Real-Time Fault Diagnostics for a Permanent Magnet Synchronous Motor Drive for Aerospace Applications

Transcription

1 Real-Time Fault Diagnostics fo a Pemanent Magnet Synchonous Moto Dive fo Aeospace Applications Milijana Odavic, Mak Sumne, Pat Wheele, Jing Li The Univesity of Nottingham Univesity Pak, NG7 2RD, Nottingham, UK milijana.odavic@nottingham.ac.uk Abstact -- This pape descibes the development and implementation of a pemanent magnet synchonous machine contolle fo safety citical applications which embeds independent eal-time mechanical-fault diagnostics into the dive contolle. The dive is intended fo aeospace applications and theefoe must employ an ASIC o FPGA as the main contol pocesso to povide a moe suitable oute to flight-poduct cetification. Moto Cuent Signatue Analysis (MCSA) is employed to indicate the development o existence of faults within the dive system and this is achieved by embedding a eal-time fequency analysis of the moto cuent within the FPGA, opeating independently of the moto contol. Expeimental esults ae povided to validate the poposed contol and condition monitoing. Index Tems Field pogammable gate aay, pemanent magnet moto, fault diagnosis, aicaft maintenance I. INTRODUCTION Pemanent Magnet Synchonous Machine (PMSM) dives ae now being consideed fo many aeospace applications as bespoke designs offe significant eduction in weight and size compaed to altenative dive solutions. The contol method usually employed vecto contol is quite complicated and is usually pogammed on a micocontolle o a Digital Signal Pocesso. Howeve, Digital Signal Pocessos (DSPs) ae constantly being econfigued fo many diffeent tasks (some digital signal pocessing oiented, some contol o potocol oiented) and these shae esouces such as pocesso coe egistes, intenal and extenal memoy, and input/output peipheals. This econfiguation ceates many oppotunities fo the tasks to inteact in unexpected ways and it is impossible to test all the possible scenaios to detemine whethe these inteactions o unexpected latencies between tasks will cause system failues. Fo safety citical applications, this is unacceptable as the opeation of the contol pocessing platfom must be fully defined. Theefoe, fo safety citical applications, altenative contol platfoms ae equied in ode to poduce a contolle with high integity and pedictability which can be cetified fo flight. One possibility is to employ a Field Pogammable Gate Aay (FPGA) which allocates specific pocessing esouces to each individual task such that the tasks can un independently. Each task is pemanently allocated the dedicated memoy it needs to allow easy identification and isolation of bugs in a logical and pedictable fashion. The dawback of using an FPGA fo eal time contol is that the pogamming of complicated algoithms such as vecto contol can be quite challenging. To futhe complicate the challenge associated with developing a PMSM dive fo safety citical applications, it is desiable to include fault detection algoithms within the dive in ode to povide a degee of condition monitoing in the dive system. This condition monitoing also needs to be, to some extent, independent of the moto contol algoithm so that it can povide an independent diagnostic capability. Condition monitoing may be used to detect faults in the moto dive itself (i.e. moto, powe convete o tansduces) o in the mechanical system being diven (beaings, geabox, o the load application itself e.g. udde), and much eseach has been undetaken to investigate signatue based [1-3], model based [4] and expet system based [5-6] techniques fo fault detection. Moto Cuent Signatue Analysis (MCSA) [1-3] is a pomising appoach to fault detection as it diectly identifies changes between the hamonic content of the moto line cuent when known types of fault occu and because the stato cuents can be easily accessible in nealy all industial applications. Howeve by embedding the MCSA into the dive contolle thee is no need fo additional monitoing of the stato cuents and theefoe the applicability of the MCSA is even wide. Faults associated with beaing failue o geabox degadation can be identified by looking at specific fequencies (which ae a function of speed) in the moto line cuent and deciding whethe thee is a significant