You are on page 1of 6

> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 1

A survey on artificial intelligence for intelligent


control and management system
Hor Kunthroza

home and smart city has developed in the trend. The appearance
Abstract—Since entering the 21st century, the rapid of smart home and smart city brings individuals more
development of artificial intelligence and internet of things have convenience and in daily life, and high-quality of life.
promoted and supported sustainable development of future smart
application and comfort residential living environment.
II. BACKGROUND OVERVIEW
development of artificial intelligence there are more and more
researches on it, such as smart home, smart city, internet of A. Artificial Intelligence Overview
healthcare thing and so on for providing better living standards,
quality of life and making people’s life more convenient by
AI encompass numerous techniques to address issues related
incorporating with technology. This survey highlights some to optimization, automation learning, recognition and
description, control and management and the reason for prediction. Artificial neural networks (ANN) and machine
emergence smart home, smart city, etc. that will allow user control learning models are two leading AI techniques that are
appliance and service whenever needed. It focuses not only on AI, extensively used in modern-day applications.
but also on the application of Internet of Things (IoT), machine I. Artificial Neural Network
learning because of the ability to recognize and classify patterns in
large data sets.
Artificial neural network (ANN) is a computational network
inspired by biological nervous systems, i.e., brain. Generally,
Keywords—Artificial Intelligence, machine learning, IoT, smart ANN comprises three layers: input (independent variables),
home, smart city. hidden, and output (dependent variables).
Firstly, the input layer receives information to be processed
I. INTRODUCTION by the hidden layer, and the output layer presents the final
In the last twenty years, developments within the ongoing output [3]. The hidden layer consists of a computational
socioeconomic digitalization have created the vision of smart processing unit known as artificial neurons that mimics the
society such as cities, which aspires to connect all aspects of biological neurons of a nervous system. This process of setting
urban life. The basis for connecting aspects of urban life in up an ANN with generalized solutions for a given input and
smart cities is being built and emerging technologies, such as given output is known as a learning or training process, and this
cloud computing, the Internet of Things and cyber-physical is completed with a given learning algorithm. An ANN can be
systems, representing the latest chain in industrial revolution, trained through supervised, unsupervised, reinforcement,
referred to as Industry 4.0. A key aspect of the technologies is offline/batch, and online learning mechanisms [3].
adopting and advancing artificial intelligence (AI) techniques, The selection of learning mechanisms depends on the data
which have proven their ability to process large amounts of data availability and nature of the problem. Further, based on the
towards developing learning rules, e.g. via machine learning interconnection and composition of the neurons in the hidden
(ML), making complex associations, and predicting outcomes layer, artificial neural networks are categorized into single-
of complex physical processes. layer feedforward networks, multilayer feedforward networks,
Smart homes are a combination of information technology, recurrent/feedback networks, and mesh networks [3]. Below
Internet of Things (IoT), and communication technologies. By fig. presents a neural network having two hidden layers.
virtue of smart home systems, the home security, lighting and
power consumption can be monitored and controlled remotely
using web interface or mobile devices. Entertainment devices
such as televisions can be connected to remote mundane
channels. Shortly, A smart home is not referring to how well a
home is built or how effectively space is utilized or how
environment friendly it is. Indeed, a smart home encompasses
all these attributes but it is the use of different interactive and
intelligent technologies that make it smart [1][2].

With the fast development of technologies, such as artificial


intelligence, internet of thing, network technology, internet of
things and artificial intelligence, artificial intelligence has
become a new trend of social development. Therefore, the smart
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 2

