A Deep Machine Learning Technology for Improving the Energy Efficiency of IoT

Devices in Long-Distance Wireless Networks applications

Abstract

Internet of Things (IoT), includes the countless physical devices in the world that are
linked to the internet, embedded with sensors, software, and other technologies. Variation in
the performance varies based on the device properties, such as enery utilization. IoT devices
are used for long-time tasks depending on applications and user requirements. The IoT
devices involved in long-time processes require high energy for efficient communication and
increased lifetime. Some of the challenges faced in utilizing the IoT devices include dropping
data transmission resulting in data loss, difficulty detecting the device during the low
condition of the battery. It can be recognized that energy consumption has to be saved
efficiently to enhance the lifespan of the IoTsystem. As a solution to these challenges, an
optimum method should be derived to optimize the energy. In this work, a novel Deep
Convolutional Neural Network (DCNN) is proposed to analyze the data acquired from IoT
devices. DCNN has a three-dimensional neural network for analysis, which helps to prevent
the node from reaching a low state, data loss, and communication breakdowns. These
objectives are met by allowing the model to learn and detect IoT devices via a dynamic
routing table. Dynamic routing tables are produced, sustained, and restructured by a routing
protocol processing in the router. This routing table will save the data and meta-data about
the IoT devices. The active routing table gets updated data after the completion of each task.
Thus, detecting the devices with lesser energy or unstable conditions will enhance the IoT
networks' Quality of Service (QoS). Experiments have shown that the suggested DCNN
outperforms conventional algorithms in terms of energy efficiency.

Keywords: IoT Networks, Energy Efficiency, Network Lifetime, Deep Learning Models,
Energy Consumption Reduction.

1. Introduction

IoT devices work similarly to other devices that transmit data through wireless mode
in the Internet of Things (IoT). The IoT devices can be daily drivers like electronic gadgets
instead of traditional devices like laptops, smartphones, and tablets. These IoT devices
communicate via Internet through Embedded Technology which is remotely controlled and
monitored. IoT devices play the role of home and industrial automation. These devices is able
to be classified into three major sections: consumer, enterprise, and industrial. In the case of
Smart Homes, the technology is meant to detect and respond to the behaviours of the
individual. The operational efficiency of the IoT devices can be achieved by advanced
programming, and the devices are termed Enterprise IoT devices. Enterprise IoT devices
comprise smart locks, smart thermostats, smart lighting, and smart security. Industrial IoT
devices are real-time sensors and robust devices implemented to automize manual works. The
work also includes automatic signaling about the functional status of the industrial
equipment.

The main working principle of IoT is connectivity among different devices that have
unique technologies for performing communication. Therefore, making all the devices
coordinate with a single technology is highly challenging. There are six widely used
technologies in the working of IoT Devices: Low Power Wide Area Networks (LPWAN),
Cellular Network, Zigbee, Bluetooth, WiFi, and Radio Frequency Identification (RFID).
LPWAN is a trending technology incorporated in IoT devices. They provide stable benefits
like affordable batteries that can function for a prolonged duration and supports remote
communication. Applications of LPWAN include asset tracking, monitoring environment,
occurrence detection, consumer monitoring, and others. However, LPWAN uses only low
bandwidth for working on the data, resulting in a higher duration for any communication.
LPWANs are widely used in cellular devices that aid in voice calls and video streaming. The
specific applications of cellular networks include fleet management, car infotainment, traffic
routing, etc. The major downside is the high cost and power requirements. The Next-
Generation technologies like 5G provide high-speed mobility and ultra-low latency, setting a
future with automated vehicles and augmented reality.
 Sample Diagram

In addition, the network coverage can be extended by spreading the sensor information over
sensor modules. Associated with the LPWAN, Zigbeeoffersadvanced data rates with reduced
power efficiency because of mesh configuration. Earlier to LPWAN, Mesh networks were
widely employed in industries and factories. Later, Bluetooth under Wireless Personal Area
Network aids in short-range communication. The data exchange in Bluetooth can be either
point-to-point or point-to-multipoint (up to seven slave nodes). Bluetooth Low Energy (BLE)
are compatible devices connected to electronic devices that assist in real-time data transfer
like smartwatches, headsets, and others. In recent days, these technologies have been replaced
by Wireless Fidelity (WiFi) and are employed abundantly in enterprises and homes as
gadgets, appliances, and other models. However, the working of WiFi is limited to high
coverage scalability and power consumption. Due to these limitations, it is highly challenging
to operate in IoT connectivities.

1.1. Energy Efficiency in IoT

Four parameters may be used to calculate the energy efficiency of IoT devices. They are: (1)
energy required for each reliable data transmission, (2) energy is consumed every transaction,
(3) delay vs. energy trade-off, and (4) network lifespan (time till the sensor's energy runs
out). The energy is used by a sensor node to perform acquisition, communication, and data
processing activities. The advancement of information collecting through the monitoring of
nodes in networks is known as data acquisition.

