Received: 15 May 2020 Revised: 3 July 2020 Accepted: 30 July 2020

DOI: 10.1002/ett.4122

SPECIAL ISSUE ARTICLE

Adaptive energy saving algorithms for Internet of Things

devices integrating end and edge strategies

Yang Wang1 Kun Yang1,2 Weixiang Wan1 Haibo Mei1

1
School of Information and
Communication Engineering, University Abstract
of Electronic Science and Technology of By taking into account not only the node specifics but also the benefits of
China, Chengdu, China
mobile edge computing, an integrated strategy for energy saving of Internet of
2
School of Computer Science and
Things (IoT) devices is proposed in this article. This strategy consists of two algo-
Electronic Engineering, University of
Essex, Colchester, UK rithms at both the end and the edge. Considering the changeable battery level
and downlink communication traffic of the battery-powered wireless nodes, an
Correspondence
Haibo Mei, School of Information and
energy efficient automatic mode switching algorithm is designed at the end.
Communication Engineering, University Three different kinds of working modes are designed based on the features of
of Electronic Science and Technology of the end nodes and the various application requirements. This algorithm tends
China, Chengdu, Sichuan, China.
Email:[email protected] to enable the end nodes to automatically select and switch to the proper work-
ing mode according to the actual conditions. At the edge server with much
Funding information
stronger processing and computing capabilities which belongs to a higher layer,
Natural Science Foundation of China,
Grant/Award Number: 61620106011; the dynamic sampling rate adjustment algorithm is designed. It can adaptively
Zhongshan City Team Project, adjust the sampling frequencies of the end nodes and thus reduce their work-
Grant/Award Number: 180809162197874
ing durations. The proposed integrated solution aims to decrease the energy
consumption of IoT devices as much as possible and thus prolong their bat-
tery lives. The simulation results have shown both the effectiveness and the
efficiency of the proposed end and edge integrated strategy in terms of energy
consumption.

1 I N T RO DU CT ION

Internet of Things (IoT) is the extension of Internet connectivity into physical devices. With various networking tech-
nologies like WiFi, Bluetooth, ZigBee, Ethernet, and so on, local devices can interact with others over the Internet and
be remotely monitored and controlled. However, as the number of deployed IoT devices increases nowadays, the high
energy consumption of the IoT devices vs the limited capacity of the batteries, which are difficult to replace or recharge,
energy saving has gradually become one of the biggest challenges.
Much research has been conducted for energy efficient operation of IoT networks. Clustering has been proven to be
one of the most efficient techniques for saving energy of wireless sensor networks (WSNs) and Rao et al1 propose an
energy efficient cluster head selection algorithm which is based on particle swarm optimization (PSO) called PSO-ECHS.
However, this work does not consider the benefit of MEC. Some work exploit MEC for energy saving of IoT nodes but
largely staying at the high level without considering IoT network and node specifics.2,3 Based on mobile edge computing,
in order to achieve energy efficiency, Yang et al3 model the energy consumption of offloading from both task computation
and communication aspects and develops an artificial fish swarm algorithm based scheme by taking into account the

Trans Emerging Tel Tech. 2020;e4122. wileyonlinelibrary.com/journal/ett © 2020 John Wiley & Sons, Ltd. 1 of 21
https://doi.org/10.1002/ett.4122
2 of 21 WANG et al.

unique property of 5G. In summary, energy-saving strategies are carried out at both edge server and end node sides but
largely in isolation. This article aims to bringing these two sides together into an integrated solution which can take the
advantages of both. In particular this article focuses on the ultimate contributor of energy consumption of end nodes, that
is, working duration, and endeavors to reduce it as much as possible. For this purpose, a two-end strategy is proposed in
this article, in consideration of not only the node specifics but also the benefits of MEC.
At the IoT device side, the concept of working mode is introduced, which categorizes the working status of an IoT
device (simply called end node or node) into three modes with increasingly longer working duration and thus more energy
consumption. Based on the proposed three working modes, an energy efficient automatic mode switching algorithm
is proposed. What is more, this part of work is mainly based on the ZigBee-enabled low-power IoT network, espe-
cially for the battery-powered nodes. As for ZigBee, it is based on IEEE 802.15.4 standard, especially for low-power, self-
organizing, high-security networks. It offers a complete IoT solution for home automation, smart energy, healthcare, and
so on.
At the MEC side, the edge server is physically closer to the IoT devices than the cloud. Thus, the communication from
IoT devices to the edge server costs much less energy than that to the cloud. Accordingly, the communication delay is
greatly reduced too. Furthermore, the edge server is at a higher layer and has an overview of the IoT networks it serves.
Based on that, an algorithm that dynamically adjusts the sampling frequencies of nodes is proposed, aiming to further
reduce a node’s working duration (sometimes also called working cycle).
By integrating both sides together, a more adaptive strategy is presented in this article, which addresses the critical
energy consumption issue from both microlevel (node) and macrolevel (edge). The main purpose is to reduce the energy
consumption of the IoT devices and prolong the devices’ battery lives. Thus, the service life of the whole system can be
prolonged.
The remainder of this article is organized as follows. Section 2 describes the related work on the energy saving efforts
for IoT devices, in the edge and on the cloud. Section 3 makes a specific clarification on the overall system architecture
based on the practical deployment and the ZigBee network architecture which the research work of this article is based on.
The proposed energy efficient automatic mode switching algorithm and the dynamic sampling rate adjustment algorithm
are respectively designed and clarified in Sections 4 and 5. Performance evaluation and analysis of the proposed algorithms
are presented in Section 6. Finally, the article concludes at Section 7.

2 RELATED WORK

In order to reduce the energy consumption of the globally deployed IoT devices, extend the service life of the battery on
the equipment and thus extend the life of the whole network, various scientific researchers or relevant alliance organiza-
tions are making continuous efforts. From all aspects of the network, many energy-saving strategies and algorithms are
proposed.
Some researches propose algorithms from PHY layer. Alassery et al4 present a receiver approach for alleviating power
consumption of WSNs. Unlike other power consumption techniques, instead of decoding every received signal at the
receiver which consume too much power the proposed approach studies the histograms of sensors’ transmitted signals
to detect collisions, so the receiver can determine when the transmitted signals can be decoded without wasting precious
power decoding transmitted signals suffering from collisions. On top of that, there is some study on the effect of CPU
buffer on energy consumption.5
Anchora et al6 design an algorithm based on MAC layer. In this work, a new MAC protocol is designed and validated
based on an efficient setting of the node’s duty cycle as a function of the transmission times of the neighbor nodes. In
a duty cycle period, each node wakes up once to transmit and N times to receive, where N is the number of neighbors,
while it remains in sleep mode for the rest of the time. A TDMA-based scheduling scheme that balances energy saving
and end-to-end delay is proposed in the paper.7
There is much work done on the network routing.8-17 From this aspect, an energy-efficient multihop transmission
technique for wireless sensor networks (WSN) based on Dijkstra algorithm is proposed.8 This article groups the sensor
nodes into clusters according to their placement in the monitoring area, then organize the nodes within each cluster
by electing the appropriate node as cluster head and classifying the remaining nodes into active nodes and sleeping
nodes. Then, it selects the set of reliable relays which cooperate to forward data with the least transmit power. A novel
energy-saving routing protocol based on two-layers routing mechanism is proposed in the paper,15 in which the first layer
adopts the remaining energy centroid algorithm to form clusters and elect the cluster heads and the second layer adopts
WANG et al. 3 of 21

