len TRANSACTIONS ON INDUSTRIAL ELECTAORICS, VOL. 1S, NO. 1 FEBRUARY (98 8s Observers for Flux Estimation in Induction Machines GEORGE C. VERGHESE, wees, 12, an SETH R. SANDERS, {apelyeasime simulations of machine equa any coreivepredicion eur. We show ta eonectv feedback canbe ‘ed to sped up convergence ofthe fg eimats, and ote hat cam leo reac he sma ofthe estimate o parameter variations. 1. Ietmoovcrion ELD-ORIENTED contro has (along with some variants) emerged as an important approach to the control of 2c ‘machines, and coatinues to be discussed and developed in the literature (se (1}-[3], [27], (28), and references therein). Rotor flux estimation from the terminal variables (stator voltage and current, and rotor speed) isa key step in the most popular implementations of such fild-oriented control for Induction machines. The estimation schemes in use for this purpose are typically real-time simulations of the dynamic equations governing rotor flux. The task of rotor flux estimation may also be expected to arse in other approaches to control and monitoring of induction machines. A framework for understanding and extending fux estima tion schemes ina untied way is provided by observer theory, whichis briefly reviewed in Section I We use this framework in Section II 10 assess the hiterature on flux estimation in induction machines, pointing out the constrained convergence rates ofthe real-time simulations typically used, Section IV begins by examining a realtime simulator based on the equations governing the rotor circuit, as this is a preferred estimation scheme for field-oriented control. We then show analytically, and verify by numerical simulations, that the ‘modifications suggested by observer theory yield fux estima- tors that converge faster. Though present field-oriented control schemes demonstrate high performance, itis believed that there will be situations where improved convergence of| the fux estimator itself wll be called for. Section V outlines a similar development for an estimation scheme based on the Stator circuit. Bot these schemes correspond to reduced-order observers; a full-rder observer is derived in Section VI. The use ofthe observer structure to achieve objectives other than Manuscript reese aly 30,1984; ie Je 4, 187, This work was appari uty the Mindat Power Eton Calle sn st purty a pan for te Gewerl Ect Company. In aiivon, ©. C erphse was supped ty the Soderberg Chi in Power Eeginsing st Mi and Sanders wis Sopp yap IBM Felons. The ahr afe wih the Labrie Yor Electromagrcte nd Elston ‘yma, Mauachatte Insite of Teccogy,Canieae, MA C219, THEE Log Nomber 97181 faster convergence rate, such a reduced sensitivity to parame- ter variations, i also mentioned in these sections. Section VIL discusses issues related to sampled-data implementation of observers. A preliminary report on the work presented in this paper appeared in the conference paper [4], while several ‘details nt shown in the present paper are given in the thesis Is1 Some pioneering work along the directions pursued inthis ‘paper can be found in papers of Bellini era. [6]-[8] and, more recently, [9]. Our results, while similar in several respects (even though initially developed completely independently), are perhaps more transparent. One reason for this is that our evelopment views an observer as a real-time simulation corrected by feedback ofa prediction error term, and thereby permits the observer approach o be directly related to existing ‘ux estimation schemes, There is also significant overlap between our work and the interesting and thorough study inthe recent thesis (in German) of Zagelein [10]. Some of these connections to (61-{10] are noted at appropriate points in the paper. Tr should be Kept in mind thatthe basic methodology of observer design is applicable to a much broader class of problems than the particular (and important) one addressed here, and similar issues may be expected 10 arise in other situations involving estimation and control for clectrical machine systems. One of the objectives of this paper, therefore, is to provide a paradigm for treatments in other 1, Essewris oF Onsenven Tutony ‘Consider a system modeled by the state-space description 2) Ax()+ Bul) + Gd) o where x(@) is an dimensional state vector, x'(0) is its ccomponent-wise derivative, u(t) is an m-dimensional vector ‘of known inputs, and a(t) is 2 k-vector of unknown inputs representing external disturbances and parameter uncerain- lies. Suppose the measured outputs of the system are modeled by HO=CH(t)+ Ha, ® where y(0) isa p-dimensional vector. Here A, B, C, G, and H are all constant matrices of appropriate dimensions Observer theory is aimed at providing a real-time estimate £(0 ofthe state (0) inthe above model, using onl the known signals u(-) and y(-). The theory is well developed in the (0278-0046188/0200-0085S01.00 © 1988 IEEE.86 {RE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL_3.NO. 1 FURUARY 1 ‘comet of linear systems (See {11}, for example) and many results ae knovn in the ease of nonlinear systems as well (ee [12], [13h for example). A straightforward approach to Doviding & state estimate for the model (1) when a(t) is unknown is via real-time simulation of (1) that fgnores the term Ga), namely SU) = ARC + Bul). a However, an observer forthe system modeled by (1) and (2) goes one step futher, in that one corrects the above real-time simulation by use of the discrepancy between the actual ‘outputs (0) of the system and the prediction C&(e) of these ‘outputs thats obtained by ignoring the term (7) in (2). This results inthe system of equations er ® AS(t)+ Bul + KICHD— OL “The term in brackets is called the prediction error, and the matrix K is termed the observer gain. When K = 0, one recovers the simple real-time simulation of (3). Given u(-) and »(-), the systom (4) can be solved by integrating forward in real time from some specified initial condition £(0), using analog and/or digital methods (e.g, using a microprocessor), thereby providing a state estimate The effectiveness ofthe observer is assessed by examining the dynamics ofthe estimation error e()=8()-2(0.