IEEE TRANSACTIONS ON MOBILE COMPUTING,

VOL. 10,

NO. 5,

MAY 2011

669

Design and Performance of an Optimal Inertial

Power Harvester for Human-Powered Devices
Jaeseok Yun, Shwetak N. Patel, Matthew S. Reynolds, and Gregory D. Abowd
AbstractWe present an empirical study of the long-term practicality of using human motion to generate operating power for bodymounted consumer electronics and health sensors. We have collected a large continuous acceleration data set from eight
experimental subjects going about their normal daily routine for three days each. Each subject is instrumented with a data collection
apparatus that simultaneously logs 3-axis, 80 Hz acceleration data from six body locations. We use this data set to optimize a firstprinciples physical model of the commonly used velocity damped resonant generator (VDRG) by selecting physical parameters such
as resonant frequency and damping coefficient to maximize the harvested power. Our results show that with reasonable assumptions
on size, mass, placement, and efficiency of VDRG harvesters, most body-mounted wireless sensors and even some consumer
electronics devices can be powered continuously and indefinitely from everyday motion. We have optimized the power harvesters for
each individual and for each body location. In addition, we present the potential of designing a damping- and frequency-tunable power
harvester that could mitigate the power reduction of a generator generalized for average subjects. We present the full details on the
collection of the acceleration data sets, the development of the VDRG model, and a numerical simulator, and discuss some of the
future challenges that remain in this promising field of research.
Index TermsHuman power harvesting, inertial generator, optimal design, tunable generator, human-powered device.

INTRODUCTION

OWER

remains a key to unlocking the potential for a

sustainable ubicomp reality, particularly for body-worn
mobile electronics. Paradiso and Starner have shown the
promise of scavenging energy to power mobile electronics
[1], [2], [3], [4]. While power harvesting from natural and
renewable sources has been a long-standing research area,
there is surprisingly little in the published literature that
demonstrates the long-term, practical capabilities of harvesting energy from daily human activity. Human activity
levels vary greatly over the course of a typical day,
including periods of sleep, quiet intervals for reading or
office work, and highly active periods of walking, running,
or participating in sports activities. Most prior research into
power harvester performance using human subjects has
been performed for short time intervals in laboratory
simulations of selected human activities. In this work, we
report the first all-day, continuous, all-activity study of
inertial power harvester performance using eight untethered human subjects going about their ordinary lives.

. J. Yun is with the U-embedded Convergence Research Center, Korea

Electronics Technology Institute, 68 Yatap-dong, Bundang-gu, Seongnam,
Gyeonggi-do 463-816, S. Korea. E-mail: jaeseok@keti.re.kr.
. S.N. Patel is with the Department of Computer Science and Engineering
and the Department of Electrical Engineering, University of Washington,
Box 352350, Seattle, WA 98195-2350.
E-mail: shwetak@cs.washington.edu.
. M.S. Reynolds is with the Department Electrical and Computer
Engineering, Duke University, 130 Hudson Hall, Box 90291, Durham,
NC 27708. E-mail: matt.reynolds@duke.edu.
. G.D. Abowd is with the School of Interactive Computing and GVU Center,
College of Computing, Georgia Institute of Technology, 329 Technology
Square Research Building (TSRB), Atlanta, GA 30332-0760.
E-mail: abowd@gatech.edu.
Manuscript received 13 May 2009; revised 13 Jan. 2010; accepted 23 June
2010; published online 19 Oct. 2010.
For information on obtaining reprints of this article, please send e-mail to:
tmc@computer.org, and reference IEEECS Log Number TMC-2009-05-0171.
Digital Object Identifier no. 10.1109/TMC.2010.202.
1536-1233/11/$26.00 2011 IEEE

We created an untethered, wearable apparatus for

gathering calibrated 3-axis acceleration data continuously
at an 80 Hz sampling rate from six different body locations.
We collected data sets spanning continuous 24-hour
collection periods, and obtained three days of ordinary
human activity from eight experimental participants. This
unique, high-fidelity data set provides the input for our
power generation experiments.
We next developed a first-principles numerical model of
the most common type of inertial power harvester, the
Velocity Damped Resonant Generator (VDRG), using
MATLAB Simulink. This model provides an accurate and
realistic estimate of achievable performance from an
inertial energy harvester based on the VDRG architecture.
For common wireless health sensors and several commonly
used consumer electronics devices (such as GPS receivers,
cell phones, and MP3 decoders), we used reasonable
assumptions on form factor, placement, and efficiency of
realizable harvesters to create models of appropriate
VDRGs that could be built into each device. By feeding
each subjects acceleration data set into our VDRG models,
we estimate available power to indicate which electronic
devices or wireless sensors can be powered continuously
and indefinitely from energy harvested over the course of a
typical days activity. We explore six different locations for
placement of the harvester-powered device, considering
devices appropriate for each location on the body.
We discuss the optimal design of a VDRG that could be
built into real objects in terms of the critical features of form
factor, proof mass, resonant frequency, and damping
coefficient. We first optimized the VDRG harvester on a
per individual basis, and then tried to generalize the design
across the entire subject pool so that the generator would
work well enough for all subjects. The power harvesters
generalized over all subjects showed different performance
according to six different simulation conditions. Thus, we
Published by the IEEE CS, CASS, ComSoc, IES, & SPS

670

IEEE TRANSACTIONS ON MOBILE COMPUTING,

discuss the potential of designing tunable power harvesters

where the resonant frequency and damping coefficient are
adjustable over time in response to an individual human
and across various activity levels. We provide full details
on the design procedure for a power harvester optimized
for a given individual and location, and for a tunable
inertial generator.
The organization of the paper is as follows: Section 2
introduces various power harvesting systems using human
power as well as several studies on everyday energy
harvesting. Section 3 presents a generic model and operating principle of a VDRG inertial generator, and explains
what physical parameters we employ to optimize VDRGbased generators. Section 4 describes the data collection
apparatus we developed for collecting acceleration data
from our participants, and discusses how to analyze the
captured sensor data for optimal design of the harvester.
Section 5 presents our simulation conditions we considered
for the experiments and the simulation results with the
power estimation procedure built in a numerical simulation
program. Section 6 describes the design of tunable power
harvesters and its performance with our acceleration data
sets. Section 7 discusses some limitations of our method and
gives remaining challenges for tunable harvesters. Finally,
Section 8 offers concluding remarks.

RELATED WORK

The focus of considerable study, energy harvesting has a

long history. Paradiso and Starner presented an extensive
survey of the various available energy sources to power
mobile devices [1], [2], [4], including solar power [5],
background radio signals [6], thermal gradients [7], vibrational excitation [8], and humans. Paradisos team at MIT
has demonstrated some clever means of exploiting human
motion, such as a self-powered, spring-loaded igniter
switch capable of wirelessly transmitting a static ID number
[9] and a shoe-mounted piezoelectric that produces enough
energy to power a wireless transmitter [10].
As these examples suggest, and summarized by Starner
[3], the human body is a tremendous energy reservoir. Our
interest is in the use of human motion to generate electrical
energy. The first use of human-induced vibration as a
practical mechanical energy source was the self-winding
pedometer watch created in the 1600s [11]. Self-winding
watches exploit human motion to wind the mechanism,
yielding enough mechanical power to operate the watch.
Since that time, several commercial products using humaninduced vibration have been invented [2]. Examples
include a shakable flashlight, which employs electromagnetic induction of a magnet moving through electrical coils
to power an LED or incandescent bulb [12], and hand
cranks, which have also been integrated into small flashlights and radios to produce temporary power. Similarly,
machine-induced vibration can also be exploited to harvest
electrical power [13], [14]. In this case, the harvesters
mechanical resonance is tuned to a particular frequency
generated from motors or vehicles, yielding electrical
power ranging from 1 mW to 45 mW, depending upon
the acceleration magnitude, easily powering industrial
wireless sensors without batteries. This capability of