Sample Diagram
The primary reason for energy emission and reception is communication among network
nodes. Intermediate nodes utilize data processing concepts to gather and consolidate data
from various sensors available in the source nodes to the sink node. Most IoT devices
consume energy from batteries and may reduce power consumption. There are five effective
methods to reduce the power consumption of IoT devices. They are
1. move the ideal nodes (nodes that are active but not performing any transactions) to the
sleep state,
2. Wake up the sleeping node (if it has to receive any transaction) by invoking the
device processor and the internal clock,
3. Receiving (Rx): The device scans to receive possible data,
4. Listen- It processes the data received and makes it ready for transmission, and
5. Transmission (Tx): It transmits the processed data.
1.2. Improving Energy Efficiency inIoT Networks

In this paper, the energy consumption of IoT devices is discussed in detail. One of the
significant hurdles for IoT technology is the cost of power consumption. However, demand
for battery power cannot always be a better solution in terms of economy. The battery
capacity is a significant concern that increases battery rates and service time consumed by
batteries. Rechargeable batteries are also an issue in usage due to their hazardous nature and
environmental effects. At the same time, continuous access to the power grid is also highly
challenging. Energy harvesting can be adopted to save energy from the surrounding
environment. They used piezoelectric materials, thermoelectric material, solar energy, and
wind energy to save energy and reduce environmental issues. Harvesting techniques such as
solar-powered IoT device kit, small solar panel, energy-efficient motherboard & Bluetooth,
USB Bridge, and energy Harvest converts kinetic energy into electrical energy are adopted in
IoT Devices. Another significant source is energy harvesting reverb energy that can generate
energy from vibration even at low-frequency vibrations like ten to 30 Hz. Rather than using
energy efficient technologies, IoT device makers make use of a specific protocol with three
types.

1. Power-Saving Mode (PSM)

2. Employing Extended Discontinuous Reception(e-DRX) protocols
3. Wakeup signals.
 Many cellular networks switch off the connectivity to save energy. However, it
requires more additional power to reconnect. This power-saving mode uses a protocol to stay
connected with the network for a certain period by ignoring the power consumed for re-
connection. e-DRX functions are similar to PSM. This method periodically shuts down the
IoT devices, leading to delayed data transmission, but it is not a challenge in many IoT
devices. Wake-up signals help the IoT devices sleep and stop the periodical checking for
signals when the device is sleeping. If it receives a message, it gets a wake-up signal to turn
on the device. This method is beneficial to devices that need to communicate for a short
period. The IoT devices are evolving with power-saving features and, combined with
Artificial Intelligence (AI), will make a significant difference in the future to save power.

1.3. Machine Learning Algorithm for Enhancing Energy in the IoT Devices

IoT devices are connected and communicate by sharing a massive volume of data
every day. Machine Learning algorithms are inevitable in process extraction through
programming for processing the collected data. Ensemble algorithms are utilized in Machine
Learning to give a solution by blending various base models and proposing a new model.
This new model uses cameras to capture samples, train various classifiers, and combine their
results to reduce accidents. The Bayesian network can be illustrated by a Directed Acyclic
Graph (DAG), which uses probability distribution for a given set of random variables. This
method is used to find the short-term traffic. Markov models use probability distribution, and
it is a stochastic sequence. Decision trees are top-down approach-based structures that
classify the task into several results, such as congestion reduction. Deep learning (DL) is a
group of ML algorithms, which makes use ofseveral layers to graduallyremove highest-level
topographies from the input. It can be considered as an ideal algorithm that suitsIoT devices.
If we consider image processing as an example, its inferior layers canrecognizeboundaries,
andsuperior layers can recognize the perceptionsappropriate to a humanoid. The term deep
learning (DL) is a part of the Machine Learning (ML) model, which functions based on the
working principle of Artificial Neural Networks (ANN). It is mainly used in resource-
intensive tasks like image pattern recognitionspeech recognition, and synthesis.

Each device in an IoT network has specific property and function. Therefore, it is
mandatory to have prior knowledge about each device to process massive data of different
characters. Deep learning algorithm helps the surveillance cameras make decisions based on
the person's activities and can be programmed to react accordingly. It provides device
security inIoT with the help of specific data classes.

1.4. Contribution of the Paper

The main objective of this paper is to optimize the IoT devices and perform Task
Scheduling in IoT-Cloud. For this, first, we have to create a sample IoT network scenario
with a defined number of IoT devices.A deep CNN architecture is created for analyzing each
IoT device in the network at a defined interval of time. From the obtained results, an energy
efficiency model, based on IoT nodes for resource allocation and task assignmentis designed.

To understand the problems a detailed literature survey was carried out and given below.

2. Literature Survey

Modern technical discoveries and the rapid fusion of industries such as sensors/sensing and
actuation applications, embedded networks, wireless networks, and information technology
are accelerating the growth of the Internet of Things (IoT). The Internet of Things is being
used to make significant revenues for corporations [1-3]. The exponential increase in data
utilisation, as well as the fast expansion of severalIoT-connected modules, demonstrates how
big data and IoT are increasing in lockstep.