II. Machine Learning B. Internet of Things Overview


According to [9], machine learning lies at the junction of There are three main major technologies that are sensor
statistics and computer science, and the core of AI and data technology, radio frequency identification technology and
science. Machine learning algorithms are broadly classified into embedded system technology for application in the Internet of
supervised, unsupervised, semi-supervised, and reinforcement Things.
learning, as discussed in the following subsections. Sensor refers to components or devices that it can acquire the
1) Supervised Machine Learning object to be detected information, and convert the information
Supervised machine learning is the most popular and obtained into the required form of information by a certain
trustworthy method due to its working principle. Supervised regularity or method.
machine learning works on a set of labeled data which is a set Radiofrequency identification technology combines radio
of data containing detailed input features and their frequency technology with embedded technology, which can
corresponding output, i.e., labels. The set of labeled data is identify specific targets by radio signals and read and write
divided into training sets and testing sets and is employed by specific target-related data.
the given learning model for training and testing purposes. Embedded system technology is a combination of computer
hardware and software, electronic application technology,
2) Unsupervised Machine Learning
integrated circuit technology, sensor technology as one
The unsupervised machine learning method is used to
technology. Embedded system consists of CPU, memory, I / O
discover hidden patterns of data within a given dataset without
and software, which meets the various functions of the
any reference to known or labeled outcomes. The purpose of application system, small size, and has high reliability and high
unsupervised machine learning is to detect distinct patterns by control processing capabilities.
exploring the structure of information and using the extracted These majors are tools for control processing of smart home
information towards improving the decision-making process. devices. It’s closely linked to many fields and has been covered
Clustering techniques and Rule-Based learning are two widely in all aspects of our life. As shown in above picture, the key
used unsupervised machine learning algorithms. areas of the Internet of things mainly involves smart home,
3) Semi-supervised Machine Learning smart schools, smart healthcare, smart transportation, smart
This method is lying at the intersection of supervised and industry and smart agriculture, among which smart home is
unsupervised machine learning since it uses both labeled and rapidly developing and is quietly entering people's lives.
unlabeled data for training purposes. Generally, a small portion
of labeled data is combined with a large portion of unlabeled C. Smart Home Overview
data, to improve learning accuracy.
4) Reinforcement Learning Smart home is based on residential platform, which
connects all the devices (lighting system, air-conditioning
Reinforcement machine learning is a method that interacts
system, security system, video system, network appliances,
with a dynamic environment where it performs certain actions
etc.) in the home through the Internet of Things technology to
towards the goal while navigating its problem space.
provide people with lighting control, security detection, remote
Reinforcement machine learning allows machine and software terminal Control, indoor and outdoor remote control and other
programs, commonly referred to as agents, to autonomously convenient features. Smart home has some features such as
determine ideal actions to achieve a goal through a trial and equipment automation, information appliances, remote
error process. Trial and error search and rewards are the key terminal control, network communications, network sharing
characteristics of reinforcement machine learning in which the and et cetera, so that it can make users more convenient and
agent is rewarded for actions that increase efficiency and are more comfortable in their daily life. As shown in Figure, the
penalized for actions that reduce it. Q-learning, Temporal smart home system consists of multiple systems for people to
Difference, and Deep Adversarial Networks are widely used provide all kinds of convenience and comfort. The service
reinforcement learning algorithms. direction of smart home is to offer people a comfortable, safe
Fig. 3 presents the details of inputs/outputs and the working and convenient home environment. Different from ordinary
principle of different machine learning algorithms. home nowadays, in addition to having the basic functions
needed in ordinary home life, smart home also has many
intelligent functions. These functions can’t be realized without
the technologies in the Internet (automation technology,
computer technology, remote terminal control technology,
Internet of Things technology, network technology, etc.).
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 3

D. Smart City Overview


The idea behind smart city is profoundly based on
optimization of costs, better living standards, preservation of
resources, integration of technology and faster transactions in
all fields. It incorporates all aspects of technology in
transforming a complicated infrastructure into a digital,
sophisticated and a simpler lifestyle. Here, the term
‘technology’ refers to the most popular and emerging fields like
AI and IoT.
Smart city is the next big thing and the government has
taken various initiatives to promote the need for a technical
infrastructure in a city. This is a platform for the states to
nominate some of their cities for transforming them into smart
cities.

B. Raspberry Pi
III. PROPOSE METHOD

A. AI-Driven Smart Home


In [5], authors proposed an AI-based smart home by
implementing framework, which consists of NILM, DSM,
Controller significantly contribute to numerous energy
efficiency applications. Implementing NILM for prediction
modeling, proposed a NILM approach to disaggregate air
conditioning or power consumption, NILM assisted platform is
presented for real-time electromechanical system diagnostics is
mostly performed through regression models or neural
networks, presented a novel NILM assisted approach to monitor
the health of elderly people. Moreover, optimization techniques
are widely adopted in the context of home energy management
systems such as Water dispenser, fridge, microwave, air
purifier, and electric kettle. It is also noted that most of the
energy efficiency applications are proposed in a standalone
approach.
As seen in Fig. 5b that NILM, DSM, and forecasting are
interlinked. In this part authors proposed the NILM outcome In this article authors proposed this model to develop the
can facilitate the DSM in numerous ways, e.g., identifying the system is Raspberry Pi Model B 512Mb Linux system to run
potential appliances for load shifting strategies using the the Brain Class and the various modules because of the
appliance-level information, i.e., corresponding operation relatively low cost of the Raspberry Pi, this has become a
status and timing. Moreover, the extracted appliances’ popular and economical solution to the more expensive
consumption pattern can also facilitate DSM towards more commercial alternatives. It listens to the user commands, acts a
robust and effective energy scheduling. local central server and communicates with the remote cloud
server. It notifies user through the notifier module using a TTS
Furthermore, the proposed AI-driven smart home is also
engine and speakers. Kinobo Usb 2.0 Microphone and self-
connected to utilities, renewable energy sources, battery
powered speakers are used as input and output devices to the
storage, and different electrical load[4][5]. raspberry pi.
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 4