VOL. 10, NO. 5,

MAY 2011

frequency-tuned harvesting provided the motivation for

this research into the feasibility of a macro-size power
harvester using human-induced vibration. However, most
of those vibration-driven resonant generators produce
electrical power by being tuned to a single resonant
frequency produced from environmental vibrational
sources. The performance of the generators would be
limited to a narrow frequency band because the amount of
energy produced is drastically reduced as the source
frequency shifts away from the generators resonant
frequency. Thus, a great deal of effort has been devoted
to developing a multifrequency or frequency-tunable resonant generator that can adapt its resonant frequency to the
source frequency [15], [16], [17], [18]. This type of adaptive
energy harvesting based on a tunable vibration-driven
generator provides the motivation for our research into the
feasibility of tunable inertial harvesters that will nearly
optimally convert kinetic energy to electrical energy during
all kinds of human activities.
Unobtrusively scavenging energy from daily human
activity is a very attractive idea in human-powered wearable
computing. However, few studies have examined closely
how human-induced vibration can be used for scavenging
energy and how much power can practically be scavenged
from vibrational energy in ordinary daily human motion.
Mitcheson et al. presented optimized architectures for
vibration-driven micropower generators with a sinusoidal
driving motion [19]. However, human motion is a complex
combination of sinusoidal components, which are difficult
to harvest simultaneously. Buren et al. measured acceleration from nine locations on the bodies of walking human
subjects and estimated the maximum output power using
simulation models of inertial microgenerators [20]. The
results of the simulations showed that the maximum power
density is around 2:1 mW=cm3 . However, these simulations
were performed with acceleration signals measured from
standard walking motion for 60 seconds on a treadmill
running at a constant 4 km/h, and there is a question of how
this result would extrapolate to normal daily activity, which
is what we want to explore. Moreover, that work did not
account for the expected power efficiency (useful electrical
power/simulated output power) of the harvester model.
Rome et al. exploited a vertical excursion of the 20-38 kg
load in a rigid backpack during normal walking and
climbing on an inclined treadmill [21]. They showed that
users were able to generate up to 7.4 W of power, depending
upon the mass of the load. Despite the attractive amount of
power harvested from the backpack, the huge mass of 38 kg
will limit its use to niche markets. Kuos team developed a
knee brace generator that produced an average of 5 W of
electricity with one device mounted on each leg [22].
However, the user has to wear a bulky orthopedic knee
brace around the knee, so it is difficult for the harvester to be
embedded into the real mobile objects. Recently, Wangs
team demonstrated energy harvesting from small-scale
dynamic muscle movement. This includes a human index
finger tapping and the body movement of a running
hamster wearing nanowire generators consisting of piezoelectric fine-wire, which generates an alternating-current
when cyclically stretched and released [23]. They showed

YUN ET AL.: DESIGN AND PERFORMANCE OF AN OPTIMAL INERTIAL POWER HARVESTER FOR HUMAN-POWERED DEVICES

671

that the voltage and current output from these two examples
(finger tapping and a running hamster) are 25 mV/150 pA
and 50-100 mV/0.5 nA, respectively.
For a self-sustaining system, Trosters team created a
button-sized solar-powered node composed of sensors, a
processor, RF transmitter, and solar cells, to collect, process,
and forward sensory data to a central wearable device,
which consumed less then 700 W in continuous operation
[24]. They showed that if the whole system is running on a
70 percent duty cycle, it could be autonomously powered
through ordinary exposure to sunlight and indoor light for
an entire day. This style of energy harvesting through
everyday activities echoes our motivation in this paper.

Fig. 1. A generic model for the velocity-damped resonant generator.

POWER HARVESTER MODEL

In order to use our measured acceleration data set to

estimate achievable electrical power output, we have
developed a model for the power harvester in MATLAB
Simulink, a numerical simulation program. We are concerned with the energy available from human body motion,
so we concentrate on inertial generator models where the
bodys acceleration imparts forces on a proof mass whose
motion is coupled to the generator itself. Mitcheson
classified inertial generators into three main categories: the
VDRG , the Coulomb-Damped Resonant Generator (CDRG),
and the Coulomb-Force Parametric Generator (CFPG),
depending on whether their operating principles are
electromagnetic, electrostatic, or piezoelectric [19]. It has
been shown that when the internal displacement travel of a
generator exceeds 0.5 mm, e.g., for macroscale (non-MEMS)
generators, the VDRG is the recommended architecture [20].
We have, therefore, concentrated on the VDRG model for
our simulations.
Vibration-driven generators are ideally represented as a
damped mass-spring system, that is, a second-order
differential equation, as in (1):
_  m
m
zt Kzt  Dzt
yt;

where m is the value of the proof mass, K is the spring

constant, D is the damping coefficient, yt is the displacement of the generator, zt is the relative displacement
between the proof mass and the generator, and t is time
[25]. In this damped mass-spring system, the electrical
energy generated is represented as the energy dissipated in
the mechanical damper. In an electromagnetic generator,
this damping is due to Lenzs law as the moving magnet
does mechanical work when traveling in and out of the
generators coils. Given a harmonic source motion yt
Y0 cos !t with the mathematical model in (1), the dissipated power in the damper can be expressed as
P

!3c Y02 !3 m
1  !2c 2 2!c 2

p
where !c !=!n , the resonant frequency !n K=m,
and the normalized damping factor  D=2m!n [19].
A generic model and operating principle for the VDRG is
shown in Fig. 1. Once the generator is accelerated with
displacement yt, the inertia of the proof mass m causes it
to remain stationary relative to the generators motion with

relative displacement zt. While work is being done against

the damping force in the mechanical damper, power that
dissipates into the damper represents the generator output
power. In particular, the VDRG is characterized by a
damping force proportional to the relative velocity of the
proof mass dzt=dt.
We must utilize several important parameters while
carrying out a simulation analysis with measured acceleration signals and the harvester model: the value of the proof
mass m, spring constant K, damping coefficient D, and
internal travel limit Zmax . Of these variables, the proof mass
m and the internal travel limit Zmax are limited to an
acceptable fraction of the mass and size of the object into
which the generator is incorporated. In our later analysis,
we explain our choice of m and Zmax based on the objects
we wish to power.
Previous studies on inertial generators have shown that
the maximum output power of a generator is proportional to
the value of the proof mass independent of the shape of the
input acceleration waveform. The spring constant is chosen
so that the generator resonates at some optimal frequency
component f such that a small bandwidth including f
maximizes the amount of energy recovered from the
complex vibration spectrum.
Accordingly, K is calculated
p
from the equation 2f K=m in terms of the selected
proof mass m and resonant frequency f. Zmax represents the
maximum travel length of the proof mass inside the inertial
generator. As previously mentioned, Zmax strongly depends
on the size of the physical object into which the power
harvester will be incorporated. A similar limitation is
imposed on estimating the value of the proof mass.
The damping coefficient D is related to the Q factor, which
refers to the ratio of the frequency at which the system
oscillates to the rate at which the system dissipates its energy.
Generally, D is chosen to be the lowest value that keeps the
proof mass from reaching the internal travel limit Zmax of the
generator. In our simulation, supposing D to be fixed in
practice by the generator mechanism, we searched for the
optimal D that resulted in the maximum generated power
over the course of each day for each subject. Of course a more
complex dynamically-damped generator could be imagined,
and this generators control system could be incorporated
into the numerical model to more accurately estimate
achievable power for any given generator.

672

IEEE TRANSACTIONS ON MOBILE COMPUTING,

VOL. 10, NO. 5,

MAY 2011

slid down while the participant was walking. The participants had the option to take off the data collection unit for
certain parts of the day if they needed to (e.g., sleeping or
taking a shower). For analysis on the variation of generated
power during different activities, all the subjects were asked
to record their activities and time on diary sheets.

