You are on page 1of 18
AAS 01-195 REPETITIVE CONTROL METHODS WHEN THE DISTURBANCE PERIOD IS NOT AN INTEGER MULTIPLE OF THE SAMPLE TIME Hao-Ping Wen! and Richard W. Longman* Repetitive control creates spethods of canceling the effects of periodic disturbances on the output of feedback control systems, Fine pointing instraments on spacecraft are often subject (o periodic disturbances from rotating wheels within the spacecraft, and repetitive control offers one option for improving the performance of active isolation systems for stich equipment. Typical repetitive control assumes that the period of the disturbance can be represented as an integer multiple of the sample time of the digital control system. In practice, it will not in general be an integer multiple. It is shown here how this can seriously degrade performance. Interpolation schemes are studied to address the situation. Methods are developed to ensure stability of the learning process over a range of speeds ‘when interpolation is used. It is seen that the process of stabilization is often done at the expense of the final error level reached. However, examples are given in which the interpolation process stabilizes an otherwise unstable repetitive control law. INTRODUCTION Repetitive control aims to improve the tracking accuracy of a feedback control system that is subject to a periodic disturbance. The approach looks at the error in the previous period or periods, and adjusts the command recursively. ‘The error in the previous repetition must be stored for this purpose, and hence implementation of any such algorithm must be done digitally. Spacecraft commonly have periodic disturbances from internal moving parts, such as momentum wheels, reaction wheels, control moment gyros, cryo pumps, etc, With fine pointing instruments on board there can be a need to isolate the instrument from the periodic vibrations produced by such sources. One can expect to get improved performance from an active isolation system instead of a passive one, but still feedback control systems normally only attenuate the effect of periodic disturbances and do not eliminate them. Repetitive control offers the potential to completely cancel the effect of a periodic 2 o where p= p~a= p @, The highest harmonic j that does note exceed Nyquist frequency must satisfy /S(p*—a)/2, so that if p* is even the maximum j=(p*/2)=1, and if it is odd, the maximum j= (p*~1)/2. Note that the 2 times the frst sinusoid on the right of (7) represents the amplitude, and around Nyquist frequency this reaches the value V2. Hence, for frequencies near Nyquist, the difference 2? —1, instead of nearly canceling to zero, actually amplifies. So far we have only considered the forcing function in the difference equation, e.g. (4a). What really matters of course, is how much steady state error is produced by this non zero forcing function. ‘This is determined by fooking at equations such as (Sa) It is saved for the example section below to Took at what the frequency transfer function looks like for this equation, Of course it has the property that the transfer function is zero for the fundamental and all harmonics for period p?'=p*T. And these notches in the frequency response can be rather sharp, Therefore, the true fundamental and harmonics for period p7 can be a long distance up the slope of the notch, sufficiently so that some ‘harmonics are amplified by the repetitive controller. This means that what might seem to be a minor matter, of an error of one half time step in representing the period of the disturbance, can in fact be a cause of seriously degraded performance. LINEAR INTERPOLATION OF THE REPETITIVE CONTROL LAW FOR THE PREVIOUS REPETITION Consider modifying the repetitive control law (3a) to use linear interpolation on each term on the tight hand side CRT) = aul (ke — YT) + Bul(k ~ P*)T) + Olael(k — B+ VT) + Belk P* +1] Be) a=pt-p; p=l-a or more generally Ula) = Hae) + @@)2" E@)] Gp Hy{z)=02+B ‘This learing law has the same form as (3d) that used a zero phase filter (and of course we could choose to include one her also), but in place of the zero phase low pass filter We have the interpolator ‘Hg(z). Then the performance of the interpolation scheme can be predicted by what we will call equations (4c,50,6c) which are obtained from (40,5b,6b) by replacing H,p(z) by H(z). Equation (Ac) predicts the behavior, (Se) gives the steady state error level in terms of its frequency components, after the convergence, and (6c) determines whether the repetitive controller is stable, Examine the behavior of the interpolation filter H(z). First consider the case when corresponding to the true period being one half way between two integers, Then the magnitude of the filter as a function of frequency is tele” = cos(aiT'/2). At zero frequency this is unity, and at Nyquist frequency @ =2/T it is zero, Hence it is a low pass filter. Note that this property implies that deliberately picking 2 sample time so that the period of the disturbance is not an integer number of time steps can be used to stabilize an otherwise unstable repetitive