IEEE INTERNET OF THINGS JOURNAL, VOL. 5, NO.

4, AUGUST 2018 2315

Motor Anomaly Detection for Unmanned Aerial

Vehicles Using Reinforcement Learning
Huimin Lu, Yujie Li, Shenglin Mu, Dong Wang, Hyoungseop Kim, and Seiichi Serikawa

Abstract—Unmanned aerial vehicles (UAVs) are used in many of disaster areas. Rescue drones are used to capture live
fields including weather observation, farming, infrastructure footage of a disaster area, provide first aid kits, and pro-
inspection, and monitoring of disaster areas. However, the cur- vide Wi-Fi links in a local area [18]. However, the use of
rently available UAVs are prone to crashing. The goal of this
paper is the development of an anomaly detection system to drones in extreme environments raises a number of issues.
prevent the motor of the drone from operating at abnormal In the USA alone, 418 accidents have been reported. Japan’s
temperatures. In this anomaly detection system, the tempera- Aeronautical Law requires drones to be operated safely when
ture of the motor is recorded using DS18B20 sensors. Then, flying over a densely populated area [17]. To the best of our
using reinforcement learning, the motor is judged to be operat- knowledge, however, no system has yet been devised to detect
ing abnormally by a Raspberry Pi processing unit. A specially
built user interface allows the activity of the Raspberry Pi to anomalies in the operation of the motor of a drone. In this
be tracked on a Tablet for observation purposes. The proposed paper, we develop a detection and protection system that lands
system provides the ability to land a drone when the motor the drone when an abnormal engine temperature is detected.
temperature exceeds an automatically generated threshold. The A common type of drone accident is the uncontrolled
experimental results confirm that the proposed system can safely descent arising from issues, such as insufficient battery capac-
control the drone using information obtained from temperature
sensors attached to the motor. ity, motor malfunctioning or direct current problems, and loss
of communication. To address these issues, an anomaly detec-
Index Terms—Anomaly detection, reinforcement learning, tem- tion system is required. In this paper, we consider the case
perature sensor, unmanned aerial vehicle (UAV).
of motor abnormalities arising in flight. One key factor in the
I. I NTRODUCTION design life of a motor is the life of the bearings. This can be
N FLIGHT, an unmanned aerial vehicle (UAV) is unat- monitored in two ways. The first detects the deterioration of
I tended and controlled remotely. A drone is a type of UAV
that is widely used for weather observation, spraying of agri-
the grease due to heat. The second estimates the mechanical
life by considering rolling fatigue. In the majority of cases,
cultural chemicals, inspection of infrastructure, and monitoring deterioration of the grease has a more significant role than
mechanical failure. Temperature is the key factor influencing
Manuscript received May 7, 2017; revised June 22, 2017 and July 4, 2017; the service life of the bearing grease. We therefore monitor the
accepted August 4, 2017. Date of publication August 8, 2017; date of
current version August 9, 2018. This work was supported in part by temperature of the motor as a means of detecting abnormali-
the Leading Initiative for Excellent Young Researcher of Ministry of ties. This assumes that motor failure will usually be associated
Education, Culture, Sports, Science, and Technology, Japan, under Grant with temperature anomalies.
16809746, in part by the Aid for Scientific Research of JSPS under
Grant 17K14694, in part by the Research Fund of State Key Laboratory The proposed motor anomaly detection and crash preven-
of Marine Geology in Tongji University under Grant MGK1608, in part tion system uses temperature sensors and an ARM Cortex-
by the Research Fund of State Key Laboratory of Ocean Engineering in A53 CPU (Raspberry Pi 2B). The temperature sensors are
Shanghai Jiaotong University under Grant 1510, in part by the Research
Fund of The Telecommunications Advancement Foundation, and in part attached to the motor unit of the drone and abnormalities
by the Fundamental Research Developing Association for Shipbuilding and are detected using the Raspberry Pi located inside the drone.
Offshore and Strengthening Research Support Project of Kyushu Institute of If an abnormal temperature is detected, the drone is landed
Technology. (Corresponding author: Huimin Lu.)
H. Lu is with the Department of Mechanical and Control Engineering, automatically and flight is resumed if the temperature returns
Kyushu Institute of Technology, Kitakyushu 8048550, Japan (e-mail: to normal. This system prevents the drone from crashing in
[email protected]). extreme environments, where failures can be catastrophic. The
Y. Li is with Yangzhou University, Yangzhou 225127, China (e-mail:
[email protected]). automatic motor anomaly detection system increases safety,
S. Mu is with the Graduate School of Science and allowing more active use of automatic piloting, and con-
Engineering, Ehime University, Matsuyama 790-8577, Japan (e-mail: sequently decreasing crashes caused by communication loss
[email protected]).
D. Wang is with the School of Information and Communication and controller error. In this paper, instead of use a static
Engineering, Dalian University of Technology, Dalian 116023, China (e-mail: threshold of temperature for abnormality detection, we use
[email protected]). reinforcement learning for monitoring the temperature at each
H. Kim is with the Department of Control Engineering, Kyushu Institute
of Technology, Kitakyushu 8048550, Japan. dynamic status. Different from supervised learning, rein-
S. Serikawa is with the Center for Socio-Robotic Synthesis forcement learning [15] is a kind of unsupervised learning
and the Department of Electrical and Electronic Engineering, methods. Therrien et al. [16] proposed a reinforcement learn-
Kyushu Institute of Technology, Kitakyushu 8048550, Japan (e-mail:
[email protected]). ing and error-based processes for motor learning. To the
Digital Object Identifier 10.1109/JIOT.2017.2737479 best of our knowledge, the proposed system is the first
2327-4662 c 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

