yundesignoptimalinertialenergyharvester11.pdf

Design and Performance of an Optimal InertialPower Harvester for Human-Powered Devices

Jaeseok Yun, Shwetak N. Patel, Matthew S. Reynolds, and Gregory D. Abowd

Abstract—We present an empirical study of the long-term practicality of using human motion to generate operating power for body-

mounted 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 first-

principles 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 Terms—Human power harvesting, inertial generator, optimal design, tunable generator, human-powered device.

Ç

1 INTRODUCTION

POWER remains a key to unlocking the potential for asustainable ubicomp reality, particularly for body-worn

mobile electronics. Paradiso and Starner have shown thepromise of scavenging energy to power mobile electronics[1], [2], [3], [4]. While power harvesting from natural andrenewable sources has been a long-standing research area,there is surprisingly little in the published literature thatdemonstrates the long-term, practical capabilities of harvest-ing energy from daily human activity. Human activitylevels vary greatly over the course of a typical day,including periods of sleep, quiet intervals for reading oroffice work, and highly active periods of walking, running,or participating in sports activities. Most prior research intopower harvester performance using human subjects hasbeen performed for short time intervals in laboratorysimulations of selected human activities. In this work, wereport the first all-day, continuous, all-activity study ofinertial power harvester performance using eight unteth-ered human subjects going about their ordinary lives.

We created an untethered, wearable apparatus forgathering calibrated 3-axis acceleration data continuouslyat an 80 Hz sampling rate from six different body locations.We collected data sets spanning continuous 24-hourcollection periods, and obtained three days of ordinaryhuman activity from eight experimental participants. Thisunique, high-fidelity data set provides the input for ourpower generation experiments.

We next developed a first-principles numerical model ofthe most common type of inertial power harvester, theVelocity Damped Resonant Generator (VDRG), usingMATLAB Simulink. This model provides an accurate andrealistic estimate of achievable performance from aninertial energy harvester based on the VDRG architecture.For common wireless health sensors and several commonlyused consumer electronics devices (such as GPS receivers,cell phones, and MP3 decoders), we used reasonableassumptions on form factor, placement, and efficiency ofrealizable harvesters to create models of appropriateVDRGs that could be built into each device. By feedingeach subject’s acceleration data set into our VDRG models,we estimate available power to indicate which electronicdevices or wireless sensors can be powered continuouslyand indefinitely from energy harvested over the course of atypical day’s activity. We explore six different locations forplacement of the harvester-powered device, consideringdevices appropriate for each location on the body.

We discuss the optimal design of a VDRG that could bebuilt into real objects in terms of the critical features of formfactor, proof mass, resonant frequency, and dampingcoefficient. We first optimized the VDRG harvester on aper individual basis, and then tried to generalize the designacross the entire subject pool so that the generator wouldwork well enough for all subjects. The power harvestersgeneralized over all subjects showed different performanceaccording to six different simulation conditions. Thus, we

IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 10, NO. 5, MAY 2011 669

. J. Yun is with the U-embedded Convergence Research Center, KoreaElectronics Technology Institute, 68 Yatap-dong, Bundang-gu, Seongnam,Gyeonggi-do 463-816, S. Korea. E-mail: [email protected].

. S.N. Patel is with the Department of Computer Science and Engineeringand the Department of Electrical Engineering, University of Washington,Box 352350, Seattle, WA 98195-2350.E-mail: [email protected].

. M.S. Reynolds is with the Department Electrical and ComputerEngineering, Duke University, 130 Hudson Hall, Box 90291, Durham,NC 27708. E-mail: [email protected].

. G.D. Abowd is with the School of Interactive Computing and GVU Center,College of Computing, Georgia Institute of Technology, 329 TechnologySquare Research Building (TSRB), Atlanta, GA 30332-0760.E-mail: [email protected].

Manuscript received 13 May 2009; revised 13 Jan. 2010; accepted 23 June2010; published online 19 Oct. 2010.For information on obtaining reprints of this article, please send e-mail to:[email protected], 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 Published by the IEEE CS, CASS, ComSoc, IES, & SPS

discuss the potential of designing tunable power harvesterswhere the resonant frequency and damping coefficient areadjustable over time in response to an individual humanand across various activity levels. We provide full detailson the design procedure for a power harvester optimizedfor a given individual and location, and for a tunableinertial generator.

The organization of the paper is as follows: Section 2introduces various power harvesting systems using humanpower as well as several studies on everyday energyharvesting. Section 3 presents a generic model and operat-ing principle of a VDRG inertial generator, and explainswhat physical parameters we employ to optimize VDRG-based generators. Section 4 describes the data collectionapparatus we developed for collecting acceleration datafrom our participants, and discusses how to analyze thecaptured sensor data for optimal design of the harvester.Section 5 presents our simulation conditions we consideredfor the experiments and the simulation results with thepower estimation procedure built in a numerical simulationprogram. Section 6 describes the design of tunable powerharvesters and its performance with our acceleration datasets. Section 7 discusses some limitations of our method andgives remaining challenges for tunable harvesters. Finally,Section 8 offers concluding remarks.

2 RELATED WORK

The focus of considerable study, energy harvesting has along history. Paradiso and Starner presented an extensivesurvey of the various available energy sources to powermobile devices [1], [2], [4], including solar power [5],background radio signals [6], thermal gradients [7], vibra-tional excitation [8], and humans. Paradiso’s team at MIThas demonstrated some clever means of exploiting humanmotion, such as a self-powered, spring-loaded igniterswitch capable of wirelessly transmitting a static ID number[9] and a shoe-mounted piezoelectric that produces enoughenergy to power a wireless transmitter [10].

As these examples suggest, and summarized by Starner[3], the human body is a tremendous energy reservoir. Ourinterest is in the use of human motion to generate electricalenergy. The first use of human-induced vibration as apractical mechanical energy source was the self-windingpedometer watch created in the 1600s [11]. Self-windingwatches exploit human motion to wind the mechanism,yielding enough mechanical power to operate the watch.Since that time, several commercial products using human-induced vibration have been invented [2]. Examplesinclude a shakable flashlight, which employs electromag-netic induction of a magnet moving through electrical coilsto power an LED or incandescent bulb [12], and handcranks, which have also been integrated into small flash-lights and radios to produce temporary power. Similarly,machine-induced vibration can also be exploited to harvestelectrical power [13], [14]. In this case, the harvester’smechanical resonance is tuned to a particular frequencygenerated from motors or vehicles, yielding electricalpower ranging from 1 mW to 45 mW, depending uponthe acceleration magnitude, easily powering industrialwireless sensors without batteries. This capability of

frequency-tuned harvesting provided the motivation forthis research into the feasibility of a macro-size powerharvester using human-induced vibration. However, mostof those vibration-driven resonant generators produceelectrical power by being tuned to a single resonantfrequency produced from environmental vibrationalsources. The performance of the generators would belimited to a narrow frequency band because the amount ofenergy produced is drastically reduced as the sourcefrequency shifts away from the generator’s resonantfrequency. Thus, a great deal of effort has been devotedto developing a multifrequency or frequency-tunable reso-nant generator that can adapt its resonant frequency to thesource frequency [15], [16], [17], [18]. This type of adaptiveenergy harvesting based on a tunable vibration-drivengenerator provides the motivation for our research into thefeasibility of tunable inertial harvesters that will nearlyoptimally convert kinetic energy to electrical energy duringall kinds of human activities.