Almost every element of technology and business is affected by the convergence of these
professions [4,5]. The process of evaluating each data segment to detect configurations,
expose hidden information, and collect useful data is known as data analytics. The primary
purpose of IoT-based information analytic systems is to increase data explanation in order to
make effective switchjudgments. These options can be utilised to accomplish and simplify a
wide range of tasks.

The IoT system needs a sophisticated data analysis mechanism that can execute duties like
organisation, assumption, bunching, recession, combination rule mining, and other
operations. Deep learning (DL) has been widely used to evaluate information produced by
IoT systems. IoT and DL have both been selected as two of the top three technological trends
for 2017 [6]. The incapacity of traditional MLsystems to meet the growing analytical
requirements of IoTmodules is the catalyst for this obsessive DL development. In numerous
aspects, deep learning models surpass traditional machine learning approaches. For starters,
they no longer need the usage of labelled data in training. As a result, deep learning systems
may swiftly harvest characteristics that are difficult for humans to recognise. Furthermore,
these approaches outperform earlier ones in terms of accuracy. DL architectures are also
well-suited to simulating complex multimodal dataset dynamics [7].

Several academics have conducted research on IoTsecurity in order to provide a helpful tool
for identifying present IoT system security flaws as well as a roadmap for future study. Most
previous surveys on IoT security, on the other hand, haven't given much attention to machine
learning and deep learning applications for IoT security. Surveys [8-14], for example,
examined existing research and categorised IoT security challenges in the areas of
encryption, verification, access management, system and application security. Granjal,
Monteiro, and Silva [15] stressed the necessity of IoT communication safety after reviewing
problems and solutions for the security of IoT communication networks.

Zarpelo et al. conducted an interruption exposure for IoT systems. [16], Weber [17] looked at
legal issues and regulatory procedures to assess if IoT frameworks met privacy and security
requirements. Roman, Zhou, and Lopez [18] examined security and privacy in the dispersed
IoT context. They also listed a number of challenges to address, as well as the advantages of
a distributed IoT technique centred on security and secrecy. The study [19] investigated how
IoT systems are growing increasingly sensitive to threats such as ransomware and security
issues. In the IoT scenario, Xiao et al. [20] looked at machine learning techniques for
information security and secrecy.

In addition, this study identified three challenges in the future implementation of machine
learning in IoT systems (i.e. calculation and communication slide, safety methods, and partial
state surveillance). Other research papers, such as [21, 22], focus on the use of data mining
and machine learning algorithms to predict intrusion vulnerability in cybersecurity. Data
mining and machine learning technologies were used to investigate cyber domain security, as
well as cyberspace misuse and anomaly detection [21, 22].
Sample Diagram

Most of the research works are have used statistical methods for analyzing numerical
datasets where they can estimate the statistical information such as mean, standard deviation,
In [23], the author has proposed a data processing application called cross-test. It is
implemented along with several machine learning algorithms. The highest accuracy is
obtained when a cross-test is implemented with the LARS Lasso algorithm. In [24],
Khanouche et al. proposed an Energy-centered and QoS-aware services Selection Algorithm
(EQSA). It utilizes relaxation techniques & a lexically based optimization strategy on Quality
of Service (QoS). AzhamHussain et al. have proposed a Genetic Algorithm based Adaptive
Offloading (GAOA) method that efficiently handles the traffic in the IoT cloud. Ding et al.
[25] have reduced the energy consumption in IoT devices by deploying multi-objective fuzzy
algorithms. In [26], Li et al. has proposed an online algorithm utilizing edge computing
techniques. Using this algorithm, the maximum efficiency in data transmission for the IoT
deep learning applications is attained. In [27], Li et al. proposed a Harmony Search
Algorithm (HSA) on the clustered nodes to Optimize Energy in a WSN. In [28], Baek et al.
has proposed a Partial-Observable Markov Decision Process (POMDP) that offers service
selection in IoT devices to cope with the dynamic environment. In [29], Ventura et al. has
achieved highly energy-efficient IoT devices with reduced eco-wastes with their proposed
model- A Real IoT Implementation of a Machine-Learning Architecture (ARIIMA) using
Autoregressive Integrated Moving Average models (ARIMA). In [30], Ke et al. Proposed a
Mobile Edge Computing (MEC) infrastructure to perform data offloading optimization and
allocation of renewable energy. Rayleigh's fading technique is used s a communication
model. In [31], Guo-Sheng Liu proposed a Genetic Algorithm (GA) deploying Link Control
Protocol (LCP) and probabilistic heuristics to reduce the time taken to process and achieve
energy optimization in the overall system. In [32], the authors proposed a multi-objective
genetic algorithm (MOGATETC) that uses dominance rules and heuristic approaches. The
computational time & energy utilized is reduced.

