Diagnosis of Stator Winding Inter-Turn Shorts in Induction Motors Fed by PWM-Inverter Drive Systems Using a Time-Series Data Mining Technique

Transcription

1 Dagoss of Stator Wdg Iter-Tur Shorts Iducto Motors Fed by PWM-Iverter Drve Systems Usg a Tme-Seres Data Mg Techque ChaChou Yeh, Studet Member, IEEE, Rchard J. Povell, Seor Member, IEEE, Behrooz Mrafzal, Studet Member, IEEE, ad Nabeel A. O. Demerdash, Fellow, IEEE Abstract---A effectve techque for dagoss of stator wdg ter-tur shorts ducto motors fed by PWMverter drve systems s proposed. Ths s doe through the use of a Tme-Seres Data Mg techque whch detfes ad extracts hdde ad heret patters (characterstcs) the mache phase currets, that ca be used for fault detfcato. Ths techque ca effectvely detect ad determe the severty of stator ter-tur faults motors by aalyzg the extracted fault sgatures of these faults comparso wth the healthy performace sgatures. I addto, t wll be see from the expermetal results that the proposed techque s mmue to motor o-dealtes such as heret maufacture-based motor asymmetry due to motor structural mperfectos, performace measuremet mperfectos, ad supply voltage ubalaces, whch result departure of performace results from those of a deally balaced 3-phase mache. I ths paper, the case-study uder vestgato s a 230-volt, 60-Hz, 2-pole, 2- hp, squrrel-cage three-phase ducto motor-drve system. The expermetal results wll demostrate the soudess ad robustess of ths techque for relable fault dagostcs. Idex Terms---Data mg, fault dagostcs, o-le codto motorg, stator wdg ter-tur shorts, tme seres, space vector, ducto motor, PWM verter. I I. INTRODUCTION N RECENT YEARS, dagoss of stator wdg tertur shorts has receved cosderable terest from dustry ad academa. Such faults may start as cpet tur-to-tur defects ad udergo a loger perod of tme before they evolve to a total sulato breakdow codto [], [2]. The stuato becomes worse whe motors are fed from verter drves, whch s due to the voltage stresses mposed by the fast swtchg of the power semcoductors of such verters. Cosequetly, such cpet wdg faults wll rapdly progress to more severe faults, such as tur-to-groud, tur-totur, or phase-to-phase faults, whch wll cause rreversble damage to these stator wdgs ad cores. The most drect cosequece of stator wdg ter-tur shorts s the asymmetry such faults troduce the motordrve crcutry ad cosequet asymmetry the mache Ths materal s based upo work supported by the Natoal Scece Foudato uder Grat No. ECS C. C. Yeh, R. J. Povell, B. Mrafzal, ad N. A. O. Demerdash are wth the Departmet of Electrcal ad Computer Egeerg, Marquette Uversty, Mlwaukee, WI 5320 USA (emal: chachou.yeh@marquette.edu, rchard.povell@marquette.edu, behrooz.mrafzal@marquette.edu, abeel.demerdash@marquette.edu). currets, flux lkages, ad voltages [3], [4]. These curret, flux lkage, ad voltage asymmetres lead to poor effcecy ad adverse codtos the mache wdgs ad magetc cores, ad cosequetly cause ultmate breakdows ad cosequet damages such maches of varous forms. Therefore, thorough ad comprehesve characterzatos of the mache performace uder stator wdg cpet tertur faulty codtos are mportat order to detect ad determe the severty of such cpet faults ad ther progresso wth tme versus the healthy codto. A substatal amout of research has bee coducted the past to dagose stator wdg ter-tur faults electrc motor-drve systems [2], [5]-[3]. Oe such method for stator ter-tur fault detecto s based o aalyzg the spectrum of the axal leakage flux compoet of the mache [5]. However, ths method requres the stallato of sesors whch are costly, ad sometmes approprate whe the motor s operatg a adverse evromet. Other methods are based o detectg ubalaces (asymmetres) of the mache such as detectg the egatve-sequece curret compoet [2], [6], [7] or egatve-sequece mpedaces [6]. Also, tertur fault detecto has bee addressed [8] through motorg the presece of certa rotor-slot-related harmocs at the termal voltage of the mache mmedately after swtch-off of the stator wdg. Furthermore, eural etwork ad artfcal tellgece methods have also bee appled to detect stator wdg ter-tur faults [9]-[]. Recetly, a ew ad uque method for fault motorg ad dagostcs ducto motors (IM) based o Tme- Seres Data Mg (TSDM) has evolved [2], [3]. Ths techque, whch s based o a tme-delay embeddg process, ca reveal ad extract hdde ad heret patters (characterstcs) the voltages ad currets that ca be used for fault detfcato. Ths techque ca effectvely detect ad dstgush the types of faults motor-drve systems by aalyzg the extracted fault sgatures of these faults comparso wth the healthy performace sgatures. The types of faults studed [2], [3] are statc ad dyamc eccetrctes, broke bars, ad broke ed-rg coectors ducto motors. Upo dstgushg the fault types, ths TSDM approach ca also reveal the severty of the faults from the fault sgatures. As revealed [2], [3], the TSDM was appled oly for drect le-start operatg codto, that s wthout the presece of ay electroc drve for torque-speed cotrol. I ths work, the TSDM techque s employed to dagose stator wdg ter-tur shorts a case of a ducto motor

2 eergzed from a PWM-verter drve system. The TSDM techque preseted here s based o a smlar cocept gve [2], [3], but a dfferet approach for fault detfcato, the detals of whch wll be clearly explaed ths paper. It s of mportace to meto that the TSDM techque preseted here s mmue to motor o-dealtes such as heret motor asymmetry due to motor structural ad maufacturg mperfectos, performace measuremet mperfectos, ad supply voltage ubalaces, whch result asymmetry of the mache as a 3-phase electrcal devce. Hece, dagostc schemes whch are based o measurg egatve-sequece curret or mpedace have to be carefully mplemeted to take to accout the heret motor odealtes. However, ths s ot the case here as these motor o-dealtes have bee corporated to the TSDM dagostc scheme through the tal acqusto of the healthy motor s performace data. Aga, ths wll be detaled later o ths paper. The mache uder vestgato here s a 230- volt, 60-Hz, 2-pole, 2-hp, squrrel-cage three-phase ducto motor-drve system, whch oe of ts phases was maufactured wth phase wdg taps for expermetal mmckg of cpet ter-tur shorts. The expermetal results wll demostrate the soudess ad robustess of ths techque for relable fault dagostcs. II. TIME-SERIES DATA MINING TECHNIQUE I ths secto, the algorthm of the TSDM techque s clearly explaed. The TSDM techque, whch s based o a tme-delay embeddg process [2], [3], trasforms the tme seres data (tme-doma waveform) to a dfferet processg state space called a recostructed phase space (RPS) [4], [5]. Gve the tme seres data: { } I =,..., = N () where, s the tme dex, the RPS matrx, I, s defed by ts row vectors whose elemets are tme-lagged versos of the orgal tme seres data, as follows: I = + ( d ) 2+ ( d ) N N ( d 2) N ( d ) where, s the orgal tme seres data, N s the umber of observatos (umber of data samples a waveform), d s the embeddg dmeso, ad s the tme lag. It s mportat to pot out that each row vector of I s a sgle pot the RPS. Here, the tme lag,, s determed usg the frst mmum of the automutual formato fucto, from whch the embeddg dmeso, d, ca be estmated usg the false earest eghbor method [6]. It had bee prove by Takes [7] that the RPS s topologcally equvalet to the orgal state space of the system, ad therefore the full dyamcs of the orgal system (2) are accessble ths RPS. Based o ths characterstc, the hdde patter that s captured from the RPS has to be resdg the orgal system, whch caot be easly detected from a frequecy spectral aalyss, or other methods of examato of the tme doma profle of the physcal pheomeo s waveform uder cosderato. As metoed earler, stator wdg ter-tur shorts troduce asymmetry (ubalace) the mache phase wdgs. Hece, t s sgfcat to use ay of the mache phase currets as the tme seres data the RPS f the locato of the stator ter-tur short s udetfed. Also, the use of the developed torque profle for TSDM s ot recommeded here. Ths s because, order to obta the torque data, oe way has to measure such torque usg a torque trasducer, whch s relatvely costly ad thereby adds a extra sesor wth cosequet addtoal ecumbrace to the dagostc system. Aother way s to measure the phase currets, the termal voltages, ad the speed of the motor, from whch the torque ca be obtaed by computg the put motor power mus the ohmc losses, ad dvded by the rotor speed. However, to obta the stator ohmc losses, oe would eed to kow the stator resstaces of all phases at ay tme because of the chages values troduced whe stator wdg tertur shorts occur. However, ths s ot a feasble approach. Therefore, a method whch s based o stator curret space vector s troduced here, whch wll clude the asymmetry effects