Investigating Sensor Data Retrieval Schemes forMulti-Sensor Passive RFID Tags

Yang Su, Asanga Wickramasinghe, Damith C RanasingheAuto-ID Lab, The University of Adelaide, Adelaide SA 5005, Australia

Email: yang.su01@student.adelaide.edu.au,{asanga.wickramasinghe,damith.ranasinghe}@adelaide.edu.au

Abstract—Measuring multiple physical quantities are increas-ingly being demanded in commercial, biomedical and generallyin ubiquitous applications. Although the recent emergence ofpassive sensor enabled RFID tags (sensor tags) provide newopportunities for these types of applications mainly due to theextended operational life and the small form factor, the energyharvesting nature of sensor tags hinders the use of multiplesensors in a single platform because of the requirement ofadditional energy to operate multiple sensors and subsequentreduction in the throughput. In this paper, we propose three,fast and energy efficient multi-sensor data retrieval approachesto obtain sensor data from sensor tags. We implemented a sensortag with two sensors, an accelerometer and a barometer. Ourextensive experiments on power consumption, operational rangeand throughput using the developed sensor tag revealed that, theproposed approaches can successfully be used for multi-sensordata retrieval and indicates that they can effectively be used in arange of real-world ubiquitous sensing applications such as fallprevention and food safety monitoring.

I. INTRODUCTION

Measuring and understanding physical quantities areparamount for commercial [1]–[3] (eg: in cold chain man-agements where temperature is constantly monitored), med-ical applications [4] and more broadly enabling ubiquitousmonitoring [5]. Recent emergence of passive sensor enabledRadio Frequency Identification (RFID) tags (here after referredto as sensor tags) such as the Wireless Identification andSensing Platform (WISP) [6] developed by Intel Researchprovide new prospects to measure physical quantities usinglow cost devices [7] without batteries. Furthermore, suchsensor tags have a small form factor and can even be embeddedpermanently within other equipment as discussed in [6].

Real-world applications demand retrieval of multiple physi-cal quantities. For instance, sensing multiple types of gasesis paramount for monitoring air pollution in different en-vironments [8], [9], monitoring temperature and humiditysimultaneously to ensure food safety [10] and retrieval of baro-metric pressure and movement information for older peoplemonitoring [11] require multi-modal sensing devices.

However, there are two main issues associated with em-bedding multiple sensors in a single passive sensing platform.Firstly, sampling multiple sensors require an excessive amountof power. For instance, operating multiple analog-to-digitalconverters (ADC) simultaneously place a larger burden onharvested power, especially due to the large peak currentrequirements of ADCs. Even though the sensors are sampled

sequentially the excessive energy consumption may still not besatisfied using harvested energy. This large electrical energyrequirements are detrimental to the effective operation rangeof sensor tags [12] as harvested power is inversely proportionalto the squared distance between the sensor tag and the RFIDreader antenna according to the Friis transmission equation[13]. Secondly, embedding multiple sensors lead to delays dueto the additional sampling time and increased data processingtime that reduce the throughput of sensor tags and, conse-quently, the overall performance of the sensor tag.

Operational range and throughput for sensor tags are impor-tant parameters that typically depend on the application. Forinstance, applications such as fall detection requires a largeroperational range where as gesture recognition applicationsfavor higher throughput to be able to capture highly dynamicmovements. Therefore, in this paper, we investigate threesensor data retrieval schemes, with differing performance char-acteristics, suitable for sensor tags with multiple embeddedsensors. We summarize our contributions below:

• We propose three sensor tag data retrieval schemes formultiple sensors embedded in a single sensor tag: i) se-quential, ii) parallel and iii) alternative. These schedulesdetermine when to sample the sensor and transmit databack to RFID readers.

• We evaluated the proposed schedules using a multi-sensorsensor tag developed based on the WISP by embeddinga barometer as an additional sensor along with theaccelerometer. Our analysis and extensive experimentsrevealed that the parallel schedule has the least delay,but has the highest power consumption; in contrast, al-ternative schedule has the highest delay and demonstratedthe largest operational range while also being the mostenergy efficient multi-sensor data retrieval schedule.

Remaining sections of the paper are organized as follows:Section II discusses related work; Section III presents ourproposed multi-sensor data retrieval approaches; Section IVdetails the implementation; Section V presents the experimentsconducted and results. We conclude the paper in Section VI.