3. Proposed Work

In the recent era of technology, IoT has grown significantly. The optimization of
energy in IoT Devices is of great importance, and thus many types of research are carried out
in this regard. This research article will study the energy optimization of IoT Devices with a
Resource Allocation Algorithm using Reinforcement Learning. The Energy-based task
scheduling in IoT architecture is the critical research subject when IoT is considered. A
DCNN model with various components for data transmission is deployed to carry out the
computational tasks effortlessly. DCNN has inbuilt storage and processing foundational
structure in the mobile network technology to facilitate ultra-low latency and increased
bandwidths for providing services rendered by the Internet of Things (IoT) devices. The
intensive computational tasks were performed with the help of the main DCNN Servers,
various base stations, and various buffers to maintain consistency in tasks. There are many
benchmark algorithms present to perform energy optimization. They are as follows, Dueling
Deep Q-Network(DQN), Dynamic Deep Q-Network (DDQN), and Greedy policy. This
proposed model employs the Resource Allocation Algorithm (RAA) and neural network-
based Reinforcement learning technique for energy optimization. It outperforms the
benchmark algorithms and results in less energy consumption, thus achieving energy
optimization in IoT Devices.
3.1. Energy Efficiency methods

The energy efficiency can be enhanced by regularizing the data transmission &
routing process. The use of natural energy sources also helps in this process. The main
methodologies involved in energy optimization are as follows,

 Statistical Methodologies
 Routing Optimization
 Scheduling Optimization
 Data Offloading
3.2. Deep Reinforcement Learning Optimization

The proposed model consists of an Deep Q-Network (DQN) algorithm that may
handle policies based on time series, to manage numerous tasks. It is done for the
conservation of energy. The tasks are nothing but offloading and execution of local data. As
queuing plays a significant role in managing the tasks of the system, more importance is
given to queue-based algorithms. The DQN algorithm is adopted considering various factors
it is a deep learning algorithm based on reinforcement learning that makes decisions based on
the situation. The Deep Deterministic Policy Gradient (DDPG) algorithm manages the
policies adopted by the system. It is used along with DQN in the proposed system. DDPG
makes decisions from the distribution of decision sets, and it is also capable of making
decisions out of smaller sets.

Sample Diagram

A decision At is considered for a device is shown as in Equation (1).

n
 *
At =Q ( s t n ∨ϕ target ) −C
n
(1)

Where ,t n is the time , st is the State−space, Q* isthe policy , ϕ target is the parameters of the
n

device network and C is represent for reduce the noise.

The neural network is updated with parameters determined by the deterministic policy
gradient which is represented in the following Equation (2).

∇ϕ tar get
*
tn [
f ( Q ( t n ) )=E S ∇ϕ target
f (Q ( St ) )∇ A Q ( St , Ak )
*
n k n ] (2)

where, A k =Q ( S t )and it shows that only discretized values are used by the deterministic
*
n

policy gradient. The least-square method is adopted by the DDPG for assigning parameters to
the device and critic network. State-space st comprises of buffer queue length L ( t n ) , channel
n

matrix H kxm ( t n ), SINR and offloading ratio. The action space contains the energy spent by the
devices as well as the bandwidth allocation ratio among the layers At . n

3.3. Reward function

Previous state st action provides the reward to the state t n. Reinforcement learning
n−1

tries to maximize the cumulative records. The cost function is frequently used to describe the
reward function. For transforming minimization issues into maximizing problems and for
convergence, the cost function is regarded a negative valued function. Finally, the reward
function comprises bandwidth, consumption of energy, and buffer queue size is depicted in
Equation (3), Equation (4), and Equation (5) respectively.
 Sample Diagram correct the diagram equation based on the paper equation

r t n=−E [ ψ 1 I 1 α 3 I 2+ ( 1−ψ 1 ) α 4 L ( t n )−P ( t ) ] (3)

K k
(4)
I 1=α 1 ∑ Ei ( t n ) +α 2 ∑ π i B m ( t n )
i=1 i=1

K N
(5)
I 2= ∑ ∑ π n B n ( t n )
i=1 j=1

Where Penalty ratio coefficient is ψ 1, Energy consumed is E k ( t n ), Weight parameters

ψ 1 ≤ 1 , α 1 , α 2 , α 3 , α 4, The bandwidthis allocated as Bm ( t n) , Bn ( t n ) , L ( t n ) is the buffer queue

length. The negative penalty factor obtained from the buffer overflow is P ( t n ).

The cumulative reward can be estimated as in Equation (6).

[ ] (6)
T−1
1
R=max lim
T→∞ T
∑ ri
i=0

Where, r i indicates the reward function at time ‘t’.

