LIMA-FILHO et al.: EMBEDDED SYSTEM INTEGRATED INTO A WIRELESS SENSOR NETWORK FOR ONLINE DYNAMIC TORQUE 405higher cost than the cost of the sensors themselves . Besidesthe high cost, the wired approach offers little ﬂexibility, mak-ing the network deployment and maintenance a harder process.In this context, wireless networks present a number of advan-tages compared to wired networks as, for example, the ease andspeed of deployment and maintenance, and low cost . Inaddition to that, wireless sensor networks (WSNs) provide self-organization and local processing capability. Therefore, thesenetworks appear as a ﬂexible and inexpensive solution for build-ing industrial monitoring and control systems.Nevertheless, the use of WSNs, when developing automationsystems for industrial environments, presents a number of chal-lenges that should be faced. Wireless networks have unreliablecommunication links, what can be aggravated with noise andinterference in the communication spectrum range. The unreli-ability of the transmission medium in wireless networks makesit difﬁcult to deﬁne quality of service guarantees. Studies on theapplication of WSNs in industrial environments, aiming at re-placing wired systems, have been extensively explored in recentyears , –.This paper presents an embedded system for determiningtorque and efﬁciency in industrial electric motors by employ-ing WSNs technology. For a set of electric motors, current andvoltage measures are gathered for later processing into an em-bedded system. Torque and efﬁciency results are then sent to abase unit for real-time monitoring. This way, preventive actionscan be taken whenever low-efﬁciency motors are detected andin cases of torque outbreaks.We have adopted the IEEE 802.15.4 standard for wirelesscommunication. This standard allows the formation of a largenetwork of sensors, in various industrial segments, where thestandard is expected to have a signiﬁcant impact . This stan-dard has been employed also in the mechatronics ﬁeld , .In comparison with other standards such as IEEE 802.11 (WiFi)and IEEE 802.15.1 (Bluetooth), the IEEE 802.15.4 standardhave advantages related to energy consumption, scalability, re-duced time for node inclusion, and low cost .Bin Lu and Gungor  identify the synergies between WSNsand the noninvasive methods for motor analysis based on electri-cal signals. The main limitation of their system derives from thelow data rate in IEEE 802.15.4 WSNs, since they do not employlocal processing. Thus, it is necessary to send a large amount ofdata when computing the desired parameters. This limits, amongother things, the frequency of data acquisition from sensors. Ina WSN with a large number of nodes, the situation becomeseven worse, since all nodes share the same physical transmis-sion medium. Furthermore, it should be taken into account theunreliability of communication inherent to wireless networks,which can cause the loss of transmitted data, hampering theparameters’ estimation process.The system proposed in this paper does all the data processinglocally, transmitting to the base unit only the targeted param-eters previously calculated. Thus, there is a large reduction inthe amount of transmitted data, enabling real-time and dynamicmonitoring of multiple motors, even with a high data rate ac-quisition in the analog-to-digital converters (ADC).This paper also presents studies on the relation betweenthe WSN performance and the quality of the communicationmedium in the network operating environment. As a result, weobserved the correlation between packet error rate (PER) andspectral occupancy1in the band used for communication. Theimpact caused by the insertion of new sources of interference inthe environment is also analyzed. Through these studies and atheoretical analysis, it was demonstrated that employing nodeswith local processing capabilities is essential for this type ofapplication, reducing the amount of data transmitted over thenetwork and allowing monitoring even in high interference sce-narios. In addition to that, our work provides insights for guidingthe development of new technologies and protocols for indus-trial WSNs.II. BACKGROUNDA. Shaft Torque EstimationIn an induction motor, the air gap is the region between sta-tor and rotor, where occurs the electromechanical conversionprocess. The AGT is the conjugate formed between the rotorand the stator magnetic ﬂux. In this study, the AGT method isused to estimate the motor shaft torque. According to (1), theestimation of the AGT can be performed noninvasively takingcurrent and voltage measurements from the electric motor :Tag =p√36(ia − ib) [vca + r(2ia + ib)]dt+(2ia + ib) [(vab − r(ia − ib)]dt (1)wherep number of motor poles;ia , ib motor line currents, in ampere;vca , vab motor power line voltages, in volt;r resistance of motor armature, in ohm.Equation (1) can be applied using instantaneous and simulta-neous acquisitions of ia , ib, vca , vab, and a measured value of r.It is valid both for motors connected in Y , with no connectionto the neutral, or Δ. Its integrals corresponding to the stator ﬂuxlinkages. AGT equations has also been used in many works thatuse other types of motors –.The torque on the shaft can be estimated by subtracting thelosses occurring after the process of electromechanical energyconversion from AGT, according to following equation :Tshaft = Tag −Lmecωr−LRslωr. (2)Mechanical losses (i.e., friction and windage Lmec) vary ac-cording to the particular motor and the industrial process towhich it belongs. If it is not possible to estimate the losses, thenit is necessary to perform a no-load test. The additional losses(i.e., stray-load loss, LRsl) result from nonlinear phenomenaof different natures, difﬁcult to quantify. These can be approx-imated by a percentage of motor power . In (2), ωr is therotor speed, in rad per second.1We use the term spectral cccupancy to denote the induced power on thechannel used by radios for communication.