II. RELATED WORKS

There are two common approaches to retrieve sensor datafrom a sensor tag: i) Read-by-Command approach (RBC) andii) Data-in-ID approach (DID). In the RBC approach, sensordata are acquired using the READ command [14] and this

(a) Read sensor data from tag memory (RBC approach)

(b) Embedding sensor data within ID (DID approach)

Fig. 1. Comparison of data retrieval from sensor tags

process is illustrated in the sequence diagram in Fig. 1(a).First, a sensor is sampled and then the data are stored ina non-volatile memory, typically in an Electrically ErasableProgrammable Read-Only Memory (EEPROM) unit of atag. Subsequently, sensor data stored in a tag’s non-volatilememory are read using the READ command by specifyinga pointer to the memory location with the desired data asdefined in the Electronic Product Code (EPC) Air Interfaceprotocol [14]. This approach requires several reader-to-tag-to-reader communication cycles as illustrated in Fig. 1(a). Storingdata into an EEPROM not only consumes considerable power(148 µW for writing [15]) but also takes a considerable amountof time. In case of sensor tags, this approach is limitedby the high power consumption and significant reduction inthroughput.

A recent study by Igarashi et al. [16] proposed a networkarchitecture to acquire data stored in an EEPROM of a sensortag by first defining a structured memory and subsequentlyaccessing a specific memory address with the desired sensordata. In this study the authors proposed a memory addressresolution scheme to reduce the time to locate the interestedsensor data and thereby reduce the overall data retrievallatency. However, their approach is not geared towards energyefficiency.

Alternatively, the Data-in-ID approach was utilized by Alan-son et al. [6] for sensor data retrieval from a WISP. In this

Fig. 2. Tag ID format when sensor data is embedded into the 96-bit.

approach, as shown in Fig.2, portion of the 96 bit EPC is re-placed by sensor data, thereby transmitting data in an inventorysession. Sensor tags such as WISPs which utilize this approachinitially sample the onboard sensors when adequate power isharvested. Upon receiving a QUERY command from an RFIDreader, the sensor tag responds with a 96-bit EPC where thesensor data is embedded within the EPC. Consequently, part ofthe EPC is used to piggyback the sensor data. It is importantto note that, if the total time taken to transmit acquired sensordata and sample on-board sensors exceeds the interval overwhich the RFID reader waits for a response, then the sensortag has to wait for a subsequent QUERY command to transmitthe sensor data; this reduces the throughput of the sensor tag.This approach enables the use of off-the-shelf RFID readers toacquire sensor data, but require special software to de-embedthe data in the ID [17].

Although the WISP developers have included a numberof on-board sensors, such as a capacitance sensor, a voltagesensor, a thermometer and a 3D accelerometer1, samplingmultiple sensors simultaneously have not been considered inthe past. All of these sensors on a WISP are analog compo-nents and are sampled using the Microcontroller’s (MCU’s)internal ADC. The current WISP firmware is designed tosample a single sensor that is specified at the firmware compiletime, hence in the present state, the WISP cannot samplemultiple sensors simultaneously. Therefore, we are motivatedto investigate a fast and energy efficient multi-sensor dataretrieval schemes from sensor tags needed for a wide arrayof real-world applications.

III. MULTI-SENSOR DATA RETRIEVAL

A multi-sensor sampling approach for passive sensorsshould have two desired properties: i) high throughput; andii) high power efficiency. As discussed in Section II, Read-by-Command (RBC) approach, which is defined in the EPCstandards, requires a considerable amount of power and in-curs significant delays. Therefore, in order to reduce powerconsumption and communication delays, we base our multi-sensor data retrieval approaches on the DID approach.

Efficiency increases the operational range while throughputdetermines the quantity of the sampled data. In order toanalyze and compare multi-sensor data retrieval approaches,we quantify the afore mentioned properties as follows. Wedenote the number of sensors embedded in an RFID platformto be n and the sampling time of ith sensor, i ∈ {1 · · ·n}, tobe tsi. We denote the data transmission time for a sensor tag

1http://wisp.wikispaces.com/Wisp+4.1+DL

TABLE ICHARACTERISTICS OF SENSOR SAMPLING SCHEDULES

Schedule Time delay Energy consumptionrate

DedicatedADC

Additionalcontrol logic

Sequential∑n

i=1 tsi + tt En/(∑n

i=1 tsi + tt) optional smallParallel maxi∈{1···n}tsi + tt En/(maxi∈{1···n}tsi + tt) necessary none

Alternative∑n

i=1(tsi + tt) En/∑n

i=1(tsi + tt) optional large

by tt. It is noteworthy that the transmission time is invariant tothe number of sensors as data is embedded in the tag ID (e.g.96 bit EPC) and successful transmission require all the 96 bitsto be transmitted. Therefore, the total time to sample sensorsand the transmission time, Tn, determine the throughput ofeach multi-sensor data retrieval approach. Denote En as thetotal amount of energy required to sample n sensors, we defineenergy consumption rate ρe = En/Tn. The quaintly En onlydepends on the type and the number of sensors embedded in asensor tag and hence for a given sensor tag En is a constant.