3.4. DCNN (Deep Convolutional Neural Network)

Deep Learning is a subset of AI and Machine Learning that mimics the human brain's ability
to process data, create patterns, and apply them to make decisions. Figure 2 shows the
comparison of machine learning (ML) and Deep learning (ML). Deep learning, also known
as deep neural networks, uses numerous layers to extract features from input data in a
progressive manner.A dataset is a collection of data that is made up of lists of data for certain
variables or objects. A dataset can also comprise a variety of documents or files in various
formats. The features that are used to construct predictions are discovered in the sets of data
contained in the observation. In a real-world scenario, the Dataset would contain observations
with various features such as Temperature, Luminosity, Infrared, Distance, Time, etc...
 Figure 2. Comparison of machine learning (ML) and Deep learning (DL).

3.5. Algorithm-1: DCNN (IoTparam, Networkparam)

Initialize all the parameters of the IoT and the network;

Initialize the memory M;

For I = 1 to max (M) do

Reset all the IoT devices, base station, and environment variables.

For j = 1 to T

Divide the data into training data and testing data

Start performing IoT devices

Calculate the network parameters REQ, RES.

If all REQ gets RES then

Allocate resources and start processing.

Else
 Resend the REQ for RES

End if

Calculate the energy of IoT and select just the most energy-
efficient IoT

End I

The proposed learning algorithm uses the real-time dataset, which includes a set of
data or information that is delivered immediately after its origin. This large amount of
collected data increases the accuracy of the testing process. The number of IoT devices used
in the experiment is varying from 10 to 150, are communicating with different base stations.
Each device is communicating with the base station with the help of sub-base stations. The
collected data or the message queue are transmitted then the garbage collector is called to
clear the queue to accept the next queue data. Each IoT device is assigned fully energized
batteries. Each device needs a small amount of energy to perform the defined task. After the
task completion, the energy is reduced. The calculation is carried out using the following
Equation (7).

Sample diagram correct the diagram equation based on the paper equation

Erem =E tot −(E Tx + E Rx + El + E sl ) (7)

where, Erem is the Remaining Energy, Etot is Total Energy, ETx -Transmission Energy, E Rx -
Receiving Energy, El -Listening Energy, and E sl-Sleeping Energy.
 The algorithm of all the IoT devices and other environmental factors are initialized
first. Each time of the new process, the network and IoT devices’ parameter values are
initialized or reset. After initializing all the parameter values are newly assigned according to
their defined performance and feed to the DCNN model, where it learns and analyzes all the
parameters and starts performing data transmission and receiving.

3.6. Mechanism of DCNN

The deep convolutional neural networks (DCNN) are bioinspired variations of multiple layer
serceptrons (MLPs). A visual cortex of an animal has two types of cells, each of which is
sensitive to a tiny portion of the visual field (receptive field). As a result, unlike classic fully
connected neural networks, a neuron in a DCNN is solitarylinked to a tiny perceptron of its
preceding layer, somewhat, complete neurons in the preceding layer. A DCNN is a heap of
three functional layers, as depicted in Figure 3, in the simplest case: Convolutional Layer,
Pooling Layer, and Fully Connected Layer.

Figure 3. Schematic diagram of DCNN architecture

The Convolutional layer is considered as the foundation of DCNN; its primary function is
convolution, which generates the dot-product of regions of interest as well as a collection of
convolutional filters (or kernels) (1). The procedure of convolution procedures is depicted in
Figure 4.
 Figure 4 . Layout of Convolution process

After the convolution processes, the pooling layers perform nonlinear downsamplings to
lower the data dimension. Max pooling and average pooling are the two most frequent
pooling algorithms. Max pooling determines the candidates' maximum value, while average
pooling determines the candidates' average value. The collected feature maps are then given
to activation functions that perform non-linear changes after downsampling processes.

Rectified Linear Unit (ReLU) is given by, f(x) = max(0, x)

Sigmoid function is explained as, f(x) = (1 + e −x )−1

2
Hyperbolic tangent (tanh) function, is described as f (x)= −2 x -1.
1+e

The completely connected layer then completes the high-level logic. This layer's neurons are
linked to all of the previous layer's activation findings. Furthermore, the loss layer is the
DCNN's latter layer, and it determines by what method the difference among anticipated and
real markers is disciplined during the system preparationprocedure. For different applications,
alternative loss functions including such softmax loss and sigmoid cross-entropy loss can be
utilized.
 Figure 5. Functional blocks in DCNN. (a) Inward Product, (b) pooling, and (c) activation.

In the software/algorithm arena, the term "neuron" is commonly used. A neuron inside the
domain of DCNNs is made up of one or more basic operations.Figure 5 depicts the
characteristics of function blocks, with xi's representing components in a region of interest
and wi's representing elements in a filter in Figure 5(a). The average pooling and maximum
pooling function blocks are shown in Figure 5(b). The activation function block (e.g.
hyperbolic tangent function) is shown in Figure 5(c). The feature extraction block, which
removes features from feature maps, is made up of an inwardslab, a pooling block, and an
initiation function block.