406 IEEE/ASME TRANSACTIONS ON MECHATRONICS, VOL. 17, NO. 3, JUNE 2012Fig. 1. Relation between torque and speed.B. Shaft Speed EstimationMeasuring directly the rotor speed ωr can be impractical insome cases. Several methods of sensorless rotor speed estima-tion have been proposed. These methods follow two categories:one employing an induction motor model, and the other derivedfrom the analysis in the frequency spectrum of voltage and elec-tric current . The method proposed by Ishida and Iwata ,based on the electrical voltage, uses techniques of digital signalprocessing to detect the harmonics generated due to the rotorslots. However, it requires high rotor speed and stability .Ferrah et al.  and Hurst and Habetler  used the fastFourier transform to extract harmonics due to the rotor slots fromthe electric current spectrum. Some limitations of such methodare that it requires a high acquisition rate from sensors and highprocessing power. The method also requires information fromthe motor, which do not appear in their factory speciﬁcations.The methods mentioned earlier do not work well when thespeed is close to the synchronous speed and in dynamic systemswith variable torque and vibration. A conventional inductionmotor has a speed variation of less than 10% to the synchronousspeed when it is being used from no load to full load. In thenormal operation region, close to synchronous speed, the motorpresents an almost linear relationship between its torque and itsangular velocity (as can be seen in Fig. 1). Thus, a procedure forcurve linearization can be adopted. To perform this linearization,two points are needed to relate torque and speed. These pointscan be when the torque is nominal and when it is zero.C. Efﬁciency EstimationThe motor efﬁciency η can be estimated by the relation be-tween the electrical power supplied to the motor (i.e., inputpower Pin) and the mechanical power supplied to the shaft bythe motor (i.e., output powerPout), according to the followingequation:η =PoutPin. (3)Pin of a three-phase induction motor can be calculated by theinstantaneous currents and voltages, according to the followingequation:Pin = ia va + ibvb + icvc = −vca (ia + ib) − vabib. (4)Pout can be determined by the estimated shaft toque and therotor speed as follows:Pout = Tshaftωr . (5)By the replacement of (4) and (5) in (3), the efﬁciency η canbe estimated as follows:η =Tshaftωr−vca (ia + ib) − vabib. (6)D. Industrial WSNWSNs are formed by devices equipped with sensors and arecapable of communicating via radio frequency. These sensorscan produce responses to changes in physical conditions suchas temperature, humidity, or magnetic ﬁeld. Speciﬁc types ofWSNs, such as for industrial monitoring, have unique charac-teristics and speciﬁc application requirements. Therefore, thedeployment of WSNs must necessarily involve considerationsof the targeted application , .In general, there are key features that should be providedby the WSN, such as security, robustness, reliability, through-put, and adequate determinism. Among these characteristics,the lack of reliability is the main reason why many users donot deploy wireless equipment , . Much of this concernis related to the interference in the spectrum used by the wire-less networks for communication. Nodes in a wireless networkmay suffer interference from the coexistence with other networknodes, from other networks, and other technologies operatingin the same frequency range.In industrial environments, there can be other sources of noise,such as thermal noise, and noise from motors and devices thatcause electrical discharge . The error characteristics pre-sented in the wireless channel depend on the propagation envi-ronment, the modulation, transmission power, frequency range,among other parameters. In general, industrial wireless systemstend to have varying and often high error rates .The IEEE 802.15.4 standard is well suited for WSN appli-cations. It provides wireless communication with low powerconsumption and low cost, for monitoring and control appli-cations that do not require high data transmission rate. Thereare some protocols that implement the network layer over theIEEE 802.15.4 standard, such as Zigbee  and MiWi . Thestandard deﬁnes three frequency bands: 868 MHz, 915 MHz,and 2.4 GHz . In this study, we have considered only the2.4-GHz band.In an IEEE 802.15.4 network, there are two types of nodes:full function device (FFD), and reduced function device (RFD).The FFD nodes can act both as network coordinator or endnode. The coordinator is responsible, among other functions,for the initialization, address allocation, network maintenance,and the recognition of all other nodes. RFD nodes work only asend nodes, which are responsible for the functions of sensingor action. FFD nodes can also perform the function of inter-mediate routers between nodes, without the intervention of thecoordinator .Nowadays, many communication devices operate in the2.4-GHz band, which is an unlicensed band within the Industrial,
LIMA-FILHO et al.: EMBEDDED SYSTEM INTEGRATED INTO A WIRELESS SENSOR NETWORK FOR ONLINE DYNAMIC TORQUE 407Fig. 2. Embedded system integrated into the WSN.Scientiﬁc, and Medical radio bands (ISM), including radios thatuse the Bluetooth technology, WiFi, and other devices such ascordless phones and microwave ovens. The ISM band was ini-tially allocated for noncommercial use, and it was later modiﬁedto allow more services, which led to the emergence of a largenumber of applications. Such applications take advantage ofISM bands for free operation .Since the ISM bands are unlicensed, no users have priorityover others. The only restriction is on the strength of the sig-nal power, for ensuring reduced interference among concurrentsystems. Since there is no protection against interference ofconcurrent users, it is necessary to develop efﬁcient coexistencetechnologies, providing a good operation of unlicensed bandsystems. Therefore, it is necessary to develop a new approach forwireless communication systems design, which should includespectrum occupation measurements, modeling of interferenceand coexistence, and performance evaluations .The operation of WSNs in harsh industrial environments hasbeen continuously evaluated. Performance analyses of the IEEE802.15.4 in industrial environments are essential before thisstandard is employed for critical industrial applications. Suchanalysis can also provide valuable and solid foundation to sup-port the development of new technologies and protocols forWSNs, and guide the design of industrial applications based onWSN .III. SYSTEM DESCRIPTIONFig. 