Unobtrusively scavenging energy from daily humanactivity is a very attractive idea in human-powered wearablecomputing. However, few studies have examined closelyhow human-induced vibration can be used for scavengingenergy and how much power can practically be scavengedfrom vibrational energy in ordinary daily human motion.Mitcheson et al. presented optimized architectures forvibration-driven micropower generators with a sinusoidaldriving motion [19]. However, human motion is a complexcombination of sinusoidal components, which are difficultto harvest simultaneously. Buren et al. measured accelera-tion from nine locations on the bodies of walking humansubjects and estimated the maximum output power usingsimulation models of inertial microgenerators [20]. Theresults of the simulations showed that the maximum powerdensity is around 2:1 mW=cm3. However, these simulationswere performed with acceleration signals measured fromstandard walking motion for 60 seconds on a treadmillrunning at a constant 4 km/h, and there is a question of howthis result would extrapolate to normal daily activity, whichis what we want to explore. Moreover, that work did notaccount for the expected power efficiency (useful electricalpower/simulated output power) of the harvester model.

Rome et al. exploited a vertical excursion of the 20-38 kgload in a rigid backpack during normal walking andclimbing on an inclined treadmill [21]. They showed thatusers were able to generate up to 7.4 W of power, dependingupon the mass of the load. Despite the attractive amount ofpower harvested from the backpack, the huge mass of 38 kgwill limit its use to niche markets. Kuo’s team developed aknee brace generator that produced an average of 5 W ofelectricity with one device mounted on each leg [22].However, the user has to wear a bulky orthopedic kneebrace around the knee, so it is difficult for the harvester to beembedded into the real mobile objects. Recently, Wang’steam demonstrated energy harvesting from small-scaledynamic muscle movement. This includes a human indexfinger tapping and the body movement of a runninghamster wearing nanowire generators consisting of piezo-electric fine-wire, which generates an alternating-currentwhen cyclically stretched and released [23]. They showed

670 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 10, NO. 5, MAY 2011

that the voltage and current output from these two examples(finger tapping and a running hamster) are 25 mV/150 pAand 50-100 mV/0.5 nA, respectively.

For a self-sustaining system, Troster’s team created abutton-sized solar-powered node composed of sensors, aprocessor, 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 a70 percent duty cycle, it could be autonomously poweredthrough ordinary exposure to sunlight and indoor light foran entire day. This style of energy harvesting througheveryday activities echoes our motivation in this paper.

3 POWER HARVESTER MODEL

In order to use our measured acceleration data set toestimate achievable electrical power output, we havedeveloped a model for the power harvester in MATLABSimulink, a numerical simulation program. We are con-cerned with the energy available from human body motion,so we concentrate on inertial generator models where thebody’s acceleration imparts forces on a proof mass whosemotion is coupled to the generator itself. Mitchesonclassified inertial generators into three main categories: theVDRG , the Coulomb-Damped Resonant Generator (CDRG),and the Coulomb-Force Parametric Generator (CFPG),depending on whether their operating principles areelectromagnetic, electrostatic, or piezoelectric [19]. It hasbeen shown that when the internal displacement travel of agenerator 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 forour simulations.

Vibration-driven generators are ideally represented as adamped mass-spring system, that is, a second-orderdifferential equation, as in (1):

m€zðtÞ ¼ �KzðtÞ �D _zðtÞ �m€yðtÞ; ð1Þ

where m is the value of the proof mass, K is the springconstant, D is the damping coefficient, yðtÞ is the displace-ment of the generator, zðtÞ is the relative displacementbetween the proof mass and the generator, and t is time[25]. In this damped mass-spring system, the electricalenergy generated is represented as the energy dissipated inthe mechanical damper. In an electromagnetic generator,this damping is due to Lenz’s law as the moving magnetdoes mechanical work when traveling in and out of thegenerator’s coils. Given a harmonic source motion yðtÞ ¼Y0 cos!ðtÞ with the mathematical model in (1), the dis-sipated power in the damper can be expressed as

P ¼ �!3cY

20 !

3m

½1� !2c �

2 þ ½2�!c�2; ð2Þ

where !c ¼ ð!=!nÞ, the resonant frequency !n ¼ffiffiffiffiffiffiffiffiffiffiffiK=m

p,

and the normalized damping factor � ¼ ðD=2m!nÞ [19].A generic model and operating principle for the VDRG is

shown in Fig. 1. Once the generator is accelerated withdisplacement yðtÞ, the inertia of the proof mass m causes itto remain stationary relative to the generator’s motion with

relative displacement zðtÞ. While work is being done against

the damping force in the mechanical damper, power thatdissipates 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 dzðtÞ=dt.We must utilize several important parameters while

carrying out a simulation analysis with measured accelera-tion 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 anacceptable 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 objectswe wish to power.

Previous studies on inertial generators have shown thatthe 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 frequencycomponent f such that a small bandwidth including f

maximizes the amount of energy recovered from the

complex vibration spectrum. Accordingly, K is calculated

from the equation 2�f ¼ffiffiffiffiffiffiffiffiffiffiffiK=m

pin 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 powerharvester will be incorporated. A similar limitation is

imposed on estimating the value of the proof mass.The damping coefficientD is related to the Q factor, which

refers to the ratio of the frequency at which the systemoscillates to the rate at which the system dissipates its energy.Generally, D is chosen to be the lowest value that keeps theproof mass from reaching the internal travel limit Zmax of thegenerator. In our simulation, supposing D to be fixed inpractice by the generator mechanism, we searched for theoptimal D that resulted in the maximum generated powerover the course of each day for each subject. Of course a morecomplex dynamically-damped generator could be imagined,and this generator’s control system could be incorporatedinto the numerical model to more accurately estimateachievable power for any given generator.