- © nis easily seen from (1), 2), and (4) that eW=(AFKO)-(G+KHA). ‘The inital condition for (6) isthe initial estimation error e@), which (even if small) is invariably nonzero because of uncertainties regarding the initial state of (1), Consider frst the case where the disturbance/uncertainty term d{t) can be taken to be 0. The behavior of (6) is then governed by the eigenvalues of A + KC. If we set K = 0, ive if the real-time simulation is nor corrected by a prediction crzor term, then the error dynamics is governed by the eigenvalues of the matrix A, and is thus the Same as that ofthe ‘underlying system (1), For @ sluggish or unstable system (1) this is typically unacceptable, because the estimate ¥(¢) will then converge only sligeishly or not at alo x(). We might ‘expect thatthe error dysamics can be modified to obtain Taster convergence of the estimate if we appropriately pick some nonzero observer gain K. Its out—see [11)—that, under a so-called observability condition on the pair of matrices A, C, an appropriate choice of K can place the eigenvalues of A KC arbitrarily (Subject only t the requirement that complex eigenvalues are specified in conjugate paits). The error dynamics can therefore, in principle, be modified arbitrarily from that of the simple real-time simulator. “The unavoidable presence of system disturbances and model uncertainties, ie. the fae that d(f) is nonzero, causes the terror behavior to nat be governed purely by the eigenvalues of| A+ KC, ass evident from (6). Thus, even if one makes the Simplifying assumption that d(¢) = 0 when carrying out the ‘observer design, one would no attempt in practice to make the fobserver overly fast, A. fast observer would require large values for the entries of K, and these large values would in turn cause large entries in'G + HK, with a corresponding accentuation of the effects of nonzero (0). In fact, the primary objective of observer design often shifts altogether to that of obtaining estimates that are Tess sensitive to distur bances/uncertaimies, rather than obtaining estimates. that converge fast. For example, the most familiar version of the celebrated Kalman filter (se reterences in (11) is precisely an ‘observer in which the gain Kis chosen to give minimum mean square estimation error when di) is modeled as a white ‘Gaussian noise process. Our results in this paper focus entirely ‘on obtaining faster flux estimates in induction machines, but ‘we shall refer to work by others tha is aimed a obtaining less sensitive estimates I is also possible to consider combining these approaches, tading off speed of convergence against lowered sensitivity ‘The review of elementary observer theory given in this section is onl intended to expose the underlying philosophy of using a real-time simulation that is eoereted by a prediction error term. There are many variants ofthe above situation that could be addressed, For example, the measured outputs may be given by a more complicated form than (2) perhaps {volving the inputs. their derivatives, and state derivatives In this case, one may want to construct the observer with ‘modified variables so a6 to avoid differentiation of signals, ‘This will be th situation forthe particular clas of observers wwe derive in Sections IV and V. but we defer the treatment of this case to those sections, rather than attempting 2 general statement here, Similatly for machines operating with vara ble speed. our models wil turnout tobe time-varying, so that the associated dynamics is no longer simply characterized by eigenvalues ofthe governing matrices. Again, we shall defer {rcatment tothe appropriate sections below, rather than trying to develop the requisite results in any generality here. LU, Assessweny of Existing Fux Esm\ation ScuEMES “The literature on flux estimation for electical machines rarely makes a clear distinction (notationally or otherwise) between the state of the system being sted and the state of the estimator self, thus obscuring the issue ofthe behavior of the estimation error. There are some notable exceptions, the sox explicit treatments being {6]-[10}, [14]. [15], and, in a Kalman filter setting, {19]-{21]. The distinction between the ‘underiying state and ts estimate is also made in {161 and (17), where the focus ison the effects of parameter errors. A further striking fact i that most Hux estimation schemes that we are aware of correspond essentially 1 real-time simulations that have mo correction term derived feom the prediction error. ‘This applies to [14] and [15] as well, even though they use the word "bserver™ in thei tiles. Schemes that do use « corrective prediction error include {1}. (6}-{10), 16]. (18}-[22]. As already mentioned in Section the closest in spirit to our work are (6}-[10], and these wi be mentioned again later. The results in {1}, {10}, [16]. and 118} are primarily concerned with reducing sensitivity 10 "uncertainties