change compaed to the healthy condition. Howeve, the key to implementing this type of condition monitoing appoach in eal-time on an FPGA is to educe the on-line computational equiements of the condition monitoing algoithm and this demands inteaction with the contolle. This pape descibes the development of a PMSM dive which can be used fo safety citical applications. The next sections descibe the fequency detection algoithms evaluated, and intoduce the appoach using multiple otating efeence fames. The fault decision algoithm is then discussed. The implementation of the vecto contol algoithm and its associated components on a hadwae contolle ae biefly descibed and expeimental esults pesented show the effectiveness of the esultant system. The

2 dive is then extended to incopoate condition monitoing based on MCSA such that opeational conditions ae used to educe the computational ovehead of the MCSA, whilst the independent diagnostic capability is still etained. II. REAL TIME FREQUENCY ANALYSIS The main task of the MCSA is to detect any change in magnitude and fequency of the sideband components of the moto cuent which elate to a paticula element in the dive, compaed with the pefomance of a healthy machine opeating unde the same conditions. Howeve the MCSA scheme should have enough esolution to detect small changes in the system, since this allows the detection of incipient faults befoe catastophic failue occus. The key to achieving this is to allow inteaction between the contol and fault detection blocks at the high level, without educing the integity of eithe algoithm at the low level. This can be achieved by using the measued speed o oto position as pat of the fault detection algoithm. This will educe the memoy and computational equiements of the fault detection algoithm as will be discussed in the next section. The amplitudes of the specific fault elated fequencies ae calculated in eal time and can then be compaed to thesholds set duing a commissioning stage. A flag can be set accodingly to indicate the development of a fault. In any sampling pocess aliasing effects must be consideed. In ode to pevent high fequencies becoming aliased down to low fequencies it is necessay to pocess the analogue signal befoe sampling by employing an antialiasing (low pass) filte whose cutoff fequency should be chosen elative to the sampling fequency. Howeve by embedding eal-time fault diagnostics into the dive contolle anti-aliasing filtes can be avoided: sampling can be easily synchonized with the switching fequency of the powe convete and this effectively means that the aveage value of the cuent ove one switching peiod is measued, eliminating all switching fequency elated hamonics in the sampled signal. Thee appoaches to eal time fequency analysis fo fault detection have been consideed the Full FFT, the Zoom FFT, and an appoach employing efeence fame tansfomations. The fequency analysis methods have been evaluated on an expeimental pemanent magnet moto dive (descibed in Section IV) which is connected to its load via one of two geaboxes. The fist geabox is healthy. The second has one tooth missing. The moto line cuents have been sampled at 1kHz and subsequently pocessed. A. FFT analysis The initial appoach to MCSA in this wok is to employ a Fast Fouie Tansfom (FFT) to the moto line cuent. Figs.1a and 1b show the time domain and fequency domain content fo the measued moto cuent when the healthy geabox and faulty geabox ae used espectively. The esults ae shown fo the case when one gea tooth is missing in the faulty geabox. The expected distubance fequencies should be n f [1-3] (n=1,2,,) whee f is the oto speed, fo this test 2Hz, and f e is the electical fequency, fo this test 6Hz). Fo this initial wok, all esults wee measued ove two second duation in steady state, giving the FFT esults a fequency esolution of.5hz. When compaing the hamonic content of the moto cuent in the pesence of the faulty geabox with the moto cuent with the healthy geabox, lage amplitudes at f e ±f and f e ±2f can be obseved in the stato cuent fo the faulty geabox with the missing tooth. The amplitudes of the components at f e ±f and f e ±2f fo the faulty geabox ae (.2A,.34A) and (.2 A,.24A) espectively. They ae moe significant than those of