YUN ET AL.: DESIGN AND PERFORMANCE OF AN OPTIMAL INERTIAL POWER HARVESTER FOR HUMAN-POWERED DEVICES 671

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

4 DATA COLLECTION AND ANALYSIS

4.1 Experimental Setup

To acquire acceleration signals from six body locations of thehuman subjects for an entire day, we have developedwearable data collection units comprising data loggers,accelerometer modules, and a lithium-ion camcorder battery(0.1 kg). Each data collection unit consists of two LogomaticSerial SD data loggers ð65 mm� 110 mm) from SparkFunElectronics, whose analog inputs are connected to six tri-axial accelerometer modules based on the ADXL330 (�3 g)and ADXL210 (�10 g) accelerometers from Analog Devices,and the MMA7260Q (�6 g) accelerometer from FreescaleSemiconductor. Each accelerometer module is packaged in asmall container custom-fabricated with a 3D printer andfilled with hot glue to protect the sensor from dust or sweat.Thin, flexible wires interconnect each accelerometer modulewith the data logger. The data loggers are contained in asmall waist-pack intended to hold a digital camera. TheLogomatic SD data loggers are configured to record eachanalog input as a time series file on a 1 GB SD memory card,with a sampling rate of 80 Hz. This configuration permitsover 24 hours of continuous operation without the need torecharge the battery or exchange SD cards.

Our experiments consisted of capturing acceleration datasets from eight participants (four men and four women)wearing the data collection unit for three days (twoweekdays and one weekend day). During this time,acceleration data was continuously recorded from six bodylocations where power harvesters could plausibly bemounted: around the neck, the upper arm, the wrist, thewaist, the knee, and the ankle (see Fig. 2a). The neckaccelerometer was worn around the neck on a lanyard, likea pendant, and the waist accelerometer was attached on thedata logger to function like a cell phone holster on the waistbelt. All the other accelerometers were fastened to the bodyusing an elastic band with Velcro ends, as shown in Fig. 2b,except the knee sensor, which was fixed to the skin by amedical bandage because an elastic band on the knee often

slid down while the participant was walking. The partici-pants had the option to take off the data collection unit forcertain parts of the day if they needed to (e.g., sleeping ortaking a shower). For analysis on the variation of generatedpower during different activities, all the subjects were askedto record their activities and time on diary sheets.

4.2 Processing the Acceleration Signals

The acceleration time series, we obtain from the body-mounted accelerometers corresponds to the accelerationexperienced by the microscopic proof masses inside eachaccelerometer unit. Since we use 3-axis accelerometers, werecord each orthogonal component of this accelerationseparately, in X, Y , and Z. These axis labels are local toeach accelerometer’s reference frame because each accel-erometer is attached to a different body part, and peoplemove around constantly during the day. Fortunately, thereference frame of the accelerometer proof mass is the sameas that of a proof mass inside a VDRG mounted at the samelocation. To obtain the displacements experienced by theaccelerometers, for each axis a time series of differences indisplacement of each accelerometer is first obtained bydouble-integrating the acceleration data set. This displace-ment time series is then used to calculate the relative motionof the proof mass inside the VDRG by feeding thedifferences in displacement into the power harvester model.

It should be noted that, we measured acceleration datawith the assumption that we would harvest energy fromdaily human activities using free motion; that is, our VDRGdoes not experience forces beyond those due to its relativedisplacement and the inertia of its proof mass. For example,the accelerometers do not measure, and a VDRG would notharvest energy from, torque generated while maneuveringa steering wheel to drive a car or pedaling a bicycle.Similarly, the ground reaction force on the feet while one iswalking or running is not considered. This force-driven (ortorque-driven) energy could be investigated with anothertype of sensor and power harvester [10], [22]. Because theyadd resistance to the body’s motion beyond their ownweight, these force- or torque-driven harvesters are notinertial in nature. In this work, we consider only kineticenergy generated from human motion as an absolutemagnitude of acceleration, assuming the mass of each partof the torso and the limb are held constant for each subject.In the case of the neck and waist accelerometers, althoughthe movement of each sensor is not exactly the motion of thesubjects, they do behave like real objects people wear (e.g.,pendants or pouches).

Even though the data sheets for the accelerometers thatwe used provide typical zero-gravity output voltage andsensitivity (volt/gravity), manual calibrations were per-formed for all accelerometers to obtain more accuratetranslation from the voltage output of each accelerometerto its acceleration value. To accomplish this calibration, the0 g voltage and sensitivity for every axis of each accel-erometer were recorded by positioning each sensitive axis(X, Y , Z) of the accelerometer at þg and �g. Becauseaccelerometers measure the static acceleration of gravity(1 g) as well as the dynamic acceleration generated by purehuman motion, gravity is always included in the measuredacceleration data as a DC component. To eliminate this


Fig. 2. (a) Illustrates our experimental setup of six 3-axis accelerometermodules and two data loggers mounted on body locations. Eachaccelerometer module was carefully chosen to match the probablerange of acceleration in the corresponding body location. (b) Anaccelerometer module mounted on the wrist and a data logger unit.

background signal, we high-pass filtered the measuredacceleration signals with a 0.05 Hz cutoff frequency. Fig. 3shows 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 aFFT (Fast Fourier Transform) on the acceleration signal afterremoving the background signal due to gravity. Fig. 4displays the spectra of the acceleration signals from sixlocations on the body of one subject during one day. Sinceeach plot is made with the same scale, we immediatelyobserve that the lower body (waist, knee, and ankle)experiences are much more pronounced than the upperbody (neck, arm, and wrist). This suggests that the lowerbody is likely to yield more electrical energy converted fromkinetic energy than the upper body. However, the upperbody spectra (especially the wrist) contain energy at lowerfrequencies (below 0.5 Hz) when compared to the lowerbody, probably because the upper body has a higher degreeof freedom when moving than the lower body (e.g., talkingwhile gesturing while sitting on a chair).

Although, ordinary human motion is not a puresinusoidal driving signal, i.e., cosð2� ftÞ, several largepeaks appear in each spectrum. Accordingly, ordinary

human motion can be approximated as a combination ofseveral dominant frequency components. This observationleads us to the potential of designing an optimal inertialpower harvester that can be accurately tuned to a selecteddominant frequency range for a given set of ordinaryhuman activities.

To algorithmically determine the dominant frequency

of each subject’s 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 by adjusting spring

constant K to satisfy the equation 2�f ¼ffiffiffiffiffiffiffiffiffiffiffiK=m

p, where m

is the value of the proof mass of the inertial generator and

f is the resonant frequency.

5 SIMULATION AND RESULTS

5.1 Simulation Setup

To estimate the electrical power that could be generated byVDRGs incorporated into everyday electronic devices, wechose a wristwatch, a cell phone, and a shoe as representa-tives of typical mobile devices. A wristwatch is a goodexample of a device for which power harvesting technologyhas already been developed. Cell phones are commonmobile devices we use every day, and their size and massare similar to those of an MP3 player such as an iPod.Although a shoe is not per se a device that consumeselectrical power, it does represent an object into whichpower harvesting technology can be embedded. A numberof inventors have developed prototype shoe-mounteddevices and described how they may be powered byenergy harvesting, though most of those harvesters are ofthe noninertial type that rely on inserting the harvesterbetween the foot and the ground.