in parameters (especially the rotor time: constant,which varies substantially with operating temperature) oF t errors in measurements (of speed, for example—see {10] and [18). OF these later four references, the schemes in [10} and [18] have the same essential form as the observers that are Aiscussed later in this paper, though the estimator in [18] isnot presented as an observer. The schemes in (1) and {16}, on the other hand, generate their prediction error terms in a slightly ‘more intricate way, and use these to correct (ona much slower ‘ime-scale than the flux dynamics) the value of the assumed ‘oo time constant in a real-time simulator: they may thus be considered slowly adapting real-time simulators, rather than observers in the sense of this paper. The papers {19]-[22] proceed via an extended Kalman filter or least squares approach and therefore automatically hve a corrective predic tion eror term. We may also note that [23] mentions the use of corrections tothe real-time simulator but does nt elaborate on this at al, ‘The discussion following (6) suggest thatthe dynamies of {he estimation error in all those schemes that are simply rel- {ime estimators is governed by the rotor time constant. We shall exhibit this fact in more detail in Section 1V, and show how to obtain faster error decay by the use of a corrective prediction error signal. An alternative estimation scheme based on the stator circuit is examined in Section V, and is found to have no error decay mechanism! Again, the Use of @ corrective prediction error signal serves to ‘remedy the ‘There are perhaps two main reasons for the neglect of error «dynamics in existing. ux estimation schemes. Festly, feld- oriented contro! has been found in practice tobe satisfactorily robust and effective without faster Mux estimation than that provided by present schemes. Despite this first reason, the isgue needs to be exposed and studied, because there will undoubtedly be applications where eror deay at rate limited by, for example, the rotor time constant may not be at all satisfactory. Secondly, existing theoretical treatments ofthis certor dynamics, such as (6)-(9], [14], {15}, [19]-[22}, ae not ‘easily penetrated. This fact gives us addtional motivation to ‘examine the question of error dynamics thas aleeady been noted in Section I that our study here of observers for faser flux estimation can also serve as a paradigm for similar treatments in the context of other estimation problems in machine systems, There i a clear trend toward replacing sensors that ate expensive, unreliable, oF hhard to install, with less costly measurement schemes and ‘more Sophisticated signal processing; see, for example, the iscussions in {1 (20), [22], (24. [27]. One will 'thus inevitably have to confront the issue of estimation errors and their dynamics in other contexts IV. Onsexver Bastp ow Tar Roror Circurr A.A Real-Time Simulator Consider the idealized two-axis model of a squirre-eage induction machine that is used throughout the literature on field-oriented control (se, for example {I]). The rotor fux «dynamics in this model satisfies ASU-UT + msINU/T IM, ” where by andi, are two-component vectors that constitute the ‘wo-axis representations (in stator-ixed coordinates) of rotor flux and stator curren, respectively, w isthe angular velocity of the rotor (we are assuming without loss of generality that there is only one pole pair), 7 isthe rotor time constant, M is the mutual inductance between rotor and stator, and -([h}eR op @ Note that, i and wean and generally wil all be funetions ‘of time, though we shall not explicitly show the time argument inorder to avoid notational clutter. I should also be mentioned that and can also be taken as complex numbers rather than real two-component vectors, with J and J then being inter preted as | and \/—T, respectively (see [25}, [27). ‘The model (7) motivates the following real-time simulation that constitutes @ common estimation schemes for rotor fux Rye (UTE WHT IM. o Ie is assumed that rotor speed and stator current are known quantities, 30 that the state-space model (9) is in principle simply implemented, with analog or digital hardware 10 perform the required real-time integration of (9) ‘The error inthe flux estimate produced by (9) is v—Wy (10) and is governed by the following state-space equation, ‘obtained by subtracting (7) from (9): =[(-UT I+ wdle. a 'As mentioned in Section IL, this equation is rarely displayed ‘or examined explicitly in the literature, with [8 (9) 15], and {16} being exceptions (and with [6], 17], [10] and [141 Lisplaying the analogous equation for the fullorder observer, see Section VI), For a given speed waveform w, (II) is a linear system. However, since w is in general a time-varying function, the convergence properties of (1) cannot in general be studi by simply taking the eigenvalues ofthe matrix in brackets. When the speed is constant, this matrix becomes constant, In this case, since the eigenvalues of the matrix are —1/T, + jw (whichis most easly seen ifone understands the ismomphism between matrices of the form al + b/ and the complex number a + jb, see [25)), the two scalar components of display an ostllation at the frequency w (¢he constant rotor speed) that is damped with atime constant of T; (the rotor time constant. Significant insight into the case of general. time-varying. isalso fairly easily obtained. Premultplying both sides of (11) by 2e*, where the * denotes transposition (or conjugation, if fe notation is being