Fig. 2. (a) Illustrates our experimental setup of six 3-axis accelerometer

modules and two data loggers mounted on body locations. Each
accelerometer module was carefully chosen to match the probable
range of acceleration in the corresponding body location. (b) An
accelerometer module mounted on the wrist and a data logger unit.

DATA COLLECTION AND ANALYSIS

4.1 Experimental Setup

To acquire acceleration signals from six body locations of the
human subjects for an entire day, we have developed
wearable data collection units comprising data loggers,
accelerometer modules, and a lithium-ion camcorder battery
(0.1 kg). Each data collection unit consists of two Logomatic
Serial SD data loggers 65 mm  110 mm) from SparkFun
Electronics, whose analog inputs are connected to six triaxial accelerometer modules based on the ADXL330 (3 g)
and ADXL210 (10 g) accelerometers from Analog Devices,
and the MMA7260Q (6 g) accelerometer from Freescale
Semiconductor. Each accelerometer module is packaged in a
small container custom-fabricated with a 3D printer and
filled with hot glue to protect the sensor from dust or sweat.
Thin, flexible wires interconnect each accelerometer module
with the data logger. The data loggers are contained in a
small waist-pack intended to hold a digital camera. The
Logomatic SD data loggers are configured to record each
analog input as a time series file on a 1 GB SD memory card,
with a sampling rate of 80 Hz. This configuration permits
over 24 hours of continuous operation without the need to
recharge the battery or exchange SD cards.
Our experiments consisted of capturing acceleration data
sets from eight participants (four men and four women)
wearing the data collection unit for three days (two
weekdays and one weekend day). During this time,
acceleration data was continuously recorded from six body
locations where power harvesters could plausibly be
mounted: around the neck, the upper arm, the wrist, the
waist, the knee, and the ankle (see Fig. 2a). The neck
accelerometer was worn around the neck on a lanyard, like
a pendant, and the waist accelerometer was attached on the
data logger to function like a cell phone holster on the waist
belt. All the other accelerometers were fastened to the body
using an elastic band with Velcro ends, as shown in Fig. 2b,
except the knee sensor, which was fixed to the skin by a
medical bandage because an elastic band on the knee often

4.2 Processing the Acceleration Signals

The acceleration time series, we obtain from the bodymounted accelerometers corresponds to the acceleration
experienced by the microscopic proof masses inside each
accelerometer unit. Since we use 3-axis accelerometers, we
record each orthogonal component of this acceleration
separately, in X, Y , and Z. These axis labels are local to
each accelerometers reference frame because each accelerometer is attached to a different body part, and people
move around constantly during the day. Fortunately, the
reference frame of the accelerometer proof mass is the same
as that of a proof mass inside a VDRG mounted at the same
location. To obtain the displacements experienced by the
accelerometers, for each axis a time series of differences in
displacement of each accelerometer is first obtained by
double-integrating the acceleration data set. This displacement time series is then used to calculate the relative motion
of the proof mass inside the VDRG by feeding the
differences in displacement into the power harvester model.
It should be noted that, we measured acceleration data
with the assumption that we would harvest energy from
daily human activities using free motion; that is, our VDRG
does not experience forces beyond those due to its relative
displacement and the inertia of its proof mass. For example,
the accelerometers do not measure, and a VDRG would not
harvest energy from, torque generated while maneuvering
a steering wheel to drive a car or pedaling a bicycle.
Similarly, the ground reaction force on the feet while one is
walking or running is not considered. This force-driven (or
torque-driven) energy could be investigated with another
type of sensor and power harvester [10], [22]. Because they
add resistance to the bodys motion beyond their own
weight, these force- or torque-driven harvesters are not
inertial in nature. In this work, we consider only kinetic
energy generated from human motion as an absolute
magnitude of acceleration, assuming the mass of each part
of the torso and the limb are held constant for each subject.
In the case of the neck and waist accelerometers, although
the movement of each sensor is not exactly the motion of the
subjects, they do behave like real objects people wear (e.g.,
pendants or pouches).
Even though the data sheets for the accelerometers that
we used provide typical zero-gravity output voltage and
sensitivity (volt/gravity), manual calibrations were performed for all accelerometers to obtain more accurate
translation from the voltage output of each accelerometer
to its acceleration value. To accomplish this calibration, the
0 g voltage and sensitivity for every axis of each accelerometer were recorded by positioning each sensitive axis
(X, Y , Z) of the accelerometer at g and g. Because
accelerometers measure the static acceleration of gravity
(1 g) as well as the dynamic acceleration generated by pure
human motion, gravity is always included in the measured
acceleration data as a DC component. To eliminate this

YUN ET AL.: DESIGN AND PERFORMANCE OF AN OPTIMAL INERTIAL POWER HARVESTER FOR HUMAN-POWERED DEVICES

673

human motion can be approximated as a combination of

several dominant frequency components. This observation
leads us to the potential of designing an optimal inertial
power harvester that can be accurately tuned to a selected
dominant frequency range for a given set of ordinary
human activities.
To algorithmically determine the dominant frequency
of each subjects activities throughout the day, we
searched for the top 20 frequency components, ranked in
order of observed energy from highest to lowest, from
each of six body locations. In Fig. 4, red circles depict the
top 20 frequency components between 0.5 Hz and 40 Hz.
We neglect frequency components below 0.5 Hz due to the
impracticality of manufacturing VDRGs with acceptable
efficiency at such low frequencies. Instead of treating the
observed accelerations as pure tones, we investigated
which frequency bins contain considerable kinetic energy
to set the resonant frequency of the VDRG. Consequently,
from the top 20 frequency components, we again sorted to
obtain the 20 frequency bins by aggregating the adjacent
frequency bins (0:25 Hz) and ranking the sums. The top
20 frequency bin values are then used to tune the VDRG
model to obtain an optimal inertial generator that yields
the maximum output power with a measured acceleration
signal. This optimization is performed bypadjusting
spring

constant K to satisfy the equation 2f K=m, where m

is the value of the proof mass of the inertial generator and
f is the resonant frequency.

5
Fig. 3. An example (the arm) of (a) unfiltered and (b) filtered acceleration
waveform over a 24-hour period.

background signal, we high-pass filtered the measured

acceleration signals with a 0.05 Hz cutoff frequency. Fig. 3
shows an example of the filtered acceleration waveform.

4.3 Data Analysis in Frequency Domain

To gain more insight about the measured acceleration data,
we generated frequency domain spectra by performing a
FFT (Fast Fourier Transform) on the acceleration signal after
removing the background signal due to gravity. Fig. 4
displays the spectra of the acceleration signals from six
locations on the body of one subject during one day. Since
each plot is made with the same scale, we immediately
observe that the lower body (waist, knee, and ankle)
experiences are much more pronounced than the upper
body (neck, arm, and wrist). This suggests that the lower
body is likely to yield more electrical energy converted from
kinetic energy than the upper body. However, the upper
body spectra (especially the wrist) contain energy at lower
frequencies (below 0.5 Hz) when compared to the lower
body, probably because the upper body has a higher degree
of freedom when moving than the lower body (e.g., talking
while gesturing while sitting on a chair).
Although, ordinary human motion is not a pure
sinusoidal driving signal, i.e., cos2 ft, several large
peaks appear in each spectrum. Accordingly, ordinary