patterns which are being repeated must be discovered and


1. Cloud Server classified. Secondly, the action mapped to the pattern is decided
The user history and activity data is stored on a remote upon. This is achieved by using a pattern recognition algorithm
server. This data is trained using machine learning algorithms along with a decision algorithm.The hidden layer nodes, as
to provide a model for recommendation and activity recognition shown in Fig. 2 handle the computations and are used to find
system because it allows large data storage and high processing behavioral patterns.
capabilities required for training a highly accurate data model. These units are composed of networks that are able to learn
a classification problem once trained with sufficient data. he
2. Brain Class ANN will learn the pattern mapped to a specific attribute value
Brain class is the central unit of the system and placed for to eventually classify new and hidden behaviors.
configuration, interaction, handling request queues, and
managing overall system performance, also deals with
communicating on the Web and request queues. This query is
interpreted using NLP and forwarded to the respective device
called.
3. Notifier class
This system processes the query, obtains the result and
outputs it using e-Speak Text To Speech engine and notifies
service which are offered include getting news updates, traffic,
emails, calendar notifications, etc. through the use of user given
data by using a list of APIs which output data in a format which
is easy to integrate.
4. APIs
API is a simple interface for users to configure their own
devices and modules. The interface requires an isValid()
method which returns true if input is valid. Once input is 8. Security Module
validated, appropriate event Handler code must be written to do A camera is situated at the door which keeps recording frame
the action required. Finally, a context will be used for user by frame and stores them on the cloud storage. Cascade
interaction. Classifier of OpenCV library is used for face recognition. The
5. Context Manager process flow is shown in Fig. 3. Whenever a person arrives at
A context is used to manage different users. This context the door, the camera module detects the face. This image is then
differentiates the user through the use of speech recognition and preprocessed and undergoes feature extraction. The OpenCV
various other factors specific to a user. User specific speech face detector is pre-trained using a very comprehensive
data is stored using Python API. database of face/non-face examples and is used for face
6. Natural Language Processing identification.
NLP is used for finding proper intent from a user command Whenever a person arrives at the door, the camera module
so it can be forwarded to the appropriate device and it is detects the face. This image is then preprocessed and undergoes
implemented in two parts, intent identification and entity feature extraction. The OpenCV face detector is pre-trained
recognition. using a very comprehensive database of face/non-face
For intent identification supervised learning is used. After a examples and is used for face identification.
command is received by brain class it's sent to the remote
server. At the server a pre-trained model is stored which has
command mapped with their intended actions. This model is
used as a data set for interpreting the commands.
Intent identification is done by testing the user command
against the pre-trained model using MaxEnt classification. The
entity extraction is done by using NER(Named Entity
Recognition). Named Entity recognition is used to find entities
or objects from the given text and also to classify these entities
into predefined subcategories. This data is stored in user
preferences and used by learning algorithms to predict most
accurate action of the user.
7. Machine Learning and ANN
In this system uses machine learning to provide 9. Activity Recognition
recommendations to the user according to user preferences. Sensor nodes are placed intelligently in order to record user
Data is stored categorically with respective intents provided by activities which are mainly physical as well as interactive with
NLP, and given for training. Content Based Recommendation objects. The sensor and camera data (is mined and used for
(CBR) algorithm is chosen as it provides personalized training the model) is sent to server where it is preprocessed,
recommendations. To accurately recognise and predict the converted into a sequence of events, which is further used by a
user’s activity, the process is divided into two parts. Firstly, the FP-growth mining algorithm [6] to extract the frequent patterns.
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 5