Based on an informal empirical survey of digital devices,and some general assumptions, we have developed whatwe believe are reasonable values of the proof mass m andinternal travel length Zmax for a harvester incorporated intoeach object: 2 g/4.2 cm, 36 g/10 cm, and 100 g/20 cm for awristwatch, a cell phone, and a shoe, respectively. Theproof mass of 2 g for the wristwatch was derived from themass of a commercialized product [2] and the internaltravel length for a wristwatch was obtained from theaverage value of the diameters of 41 SWATCH wrist-watches; the proof mass of the cell phone was derived fromone-third of the average mass of 40 NOKIA cell phones;


Fig. 3. An example (the arm) of (a) unfiltered and (b) filtered acceleration

waveform over a 24-hour period.

and the proof mass of the shoe was obtained from anassumption that shoe length and mass for the normal adultwould be greater than 20 cm and 300 g, respectively.

Based on the chosen objects and the body locations, weselected the appropriate power harvester configuration forsimulation. We could not imagine any useful digital objectmountable on the knee, so we excluded this case from thesimulation. We considered all other configurations asplausible conditions in daily life. The wristwatch casesrepresented either a pendant worn around the neck or awatch worn on the wrist; the cell phone cases represented aphone suspended on a necklace around the neck, inside anarm band, or in a waist holster; and the shoe-size objectrepresented a device that could be worn on the ankle sincesuch devices would not typically be placed on other parts ofthe body.

5.2 Power Estimation Procedure

We have developed the following procedure for using themeasured acceleration data set, in combination with theVDRG model, to estimate the available power (see Fig. 5).First, an acceleration data set for an entire day is dividedinto 8;640 ð¼ 24 hr � 60 min=hr � 60 sec=min � 1=10 secÞten-second interval fragments. Each fragment is fed intothe generator model so that the displacement of thegenerator frame (i.e., the accelerometer container) yðtÞ canbe obtained by double integrating the acceleration signal.Subsequently, the relative displacement zðtÞ and derivativedzðtÞ=dt of the proof mass can be calculated from yðtÞ. Asmentioned in the previous section, the electrical powergenerated in the inertial generators can be represented bythe power dissipated in the mechanical damper shown inFig. 1. Accordingly, the energy dissipated in the damper,Ed is:

Ed ¼Z Z2

Z1

F � dz; ð3Þ

where F ¼ D _z 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 ¼Z Z2

Z1

D _z � dz ¼Z T

t¼0

Ddz

dt� dzdtdt ¼ D

Z T

t¼0

ð _zÞ2dt: ð4Þ

The average power during time interval T ¼ 10 sec is

Paverage ¼1

TEd ¼

D

T

Z T

t¼0

ð _zÞ2dt: ð5Þ

By applying this procedure to all fragments successively,the average power value for every 10-second fragment isobtained. The power calculation procedure has been

implemented together with the VDRG model in MATLABSimulink.

Because the accelerometer modules, we deployed in ourexperiments recorded the accelerations along X, Y , and Zaxes, we assumed that a practical VDRG unit would be builtwith a single axis in alignment with one of the axes of theaccelerometers. The preferred orientation of the VDRG isthen chosen based on which one of the three accelerometeraxes yields the maximum output power. However, theorientation of the mobile devices mounted on the humanbody would not necessarily be constant over an entire day.Moreover, the subjects may wear mobile devices in differentways. For example, people wear wristwatches in differentways (right/left hand or with the face up/down), and theorientation of a cell phone in a waist-band holster might be


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.

different according to the form factor and position of theholster on the belt.

5.3 Optimization and Results

Given m and Zmax as derived for each of the selected mobiledevices, and K chosen based on the FFT analysis for eachharvester configuration, we optimized the power harvestersby searching for the optimal damping coefficient D whichmaximizes average output power P . P then represents theelectrical power generated while the subjects wore the datacollection unit over the entire measurement time. First, weoptimized the power harvesters on a per individual basis.We concatenated the acceleration data sets for three daysfor each of our eight subjects to a single time series, andthen performed FFT analysis on it. By doing so, we wereable to obtain a single dominant frequency fo or springconstant Ko for three days of activities for a givenindividual. After optimizing the damping coefficient Do interms of the selected spring constant Ko, the powerharvester would work well for a given individual. It is stillpossible to design a power harvester optimized on a givenday, and its efficiency of harvesting electrical power wouldobviously be better than that of the harvester optimizedover three entire days on a given individual. However,generally supposing K or D to be held constant in a realpower generator, it is not practical to use different K or Dvalues in the harvesters on given days.

With the harvesters optimized for the given configurationfor each of our subjects, we are able to estimate the electricalpower for a given form factor. However, even though theharvesters have been theoretically optimized, other practicalquestions may need to be addressed. First, we must estimatethe expected mechanical-to-electrical conversion efficiencyof the harvester, in terms of useful electrical power obtainedas a fraction of the amount of power dissipated in the

damper. Although the dissipated power represents thegenerated power of the harvester in Fig. 1, the two valuesmight differ due to the mechanical aspects of implementa-tion and electrical circuitry efficiency. Therefore, by apply-ing the characteristics of the PMG 27 [13] harvester to oursimulation method and then comparing the result to thedata sheet, we estimated that the typical mechanical toelectrical conversion efficiency would be around 20 percent.Consequently, we assumed that only 20 percent of theenergy dissipated in the harvester’s damper would beprovided to the electronics it was powering. Second, weassumed that all energy generated over the day could bestored in some form of storage and would be used to powerwearable electronics. Our last assumption is that thedirection of the power harvester in real objects could becontinuously aligned with the axis that generates themaximum output for a whole day. Based on these assump-tions and the simulation results, the average electrical powerexpected from a VDRG embedded in each object was155� 106 �W, 1:01� 0:46 mW, and 4:9� 3:63 mW for thewristwatch, the cell phone, and the shoe, respectively,depending upon the body location on which the powerharvester was mounted. Table 1 illustrates the averageoutput power generated from the harvesters optimized for agiven individual, and corresponding dominant frequencyand damping coefficient. We also provide a graphicalanalysis to show the feasibility of using the harvested powerto power various mobile electronics at different simulationconditions (see Appendix).