2 depicts the WSN proposed in this paper. End nodesare composed by the embedded systems located close to theelectric motors. The values of motor voltage and current areobtained from the sensors, and the embedded system performsthe processing for determining the values of torque, speed, andefﬁciency. Information obtained after the processing are trans-mitted to the base station through the WSN.Depending on the distance between end nodes and the coor-dinator, it may not be possible to achieve direct communication,due to the radio’s limited range and the interference present onthe environment, among other factors. Therefore, the communi-cation among nodes and coordinator can be done with assistanceof routers.Fig. 3. Block diagram of the embedded system.Fig. 4. Activity diagram.Fig. 3 shows a simpliﬁed block diagram of the proposed em-bedded system. For current measurement, Hall Effect sensors areemployed due to their robustness and noninvasiveness. Trans-formers with grain-oriented core are used to measure the voltagebetween phases, which provide the voltages in the secondary andprimary without delay. The acquisition and data processing unit(ADPU) is responsible for data acquisition and conversion, be-sides the data processing. Theprinted boards power supply sup-plies the current and voltage for the sensors, the IEEE 802.15.4transceiver, and the ADPU.The main element of the ADPU is a dsPIC33FJ64GP706,which is a digital signal controller designed for applicationsthat require high processing capacity. It has two integratedADC, which perform simultaneous acquisition of the voltageand current sensors. The input/output channels can be used foruser interface, and possible connections to auxiliary sensors andactuators. The values of torque and motor efﬁciency are trans-mitted using the IEEE 802.15.4 Transceiver. We have used anMRF24J40 transceiver, designed by Microchip. The connectionbetween the transceiver and the dsPIC is accomplished using aSerial Peripheral Interface Bus.The internal operation of the embedded system is illustratedby the activity diagram shown in Fig. 4. When the systemstarts, the embedded system parameters are conﬁgured. These
408 IEEE/ASME TRANSACTIONS ON MECHATRONICS, VOL. 17, NO. 3, JUNE 2012Fig. 5. Workbench employed for system analysis.parameters include the wireless network settings (e.g., address,channel), and the ADC settings. To obtain good accuracy froma simple numerical integration method, such as trapezoidal(used to implement the algorithm), a sample rate greater than2 kHz  should be used. In our system, we set the ADC tooperate with 3 kHz and 10 bits of resolution.After the ﬁrst step, the system connects to the WSN. Theembedded system only begins to acquire and process data aftersuccessfully connecting to a coordinator operating in the samechannel. Then, the system gets into the acquisition loop, pro-cessing, and transmitting data, which is repeated until the systemshuts down. The voltage and current values, after acquired, mustbe adjusted to reﬂect the real values measured from the sensors.After that, the algorithm is executed to compute the AGT, ac-cording to (1). After that, the losses are removed, and the shafttorque is estimated according to (2). Using the shaft torquevalues, the system estimates the motor speed and efﬁciency.The embedded systems were conﬁgured to calculate a set of360 values (2 bytes each) of torque and efﬁciency, and thentransmit these values aggregated into 20 packets with 72 bytesof payload each. The time necessary to acquire the signals andcalculate the 360 values of torque and efﬁciency is about 11 s (6s to acquire 360 cycles of current and voltage, and 5 s to performthe calculations). Thus, the system transmits data in burst mode,spending only about 8% of the time transmitting data, at a rateof 20 packets/s (about 14 kb/s, including control overhead).IV. EXPERIMENT METHODOLOGYA. Workbench for System AnalysisThe workbench was designed to obtain a known and variabletorque on the motor shaft. Fig. 5 shows its sketch, which consistsin a 550-W induction motor, with nominal rotation speed of1680 RPM, coupled to a reducer that provides an output speedof around 15 RPM. A steel disc was ﬁtted on the output shaft,symmetrically coupled with a metal bar. At one end of this barcan be inserted masses, according to the desired torque value,as shown in Fig. 5.When the motor is turned ON and the system (reducer shaft,steel disc, and steel bar) begins to rotate, the masses placed at theend of the steel bar impose a variable torque on the motor, andconsequently on the shaft of the reducer. The resultant torquehas sinusoidal shape with values dependent on the weights andthe position in which the bar is ON. The mathematical model ofthe reducer shaft torque (Treducer) can be obtained by the dy-namic equilibrium analysis, performing the sum of the momentsaround the center of the reducer shaft. We have that(M0)ext = (M0)inertia (7)where (M0)ext are the external moments and (M0)inertia arethe inertia moments components. Applying (7) for the systemof Fig. 5, it results inTreducer = LWr sin(θ) + α(Id + Ib + L2m) (8)where L is the distance between the masses center and the re-ducer shaft, Wr is the reference weight, m is the reference mass,θ is the angular position of the bar, and α is the angular accel-eration on the reducer output. The moment of inertia relativeto the disk and bar are the variables Id and Ib, respectively.The model of torque transformation between the low-speed sideto the high-speed side is obtained according to the followingequation:Tref =Treducerωreducerωr+ Jreducerα (9)where ωreducer is the reducer angular speed, Jreducer is thereducer inertia with respect to the high-speed side, and α is therotor angular acceleration. The workbench was instrumented tomeasure θ, ωreducer, ωr and α, used in torque and efﬁciencyequations, by using Hall Effect sensors and magnets. Jreduceris obtained from the manufacturer, and tests were performed toprovide the torque as a function of the load.According to (8), in the ﬁrst quarter circle, Treducer rangesfrom zero to its maximum value. Since the workbench efﬁciencymodel is determined by substitution of Tshaft by Tref in(6), theefﬁciency curve of the motor can be obtained for all of theiroperating range, using an appropriate Wr .B. Methodology of WSN Performance EvaluationIt is very important to conduct performance studies of wire-less systems in industrial environment, mainly due to the lackof reliability inherent to wireless networks. Therefore, this pa-per presents a study on the performance of the proposed WSNin order to observe its limitations and provide recommenda-tions when developing new solutions for achieving better per-formance of such systems.The communication performance among end nodes and thecoordinator was evaluated, while performing spectrum analy-sis in the surrounding environment. The PER was the chosenperformance metric. The spectrum was divided into channels,according to the IEEE 802.15.4 speciﬁcation. A comparisonwas made between the power values induced for each channeland the corresponding performance results. Thus, the effect ofinserting new interference sources in the environment was stud-ied, by verifying its impact on the spectrum occupation and onthe communication performance within each channel. Follow-ing are the detailed factors and response variables considered inthe experiment:1) Primary Factors (PF)