of the wdgs to a sgle set of tme seres data. The space vector of the stator currets s defed as the sum of space vectors of dvdual phases, as follows [8]: where, sa 2 2 s() t = { sa() t + αsb() t + α sc() t } (3) 3, sb, ad sc are mache stator phase currets, ad 2 π /3 α = e j s a complex space operator. The curret space vector (3), whch s a fucto of tme, s a complex umber. Therefore, the absolute value of the curret space vector wll be used here as the tme seres data for the TSDM. Upo geeratg the RPS usg the tme seres of curret space vector, the ext step s to categorze the curret space vector sgals to dfferet classes (healthy ad faulty, where the faulty classes ca be further dvded to the umber of shorted turs) usg Gaussa Mxture Models (GMMs) [5]. Ths s doe by learg the GMM probablty dstrbuto for each class the RPS, whch s defed as follows: M ( ) = m m( ) = m ( ; m, m) M p x w p x w N x µ Σ (4) m= m= where, M s the umber of mxtures, ( ;, ) ormal dstrbuto wth mea, N xµ Σ s a, ad covarace matrx, Σm, ad wm s the mxture weght havg the costrat of w m = [5]. The umber of mxtures, M, ca be determed based o the classfcato accuracy whch teds to crease wth the creased umber of mxtures, provded there s suffcet data to tra o the GMM. The remag parameters of (4) ca be estmated usg the well-kow µ m m m

3 Expectato Maxmzato (EM) method [9]. After developg a GMM for each sgal class, the ext step s to classfy the test curret space vector sgal based o the traed (leared) models. Ths s doe usg a Bayesa maxmum lkelhood classfer, whch computes the codtoal lkelhoods of the test sgal uder each traed model ad selects the model wth the hghest lkelhood. The lkelhoods are computed o a pot-by-pot bass from the traed models, as follows [5]: N p( X c) = p(x c ) (5) =+ ( d ) where, X s a RPS matrx of dmeso, d, ad tme lag,, of the test sgal, s a pot the RPS, ad p x c s x ( ) the probablty of x gve the th class, whch s calculated usg (4). The classfcato s: cˆ = arg max p X c (6) where, ĉ s the maxmum lkelhood class. I summary, there are two stages the TSDM process, a block dagram of whch s depcted Fg.. The frst stage s the trag (learg) stage whch oly the traed data (sgals) of the dfferet classes (healthy ad faulty) uder vestgato, as well as the umber of mxtures, M are requred. From ths formato, a GMM for each class s developed. The secod stage s the motorg (classfyg) stage whch the mache phase currets are motored, measure ( ) measured, ad classfed based o all the traed models usg the Bayesa maxmum lkelhood classfer. The mache phase currets that are used the trag ad motorg stages of the TSDM have to be frst trasformed to curret space vectors before beg used as the subject tme seres data the TSDM process. As metoed earler, ths techque preseted here s somewhat dfferet from the method used [2], [3] where the radus of gyrato aroud the ceter of mass of the pots the RPS was used as a fault detfcato parameter. For more formato, [2], [3] should be cosulted. Also, aother advatage of ths techque over the oe gve [2], [3] s that oly mache phase currets are requred here for measuremet, whereas the less accessble torque profle was used as the tme seres data [2], [3]. Aga, as metoed earler, oe way to obta the torque data s to use a torque trasducer, whch wll add to the cost of the dagostc system, ad perhaps troduce extra dagostc system relablty questos that oe ca do wthout. Aother way s to measure the phase currets, the termal voltages, ad the motor speed, from whch the torque ca be calculated. I earler work of [2], [3], the values of stator resstaces were kow sce the faults that were beg vestgated dd ot have ay effect o the stator resstaces, ad therefore they were ascertaed beforehad. I other words, resortg to dagostcs based o torque calculato from motored speed, currets, ad voltages or drect torque measuremets, the relablty of the dagostc system wll degrade sce more parameters measuremet wll lkely troduce more measuremet errors to the system, ad cosequetly more errors to the dagostc algorthm. Expermetal results are preseted ext to demostrate the valdty of the proposed TSDM techque. PWM Drve Iducto Motor Utlty sa sb sc Load sa sb sc Space Vector Trasformato s Secod Stage TSDM ONLINE MONITORING & DIAGNOSING Data Acqusto Frst Stage OFFLINE TRAINING Database of Healthy ad Faulty GMMs Fg.. Block dagram of TSDM dagostc process.