used), and using the ere)’ and e*Je = 0, we see thatthe squared magnitude ofthe eror satisfies the equation (ere) = -20*ey/T, 2) ‘The magnitude ofthe error thus decays with the time constanta8 Inet TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 35 NO. FEBRUARY 98 fof the rotor. This analysis is really the Lyapunov function ‘analysis in (15), with e¥e as Lyapunov function. B. Using Prediction Error ‘Suppose now that we wish to have an estimate th ‘converges faster than the one above. The philosophy underly ing observer theory naturally suggest that 2 corrective signal derived from a prediction error should be added to the estimator in (9), We are thereby directly led t0 examine the following companion equation to (7). which completes the ‘idealized deseription of the electromagnetic (as opposed 10 ‘echanical) variables associated with the machine, and whic relates stator voltage ¥, (@ two-component vector) f0 stator current and rotor Tus = (M/LDN; +(0L a + Rls «a where o = 1 M3/¢LL), the leakage parameter, with Ly, Zy.and R, being the rotor and stator inductances andthe aor resistance, respectively’ not also that T> = Ly/Ryy where Ry isthe rotor resistance. If the stator volage is measured, ean ‘be compared with the stator voltage predicted on the basis of (13), bur using 4, instead of Be (M/LDRY +L di, 4 Rd «sy ‘The resulting observer then takes the form [UT + WIR H/T Mi, + KE) 1S) where es measured and fis computed from (14). Kis a2 x 2 matrix of observer gains. This is essentially the same ‘observer a6 that of [8], {9]. though our approach to itis somewhat different, It is also closely related to the flux estimator in [18], which willbe Further discussed later i this section ‘We shall shorily show how to avoid having to take derivatives. in implementing (14). (15). Firs however, @ Strnightforwaed calculation shows that the system (11) that governed error dynamics of the real-time simolation is now replaced by (UT) wd les (M/LdKe’ (10) =H (UT )I+wsle. 07) Its evident that different choices ofthe observer gain matrix K wil lead to different error dynamics. Suppore, for illustration, that we pick Kear as) where kis a scalar observer gain parameter. Irthe rotor speed ‘wis constant, then (17) isa time-invariant linear system, and the eigenvalues that govern i ae seen from (17), (18) £0 be (1 EMILY M=V/T,£30) 9) “Thus, the eigenvalues of the error dynamics are sealed up by the factor (I~ KM/L,)-', i, he time constant that governs the error decay is scaled down from that of the convention ‘observer by this factor, while the frequency’ of oscillation in the error decay waveform is sealed up by the same factor. For the more general time-varying case, we proceed as earlier 10 find that (12) i replaced by (orey’ = -20 AML) /T, Wee) 20) so tha the error magnitude now decays with atime constant of a It is evident that & cam be chosen to make ths time constant considerably smaller than T,. (The analysis of the dynamics of the closely related observer in [8], (9] is carried out by invoking Wazewskis inequality, which involves essentially the same ideas.) ‘We now consider how to avoid having to take derivatives in implementing (4), (15). First define the auxiliy variable 1-K(MIL IKK (aL. 2 Then 2” is simply obtained by grouping together all the terms of (15) that contain derivatives (after substituting for 6, from (14), so that (-kM/LOT, L/T)L+ wd |R-+(/T.)Mi,* K(RGi= 0). (23) Now (22) shows that Kee UI K(M/LA e+ K(oL id oH We can substitute this expression for Kin (23) to ge a state equation forthe veto & that is driven only by wand not their derivatives. This can be solved forward from whatever inital condition 2() is imposed by the choice of (0) in 22), To then find A, from 2, one simply uses (24)- Thus. 00 ferentiation of signals s needs to implement this observer. CC. Mlustrative Simulation Results ‘Some results of numerical simulations of the above observer ‘design (using the program SIMINON, sce description in (26)) are shove in Fig, 1. The machine (whose parameters are given in the Figure) was considered to be excited by a sinusoidal 60- He voltage, and had a rotor time constant of 0.18 s, The trace ina shows the speed waveform for operation around 377 rads The trace in (b) isthe actual rotor ux component along axis 1 The traces associated with axis 2, both here and for the ‘numerical simulations inthe rest ofthe paper. are very simi (being, approximately phase-shitted versions of the axis 1 traces), and are not shown, The trace in (¢) shows the rotor flux estimate produced by the conventional realtime simula- tion scheme; (4) shows the estimate produced by our improved scheme above, with the choice K = (L,/2M to give an error {ime constant of 7,/2, according t0 (20) e) plots the error in the conventional estimate; and (Q) shows the error in our improved scheme, where the magnitude of the error decays with ime constant 7,/2, but oscillates a a frequency of 2. It 's evident that our scheme converges significantly faster. D, Extensions ‘The particular gain in (18) was chosen for ease oF illustra tion, Amore general gin, suggested tous by results in [8]. isNw y ‘ew » Penns © “hove ip LAVIN, ae oo 1 a 1 a a Coo a tite (20) @ ror mcomeaioas erat (1) (9 toe melee sense aku ls hd es) If the rotor speed wis constant, the corresponding error system is again a time-invariant linear system. The goveraing eigenvalues are easily computed to be ~(eV/T.