LIMA-FILHO et al.: EMBEDDED SYSTEM INTEGRATED INTO A WIRELESS SENSOR NETWORK FOR ONLINE DYNAMIC TORQUE 409PF-1—Channel: this factor is categorical and it has threelevels: channel 13, channel 18, and channel 24.PF-2—Interference Sources: this factor is categorical andit has three levels: “Only Microwave oven ON,” “OnlyIEEE 802.11g ON (channel 6)” and “Microwave ovenOFF and IEEE 802.11g OFF.”2) Response Variables (RV)RV-1—Spectrum Occupancy (SO): this response variableis obtained from the average power induced on the spec-trum range of each channel.RV-2—PER: this response variable is the number of incor-rectly transmitted data packets divided by the total numberof transmitted data packets.We decided to compare SO and PER considering only threechannels. This way, the time needed for the experiments wasreduced and allowed us to get a good analysis of the spectrumoccupation distribution and the corresponding relationship be-tween spectral occupancy and communication performance. Toincrease the signiﬁcance level of the ﬁndings, three replicationswere performed for every experiment. The conﬁdence level usedin all statistical tests was 95%.1) Calculationof Spectral Occupancy: Let fj beafrequencycomponent of the spectrum and F be the set of all frequencieson the spectrum range considered in the experiment.cr is a communication channel, and Cr is the set of frequencycomponents of the channel, where Cr ⊂ F.The induced power on the frequency component f in thespectrum at a given instant t is denoted by pt(fj ), and the in-duced power on the channel cr , denoted by Pt(cr ), obtainedfrom the average among the powers of the frequency compo-nents belonging to Cr , is given as follows:Pt(cr ) =fj ∈Crpt(fj )|Cr |. (10)The spectral occupancy on channel cr , denoted by SO(cr )w ,in a given discrete time interval Δtw is obtained through theaverage values of Pt(cr ), measured at each instant during thetime interval Δtw . The values of Pt(cr ) considers the powerinduced by all devices in the frequency range of cr , includingthe IEEE 802.15.4 radios and the interference sources.2) Threats to Validity: The exact moment that the experi-ment is performed can affect the conclusions, because the spec-trum occupation pattern can vary along time. In addition to that,temporal variations during the measurements can affect the re-sults, due to uncontrolled external factors, such as temperatureand humidity. However, our experiment was replicated threetimes, allowing observing the system behavior at different timeintervals, thus avoiding restricting the conclusions to a speciﬁcmeasure.3) Industrial Environment: The experiments were con-ducted inside a shed, with typical characteristics of industrialenvironments, such as the presence of large amounts of metallicdevices. Fig. 6 shows the environment where the experimentswere performed.4) Experiment Setup: We used three nodes to test the WSN,forming a star topology with one coordinator and two end nodes.Fig. 6. Industrial environment.The ﬁrst node, N1, was set 16 m away from the coordinator,while the second node, N2, was set 13 m away from the coor-dinator. Another important detail is that node N1 had no lineof sight to the coordinator, and it was among several metal ob-jects; while node N2 had a line of sight to the coordinator. Theend nodes were conﬁgured to transmit with an output power of0 dB·m.During the experiments, to verify the impact of an IEEE802.11g network, we performed a ﬁle transfer between twoIEEE 802.11g nodes connected to a base station. One node wasabout 1 m away from the IEEE 802.15.4 end nodes, while theother was placed next to the coordinator. The IEEE 802.11gnodes were set to transmit at power level of +15 dB·m.The embedded systems were conﬁgured to transmit 2000packets, in each replication. During the time spent to estimateand transmit all values, it was performed spectrum measure-ments. The values of SO(cr )w in the channel cr used for com-munication was calculated for this time interval Δtw .5) Instrumentation: For spectrum power acquisition, wehave used the Airview2/EXT  spectrum analyzer. For theIEEE 802.11g network deployment, a D-Link DI-524 router andtwo personal computers equipped with DWL-AG132 WirelessUSB Adapter were employed. The microwave oven used was aConsul CMS25ABHNA model, with a 700-W output power.V. RESULTS AND DISCUSSIONA. Analysis of the Embedded System Estimated ValuesFig. 7 shows the workbench used to analyze the system withall its components. The embedded system was placed near themotor to acquire current and voltage data. Torque and efﬁciencyare calculated by the ADPU module and are then transmittedthrough the WSN using the IEEE 802.15.4 transceiver. Thetorque and efﬁciency values are received at the monitoring basestation, where they can be visualized and stored.Fig. 8 shows the estimated torque curves read at the monitor-ing base station Tshaft, calculated in ADPU using (2), and thereference torque obtained from the workbench dynamic modelTref , obtained from (9).The curves in Fig. 8 were obtained for the ﬁrst half cycleof the steel bar (see Fig. 5), using two different masses. Thecurves comply with the dynamic model of the workbench that
410 IEEE/ASME TRANSACTIONS ON MECHATRONICS, VOL. 17, NO. 3, JUNE 2012Fig. 7. Experimental setup for the torque and efﬁciency analysis.Fig. 8. Comparison between estimated and reference torque measurementsfor two masses.consists of a sinusoid, corresponding to the ﬁrst part of (8),and modulated by the acceleration components, regarding thesecond part of (8) and (9).As shown in the curve regarding the reference weight withmass equal to 10 kg, the estimated torque follows the referencetorque, and it captures the workbench vibration. The relativeerror between the two curves is less than 2%. For the curveregarding reference weight with mass equal to 40 kg, we noticethat the torque amplitudes begin to diverge near the peak region,when the workbench presents an increase in vibration amplitude.With regard to the system’s efﬁciency, we have used the differ-ent reference weights for computing the peak reference torqueand the peak estimated torque. The corresponding speed andpower input values were also used in the calculation. Thus, thereference and estimated efﬁciencies were calculated using (6),by replacing Tshaft by the reference torque Tref [see (9)], andFig. 9. Comparison of estimated and reference efﬁciencies versus load.Fig. 10. Comparison of estimated and measured motor speed.the estimated torque [see (2)], respectively. Fig. 9 illustratesthe reference curve and the values estimated by the embed-ded system. On the X-axis we have the engine load range. Be-tween 0% and 85% of nominal power, the maximum error didnot exceed 2%. Even with the embedded processing and wire-less transmission, this result corroborates to other works thatuse the AGT efﬁciency method , .For other operating ranges, it was also possible to obtain rel-atively accurate efﬁciency with respect to the reference torque,even in the presence of strong workbench vibrations, whichoccur with greater intensity in the region near the nominal load.The motor speed was estimated through a linear approxima-tion using the AGT method. Fig. 10 compares the measuredspeed using the effect hall sensors and magnets, and the esti-mated speed for the two masses. We observed a maximum errorof 0.26% for the reference weight with mass equal to 10 kg and0.4% for the reference weight with mass equal to 40 kg.It was developed a software that runs in the the monitoringbase station. The system allows viewing the values obtainedfrom all embedded systems connected to the WSN in real time.