The Camera Module was tested under varying light and hue the features through a forward neural network. Finally, it
conditions. Initial experiments were directed towards obtaining predicts the command, as illustrated in “Fig. 1”.
an optimal size of the face region samples used for recognition. The first phase of the model involves extracting the features
The sizes of the faces can vary significantly. Fig. 5. shows the of both voice and image datasets for analyzing and finding
recognition accuracy corresponding to different image sizes. relations between different aspects into the same vector
Fig. 6. is a graphical representation of the same. dimensions. ex,
For the voice dataset in our experiment, we used the open-
source spectrogram algorithm from the SciPy library that
representation of the frequency spectrum of sounds or other
signals that vary with time. Then, we resized the image data
during the pre-processing stage so that both the image and voice
data are the same vector.
The second phase involved building a feed-forward neural
network that consists of four layers, as illustrated in “Fig. 2”.
The first layer is the input that will receive the two feature
vectors. The second layer takes as input, a list of identically
shaped tensors and returns a single tensor that is the
concatenation of all inputs. The third layer convert the data into
a one-dimensional array. Finally, the output predicted the
command. The activation function used was softmax with 10
classes, each class representing the command.

As can be seen, it has exceeded the results of both of [8]


and [9]. For [9] is the use of deep learning approach in training,
which helps increase accuracy. Similarly, in this model
exceeded that of [8] in the size of datasets used for training
approach the accuracy of a MUI model.

C. Fusion Model
In this article authors propose the data set used in the
experiment with propose model named Fusion Model.
1. Data Set
In this experiment was processed using two separate voice
and image data sets. For voice data set [7] contained one-second
WAV audio files, each with a single spoken English word. The
ten voice commands used in this experiment were “down,”
“go,” “left,” “no,” “off,” “on,” “right,” “stop,” “up,” and “yes,”
with a total of 2000 samples. For image dataset [32] was
composed of 10 different hand gestures performed by 10
different subjects (5 men and 5 women).
The authors use 60% for of the datasets were used for training D. Sensor-Based PH system
and 40% for testing to make each voice command corresponded In this article authors describe a scenario of using this sensor
to the respective for making the model fit the training dataset as to collect various information such as heart rate, EEG, blood
well as the test dataset hand gesture. sugar, pressure etc. and send it to a database. The data is stored
there for the involved parties to access. At the same time, the
2. Feature Level Fusion Model database uses some machine learning algorithm to analyses the
The model has two phases. First it extracts the features of collected data to determine a patient’s risk factor, health
voice and image datasets. Second, it carries out data fusion of
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 6