multihop transmission to guarantee that the cluster heads can transmit data to the base station along an optimal path.
For monitoring burst events in reactive wireless sensor networks, the paper17 proposes an novel energy-efficient dynamic
voting cluster (EE-DVC) algorithm.
There is more effort at the application layer.18-24 Carmona et al18 propose a transmission power management algorithm
based on link quality parameter to obtain savings in node energy consumption. The radio operation is the most power
consuming task in a wireless node network; therefore, the proposed algorithm tries to determine the optimum transmis-
sion power level to use only the required energy. A transmission power control strategy that is used to reduce high power
consumption in wireless sensor networks (WSNs) is proposed in the paper.19 Le tries to limit the transmission of infor-
mation over the wireless network and ensures that the sensor node transmits enough information about the host at the
same time.20 This article introduces a lossless data compression algorithm to suit the wireless sensor network and meet
the transmission requirements.
There is also a lot of research which not only considers the issue of energy consumption but also how to make the
devices maintain the normal operation by energy harvesting.25-28 Hu et al25 make a comprehensive survey to such prob-
lems. Different from the inductive coupling and magnetic resonance-based wireless charging techniques which restrict
the freedom of movement, radio frequency (RF) signals are capable of supplying energy over distances, which are gradu-
ally inclining closer to the ultimate goal-charging anytime and anywhere. However, the RF signal-based wireless energy
transfer (WET) may impair the wireless information transfer (WIT) operating in the same spectrum, which results in
many emerging research directions. Hence, it is crucial to coordinate and balance WET and WIT for simultaneous wire-
less information and power transfer, which evolves to the integrated data and energy communication networks (IDENs).
After coming up with a new power supply for the battery-free devices, the paper27 designs an energy conservation scheme
for the nodes to save energy. Considering contemporary wireless sensor network (WSN) adopts IEEE 802.15.4 standard
in its medium access control (MAC) layer and physical (PHY) layer, the 802.15.4 based battery-free (BF)-WSN is modeled
using Markov chain.
In addition, it has been more and more popular for a mobile device to have both WiFi and other low-power wireless
interfaces such as Bluetooth and ZigBee, Qin and Zhang29 propose a ZigBee-assisted power saving management (ZPSM)
scheme, leveraging the ZigBee interface to wake up WiFi interface on demand to improve energy efficiency without
violating delay requirements.
It is also an important perspective to consider the optimization of network energy saving from the nodes’ data-
sampling process.30-32 The sampling data often have time and space correlation, reducing the frequency of data sampling
can effectively reduce the network energy consumption. Ren et al30 investigate the network utility maximization problem
in energy harvesting cognitive radio sensor networks (CRSNs). A joint channel access and sampling rate control scheme,
named JASC, is proposed, aiming at maximizing the network utility by jointly controlling the sampling rates and chan-
nel access of sensor nodes, under the energy consumption, channel capacity, and interference constraints. In wireless
body sensor networks (WBSNs), biosensors collect periodically physiological measures and send them to the coordinator
where the data fusion process takes place.32 Processing the huge amount of data captured by the limited lifetime biosen-
sors and taking the right decisions when there is an emergency are major challenges. Based on that, a biosensor data
management framework composed of an adaptive data collection approach on the biosensor node level and a data fusion
model on the coordinator level is introduced.
Apart from the network itself, many researchers focus their attention to the increasingly enhanced cloud and edge. The
emergence of multiaccess edge computing (MEC) technology aims at extending cloud computing capabilities to the edge
of the radio access network, hence providing real-time, high-bandwidth, low-latency access to radio network resources.33
MEC is able to provide cloud platform and gateway services at the network edge and it inspires the development of myriads
of applications and services with demand for ultralow latency and high quality of service due to its dense geographical
distribution and wide support for mobility. Taking full advantages of cloud radio access network (C-RAN) which is a
potential candidate of the next generation access network technology to address the increasing mobile traffic and mobile
cloud computing (MCC) which offers a prospective solution to the resourcelimited mobile user in executing computation
intensive tasks, Wang et al34 present C-RAN with MCC to enhance both performance and energy efficiencies. In particular,
the joint energy minimization and resource allocation in C-RAN with MCC under the time constraints of the given tasks
are studied.
There are many strategies and algorithms studied separately on the end nodes and the edge server. However,
the researchers seldom consider the two sides’ both benefits. This article proposes an integrated solution to take the
advantages of both.
4 of 21 WANG et al.

3 OV ERALL SYSTEM A RCHITECTURE

As described in Figure 1, based on the actual network deployment in the home or industrial environment, various IoT
devices are connected to the cloud through the gateways. Accordingly, users or administrators can interact with local
devices by accessing the cloud. In addition, an edge server is placed in the system in this article to process data uploaded
from the underlying network and only the processing results will be sent to the cloud. Consequently, both the energy
consumption and the amount of data uploaded to the cloud are significantly reduced, and the response speed of the system
is greatly improved meanwhile. The edge server is at the same level as the gateway and the two parts together constitute
the gateway edge server (GES).

3.1 IoT network

The proposed algorithm at end nodes of this article is based on the ZigBee-enabled low-power IoT network. There are
three types of devices in a ZigBee network: the ZigBee coordinator (ZC), the ZigBee router (ZR),and the ZigBee end device
(ZED). The specification supports at most one coordinator, multiple routers, and multiple end devices within a single
network.
As shown in Figure 1, dependent on the various application requirements and power consumption levels, the end
nodes can be individually designed as lamps, fire detection sensors, temperature sensors, sockets and curtains, and so on.
Generally, the end nodes are designed based on ZigBee end devices which have no specific responsibility for maintaining
the network infrastructure, consume relatively much less energy, and are commonly battery-powered. The end nodes
of the IoT network commonly cannot access to the cloud directly. Then the gateway plays a vital role as a bridge to connect
the local network to the cloud.