LIMA-FILHO et al.: EMBEDDED SYSTEM INTEGRATED INTO A WIRELESS SENSOR NETWORK FOR ONLINE DYNAMIC TORQUE 411Fig. 11. Base monitoring system.Fig. 12. Impact of a microwave oven.Fig. 11 shows the torque and efﬁciency curves received in realtime at the monitoring base station.B. WSN Performance EvaluationWe investigated the impact on spectral occupancy and PERmetrics due to the insertion of interference sources. In this sec-tion, we present also a theoretical analysis to compare the ap-proaches with local processing and without local processing.Figs. 12 and 13 show the impact of a microwave oven and anIEEE 802.11g network (operating on channel 6), respectively.The graphs on the left show the impact on node N1, and thegraphs on the right show the impact on node N2. The x-axiscontains the channels cr considered in the experiment, and inthe y-axis, on the left, we have PER values and in the y-axis, onthe right, we have SO(cr )w (in dB·m).We can note by the graphs from Figs. 12 and 13 that theinclusion of new interference sources resulted in a signiﬁcantperformance drop, mainly for node N1. Overall, node N2 didnot suffered great performance losses compared to node N1.This is because node N1 was farther away from the coordinator,without having a line of sight path to it.When analyzing the impact of the microwave oven, we ob-served that node N2 did not suffer a great drop in performance,but when operating on channel 24, the PER presented a smallFig. 13. Impact of an IEEE 802.11g network.variance of up to 15%. Node N1 suffered a signiﬁcant impactin PER, mainly when operating on channels 18 and 24. Thevariance was also large for these channels, reaching 80% onchannel 18% and 60% on channel 24.When analyzing the impact of the IEEE 802.11g network,we observed that there is a larger correlation between SO andPER. Again, node N1 experienced a bigger drop in performancecompared to node N2. For an 802.11g network operating onchannel 6, the IEEE 802.15.4 nodes experienced a signiﬁcantimpact only when they were operating on channel 18. On thischannel, node N2 presented a best performance, but with a PERvariance of up to 85%.When operating on channel 13, the two end nodes kept a goodquality of communication, in all scenarios under consideration.Fig. 14 shows the induced power on the frequency compo-nents in each scenario for one of the replications performedfor the channel 13. We note that the average power inducedby the microwave oven can be larger than the average powerinduced from IEEE 802.11g network, but the variance is alsomuch larger.When the microwave oven is turned ON, the SO(cr )w valuesdoes not present a very high variance, as can be seen in thegraphs 12 and 13, but when analyzing the frequency componentsindividually, we see that the variance of pt(fj ) is very largeduring Δtw , mainly in the more affected components.This may explain the lower correlation between the spectraloccupancy and PER for this scenario. It also explains why theinterference of the IEEE 802.11g network leads to a greaterperformance drop both in N1 and N2, even inducing a loweraverage power on the channels. We can see that when the IEEE802.11g network is turned ON, the noise level in the affectedfrequency range remains high all the time, with little variance.When the microwave oven is turned ON, the noise level inthe affected frequency range ﬂuctuates between very low noiselevels and very high noise levels, increasing the fraction of timein which the medium remains free.In high interference scenarios, the use of local processingbecomes even more important. Each node obtains data fromtwo current sensors and two voltage sensors with an acquisitionrate of 3 kHz, where each value obtained from the sensors has10 bits. To calculate a value of AGT, a full cycle of voltage andcurrent is needed. Therefore, to obtain one efﬁciency value and
412 IEEE/ASME TRANSACTIONS ON MECHATRONICS, VOL. 17, NO. 3, JUNE 2012Fig. 14. Induced power on the frequency components.one torque value, 50 samples are required from each sensor. LetQ be the number of bits that must be transmitted to estimate thevalues, we have that Q = 4 × 50 × 10 = 2000 bits.With local processing, to obtain one torque value and oneefﬁciency value, it is always necessary to transmit a constantnumber of bits, regardless the ADC’s acquisition rate and reso-lution. In our case, 4 bytes (32 bits) are used to store a value oftorque and efﬁciency. Therefore, it is necessary only one packetto transmit such data. When local processing is not used, it isnecessary to transmit more packets through the WSN. Whereasthe packet payload is 118 bytes (944 bits), the number of packetsthat must be transmitted, Qp , is Qp = Q/944 + 1 = 3.In the best case, the total number of transmitted packets isthree times larger in the scenario without local processing. How-ever, when considering an unreliable transmission medium, thedifference can be greater. The transmission of a packet con-sists of a Bernoulli event with successful probability p, and thenumber of trials until the ﬁrst success is deﬁned by a geometricdistribution. In a geometric distribution, the average number ofevents until the ﬁrst success is 1/p.Since Qp = 3, when local processing is not used the proba-bility of successfully transmitting the data necessary to estimatethe values is p3, and the number of attempts before the ﬁrstsuccess is 1/p3, considering a scenario without retransmissionof lost packets. As in each attempt three packets are transmitted,then we have an average of 3/p3transmissions