From the FFT analysis, we have shown that the dominantfrequencies strongly depend on the subject’s activities on agiven day. In contrast, from Table 1 we observed thatdamping coefficients vary according to the value of the proofmass of a VDRG embedded in real objects. The mean valueand standard deviation of the dominant frequencies for six


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

simulation conditions are graphically shown in Fig. 6. It isimportant to note that the dominant frequencies in thesimulation conditions C2 and C6, a power harvester on thewrist and a shoe-harvester located on the ankle, distributeevenly (i.e., within a small bandwidth, at approximately 1and 2 Hz, respectively) when compared to other cases,probably because several ambulatory movements of thewrist and ankle are taking place during individuals’activities in daily life. This suggests the potential ofdesigning an optimal power harvester that would be tunedfor a given location and work well enough for an entiresubject pool, though some variation would obviously appearin the performance of the harvester across subjects. For othercases, we could not find any particular pattern for the

dominant frequencies, so we expect that harvesters opti-mized with a single K-D combination for all subjects wouldnot work well for those simulation conditions in practice.

As mentioned previously, we could design a powerharvester with a single K-D combination for a givenlocation for all of our subjects. The theoretical methodwould be analogous to that used in optimizing the harvest-er for a given individual. We can first concatenate the dataof all the days into a single time series; then we can performFFT analysis; and finally we might be able to obtain a singledominant frequency for a given location. After optimizingD with the single K value, a harvester for all subjects can bedesigned for a given location. However, this optimizationmethod requires a significant amount of computingrequirements, in particular for performing FFT analysis on24 days of acceleration data sets (166 million floating-pointvalues). As an alternative, we chose a simpler way to find asingle K-D combination for all eight subjects by calculatingthe average of K-D values, excluding their highest andlowest values. Based on the six average K-D valuescalculated corresponding to each of the simulation condi-tions, we are able to estimate a new average output powergenerated from a harvester having a single K-D combina-tion for a given location. Fig. 7 shows the energy generatedwith new harvesters as a fraction of the amount of energygenerated from the harvesters optimized on a givenindividual. The new power harvester on the wrist showsapproximately 100 percent power recovery given the evendistribution on the dominant frequency. A shoe-harvesteralso shows more than 90 percent recovery, compared to theharvester optimized on a given individual. However, thepower from a harvester on the arm also shows goodrecovery (about 95 percent) even though the distribution ofdominant frequencies illustrated in Table 1 does not look


TABLE 1Generated 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.

Fig. 6. Mean and standard deviation of the dominant frequencies on thebody locations: C1-A watch hanging around the neck, C2-A watch on thewrist, 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 inertialharvesters may be less influenced by changing the resonantfrequency than other cases, which might be explained bylooking into the spectrum on the arm where large peaksrarely appear unless individuals are moving their armsvigorously, e.g., running. In contrast, three cases C1, C3,and C5 show considerable reduction in the amount of theenergy recovered with the new harvesters, which suggeststhat the harvesters optimized for the given simulationconditions C1, C3, and C5 would not work well for allsubjects, and we would have to investigate anotherharvesting technique. In the next section, we discuss anew type of power harvester-an adaptively-tunable VDRG-based inertial generator.

6 TUNABLE INERTIAL HARVESTER

The most important feature of vibration-driven resonantgenerators is that the output power harvested with thosegenerators strongly depends on how well the resonantfrequency of the generator is matched with the frequency ofexcitational vibration from its environment. However, thespectrum of the environmental vibration does not show anideally sharp impulse due to the variation of source vibrationor noise distortion. Moreover, as our FFT analysis shows inSection 4.3, the spectra of the acceleration waveformscollected from daily human activities can be represented asa combination of sinusoidal components, which are difficultto harvest at the same time, but rather would have to begarnered by changing the resonant frequency of the inertialgenerator. Thus, to improve the efficiency of vibration-driven resonant generators under the variable vibrationalsources over time (e.g., everyday human motion), we discusstunable inertial harvesters with the assumption that thespring constant K and damping coefficient D could beadjusted mechanically or electronically.

We divide an acceleration data set for an entire day into aseries of small time fragments. By successively searching fora dominant frequency and an optimal damping coefficientfor 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 indivi-

duals 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. How-

ever, 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 environ-

ment 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 Lenz’s 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 constantD. 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 implementa-

tion. 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 damp-

ing-tunable harvester would be reasonable in practice unless

changing the load impedance of the generator coils requires

too much electrical power.


Fig. 7. The energy generated with the harvesters optimized for a givenlocation: 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-Aphone on the waist, and C6-A shoe on the ankle.

Fig. 8. Searching for a dominant frequency and damping coefficient for asmall time interval fragment.

In order to estimate the achievable power with thedamping-tunable harvester from our subjects, we fixed thedominant frequency that we obtained in optimizing on agiven individual and searched for optimal dampingcoefficient for every fragment. The left column in Table 2shows the average improvement on the amount of theenergy generated with the tunable harvester when com-pared with the harvester optimized on a given individual.The tunable harvester slightly improves harvester efficiencyin most cases, with the exception of the watch hangingaround the neck.

6.2 Frequency-Tunable Harvester

Damping-tunable harvesters do not retain their efficiency asthe source frequency shifts far from the resonant frequencyof the generator. To address this case, we propose alteringthe resonant frequency of the power harvester by adaptingthe spring constant of the spring suspending the proof massinside a VDRG. Several studies have been devoted to thetheory of an adaptive frequency-tunable harvester. Roundyand Zhang presented a frequency-tuning technique to alterthe resonance frequency by applying an electrical voltageinput across a piezoelectric bimorph, which results in achange in the stiffness of the bimorph [15]. However, thismethod requires a continuous source of input power foractive tuning of the beam, reducing the efficiency of theenergy harvesting device. Leland and Wright showedanother technique that adapts the resonant frequency ofthe device by applying an axial compressive load andaltering the stiffness of the supported beam [17]. Malkin andDavis demonstrated a multifrequency piezoelectric energyharvester consisting of a number of cantilevered beams,which generates electrical energy over a larger frequencyrange [16]. However, the expected efficiency of such deviceswould obviously be reduced because a large subset of beamsin the generator will not be in resonance at a given frequency.Recently, Challa et al. showed the design and testing of afrequency tunable energy harvesting device using a mag-netic force technique, in which the applied attractive andrepulsive magnetic force induces an additional stiffness inthe system and alters the overall effective stiffness and thecorresponding resonant frequency of the device [18]. With aprototype device with a 26 Hz resonant frequency, theyshowed that the magnetic force technique enables resonancetuning to �20 percent relative to an untuned resonantfrequency (26 Hz) and a continuous power output over theentire frequency range tested (22-32 Hz). They also pre-sented another magnetic-coupling tunable generator where

either higher power density or a wider tunable frequencyrange is achievable using a new geometrical structure-twoindependent single degree of freedom cantilever beams [26].Even though they showed a piezoelectric technique for theresonant frequency tunable energy harvesting device, theyreported that their tuning technique can be applied to any ofthe three energy harvesting techniques including electro-magnetic generators, which we consider in this study. Thus,we call this harvester a frequency-tunable harvester whosespring constant is adjustable for every time fragment, withan adaptive damping coefficient.