Authorized licensed use limited to: Universidad Federal de Pernambuco. Downloaded on January 26,2023 at 16:53:18 UTC from IEEE Xplore. Restrictions apply.
2316 IEEE INTERNET OF THINGS JOURNAL, VOL. 5, NO. 4, AUGUST 2018

Fig. 2. Thermal imaging system based on abnormal temperature detection.

Fig. 1. Inspection system using abnormal sound detection.

work that solves the motor failure in real time without prior
experience.
The remainder of this paper is organized as follows.
In Section II, we review related work on motor anomaly
detection. Section III introduces the configuration of the
proposed system. Experiments and results are reported in
Section IV. Finally, Section V presents the conclusions. Fig. 3. Outline of the fan sensor system.

II. R ELATED W ORK abnormality or a change in the air speed of the drone. The fan
sensor also influences the flight characteristics of the drone.
Many methods are available for detecting motor anomalies.
As noted above, all conventional anomaly detection meth-
In this section, we summarize these methods and discuss the
ods have drawbacks. This paper addresses the practicality of
advantages and disadvantages of each.
a drone-mounted anomaly detection system using temperature
A. Abnormal Sound Detection sensors (DS18B20) and a Raspberry Pi 2B.
One type of system is based on abnormal sound
detection [1] using case-based identification algorithms. As III. S YSTEM C ONFIGURATION
illustrated in Fig. 1, in this method it is necessary to first The system developed in this paper is designed to land the
calculate the sound data by feature vector and then apply an drone if the motor temperature exceeds a threshold, defined as
establishment distribution model, which is a form of super- representing an abnormality. Previous studies have suggested
vised learning. Supervised learning requires a training dataset that, in the case of a drone in sustained flight, this abnormal
of sample elements and an appropriate classification for each temperature is between 70 ◦ C and 80 ◦ C. As the flight time
of the samples. However, in the majority of cases it is dif- used in this paper was as short as 10 min, 28 ◦ C was adopted
ficult to train the datasets in an extreme environment. False as the initial threshold.
detections therefore arise when using sound-based inspection
systems against a noisy background. A. Overview
B. Infrared Imaging Detection The experimental setup used in this paper comprised two
A second type of system collects infrared images of the devices. The drone is equipped with temperature sensors and
motor and triggers an alarm when the temperature appears a Raspberry Pi 2B CPU; a Tablet is used to collect the obser-
to equal or exceed a predetermined threshold [2], [3] (see vational data. The concept map of the system components is
Fig. 2). A problem with this type of system is that it is diffi- presented in Fig. 4. We define the landing and flying model
cult to mount a compact thermal imaging camera on a drone by the following equation:

with sufficiently accurate infrared imaging to allow changes Landing: t ≥ T̃
in temperature to be determined. (1)
Flying: t < T̃
C. Abnormal Rotation Detection where t is the real-time measured temperature and T̃ is the
A third system uses a sensor to monitor the rotation speed threshold temperature for determining deceleration. Notice
of the fan and triggers an alarm when abnormal rotation that T̃ is not a fixed value, it changes with flight time. In
is detected [4], [5] (see Fig. 3). However, it is difficult to Section III-C, we use reinforcement learning for determining
determine whether a reduction in rotation speed is due to an the threshold temperature.