per efﬁciencyand torque value obtained.When local processing is not employed, packet retransmis-sion improves performance, since only individual packets witherrors need to be retransmitted. However, recurrent retransmis-sions increase overall delay, possibly failing to obtain the mostcurrent data. Considering that the successful probability of anacknowledgment packet is also p, then the successful probabil-ity of transmit a packet becomes p2, and the average number oftransmissions to the ﬁrst hit is 3/p2. In the scenario with localprocessing and retransmission, the average number of transmis-sions until the ﬁrst hit is 1/p2. Therefore, when using localprocessing, retransmissions increase the amount of transmittedpackets.Table I shows the average number of transmissions requiredto obtain a torque value and an efﬁciency value, for various val-TABLE 1TRANSMISSIONSues of p in each scenario. From the table, we note that in highinterference scenarios, there is a large reduction in the amountof data being transmitted when using local processing. Whenthe PER is too high, it is very difﬁcult to perform monitoringwithout using local processing. For example, in the scenariowith an IEEE 802.11g network operating on channel 6, we seethat the average PER for node N1 (on channel 18) was 90%. Inthis scenario, without local processing it would be required, onaverage, 300 transmissions to obtain data on the target, consid-ering the use of retransmissions. Using local processing withoutretransmission, only ten transmissions are needed in average.Node N2 presents best performance in this scenario, but thePER variance was high (up to 85%).Moreover, as a torque and an efﬁciency value occupy only 4bytes, we can aggregate multiple measures in a single packet.The embedded system was conﬁgured to transmit a set of 18measures of torque and 18 measures of efﬁciency in a singlepacket, with a total payload of 72 bytes. Besides that, it is impor-tant to note that due to an increase in the number of transmittedpackets, considering an approach without local processing, thePER tends to increase.From these results, we note that the deployment of industrialWSN still presents serious challenges related to communicationreliability. In order to keep a certain level of quality of service,radios need to be aware of the environment where they are op-erating, adopting a dynamic spectrum allocation approach, notonly at the beginning of its operation, but also during the entireperiod of operation. When there are changes in the distribution
LIMA-FILHO et al.: EMBEDDED SYSTEM INTEGRATED INTO A WIRELESS SENSOR NETWORK FOR ONLINE DYNAMIC TORQUE 413of spectrum usage, probably resulting from other sources ofinterference, the distribution of spectrum usage along the avail-able channels, and the communication performance on eachchannel, will change, what may imply on switching to a lesspolluted channel.Despite high PER in some scenarios, it is important to notethat due to local processing capability of the embedded systems,all the data arriving at the coordinator are useful information thatcan be employed for decision making. Without local process-ing, probably it would not have been possible to obtain usefulinformation via the WSN, in some scenarios.Some studies ,  have proposed solutions for miti-gating the interference effects in IEEE 802.15.4 networks. Bycombining the local processing capability, as explored in thisstudy, with dynamic spectrum allocation techniques and tech-niques for mitigating the effects of interference, it may be pos-sible to achieve a good quality of service in motor monitoringapplications based on WSNs.VI. CONCLUSION AND FUTURE WORKThis paper presented an embedded system integrated into aWSN for online dynamic torque and efﬁciency monitoring ininduction motors. We used the AGT method to estimate shafttorque and motor efﬁciency. The calculations for estimatingthe targeted values are done locally and then transmitted to amonitoring base unit through an IEEE 802.15.4 WSN.Experimental tests were performed to analyze the torque val-ues obtained by the system, and then compared with torquevalues based on the workbench dynamic model. The estimatedefﬁciency was compared with the reference efﬁciency, present-ing an error smaller than 2.0% in the range of 0–85% loading.This paper also showed an experimental study aiming to iden-tify the relation between spectral occupancy and PER for theproposed WSN. The experiments were conducted inside a shed,with typical characteristics of industrial environments.The study demonstrated that the addition of new interferencesources can signiﬁcantly affect the spectral occupancy, by alsohaving a direct impact on the communication performance.Even with the difﬁculties in data transmission using the WSNin some scenarios, the system was able to provide useful mon-itoring information, since all processing is done locally (i.e.,only the information already computed is transmitted over thenetwork). Without local processing, it might be impossible touse the WSN technology for this particular application, con-sidering an unreliable transmission medium. Allied to the lo-cal processing capacity, other techniques can be developed tomitigate interference in those environments, leading to bettercommunication performance.As future work, we intend to conduct more detailed per-formance studies, considering a network with a larger numberof nodes in an industrial plant. Finally, we intend to developspectrum-aware protocols to allow the radios to