3.2 Gateway edge server

Putting all the computing tasks on the cloud has been proved to be an efficient way for data processing since the com-
puting power on the cloud outclasses the capability of the end nodes.35 However, compared with the fast developing data

FIGURE 1 Overall system architecture

WANG et al. 5 of 21

processing speed, the bandwidth of the network has come to a standstill. With the growing quantity of data generated at
the edge, speed of data transportation is becoming the bottleneck for the cloud-based computing paradigm.
Presently, data are increasingly produced at the edge of the network. Therefore, it would be more efficient to also
process the data at the edge of the network and that is the work the edge server does in the overall system. In other words,
it is called edge computing.36 Accordingly, apart from the work at the end nodes, this article proposes a dynamic sampling
rate adjustment algorithm at the edge server in order to take advantage of the strong processing and computing capability
of it. As a result, the energy consumption of the IoT devices can be further reduced.

4 ENERGY EFFICIENT AU TO MATIC MODE SWITCHING ALGORITHM AT

END NODES

The standard ZigBee protocol does not define the working mode of ZigBee devices. As for different devices with different
application requirements in a ZigBee network, their uplink communication modes are similar but the downlink receiving
window is opened differently. Thus, they have different downlink reliabilities and energy consumption levels. Aiming to
enhance the downlink reliability and optimize the energy consumption, three kinds of working modes of end devices are
defined in this article. Furthermore, an end node based mode switching algorithm is proposed to specify how to select and
switch to the proper working mode under different conditions by building the energy consumption model and analyzing
the changeable battery level and downlink traffic. Finally, the performance of the mode switching algorithm is verified
in Section 6.

4.1 Mode definition

Low power consumption is one of the most significant features of ZigBee which is mainly reflected in the battery-powered
ZigBee end devices. In order to adapt to diverse application environments, in the latest ZigBee 3.0 specification, the Zig-
Bee end device has been given various characteristics which mainly aim to achieve lower energy consumption and better
communication performance. Considering the sleeping modes, the receiving window, and so on, based on the standard
features of ZigBee End Devices, three kinds of working modes of end devices are defined. For the convenience of descrip-
tion, the three kinds of working modes are named as Mode A, Mode B, and Mode C. The working schedule of the three
kinds of working modes are illustrated in Figure 2.

• Mode A: An end device of Mode A is almost in sleep all the time after it joins the network and the parent device is
responsible for buffering the messages for its sleeping children. Mode A is of two-way communication. But for the sake
of saving energy, the end device of Mode A will not try to receive the downlink message until it launches an uplink
transmission. Only when the trigger like an external interrupt or a preset internal timer trigger happens will the end
device launch an uplink transmission. Right after the transmission, the end device will open the receiving window
for data RX (receiving) if there are some messages buffered by the parent node. After that, the end device enters sleep
again.
• Mode B: The end device of Mode B sleeps most of the time too. Differently, once it joins the network, it will periodically
send a data request to its parent node and then open the receiving window to retrieve the messages buffered by the

(A) (B)

FIGURE 2 Working schedule of Mode

A/B/C (C)
6 of 21 WANG et al.

parent. So the receiving event is irrelevant to the uplink communication for end devices of Mode B. Certainly, no data
are sent to the sleeping end device until the end device requests for it.
• Mode C: The end device of Mode C does not route messages for other devices, but its radio remains on all the time to
monitor the wireless data packet except when it is under a uplink transmission. Therefore, it can always obtain the
messages sent to them in time.

4.2 Problem statement

As far as the three kinds of working modes of end devices defined above is concerned, their uplink communication are
identical, but when it comes to downlink communication, it takes the end device of Mode A the least time to open the
receiving window. Thus, it has the lowest energy consumption but also the lowest downlink reliability. The end device of
Mode C spends the most time opening the receiving window. Accordingly, it has the highest downlink reliability along
with the highest energy consumption. For Mode B, all those aspects of it are between Mode A and Mode C.
In the scenarios of Internet of Things with a large amount of downlink data and high requirements for communication
performance, the end devices need to extend the opening time of receiving window, that is, switch to Mode B even Mode
C to improve the reliability of downlink communication. For the purpose of saving energy, once the downlink traffic
decreases, the end device should reduce the opening time for the receiving window and switch back to Mode A.
It can be seen from Figure 3 that from Mode A to Mode C, the downlink communication reliability is gradually
improved along with the gradually extended receiving window duration. However, the energy consumption is also
increased at the same time. The changeable battery level and downlink traffic are taken into consideration to realize the
switching algorithm among Mode A, Mode B, and Mode C as depicted in Figure 3.

4.3 Energy consumption model

As a typical low-cost and low-power wireless communication technology, energy consumption of end devices is a sig-
nificant index to measure the performance of ZigBee network. This subsection uses energy consumption model to
analyze network energy consumption, including establishing the energy consumption model of a single battery-powered
end device and the overall battery-powered end devices in the network (Figure 4). As illustrated in Figure 4, energy
consumption models for different working modes are established in this sub-section to analyze the energy consumption.

4.3.1 Mode A

The energy consumption model of the end device of Mode A in one transmission cycle is shown above. The total energy
consumption of one transmission cycle of Mode A is defined as ETA . According to the description above, ETA is:

ETA = Ptx Ttx + Prx Trx + Psleep (TA − Ttx − Trx ). (1)

FIGURE 3 Mode switching process

WANG et al. 7 of 21

FIGURE 4 Energy consumption model of

Mode A/B/C

(A)

(B)

(C)

4.3.2 Mode B

The energy consumption model of the end device of Mode B in one transmission cycle is shown above. The total energy
consumption of one transmission cycle of Mode B is defined as ETB . According to the description above, ETB is:

ETB = Ptx (Ttx1 + Ttx2 ) + Prx Trx + Psleep (TB − Ttx1 − Ttx2 − Trx ). (2)

4.3.3 Mode C

The energy consumption model of the end device of Mode C in one transmission cycle is shown above. The total energy
consumption of one transmission cycle of Mode C is defined as ETC . According to the description above, ETC is:

ETC = Ptx Ttx + Prx (TC − Ttx ). (3)

4.4 Mode switching algorithm

In this article, based on the different power ranges, the battery power level is defined as BL1 (Battery Level 1), BL2 (Battery
Level 2), and BL3 (Battery Level 3). The battery power capacity is marked with 0, BLL (Battery Level Low), BLH (Battery
Level High), and 100%, from low to high. When the power level is lower than BLL, the power level is defined as BL1.
8 of 21 WANG et al.