Authorized licensed use limited to: Universidad Federal de Pernambuco. Downloaded on January 26,2023 at 16:53:18 UTC from IEEE Xplore. Restrictions apply.
LU et al.: MOTOR ANOMALY DETECTION FOR UAVs USING REINFORCEMENT LEARNING 2317

Fig. 4. System working diagram.

Fig. 6. Circuit diagram of the sensors.

Fig. 5. Conceptual diagram of the system.

Fig. 7. Placement of the temperature sensor.

The processing flow of the proposed system is as follows.

• The Raspberry Pi is made operable from the Tablet.
computer for educational use and is less expensive than the
• After takeoff, the drone assumes command from
standard Raspberry Pi. It is the lightest ARM series processor
Raspberry Pi to drone.
and is fully capable of meeting the requirements of light load
• Drone advances after takeoff.
and efficient real-time processing. A special feature is that it
• The temperature of the motor is measured in flight.
can use SD card storage rather than an internal hard drive.
• The drone is landed automatically when the measured
The control unit has two functions.
temperature exceeds the threshold temperature.
• To implement control of the drone by the Raspberry Pi.
• To decide whether to land the drone based on the
B. Device Setting data obtained from the temperature sensor.
1) Drone Composition: The experimental drone comprised 4) Communication Unit: A Wi-Fi connection is used to
four units: 1) a detection unit (the temperature sensors); 2) the allow the Raspberry Pi to communicate with the drone as
drone itself; 3) a control unit; and 4) a communication unit. For well as for the communication between the Tablet and drone
the control unit, we propose a novel reinforcement-learning using VNC [8]. The service set identifier of the drone used in
algorithm in the Raspberry Pi. This algorithm is described in this paper was ardrone2-132506 and the IP address was set as
detail in Section III-C. The communication unit implements 192.168.1.1. The flowchart of the system is displayed in Fig. 8.
virtual network computing (VNC) to enable communication In this system, when the drone takes off, the Raspberry Pi
between the Raspberry Pi and the Tablet. The configuration is automatically measures the temperature. If the real-time mea-
illustrated in Fig. 5. sured temperature t is greater than the threshold temperature
2) Detection Unit: The detection circuit is displayed in T̃, the drone is landed; otherwise, the drone continues to fly,
Fig. 6. The temperature sensor is attached to an exposed monitoring the motor temperature.
area of the motor at the lower side of the motor case,
as indicated in Fig. 7. As the 1-wire temperature sensor
(DS18B20) has a built-in A/D converter, there is no signal C. Threshold Temperature Update by Reinforcement
deterioration due to the wiring length; the temperature data is Learning
converted to digital inside the sensor. In the Raspbian OS, this The algorithm adjusts the rate of the temperature increase
sensor is detected as a 1-wire device simply by loading the based on changes of speed; this is the key for measuring
appropriate module. the threshold temperature. In this paper, the optimal speed
3) Control Unit: A Raspberry Pi model 2B is used in the adjustment to suppress the rise in temperature is derived
control unit for processing. It was developed as a single-board by comparison with the previously measured temperatures.

Authorized licensed use limited to: Universidad Federal de Pernambuco. Downloaded on January 26,2023 at 16:53:18 UTC from IEEE Xplore. Restrictions apply.
2318 IEEE INTERNET OF THINGS JOURNAL, VOL. 5, NO. 4, AUGUST 2018

Fig. 10. Relationship between time and motor temperature (from 0 to 50 s).

standard deviation σn+1 , a normal condition exists. The drone

continues to fly and updates the threshold temperature
Fig. 8. Flowchart of abnormal temperature detection.
T.
Definition 3: If the measured value t is less than the subtrac-
tion of the mean values T n+1 and standard deviation σn+1 , an
ideal condition exists. The drone continues to fly and records
the deceleration value.
According to the above definitions, based on the measured
values and previous statistic values, the proposed system has
three conditions. These are defined as the ideal condition, nor-
mal condition, and rising abnormal condition. We discuss the
processing in the case of a rising abnormal condition and
Fig. 9. Flowchart of the system for detecting abnormal increase in
ideal condition together. There exist two types of condition
temperature and controlling drone speed. for becoming an ideal condition.
• The drone is in transition from a normal condition to an
ideal condition or is in an ideal condition. It must only
Fig. 9 is the flowchart of the system for detecting an increase update the threshold temperature.
in temperature. • The drone is in transition from an abnormal condi-