The definitions of the desired properties higher throughputand energy efficiency are inversely related and therefore,we need to have a compromise between these properties.To this end, we propose three multi-sensor data retrievalapproaches: i) Sequential; ii) Parallel; and iii) Alternative.These three approaches are characterized by the schedulingof sensor sampling and transmission operations. To avoidconfusion with DID and RBC data retrieval approaches, werefer to the proposed multi-sensor data retrieval approachesas sensor sampling schedules. An illustration of these threesensor sampling schedules are given in Fig. 3 and Table Isummarizes the properties of each sensor sampling schedule.

a) Sequential: This is the most basic schedule wheremultiple sensors are sampled within a single session sequen-tially as shown in Fig. 3(a). After acquiring sensor data fromall the sensors, the data are transmitted as the reply to the firstquery command after sampling.

b) Parallel: In this schedule sensors are sampled simul-taneously as shown in Fig. 3(b) and hence reduces the delay.This delay is constrained by the maximum sampling timeamong all sensors (see Table I); therefore it is invariant to thenumber of sensors embedded in a sensor tag. As all the sensorsare operating simultaneously, this schedule has the highestpower consumption. Consequently, the probability of a sensortag to brownout is higher. Therefore, this schedule is expectedto have a lower operational distance and a lower throughputparticularly when power harvesting is limited. To realize thisschedule each senor requires a separate ADC.

c) Alternative: This approach is illustrated in Fig. 3(c).Sequential and parallel schedules sample all the sensors andthen transmit data from all the sensors, thus have a sin-gle sample-transmit session. In contrast, alternative scheduledistributes the sampling and transmission of data from nsensors into n sample-transmit sessions. In such sample-transmit session, one sensor is sampled and the data pertainingto that sensor is transmitted. Although, this scheme increasesthe delay (see Table I), it reduces the energy consumption rateand subsequently sensor sampling failures, particularly whenthe energy to be harvested is scarce.

Fig. 4. Block diagram of the Multi-Sensor WISP, developed bases on WISP4.1DL.

TABLE IISPECIFICATIONS OF THE SENSORS USED IN MULTI-SENSOR WISP

Sensor Typical sample time Peak currentAccelerometer (ADXL330) 4 ms [18] 1.05 mA [18]

Barometer (MS5611-01BA03) [19] 8.2 ms 1.4 mA

condition: 2.2V supply, 25 C◦, maximum resolution.

When these sensor sampling schedules are considered, itis expected for the alternative schedule to have the greatestoperational range as it has the lowest energy consumptionrate. The parallel schedule is expected to have the highestthroughput as the time to sample sensors is minimum but onlywhen the sensor tag is adequately powered, particularly whenthe sensor tag is near an antenna. The sequential scheduleprovide a compromise between the parallel schedule and thealternative schedule.

IV. IMPLEMENTATIONS

In this section, we first discuss the hardware developmentsand later we discusses the upgrades to the WISP firmwareto implement the three proposed multi-sensor data retrievalschedules.

A. Hardware modification

In order to test the multi-sensor performance, we built anew platform based on the WISP 4.1DL2. As shown in Fig. 4,a new digital barometer is added alongside the existing 3-axis accelerometer. Both sensors acquire power from the MCUGeneral Purpose Input/Output (GPIO) ports while the 3-axisaccelerometer is sampled by the MCU’s internal ADC, but thebarometer uses its own ADC. The digitized barometer data aretransferred to the MCU through a Serial Peripheral Interface(SPI) bus. Specifications of these sensors are given in Table II.All the other parts of the circuit are identical to the originalWISP 4.1DL.

The Fig.5 compares the new Multi-Sensor WISP to theoriginal WISP 4.1DL. Due to the fabrication issues the outerborder of the circuit board was kept and hence our Multi-Sensor WISP is visually larger than the WISP.