4. Results and Discussion

The accuracy of the model is compared to that of standard feature-based classification

methods. The results indicate that DCNN models outperform their counterpart methods in
terms of classification accuracy, implying that they are extremely effective tools for detecting
and identifying coexisting devices in the presence of severe overlap and interference. By
improving the accuracy of automatic detection and identification, the proposed approach aims
to advance cognitive wireless awareness.The DCNN indicates how the transmission occupies
the spectrum over sending data. This shows the spectral efficiency of the algorithm over a
long period of time, and how much airtime is required in the process. Finally, the energy
efficiency shows how much energy needs to transmit a bit of data. As the total data
transmitted in the network is fixed, the higher the energy efficiency, the longer the network
lasts. Hence, the networkis more sustainable.
 Sample Diagram

The performance of the suggested strategy is assessed by comparing it to data

offloading and resource allocation. The neural architecture of the Duelling DQN and DDQN
is the same with its learning rate at10−3 . The state-bound and action are set as 5. The system’s
residual energy is connected to the coefficient’s value α 1. The parameter α 1value is assigned
as 0.5 by comparing the current energy levels in the systems. The experiments considered are
conducted using different parameters that are given in Table-1.

In terms of power usage, the proposed approach is quite efficient. The RLODORA is
also nearly efficient but not compatible with multiple environment support. The Greedy
algorithm is the least efficientbecause of its higher power consumption. The Duelling DQN
and DDQN are also relatively low compared to the proposed algorithm. The four algorithms,
such as RLODORA, Dueling DQN, DDQN, Greedy algorithms are compared with the
proposed algorithm. Theirimplementation is frequently analysedusing the quantization levels
At , compatible with other algorithms.
n

Figure6 shows that the cumulative rewards on average received by Duelling DQN and
DDQN increase with the increase in the discretized value with ω 1=0.5as the coefficient of
energy consumption, calculated from equation (6). The energy consumption coefficients for
the greedy and RLODORA algorithm are -1.41 and -1.32, shown in Table 2, and the reward
levels are relatively high, irrespective of the discretized levels. It shows that the proposed
system is consistent at all levels. The base stations and MEC servers face several
computational tasks. The computational time also varies with the size of the data and the
workload. It also affects the arrival time of the tasks.

Figure 6.Comparing discretized levels with Average cumulative rewards

Table-2. Comparing Discretized levels with the average cumulative rewards

Figure 3 shows that the suggestedalgorithm provides dominating results in with respect to the
arrival of tasks. By keeping the coefficient of energy consumption ω 1=0.5, the Dueling
DQN, DQN, RLODORA, and the proposed method received the cumulative rewards are -
3.14823, -3.09657, -3.2323, and -2.8273 respectively. The calculation of cumulative rewards
using RLODORA, Dueling DQN, DDQN, Greedy algorithms always showed negative
values, and among that the results of the proposed DCNN algorithm possess better results.

Figure7.Comparison of average cumulative rewards and arrival rate of computing tasks.

Figure 7 shows that the proposed algorithms have the highest average cumulative
rewards and decrease with the arrival rate of tasks. The greedy algorithm performs the least
that can be seen in the graph. The rest of the algorithms perform relatively well, but the
proposed algorithm performs well in both cumulative rewards and cost-efficiency.

Table-3. A comparison table of arrival rate and average cumulative rewards.

 The proposed algorithm is compared using RLODORA, Dueling DQN, DDQN,
Greedy algorithms,based on their rewards and computing tasks. Table 3 shows the
comparison of the arrival rate of tasks and average cumulative rewards. The greedy algorithm
shows the worst performance compared to all various algorithms. It makes the optimal choice
and attempts to find the optimal way, but it fails to collect the cumulative rewards of the
arriving computing tasks. Next, both the Duelling DQN and DDQN algorithms perform
relatively better than the greedy algorithm. The Q-value estimators in both the algorithms
make unbiased decisions and perform well in providing cumulative rewards. However, the
proposed algorithm makes effective decisions by predefined variables, optimizes energy
usage, and periodically assesses the networks' performance.

Table.4Comparison cost of energy consumption by different algorithms in the context of

Arrival Rate

(Joule)

Table-4 shows the energy consumed by all the algorithms considered. The
RLODORA algorithm performs relatively well than the other algorithms because it uses
optimization variables to manage the tasks. The proposed algorithm also follows the
technique but with more variables that provide more room for decentralization.
Figure 8.Comparison of the Arrival rate of Computing task and cost of energy consumption

The proposed algorithm consumes the least energy compared to the rest of the
algorithms in energy consumption. The worst is the greedy algorithm. Dueling DQN also
consumes more energy because the energy consumption relies on the optimization and the
resources required to run the algorithm. In that way, the proposed algorithm consumes less
energy than other algorithms and the rest lack optimization.

Table-5. Comparison between Algorithms based on the average length of the computing

tasks
 Generally, in IoT networks, the tasks to be computed will be higher, as it deals with
heterogeneous devices. The length of the computation queue shows a better view of the
overall performance obtained using the algorithm is given in Table 5. The proposed algorithm
relatively holds less number of tasks compared to other algorithms and the graphical
representation is given in Figure 9.