control law, in the same way that the zero phase filter is used for stabilization. Examples of this are given in the example section below. Of course, such an approach is not particularly recommended. It requires that you have the ability to pick the sample time according to the disturbance period, and that the period does not change thereafter. Even if one has these requirements satisfied, the performance is degraded because the low pass filter H(z) is not a zero phase filter. Looking atthe forcing function in equation (4b), using something like a high order zero phase Butterworth filter makes the forcing function very close to zero up tothe filter cut off frequency, and then the forcing function is allowed through, When the Hep(2) is replaced with H(z), the low pass characteristic is no longer a nice sharp frequency cut off, and in addition the phase changes in H(z) further compromise the elimination of the forcing function below any cut off More generally, the magnitude of the H(z) varies with a, 8, and @ according to [Hotel |= 1a? + 20Bcosr + p21! @) ‘and at Nyquist frequency this gives |B ~ cf. Thus, the high frequency attenuation is greatest when the true period is an integer number of time steps plus one half of a time step. IF course it is this attenuation that can produce stability in an otherwise unstable repetitive controller. Graphs of the frequency response are given in the example section below. HIGHER ORDER INTERPOLATION IN THE PREVIOUS REPETITION As the frequency goes up, a linear interpolation of a sinusoid gives an increasingly inaccurate estimate of the value of the sinusoid at intermediate points. By going to a higher order interpolations scheme one should get improved interpolation results, at least for a low to middle range of frequencies. Consider the use of two data points behind and two data points ahead of the point of interest, making a cubic interpolation instead of a linear interpolation. Then Wh = U(kT) + de((k +1)7) Ge) BET) = ay = B -1)T) + agul (ke P°YT) + agul(k= P* +) T)+aqul(k— p* +2)0) a, =~(1/6)8(1 + 8) ay = (1/214 a) BC+) L/2)(L+ ean(l +B) “C/G + aap with 2((k +107) defined analogously. More generally PUG) = MeV + 2 EO) om Hy@)= aye! + ay age + age? Again, the interpolation creates a learning law with the same form as (34), with an interpolation filter taking the place of the zero phase low pass filter. We call equations (4d,5d,6d) the equations analogous to (4b,5b,6b), obtained by replacing Hyp(z) by H(z), and of course these equations predict the behavior of the higher order interpolation scheme, tell how to tune the learning law to obtain stability (using a cutoff if necessary), and telling what the final errorlevel will be. At frequency zero, the filter gives Hg(1)=1, and at Nyquist frequency it gives Hg(-l) =a, +03 ~ a3 +04 =(B-@)[(2/3)a8 +1]. Again, when 05 the result is zero. Graphs of the frequency response are given in the example section below, and they show that there is less attenuation at che lower frequencies when using the higher order interpolation, This should make the forcing function of the difference equation nearer zero for these frequencies, with corresponding better final erro levels. ‘THE SPREADING EFFECT OF INTERPOLATION IN RECURSIVE RC ‘The basic principle used to obtain zero tracking error in these Teaming laws is the discrete form of integration of the errors for each time step in a period, as described by the surnmation in the last of equations (3a). Ifthe ervor is not yet zero, the learning law keeps adding to the control action until it does become zeto (or the system goes unstable). The first form in equation (3a), the recursive form used in real time implementation, is equivalent to the last of equations (3a), the summation form, when there is no interpolation. When interpolation is introduced as in (3e), then the recursive form is no longer equivalent to the desired sum, and the difference should have an influence oa the final error levels reached. It is easiest to understand the effect by looking at an example. Consider time step &=32 and petiod BT =7.67. Then the discrete integral control action desire uG2T)= $le(25.47) + e(17.87) + (10.27) + €(2.67)] ©) where for simplicity we have set the initial period w terms to zero (they are subject to the same spreading process as the errors are). None of these previous ertors are at sample times, so each would have to be interpolated, but it is clear how one would do these interpolations. Now consider what the recursive interpolation formula produces at time step 32 when written out in terms of the associated surnmation G27) = glae(26T)+ Be(2ST)]+ gla*e(19T) + 20fe(187) + Be(17T)]+ GlaP e(12T) + 3a” Bell 1T) + 3082407) + Be(9T)]+ #ar*e(ST) +407 Be(47) + 60 B*e(3T) + 4ap*e2T) + Be) (10) where c= 0.4 and B=0.6, The terms in order are trying to approximate the terms the terms in equation (9). The first term matches. The thd term uses Four points as does the interpolation formula of the previous section, but note that the coefficients are not correct for a four point interpolation. We conclude that using a linear