From the frequency domain analysis, we have shownthat the dominant frequency strongly depends on theindividual’s activities. Thus, adapting the spring constantappropriately for every time fragment implies that thepower harvester would work well for all kinds of activitiesof all individuals. To estimate the achievable power withthe frequency-tunable harvester from our subjects, wefound the dominant frequency for each fragment usingthe same method described in the FFT analysis section, andsearched for the optimal damping coefficient. The result isshown in the right column of Table 2. The results from awatch hanging around the neck, a phone hanging aroundthe neck, and a phone on the waist show considerableimprovement with the tunable-spring technique. This isprobably because they are hanging objects which are notfastened on the human body, thus forming a pendulum. Inthis scenario, many large peaks may occur in the spectrumdue to cross coupling between excitation from humanmotion and the pendulum’s period. In contrast, the powergenerated on a watch on the wrist shows no improvementbecause most of the energy may be garnered within a smallbandwidth as shown in Fig. 6, and changing the resonantfrequency does not have a useful effect.

7 DISCUSSION

Although we have theoretically designed a tunable inertialharvester for these simulation conditions, this harvestingmethod would not be directly deployed as a practicalharvester. This is because the updating procedure needs tobe performed in real time and the prior analysis for everytime fragment is necessary for adjusting parameters of thefollowing fragment, as illustrated in Fig. 9. In a practicalscenario, we require a causal mechanical system. A simpleimplementation would be both causal and memoryless,where past human activity is not considered in setting theharvester’s parameters in a given time instant. Givenacceleration data during time interval Ti�1, we try to searchfor dominant frequency components and select a resonant


TABLE 2Average Improvement on the Amount of the Energy

Generated with the Ideal Tunable Harvester when Comparedwith the Harvester Optimized on a Given Individual

Fig. 9. Predicting a dominant frequency and damping coefficient for thefollowing acceleration fragment.

frequency fi�1. Next, based on fi�1, we can updateKi andDi,which will be used for the next time fragment Ti. In ourexperiment, we have set T to 10 seconds because oursimulation model implemented in MATLAB Simulink wasdesigned with 10-second simulation interval. Table 3 showsthe average improvement on the amount of the energygenerated with the practical tunable harvester when com-pared with the harvester optimized on a given individual.

In the purely-causal case without memory, we observe areduction in the amount of harvested energy. One possiblereason for the reduction is the 10-second time interval whichmight be too long to appropriately predict the harvesterparameters for each of all time fragments. Accordingly, wemight be able to use shorter time intervals for betterprediction. However, there are several important issues weshould carefully consider when shortening the time intervalfor adjusting physical parameters of VDRG harvesters. First,shortening the time interval will increase the number ofcalculations for FFT processing as well as the timingconstraint within which the measured signal should beanalyzed and the optimal parameters should be estimated,causing an increased computing power overhead. Depend-ing on the energy cost of computation in a particularadaptive harvester, compute energy cost may outweigh theimprovement in harvested energy. Second, since humanbody movements are usually located in low frequencybandwidth (0.5 Hz to 4 Hz), shown in Table 1, there are only5 to 20 cycles included within the 10-second time window.Accordingly, as the time interval continues to shrink, therewill be fewer cycles located, and we could not expect highfrequency resolution in the frequency domain analysis.

For the overall efficiency of energy harvesting system,the most critical feature of the tunable harvester is theactuating power for adapting damping coefficient andspring constant. Roundy and Zhang demonstrated twotypes of frequency tuning approaches: active tuning andpassive tuning. Active tuning can be accomplished byactive actuators that are on all the time to change theresonant frequency of the generator in real time. Incontrast, passive tuning can be accomplished by actuatorsthat turn on to tune the resonant frequency of the generatorand then turn off while keeping matched the newly tunedfrequency with the excitational frequency [27]. For ourtunable harvester, we could deploy the passive tuningtechnique; from the FFT analysis a microcontroller firstobtains a dominant frequency and an optimal dampingcoefficient for a 10-second time fragment; and then, if

needed, the controller turns on and drives the actuator toadjust the stiffness and electronically-induced resistance sothat the power harvester could be appropriately tuned tothe excitational vibration; finally, the controller turns off theactuator and waits until the next 10-second interval is puton the queue of FFT analysis. This self-tuning mechanismenables the tunable power harvester to autonomouslycontrol its resonant frequency and damping coefficient toaccount for the change in the vibration characteristics overtime. However, we would have to very carefully considerthe overall energy cost so that the amount of energyconsumed for analyzing acceleration data and driving thetuning actuator does not overwhelm the electrical energyharvested during the preceding 10-second interval.

As shown in Table 1 and Fig. 6, all of the dominantfrequencies of the six simulation conditions are between0.5 Hz and 5 Hz, and thus the frequency band (0.5 Hz and5 Hz) would be of interest for the frequency-tunableharvester. If we consider the magnetic force techniquepreviously described in the frequency-tunable harvestersection [18], the controllable range of resonant frequencywould be 1.6 Hz to 2.4 Hz, assuming that the untunedresonant frequency is 2 Hz and �20 percent of untunedresonant frequency is tunable by the technique. Althoughthe tunable frequency range is not wide enough to coverthe frequency range necessary for our harvesters, we canimagine achieving a full coverage of the frequency bandof our interest by applying the advanced techniquepresented in [26].

8 CONCLUSION

We have presented the first 24-hour in situ, continuousstudy of inertial power harvester performance from every-day activities. We collected realistic, long-term accelerationdata from eight experimental participants performingeveryday activities for three days each. We discussed thedesign of a specific type of inertial harvester, known as avelocity damped resonant generator. Based on the accel-eration data collected from six body locations from each ofour eight participants and the first-principles numericalmodel of VDRG implemented in MATLAB, we estimatedthe amount of energy that can be garnered at thoselocations. The output power from the simulation resultshows that it is practical to continuously operate motion-powered wireless health sensors and other low-powerdevices, and that motion-generated power may providean intermittent power source for more power intensivedevices like an MP3 player or cell phone. For optimaldesign of VDRGs that will be embedded into real mobiledevices, we considered selecting optimal spring constantand damping coefficient of VDRGs, depending upon formfactor and placement of the mobile devices. Our resultsshow that it is feasible to develop inertial power harvestersoptimized on a given location as well as a given individual,though some variation appeared in the performance of theharvesters across the six simulation conditions. We alsodiscussed the potential of designing a tunable powerharvester whose resonant frequency can be automaticallyadjusted to the vibrational excitation of its environment.With the advent of new low-power integrated circuits and


TABLE 3Average Improvement on the Amount of the Energy Generated

with the Practical Tunable Harvester when Comparedwith the Harvester Optimized on a Given Individual

sensors, or the development of adaptively tuned VDRG, wecan imagine supporting a larger number of devices in thenear future.