B. Sequential schedule

We first sample the barometer and then the accelerometerto realize the sequential schedule. The barometer consumes

2http://wisp.wikispaces.com/Wisp+4.1+DL

(a) Sequential (b) Parallel (c) Alternative

Fig. 3. Proposed sensor sampling schedules

Fig. 5. The Multi-Sensor WISP (upper) and WISP 4.1DL (lower).

Fig. 6. Sequence diagram, showing two sensors running sequentially.

higher power (see Table II) and successful sampling of thebarometer increases the probability of the sensor tag to samplethe less energy consuming accelerometer. The total time tosample both the sensors are approximately 12 ms, which isthe sum of the individual sensor sampling times.

Figure 7 illustrates the format of the tag ID used to transmitdata using the sequential sampling schedule. The data fromthe accelerometer is stored using 30 bits and the data from thebarometer is stored using 24 bits.

C. Parallel schedule

As mentioned in Section IV-A, the accelerometer is sampledby the ADC integrated in the MCU and the barometer is sam-pled by its on-chip ADC. Having two ADCs to sample eachsensor enabled the implementation of parallel sensor samplingschedule. We designed the parallel process as shown in Fig. 8.First, we setup the barometer and then the accelerometer.

.

Fig. 7. Format of the 96-bit EPC tag ID used for sequential and parallelschedules

Fig. 8. Sequence diagram, showing two sensors running in parallel.

The latter is coordinated by soft instructions and completeearly. The barometer takes a longer time (see Table II) andraise a hardware interruption to notify the MCU to receiveresults. Generally, the sensor response time is trivial becauseby specifying interruption priorities any sensor can be sampledas described in the parallel schedule. In this schedule, data issent using the same format as in the sequential schedule whichis shown in Fig 7.

D. alternative schedule

The alternative schedule is similar to the single sensorcase, but each sensor is sampled and its data are transmittedseparately in each session as shown in Fig.9. In our imple-mentation, upon powering of the sensor tag, we first samplethe accelerometer and then the barometer. The sensors arealternatively sampled while the tag is alive and this processrepeats each time the sensor tag is powered.

Figure 10 illustrates the format of the tag ID for the al-ternative schedule. This organization is similar to the tag ID

Fig. 9. Sequence diagram, showing two sensors running alternatively.

Fig. 10. Format of the 96-bit EPC tag ID used for the alternative schedule.

organization of both sequential and parallel schedules, but theunused sensor region is filled with zeros.

V. EXPERIMENT AND RESULTS

In our experiments, we mainly evaluate the throughputand efficiency of each sensor data retrieval approach: i) DIDand ii) RBC and the proposed multi-sensor data retrievalschedules. We conducted four types of experiments: i) timedelay; ii) power consumption; iii) communication range; andiv) read rate.

All the experiments were conducted using an Impinj Speed-way Revolution R420-GX11M reader attached to a 6 dBi cir-cularly polarized antenna. We employed two identical originalWISP 4.1DL sensor tags from the same batch to evaluate theperformance of DID and RBC data retrieval methods. Dueto the manufacturing issues, the efficiency of RF front endvaries for each individual Multi-Sensor WISP tags and henceone multi-sensor tag (see Fig. 5) was used. We reprogram thesame mutli-sensor tag to implement the three sensor samplingschedules for our experiments. We considered acquiring datafrom all the sensors as a successful reading. In the caseof alternative schedule, retrieval of data from both sensors(barometer and accelerometer) sequentially in two sample-transmit cycles were considered as a successful reading.

1) Time delay: We utilized two standard WISPs and pro-grammed them to operate on DID and RBC. For DID ap-proach, we measured the time taken to inventory a sensor tag.In the case of RBC approach, we measured the EEPROM writetime, time taken to inventory the tag, EEPROM read time and

Fig. 11. Power consumption test setup, the digital oscilloscope monitors thevoltage across the energy storing capacitor of WISP

time to read the memory of the sensor tag by specifying theexact memory location (offset: 0) and the required amount ofinformation (2 words which corresponds to 32 bits). During theexperiment, the WISP is powered by an MSP430 debuggerwith a constant 2.5 V Vcc to eliminate timing errors causedby low power sleep and failed EEPROM operations as theEEPROM is sensitive to the voltage level during writing.During the experiment the distance between sensor tags andthe RFID antenna is fixed at 40 cm. It is not possible to directlymeasure the EEPROM writing time using the data from anRFID reader. Therefore, we modified the WISP firmware toraise a GPIO pin on sensor data retrieval and lower it upona backscatter event. The fluctuation on the GPIO pin can beeasily captured by a logic analyzer or a digital oscilloscope toobtain the required measurements.