the healthy geabox which ae (.1A and.27a) at f e ±f and (.16A and.177a) at f e ±2f espectively. These ae vey small changes but if these specific speed elated fequencies can be accuately measued, geabox condition monitoing can be achieved. Amplitude[A] FFT spectum[a] PM moto cuent[a]:healthy Condition Time[S] Fe-2F Fe-F.6 Fe-2F Fe-F Fe+F Fe Fe+2F 2Fe 4Fe 5Fe-F FFT esult 5Fe 7Fe fequency[hz] a) healthy geabox PM moto stato cuent:[a:]:with Faulty Geabox 11Fe Time:[S] Fe Fe+F Fe+2F 2Fe 4Fe 5Fe-F 5Fe Fequency:[Hz] b) faulty geabox Fig. 1 Measued stato cuent esults with f =2Hz and f e=6hz fo a) the healthy and b) the faulty geabox. A two-second duation fo the sampled stato cuent with a sampling fequency of 1kHz (a typical switching fequency) equies a 2k-point FFT which goes beyond most pactical DSP and FPGA computing capabilities. The implementation aspects of a pactical FFT ae discussed 7Fe 11Fe 13Fe 13Fe

3 below. The esults shown in Fig.1 ae sampled and ecoded within the FPGA contolle and ae pocessed off-line using the MATLAB envionment. When using the FFT to pefom a fequency analysis, two associated poblems must be consideed: 1) the measuement time must be limited to a ealistic value and 2) the FFT povides esults only fo cetain discete fequencies. The FFT makes an implicit assumption about the signal outside the measuement inteval by assuming that the signal within the measued time epeats fo all time. This can add discontinuities into the analysed signal that do not eally exist. In the fequency domain this will cause the signal's fequency spectum to spead out and will especially effect the detectability of a small signal in the pesence of a lage one. The spectal leakage is of paticula inteest when the machine opeates at low speed as the detection of fault fequencies (i.e. sidebands hamonics) may be lost due to the leakage aound the cuent fundamental fequency, as demonstated in Fig.2 fo the otational speed F of 4.5Hz and the coesponding electical fequency F e of 13.5Hz. The sideband hamonics ae difficult to detect due to the spectal leakage effect as seen in Fig.2a. i [A] i [A] f [Hz] F 2F 3F a) Leakage effect f [Hz] b) q component of the measued stato cuents Fig.2 Measued stato cuent fequency spectum with f =4.5Hz and f e=13.5hz a) in the stationay efeence fame and b) in the synchonous efeence fame. One solution to educe leakage effects is to apply a technique known as windowing to educe the leakage effect. Howeve, this adds to pocessing equiements. Altenatively it can be shown that the spectal leakage can be significantly educed by analyzing the coesponding components of the stato cuents in a synchonously otating fame (which otates synchonously at the fundamental fequency), as shown in Fig.2b. In this fame, the stato cuent consists of small oscillations at otational (and its multiple) fequency supeposed onto a dc component (the tansfomed fundamental ). The impoved claity of the sidebands compaed to Fig. 2a is evident. It must be stessed howeve that this appoach uses a vey lage data set to achieve the desied fequency esolution fo fault detection. Due to the limited pocessing capability of DSP/FPGA pocessos, only the shot length FFT (i.e. using a maximum of 248 data points) can be implemented in pactice. One pactical appoach could be to divide the signal into seveal pats and then to apply a limited-sample FFT to each pat. The esults ae simila to the full FFT but this appoach suffes fom poo esolution [7]. With a 248-point FFT and the sampling fequency of 1 khz the fequency esolution is only Δf = f s /N= 4.9Hz. Howeve the specific fault elated fequencies, which ae of inteest fo this wok, ae sidebands aound fundamental and/o mesh fequency of the stato cuent. Theefoe athe then calculating the FFT fo the entie fequency ange, it can be calculated fo a cetain fequency band. This appoach (i.e. the fequencies zoom) is divided into the following steps: fequency shifting, ideal low-pass filteing and down sampling [7], [8]. By applying the zoom FFT on a naow fequency band (i.e. 5Hz) aound the fundamental/mesh fequency of the stato cuent, the sampling fequency can be educed (i.e. 1kHz) which effectively impoves the fequency esolution (fo this case Δf=.49Hz). B. Fequency extaction using fault specific efeence