+e")+ilew—8/T) 260) where 21=(1—Mky/L,)/((~Mky/L,)? +(Mkx/L)*) (266) B= (Mks/L)/A(—Mk\/L)+(Mky/L,)*1, 266) It isnot hard to see from (26) that by proper choice of ky and ‘ky one can place these eigenvalues at any specified pair of| conjugate locations. It isin principle also possible to vary ky and k; asa function of (a slowly varying) operating speed, to control the variation of error dynamics with the operating point in some desired fashion, in accordance with (26); this Soft of idea is explored (for a fullorder observer, but similar ‘ideas would apply here) in (6). When w is not constant (and not slowly varying), a Lyapunov analysis ofthe kind used t0 obtain (12) and (20) can be carried out fr the present case of| the gain in 25), but we omit the details E, Effects of Parameter Uncertainty and Measurement Errors I is important 10 examine the effects on the observer of errors in measured speed, in the current and volage signal and in the machine parameter values. We shall illustrate an approach to studying such effects by examining the case of uncertainty inthe rotor resistance, a parameter that is known to vary coasiderably withthe operating temperature [1], {16] References to some of the literature on the effects of ‘measurement and parameter errors will also be given, 9 Consider the implementation ofthe rotor ux observer (15) in te case where the assumed value of rotor resistance R, is ‘ot equal tothe actual value R,. The observer then becomes ke =~ Ry/L I+ WIR, + (R/L Mie K (6, ~0). a By subtracting (7) from (27) and using the caleuation REAR = RA BARD RN Rees (RR) es) we obtain the error dynamics ef =U (MILKY (RL NI+ De $US (MILKY 4 MiYR-RVLe. 29) Its thus seen thatthe error dynamics contains a driving term that is associated with the uncertainty in the rotor resistance ‘The error system (29) has the form of the error system (6) discussed in the section on observer theory. Note that the strength ofthe driving disturbance is dependent upon the state ‘or operating point of the induction machine. ‘There are two general routes that may be followed in analyzing this error dynamics. If prior bounds on the machine state variables 2,, i, and on the uncertainty in the rotor resistance ean be determined, then by picking an appropriate Lyapunov function (e.g, €*e) the error e can be shown to be ssympiotically confined toa certain bounded region about the origin, A second approach isto consider the behavior of (29) when the machine is in the sinusoidal steady state. The literature contains numerous studies using this later approach, Garces [16] displayed a similar error system to that of (29) with gain K = 0, but the focus in [16] (and in {1]) was 0 lowly update the assumed value R, 10 reduce steady-state certors in the rotor fux estimate. Vagati and Villata (18) proposed an observer similar tothe one here (though they did ‘ot label ian observer), witha particular gain that resulted in reduced sensitivity to proportional variations in R, and Ry Simultaneously and 10 measurement errors in the speed. tis worth mentioning that this particular gain selection actually lead toa slower estimation error convergence rate than that of the real-time simulator. Bellin ef al, [9] studied the effects of uncertainties in the stator and rotor resistances and the ‘machine inductance parameters. In [9] steady-state estimation error results are tabulated for various operating speeds and for various parameter errors. The gains suggested in (9] for ‘minimizing sensitivity to parameter uncertainties are generally different from K = KI. The thesis of Zagelein [10] contains the most comprehensive discussion of steady-state estimation certor in conventional estimation schemes and in observers of| the type discussed in this paper. It also demonstrates the improved performance attainable by use of an observer in fiek-oriened control of induction machines. ‘There are many other directions in which the ahove scheme needs f0 be further investigated and developed. The study of issues related to microprocessor implementation is of partic0 {EE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL 1,801 FERUARY 1 tar vil terest, see [9], [27], 281, and the comments in Section LV, Ossenvex Basti ov THE Seaton Cicer There are many flux estimation sehemes that may be considered, #8 evident front the review in [10], One natural faternative tothe schemes discussed in Section TV is based on rewriting (13) a8 A= (L/W Ril) (olL/MVi; G0) which leads tothe realtime simlation RP=(L/MM,- Ri) (LLM. 1) “The limitation ofthis scheme that is typically quoted in the literature is the poor behavior at low rotor speeds, Note the auitional fact, however, that the estimation error remains Constant, because (if the scheme is actually implemented as fove) the derivatives of actual and estimated flux are equal ‘here is no mechanism for decay’ of ervor induced by initia uncertainty oF nose during operation. ‘One can now attempt to improve this estimator by feeding in corrective prediction error term. [nhs ease, its (7) that we turn to forthe prediction eeror term. The resulting observer them as the Form Ky (LMM Ri) —(obobs/ MD; + KG) 82) where v andi ae measured andj is obtained frm (7) as eCTIMMK, H/T WIR). G3 The associated error sytem is then eR (TIMDK (e WEVT OF we} BH LU /M)K | MRT IE She. 65) ‘The parallel between this error model and the one in the previous section is evident, Once again, simple choices of K Sch as those in (78) or (25) wil lead to error dynamies that is Substantially different from that of the uncorrected. system, Also, by use of appropriate auxiliary variables. it is again Straightforward to