2) Power Consumption: As specified in WISP circuitschematic, the harvested energy is stored in a 10µF monolithiccapacitor. The electrical energy in a capacitor with capacitanceC is proportional to the square of capacitor’s terminal voltage,Vcap, as given by Ecap = 1

2CV2cap. According to the descrip-

tion in Sample et al. [6], the hardware voltage supervisormonitors Vcap and generates an interrupt when Vcap > 2 Vto switch the MCU from Power Save Mode to Active Mode.The system then starts sampling sensors and RFID protocolrelated calculations. When the required power is less thanthe harvesting power, the energy storing capacitor is drainedto provide the deficit and Vcap varies as the WISP processinto sensor sampling event. Therefore, measuring the voltagedifference from MCU become alive (when Vcap > threshold)and the end of sample-transmit session (end of backscatter)will give a quantitative indication of power consumption. Thisvoltage variation can be easily measured as shown in Fig. 11 byusing a Digital Oscilloscope internal math function 2V −Vmin

and we take the mean value over 256 session.3) Operational Range: This test is designed to measure

the maximum operational distance of the sensor tag underDID and RBC configuration as well as the proposed sensorsampling schedules. The experiment is conducted inside ananechoic box as shown in Fig.12. Initially, we placed a sensorat position 0 cm from the reader antenna and then the tag isgradually moved away from the antenna. When the intervalbetween two sensor readings exceed 5 s, we consider thisdistance to be the maximum operating range. In order to avoid

Fig. 12. Range test setup, view towards the reader antenna.

TABLE IIIDELAY TEST RESULT FOR RBC AND DID APPROACHES SHOWING THE

STATISTICS FOR EACH MEASURABLE STAGE.

Approach Session Command / Response Time(ms) Total time(ms)DID Inventory Query (Rx)

RN16 (Tx)7.35 ± 3.92 7.35 ± 3.92Ack (Rx)

ID (Tx)

RBC W.EEPROM – 9.82 ± 0.12Inventory Query (Rx)

8.45 ± 3.27

21.41 ± 3.29

RN16 (Tx)Ack (Rx)ID (Tx)

R.EEPROM – 0.287 ± 0.009Read ReqRN (Rx)

2.845 ± 0.077Rn16 (Tx)Read (Rx)Data (Tx)

any possible interference, e.g. multi-path reflection and powerbackscatter from another WISP, measurements related to eachWISP were taken separately from other WISPs.

4) Read rate: Read rate directly measures the throughputof the system. A higher read rate is indicative of sensor tag’sability to capture detailed information in terms of timelinessand resolution. This experimental setup is similar to the delaytest where the tag is fixed at 1 m from the RFID reader antennainside the anechoic box. Here, sensor tags were powered bythe harvested energy unlike in delay evaluation. We use an in-house developed data logging application to obtain readingsand subsequently to calculate the read rates.

A. Data-in-ID approach versus Read-by-Command approach

Here, we measured the throughput for sensor data retrieval,energy consumption and read range. The main goal of theseexperiments is to empirically compare and contrast the DIDand RBC approaches.

Table III show the results for the time delay experiment foracquiring acceleration data using DID and RBC approaches. Itis clear from this table that the latency for RBC is on average21.41 ms, approximately 3 times that of DID approach. There-fore, using DID approach is unfavorable for high throughputapplications.

Table IV lists the results for power consumption test andread range test for DID and RBC approaches. From theseresults we can see that there is a considerable change in∆Vcap and hence in ∆E in case of RBC. This suggests

TABLE IVPOWER CONSUMPTION AND OPERATIONAL RANGE FOR RBC AND DID

APPROACHES

Approach DID RBC∆Vcap (mV) 400 660∆E (µJ) 0.80 2.72Range (cm) 188 65

that there is a higher probability for the WISP to brownoutwhile RBC is used compared to when DID is used. Thehigh electrical energy requirement for RBC will adverselyimpact the communication range, mainly because at a higherdistance the harvested power is insufficient to support the RBCapproach. As shown in Table IV, the operational range of DIDis approximately three fold higher than RBC.

From these experiments, i.e. delay, power consumption andoperational range, it is evident that the DID performed betterthan RBC approach. Therefore, DID approach can be used asan energy efficient multi-sensor data retrieval approach.