fames This wok is focused on finding the best method fo ealtime fault detection fo PMSM machines which can be embedded within the contolle. One appoach to this poblem employs multiple efeence fame filtes; one fo each specific fault elated fequency. The fault elated fequencies in the stato cuent can be defined as [3]: F = F ± m F ± q F fault e mesh = p F ± m F ± q N F (1) = F ( P ± m ± q N ) = F i whee P denotes the pole pais of the PMSM, N denotes the numbe of gea teeth and m,q =,1,2,. In any otating efeence fame the synchonous hamonic (i.e. the one used fo the tansfomation fom the stationay to the otating efeence fame) will be seen as a dc value, while all othes hamonics in this efeence fame appea as altenating components. The fequency of an altenating component is equal to its own fequency plus the synchonous fequency if they both otate in the same diection o its own fequency minus the synchonous one if thei sequences ae opposite. These altenating components can be filteed out using a low pass filte o a Moving Aveage Filte (MAF) with an adjustable window elated to the otational fequency of the PMSM. The cuents ae tansfomed fom the abc-stationay efeence fame to the i th synchonous otating efeence

4 fame by the tansfomation matix: cos( i Θ ) cos( i Θ ) cos( i Θ + ) (2) sin( i Θ ) sin( i Θ ) sin( i Θ + ) whee Θ denotes the instantaneous mechanical oto position angle given by: Θ = Θ + ω t (3) whee Θ is the pemanent magnet flux position at t= which geneally is not known. In ode to detemine an initial position fo the oto flux angle, a small i d cuent (epesenting a flux efeence) is demanded at low voltage and Θ =. This will move the oto d-axis towads the imposed stato flux wheeby this movement pesents the oto flux offset Θ. This can be epeated fo a seveal oto positions and the aveage value of the oto position eadings is consideed as a oto flux offset. As a geneal consideation, failues in the mechanical system being diven by the PMSM dive (such as failues in the beaings, geabox, o the load application itself) cause load toque oscillations. Theefoe the measued oto position angle contains the same fequency vaiations as the speed due to this vaiation in load toque. The fequency spectum of the moto speed when one tooth of the geabox is emoved is illustated in Fig otating speed W [ps] obtained fom a synchonous filte otating at (P+1)F is shown in Fig.5. The initial tansient peiod obseved depends on the MAF window length. Howeve this efeence fame is aligned to the fundamental positive-sequence oto position initial angle and theefoe both d and q dc components exists. In ode to obtain the amplitude of the hamonic at F e +F, the squae oot of sum of the squaes of the d and q components needs to be calculated. Fig.4 Block diagam of the multiple efeence fame filtes i [A] X:.7294 Y: t [s] Fig.5 Amplitude of the stato cuent hamonic component F e+f fo the faulty geabox using the synchonous filte otating at (P+1)F Amplitude[ps] FFT spectum F Time[S] 2F Fe 4F 5F 2Fe 8F 3Fe 11F 4Fe fequency[hz] Fig.3 Measued speed with the faulty geabox (F -otational fequency, F e- electical fequency) Theefoe the oto position angle needs to be filteed to obtain the positive sequence component of the fundamental oto position Θ _ 1. This is obtained using a Moving Aveage Filte (MAF) with an adjustable window T w elated to the otational fequency of the PMSM. Since the angle inceases with a constant slope k, the output of the MAF includes a constant eo, kt w /2, which needs to be compensated fo [7]. The block diagam of the poposed fequency detection algoithm is shown in Fig.4.. The amplitude of the fault elated hamonic F e +F 6Fe III. CONDITION MONITORING Fom the pevious discussion it can be seen that the use of fault specific otating efeence fames povides an appoach to condition monitoing which can povide an accuate estimate of the amplitude of a fequency component with a vey good esolution without the extensive pocessing and memoy equiements of the FFT with a compaable fequency esolution. The pocesses used fame tansfomations and filtes ae standad blocks used by moto contolles. The impoved use of pocessing capabilities is achieved by using a measue of the instantaneous oto position, available fom the contolle itself, o fom the position esolve itself. Subsequent fault detection pocessing can be made independent of the contolle, and seveal paallel otating efeence fames with diffeent fault elated fequencies can be employed independently in paallel. The amplitudes of faulty fequency components obtained by the synchonous fames filtes can then be sent to a highe level contol boad (e.g. a Remote Electonic Unit) whee fault diagnosis can be made. This appoach is vey simple and theefoe shows a geat potential to be embedded into the dive contolle to povide eal-time fault indication of the