Figure 9. Energy consumption at each node

Figure 6 illustrates the energy consumption based on the payload size of three nodes.
The first and second hops indicate the routers, and the third hop represents the server node. In
the experiments which use internet protocols (IP) like IPv4 and IPv6, the energy consumption
differences of every node are to be considered for the demand. Because of the transmissions
and receptions on both sides, the routers are extremely energy-concentrated in comparison to
the server. The router of the first hop consumes less energy when compared to the second hop
and loses its energy whenever it switches on the radio. But, the second hop node does not
lose energy while transmitting multiple strobes.

4.1. Experimental Results and Discussion

In the experiment, to analyze the results and to construct the environment, some of the
parameters are initialized with some constant values, and it is given in Table-1. The
performance of the proposed method depends on the values given in Table-1. Also,
depending on the values given in Table-1, the proposed algorithm is implemented in any
computer programming language and the results are verified. For understanding the entire
process, the pseudo-code for the deep learning algorithm is given in Algorithm-1.

Table-1. List of Environment Parameters

Parameters Values

Queue length 1100 m

No. of IoT devices 130

No. of small base stations 2

Time slot length 2ms

The bandwidth of Multi-Access Edge 7 Mbps

Computing(MEC) Server
The bandwidth of small base stations 4Mbps

Static power of MEC server 0.5 W

( α 1 , α2 , α3 , α 4 ) (0.5, 0.5, 0.5, 0.5,)

Depending upon the values provided in Table-1, the proposed algorithm is

implemented and the results are verified using the deep learning algorithm,which is given in
Algorithm-1.

5. Conclusion

In this paper, an unique MEC module is created to construct an IoT device network
with tiny base stations and MEC servers employing heterogeneous components. Processing,
transporting, and computing jobs demand more energy and time. To reduce this, DL
optimization techniques are presented, which may also provide dynamic solutions by
calculating data from many sensors in the IoT network. In terms of regulating the dynamics
of the arrival of computing tasks, the suggested algorithm also has a higher learning rate. The
suggested approach outperforms other standard frameworks in terms of optimization, energy
consumption, and computing efficiency, according to simulation and comparison of the
above-mentioned works. As a result, recognising devices with lower energy or in unstable
settings would improve the Quality of Service of IoT networks (QoS). The suggested DCNN
has a greater energy efficiency, according to the experiments. Because it can always pick the
most energy-efficient method, the recommended algorithm's upper bound benchmark has a
significant benefit. This might be the quickest route, resulting in a significant increase in
efficiency. We may infer that using DCNN helps increase the performance and energy
efficiency of IoT networks that are monitored remotely.

REFERENCES

[1] Al-Fuqaha, A., Guizani, M., Mohammadi, M., Aledhari, M., &Ayyash, M. (2015).
Internet of things: A survey on enabling technologies, protocols, and applications. IEEE
communications surveys & tutorials, 17(4), 2347-2376.
[2] Zahra, S. R., &Chishti, M. A. (2019). Assessing the services, security threats, challenges
and solutions in the Internet of Things. Scalable Computing: Practice and Experience, 20(3),
457-484.
[3] Zahra, S. R., &Chishti, M. A. (2019, January). Ransomware and internet of things: A new
security nightmare. In 2019 9th international conference on cloud computing, data science &
engineering (confluence) (pp. 551-555). IEEE.
[4] Ahmed, E., Yaqoob, I., Hashem, I. A. T., Khan, I., Ahmed, A. I. A., Imran, M.,
&Vasilakos, A. V. (2017). The role of big data analytics in Internet of Things. Computer
Networks, 129, 459-471..
[5] Zahra, S. R., &Chishti, M. A. (2020). Fuzzy logic and fog based secure architecture for
internet of things (flfsiot). Journal of ambient intelligence and humanized computing, 1-25.
[6] Panetta, K. (2017). Gartner Top 10 Strategic Technology Trends for 2018-Smarter With
Gartner. Gartner, Inc.
[7] Chen, X. W., & Lin, X. (2014). Big data deep learning: challenges and perspectives. IEEE
access, 2, 514-525.
[8] Sfar, A. R., Natalizio, E., Challal, Y., &Chtourou, Z. (2018). A roadmap for security
challenges in the Internet of Things. Digital Communications and Networks, 4(2), 118-137.

[9] Sicari, S., Rizzardi, A., Grieco, L. A., &Coen-Porisini, A. (2015). Security, privacy and
trust in Internet of Things: The road ahead. Computer networks, 76, 146-164.

[10] Alaba, F. A., Othman, M., Hashem, I. A. T., &Alotaibi, F. (2017). Internet of Things
security: A survey. Journal of Network and Computer Applications, 88, 10-28.
[11] Zhao, K., &Ge, L. (2013, December). A survey on the internet of things security. In
2013 Ninth international conference on computational intelligence and security (pp. 663-
667). IEEE.

[12] Kumar, J. S., & Patel, D. R. (2014). A survey on internet of things: Security and privacy
issues. International Journal of Computer Applications, 90(11).