4 III. EXPERIMENTAL SETUP Utlty PWM Drve Motor Dyamometer Expermetal data were obtaed o a specally rewoud 230-volt, 2-pole, 2-hp squrrel-cage ducto motor fed from a commercally avalable PWM-verter drve. The motor has a phase wdg that was prepared wth taps to eable expermetal mmckg of cpet ter-tur faults by exterally short-crcutg segmets of the wdg usg exteral resstors of relatvely small value, see the schematc wdg dagram of Fg. 2 ad the actual motor wth tapped wdg Fg. 3. There are a total of sx taps o the wdg, as show Fg. 2, whch allows oe to smulate oe through fve shorted turs out of 08 turs per phase. The stator phase currets data were obtaed at a data acqusto samplg frequecy of 50 KloSample/sec. A complete laboratory setup of the motor-drve system s llustrated Fg. 4. The stator phase currets of the motor uder healthy ad oe shorted tur codto (through a exteral shortg resstace of 0.Ω) are show Fgs. 5 ad 6, respectvely. As may be observed Fg. 5, the three phase stator currets are ubalaced, whch s due to the heret motor structure mperfectos. These heret motor maufacturg mperfectos ad performace measuremet mperfectos Scope Curret Sesors Fg. 4. Laboratory setup of motor-drve system. resstor fuse B A C Fg. 5. Stator phase currets uder healthy codto. Fg. 2. Schematc dagram of stator wdgs wth taps. Fg. 6. Stator phase currets uder oe shorted tur codto. Fg. 3. Actual motor wth tapped wdg. are aturally corporated to the TSDM durg the trag stage, ad hece produces a more relable database for purposes of fault dagostcs through the comparsos wth the data of the fault cases whch follow. By examg Fgs. 5 ad 6, t s deed dffcult to coclude whether a stator fault has occurred because of the above metoed motor mperfectos whch cause phase curret ad flux lkage asymmetres uder

5 the healthy codto. Therefore, dagostc methods such as measurg the egatve-sequece curret or mpedace wll be adequate ad leave much to be desred f the motor mperfectos are ot take to accout. The followg secto wll descrbe the TSDM applcato to the mache phase currets, through ts curret space vector cocept, to accurately detect ad dstgush the type of fault at had. IV. TIME SERIES DATA MINING APPLICATION AND RESULTS The tme-delay embeddg process, whch was brefly descrbed above, was appled to the curret space vector of (3) for the cases of healthy ad oe through fve shorted turs. I addto to the trag set of data, two sets of test data were obtaed usg the same ducto mache for the healthy ad oe through fve shorted turs cases at the same samplg frequecy of 50 KloSample/sec. These two sets of tme seres test data were obtaed from two separate groupgs of samples for each fault over the three secod perod for whch data was recorded. The trag set ad the two test sets were sequetally staggered tme as show Fg. 7. These test data were the processed ad aalyzed usg the TSDM techque. The probablty dstrbutos of these test data wth respect to the trag data are show Tables I ad II, respectvely. It ca be see from Tables I ad II that the classfcato accuracy of the two sets of test data s 00%. Therefore, t ca be cocluded that the TSDM techque ca accurately detect ad classfy the type of fault. Ths preset study, whch establshed the proof of prcple of ths method, wll be exteded ear future work to clude factors such as maufacturg qualty varablty wth the same class of motors of detcal desg ad ratgs. Future work wll also clude usg the TSDM techque for power codtoer fault dagoss, whch wll be very useful from the stadpot of mproved drve desg, relablty, protecto, ad fault tolerat cotrol. Trag Data Trag Data TABLE I PROBABILITY DISTRIBUTIONS OF TEST DATA # Test Data # Healthy Tur 2 Turs 3 Turs 4 Turs 5 Turs Healthy Tur Turs Turs Turs Turs TABLE II PROBABILITY DISTRIBUTIONS OF TEST DATA #2 Test Data #2 Healthy Tur 2 Turs 3 Turs 4 Turs 5 Turs Healthy Tur Turs Turs Turs Turs V. CONCLUSION I ths work, a proposed fault dagostcs method, based o TSDM cocepts appled to the stator curret space vector, for detfcato of wdg ter-tur shorts of ducto motors eergzed from PWM-verter drve systems was preseted. The TSDM techque, whch s based o the tme-delay embeddg process, trasforms the mache phase curret space vector to a recostructed phase space, where hdde patters or formato the mache phase currets uder healthy ad stator ter-tur short-crcut codtos ca be extracted ad classfed. It was show from the expermetal results that the TSDM techque possess a very hgh degree of classfcato accuracy uder healthy ad stator fault codtos, ad hece ca be cosdered as a promsg tool for stator ad rotor fault dagostcs of varous kds. ACKNOWLEDGMENT The authors wsh to express ther deep apprecato for the support redered by Mr. S. J. Dellger, Dr. D. M. Ioel, ad Mr. R. P. Bartos of A. O. Smth Corporato provdg the test motors, as well as the support by Natoal Scece Foudato uder Grat No. ECS Fg. 7. Tme seres trag ad test sets data.