APPENDIX

POWERING MOBILE ELECTRONICS

Given the power garnered from humans, we are able toimagine powering devices such as consumer electronicsand self-sustaining body sensor networks. The tappedpower can be an alternative to batteries for consumerelectronics such as watches, MP3 decoders, cell phones, andGPS (global positioning system) receivers. Another possibleapplication of human power harvesting is to powerwearable sensors for monitoring human vital signs.Monitoring of the status of human health could play akey role in the future health care services for diseaseprevention as well as the treatment for chronically illpatients. The required power for various electronics hasbeen tabulated and presented in Table 4. Thus, from thegenerated power in Table 1, we can determine whether thewearable electronics in Table 4 can be powered for a givensubject on a given day, depending upon the simulationconditions. Fig. 12 shows the result of using the garneredoutput power to power different wearable electronics atvarious parts on the body. In a number of cases, the outputpower is insufficient to continuously power the highestdemanding electronics such as the MP3 decoder, the RFreceiver, and the GPS receiver. However, the VDRG can beused to charge a battery for intermittent operation of thosedevices. One can imagine using the garnered power tocharge a battery when a standard recharging power sourceis not available.

As mentioned above, a self-sustaining body sensornetwork for monitoring the status of human health is oneof the promising applications of power harvesting technol-ogy. A generic architecture of a self-powered sensor node isshown in Fig. 10. Since the basic operation of a wirelesssensor node is to measure raw data, convert it to a digitalvalue, and wirelessly transmit the digital sensory data to a

central processing device, a single wireless sensor nodewould be composed of a sensor, an A-D converter, and a RFtransmitter. For a self-sustaining sensor node, the powerharvested from inertial generators would have to be largerthan the total amount of power consumed in the node,which is the sum of the power consumed in eachcomponent in Fig. 10. It was shown that 1 kbit/s is enoughfor body sensor data transmission [6], and this can beprovided with sub-�W power level within 1 m distance[41]. In addition, A-D conversion can be achievable with1 �W [40]. Thus, with these ultra-low power-consumingcomponents, the most important power usage in a self-powered node is for sensors presented in Table 4. Fig. 11shows various wearable sensors and the power require-ments for monitoring human vital signs: temperaturesensor [37], PPG (photoplethysmograph) sensor for mon-itoring blood pressure, heart rate, and respiration rate [33],3-axis accelerometer [36], humidity [34] and pressuresensors [35], and a central processing device composed ofa micro controller, RF receiver [29], and memory [39]. Fromthe output power in Table 1, a temperature sensor node canbe continuously operated from the power harvested from aphone on the upper arm. In a PPG sensor node, its powerrequirement exceeds the power garnered in a wristwatchon the wrist, but the power shortage might be able to besupplied from the power on the upper arm. While anaccelerometer node works well with the power from aphone-size harvester on the waist, the 6D motion sensorwould not be able to be powered. A pressure sensor andhumidity sensor could be continuously powered in a shoe-size power harvester. We imagine that a central processingdevice would be located and powered in a shoe since ashoe-size harvester shows the largest amount of outputpower among six simulation conditions. However, the RFreceiver chip may not be powered from our results, and thepower shortage could be mitigated in several ways. First,


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

TABLE 4Required Power for Consumer Mobile

Devices and Health Sensors

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

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

because the output power of the harvester is proportionalto the value of the proof mass regardless of the accelerationwaveform and the architecture of the harvester, we canincrease the output power with a heavier proof mass in theharvester. Second, we could build a specific RF receiversatisfying the power requirement using off-the-shelf com-ponents. As new low-power technologies emerge, we canimagine many more senors being able to be supported bythis harvesting approach. Thus, we begin to see the

feasibility of developing wearable electronics powered

from normal everyday activities.

REFERENCES

[1] J. Paradiso, “Systems for Human-Powered Mobile Computing,”Proc. Design Automation Conf. (DAC ’06), pp. 645-650, 2006.

[2] J. Paradiso and T. Starner, “Energy Scavenging for Mobile andWireless Electronics,” IEEE Pervasive Computing, vol. 4, no. 1,pp. 18-27, Jan.-Mar. 2005.


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 ¼ 24 days), and the Y-index represents the wearable electronics considered in Table 4.

[3] T. Starner, “Human-Powered Wearable Computing,” IBM SystemsJ., vol. 35, nos. 3/4, pp. 618-629, 1996.

[4] T. Starner and J. Paradiso, “Human-Generated Power for MobileElectronics,” Low-Power Electronics Design, pp. 1-35, CRC Press,2004.

[5] W. Huynh, J. Dittmer, and A. Alivisatos, “Hybrid Nanorod-Polymer Solar Cells,” Science, vol. 295, no. 5564, pp. 2425-2427,2002.

[6] E. Yeatman, “Advances in Power Sources for Wireless SensorNodes,” Proc. Workshop Body Sensor Networks (BSN ’06), pp. 20-21,2004.

[7] M. Strasser, “Miniaturized Thermoelectric Generators Based onPoly-Si and Poly-SiGe Surface Micromachining,” Sensors andActuators A: Physical, vols. 97-98, pp. 535-542, 2002.

[8] R. Amirtharajah and A. Chandrakasan, “Self-Powered SignalProcessing Using Vibration-Based Power Generation,” IEEE J.Solid-State Circuits, vol. 33, no. 5, pp. 687-695, May 1998.

[9] J. Paradiso and M. Feldmeier, “A Compact, Wireless, Self-Powered Pushbutton Controller,” Proc. Int’l Conf. UbiquitousComputing (Ubicomp ’01), pp. 299-304, 2001.

[10] N. Shenck and J. Paradiso, “Energy Scavenging with Shoe-Mounted Piezoelectrics,” IEEE Micro, vol. 21, no. 3, pp. 30-42,May/June 2001.

[11] A. Chapuis and E. Jaquet, The History of the Self-Winding Watch.Roto-Sadag, 1956.

[12] S.R. Vetorino, J.V. Platt, and D.A. Springer, “Renewable EnergyFlashlight,” Patent and Trademark Office, 2001.

[13] PMG �-generator, http://www.perpetuum.co.uk, 2011.[14] VEH-360 electromechanical generator, http://www.ferrosi.com/

energy-harvesters.html, 2011.

[15] S. Roundy and Y. Zhang, “Toward Self-Tuning AdaptiveVibration-Based Microgenerators,” Proc. SPIE , pp. 373-384, 2005.