implement the observer without use of silferentiators “The question thal now arises Is how one is. 10 choose between the observer inthis section and that in the previous section, The two are closely related. since they actually result from using the same sts of equations. and the error dynamics ‘obiained by ay particular choice of gain matrix in one yeheme ‘can typically be obtained by some appropriate choice of gain tnatriv in the other. However, it may be that one scheme is tore easily implemented than the other in given situation For example, if one wanted the time constant of error decay 10 be T,. with a gain ofthe form KY, then X would be O in the Hist Scheme bat infinite inthe second, We have not eatried out any Getailed comparative study. VI. Fuut-Oanen Onssaver Te observers presented above are actually examples of what are termed redueed-order observers, because they 40 rot aterapt to estimate oll the state variables in our model of the aystem, A full-order observer would attempt t0 estimate, forexample, i, and w in addition wo , because it isall ofthese variables (or equivalent combinations of then) that together make up the state variables in the model of our sysiem, While fy and ¥ are generally easier (0 measure than yy and are therefore often assumed (asin the previous two sections) t0 be available to us, the price one usually pays with the resulting reduced-order observers for Nis @ strong sensitivity of the festimate A, to any noise in the measurements of J and w. For example, (24) shows that any noise in i, appears (scaled but) unfitered in, For this reason, even if i and w are available as measured signals, one might be motivated t0 construct « fullorder ‘observer, The estimate of i, and that the fullorder observer produces will be filtered versions of the possibly noisy ‘measurements ofthese variables. and therefore could (i the ‘observer is properly designed) be preferable to the raw ‘measurements for control purposes, e.g current: oF speed- ‘control loops. A proper enalysis of this possibilty, however ishard, It would require introducing and studying the effet of noise models in the context of nonlinear models. References {19} and 20} represent beginnings in his direction, Our aim in this section is more modest ‘We shall, inthis section, consider fourth-order observer that produces a filtered estimate of J, along with one of Nox given measurements of J, and w. We shall not attempt 0 produce a filtered estimate of w, a this would get us into the domain of nonlinear (38 opposed to linear, time-varying) models, For this reason, our observer is actally “fll onder” ‘only in the sense that it produces estimates of all the ‘electromagnetic variables in the model. Assuming noise-free ey itis not so hard (0 bwin expressions for the effect of (certain types of) noise in wand i, on the resulting estimates, but we omit this. Its considerably harder to assess the effect (of ine in , because of the way w enters the equations ofthe induction machine model, and we do-not address this further here but see 29) A. The Observer in [14] ‘The “observer proposed in [14] isa fourth-order real-time simulation, without any corrective prediction error term. It generates estimates of both slator and rotor fluxes, using ‘measured values of stator voltage and rotor speed. The present subsection reviews the analysis of [14] (displaying its key Features ina much more direct way than [14)). The next subsection then examines the use of & corrective prediction terror term, based on measurements of stator current The fourth-order system modal under consideration is given Teme 8 SES Ea) where Au. Av mo ae 6o-component vectors representingstator flux, rotor flux. and stator voltage, respectively. The matrices R and L represent winding resistances and machine inductances, respectively, RI 0 Ll MP ed cae 1 the observer is simply a realtime simulation ofthe above system model, as iti in [14], then th error mode is easily seen tobe given by [aren [Sse eo where ¢ is now the four-component vector denoting the error in the flux vector estimate If w is (nearly) constant, the dynamics ofthe errr system is (approximately) governed by the eigenvalues ofthe system in G8). The special structure of (38), see [25], allows us o write the eigenvalues down explicly, as the roots py, Py of the polynomial PHU/To)+(W/T.0)=jwlp+(I-jTAT,T.0) (9) along with their complex conjugates pp, where T, i the ator time constant L,/R, Fig. 2 shows resills of numerical simulations of the ‘observer error transients fo the same machine used for Fig. 1 Near 0 speed, the above eigenvalues are approximately — 2.77 (vice) and ~ 182.0 (owice). These values are reflected in the fertor transient associated with the rotor flux estimate on axis 1, shown in (a) the response is ultimately dominated by the larger time constant, which i 0.36 s (= 1/2.77), Ata speed of 377 radis, the eigenvalues are computed to be at ~93.0 + 4354.0 and ~91.7 + 22.7. Again, these values are reflected in the rotor fux error waveform for axis 1, shown in (6); the time constant of the envelope is now 0.01 s (= 1/917), If w is not constant, a natural route t follow is Lyapunov analysis. Instead of examining the derivative of the squared length e¥e as before, we examine that ofa weighted squared length, e* R-'e. ICs not hard to show from (38) that (eRe) = -2e*L'e) <0 for exo 0) where the inequality follows from the positive definiteness of L (or, les technically, from the fact that L, > O and LL, > Mf). It can be shown from (40) that the