SIMULATION AND RESULTS

5.1 Simulation Setup

To estimate the electrical power that could be generated by
VDRGs incorporated into everyday electronic devices, we
chose a wristwatch, a cell phone, and a shoe as representatives of typical mobile devices. A wristwatch is a good
example of a device for which power harvesting technology
has already been developed. Cell phones are common
mobile devices we use every day, and their size and mass
are similar to those of an MP3 player such as an iPod.
Although a shoe is not per se a device that consumes
electrical power, it does represent an object into which
power harvesting technology can be embedded. A number
of inventors have developed prototype shoe-mounted
devices and described how they may be powered by
energy harvesting, though most of those harvesters are of
the noninertial type that rely on inserting the harvester
between the foot and the ground.
Based on an informal empirical survey of digital devices,
and some general assumptions, we have developed what
we believe are reasonable values of the proof mass m and
internal travel length Zmax for a harvester incorporated into
each object: 2 g/4.2 cm, 36 g/10 cm, and 100 g/20 cm for a
wristwatch, a cell phone, and a shoe, respectively. The
proof mass of 2 g for the wristwatch was derived from the
mass of a commercialized product [2] and the internal
travel length for a wristwatch was obtained from the
average value of the diameters of 41 SWATCH wristwatches; the proof mass of the cell phone was derived from
one-third of the average mass of 40 NOKIA cell phones;

674

IEEE TRANSACTIONS ON MOBILE COMPUTING,

VOL. 10, NO. 5,

MAY 2011

Fig. 4. The spectra of the acceleration signals from six locations from one subject over a 24-hour period. (a) Neck. (b) Arm. (c) Wrist. (d) Waist.
(e) Knee. (f) Ankle.

and the proof mass of the shoe was obtained from an

assumption that shoe length and mass for the normal adult
would be greater than 20 cm and 300 g, respectively.
Based on the chosen objects and the body locations, we
selected the appropriate power harvester configuration for
simulation. We could not imagine any useful digital object
mountable on the knee, so we excluded this case from the
simulation. We considered all other configurations as
plausible conditions in daily life. The wristwatch cases
represented either a pendant worn around the neck or a
watch worn on the wrist; the cell phone cases represented a
phone suspended on a necklace around the neck, inside an
arm band, or in a waist holster; and the shoe-size object
represented a device that could be worn on the ankle since
such devices would not typically be placed on other parts of
the body.

5.2 Power Estimation Procedure

We have developed the following procedure for using the
measured acceleration data set, in combination with the
VDRG model, to estimate the available power (see Fig. 5).
First, an acceleration data set for an entire day is divided
into 8;640 24 hr  60 min=hr  60 sec=min  1=10 sec
ten-second interval fragments. Each fragment is fed into
the generator model so that the displacement of the
generator frame (i.e., the accelerometer container) yt can
be obtained by double integrating the acceleration signal.
Subsequently, the relative displacement zt and derivative
dzt=dt of the proof mass can be calculated from yt. As
mentioned in the previous section, the electrical power
generated in the inertial generators can be represented by
the power dissipated in the mechanical damper shown in
Fig. 1. Accordingly, the energy dissipated in the damper,
Ed is:

Ed

Z2

F  dz;

Z1

where F Dz_ is the damping force, and Z1 and Z2 are the

start and end positions, respectively, for the line integral.
Equation (3) can be expanded as follows:
Ed

Z2

Dz_  dz

D
t0

Z1

dz dz
 dt D
dt dt

_ 2 dt:
z

t0

The average power during time interval T 10 sec is

Paverage

1
D
Ed
T
T

_ 2 dt:
z

t0

By applying this procedure to all fragments successively,

the average power value for every 10-second fragment is
obtained. The power calculation procedure has been
implemented together with the VDRG model in MATLAB
Simulink.
Because the accelerometer modules, we deployed in our
experiments recorded the accelerations along X, Y , and Z
axes, we assumed that a practical VDRG unit would be built
with a single axis in alignment with one of the axes of the
accelerometers. The preferred orientation of the VDRG is
then chosen based on which one of the three accelerometer
axes yields the maximum output power. However, the
orientation of the mobile devices mounted on the human
body would not necessarily be constant over an entire day.
Moreover, the subjects may wear mobile devices in different
ways. For example, people wear wristwatches in different
ways (right/left hand or with the face up/down), and the
orientation of a cell phone in a waist-band holster might be

YUN ET AL.: DESIGN AND PERFORMANCE OF AN OPTIMAL INERTIAL POWER HARVESTER FOR HUMAN-POWERED DEVICES

675

Fig. 5. A flow chart of the signal processing pipeline for the power estimation procedure.

different according to the form factor and position of the

holster on the belt.

5.3 Optimization and Results

Given m and Zmax as derived for each of the selected mobile
devices, and K chosen based on the FFT analysis for each
harvester configuration, we optimized the power harvesters
by searching for the optimal damping coefficient D which
maximizes average output power P . P then represents the
electrical power generated while the subjects wore the data
collection unit over the entire measurement time. First, we
optimized the power harvesters on a per individual basis.
We concatenated the acceleration data sets for three days
for each of our eight subjects to a single time series, and
then performed FFT analysis on it. By doing so, we were
able to obtain a single dominant frequency fo or spring
constant Ko for three days of activities for a given
individual. After optimizing the damping coefficient Do in
terms of the selected spring constant Ko , the power
harvester would work well for a given individual. It is still
possible to design a power harvester optimized on a given
day, and its efficiency of harvesting electrical power would
obviously be better than that of the harvester optimized
over three entire days on a given individual. However,
generally supposing K or D to be held constant in a real
power generator, it is not practical to use different K or D
values in the harvesters on given days.
With the harvesters optimized for the given configuration
for each of our subjects, we are able to estimate the electrical
power for a given form factor. However, even though the
harvesters have been theoretically optimized, other practical
questions may need to be addressed. First, we must estimate
the expected mechanical-to-electrical conversion efficiency
of the harvester, in terms of useful electrical power obtained
as a fraction of the amount of power dissipated in the

damper. Although the dissipated power represents the

generated power of the harvester in Fig. 1, the two values
might differ due to the mechanical aspects of implementation and electrical circuitry efficiency. Therefore, by applying the characteristics of the PMG 27 [13] harvester to our
simulation method and then comparing the result to the
data sheet, we estimated that the typical mechanical to
electrical conversion efficiency would be around 20 percent.
Consequently, we assumed that only 20 percent of the
energy dissipated in the harvesters damper would be
provided to the electronics it was powering. Second, we
assumed that all energy generated over the day could be
stored in some form of storage and would be used to power
wearable electronics. Our last assumption is that the
direction of the power harvester in real objects could be
continuously aligned with the axis that generates the
maximum output for a whole day. Based on these assumptions and the simulation results, the average electrical power
expected from a VDRG embedded in each object was
155  106 W, 1:01  0:46 mW, and 4:9  3:63 mW for the
wristwatch, the cell phone, and the shoe, respectively,
depending upon the body location on which the power
harvester was mounted. Table 1 illustrates the average
output power generated from the harvesters optimized for a
given individual, and corresponding dominant frequency
and damping coefficient. We also provide a graphical
analysis to show the feasibility of using the harvested power
to power various mobile electronics at different simulation
conditions (see Appendix).
From the FFT analysis, we have shown that the dominant
frequencies strongly depend on the subjects activities on a
given day. In contrast, from Table 1 we observed that
damping coefficients vary according to the value of the proof
mass of a VDRG embedded in real objects. The mean value
and standard deviation of the dominant frequencies for six

676

IEEE TRANSACTIONS ON MOBILE COMPUTING,

VOL. 10, NO. 5,

MAY 2011

TABLE 1
Generated Power, Dominant Frequency, and Damping Coefficient: C1-A Watch Hanging around the Neck, C2-A Watch on the
Wrist, C3-A Phone Hanging around the Neck, C4-A Phone on the Arm, C5-A Phone on the Waist, and C6-A Shoe on the Ankle

The unit of P in C1 and C2 is W . All other P are mW. The unit of fo and Do are Hz and N/ms.

simulation conditions are graphically shown in Fig. 6. It is

important to note that the dominant frequencies in the
simulation conditions C2 and C6, a power harvester on the
wrist and a shoe-harvester located on the ankle, distribute
evenly (i.e., within a small bandwidth, at approximately 1
and 2 Hz, respectively) when compared to other cases,
probably because several ambulatory movements of the
wrist and ankle are taking place during individuals
activities in daily life. This suggests the potential of
designing an optimal power harvester that would be tuned
for a given location and work well enough for an entire
subject pool, though some variation would obviously appear
in the performance of the harvester across subjects. For other
cases, we could not find any particular pattern for the