[13] Suo, H., Wan, J., Zou, C., & Liu, J. (2012, March). Security in the internet of things: a
review. In 2012 international conference on computer science and electronics engineering
(Vol. 3, pp. 648-651). IEEE.

[14] Kouicem, D. E., Bouabdallah, A., &Lakhlef, H. (2018). Internet of things security: A
top-down survey. Computer Networks, 141, 199-221.

[15] Granjal, J., Monteiro, E., & Silva, J. S. (2015). Security for the internet of things: a
survey of existing protocols and open research issues. IEEE Communications Surveys &
Tutorials, 17(3), 1294-1312.

[16] Zarpelão, B. B., Miani, R. S., Kawakani, C. T., & de Alvarenga, S. C. (2017). A survey
of intrusion detection in Internet of Things. Journal of Network and Computer Applications,
84, 25-37.

[17] Weber, R. H. (2010). Internet of Things–New security and privacy challenges. Computer
law & security review, 26(1), 23-30.

[18] Roman, R., Zhou, J., & Lopez, J. (2013). On the features and challenges of security and
privacy in distributed internet of things. Computer Networks, 57(10), 2266-2279.

[19] Yaqoob, I., Ahmed, E., urRehman, M. H., Ahmed, A. I. A., Al-garadi, M. A., Imran, M.,
&Guizani, M. (2017). The rise of ransomware and emerging security challenges in the
Internet of Things. Computer Networks, 129, 444-458.

[20] Xiao, L., Wan, X., Lu, X., Zhang, Y., & Wu, D. (2018). IoT security techniques based
on machine learning: How do IoT devices use AI to enhance security?. IEEE Signal
Processing Magazine, 35(5), 41-49.

[21] Buczak, A. L., &Guven, E. (2015). A survey of data mining and machine learning
methods for cyber security intrusion detection. IEEE communications surveys & tutorials,
18(2), 1153-1176.
[22] Mishra, P., Varadharajan, V., Tupakula, U., & Pilli, E. S. (2018). A detailed
investigation and analysis of using machine learning techniques for intrusion detection. IEEE
Communications Surveys & Tutorials, 21(1), 686-728.

[23] Chilipirea, C., Ursache, A., Popa, D. O., & Pop, F. (2016, September). Energy efficiency
and robustness for IoT: Building a smart home security system. In 2016 IEEE 12th
international conference on intelligent computer communication and processing (ICCP) (pp.
43-48). IEEE.

[24] Hahm, O., Baccelli, E., Schmidt, T. C., Wahlisch, M., &Adjih, C. (2016, December). A
named data network approach to energy efficiency in IoT. In 2016 IEEE Globecom
Workshops (GC Wkshps) (pp. 1-6). IEEE.

[25] Lin, K., Wang, D., Xia, F., &Ge, H. (2017). Device clustering algorithm based on
multimodal data correlation in cognitive Internet of Things. IEEE Internet of Things Journal,
5(4), 2263-2271.

[26] Zhang, X., Ramachandran, A., Zhuge, C., He, D., Zuo, W., Cheng, Z., ...& Chen, D.
(2017, November). Machine learning on FPGAs to face the IoT revolution. In 2017
IEEE/ACM International Conference on Computer-Aided Design (ICCAD) (pp. 894-901).
IEEE.

[27] Shao, X., Yang, C., Chen, D., Zhao, N., & Yu, F. R. (2018). Dynamic IoT device
clustering and energy management with hybrid NOMA systems. IEEE Transactions on
Industrial Informatics, 14(10), 4622-4630.

[28] Feng, X., Zhang, J., Ren, C., & Guan, T. (2018). An unequal clustering algorithm
concerned with time-delay for internet of things. IEEE access, 6, 33895-33909.

[29] Xiong, J., Ren, J., Chen, L., Yao, Z., Lin, M., Wu, D., &Niu, B. (2018). Enhancing
privacy and availability for data clustering in intelligent electrical service of IoT. IEEE
Internet of Things Journal, 6(2), 1530-1540.

[30] Zhao, S., Rengasamy, P. V., Zhang, H., Bhuyan, S., Nachiappan, N. C.,
Sivasubramaniam, A., ...& Das, C. (2019, July). Understanding energy efficiency in iot app
executions. In 2019 IEEE 39th international conference on distributed computing systems
(ICDCS) (pp. 742-755). IEEE.
[31] Ciccia, S., Giordanengo, G., &Vecchi, G. (2019). Energy efficiency in IoT networks:
Integration of reconfigurable antennas in ultra low-power radio platforms based on system-
on-chip. IEEE Internet of Things Journal, 6(4), 6800-6810.

[32] Song, L., Chai, K. K., Chen, Y., Loo, J., Jimaa, S., & Iraqi, Y. (2019). Energy efficient
cooperative coalition selection in cluster-based capillary networks for CMIMO IoT systems.
Computer Networks, 153, 92-102.