improvement and suggest further actions based on that. Figure they use Deep Q learning method for Learning-based Deep Q
3.1 n/w to manage health data and reduce malware attacks purpose.
1. IoT related issues in PH
BAN uses sensors to collect body related health information IV. CONCLUSION
from users, effectively transmit and transform sensed
phenomena into valuable information, and it needs to ensure In this article, various AI models has been evaluated for
that it will meet other system requirements, such as energy intelligent control and smart system management in various
efficiency. Furthermore, its ability to selectively process and scenarios, including smart home, energy management,
deliver information at fidelity levels and rates appropriate to the industrial IoT. We evaluated that with AI-enhanced model, the
data's destination. It is required to ensure the data integrity, recommendation and prediction are accurate.
energy efficiency of the sensors and ability to integrate with
existing infrastructure. REFERENCE
[1] R. HARPER, INSIDE THE SMART HOME. SPRINGER SCIENCE &
2. ML related issue in PH BUSINESS MEDIA,2006.
[2] B. K. SOVACOOL AND D. D. F. DEL RIO, "SMART HOME TECHNOLOGIES
The benefit of using ML is strategy analogy, decision making IN EUROPE: A CRITICAL REVIEW OF CONCEPTS, BENEFITS, RISKS AND
from existing data and predict from previous experience. ex, for POLICIES," RENEWABLE AND SUSTAINABLE ENERGY REVIEWS, VOL.
monitoring a patient, the ML-based method will analyses the 120, P. 109663, 2020.
situation according to the trained dataset. The training dataset [3] I. N. DA SILVA, D. H. SPATTI, R. A. FLAUZINO, L. H. B. LIBONI,
AND S. F. DOS REIS ALVES, "ARTIFICIAL NEURAL NETWORKS,"
most of the time plays a crucial role in predicting the future
CHAM: SPRINGER INTERNATIONAL PUBLISHING, P. 39, 2017.
trend of a given new problem successfully. for using ML-based [4] K. CARRIE ARMEL, A. GUPTA, G. SHRIMALI, AND A. ALBERT, "IS
PH service, predictive analysis can assist the hospital released DISAGGREGATION THE HOLY GRAIL OF ENERGY EFFICIENCY? THE
patients who may need to be readmitted to the hospital. CASE OF ELECTRICITY," ENERGY POLICY, VOL. 52, NO.
Predictive analysis is convenient to create a risk stratification SUPPLEMENT C, PP. 213-234, 2013/01/01/ 2013.
model where certain patients with higher risk are managed with
additional effort. [5] A. U. REHMAN ET AL., "APPLICATIONS OF NON-INTRUSIVE LOAD
MONITORING TOWARDS SMART AND SUSTAINABLE POWER GRIDS:
A SYSTEM PERSPECTIVE," IN 2019 29TH AUSTRALASIAN
UNIVERSITIES POWER ENGINEERING CONFERENCE (AUPEC),
2019, PP. 1-6.
[6] AGARWAL R., SHRIKANT R, “FAST ALGORITHMS FOR MINING
ASSOCIATION RULES; PROCEEDINGS OF THE INTERNATIONAL
CONFERENCE ON VERY LARGE DATA BASES” IN VLDB '94
PROCEEDINGS OF THE 20TH INTERNATIONAL CONFERENCE ON
VERY LARGE DATA BASES , SEPTEMBER 1994 , PP. 487-499
[7] BUCHNER AND JOHANNES, “SYNTHETIC SPEECH COMMANDS: A
PUBLIC DATASET FOR SINGLE-WORD SPEECH RECOGNITION.,” 2017.
ACCESSED: SEP. 23, 2019. [ONLINE]. AVAILABLE:
HTTPS://WWW.KAGGLE.COM/JBUCHNER/SYNTHETIC-SPEEC
COMMANDS-DATASET/.
[8] Y. JUNG, M. HAN, J. S. PARK, AND S. J. LEE, “DEVELOPMENT OF
RECOGNITION SYSTEM USING FUSION OF NATURAL
GESTURE/SPEECH,” IN 2008 DIGEST OF TECHNICAL PAPERS -
INTERNATIONAL CONFERENCE ON CONSUMER ELECTRONICS, PP.
E. AI Security in Healthcare 1–2, JAN. 2008, DOI: 10.1109/ICCE.2008.4588016.
1. AI Security in Healthcare [9] Y. JUNG, M. HAN, J. S. PARK, AND S. J. LEE, “DEVELOPMENT OF
RECOGNITION SYSTEM USING FUSION OF NATURAL
Authors present a study of paper related to how AI security GESTURE/SPEECH,” IN 2008 DIGEST OF TECHNICAL PAPERS -
services can assist healthcare and further observations are INTERNATIONAL CONFERENCE ON CONSUMER ELECTRONICS, PP.
discussed in this section [10] and 11] they use SVM classifier 1–2, JAN. 2008, DOI: 10.1109/ICCE.2008.4588016.
model Robust watermarking model for maintaining the [10] RAI A AND SINGH H V 2017 SVM BASED ROBUST WATERMARKING
FOR ENHANCED MEDICAL IMAGE SECURITY MULTIMED. TOOLS
robustness of embedded data using double-layer security
APPL. 76 18605–
purpose and Tiger Hash sign based AdaBoost with SVM
[11] VENKATESAN R, SRINIVASAN B AND RAJENDIRAN P 2019 TIGER
classifier to improve the security of multicasting routing in HASH BASED ADABOOST MACHINE LEARNING CLASSIFIER FOR
MANETs to leans more towards preservation of patient privacy SECURED MULTICASTING IN MOBILE HEALTHCARE SYSTEM
and secure healthcare. CLUSTER COMPUT. 22 7039–53
[12] MOHAMED SHAKEEL P, BASKAR S, SARMA DHULIPALA V R,
2. IoT Security in Healthcare using AI
MISHRA S AND JABER M M 2018 MAINTAINING SECURITY AND
In this section [12], authors provide a study of how AI can be PRIVACY IN HEALTH CARESYSTEM USING LEARNING BASED
used for IoT security in Healthcare and further observations. DEEP-Q-NETWORKS J. MED. SYST. 42DF

You might also like