When the power level is higher than BLL but lower than BLH, the power level is defined as BL2. When the power level
is higher than BLH, the power level is defined as BL3.
Furthermore, based on the number of downlink packets received in cycle T, the downlink traffic level is defined as
DTL1 (Downlink Traffic Level 1), DTL2 (Downlink Traffic Level 2), and DTL3 (Downlink Traffic Level 3). When no packet
is received in cycle T, the downlink traffic level is defined as DTL1. When the packets received in cycle T is more than
zero but less than DTLT (Data Traffic Level Threshold), the downlink traffic level is defined as DTL2. When the packets
received in cycle T is more than DTLT, the downlink traffic level is defined as DTL3.
This subsection describes the specific implementation process of the mode switching algorithm. According to the
change of the battery level and the downlink traffic of the end devices, the automatic mode switching algorithm described
in Figure 3 is designed and the pseudocode is described in Algorithm 1.

Algorithm 1. Automatic mode switching algorithm

Input: Device’s initial mode DevInitMode; Device’s current mode DevCurrentMode; Battery level BLevel; Downlink traffic
level DLTLevel
Output: None
1: DevCurrentMode ← DevInitMode
2: DLTLevel ← DTL1
3: loop
4: Get BLevel
5: if It has been a period of T then
6: Get DLTLevel
7: end if
8: if DevCurrentMode is Mode A then
9: if BLevel is BL2 and DLTLevel ≥ DTL2 then
10: DevCurrentMode ← Mode B
11: else if BLevel is BL3 and DLTLevel ≥ DTL2 then
12: DevCurrentMode ← Mode C
13: end if
14: else if DevCurrentMode is Mode B then
15: if BLevel is BL3 and DLTLevel ≥ DTL2 then
16: DevCurrentMode ← Mode C
17: else if BLevel is BL1 or DLTLevel is DTL1 then
18: DevCurrentMode ← Mode A
19: end if
20: else if DevCurrentMode is Mode C then
21: if BLevel is BL2 and DLTLevel ≥ DTL2 then
22: DevCurrentMode ← Mode B
23: else if BLevel is BL1 or DLTLevel is DTL1 then
24: DevCurrentMode ← Mode A
25: end if
26: end if
27: Delay for tdelay
28: end loop

When the end device is in Mode A, if it detects that its battery level is BL2, it will switch to Mode B as long as it detects
that the downlink traffic level is DTL2 or DTL3 at the same time so that it can achieve better communication performance.
If it detects that its battery level is BL3, it will switch to Mode C as long as it also detects that the downlink traffic level is
DTL2 or DTL3 so it can have even better downlink reliability than end devices of Mode B. In other cases, the end device
remains in Mode A.
When the end device is in Mode B, if it detects that its battery level is BL1, or it detects that its received downlink traffic
level is DTL1, it will switch to Mode A. Consequently, it can decrease the energy consumption. If it detects that its battery
level is BL3, it will switch to Mode C as long as it detects that the downlink traffic level is DTL2 or DTL3. As a result,
WANG et al. 9 of 21

the energy consumption is increased while the downlink latency is reduced. In other cases, the end device remains in
Mode B.
When the end device is in Mode C, it has highest downlink reliability but also the highest energy consumption. If it
detects that its battery level is BL1, or it detects that its received downlink traffic level is DTL1, it will switch to Mode A. If
it detects that its battery level is BL2, it will switch to Mode B as long as it detects that the downlink traffic level is DTL2
or DTL3. In other cases, the end device remains in Mode C.

5 DY NA MIC SA MPLING RATE ADJUSTMENT ALGORITHM AT EDGE

SERV ER

5.1 Problem statement

An IoT device is generally composed of battery, sensor (including ADC), RF transmission module, and processor (includ-
ing microcontroller and memory). For ZigBee devices, the battery life is one of the most important challenges to provide
network services. Each component of the device has an impact on its power consumption. Energy of the battery is mainly
consumed by the sensor unit, the processor unit, and the RF unit. Among them, the wireless communication consumes
more energy than data collection and computation. Some researches ignore the energy consumption of the data collec-
tion, but for different types of sensor, the energy consumption differs a lot. Because the time of data processing is the least,
the energy consumption is the lowest. If the data processing lasts for too much time, the energy consumption will be far
greater than that of data collection and wireless communication.
The workflow of an IoT node can be divided into three parts, as shown in Figure 5, which are the sampling process,
data transmission process, and the process of keeping the device dormant when there is no task. In the sleep mode, the
ZigBee device suspends the power supply of most components to minimize the energy consumption and the watchdog
timer of MCU performs the role of waking up the device regularly. In the process of sampling and data transmission, all
components of the device will be awakened due to many computing tasks, and the energy consumption is far greater than
that in the sleep mode. In addition, for different applications, the energy consumption of the sampling process and data
transmission process will also vary greatly. If the energy consumption is simply considered from the data transmission
level and the energy consumption brought by the device’s sensor data collection and data transmission is ignored, the
energy-saving effect will be far from good. Therefore, for the energy-efficient strategy of ZigBee network, this article
studies on how to reduce the frequency of the sampling and data transmission so as to extend the sleep time of the device
and reduce the energy consumption.
This article designs an energy-efficient scheduling strategy for ZigBee network based on the sampling data character-
istics. As can be seen from the energy consumption comparison at each stage of ZigBee devices in Figure 6, the sampling
and communication processes of devices are the most energy consuming, accounting for more than 90% of the total energy
consumption, while the energy consumption in the dormant mode is much less than that of other stages of devices. There-
fore, the energy-efficient strategy should enable the device to switch between the active mode and the dormant mode to
extend the dormant time of the device as long as possible. It is necessary to reduce the frequency of data sampling and
transmission while ensuring the data accuracy. Consequently, the energy consumption is reduced, which is the main
purpose of the adaptive sampling strategy.

FIGURE 5 Workflow of an IoT node

10 of 21 WANG et al.

F I G U R E 6 Energy
consumption of different
stages

5.2 Design of adaptive sampling scheduling algorithm

This section will explain the design of the adaptive sampling scheduling algorithm in detail and explain how the above
theoretical concepts are applied to the algorithm. The algorithm is divided into two phases, as shown in Figure 7.
In the first stage, the sampling step size (the interval between two consecutive sampling data) is determined based
on the error between the measured data and the predicted data. The next sampling moment is predicted with the k-step
quadratic exponential smoothing algorithm so as to determine the sampling step size to be used at the next time. If the
prediction error is lower than the threshold, the sampling step size should be increased. Consequently, tsleep is extended
and the energy consumption is reduced. In the second stage, the feedback mechanism of EWMA-based emergency
detection is implemented. If an emergency is detected, the sampling step is reduced so that the devices can sample
at a smaller interval. Then the environmental changes can be captured in time and the accuracy of the sampling can
be ensured.