We next introduce the details of the system for adjusting the tion to an ideal condition. The temperature increase
speed of the drone, which is based on the detection of a rapid rate is less than that already recorded; the temperature
increase in temperature. First, let the temperature rise of the increase rate, threshold temperature value, and tempera-
drone’s motor be recorded every 2 s for five times be {T 1 , T 2 , ture increase rate are recorded. Finally, the deceleration
. . . , T n }. The first five measurements are used to derive the value is recorded as the minimal value of the deceleration
average and standard deviation of the temperature rise. From formed by the deceleration.
the sixth measurement onward, the mean value T n+1 and the In the processing, from counts 1 to 25 (the first 50 s of
standard deviation σn+1 is used to define the threshold value measurement), the temperature value at a fixed time is rising
abnormally from the initial translational speed 0.05 m/s in the
1 normal condition as the drone speed increases. Fig. 10 dis-
n
T̄n+1 = Ti , i = 1, . . . , n (2) plays the relationship between time and motor temperature.
n

i=1 Decelerate from the temperature increase rate rapidly as shown
 n in Fig. 11. By comparing the temperature rise under decelera-
 1  2
σn+1 = Tn+1 − T̄n+1 . (3) tion in the ascending abnormal condition with that in the ideal
n condition, the optimum deceleration value can be determined.
i=1
For counts 1 to 25, we define the deceleration rate as (0.001
To update the threshold temperature and record the decel- + 0.0005×n) m/s when the nth abnormality occurs. Let x0 be
eration value, we have the following definitions. the optimal value for deceleration at the time between counts
Definition 1: If the measured value t is greater than or equal 1 to 25. The optimal value of deceleration is 0.008 (n = 14).
to the sum of the mean values T n+1 and standard deviation Fig. 11 displays the relationship between the time and speed
σn+1 , a rising abnormal condition exists. The drone performs changes.
deceleration. From Figs. 10 and 11, we determined that the temperature
Definition 2: If the measured value t is less than the sum rising rate exceeds the threshold of temperature rising rate
of the mean values T n+1 and standard deviation σn+1 , yet between 10 and 12 s and the deceleration process is initiated.
is greater than the subtraction of the mean values T n+1 and The deceleration process continues for 22 s. From Fig. 11, we

Authorized licensed use limited to: Universidad Federal de Pernambuco. Downloaded on January 26,2023 at 16:53:18 UTC from IEEE Xplore. Restrictions apply.
LU et al.: MOTOR ANOMALY DETECTION FOR UAVs USING REINFORCEMENT LEARNING 2319

Fig. 11. Relationship between time and speed (from 0 to 50 s). Fig. 14. Relationship between time and motor temperature (from 71 to 90 s).

Fig. 12. Relationship between time and motor temperature (from 51 to 70 s). Fig. 15. Relationship between time and speed (from 71 to 90 s).

nth abnormality is a deceleration rate x2 obtained by (x0 −

0.0001×n) m/s from the initial speed 0.05 m/s. The change of
motor temperature and speed are displayed in Figs. 14 and 15,
respectively.
Finally, we have three values from 0 to 90 s for detecting
abnormal motor temperature. We choose the optimal threshold
temperature value of the motor as

T̃ = min{Tx0 , Tx1 , Tx2 } (4)

where Tx is the average temperature change of the motor at

different deceleration rates of different operation conditions.
Fig. 13. Relationship between time and speed (from 51 to 70 s).
From (4), we determined that the optimal threshold value
of the motor temperature changes with time. This is superior
to using a static threshold temperature for detecting.
determined that the deceleration rate increases continuously
(from 0.0015 to 0.002 m/s).
Then, for counts 26 to 35 (from 51 to 70 s of the measure- IV. S IMULATION R ESULTS
ment), we define the deceleration rate x1 as (x0 + 0.0001 × In this simulation experiment, a system was constructed to
n) m/s when the nth abnormality occurs. The change of land the drone when the motor temperature exceeded a certain
motor temperature and speed are indicated in Figs. 12 and 13, value. Simulations confirmed that this objective was achieved.
respectively. When the system was operated with an initial threshold of
For counts 36 to 45 (from 71 to 90 s of the measurements), 28 ◦ C, the Raspberry Pi could detect abnormalities and the
we consider x0 as the optimal value and set the deceleration drone was landed. We calculated the deceleration rates, for
value as −0.0001. Therefore, the speed at the time of the determining and testing the system performance.