weighted squared length of e decays at least as fast as an exponential of time constant (imax eigenvalue of L)/2(min {R., R/}). (41) Applying this bound tothe constant speed case, and comparing with the exact eigenvalues given in the previous paragraph, shows that the bound in (41 tend tobe conservative at higher speeds, but good at lower ones, B. Using Prediction Error We now consider improving the convergence of the real- tion above by use of a prediction error term ot « » Fig. 2 Founder “ober wiht srstine prion ee em oor estan err om ms 138). (a) tO ra and 37a derived from measurements of stator curent. This approach has been taken earlier in [7], but our treatment is rather diferent. is convenient for our purpose to work with a model whose Sate variables are stator current and rotor flux. This is also a Practical choice, since many contol schemes need accurate ‘estimates of these variables, The model is given by EDP 2 averr yf fLewtor Eve aaa a where L2R+MPR\MbL)) «) and Ls w We propose an observer ofthe form a) -al (M/bT)T & ] | Lawn Cytron, (2 ea 0 J i Pero) Patek] oe iy 0} Lasrekeng | EO 4s) here the kare scalars, Note the use of spood-dependent gains in (5). The resulting model forthe observer error dynamics is then eel] cor uweryr | Ltetant on tar ow [bg oy l]- us2 Note that we can freely determine the scalar cooicent in the Tefthand blocks of the two matrices in (46). IF ky and ky are selected such that = STO aE any eee tne save i ra [mr] eats 209=[ OB ar Brew] 89 “The freedom that we have in selecting kz and ky can be vsed to plac the eigenvalues of A in pairs at stbtrary locations, as is easily verified by noting thatthe characteristic polynomial of Ais [p04 kapha RMB 49) Ifthe eigenvalues of A are py (tice) and py (wee), then the eigenvalues of AQ(0) can be shown to be (UT) = poly, and (VT) HI. (50) Inence ifthe machine speed w is (nearly) constant, the error dynamics is (approximately) governed by the eigenvalues in (60), I is therefore clear that inthis case the observer error may be made to decay 10 zero exponentially fast, with a specified time constant. Fig, 3 shows the results of numeric Simulations ofthe observer above, with p; = 2, pa = 10. The ‘waveforms in (a) and (b) are obtained fora speed near 0, and correspond respectively to errors in the estimates of stator current and rotor flux on axis 1. Note thatthe error decay has time constant 0.09 s (= T,/p)). The waveforms in(c) and (4) correspond to these same two quantities, but are obtained for a Speed near 377 radis. The results correlate well withthe above analysis, in that the visible oscillation frequencies are 2 x 377 rad/s and 10 377 rad’s, while the envelope ultimately decays with time constant 0.09 s, as before. fw is time-varying, the eigenvalues do not directly give information on dynamics, and it is natural to attempt a Lyapunov analysis. I isnot hard to show that picking and ‘to give real, positive py and p results in an exponential ‘pounded decay of the error, where the bound has time constant T./{min Py, B:}. Thus, improvement in convergence rate ‘over the observer in [1d] ean be guaranteed in this case as well. Further details may be found in (5) VIL, Sawrtso-Dara IvruewesaTion oF Onseevens In this section we brifly examine some issues related to design of observers for sampled-data implementation. Such an | © jee non » @ Fp. 3. Fuori burr wih corecvepedtio torte, (8. fay Sur coven etna cor om ae Ta Ors () Ror O05 ‘mtn ert ra () Sot caret ematon ror at 377 ral (Ror x eairaon cor 37 al Jimplementaton is natural in the context of microprocessor conteol. Consider first the linear, time-invariant, continuous: {ime system (1) with d(@) = 0, and suppose that its input u(e) is constant over intervals of length 7, starting at ¢ = 0 n=O, 10 un (nT) for nTetetn+ DT, on Its then well known, see for example (26), that the evolution ‘of the sampled stato x(77) is described by the linear, time- ‘variant, discrete-time model MoT sT=FaaT Guta) — 6) ste stare G=[lon Annas 62) entgyersagr erates 60) “The latter matrix scaled the matrix exponential. Note that F and G are independent ofthe diseret ime index n. The model fn (52) is termed the sampled-data model corresponding to the continuous-time system (1) If the underlying continuous-time system i time-varying, of the form (=A BEOMCO, 2) then the above procedure cannot be followed. However, if ‘Ald) and B(t) can be taken as piecewise constant over intervals of length 7, with values A, and By in the mh {interval then one can still obtain a sampled model, except that itis now ime-varying MOTH T)=FextnT)4GuulnT) (54a) where Feces (Aston Go [xp eB.‘The major difficulty with this isthe neod to reeompute the matrices in (S4) at each time step to obtain the model that applies to that step. The main point we wish to make inthis section is thatthe observers we have been considering so far have the feature that very litle new computation is needed at cach step, ‘To se this, return wo the reduced-order observer (23), (24) ‘of Section TV. If we assume that w is constant, and that 0, and J, are piecewise constant, then a time-invariant sampled-data implementation of the form (52) is readily obtained by applying the above procedure. With the choice of gain K given in (18), the matrix exponential involved is exp {(U-RMIL) (1/7) wT} i cos (dT) ~sin (dT) soe [San Stn] 6 where c= (AML) VT, (550) d= (1 kMIL,)'w, (950) ‘The more interesting case occurs when w is nor constant Assuming that w is piecewise constant, taking the value wy in the mth interval