkeley, CA.
https://doi.org/https://doi.org/10.1007/978-1-4842-6510-9_1
Paredes-Valverde, M. A., Alor-Hernández, G.,
García-Alcaráz, J. L., Salas-Zárate, M. del P., Colombo-Mendoza, L. O., &
Sánchez-Cervantes, J. L. (2020). IntelliHome: An internet of things-based
system for electrical energy saving in smart home environment. Computational
Intelligence, 36(1), 203–224. https://doi.org/10.1111/coin.12252
Patel, H. H., & Prajapati, P. (2018). Study and
Analysis of Decision Tree Based Classification Algorithms. International
Journal of Computer Sciences and Engineering.
Pavithra, D., & Balakrishnan, R. (2015). IoT based
monitoring and control system for home automation. 2015 Global Conference on
Communication Technologies (GCCT).
https://doi.org/10.1109/GCCT.2015.7342646
Peng, Y., Peng, J., Li, J., & Yu, L. (2019). Smart
Home System Based on Deep Learning Algorithm. Journal of Physics: Conference
Series, 1187(3). https://doi.org/10.1088/1742-6596/1187/3/032086
Raju, L., Sowmya, G., Srividhya, S., Surabhi, S.,
Retika, M. K., & Reshmika Janani, M. (2021). Advanced home automation using
raspberry pi and machine learning. Proceedings of the 7th International
Conference on Electrical Energy Systems, ICEES 2021, 600–605.
https://doi.org/10.1109/ICEES51510.2021.9383738
Raspberry Pi Documentation. (n.d.). Retrieved December 14, 2022, from
https://www.raspberrypi.com/documentation/computers/raspberry-pi.html#gpio-and-the-40-pin-header
Saleem, A. A., Hassan, M. M., & Ali, I. A. (2022).
Smart Homes Powered by Machine Learning: A Review. Proceedings of the 2nd
2022 International Conference on Computer Science and Software Engineering,
CSASE 2022, 355–361. https://doi.org/10.1109/CSASE51777.2022.9759682
Sarker, I. H. (2021a). Machine Learning: Algorithms,
Real-World Applications and Research Directions. SN Computer Science, 2(3).
https://doi.org/10.1007/s42979-021-00592-x
Sarker, I. H. (2021b). Machine Learning: Algorithms,
Real-World Applications and Research Directions. SN Computer Science, 2(3),
1–21. https://doi.org/10.1007/s42979-021-00592-x
Singh, P. K., Kar, A. K., Singh, Y., Kolekar, M. H.,
& Tanwar, S. (2020). Proceedings of ICRIC 2019: Recent Innovations in
Computing. In P. K. Singh, A. K. Kar, Y. Singh, M. H. Kolekar, & S. Tanwar
(Eds.), Lecture Notes in Electrical Engineering (First Edit). Springer
Nature. https://doi.org/https://doi.org/10.1007/978-3-030-29407-6
Sunrise and sunset Zakho Dahuk Iraq. (n.d.). Retrieved January 7, 2022, from
https://www.weatheravenue.com/en/asia/iq/dahuk/zakho-sunrise.html
Taiwo, O., Ezugwu, A. E., Oyelade, O. N., &
Almutairi, M. S. (2022). Enhanced Intelligent Smart Home Control and Security
System Based on Deep Learning Model. Wireless Communications and Mobile
Computing, 2022. https://doi.org/10.1155/2022/9307961
Tatsat, H., Puri, S., & Lookabaugh, B. (2020).
Machine Learning and Data Science Blueprints for Finance. In O’Reilly Media
(First Edit). O’Reilly Media, Inc.
van Engelen, J. E., & Hoos, H. H. (2020). A survey
on semi-supervised learning. Machine Learning, 109(2), 373–440.
https://doi.org/10.1007/s10994-019-05855-6