5 PMSM when known types of faults occu. An example of a eal implementation of the PMSM contolle togethe with the fault diagnostic algoithm is given in the next section. IV. EXPERIMENTAL IMPLEMENTATION A Field Pogammable Gate Aay (FPGA) has been employed as the main contol pocesso fo the pototype PMSM dive. The block diagam of the FPGA-based sevodive system fo the PMSM dive is shown in Fig.6. The system includes an uncontolled diode bidge, a DC link capacito and a PWM invete, a 4kW synchonous moto, a DC moto woking as the load and a powe convete fo the DC moto. The machine is a 6-pole pemanent magnet suface-mounted moto. The FPGA based contol pocessing platfom is shown in Fig.7. The caie fequency of the PWM invete is 5kHz. A. FPGA-based contol A low cost Actel Poasic3 A3P1 FPGA was used as the taget component fo the implementation of the contolle. The algoithm is implemented using fixed-point aithmetic with a vaiable wod-size. A vaiable-wod is used to impove a calculation pecision without inceasing extensively the FPGA-chip aea. A block diagam of the FPGA-based contol algoithm is pesented in Fig.8. Fig.8 Block diagam of the poposed FPGA-based achitectue of PMSM motos Fig.6. Block Diagam of the Moto Dive System Fig.7. FPGA contol boad The PMSM moto is contolled using a standad vecto contol algoithm [9] enhanced by incopoating on-line mechanical fault detection implemented sepaately inside the FPGA. The aim of the vecto contol appoach is to contol toque and flux independently by contolling a toque poducing component of cuent (iq) and a flux poducing component of cuent (id). id and iq ae deived fom mathematical abc-dq tansfomations made on the instantaneous measued moto line cuents (ia, ib, ic), using the instantaneous (electical) oto position angle θ e. The toque demand is deived fom the oute speed loop. Fom this block diagam it can be obseved that computation concuency can be used to pefom the mechanical-fault diagnostic module independently of the machine contol module. The ADC inteface module contols the analogue to digital convesion and aanges the conveted data into the 14-bit fomat equied by the PMSM contol module. In this system, samples of thee line cuents and the dc-link voltage ae fed to the PMSM contol module at the beginning and in the middle of a PWM peiod. The esolve inteface module computes the electical position fom the absolute oto position; futhemoe it adds an offset position to align the calculated electical position with the diect d axis and finally aanges the esulting position into a fomat expected by the PMSM module. The contol algoithm is coded in Vey high speed integated Hadwae Desciption Language (VHDL) [1-12]. The algoithm is executed with an FPGA clock peiod of 2ns leading to a low computation time fo each elementay module - the PMSM contol module takes less then 1us to pefom calculations in spite of its complexity. This module compises: the speed PI contolle including speed calculation fom the oto position, the electical position sine and cosine calculations, and the d and q PI cuent contolles combined with decoupling tems to povide efeence voltage demands fo the Space Vecto PWM module. The algoithm is executed in a sequential ode. The SVPWM module povides an efficient conventional space vecto pulse-width modulation (SV- PWM) algoithm implemented in a way to educe the computation [13]. The expeimental vecto contol scheme