one can obtain a time-varying sampled data ‘model ofthe form (54). Note from ($8) thatthe new model for cach time step is easily computed when the new value of w is obtained. The reason for tis is that the matrix exponential in this case is a simple function of w. Similar comments can be made forthe other abservers we have considered inthis paper; see [5] for details Fig. 4 compares the performance of the continuous-time and sampled-data implementations of the reduced-order observer in Section TV. The waveform in (a) is just Fig. 1() repeated, showing the error in the rotor flux estimate produced by the continuous-time observer, while (6) shows the estimate produced by a sampled-lata observer sampling a 10 KH. eis evident that the sampled-data implementation performs well, For comparison, the waveform in (c) shows the eror obtained if one attempts to get away without computing. matrix exponentials at all, but simple uses forward differences 10 approximate derivatives (ie., uses the “forward Euler" method); the result in this instance isa disaster! Further issues related to microprocessor implementation are discussed in (9], 271, and (28), VILL Covewesion This paper has used the perspectives of observer theory 10 ‘examine the problem of flux estimation in induction machines. Related work inthe literature has also been noted, There are several interesting directions in which further studies of state estimation for electrical machines may be usefully pursued, ‘Among these are the development of adaptive observers, extending the work in [1] and [16], and the development of nonlinear observers for estimation of rotor speed as well, extending results such as [19]-(21} Note added 93 ® © Fig, 4. (@ Same as Fig. 1). Eon a sample at inperenttion of iverer, IGAME sapling (e) Err in approtinate sano Implenenaton sing forward Euler mea ‘consult the following very recent work, which extends the results in this paper: 1) ¥. Hori, V. Couer, and ¥. Kaya, “A novel induction machine ux observer and its application to 2 high performance ac drive system," presented at IFAC, 10th World Congress on Automatic Control, Munich, July 1987 2) V. Cotter, “Implementation ofa sampled-data secondary flux observer for an induction motor and minimization ofits sensitivity to parameter variation,” Master's thesis, Uni Tokyo, 1987, Rerenences (MR. Gobel dW Lanta, "Mpc ctl of indton fain Proc. Int Sonicndactor Poses Conrarer Con? {Gri Fp, 3-08 Sa 12) "Kawamori Ho Ae amt of indacson mae fel Steno vec est” Ie Poe Tah BEE PES Rod soy Geass Ml 8a, ee MP cw ant HI” Rips“ Comparon of dae ehinor of eueny cone fed indston machine ve sh ro. nd AC Simp Corral Foner Beco ities! Drs naan p38, So 1 (6) OE" Vegas ind 8 Sat, Obes fr taser ta sein tn thi” a BBE PESC ed 151 SR Sande se tino int Machies"S Mes, ets Doe MT, Ca Mac e188 (6), Belt wad. Fg “A itm eae af he ae of he incon mace," Rete Aural Typ 1 Sti ne (7) Bein. Fat and 7. Lomi, An apc indtion ‘ices yen obec spc hy Con ‘cra Mache ret, ESE Un Ai eho, in azar aes of he nc machin Bro dT Con Blo! Yona Sed ones nk). pe 1788 89, 0) ar mceperer Sl erie be toc cone ai a hPa Eunyon Co Pot Eero ond App rs Oo 9.o 19} wo 1) ue ws) i} eo, Ww. Zien, “Speed epltion of amzeons motors sng SRAM ees urmecr stip.” PRD. diseraton, area xe nest Etangen Nrwberg, 18 60 (erm. Rath, Leer Systems Dew, P Stevens and E, Nos, “Observes fr linear ates ‘oan ded nt Tt. Sptems Scene 10, pp. 8-468 Assan sh ian of omer ters wig SEGSRMEEE Tome atom. Con, vl. ACS, pp. Oe 50, June 1980. Prove xsene of in jl a sbiation of indaction aor My Ree CR ste ober” [EEE Tron Automat. Cont, YOUNES py 21, fe 979 Pie SEiseaion of cooled caren suction mor sive yacela now mniar sate oerer [PEP Trans Ind. Els: Peeve trum a HEC 37, p77, May 198, {PF ce Pua gain fore Seen sie AC Teac ln asquieleage inition mor" TEEE Tons. Ind. ABP Satag-toupp 3-178 Mars ape 198, Niue iota Hebi and. Karosawa, “An pra 0 Aas col of ntcton motor oper wi arale-sreueey Power ees IEEE Tron fad Appl 8 1AT6. pp. 2-39, ey) CN apa ant Vl, "AC servo ste for psa come Ganftetcca Machines Laan). B7-¥T4 Set 984, $e nd "Cond parce a te exit of Sogo Proc AEEE [4S Conf (Cleveland, OF), (Biot. Sem FP illenban, Are for desig the spe and oo a of rgewood Cis, NI Pete Maly In co} pa ey pa es bs en es ry the ayctonaus machine y meus ee ques. In Pros trd AC Symp. Conta! Power Ekeronis and Ekeircol Drives Lasane), pp 8-62 Se 1983 cate F Henan, aed Ch Landpra “A propel oe vote the ie eturoret of sped and ant poste of Be Jochen chine "in roe Ted TEAC Symp. Conrol Power ‘Btcomes Eleseal Drier hasan), pp. 68-08. Sep 1983 Bre onel Ch. Reiss. and SM. Bota, "Ses coal 9 J specs or a nao thie fea by 2 ate soovere" ‘Prac rd trac Symp Contol Power Etonic eral Dives (Cass, p02. 78, Sp. 198%. Minne Nt Hsin. Yas, and T, Nakano, “Perforce Movement cre souce ives fed ition to eu TEE Trans Inds pots wl Anite ap. 205-711, ow De. 1982 Jct ond G. laser, “Cont thas for god eyeame ‘edo idacign nr dees fas on cre and wage 3 Fessred ganic” ZEEE Trans Ind. App 80 1ASI9, p38 5, ane 983. DW Nove and J. H. Woes, “Indacn machine ase ‘Ras snd Jeans pune by mean of cage ine arals IEE Tans Power Appl Sst Yl PASOS. 0.4 pp 125-130, Ihiynag. 1976. Ke Riven and 8, Wiens, Computer Conrled Stems Engiowod Chi, Ni Prem Hal (98 ‘Wi Ceomhr "Msospuier cove! of igh dani efomance eatines A ney taromatieg, wl 22, 0.1 pe 1-20. BF, Bos, Power Ecronis and AC Drives. Bngewoot Cis ic remse a 186 TR's Witty and $1 Marcon, Analsis olin ie od in leads" Frole Iscvol3, 08, 12,29 108 ean Fes 97,