FIGURE 7 Framework of the

algorithm
WANG et al. 11 of 21

FIGURE 8 Flowchart of adaptive sampling

5.2.1 The first stage

In the first stage, f e is defined as the error between the predicted value of the prediction model and the current measured
value. According to this error, the devices will determine the step size kt of the next sampling moment at every current
sampling moment. The initial value of kt is set to 1, and the maximum value of kt is set to kMAX . The maximum value of
step size is preset by the user according to the data accuracy requirements. At each sampling time, the devices compare
f error with the predefined prediction error tolerance 𝛿. If it is lower than the prediction error tolerance, the step size will
be increased. Each increment is one step greater than the previous sampling point, that is, increasing the sleep time of
the device. It will be repeated until kt equals to kMAX . When the maximum sampling interval of kMAX is reached, kt stops
increasing. This process is similar to the slow start process of TCP congestion control. The sampling time tsample and
data transmission time ttransmit are reduced in this process, and the sampling interval is prolonged. The sleep time tsleep is
increased, and thus the energy consumption is greatly reduced.
Once f error is greater than the prediction error tolerance 𝛿, kt will be reduced to half of the current sampling interval
until f error is less than 𝛿 or kt is reduced to the initial value. At the same time, the increment of step size starts from 1.
This stage is similar to congestion avoidance algorithm, and the dynamic adjustment of sampling interval is depicted in
Figure 8.

5.2.2 The second stage

In the second stage, when the sampling step reaches the maximum step size, that is, the device has the maximum sleep
time. If the prediction error remains below the prediction error tolerance, the sampling step size will remain at the maxi-
mum sampling step size kMAX . But at the same time, the emergency detection based feedback mechanism will be triggered.
This detection is essential to minimize the incidence of missed detection. Missed detection refers to that when the sam-
pling environment changes suddenly, the sampling algorithm runs at the maximum step size, which may miss some
changes of the sampling environment and fail to report that to the coordinator in time. To some extent, the missing rate
reflects the accuracy of the sampling algorithm, which is one of the indexes to evaluate the performance of the algorithm.
The emergency detection feedback mechanism is based on the exponentially weighted moving average algorithm
(EWMA), which can judge whether the sampling environment changes suddenly, so as to reduce the missing rate.
First of all, two different smoothing coefficients 𝛼long and 𝛼short are defined. Then the short-term index weighted mov-
ing average Sshort and the long-term index weighted moving average Slong of the sampling sequence are calculated. For
12 of 21 WANG et al.

FIGURE 9 Flowchart of feedback based

on emergency detection

each sampling data, the ratio of the two moving average values which is defined as 𝜂 is calculated. If the ratio exceeds
the threshold 𝜏, it means that there is a sudden change in the sampling environment. The sampling step kt should be
immediately adjusted to the initial value of 1, and the increment starts from 1 again.

Sshort
𝜂= . (4)
Slong

This stage mainly works on making the device quickly recover the high-frequency and high-precision sampling when
the environment suddenly changes. Although the energy consumption increases, the missing rate is reduced, so as to
achieve the purpose of self-adaptation. Figure 9 shows the flow of event detection algorithm.

5.3 Procedure of adaptive sampling scheduling algorithm

Algorithm 2 shows the process of the device adaptive sampling algorithm. In the process of data collection, the improved
quadratic smoothing algorithm is used to predict the data of future sampling points. If the prediction error f error is less
than the error tolerance 𝛿, the step size will increase. The sleep duration will increase, avoiding unnecessary sampling
and saving the battery energy.
On the other hand, if the prediction error exceeds the error tolerance or the environment changes dynamically, the
step size will be reduced and the sampling will be resumed at a shorter step. Therefore, it can be seen that the device
adaptive sampling algorithm proposed in this article is easy to operate and can be easily implemented on ZigBee devices
with limited resources. This algorithm with proper parameters can not only meet the needs of energy saving but also
ensure the accuracy of data sampling.
WANG et al. 13 of 21

Algorithm 2. Adaptive sampling scheduling algorithm

Input: St time series array
Output: The Sampling step size kt after dynamic adjustment
1: initialize Variables kt = 1, KMAX , 𝛿
2: initialize DES parameters 𝛼, 𝛽
3: initialize EWMA parameters 𝛼short , 𝛼long
4: initialize St ← yt , bt ← yt+1 − yt , Vt ← 1 − power((1 − 𝛼), k), Ut ← 1 − power((1 − 𝛽), k), t ← 1
5: generate data prediction module
6: if 0>𝛼>1 and 0>𝛽>1 then
7: while kt > total data set do
8: y(t) = Sense data
9: DES predictionS(t) = yt+1 = St + kt bt
10: evaluate forecast accuracy error
11: if S(t) − y(t)>𝛿 then
12: if kt >kMAX then
13: kt increment
14: end if
15: stay at KMAX and check for event detection
16: end if
17: Sampling window adjustment
18: kt drop to half of the current interval
19: Event Detection EWMA
20: Stshort = 𝛼short yt−1 + (1 − 𝛼short )St−1
21: Stlong = 𝛼long yt−1 + (1 − 𝛼long )St−1
22: 𝜂 = Sshort ∕Slong
23: if 𝜂<𝜏 then
24: Sampling window adjustment
25: kt ← 1
26: Count missed false
27: reset Sshort and Slong
28: Sense data
29: end if
30: end while
31: end if

6 PERFORMANCE EVA LUAT ION AND ANALYSIS

In this section, two types of devices are designed to carry out this part of simulation. They are the ZigBee gateway which
is capable of date receiving, processing, and relaying and the ZigBee end device which is able to collect the environment
temperature and humidity.

6.1 Evaluation of mode switching algorithm

The simulation in this section is mainly based on the star network. As described above, the energy consumption of a
single device or multiple devices in a whole network is related to the mode selection, the number of end devices, and the
parameters set. It is assumed that T A , T B , and T C are the same and defined as T.
From Figure 10A, we can see that the device working in Mode A has the lowest energy consumption level with the
same transmission frequency. Apart from energy consumption, the packet loss rate is also a vital performance index and
it is evaluated in this subsection. The simulation result is depicted in Figure 10B. As seen in the figure, it is obvious that
14 of 21 WANG et al.