interpolation formula on the previous repetition, causes the summation to spread over @ wider and wider set of time steps as one goes back further in time. One time step is added for every repetition back. This spreading will be faster when a higher order interpolation scheme is used, Since the desired integrating action for each time step is being compromised by the recursive interpolation, we expect the final error levels to be detrimentally affected by the interpolation. Hence, looking at the summation involved is a different approach to trying to understand the final error levels reached, ie. the particular solution of the difference equation. HIGHER ORDER REPETITIVE CONTROL LAWS TO MINIMIZE SPREADING ‘One means of limiting the spreading of the dependence going back in time is to perform the desired interpolation as appropriate for the expression (9) for as many repetitions back as one can handle in real time, and then use the recursive form to include the approximate surn of errors in the earlier epetitions. For integral control based leaming, such a law going back N repetitions takes the focm| WRT) = Glaye((K + 1~ BY) + Bye((k+1- pty) + + axye(k-+1~(NBY JT) + Bye((k +1-(NB)*)TY) oD “+ onyt((k — (BY YT) + Bel — (NBII) In the example seation we will examine in more detail what happens when N is set to 2 repetitions, Noting that p* =f” +1 and (28)" =(@A)" +1, produces the following z-transform version of (3)) (generalized to include ©(2)2” one could also generalize to include a zero phase filter) 22° UG) = Hx MUE)+ Hx @O(227 EO) BH Haale) = ane fa 214 22+ AL 2p)" yale): the ‘Then the difference equation the predicts the performance, the frequency transfer function to look at steady state final error levels, and the sufficient condition for stability are given as (22° — Hay(2)+ Haale) Fine) P22" GRIEG) =[22°" — Hyp()l%y(2)- VO) (4) (apy _ (2) B(e) =| Bn) ____lj—veey (se) 22 — Hyp(2)+ Hop 2)Hya(2)®G@)2" 6), (2B) 12) = Hyg(e)ll— Hyg(2)(2)2" Gte)NEC2 22" Ee) = Hyp (ll = Hig(W2" G@)IE) io {eal ~ Hal)" Gl <1 V @,2= 6 This is a slightly more general form than any of the previous laws, having the extra transfer function ‘Hyz(2). Note that the second term in this transfer function has a term of z to the power (2p)" — pt which is equal to z to the power p or #* depending on whether f is less than or greater than 0.5, ‘This is generally a large power, and it has the effect of causing the phase of this tetm to rotate through ‘many revolutions, producing loops in the Nyquist plot. As seen in the example section below, these loops aggravate the process of trying to satisfy the sufficient stability condition (6e). This appears to be a disadvantage of such higher order repetitive controllers, RELATIONSHIP OF THE APPROXIMATE MONOTONIC DECAY CONDITION TO THE ‘TRUE STABILITY BOUNDARY In all of the above learning laws we have developed the heuristic stability conditions with equation numbers (6) which indicate monotonic decay of the amplitudes of all steady state frequency components of the error from one repetition to the next. Hence, in so far as the behavior in each repetition can be characterized by steady state response, this condition in a monotonic decay condition, suggesting good transients ofthe learning process. This section shows that these conditions are also sufficient conditions for asymptotic stability, independent of whether one can actually think of the error in each repetition as being characterized by steady state frequency response, This section parallels developments given in (4,7). Consider the characteristic polynomial P(e) PC) =a) Hp()P)E Mp2” ~ [Hayy )P)ERp + Ha) Hy 2P2)G)hy 2 = 0 Paz F HDL Hy )OQTER]=1 PQ a2 ¥ Hgla\ll Hye" GO) With proper choice of the quantities p', Ha(z),Alp(z), the above P(e) represents the characteristic polynomial for each of the difference equations (4a) through (4e). As before the subscripts Nand D denote the numerator and denominator polynomials of the associated transfer functions. Using the principle of the argument as in the development of the Nyquist stability criterion, one cam let z g0 around a contour given by the unit circle, a branch cut along the negative real axis out to infinity, encircling at infinity and then back along the branch cut, Ifthe values of P(z) do not encircle the origin as z goes around this contour, then there are no roots of the characteristic equation outside the unit circle and the associated difference equation is stable. One difference between P (2) and P(2) is the division by 2”, all roots of which are inside the unit circle, so this division does not influence the encirclement condition. The remaining difference is the division by the denominators [Ha(2)Hy(2)®(2)GC2)]p. and provided there are no roots of this polynomial outside the unit circle, the same encirclement condition again applies. Note that G(z) is a closed loop feedback control system transfer function, soit should not contain any poles outside the unit circle, The other factors in this denominator are chosen by the repetitive control system designer so that he knows if there are any roots