B. Proposed multi-sensor data retrieval schedules

We conducted three experiments: i) power consumption;ii) communication range; and iii) read rate to analyze the threeproposed multi-sensor sampling schedules based on the DIDapproach.

1) Power consumption: We conducted the power consump-tion experiment as described earlier in Section V-2. Figure 13illustrates the variation of terminal voltage of the energystoring capacitor, Vcap, and the backscatter events for eachmulti-sensor data retrieval schedule. From the combination ofthe two measurements, we can infer the internal state of theWISP. In Fig. 13, successful sampling of the accelerometercan be identified from the stylized three-peak saw tooth. Thissaw tooth reflects the successive sampling of the three axesof the accelerometer. A successful sample-transmit session ofthe barometer is indicated by a large slope which correspondsto a drawing of more power. Backscatter event suggests asuccessful transmission of sensor data.

When the Multi-Sensor WISP operates in the sequentialschedule, (see Fig. 13(a)) there are some failed readings due toinsufficient power or a SPI transmission error. For example, attime ≈ 0.4 s the MCU failed to receive the interruption fromthe barometer, then the system was reset as a consequence ofthe Watchdog Timer (WDT) timeout. Figure. 13(a) shows areset at time ≈ 0.3 s caused by a low voltage as a result ofthe barometer drawing a large current. Similar situations canalso be observed at time ≈ 0 s and time ≈ 0.45 s where theaccelerometer was drawing excessive power while reading thesecond axis resulted in system resets.

In case of the parallel schedule shown in Fig. 13(b), we cansee that there are larger voltage drops due to the simultaneoussampling of the accelerometer and the barometer. There arefour successful retrieval-report sessions. At time ≈ 0.15 s thesystem was reset due to an SPI error. The label 3© showsan invalid session. After the accelerometer data is retrieved,the barometer failed to respond the SPI read command and

1.5

1.6

1.7

1.8

1.9

2

2.1

2.2

2.3

2.4

2.5

0 0.10 0.20 0.30 0.40 0.50 0.60

Back

scat

ter e

vent

(on\

off)

Volta

ge a

cros

s en

ergy

sto

ring

capa

cito

r (V)

Time (s)

VcapThresholdBackscatter event

①① ①①

②

②②② ② ③

④④ ①④

②④④

(a) Sequential

③

②槐

1.5

1.6

1.7

1.8

1.9

2

2.1

2.2

2.3

2.4

2.5

0 0.05 0.10 0.15 0.20 0.25 0.30 0.35 0.40

Back

scat

ter e

vent

(on\

off)

Volta

ge a

cros

s en

ergy

sto

ring

capa

cito

r (V)

Time (s)

VcapThresholdBackscatter event

① ① ①

② ②

③

①

②

③

(b) Parallel

1.5

1.6

1.7

1.8

1.9

2

2.1

2.2

2.3

2.4

2.5

0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4

Back

scat

ter e

vent

(on/

off)

Volta

ge a

cros

s en

ergy

sto

ring

capa

cito

r (V)

Time (s)

VcapThresholdBsckscatter event

① ① ①② ② ②④④ ④

(c) Alternative

Fig. 13. Power consumption test (variation of the voltage is measured across the energy storing capacitor, see Fig 11), ( 1©: Valid accelerometer reading; 2©:valid barometer reading; 3© failed accelerometer reading; and 4© failed barometer reading)

TABLE VPOWER CONSUMPTION CALCULATED FROM Vcap VARIATION

Approach ∆Vcap(mV) ∆E(µJ) t256(s) ρe(µJ/s)Sequential 432 0.93 20.97 11.35

Parallel 460 1.06 19.93 13.62Alternative 440 0.97 31.41 7.91

then the system times out; therefore no backscatter event wasgenerated in this session.

Figure 13(c) shows the terminal voltage (Vcap) of the energystoring capacitor and the backscatter activities of the Multi-Sensor WISP when it is operating under the alternative sched-ule. We can observe that, there are some unsuccessful readingsat time ≈ 0.1 s and time ≈ 0.25 s. They are failed barometerreading sessions and indicate that the barometer is more likelyto fail as it can consume up to 1.4 mA (peak current) the analogto digital conversion (see Table II).

Table V shows the power consumption calculated usingVcap. Using the total time, t256, taken to complete 256 readingcycles, the energy consumption rate ρe can be calculated asρe =

Eavg·256t256

. We can see that the sequential approach has aslightly lower voltage variation, consequently a lower energychange per cycle. However, the alternative schedule displaysthe minimum energy consumption due to the lower duty cycle.