F I G U R E 10 Energy consumption and downlink

packet loss rate of different modes

the device working in Mode A has the highest packet loss rate when the same number of downlink packets are sent while
the device in Mode C has the lowest packet loss rate.
The energy consumption and packet loss rate of the whole network is also evaluated and depicted separately in
Figure 11A,B. We can see that in a whole network, the mode switching algorithm does save more energy than usual
network without mode switching and the packet loss rate is much improved compared with the situation where all the
devices work in Mode A.

6.2 Evaluation of the overall proposed algorithm

For the adaptive sampling scheduling algorithm, this article uses different parameters to evaluate the energy-saving effect
and comprehensive performance of the algorithm. The energy-saving effect of the algorithm can be evaluated by using
the sampling ratio of ZigBee devices. The accuracy can be judged by the rate of missing detection. Combined with these
two factors, the performance of the proposed algorithm can be evaluated comprehensively.
WANG et al. 15 of 21

F I G U R E 11 Energy consumption and downlink

packet loss rate of the whole network

The verification of the proposed algorithm in this article is implemented with MATLAB, and the algorithm perfor-
mance is verified with the open-source data set from Intel Berkeley Laboratory. The temperature and humidity data
collected from device 21 in 36 hours on 1 March 2004 is selected as the verification data set of this scheme. The temperature
and humidity data have obvious changes during the 36 hours.

6.2.1 E-Sense

E-Sense is a framework that utilizes knowledge of the underlying data streams as well as application data quality require-
ments to conserve energy of sensor nodes. E-Sense employs a stochastic scheduling algorithm to dynamically control the
operating modes of the sensor node components. This scheduling algorithm enables an adaptive sampling strategy that
aggressively conserves power by adjusting sensing activity to the application requirements.37
In order to display the energy-saving characteristics of the proposed algorithm more intuitively, in addition to com-
paring the energy consumption of the adaptive sampling double exponential smoothing (AS-DES) algorithm with the
16 of 21 WANG et al.

conventional sampling scheme, this article also compares the energy consumption of AS-DES with the scheduling scheme
E-Sense based on the model dynamic sampling and evaluates the advantages of AS-DES.

6.2.2 Energy consumption evaluation

The algorithm described above shows that the sampling step of ZigBee devices has a direct impact on the energy con-
sumption of devices. The larger the sampling step is, the larger the sampling interval is. On the one hand, the sampling
times will decrease. On the other hand, the sleep duration will increase, and the energy consumption of devices will also
decrease. According to the analysis of device energy consumption above, the data sampling and transmission process of
devices consumes nearly 90% energy. Therefore, the energy consumption of different schemes can be visually compared
from the sample size of devices. The energy consumption for sampling once is defined as esample , and the number of sam-
pling is defined as n. The smaller the sampling ratio is, the fewer the sampling points are, and the corresponding sampling
algorithm can save more energy. The energy conversion formula can be described as follows:

Esample = n ⋅ esample . (5)

Figure 12A,B shows the energy consumption comparison between different sampling schemes based on temper-
ature and humidity data sets, corresponding to the fixed frequency scheme, namely, the common sampling scheme,
the model-based dynamic sampling scheme, and the adaptive sampling schedules. The device with the fixed-frequency
scheme collects data at the frequency of once per second. The AS-DES scheme has a maximum step size of 30 seconds,
and E-Sense has a maximum delay of 30 seconds. For conventional sampling schemes, a single device sends 129 600 sam-
pling data to the coordinator within 36 hours. If the device collects the ambient humidity, it will consume 7.344 mA⋅h
of energy. The device with an AAA battery of 1300 mA⋅h can be used for 170 days plus the energy consumption in the
dormant state. Compared with conventional sampling, AS-DES scheme and E-Sense scheme can extend the duration of
dormancy state and reduce the number of sampling times due to real-time adjustment of sampling intervals. In addi-
tion AS-DES and E-Sense consume significantly less energy than conventional sampling and the battery life of ZigBee
devices is further extended. The scheme proposed in this article (AS-DES) consumes only 30%-40% of the energy of the
conventional sampling scheme.
The energy consumption ratio of AS-DES and E-Sense is further compared. The energy consumption ratio 𝜅 is defined
as the ratio of the energy consumed by the energy saving sampling algorithm EAS − DES to the energy consumed by the fixed
sampling frequency Edefault . That is, the ratio of the total number of data samples nAS − DES obtained by the energy-saving
sampling algorithm to the total number of data points available in the real data set ndefault .

EAS−DES n
𝜅= = AS−DES . (6)
Edefault ndefault

Through the simulation validation of the above schemes, we can see AS-DES proposed in this article has obvious
energy efficiency. Compared with the sampling scheme with fixed sampling interval (1 second), the sampling scheme
with a maximum step of 30 seconds can save about 70% of the battery energy.

6.2.3 Comprehensive evaluation

In order to evaluate the energy saving characteristics of the algorithm proposed in this article more comprehensively, this
subsection will evaluate the energy saving performance of the scheme based on different sampling control windows. The
larger the sampling interval is, the longer the sleep time is and the lower the corresponding energy consumption is. In
the AS-DES scheme proposed in this article, the sampling interval is controlled by an adaptive sampling control window
whose maximum value is also the maximum step kMAX . The energy consumption of devices is different for different
window sizes. Meanwhile, the missed detection of emergencies cannot be ignored. Increasing the sampling interval but
ignoring emergencies in the sampling environment can affect the service quality of the overall network. Therefore, it
is necessary to evaluate the comprehensive performance of the proposed scheme by combining the energy-saving effect
under different sampling windows and the success rate of emergency capture (ie, the rate of missed detection).
WANG et al. 17 of 21

F I G U R E 12 Comparison of energy consumption

If the monitoring environment changes but the device does not capture the change in time, it means missed detection.
The rate of missed detection is defined as 𝜃. The total number of sampling points is defined as n. The number of undetected
events is defined as nf . The missed detection rate can be used to judge the response ability of the algorithm to emergencies
on the premise of energy saving, so as to determine whether the algorithm can meet the low-power requirements and
provide accurate data sampling services at the same time.

nf
𝜃= . (7)
n

In different sampling control windows, that is, the maximum sampling interval, the energy consumption, and missed
detection rate of different sampling schemes are displayed, which can not only compare the comprehensive performance
of different sampling schemes, such as energy consumption and accuracy, but also select reasonable parameters and
appropriate energy-saving sampling schemes for different needs. Based on the data set of the sampling temperature, the
error tolerance of AS-DES is set to 0.01 and that of E-Sense is set to 0.1. The maximum step kMAX and maximum delay
MaxDelay change from 10 to 60 seconds. The device’s temperature and humidity data of 1 hour is used for self-training
for E-Sense. The energy-saving performance of fixed-interval sampling scheme, E-Sense, and adaptive sampling scheme
on temperature data sets are respectively shown in Figure 13A-C.
18 of 21 WANG et al.