[16] M.C. Malkin and C.L. Davis, “Multi-Frequency PiezoelectricEnergy Harvester,” J. Acoustics Soc., vol. 118, no. 1, pp. 24-24, July2005.

[17] E.S. Leland and P.K. Wright, “Resonance Tuning of PiezoelectricVibration Energy Scavenging Generators Using Compressive AxialPreload,” Smart Materials and Structures, vol. 15, no. 5, pp. 1413-1420, Oct. 2006.

[18] V.R. Challa, M.G. Prasad, Y. Shi, and F.T. Fisher, “A VibrationEnergy Harvesting Device with Bidirectional Resonance Fre-quency Tunability,” Smart Materials and Structures., vol. 17, no. 1,Feb. 2008.

[19] P. Mitcheson, T. Green, E. Yeatman, and A. Holmes, “Architec-tures for Vibration-Driven Micropower Generators,” IEEE J.Microelectromechanical Systems, vol. 13, no. 3, pp. 429-440, June2004.

[20] T. Buren, P. Mitcheson, T. Green, E. Yeatman, A. Homes, and G.Troster, “Optimization of Inertial Micropower Generators forHuman Walking Motion,” IEEE J. Sensors, vol. 6, no. 1, pp. 28-38,Feb. 2006.

[21] L. Rome, L. Flynn, E. Goldman, and T. Yoo, “GeneratingElectricity while Walking with Loads,” Science, vol. 309, no. 5741,pp. 1725-1728, 2005.

[22] J. Donelan, Q. Li, V. Naing, J. Hoffer, D. Weber, and A. Kuo,“Biomechanical Energy Harvesting: Generating Electricity duringWalking with Minimal User Effort,” Science, vol. 319, no. 5864,pp. 807-810, 2008.

[23] R. Yang, Y. Qin, C. Li, G. Zhu, and Z.L. Wang, “ConvertingBiomechanical Energy into Electricity by a Muscle-Movement-Driven Nanogenerator,” Nano Letters, vol. 9, pp. 1201-1205, Feb.2009.

[24] N. Bharatula, S. Ossevoort, M. Stager, and G. Troster, “TowardsWearable Autonomous Microsystems,” Proc. Int’l Conf. PervasiveComputing (Pervasive ’04), pp. 225-237, 2004.

[25] C. Williams and R. Yates, “Analysis of a Micro-Electric Generatorfor Microsystems,” Sensors and Actuators A: Physical, vol. 52, pp. 8-11, 1996.

[26] V.R. Challa, M.G. Prasad, and F.T. Fisher, “High Efficiency EnergyHarvesting Device with Magnetic Coupling for ResonanceFrequency Tuning,” Proc. SPIE, Mar. 2008.

[27] S. Roundy, E.S. Leland, J. Baker, E. Carleton, E. Reilly, E. Lai, B.Otis, J.M. Rabaey, P.K. Wright, and V. Sundararajan, “ImprovingPower Output for Vibration-Based Energy Scavengers,” IEEEPervasive Computing, vol. 4, no. 1, pp. 28-36, Jan.-Mar. 2005.

[28] Samsung Helix, Nexus and Pioneer Inno Power Consumption,http://www.orbitcast.com/archives/samsung-helix-nexus-and-pioneer-inno-power-consumption.html, 2006.

[29] ASK Single Conversion Receiver TDA 5200, http://www.infineon.com/cms/at/product/channel.html?channel=ff80808112ab681d0112ab6909cd00de, 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 CMOSWireless MEMS Motion Sensor for Physiological Activity Mon-itoring,” IEEE Trans. Circuits Systems, vol. 52, no. 12, pp. 2539-2551,Dec. 2005.

[32] Cell Phone Power Management Requires Small Regulators withFast Response, http://www.commsdesign.com/printableArticle/?articleID=12801994, 2002.

[33] S. Rhee, B.-H. Yang, and H. Asada, “Artifact-Resistant, Power-Efficient 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 forMedical 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-LowPower Data Storage for Sensor Networks,” Proc. Conf. InformationProcessing in Sensor Networks (IPSN ’06), pp. 374-381, 2006.

[40] J. Sauerbrey, S. Landsiedel, and R. Thewes, “A 0.5-v, 1-�wSuccessive 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 degreesin mechatronics from the Gwangju Institute ofScience and Technology, South Korea, in 1999and 2006, respectively. Currently, he is workingas a senior researcher in the U-embeddedConvergence Research Center at the KoreaElectronics Technology Institute (KETI), Seong-nam, South Korea. He was a research scientist/postdoctoral researcher with the UbiquitousComputing Research Group in the College of

Computing and the GVU Center at the Georgia Institute of Technologyfrom 2006 to 2009. His research interests are in ubiquitous computing,human-computer interaction, and context-based interactive environ-ments. He is particularly interested in human activity sensing and itsapplications.

Shwetak N. Patel received the BS and PhDdegrees in computer science from the GeorgiaInstitute of Technology (Georgia Tech) in 2003and 2008, respectively. Currently, he is workingas an assistant professor in the Departments ofComputer Science and Engineering and Elec-trical Engineering at the University of Washing-ton. His research interests are in the areas ofsensor-enabled embedded systems, human-computer interaction, ubiquitous computing,

and user interface software and technology. He is particularly interestedin developing easy-to-deploy sensing technologies and approaches forlocation and activity recognition applications. He was the cofounder ofUsenso, Inc., a recently acquired demand side energy monitoringsolutions provider. He was also the assistant director of the AwareHome Research Initiative at Georgia Tech. He was a TR-35 awardrecipient for 2009.


Matthew S. Reynolds received the SB, MEng,and PhD degrees from the MassachusettsInstitute of Technology in 1998, 1999, and2003, respectively. Currently, he is working asan assistant professor in Duke University’sDepartment of Electrical and Computer Engi-neering. He is a cofounder of the RFID systemsfirm ThingMagic Inc. and the demand-sideenergy conservation technology firm Zensi. Hisresearch interests include the physics of sensors

and actuators, RFID, and signal processing. He is a member of theSignal Processing and Communications Group and the ComputerEngineering Group at Duke, as well as the IEEE Microwave Theory andTechniques Society.

Gregory D. Abowd received the MS and PhDdegrees in computation from the University ofOxford, United Kingdom, in 1987 and 1991,respectively. He is a distinguished professor inthe School of Interactive Computing at theGeorgia Institute of Technology (Georgia Tech)and the W. George professor and director of theGeorgia Tech/Emory University Health SystemsInstitute. His research interests are in applica-tions-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.


yundesignoptimalinertialenergyharvester11.pdf

Documents

human motion

power harvesters

hz acceleration data

acceleration data sets

human activitylevels

operating power

power reduction

harvested power