choose theiroperation channels dynamically, allowing the embedded sys-tems to self-adapt to the operating environment, improving thequality of service of the network. It is also desirable to con-duct more detailed dynamic analysis of the workbench used forvalidation, especially with regard to reducing losses at nominalload.REFERENCES R. Hanitsch, “Energy Efﬁcient Electric Motors,” in Proc. RIO 02—WorldClimate Energy Event, 2002, pp. 6–11. K. Kim and A. G. Parlos, “Induction motor fault diagnosis based on neu-ropredictors and wavelet signal processing,” IEEE/ASME Trans. Mecha-tronics, vol. 7, no. 2, pp. 201–219, Jun. 2002. A. C. Lima-Filho, F. A. Belo, and R. D. Gomes, “Tests prove, self-powered,wireless, pump torquemeter,” Oil Gas J., vol. 106, no. 46, pp. 43–48, 2008. R. B. Reich, “Rotary transformer,” U.S. Patent 4 412 198, Oct. 25, 1983. W. F. Buchele, “Strain-gauge brushless torque-meter,” U.S. Patent3 881 347, May 6, 1975. Z. Meng and B. Liu, “Research on torque real time monitoring system ofrotary machine,” Chin. J. Sci. Instrum., vol. 26, pp. 38–39, 2005. A. C. Lima-Filho, F. A. Belo, J. L. A. Santos, and E. G. Anjos, “Exper-imental and theoretical study of a telemetric dynamic torque meter,” J.Brazilian Soc. Mech. Sci. Eng., vol. 32, pp. 241–249, 2010. J. S. Hsu and A. M. A. Amin, “Torque calculations of current-sourceinduction machines using the 1-2-0 coordinate system,” IEEE Trans. Ind.Electron., vol. 37, no. 1, pp. 34–40, Feb. 1990. J. S. Hsu, “Capacitor effects on induction motors fed by quasi rectangularcurrent sources,” IEEE Trans. Energy Convers., vol. 7, no. 3, pp. 509–516,Sep. 1992. T. H. Lee, T.-S. Low, K.-J. Tseng, and H. K. Lim, “An intelligent indirectdynamic torque sensor for permanent magnet brushless DC drives,” IEEETrans. Ind. Electron., vol. 41, no 2, pp. 191–200, Apr. 1994. A. C. Lima-Filho, F. A. Belo, J. L. S. Santos, and E. G. Anjos, “Self-powered telemetric torque meter,” J. Dyn. Syst., Meas., Control, vol. 133,pp. 1–7, 2011. IEEE Standard Test Procedure for Polyphase Induction Motors and Gen-erators, IEEE Standard 112-1996, Nov. 2004. J. S. Hsu, J. D. Kueck, M. Olszewski, D. A. Casada, P. J. Otaduy, andL. M. Tolbert, “Comparison of induction motor ﬁeld efﬁciency evaluationmethods,” IEEE Trans. Ind. Appl., vol. 34, no. 1, pp. 117–125, Jan./Feb.1998. Washington State University-Energy Program, “In-service motor testing,”Res. Rep. E99-040, 1999. J. D. Kueck, M. Olszewski, D. A. Casada, J. S. Hsu, P. J. Otaduy, and L.M. Tolbert, “Assessment of methods for estimating motor efﬁciency andload under ﬁeld conditions,” Lockheed Martin Energy Res. Corp., OakRidge, TN, Rep. ORNL/TM-13165, 1996. J. S. Hsu and B. P. Scoggins, “Field test of motor efﬁciency and loadchanges through air-gap torque,” IEEE Trans. Energy Convers., vol. 10,no. 3, pp. 471–477, Sep. 1995. B. Lu, T. G. Habetler, and R. G. Harley, “A nonintrusive and in-servicemotor-efﬁciency estimation method using air-gap torque with considera-tions of condition monitoring,” IEEE Trans. Ind. Appl., vol. 44, no. 6,pp. 1666–1674, Nov./Dec. 2008. B. Lu and V. C. Gungor, “Online and remote motor energy monitoringand fault diagnostics using wireless sensor networks,” IEEE Trans. Ind.Electron., vol. 56, no. 11, pp. 4651–4659, Nov. 2009. P. Baronti, P. Pillai, V. W. C. Chook, S. Chessa, A. Gotta, and Y. F. Hu,“Wireless sensor networks: A survey on the state of the art and the 802.15.4and ZigBee Standards,” Comput. Commun., vol. 30, no. 7, pp. 1655–1695,2007. F. Salvadori, M. de Campos, P. S. Sausen, R. F. de Camargo, C. Gehrke,C. Rech, M. A. Spohn, and A. C. Oliveira, “Monitoring in industrialsystems using wireless sensor network with dynamic power management,”IEEE Trans. Instrum. Meas., vol. 58, no. 9, pp. 3104–3111, Sep. 2009. A. Willig, “Recent and emerging topics in wireless industrial communica-tions: A selection,” IEEE Trans. Ind. Informat., vol. 4, no. 2, pp. 102–124,May 2008. V. C. Gungor and G. P. Hancke, “Industrial wireless sensor networks:Challenges, design principles, and technical approaches,” IEEE Trans.Ind. Electron., vol. 56, no. 10, pp. 4258–4265, Oct. 2009. H. Wei, Y. Chen, J. Tan, and T. Wang, “Sambot: A self-assembly modularrobot system,” IEEE/ASME Trans. Mechatronics, vol. 16, no. 4, pp. 745–757, Aug. 2011. J. Takahashi, T. Yamaguchi, K. Sekiyama, and T. Fukuda, “Communica-tion timing control and topology reconﬁguration of a sink-free meshedsensor network with mobile robots,” IEEE/ASME Trans. Mechatronics,vol. 14, no. 2, pp. 187–197, Apr. 2009.
414 IEEE/ASME TRANSACTIONS ON MECHATRONICS, VOL. 17, NO. 3, JUNE 2012 J. L. A. Santos, R. C. C. Araujo, A. C. L. Filho, F. A. Belo, and J. A.G. Lima, “Telemetric system for monitoring and automation of railroadnetworks,” Transp. Plann. Technol., vol. 34, pp. 593–603, 2011. K. Gulez, A. A. Adam, and H. Pastaci, “A novel direct torque controlalgorithm for IPMSM with minimum harmonics and torque ripples,”IEEE/ASME Trans. Mechatronics, vol. 12, no. 2, pp. 223–227, Apr. 2007. S. B. Ozturk and H. A. Toliyat, “Direct torque and indirect ﬂux control ofbrushless DC motor,” IEEE/ASME Trans. Mechatronics, vol. 16, no. 2,pp. 351–360, Apr. 2011. Q. D. Nguyen and S. Ueno, “Modeling and control of salient-pole perma-nent magnet axial-gap self-bearing motor,” IEEE/ASME Trans. Mecha-tronics, vol. 16, no. 3, pp. 518–526, Jun. 2011. R. M. Bharadwaj and A. G. Parlos, “Neural speed ﬁltering for inductionmotors with anomalies and incipient faults,” IEEE/ASME Trans. Mecha-tronics, vol. 9, no. 4, pp. 679–688, Dec. 2004. M. Ishida