outside the unit circle, and can adjust the encirclement condition accordingly. ‘The third form P* (2) differs by dropping the 1, and hence we ask if it encircles the point +1 instead of the origin to determine whether the system is stable or not. ‘This stability condition defines the stability boundary and it depends on the parameter p'. ‘The ‘good transient conditions (6a) to (Ge) ask that [Holt My2)Oe)2G@)] <1 0,2 ay In practice, the difference between these two conditions is likely to be quite sal fora reasonable number of time steps per repetition, as is shown for certain cases in (8). Suppose that this condition is satisfied and that all roots of [H,(2)Hy(2}®(2)G(@)Ip are inside the unit circle, and that there are More roots of this denominator than the associated numerator (normally required for causality). Then the branch eut and contour at inioity have no influence, and all that is needed forthe encirclement condition is to look at the map of the unit circle, Note that on the unit circle the term 2?” has ‘magnitude equal one, making the magnitude of P"() less than one along the contour. Tt therefore cannot encircle the point +1. We conclude that (11) becomes a sufficient condition for stability of the repetitive controller, in addition to being a condition aiming for good transients in the learning process. It is important to note that condition (11) is independent of the value of p’. Hence, under the conditions of the previous paragraph it establishes stability of the repetitive controller independent of the period of the disturbance or command. This is of course, an important type of invariance for practical applications, Now consider the design process when we want the repetitive controller to function in a stable manner for a range of speeds. And suppose that we are using interpolation in the previous repetition as, in GF) or (Bh). Since H(z) and H4(z) do not have denominators, and assuming wé do not pick an unstable compensator (2) or have an unstable feedback conteol system G2), then condition (11) is a sufficient condition for stability, Although this condition does not depend on p’, it does depend on ©, so that we want it to be satisfied forall at between zero and one. In practice, it is difficult to satisfy (11) for all frequencies up to Nyquist, and therefore one makes use of a zero phase filter H,,(z) in order to cat off the learning at some frequency below that frequency at which condition (11) is first violated, When the true period is an integer multiple of sample times, the parameters of the learning law can be adjusted to make this cutoff at as high a frequency as possible as discussed in [4]. For the current problem where we want the system to work for all periods, not just periods that are an integer multiple of the sample time, the new aspect is to adjust the cutoff so that (IL) is satisfied up to the cutoff, forall ev between zero and one. This forms a simple and effect design approach, EXAMPLES In this section we apply the results of this paper to an example system that is the digital ‘equivalent ofthe transfer function a ot when fed by a zero order hold, and using a sample rate of 50 Hz forming a Nyguist frequency of 25 Hz. The a corresponds to a break frequency of 1.4 Hz, and the @, corresponds to an undamped resonant frequency at 5.6 Hz. This is & rather good model of the closed loop behavior of the feedback controller for each axis of a Robotics Research Corporation robot [4,6]. For our purposed here, it forms a reasonably simple transfer function, but one that is sufficiently complex to be a test of repetitive control techniques. For some examples we will use a different transfer function G2) = 1Mz— 0.9) for simpiicity. Figures 1 and 2 show the Nyquist plot of the stability condition for nearest integer (6a), for linear interpolation (6c), for cubie interpolation (6d), and for linear interpolation two repetitions back (6). Pure integral control based learning is used (3a) with a learning gain of §=0.5. The figures vary the true period from 49.5 to 50.5. For the nearest integer method p is set to 50 time steps, and of course for this case the stability condition is independent of the true period and a and always onds at +1, and appears as the solid line in each plot, The plot of 50.5 is always identical 10 49.5, and in the case of interpolation two repetitions back these are also identical to period 50. Also, for interpolation {wo repetitions back, the plot for period 50.1 is identical to that for 49.6. When interpolation is introduced, the end point is brought in as predicted analytically. Now consider a true period of 49.6 time steps, and vary the linear phase lead parameter 77. Using the learning law (z)zY = 0.52 and varying 7 between two and seven produced guaranteed stability for various values of the phase lead, Figure 2 shows the best case of 7” =4. Linear interpolation using (31) produces a stable learning law for all values of 7, whereas the law without interpolation does not satisfy the stability condition for any values of 7. Hence, we have demonstrated the ability of the interpolation