2) Operational range: This experiment is conduced in ananechoic box as described in Section V-3. Table VI shows themean values of the maximum operational distances of the threemulti-sensor data retrieval schedules.

From Table VI, we can see that both sequential and parallelschedules have a shorter operational distance compared tothe alternative schedule. It is important to note that there isan inverse relation between the operational distance and theenergy consumption rate. The alternative schedule depicts thehighest operational range and the minimum energy consump-tion rate. Similarly, the parallel schedule showed the minimumoperational range and maximum energy consumption rate. Asexpected (see Section III), lower energy consumption ratesresult in higher operational distances (see Table VI).

3) Read rate: Figure 14 illustrates the read rate of the WISPas the reader transmission power is decreased. We can seethat, when adequate power is present the parallel schedule

TABLE VIOPERATIONAL RANGE OF MULTI-SENSOR SAMPLING SCHEDULES

Sequential (cm) Parallel (cm) Alternative (cm)46 41 173

0

2

4

6

8

10

12

14

192021222324252627282930

Effe

ctiv

e re

ad ra

te (1

/s)

Reader Tx Power (dBm)

DID.SequentialDID.ParallelDID.Alternating

Fig. 14. Reporting rate versus Reader power, illustrates the throughput offour different arrangement various as the decreasing harvested power.

has the highest read rate, but it is closely followed by thesequential schedule. This is due to the fact that, both scheduleshave similar energy consumption rates as shown in Table V.The lower read rate for the alternative schedule is because itrequires multiple inventory sessions (two in this experiment astwo sensors are used) to transmit the data from all the sensorsembedded in the sensor platform.

It is also important to note that, as expected, the alternativeschedule can perform even when there is less power to beharvested. When transmission power is at 20 dBm, both theparallel schedule and the sequential schedule were unrespon-sive. At this instance the alternative schedule was responding,but at a very low read rate. These results closely correlatewith the range test results (see Table VI) where the harvestedpower is reduced as the tag moves further away from the RFIDantenna.

VI. CONCLUSION AND FUTURE WORK

In this paper, we presented the design, implementation andevaluation of three data retrieval approaches for Multi-Sensor

Passive RFID Tags. These approaches were designed based onthe Data-in-ID approach.

From analysis and extensive experiments, we show thatthe alternative schedule has a longer operational range. Bothparallel and sequential schedules performed similarly in termsof operational range and power consumption. Although the se-quential schedule showed a slightly lower read rate comparedto the parallel schedule, sequential schedule was able to sampledata at a lower energy than required by the parallel schedule.This indicates that the sequential schedule perform marginallybetter than the parallel schedule when two sensors are em-bedded in a sensor tag. These properties may change whenmore sensors are used as indicated in Section III. These threesensor sampling schedules are energy efficient and faster thanthe traditional sensor data retrieval approach which is Read-by-Command. Proposed multi-sensor data retrieval approacheshave different characteristics and hence the most suitable onecan be selected according to application needs.

Although we have implemented these sensor samplingschedules using only two sensors, they can be used to samplemore than two sensors as these sensor sampling schedules bareno assumptions on the number of sensors used. To accountfor multiple sensors and more data requirements, 256 bit EPCor a larger EPC can be used instead of the 96 bit EPC usedcurrently. Incorporating more sensors, we are expecting toobserve the same, but more distinctive properties. For instance,read rate of the sequential schedule and the parallel scheduleis similar with two sensors (see Fig. 14), but we expect theread rate of the sequential schedule to be significantly lowerthan the parallel schedule.

There are two main limitations of this study. First, the sizeof the data that can be transmitted is limited by the EPC.Although a larger EPC could be used, transmitting data inorders of several kilobytes may require multiple inventorycycles and subsequently result in larger delays. However, usingRead-by-Command approach can acquire such data by usingone inventory command and one read command, which ismore efficient. Secondly, we have evaluated these multi-sensordata retrieval schedules using only two sensors. To furtheranalyze the properties of the proposed approaches empirically,a sensor tag with multiple sensors need to be implemented.

In conclusion, this study provides new prospects for usingsensor tags in multi-modal sensing to realize a wide range ofapplications.

VII. ACKNOWLEDGEMENT

This research was supported by grant funding from theAustralian Research Council (DP130104614) and the Depart-ment of Further Education, Employment, Science and Tech-nology (DFEEST) under the Collaboration Pathways Program,Government of South Australia. The researchers, in particularAlanson Sample, responsible for developing the WISP alsodeserve our gratitude.