and K. Iwata, “Steady-state characteristics of a torque andspeed control system of an induction motor utilizing rotor slot harmonicsfor slip frequency sensing,” IEEE Trans. Power Electron., vol. PE-2, no3, pp. 257–263, Jul. 1987. B. Lu, T. G. Habetler, and R. G. Harley, “A survey of efﬁciency-estimationmethods for in-service induction motors,” IEEE Trans. Ind. Appl., vol. 42,no. 4, pp. 924–933, Jul./Aug. 2006. A. Ferrah, K. Bradley, and G. Asher, “Sensorless speed detection of in-verter fed induction motors using rotor slot harmonics and fast Fouriertransform,” in Proc. IEEE PESC, 1992, vol. 1, pp. 279–286. K. D. Hurst and T. G. Habetler, “Sensorless speed measurement usingcurrent harmonic spectral estimation in induction machine drivers,” IEEETrans. Power Electron., vol. 11, no. 1, pp. 66–73, Jan. 1996. A. A. F. Loureiro, J. M. S. Nogueira, L. B. Ruiz, R. A. de Freitas Mini,E. F. Nakamura, and C. M. S. Figueiredo, “Redes de sensores sem ﬁo,”in Proc. Simp´osio Brasileiro de Redes de Computadores, (in Portuguese),2003, pp. 1–48. I. F. Akyildiz and M. C. Vuran, Wireless Sensor Networks, 1st ed. NewYork: Wiley, 2010. J. Yick, B. Mukherjee, and D. Ghosal, “Wireless Sensor Network Survey,”in Proc. Comput. Netw., 2008, pp. 2292–2330. B. Fette, R. Aiello, P. Chandra, D. Dobkin, D. Bensky, D. Miron, D. Lide,F. Dowla, and R. Olexa, RF & Wireless Technologies: Know It All. Am-sterdam, The Netherlands: Elsevier, 2007. A. Willig, K. Matheus, and A. Wolisz, “Wireless technology in industrialnetworks,” in Proc. IEEE, vol. 93, no. 6, 1130–1151, 2005. Zigbee Alliance. (2011). [Online]. Available: http://www.zigbee.org. Microchip Technology. (2011). [Online]. Available: http://www.microchip.com. L. Cao, W. Jiang, and Z. Zangh, “Network wireless meter reading systembased on zigbee technology,” in Proc. IEEE Control Decis. Conf., 2008,pp. 3455–3460. H. Karl and A. Willig, Protocols and Architectures for Wireless SensorNetworks. New York: Wiley, 2005. H. Li, M. Syed, Y.-D. Yao, and T. Kamakaris, “Spectrum sharing inan ISM band: Outage performance of a hybrid DS/FH spread spectrumsystem with beamforming,” EURASIP J. Adv. Signal Process., 2009, pp.1–12. AirView - Ubiquiti Networks. (2011). [Online]. Available: http://www.ubnt.com/airview. J. Huang, G. Xing, G. Zhou, and R. Zhou, “Beyond co-existence: Exploit-ing WiFi white space for zigbee performance assurance,” in Proc. IEEEInt. Conf. Netw. Prot., 2010, pp. 305–314. C. Liang, N. Priyantha, J. Liu, and A. Terzis, “Surviving Wi-Fi interferencein low power zigbee networks,” in Proc. ACM Conf. Embedded Netw. Sens.Syst., 2010, pp. 309–322.Abel C. Lima-Filho was born in Macei´o, Brazil,in 1980. He received the B.S. degree in electricalengineering from the Federal University of Camp-ina Grande, Campina Grande, Brazil, in 2004, andthe M.S. and D.S. degrees in mechanical engineeringfrom the Federal University of Para´ıba, Jo˜ao Pessoa,Brazil.Since 2010, he has been with the Mechanical En-gineering Department, Federal University of Para´ıba,he is currently a Professor. His research interests in-clude telemetry, embedded systems, and electronicinstrumentation.Ruan D. Gomes was born in Jo˜ao Pessoa, Brazil,in 1988. He received the B.S. degree in computerscience from the Federal University of Para´ıba, Jo˜aoPessoa, Brazil, in 2010, and the M.S. degree in com-puter science from the Federal University of CampinaGrande, Campina Grande, Brazil, in 2012.He is currently a Researcher at the Federal Uni-versity of Para´ıba. His main research interests includewireless sensor networks and embedded systems.Marc´eu O. Adissi was born in Jo˜ao Pessoa, Brazil,in 1984. He received the B.S. degree in electricalengineering from the Federal University of Camp-ina Grande, Campina Grande, Brazil, in 2009, andthe M.Sc. degree in mechanical engineering from theFederal University of Para´ıba, Jo˜ao Pessoa, Brazil, in2012.He is currently a Researcher at the Federal Univer-sity of Para´ıba. His main research interests includeinstrumentation, electrical machinery, and controlsystems.T´assio Alessandro Borges da Silva was born in Jo˜aoPessoa, Brazil, on March 9, 1989. He received theB.S. degree in industrial automation technology fromthe Federal Institute of Education, Science and Tech-nology of Para´ıba, Jo˜ao Pessoa, Brazil, in 2011. Heis currently working toward the degree in electricalengineering at the Federal University of Para´ıba, Jo˜aoPessoa.He is also a Researcher at the Federal Universityof Para´ıba. His main research interests include em-bedded systems and thermophysical properties.Francisco A. Belo received the B.S. degree in elec-tronic engineering from the Aeronautics Technolog-ical Institute, Sa˜ao Paulo, Brazil, the M.S. degreein mechanical engineering from the Federal Univer-sity of Para´ıba, Jo˜ao Pessoa, Brazil, and the D.S.degree in mechanical engineering from the Univer-sity of Campinas, Campinas, Brazil.He is a Professor at the Federal University ofParaiba, Jo˜ao Pessoa. He has experience in the ﬁeldof mechanical engineering with emphasis on trans-port phenomena. His research interests include mul-tiphase ﬂow, tomography, telemetry, and electronic transduction.Marco A. Spohn received the Ph.D. degree in com-puter science from the University of California, SantaCruz, in 2005.He is currently an Associate Professor at the Fed-eral University of Fronteira Sul, Chapec´o, Brazil. Hiscurrent research interest is the design of algorithmsand protocols for wireless ad hoc and sensor net-works, grid computing, and distributed systems