Authorized licensed use limited to: Universidad Federal de Pernambuco. Downloaded on January 26,2023 at 16:53:18 UTC from IEEE Xplore. Restrictions apply.
2320 IEEE INTERNET OF THINGS JOURNAL, VOL. 5, NO. 4, AUGUST 2018

TABLE I TABLE III

S PEED AND T EMPERATURE C HANGES F ROM 0 TO 50 s S PEED AND T EMPERATURE C HANGES F ROM 71 TO 90 s

TABLE IV
S PEED AND T EMPERATURE C HANGES F ROM C OUNTS 46 TO 55
(D ECELERATION R ATE I S 0.008 m/s)

TABLE II
S PEED AND T EMPERATURE C HANGES F ROM 51 TO 70 s
TABLE V
S PEED AND T EMPERATURE C HANGES F ROM C OUNTS 56 TO 65
(D ECELERATION R ATE I S 0.0084 m/s)

A. Calculation of Deceleration Rates

In the rising abnormality state, the process changes based
on the number of times the temperature rise is measured. The
was 0 ◦ C, where the optimal deceleration rate was 0.0079 m/s
real time speed and the changes of temperature of counts 1 to
(=0.05–0.0421 m/s).
25 are presented in Table I.
From Table I, we identified that the minimum temperature
change was −0.188 ◦ C, where the optimal deceleration rate B. Calculation of Optimal Deceleration Rate
was 0.008 m/s (=0.05–0.042 m/s). In the next experiment, we In the above stage, we obtained three deceleration rates
monitored the real time speed and the changes of temperature (0.008 m/s, 0.0084 m/s, and 0.0079 m/s) in different operation
of counts 26 to 35; this is displayed in Table II. conditions. To calculate the optimal deceleration rate, we set
From Table II, we determined that the minimum tempera- the three deceleration rates for testing. In the test experiment
ture change below 0.008 m/s at the first time measurement (from counts 46 to 55), we obtained the results in Table IV.
was −0.125 ◦ C, where the optimal deceleration rate was From Table IV, we determined that the average temperature
0.0084 m/s (=0.05–0.0416 m/s). In the next experiment, we change was −0.062 ◦ C at the deceleration rate of 0.008 m/s.
monitored the real time speed and the changes of temperature In Tables V and VI, we indicate that the average change was
of counts 36 to 45; this is displayed in Table III. 0.125 ◦ C at the deceleration rate of 0.0084 m/s and the average
From Table III, we identified that the minimum tempera- change was −0.061 ◦ C at the deceleration rate of 0.0079 m/s,
ture change above 0.008 m/s at the first time measurement respectively.

Authorized licensed use limited to: Universidad Federal de Pernambuco. Downloaded on January 26,2023 at 16:53:18 UTC from IEEE Xplore. Restrictions apply.
LU et al.: MOTOR ANOMALY DETECTION FOR UAVs USING REINFORCEMENT LEARNING 2321

TABLE VI TABLE VIII

S PEED AND T EMPERATURE C HANGES F ROM C OUNTS 66 TO 75 S PEED AND T EMPERATURE C HANGES F ROM C OUNTS 101 TO 125
(D ECELERATION R ATE I S 0.0079 m/s) U NDER FALLING C ONDITION OF THE D RONE