Fig. 6. Mean and standard deviation of the dominant frequencies on the

body locations: C1-A watch hanging around the neck, C2-A watch on the
wrist, C3-A phone hanging around the neck, C4-A phone on the arm,
C5-A phone on the waist, and C6-A shoe on the ankle.

dominant frequencies, so we expect that harvesters optimized with a single K-D combination for all subjects would
not work well for those simulation conditions in practice.
As mentioned previously, we could design a power
harvester with a single K-D combination for a given
location for all of our subjects. The theoretical method
would be analogous to that used in optimizing the harvester for a given individual. We can first concatenate the data
of all the days into a single time series; then we can perform
FFT analysis; and finally we might be able to obtain a single
dominant frequency for a given location. After optimizing
D with the single K value, a harvester for all subjects can be
designed for a given location. However, this optimization
method requires a significant amount of computing
requirements, in particular for performing FFT analysis on
24 days of acceleration data sets (166 million floating-point
values). As an alternative, we chose a simpler way to find a
single K-D combination for all eight subjects by calculating
the average of K-D values, excluding their highest and
lowest values. Based on the six average K-D values
calculated corresponding to each of the simulation conditions, we are able to estimate a new average output power
generated from a harvester having a single K-D combination for a given location. Fig. 7 shows the energy generated
with new harvesters as a fraction of the amount of energy
generated from the harvesters optimized on a given
individual. The new power harvester on the wrist shows
approximately 100 percent power recovery given the even
distribution on the dominant frequency. A shoe-harvester
also shows more than 90 percent recovery, compared to the
harvester optimized on a given individual. However, the
power from a harvester on the arm also shows good
recovery (about 95 percent) even though the distribution of
dominant frequencies illustrated in Table 1 does not look

YUN ET AL.: DESIGN AND PERFORMANCE OF AN OPTIMAL INERTIAL POWER HARVESTER FOR HUMAN-POWERED DEVICES

677

Fig. 8. Searching for a dominant frequency and damping coefficient for a

small time interval fragment.

Fig. 7. The energy generated with the harvesters optimized for a given
location: C1-A watch hanging around the neck, C2-A watch on the wrist,
C3-A phone hanging around the neck, C4-A phone on the arm, C5-A
phone on the waist, and C6-A shoe on the ankle.

uniform. Thus, we know that arm mounted inertial

harvesters may be less influenced by changing the resonant
frequency than other cases, which might be explained by
looking into the spectrum on the arm where large peaks
rarely appear unless individuals are moving their arms
vigorously, e.g., running. In contrast, three cases C1, C3,
and C5 show considerable reduction in the amount of the
energy recovered with the new harvesters, which suggests
that the harvesters optimized for the given simulation
conditions C1, C3, and C5 would not work well for all
subjects, and we would have to investigate another
harvesting technique. In the next section, we discuss a
new type of power harvester-an adaptively-tunable VDRGbased inertial generator.

TUNABLE INERTIAL HARVESTER

The most important feature of vibration-driven resonant

generators is that the output power harvested with those
generators strongly depends on how well the resonant
frequency of the generator is matched with the frequency of
excitational vibration from its environment. However, the
spectrum of the environmental vibration does not show an
ideally sharp impulse due to the variation of source vibration
or noise distortion. Moreover, as our FFT analysis shows in
Section 4.3, the spectra of the acceleration waveforms
collected from daily human activities can be represented as
a combination of sinusoidal components, which are difficult
to harvest at the same time, but rather would have to be
garnered by changing the resonant frequency of the inertial
generator. Thus, to improve the efficiency of vibrationdriven resonant generators under the variable vibrational
sources over time (e.g., everyday human motion), we discuss
tunable inertial harvesters with the assumption that the
spring constant K and damping coefficient D could be
adjusted mechanically or electronically.
We divide an acceleration data set for an entire day into a
series of small time fragments. By successively searching for
a dominant frequency and an optimal damping coefficient
for each of the time fragments, we are able to design a

tunable harvester that generates the maximum energy for

each time fragment, as illustrated in Fig. 8. Namely, the
tunable inertial harvesting technique would allow individuals to garner the electrical power nearly optimally over all
their activities. In the optimization procedure described
previously, we assumed that the direction of the VDRG
embedded in real objects would be aligned with the axis that
generates the maximum output power over the entire
measurement time. We used the same assumption for the
tunable inertial power harvester because the direction of the
power harvester inside the real objects would not be
changeable for every fragment due to the harvester
mechanism.

6.1 Damping-Tunable Harvester

As the source frequency of the environment moves away
from the resonant frequency of the inertial generator, the
electrical power harvested is reduced dramatically. However, even though both frequencies do not perfectly match, if
they are still very close, we could increase the output power
of the generator by adjusting to a higher damping
coefficient, i.e., lowering the Q factor, which means the ratio
of the stored energy to the energy dissipated in each
oscillatory cycle. It has been shown that changing damping
characteristics will not alter the frequency at which the
power output is maximized, and thus increasing the
damping coefficient (lowering the Q factor results in having
a wider bandwidth for harvesting in the spectrum) could
improve the power output, though the resonant frequency
of the generator and the driving frequency of its environment do not match [15]. We expect that a typical VDRG
harvester will be implemented with a permanent magnet
(serving as the proof mass) moving through a stationary coil.
Due to Lenzs law, the damping force of the harvester can be
adjusted by changing the effective load resistance presented
to the generator coil. This variable-load-impedance method
permits dynamic control of the damping constant D. It could
be implemented with either an analog or a digital control
system, though probably a digital control system using a
pulse width modulation technique to dump energy into a
storage capacitor or battery for a certain portion of the proof
mass travel would be the most straightforward implementation. We call this harvester a damping-tunable harvester
whose damping constant is adjustable for every time
fragment to generate the maximum output power, but the
resonant frequency is held constant. Developing the damping-tunable harvester would be reasonable in practice unless
changing the load impedance of the generator coils requires
too much electrical power.

678

IEEE TRANSACTIONS ON MOBILE COMPUTING,

VOL. 10, NO. 5,

MAY 2011

TABLE 2
Average Improvement on the Amount of the Energy
Generated with the Ideal Tunable Harvester when Compared
with the Harvester Optimized on a Given Individual

Fig. 9. Predicting a dominant frequency and damping coefficient for the

following acceleration fragment.

In order to estimate the achievable power with the

damping-tunable harvester from our subjects, we fixed the
dominant frequency that we obtained in optimizing on a
given individual and searched for optimal damping
coefficient for every fragment. The left column in Table 2
shows the average improvement on the amount of the
energy generated with the tunable harvester when compared with the harvester optimized on a given individual.
The tunable harvester slightly improves harvester efficiency
in most cases, with the exception of the watch hanging
around the neck.