6 woks successfully as pesented in Fig [pu] load change toque demand speed t [s] Fig.9. Expeimental esults of the laboatoy setup moto dive system B. FPGA-based fault detection module The fault-detection module pocesses measued moto cuents and a oto position angle. The cuents ae sampled twice pe switching peiod: in the beginning and in the middle of the peiod and then passed to the synchonous filtes (efe to Fig. 4) to extact faulty-fequency-hamonic amplitudes. The oto position angle is filteed though the MAF with an adjustable window (elated to the cuent speed of the PMSM) to obtain a positive-sequence oto position angle. Futhemoe to obtain both sine and cosine fom the electical angle, a sine look-up table (LUT) has been implemented. The outputs fom the synchonous filtes ae filteed though the MAFs with an adjustable window to extact thei dc values and to obtain the coesponding hamonic amplitude. At pesent the outputs fom the synchonous filtes ae pesented to simple compaatos to detemine if the theshold fo a paticula hamonic has been exceeded. The theshold can be adjusted if necessay as a function of speed and load (if iq is passed fom the contolle to the fault detection outine). The fault decision making algoithm is shown in Fig 1. Fig.1 One channel of the fault decision making algoithm. V. CONCLUSIONS This pape has pesented an appoach to eal time fault diagnostics fo the PMSM moto dive systems poposed fo moe electic aicaft. The algoithm can be embedded into the moto contol FPGA and can opeate independently of the contol, using measuements of moto cuent and moto position and can be used to detect one of seveal diffeent types of mechanical fault. The algoithm uses otating efeence fames employed at the fequencies of known key faults fo the dive system. This appoach uses simple pocessing block aleady developed fo moto contol, and does not need lage data sets to obtain a suitable fequency esolution. A simple diagnostics algoithm has been poposed based on compaing the outputs of each otating efeence fame with a theshold, which may be changed as a function of fequency and load. Futhe wok is equied to detemine how thesholds can be eliably set fo diffeent moto dives, and whethe futhe noise eduction can be intoduced in ode to educe the sensitivity of the fault thesholds. ACKNOWLEDGMENT The authos gatefully acknowledge the suppot fo this wok fom the EU as pat of the SCARLETT poject. REFERENCES [1] M. Blodt, M. Chabet, J. Regnie, J. Fauche, Mechanical load fault detection in induction motos by stato cuent time-fequency analysis, IEEE Tansactions on Industy Applications, vol.42, no.6, pp , Nov./Dec. 26. [2] Chinmaya Ka and A.R. Mohanty, Monitoing gea vibations though moto cuent signatue analysis and wavelet tansfom, Mechanical Systems and Signal Pocessing, Volume 2, Issue 1, Januay 26, Pages [3] Shahin Hedayati Kia, Humbeto Henao, Géad-Andé Capolino, Analytical and Expeimental Study of Geabox Mechanical Effect on the Induction Machine Stato Cuent Signatue, IEEE Tansactions on Industy Applications, vol. 45, no. 4, July/August 29 [4] Rolf Isemann, Model-Based Fault Detection and Diagnosis-Status and Applications, Annual Reviews in Contol, Vol. 29, pp.71 85, Elsevie, 25. [5] P. Ballal, A. Ramani, M. Middleton, C. McMuough, A. Athamneh, W. Lee, C. Kwan, and F. Lewis, "Mechanical fault diagnosis using wieless senso netwoks and a two-stage neual netwok classifie," pesented at Aeospace confeence, 29 IEEE, 29. [6] Bellini, A.; Filippetti, F.; Tassoni, C.; Capolino, G.-A.; Advances in Diagnostic Techniques fo Induction Machines, IEEE Tanssactions on Industial Electonics, vol. 55, no. 12, Decembe 28. [7] Richad G. Lyons, Undestanding Digital Signal Pocessing, New Jesey, Pentice Hall, 24. [8] Robet B. Randall, State of the At in Monitoing Rotating Machiney Pat 1, Sound and Vibation, Mach 24. [9] Pete Vas, Vecto Contol of AC Machines, Oxfod: Univesity Pess, 199. [1] H.C.Roth, Digital System Design Using VHDL, in Poc.PWS Conf.,Boston, MA, [11] Mohamed Wissem Naoua, Ahmad Amma Naassani, Eic Monmasson, Ilhem Slama-Belkhodja, FPGA-Based Pedictive Cuent Contolle fo Synchonous Machine Speed Dive, IEEE Tansactions on Powe Electonics, vol. 23, no. 4, July 28. [12] Ying-Shieh Kung, Ming-Hung Tsai, FPGA-Based Speed Contol IC fo PMSM Dive with Adaptive Fuzzy Contol, IEEE Tansactions on Powe Electonics, vol. 22, no. 6, Novembe 27. [13] Zeliang Shu, Jian Tang, Yuhua Guo, and Jisan Lian, An Efficient SVPWM Algoithm With Low Computational Ovehead fo Thee-Phase Invetes, IEEE Tansactions on Powe Electonics, Vol. 22, No. 5, pp , Septembe 27.