TABLE VII
S PEED AND T EMPERATURE C HANGES F ROM C OUNTS 76 TO 100
U NDER C LIMBING C ONDITION OF THE D RONE

V. C ONCLUSION
In this paper, we proposed a reinforcement learning-based
motor-temperature anomaly detection system for UAVs. Motor
malfunction is a major cause of drone crashes. The proposed
system employed temperature sensors and a Raspberry Pi
CPU for processing. In simulations using a dynamic motor-
temperature threshold, execution of successful automatic land-
ing was demonstrated. This system facilitates the avoidance
of motor failure, allowing future drones to fly for longer peri-
ods. In future, we will consider implementing deep learning
methods [6], [7], [11]–[14] to select the threshold tempera-
ture. Further, we will attempt to use PID [9], [10] rather
than Raspberry Pi for reducing the weight and improving the
computing ability.
From Tables IV to VI, we can conclude that the deceleration
rate 0.008 m/s is the optimal deceleration value, having the R EFERENCES
lowest average temperature change, −0.062 ◦ C. [1] Started Providing IoT Solution to Detect Abnormality From Working
Sound, NTT DATA, Tokyo, Japan, accessed on Nov. 3, 2016. [Online].
Available: https://latte.la/column/43321221
C. Experimental Certification [2] K-1. Motor Abnormal Temperature Detection System—Infrared
Thermography, Nippon Avionics Co., Ltd., Tokyo, Japan, 1999.
To confirm the effectiveness of this result, we tested the [Online]. Available: http://www.avio.co.jp/products/infrared/solution/
optimal deceleration value 0.008 m/s in the conditions of k-heavy/k-01.html
climbing and falling of the drone. The results are presented in [3] Metal Feather Fan Sensor, Hirozawa Seiki Seisakusho Co., Ltd.,
Ibaraki, Japan, accessed on Mar. 15, 2017. [Online]. Available:
Tables VII and VIII. http://www.hirosawaseiki.co.jp/ikurafan/pages/Cat/IkuraFan1406/
From Table VII, we identified that the temperature increased textunderscore084.pdf
from 19.375 ◦ C to 27.687 ◦ C, which is less than the ini- [4] Design Life of Motor—Technical Data, Oriental Motor Co., Ltd.,
Bengaluru, India, accessed on Mar. 18, 2017. [Online]. Available:
tial threshold of 28 ◦ C. From Table VIII, we determined https://www.orientalmotor.co.jp/ tech/reference/life01/
that the temperature of the motor increased from 18.75 ◦ C [5] T. Yuuto. Detection and Deterioration Diagn0.313osis of
to 25.312 ◦ C, which is also below the initial threshold of Electrical Equipment Based on Current Waveform Pattern
Analysis. Accessed on Mar. 19, 2017. [Online]. Available:
28 ◦ C. Hence, we confirmed that the optimal deceleration http://vision.kyoto-u.ac.jp/japanese/happyou/pdf/Y-Tamura/
value of 0.008 m/s can suppress a temperature increase. textunderscoreASN/textunderscore 20131129.pdf

Authorized licensed use limited to: Universidad Federal de Pernambuco. Downloaded on January 26,2023 at 16:53:18 UTC from IEEE Xplore. Restrictions apply.
2322 IEEE INTERNET OF THINGS JOURNAL, VOL. 5, NO. 4, AUGUST 2018