REFERENCES

[1] D. C. Ranasinghe, R. L. Shinmoto Torres, K. Hill, and R. Visvanathan,“Low cost and batteryless sensor-enabled radio frequency identification

tag based approaches to identify patient bed entry and exit posturetransitions,” Gait & Posture, vol. 39, pp. 118–123, 2014.

[2] R. L. Shinmoto Torres, D. C. Ranasinghe, Q. Shi, and A. P. Sample,“Sensor enabled wearable RFID technology for mitigating the risk offalls near beds,” in IEEE International Conference on RFID, 2013, pp.191–198.

[3] V. Talla, M. Buettner, D. Wetherall, and J. Smith, “Hybrid analog-digitalbackscatter platform for high data rate, battery-free sensing,” in IEEETopical Conference on Wireless Sensors and Sensor Networks, 2013, pp.1–3.

[4] A. Dementyev and J. Smith, “A wearable UHF RFID-based EEGsystem,” in IEEE International Conference on RFID, 2013, pp. 1–7.

[5] A. Pietrabissa, C. Poli, D. G. Ferriero, and M. Grigioni, “Optimalplanning of sensor networks for asset tracking in hospital environments,”Decision Support Systems, vol. 55, no. 1, pp. 304–313, 2013.

[6] A. P. Sample, D. J. Yeager, P. S. Powledge, A. V. Mamishev, andJ. R. Smith, “Design of an rfid-based battery-free programmable sensingplatform,” IEEE Trans. Instrum. Meas., vol. 57, no. 11, pp. 2608–2615,2008.

[7] M. Buettner, B. Greenstein, A. Sample, J. R. Smith, and D. Wetherall,“Revisiting smart dust with RFID sensor networks,” in ACM Workshopon Hot Topics in Networks, 2008.

[8] T. Le, V. Lakafosis, Z. Lin, C. P. Wong, and M. Tentzeris, “Inkjet-printed graphene-based wireless gas sensor modules,” in IEEE ElectronicComponents and Technology Conference, 2012, pp. 1003–1008.

[9] J. Y. Kim, C. H. Chu, and S. M. Shin, “Designing integrated sensing sys-tems for real-time air quality monitoring,” in International Conferenceon Information Science and Applications (ICISA), 2014, pp. 1–6.

[10] Z. Zou, Q. Chen, I. Uysal, and L. Zheng, “Radio frequency identificationenabled wireless sensing for intelligent food logistics,” PhilosophicalTransactions of the Royal Society of London A: Mathematical, Physicaland Engineering Sciences, vol. 372, no. 2017, 2014.

[11] F. Bianchi, S. Redmond, M. Narayanan, S. Cerutti, and N. Lovell, “Baro-metric pressure and triaxial accelerometry-based falls event detection,”IEEE Trans. Neural Syst. Rehabil. Eng., vol. 18, no. 6, pp. 619 –627,2010.

[12] T. Umeda, H. Yoshida, S. Sekine, Y. Fujita, T. Suzuki, and S. Otaka, “A950-mhz rectifier circuit for sensor network tags with 10-m distance,”IEEE J. Solid-State Circuits, vol. 41, no. 1, pp. 35–41, 2006.

[13] C. A. Balanis, Antenna Theory: Analysis and Design. John Wiley &Sons, 2012.

[14] EPCglobal, “EPC Radio-Frequency Identity Protocols Generation-2UHF RFID version 2,” 2013.

[15] Z. Cheng, X. Zhang, Y. Dai, and Y. Lu, “Design techniques of low-power embedded EEPROM for passive RFID tag,” Analog IntegratedCircuits and Signal Processing, vol. 74, no. 3, pp. 585–589, 2013.

[16] Y. Igarashi, K. Miyazaki, Y. Sato, and J. Mitsugi, “A network architec-ture for fast retrieval of user memory data from sensor rf tags,” in IEEEInternational Conference on RFID, 2013, pp. 184–190.

[17] A. Wickramasinghe, D. C. Ranasinghe, and A. P. Sample, “Windware:Supporting ubiquitous computing with passive sensor enabled rfid,” inIEEE International Conference on RFID, 2014, pp. 31–38.

[18] Texas Instruments, “Mixed Signal Microcontroller,” MSP430F21x2datasheet, 2012.

[19] Measurement Specialties, “Barometric pressure sensor, with stainlesssteel cap,” MS5611-01BA03 datasheet, 2012.