6.2 Frequency-Tunable Harvester

Damping-tunable harvesters do not retain their efficiency as
the source frequency shifts far from the resonant frequency
of the generator. To address this case, we propose altering
the resonant frequency of the power harvester by adapting
the spring constant of the spring suspending the proof mass
inside a VDRG. Several studies have been devoted to the
theory of an adaptive frequency-tunable harvester. Roundy
and Zhang presented a frequency-tuning technique to alter
the resonance frequency by applying an electrical voltage
input across a piezoelectric bimorph, which results in a
change in the stiffness of the bimorph [15]. However, this
method requires a continuous source of input power for
active tuning of the beam, reducing the efficiency of the
energy harvesting device. Leland and Wright showed
another technique that adapts the resonant frequency of
the device by applying an axial compressive load and
altering the stiffness of the supported beam [17]. Malkin and
Davis demonstrated a multifrequency piezoelectric energy
harvester consisting of a number of cantilevered beams,
which generates electrical energy over a larger frequency
range [16]. However, the expected efficiency of such devices
would obviously be reduced because a large subset of beams
in the generator will not be in resonance at a given frequency.
Recently, Challa et al. showed the design and testing of a
frequency tunable energy harvesting device using a magnetic force technique, in which the applied attractive and
repulsive magnetic force induces an additional stiffness in
the system and alters the overall effective stiffness and the
corresponding resonant frequency of the device [18]. With a
prototype device with a 26 Hz resonant frequency, they
showed that the magnetic force technique enables resonance
tuning to 20 percent relative to an untuned resonant
frequency (26 Hz) and a continuous power output over the
entire frequency range tested (22-32 Hz). They also presented another magnetic-coupling tunable generator where

either higher power density or a wider tunable frequency

range is achievable using a new geometrical structure-two
independent single degree of freedom cantilever beams [26].
Even though they showed a piezoelectric technique for the
resonant frequency tunable energy harvesting device, they
reported that their tuning technique can be applied to any of
the three energy harvesting techniques including electromagnetic generators, which we consider in this study. Thus,
we call this harvester a frequency-tunable harvester whose
spring constant is adjustable for every time fragment, with
an adaptive damping coefficient.
From the frequency domain analysis, we have shown
that the dominant frequency strongly depends on the
individuals activities. Thus, adapting the spring constant
appropriately for every time fragment implies that the
power harvester would work well for all kinds of activities
of all individuals. To estimate the achievable power with
the frequency-tunable harvester from our subjects, we
found the dominant frequency for each fragment using
the same method described in the FFT analysis section, and
searched for the optimal damping coefficient. The result is
shown in the right column of Table 2. The results from a
watch hanging around the neck, a phone hanging around
the neck, and a phone on the waist show considerable
improvement with the tunable-spring technique. This is
probably because they are hanging objects which are not
fastened on the human body, thus forming a pendulum. In
this scenario, many large peaks may occur in the spectrum
due to cross coupling between excitation from human
motion and the pendulums period. In contrast, the power
generated on a watch on the wrist shows no improvement
because most of the energy may be garnered within a small
bandwidth as shown in Fig. 6, and changing the resonant
frequency does not have a useful effect.

DISCUSSION

Although we have theoretically designed a tunable inertial

harvester for these simulation conditions, this harvesting
method would not be directly deployed as a practical
harvester. This is because the updating procedure needs to
be performed in real time and the prior analysis for every
time fragment is necessary for adjusting parameters of the
following fragment, as illustrated in Fig. 9. In a practical
scenario, we require a causal mechanical system. A simple
implementation would be both causal and memoryless,
where past human activity is not considered in setting the
harvesters parameters in a given time instant. Given
acceleration data during time interval Ti1 , we try to search
for dominant frequency components and select a resonant

YUN ET AL.: DESIGN AND PERFORMANCE OF AN OPTIMAL INERTIAL POWER HARVESTER FOR HUMAN-POWERED DEVICES

TABLE 3
Average Improvement on the Amount of the Energy Generated
with the Practical Tunable Harvester when Compared
with the Harvester Optimized on a Given Individual

frequency fi1 . Next, based on fi1 , we can update Ki and Di ,

which will be used for the next time fragment Ti . In our
experiment, we have set T to 10 seconds because our
simulation model implemented in MATLAB Simulink was
designed with 10-second simulation interval. Table 3 shows
the average improvement on the amount of the energy
generated with the practical tunable harvester when compared with the harvester optimized on a given individual.
In the purely-causal case without memory, we observe a
reduction in the amount of harvested energy. One possible
reason for the reduction is the 10-second time interval which
might be too long to appropriately predict the harvester
parameters for each of all time fragments. Accordingly, we
might be able to use shorter time intervals for better
prediction. However, there are several important issues we
should carefully consider when shortening the time interval
for adjusting physical parameters of VDRG harvesters. First,
shortening the time interval will increase the number of
calculations for FFT processing as well as the timing
constraint within which the measured signal should be
analyzed and the optimal parameters should be estimated,
causing an increased computing power overhead. Depending on the energy cost of computation in a particular
adaptive harvester, compute energy cost may outweigh the
improvement in harvested energy. Second, since human
body movements are usually located in low frequency
bandwidth (0.5 Hz to 4 Hz), shown in Table 1, there are only
5 to 20 cycles included within the 10-second time window.
Accordingly, as the time interval continues to shrink, there
will be fewer cycles located, and we could not expect high
frequency resolution in the frequency domain analysis.
For the overall efficiency of energy harvesting system,
the most critical feature of the tunable harvester is the
actuating power for adapting damping coefficient and
spring constant. Roundy and Zhang demonstrated two
types of frequency tuning approaches: active tuning and
passive tuning. Active tuning can be accomplished by
active actuators that are on all the time to change the
resonant frequency of the generator in real time. In
contrast, passive tuning can be accomplished by actuators
that turn on to tune the resonant frequency of the generator
and then turn off while keeping matched the newly tuned
frequency with the excitational frequency [27]. For our
tunable harvester, we could deploy the passive tuning
technique; from the FFT analysis a microcontroller first
obtains a dominant frequency and an optimal damping
coefficient for a 10-second time fragment; and then, if

679

needed, the controller turns on and drives the actuator to

adjust the stiffness and electronically-induced resistance so
that the power harvester could be appropriately tuned to
the excitational vibration; finally, the controller turns off the
actuator and waits until the next 10-second interval is put
on the queue of FFT analysis. This self-tuning mechanism
enables the tunable power harvester to autonomously
control its resonant frequency and damping coefficient to
account for the change in the vibration characteristics over
time. However, we would have to very carefully consider
the overall energy cost so that the amount of energy
consumed for analyzing acceleration data and driving the
tuning actuator does not overwhelm the electrical energy
harvested during the preceding 10-second interval.
As shown in Table 1 and Fig. 6, all of the dominant
frequencies of the six simulation conditions are between
0.5 Hz and 5 Hz, and thus the frequency band (0.5 Hz and
5 Hz) would be of interest for the frequency-tunable
harvester. If we consider the magnetic force technique
previously described in the frequency-tunable harvester
section [18], the controllable range of resonant frequency
would be 1.6 Hz to 2.4 Hz, assuming that the untuned
resonant frequency is 2 Hz and 20 percent of untuned
resonant frequency is tunable by the technique. Although
the tunable frequency range is not wide enough to cover
the frequency range necessary for our harvesters, we can
imagine achieving a full coverage of the frequency band
of our interest by applying the advanced technique
presented in [26].

CONCLUSION

We have presented the first 24-hour in situ, continuous

study of inertial power harvester performance from everyday activities. We collected realistic, long-term acceleration
data from eight experimental participants performing
everyday activities for three days each. We discussed the
design of a specific type of inertial harvester, known as a
velocity damped resonant generator. Based on the acceleration data collected from six body locations from each of
our eight participants and the first-principles numerical
model of VDRG implemented in MATLAB, we estimated
the amount of energy that can be garnered at those
locations. The output power from the simulation result
shows that it is practical to continuously operate motionpowered wireless health sensors and other low-power
devices, and that motion-generated power may provide
an intermittent power source for more power intensive
devices like an MP3 player or cell phone. For optimal
design of VDRGs that will be embedded into real mobile
devices, we considered selecting optimal spring constant
and damping coefficient of VDRGs, depending upon form
factor and placement of the mobile devices. Our results
show that it is feasible to develop inertial power harvesters
optimized on a given location as well as a given individual,
though some variation appeared in the performance of the
harvesters across the six simulation conditions. We also
discussed the potential of designing a tunable power
harvester whose resonant frequency can be automatically
adjusted to the vibrational excitation of its environment.
With the advent of new low-power integrated circuits and

680

IEEE TRANSACTIONS ON MOBILE COMPUTING,

VOL. 10, NO. 5,

MAY 2011

TABLE 4
Required Power for Consumer Mobile
Devices and Health Sensors

Fig. 10. A generic architecture for a self-powered sensor node.

a. Photoplethysmograph sensor, that measures the blood pressure, the

heart rate, and the respiration rate.
b. With 1 kbit/s, 1 m distance data transmission.
c. With 1 kbit/s, 1 m distance data transmission.

sensors, or the development of adaptively tuned VDRG, we

can imagine supporting a larger number of devices in the
near future.