[6] Y. Li et al., “Underwater image de-scattering and classification by deep Shenglin Mu received the B.E. degree in
neural network,” Comput. Elect. Eng., vol. 54, pp. 68–77, Aug. 2016. mechanical engineering and automation from the
[7] H. Lu et al., “Wound intensity correction and segmentation with con- Northeastern University, Shenyang, China, in 2007,
volutional neural networks,” Concurrency Comput. Pract. Exp., vol. 29, and the M.E. and Ph.D. degrees in electronic and
no. 6, 2017, Art. no. e3927. information systems engineering from Yamaguchi
[8] Raspberry Pi Setting. Accessed on Jan. 11, 2017. [Online]. Available: University, Yamaguchi, Japan, in 2010 and 2013,
http://yamaryu0508.hatenablog.com/entry/2014/08/16/202441 respectively.
[9] H. Lu and S. Serikawa, “Design of freely configurable safety light cur- In 2013, he joined the Department of Electronic
tain using hemispherical mirrors,” IEEJ Trans. Elect. Electron. Eng., Control Engineering, National Institution of
vol. 8, pp. 110–111, Oct. 2013. Technology, Hiroshima College, Hiroshima, China.
[10] H. Lu, Y. Li, Y. Li, S. Serikawa, and H. Kim, “Highly accurate energy- In 2017, he joined the Graduate School of Science
conserving flexible touch sensors,” Sensors Mater., vol. 29, no. 6, and Engineering, Ehime University, Matsuyama, Japan. His current research
pp. 611–617, 2017. interests include control engineering, especially intelligent control, intelligent
[11] M. Chen, Y. Hao, K. Hwang, L. Wang, and L. Wang, “Disease prediction algorithms, mechatronics, robotics, and sensing technology.
by machine learning over big data from healthcare communities,” IEEE
Access, vol. 5, pp. 8869–8879, 2017.
[12] M. Chen, J. Yang, Y. Hao, S. Mao, and K. Hwang, “A 5G cognitive
system for healthcare,” Big Data Cogn. Comput., vol. 1, no. 1, pp. 1–15,
2017, doi: 10.3390/bdcc1010002.
[13] M. Chen, Y. Qian, S. Mao, W. Tang, and X. Yang, “Software-
defined mobile networks security,” Mobile Netw. Appl., vol. 21, no. 5,
pp. 729–743, 2016.
[14] M. Chen et al., “Wearable 2.0: Enabling human-cloud integration in next Dong Wang received the B.E. degree in electronic
generation healthcare systems,” IEEE Commun. Mag., vol. 55, no. 1, information engineering and Ph.D. degree in sig-
pp. 54–61, Jan. 2017. nal and information processing from the Dalian
[15] R. S. Sutton and A. G. Barto, Reinforcement Learning. Cambridge, MA, University of Technology (DUT), Dalian, China, in
USA: MIT Press, 2016. 2008 and 2013, respectively.
[16] A. S. Therrien, D. M. Wolpert, and A. J. Bastian, “Effective reinforce- He is currently an Associate Professor with
ment learning following cerebellar damage requires a balance between the School of Information and Communication
exploration and motor noise,” Brain, vol. 139, pp. 101–114, Jan. 2016. Engineering, DUT. His current research interests
[17] K. Yang et al., “An efficient and fine-grained big data access control include face recognition, person reidentification, and
scheme with privacy-preserving policy,” IEEE Internet Things J., vol. 4, object tracking.
no. 2, pp. 563–571, Apr. 2017.
[18] D. Chen et al., “S2M: A lightweight acoustic fingerprints-based wireless
device authentication protocol,” IEEE Internet Things J., vol. 4, no. 1,
pp. 88–100, Feb. 2017.

Hyoungseop Kim received the B.A. degree in

Huimin Lu received the B.S. degree in electronics
electrical engineering, and master’s and Ph.D.
information science and technology from Yangzhou
degrees from the Kyushu Institute of Technology,
University, Yangzhou, China, in 2008, the M.S.
Kitakyushu, Japan, in 1994, 1996, and 2001,
degrees in electrical engineering from the Kyushu
respectively.
Institute of Technology, Kitakyushu, Japan, and
He is a Professor with the Department of Control
Yangzhou University, in 2011, and the Ph.D. degree
Engineering, Kyushu Institute of Technology. His
in electrical engineering from the Kyushu Institute
current research interest includes medical application
of Technology, in 2014.
of image analysis.
From 2013 to 2016, he was a JSPS Research
Fellow (DC2, PD, and FPD) with the Kyushu
Institute of Technology, where he is currently an
Assistant Professor. He is an Excellent Young Researcher with MEXT-Japan,
Tokyo, Japan. His current research interests include computer vision, robotics,
artificial intelligence, and ocean observing.

Seiichi Serikawa received the B.S. and M.S.

degrees in electronic engineering from Kumamoto
Yujie Li received the B.S. degree in computer University, Kumamoto, Japan, in 1984 and 1986,
science and technology from Yangzhou University, respectively, and the Ph.D. degree in electronic engi-
Yangzhou, China, in 2009, the M.S. degrees in neering from the Kyushu Institute of Technology,
electrical engineering from the Kyushu Institute Kitakyushu, Japan, in 1994.
of Technology, Kitakyushu, Japan, and Yangzhou He is currently a Vice President with the
University, in 2012, and the Ph.D. degree from the Kyushu Institute of Technology, where he serves
Kyushu Institute of Technology, in 2015. as a Professor with the Center for Socio-
She is currently a Lecturer with Yangzhou Robotic Synthesis and the Department of Electrical
University. Her current research interests include and Electronic Engineering. His current research
computer vision, sensors, and image segmentation. interests include computer vision, sensors, and robotics.

Authorized licensed use limited to: Universidad Federal de Pernambuco. Downloaded on January 26,2023 at 16:53:18 UTC from IEEE Xplore. Restrictions apply.