F I G U R E 13 Comparison of energy consumption

WANG et al. 19 of 21

As seen from Figure 13A, the fixed sampling interval sampling scheme is a common sampling scheme, which con-
sumes more energy than the other two sampling schemes. This is because the device maintains a fixed sampling interval,
which is the interval of the worst case (ie, when the sampling environment changes dramatically). It ensures that the
device can report the sampling data in time in all cases. However, it sacrifices the dormancy time of the devices. Some
situations do not require such high-frequency sampling, which results in a large amount of energy waste.
From Figure 13B, the energy consumption and missed detection rate of the E-Sense scheme remain basically sta-
ble after a maximum delay of 20 seconds, which indicates that the energy consumption of E-Sense cannot be further
reduced after that. Comparing Figure 13B,C, when the sampling control window is smaller than about 30 seconds, the
comprehensive performance of AS-DES scheme is better than that of E-Sense scheme.
For the sampling scheme (AS-DES) proposed in this article, the energy consumption decreases along with the increase
of the sampling control window kMAX , which indicates that the sampling size is decreasing and the dormancy time is
increasing. However, the missing rate of sampling is also increasing at the same time, because the larger the sampling
interval is, the greater the probability of missed detection will be, as shown in Figure 13C. Compared with the conventional
sampling scheme, the energy-saving effect of the proposed scheme in this article is quite obvious. In most cases, the steady
change of the sampling environment does not require too much sampling. Therefore, the adaptive sampling scheduling
scheme is proposed to effectively avoid redundant sampling and save the battery energy.
Through the above simulation experiments, the proposed AS-DES sampling scheme can effectively reduce the sam-
pling frequency of devices, increase the dormancy time, and effectively reduce the energy consumption by 50%-70%
compared with the conventional sampling scheme. Under the appropriate sampling control window, the missed detection
rate can be controlled at about 10% and the energy-saving performance is better than the model-based E-Sense sampling
scheme. AS-DES with recursive prediction equation can adjust the sampling step adaptively according to the trend in
time series without adding any model-training cost. Such lightweight computation does not cause too much computing
burden on devices. Therefore, applying the adaptive sampling scheduling scheme proposed in this article to devices in
ZigBee network can effectively reduce the data collection and transmission, reduce the energy consumption, and prolong
the battery life.

7 CO N C LU S I O N

This article proposes an integrated solution to take the advantages of both the end node and the edge server. From the
microlevel, an energy efficient automatic mode switching algorithm is proposed and run on the end node. It enables the
IoT devices dynamically select and switch among the defined working modes according to the changeable battery level
and downlink communication traffic. From the macrolevel, an dynamic sampling rate adjustment algorithm is proposed
and run on the edge server so that the IoT devices can dynamically adjust their data-sampling rate to minimize the
energy consumption. The simulation results have shown the effectiveness and the efficiency of both the node and edge
algorithms.
Apart from the further effort on energy saving, wireless power transmission will be introduced in our future work to
add a new perspective to the conventional power supply. On top of that, the resource allocation problem will be further
studied under the time and energy constraints.

ACKNOWLEDGMENT
The work in this article was partly supported by Natural Science Foundation of China (Grant No. 61620106011) and
Zhongshan City Team Project (Grant No. 180809162197874).

ORCID
Haibo Mei https://orcid.org/0000-0001-8093-7175

REFERENCES
1. Rao PS, Jana PK, Banka H. A particle swarm optimization based energy efficient cluster head selection algorithm for wireless sensor
networks. Wirel Netw. 2017;23(7):2005-2020.
2. Guo J, Song Z, Cui Y, Liu Z, Ji Y. Energy-efficient resource allocation for multi-user mobile edge computing. Paper presented at:
Proceedings of the GLOBECOM 2017-2017 IEEE Global Communications Conference; 2017:1-7; Singapore, Singapore: IEEE.
3. Yang L, Zhang H, Li M, Guo J, Ji H. Mobile edge computing empowered energy efficient task offloading in 5g. IEEE Trans Veh Technol.
2018;67(7):6398-6409.
20 of 21 WANG et al.