APPENDIX
POWERING MOBILE ELECTRONICS
Given the power garnered from humans, we are able to
imagine powering devices such as consumer electronics
and self-sustaining body sensor networks. The tapped
power can be an alternative to batteries for consumer
electronics such as watches, MP3 decoders, cell phones, and
GPS (global positioning system) receivers. Another possible
application of human power harvesting is to power
wearable sensors for monitoring human vital signs.
Monitoring of the status of human health could play a
key role in the future health care services for disease
prevention as well as the treatment for chronically ill
patients. The required power for various electronics has
been tabulated and presented in Table 4. Thus, from the
generated power in Table 1, we can determine whether the
wearable electronics in Table 4 can be powered for a given
subject on a given day, depending upon the simulation
conditions. Fig. 12 shows the result of using the garnered
output power to power different wearable electronics at
various parts on the body. In a number of cases, the output
power is insufficient to continuously power the highest
demanding electronics such as the MP3 decoder, the RF
receiver, and the GPS receiver. However, the VDRG can be
used to charge a battery for intermittent operation of those
devices. One can imagine using the garnered power to
charge a battery when a standard recharging power source
is not available.
As mentioned above, a self-sustaining body sensor
network for monitoring the status of human health is one
of the promising applications of power harvesting technology. A generic architecture of a self-powered sensor node is
shown in Fig. 10. Since the basic operation of a wireless
sensor node is to measure raw data, convert it to a digital
value, and wirelessly transmit the digital sensory data to a

Fig. 11. A body sensor network for monitoring human vital signs.

central processing device, a single wireless sensor node

would be composed of a sensor, an A-D converter, and a RF
transmitter. For a self-sustaining sensor node, the power
harvested from inertial generators would have to be larger
than the total amount of power consumed in the node,
which is the sum of the power consumed in each
component in Fig. 10. It was shown that 1 kbit/s is enough
for body sensor data transmission [6], and this can be
provided with sub-W power level within 1 m distance
[41]. In addition, A-D conversion can be achievable with
1 W [40]. Thus, with these ultra-low power-consuming
components, the most important power usage in a selfpowered node is for sensors presented in Table 4. Fig. 11
shows various wearable sensors and the power requirements for monitoring human vital signs: temperature
sensor [37], PPG (photoplethysmograph) sensor for monitoring blood pressure, heart rate, and respiration rate [33],
3-axis accelerometer [36], humidity [34] and pressure
sensors [35], and a central processing device composed of
a micro controller, RF receiver [29], and memory [39]. From
the output power in Table 1, a temperature sensor node can
be continuously operated from the power harvested from a
phone on the upper arm. In a PPG sensor node, its power
requirement exceeds the power garnered in a wristwatch
on the wrist, but the power shortage might be able to be
supplied from the power on the upper arm. While an
accelerometer node works well with the power from a
phone-size harvester on the waist, the 6D motion sensor
would not be able to be powered. A pressure sensor and
humidity sensor could be continuously powered in a shoesize power harvester. We imagine that a central processing
device would be located and powered in a shoe since a
shoe-size harvester shows the largest amount of output
power among six simulation conditions. However, the RF
receiver chip may not be powered from our results, and the
power shortage could be mitigated in several ways. First,

YUN ET AL.: DESIGN AND PERFORMANCE OF AN OPTIMAL INERTIAL POWER HARVESTER FOR HUMAN-POWERED DEVICES

681

Fig. 12. The ratio of the electrical output power of the harvester to the required power for wearable electronics: a watch hanging around the neck, a
watch on the wrist, a phone hanging around the neck, a phone on the waist, a phone on the arm, and a shoe on the ankle. The X-index represents
the subjects (8 subjects  3 days 24 days), and the Y-index represents the wearable electronics considered in Table 4.

because the output power of the harvester is proportional

to the value of the proof mass regardless of the acceleration
waveform and the architecture of the harvester, we can
increase the output power with a heavier proof mass in the
harvester. Second, we could build a specific RF receiver
satisfying the power requirement using off-the-shelf components. As new low-power technologies emerge, we can
imagine many more senors being able to be supported by
this harvesting approach. Thus, we begin to see the

feasibility of developing wearable electronics powered

from normal everyday activities.

REFERENCES
[1]
[2]

J. Paradiso, Systems for Human-Powered Mobile Computing,

Proc. Design Automation Conf. (DAC 06), pp. 645-650, 2006.
J. Paradiso and T. Starner, Energy Scavenging for Mobile and
Wireless Electronics, IEEE Pervasive Computing, vol. 4, no. 1,
pp. 18-27, Jan.-Mar. 2005.

682

[3]
[4]

[5]

[6]

[7]

[8]

[9]

[10]

[11]
[12]
[13]
[14]
[15]
[16]

[17]

[18]

[19]

[20]

[21]

[22]

[23]

[24]

[25]

[26]

[27]