6 REFERENCES [] A. H. Boett ad G. C. Soukup, Cause ad Aalyss of Stator ad Rotor Falures Three-Phase Squrrel-Cage Iducto Motors, IEEE Trasactos o Idustry Applcatos, Vol. 28, pp , Jul./Aug [2] G. B. Klma, W. J. Premerla, R. A. Koegl, ad D. Hoeweler, A New Approach to O-Le Tur Fault Detecto AC Motors, Proceedgs Coferece IEEE Idustry Applcatos Aual Meetg, Vol., pp , 996. [3] H. A. Tolyat ad T. A. Lpo, Traset Aalyss of Cage Iducto Maches Uder Stator, Rotor Bar ad Ed Rg Faults, IEEE Trasactos o Eergy Coverso, Vol. 0, No. 2, pp , Ju [4] R. M. Tallam, T. G. Habetler, ad R. G. Harley, Traset Model for Iducto Maches Wth Stator Wdg Tur Faults, IEEE Trasactos o Idustry Applcatos, Vol. 38, No. 3, pp , May/Ju [5] J. Pema, H. G. Seddg, B. A. Lloyd, ad W. T. Fk, Detecto ad Locato of Itertur Short Crcuts the Stator Wdgs of Operatg Motors, IEEE Trasactos o Eergy Coverso, Vol. 9, pp , Dec [6] J. L. Kohler, J. Sottle, ad F. C. Trutt, Alteratves for Assessg the Electrcal Itegrty of Iducto Motors, IEEE Trasactos o Idustry Applcatos, Vol. 28, pp. 09-7, Sep./Oct [7] R. M. Tallam, T. G. Habetler, ad R. G. Harley, Stator Wdg Tur-Fault Detecto for Closed-Loop Iducto Motor Drves, IEEE Trasactos o Idustry Applcatos, Vol. 39, No.3, pp , May/Ju [8] S. Nad ad H. A. Tolyat, Novel Frequecy-Doma-Based Techque to Detect Stator Itertur Faults Iducto Maches Usg Stator-Iduced Voltages After Swtch-Off, IEEE Trasactos o Idustry Applcatos, Vol. 38, No., pp. 0-09, Ja./Feb [9] F. Flppett, G. Fracesch, C. Tasso, ad P. Vas, AI Techques Iducto Maches Dagoss Icludg The Speed Rpple Effect, IEEE Trasactos o Idustry Applcatos, Vol. 34, pp , Ja./Feb [0] A. Murray ad J. Pema, Extractg Useful Hgher Order Features for Codto Motorg Usg Artfcal Neural Networks, IEEE Trasactos o Sgal Processg, Vol. 45, pp , Nov [] A. Berer, G. Betta, ad C. Lguor, O-le Fault Detecto ad Dagoss Obtaed by Implemetg Neural Algorthms o a Dgtal Sgal Processor, IEEE Trasactos o Istrumet Measuremet, Vol. 45, pp , Oct [2] J. Bagura, R. Povell, N. Demerdash, ad R. Brow, Dagostcs of Eccetrctes ad Bar/Ed-Rg Coector Breakages Polyphase Iducto Motors Through a Combato of Tme-Seres Data Mg ad Tme-Steppg Coupled FE-State-Space Techques, IEEE Trasactos o Idustry Applcatos, Vol. 39, No. 4, Jul./Aug [3] R. Povell, J. Bagura, N. Demerdash, ad R. Brow, Dagostcs of Bar ad Ed-Rg Coector Breakage Faults Polyphase Iducto Motors Through a Novel Dual Track of Tme-Seres Data Mg ad Tme-Steppg Coupled FE-State Space Modelg, IEEE Trasactos o Eergy Coverso, Vol. 7, No., Mar [4] H. D. I. Abarbael, Aalyss of Observed Chaotc Data. New York: Sprger, 996. [5] R. J. Povell, M. T. Johso, A. C. Ldgre, ad J. J. Ye, Tme Seres Classfcato usg Gaussa Mxture Models of Recostructed Phase Spaces, ( press) Submsso to IEEE Trasactos o Kowledge ad Data Egeerg. [6] H. Katz ad T. Schreber, Nolear Tme Seres Aalyss. Cambrdge: Cambrdge Uversty Press, 997. [7] F. Takes, Detectg Strage Attractors Turbulece, preseted at the Dyamcal Systems ad Turbulece, Warwck, U.K., 980. [8] P. Vas, Electrcal Maches ad Drves: A Space-Vector Theory Applcato. New York: Oxford Uversty Press, 992. [9] T. K. Moo, The Expectato-Maxmzato Algorthm, IEEE Sgal Processg Magaze, pp , 996.