4. Alassery F, Ahmed WKM, Sarraf M, Lawrence V. A novel scheme for power saving in wireless sensor networks with packet collision.
Paper presented at: Proceedings of the WAMICON 2014; 2014:1-4; Tampa, FL: IEEE.
5. Jianwu Z, Jiji Z, Mingliang Y. Analysis and study of the relationship of energy consumption and CPU buffer of a wireless sensor node.
Paper presented at: Proceedings of the 2007 International Symposium on Microwave, Antenna, Propagation and EMC Technologies for
Wireless Communications; 2007:98-101; Hangzhou, China: IEEE.
6. Anchora L, Capone A, Mighali V, Patrono L, Simone F. A novel mac scheduler to minimize the energy consumption in a wireless sensor
network. Ad Hoc Netw. 2014;16:88-104.
7. Zhang Y, Zheng S, Xiong S. A scheduling algorithm for tdma-based mac protocol in wireless sensor networks. Paper presented at: Pro-
ceedings of the 2009 1st International Workshop on Education Technology and Computer Science; vol 3, 2009:148-151; Wuhan, Hubei,
China: IEEE.
8. Abderrahim M, Hakim H, Boujemaa H, Touati F. Energy-efficient transmission technique based on dijkstra algorithm for decreasing
energy consumption in wsns. Paper presented at: Proceedings of the 2019 19th International Conference on Sciences and Techniques of
Automatic Control and Computer Engineering (STA); 2019:599-604; Sousse, Tunisia, Tunisia: IEEE.
9. Samani M. A new routing algorithm to reduce consumed energy in wireless sensor. Paper presented at: Proceedings of the 2011 Third
International Conference on Computational Intelligence, Communication Systems and Networks; 2011:95-98; Bali, Indonesia: IEEE.
10. Xie M, TL Huang, XS Zhu. A novel routing algorithm for energy-efficient in wireless sensor networks. Paper presented at: Proceedings of
the 2009 Third International Conference on Genetic and Evolutionary Computing; 2009:65-68; Guilin, China: IEEE.
11. Alwakeel S, Prasetijo A, Alnabhan N. An adaptive energy-saving routing algorithm for mobile wireless sensor networks. Paper presented
at: Proceedings of the 2015 International Conference on Electrical and Information Technologies (ICEIT); 2015:104-108; Marrakech,
Morocco: IEEE.
12. Song XO. An energy-saving geographic routing in wireless sensor network. Paper presented at: Proceedings of the 2015 International
Conference on Intelligent Networking and Collaborative Systems; 2015:390-393; Taipei, Taiwan: IEEE.
13. Pan J, Wang L, Zou Q, Huang W. Efficient energy-saving routing algorithms of wireless sensor network. Paper presented at: Proceedings
of the 2012 8th International Conference on Wireless Communications, Networking and Mobile Computing; 2012:1-4; Shanghai, China:
IEEE.
14. Jing L, Liu F, Li Y. Energy saving routing algorithm based on spin protocol in wsn. Paper presented at: Proceedings of the 2011 International
Conference on Image Analysis and Signal Processing; 2011:416-419; Hubei, China: IEEE.
15. Zeng Y, Wang L. Energy-saving routing protocol for wireless sensor networks. Paper presented at: Proceedings of the The 26th Chinese
Control and Decision Conference (2014 CCDC); 2014:4868-4872; Changsha, China: IEEE.
16. Chan HF, Rudolph H. New energy efficient routing algorithm for wireless sensor network. Paper presented at: Proceedings of the TENCON
2015-2015 IEEE Region 10 Conference; 2015:1-5; Macao, China: IEEE.
17. Zhang M, Gong C, Feng Y, Liu C, Feng H. A novel energy-efficient dynamic voting cluster algorithm for wireless sensor networks. Paper
presented at: Proceedings of the 2009 International Conference on Networks Security, Wireless Communications and Trusted Computing;
vol 2, 2009:631-634; Wuhan, Hubei, China: IEEE.
18. Carmona C, Alorda B, Ribot MA. Energy consumption savings in zigbee-based WSN adjusting power transmission at application layer.
Paper presented at: Proceedings of the 2014 24th International Workshop on Power and Timing Modeling, Optimization and Simulation
(PATMOS); 2014:1-6; Palma de Mallorca, Spain: IEEE.
19. Mirza M Masood Y, Ahmed G, Khan NM. Modified on demand transmission power control for wireless sensor networks. Paper presented
at: Proceedings of the 2011 International Conference on Information and Communication Technologies; 2011:1-6; Karachi, Pakistan:
IEEE.
20. Le TL, Vo MH. Lossless data compression algorithm to save energy in wireless sensor network. Paper presented at: Proceedings of the 2018
4th International Conference on Green Technology and Sustainable Development (GTSD); 2018:597-600; Ho Chi Minh City, Vietnam:
IEEE.
21. Jeong KC, Choi WS, Choi SG. A wireless AP power saving algorithm by changing operating mode and altering transmission power in
ieee 802.11 wlan. Paper presented at: Proceedings of the 2014 International Conference on Information and Communication Technology
Convergence (ICTC); 2014:9-10; Busan, South Korea: IEEE.
22. Nga DTT, Kim MG, Kang M. A delay constraint energy saving algorithm in ieee 802.16 e wireless man. Paper presented at: Proceedings
of the 2007 Wireless Telecommunications Symposium; 2007:1-6; Pomona, CA: IEEE.
23. Nagareda R, Hasegawa A, Shibata T, Obana S. A proposal of power saving scheme for wireless access networks with access point shar-
ing. Paper presented at: Proceedings of the 2012 International Conference on Computing, Networking and Communications (ICNC);
2012:1128-1132; Maui, HI: IEEE.
24. Lee S, Yoo M. Power saving algorithm using data lifetime in wireless sensor network. Paper presented at: Proceedings of the 2008 10th
International Conference on Advanced Communication Technology; vol 2, 2008:1328-1332; Gangwon-Do, South Korea: IEEE.
25. Hu J, Yang K, Wen G, Hanzo L. Integrated data and energy communication network: a comprehensive survey. IEEE Commun Surv Tutor.
2018;20(4):3169-3219.
26. Lee C, Lee J. Harvesting and energy aware adaptive sampling algorithm for guaranteeing self-sustainability in wireless sensor networks.
Paper presented at: Proceedings of the 2017 International Conference on Information Networking (ICOIN); 2017:57-62; Da Nang, Vietnam:
IEEE.
WANG et al. 21 of 21

27. Zhu YH, Lv H, Li Y, Li E, Chi K. Energy conservation scheme for ieee 802.15. 4 based battery-free wireless sensor networks. Paper presented
at: Proceedings of the 2016 International Conference on Networking and Network Applications (NaNA); 2016:342-348; Hakodate, Japan:
IEEE.
28. Wang C, Li J, Yang Y, Ye F. Combining solar energy harvesting with wireless charging for hybrid wireless sensor networks. IEEE Trans
Mob Comput. 2017;17(3):560-576.
29. Qin H, Zhang W. Zigbee-assisted power saving management for mobile devices. IEEE Trans Mob Comput. 2013;13(12):2933-2947.
30. Ren J, Zhang Y, Deng R, Zhang N, Zhang D, Shen XS. Joint channel access and sampling rate control in energy harvesting cognitive radio
sensor networks. IEEE Trans Emerg Top Comput. 2016;7(1):149-161.
31. Goel S, Imielinski T. Prediction-based monitoring in sensor networks: taking lessons from mpeg. ACM SIGCOMM Comput Commun Rev.
2001;31(5):82-98.
32. Habib C, Makhoul A, Darazi R, Salim C. Self-adaptive data collection and fusion for health monitoring based on body sensor networks.
IEEE Trans Ind Inform. 2016;12(6):2342-2352.
33. Porambage P, Okwuibe J, Liyanage M, Ylianttila M, Taleb T. Survey on multi-access edge computing for internet of things realization.
IEEE Commun Surv Tutor. 2018;20(4):2961-2991.
34. Wang K, Yang K, Magurawalage CS. Joint energy minimization and resource allocation in c-ran with mobile cloud. IEEE Trans Cloud
Comput. 2016;6(3):760-770.
35. Kumar P, Kumar R. Issues and challenges of load balancing techniques in cloud computing: a survey. ACM Comput Surv (CSUR).
2019;51(6):1-35.
36. Shi W, Cao J, Zhang Q, Li Y, Xu L. Edge computing: vision and challenges. IEEE IoT J. 2016;3(5):637-646.
37. Liu H, Chandra A, Srivastava J. Esense: energy efficient stochastic sensing framework for wireless sensor platforms. Paper presented at:
Proceedings of the 2006 5th International Conference on Information Processing in Sensor Networks; 2006:235-242; Nashville, TN: IEEE.

How to cite this article: Wang Y, Yang K, Wan W, Mei H. Adaptive energy saving algorithms for Internet of
Things devices integrating end and edge strategies. Trans Emerging Tel Tech. 2020;e4122. https://doi.org/10.1002/
ett.4122