IEEE TRANSACTIONS ON MOBILE COMPUTING,

T. Starner, Human-Powered Wearable Computing, IBM Systems

J., vol. 35, nos. 3/4, pp. 618-629, 1996.
T. Starner and J. Paradiso, Human-Generated Power for Mobile
Electronics, Low-Power Electronics Design, pp. 1-35, CRC Press,
2004.
W. Huynh, J. Dittmer, and A. Alivisatos, Hybrid NanorodPolymer Solar Cells, Science, vol. 295, no. 5564, pp. 2425-2427,
2002.
E. Yeatman, Advances in Power Sources for Wireless Sensor
Nodes, Proc. Workshop Body Sensor Networks (BSN 06), pp. 20-21,
2004.
M. Strasser, Miniaturized Thermoelectric Generators Based on
Poly-Si and Poly-SiGe Surface Micromachining, Sensors and
Actuators A: Physical, vols. 97-98, pp. 535-542, 2002.
R. Amirtharajah and A. Chandrakasan, Self-Powered Signal
Processing Using Vibration-Based Power Generation, IEEE J.
Solid-State Circuits, vol. 33, no. 5, pp. 687-695, May 1998.
J. Paradiso and M. Feldmeier, A Compact, Wireless, SelfPowered Pushbutton Controller, Proc. Intl Conf. Ubiquitous
Computing (Ubicomp 01), pp. 299-304, 2001.
N. Shenck and J. Paradiso, Energy Scavenging with ShoeMounted Piezoelectrics, IEEE Micro, vol. 21, no. 3, pp. 30-42,
May/June 2001.
A. Chapuis and E. Jaquet, The History of the Self-Winding Watch.
Roto-Sadag, 1956.
S.R. Vetorino, J.V. Platt, and D.A. Springer, Renewable Energy
Flashlight, Patent and Trademark Office, 2001.
PMG -generator, http://www.perpetuum.co.uk, 2011.
VEH-360 electromechanical generator, http://www.ferrosi.com/
energy-harvesters.html, 2011.
S. Roundy and Y. Zhang, Toward Self-Tuning Adaptive
Vibration-Based Microgenerators, Proc. SPIE , pp. 373-384, 2005.
M.C. Malkin and C.L. Davis, Multi-Frequency Piezoelectric
Energy Harvester, J. Acoustics Soc., vol. 118, no. 1, pp. 24-24, July
2005.
E.S. Leland and P.K. Wright, Resonance Tuning of Piezoelectric
Vibration Energy Scavenging Generators Using Compressive Axial
Preload, Smart Materials and Structures, vol. 15, no. 5, pp. 14131420, Oct. 2006.
V.R. Challa, M.G. Prasad, Y. Shi, and F.T. Fisher, A Vibration
Energy Harvesting Device with Bidirectional Resonance Frequency Tunability, Smart Materials and Structures., vol. 17, no. 1,
Feb. 2008.
P. Mitcheson, T. Green, E. Yeatman, and A. Holmes, Architectures for Vibration-Driven Micropower Generators, IEEE J.
Microelectromechanical Systems, vol. 13, no. 3, pp. 429-440, June
2004.
T. Buren, P. Mitcheson, T. Green, E. Yeatman, A. Homes, and G.
Troster, Optimization of Inertial Micropower Generators for
Human Walking Motion, IEEE J. Sensors, vol. 6, no. 1, pp. 28-38,
Feb. 2006.
L. Rome, L. Flynn, E. Goldman, and T. Yoo, Generating
Electricity while Walking with Loads, Science, vol. 309, no. 5741,
pp. 1725-1728, 2005.
J. Donelan, Q. Li, V. Naing, J. Hoffer, D. Weber, and A. Kuo,
Biomechanical Energy Harvesting: Generating Electricity during
Walking with Minimal User Effort, Science, vol. 319, no. 5864,
pp. 807-810, 2008.
R. Yang, Y. Qin, C. Li, G. Zhu, and Z.L. Wang, Converting
Biomechanical Energy into Electricity by a Muscle-MovementDriven Nanogenerator, Nano Letters, vol. 9, pp. 1201-1205, Feb.
2009.
N. Bharatula, S. Ossevoort, M. Stager, and G. Troster, Towards
Wearable Autonomous Microsystems, Proc. Intl Conf. Pervasive
Computing (Pervasive 04), pp. 225-237, 2004.
C. Williams and R. Yates, Analysis of a Micro-Electric Generator
for Microsystems, Sensors and Actuators A: Physical, vol. 52, pp. 811, 1996.
V.R. Challa, M.G. Prasad, and F.T. Fisher, High Efficiency Energy
Harvesting Device with Magnetic Coupling for Resonance
Frequency Tuning, Proc. SPIE, Mar. 2008.
S. Roundy, E.S. Leland, J. Baker, E. Carleton, E. Reilly, E. Lai, B.
Otis, J.M. Rabaey, P.K. Wright, and V. Sundararajan, Improving
Power Output for Vibration-Based Energy Scavengers, IEEE
Pervasive Computing, vol. 4, no. 1, pp. 28-36, Jan.-Mar. 2005.

VOL. 10, NO. 5,

MAY 2011

[28] Samsung Helix, Nexus and Pioneer Inno Power Consumption,

http://www.orbitcast.com/archives/samsung-helix-nexus-andpioneer-inno-power-consumption.html, 2006.
[29] ASK Single Conversion Receiver TDA 5200, http://www.
infineon.com/cms/at/product/channel.html?channel=ff80808112
ab681d0112ab6909cd00de, 2011.
[30] GNR1040 GPS Receiver, http://www.glonavgps.com/images/
glonav_gnr1040_pb.pdf, 2004.
[31] A. Sadat, H. Qu, C. Yu, J.S. Yuan, and H. Xie, Low-Power CMOS
Wireless MEMS Motion Sensor for Physiological Activity Monitoring, IEEE Trans. Circuits Systems, vol. 52, no. 12, pp. 2539-2551,
Dec. 2005.
[32] Cell Phone Power Management Requires Small Regulators with
Fast Response, http://www.commsdesign.com/printableArticle/
?articleID=12801994, 2002.
[33] S. Rhee, B.-H. Yang, and H. Asada, Artifact-Resistant, PowerEfficient Design of Finger-Ring Plethysmographic Sensors Part I:
Design and Analysis, Proc. Eng. in Medicine and Biology Soc. Conf.
(EMBS 00), pp. 2792-2795, 2000.
[34] HIH-3610 Series, http://content.honeywell.com/sensing/
prodinfo/humiditymoisture/009012_2.pdf, 2011.
[35] C. Hierolda, Low Power Integrated Pressure Sensor System for
Medical Applications, Sensors and Actuators A: Physical, vol. 73,
nos. 1/2, pp. 1261-1265, 1999.
[36] ADXL330 3-Axis Accelerometer, http://www.analog.com/en/
prod/0,2877,ADXL330,00.html, 2011.
[37] FM20 Low-Voltage Analog Temp Sensor, http://www.
fairchildsemi.com/pf/FM/FM20.html, 2011.
[38] Seiko AGS Quartz Watch, http://www.epson.co.jp/e/company/
milestones/19_ags.pdf, 1998.
[39] G. Mathur, P. Desnoyers, D. Ganesan, and P. Shenoy, Ultra-Low
Power Data Storage for Sensor Networks, Proc. Conf. Information
Processing in Sensor Networks (IPSN 06), pp. 374-381, 2006.
[40] J. Sauerbrey, S. Landsiedel, and R. Thewes, A 0.5-v, 1-w
Successive Approximation ADC, IEEE J. Solid-State Circuits,
vol. 38, no. 7, pp. 1261-1265, July 2003.
[41] D. Yates and A. Holmes, Micro Power Radio Module,
ORESTEIA IST Project, 2003.
Jaeseok Yun received the MS and PhD degrees
in mechatronics from the Gwangju Institute of
Science and Technology, South Korea, in 1999
and 2006, respectively. Currently, he is working
as a senior researcher in the U-embedded
Convergence Research Center at the Korea
Electronics Technology Institute (KETI), Seongnam, South Korea. He was a research scientist/
postdoctoral researcher with the Ubiquitous
Computing Research Group in the College of
Computing and the GVU Center at the Georgia Institute of Technology
from 2006 to 2009. His research interests are in ubiquitous computing,
human-computer interaction, and context-based interactive environments. He is particularly interested in human activity sensing and its
applications.
Shwetak N. Patel received the BS and PhD
degrees in computer science from the Georgia
Institute of Technology (Georgia Tech) in 2003
and 2008, respectively. Currently, he is working
as an assistant professor in the Departments of
Computer Science and Engineering and Electrical Engineering at the University of Washington. His research interests are in the areas of
sensor-enabled embedded systems, humancomputer interaction, ubiquitous computing,
and user interface software and technology. He is particularly interested
in developing easy-to-deploy sensing technologies and approaches for
location and activity recognition applications. He was the cofounder of
Usenso, Inc., a recently acquired demand side energy monitoring
solutions provider. He was also the assistant director of the Aware
Home Research Initiative at Georgia Tech. He was a TR-35 award
recipient for 2009.

YUN ET AL.: DESIGN AND PERFORMANCE OF AN OPTIMAL INERTIAL POWER HARVESTER FOR HUMAN-POWERED DEVICES

Matthew S. Reynolds received the SB, MEng,

and PhD degrees from the Massachusetts
Institute of Technology in 1998, 1999, and
2003, respectively. Currently, he is working as
an assistant professor in Duke Universitys
Department of Electrical and Computer Engineering. He is a cofounder of the RFID systems
firm ThingMagic Inc. and the demand-side
energy conservation technology firm Zensi. His
research interests include the physics of sensors
and actuators, RFID, and signal processing. He is a member of the
Signal Processing and Communications Group and the Computer
Engineering Group at Duke, as well as the IEEE Microwave Theory and
Techniques Society.

683

Gregory D. Abowd received the MS and PhD

degrees in computation from the University of
Oxford, United Kingdom, in 1987 and 1991,
respectively. He is a distinguished professor in
the School of Interactive Computing at the
Georgia Institute of Technology (Georgia Tech)
and the W. George professor and director of the
Georgia Tech/Emory University Health Systems
Institute. His research interests are in applications-driven challenges of ubiquitous computing,
with particular interest in home and health topics.

. For more information on this or any other computing topic,

please visit our Digital Library at www.computer.org/publications/dlib.