to stabilize an otherwise unstable system, When a cubie interpolation is used instead of linear (3h), the low pass filter characteristic of the interpolation filter is not at sharp (see below), and not all linear phase leads are stable, although leads of 2, 3 and 4 are stabilized by the interpolation filter. Linear interpolation in not enough to produce guaranteed stability from the frequency response condition when using linear interpolating two repetitions back. Figure 4 gives the corresponding plot using (3k) for 7=7. As expected the fast phase change in the second part of #,2(z) creates Loops that make it harder to satisfy this stability condition, and no value of 7 produces stability with this gain. However, decreasing the learning gain to (2) 05 with a linear phase lead of 77=7 does result in stabilization by interpolation, Although with 7 =5 and the original gain, the plot does not seem to go outside the unit circle as far as with 7 = 7, we mast reduce the gain further to (2)= $= 0.01 to produce stability Figure 5 studies the final error level information associated with the frequency transfer functions from command and disturbance to errors, the equations (5). The figure gives the magnitude of the transfer functions in (5a) for nearest integer, (Se) for linear interpolation one repetition back, (Se) for cubic interpolation one repetition back, and (Se) using linear interpolation but two repetitions ‘back. To simplify the interpretation of the plots, the transfer function G(z) =1(z 0.9) is used in this figure. Ifthe repetitive controller did a perfect job at all frequencies, this transfer function would have zero magnitude. Any frequency for which the amplitude is less than unity in Fig, 5 indicates that once steady state is reached (assuming the repetitive controt provess is asymprotically stable), the error at that frequency is reduced by the repetitive control action, As expected, the cubic interpolation produces better attenuation at lower frequencies. But one may pay for this by greater amplification at higher frequencies. OF course, this second part may not be important, since one may be using a ze10 phase filter to cutoff the learning at some lower frequency in order to satisfy the associated stability condition (6). The point of going back more than one repetition before applying the interpolation was to improve this final steady state error level reached. Observe that by comparison to the linear interpolation one repetition back, doing it two repetitions back (Se) does in fact decrease the final value of the ertor, and it does so at all frequencies. However, at lower frequencies it is not as goad as the cubic interpolation in the previous repetition (Sd). This suggests that one might have still beter final ercor levels using cubic interpolation two repetitions back. Of course, going back more than one repetition aggravates the adjustment of patameters for stability, and for this reason my not be. advantageous in practice Figure 5 studies only frequencies that we are interested in learning, the fundamental and harmonics of the true period up to Nyquist, In practice there can be disturbances at other frequencies as well, and itis of interest to know what the repetitive controller does to these other frequencies. The waterbed effect in feedback control system design indicates that one normally must pay for good attenuation at some frequencies by having amplification at some other frequencies. Figure 6 is equivalent to Fig. 5 (but without the repetitive control law interpolating two repetitions back), but shows all intermediate frequencies. Note that the vertical axis in not logarithmic for this plot. For the nearest integer control law, one has notches to zero at the nearest integer fundamental and all harmonics. Since the actual period does not coincide precisely, the true fundamental and its harmonics are shown by an asterisk, and it is these points that appear in Fig. 5. As expected, at low frequencies this partial time step error in the period does not send the asterisk far up from the notch, but as the frequency increases it can go up enough to actually amplify the error. The circles indicate the values when using finesr interpolation in the previous repetition, and the + signs are for cubic interpolation, for those frequencies associated with the true petiod. Looking at the 7 = 2 case, we see the expected ‘behavior nearest point, linear interpolation, and then cubic interpolation giving improved performance in that order for the points indicated. However, the cubic interpolation is capable of producing substantial amplification of other frequencies in the high ftequeney range. In a practical design these might be above the cutoff of a zero phase filter and therefore might not be detrimental In practice, one will usually need a zeto phase filter cutoff in order to obtain stability. Figures 7 through 9 return to the third order example system, and study the frequency at which the plot goes outside the unit circle centered at +1. One must pick the zero phase filter cutoff below this point in order to stabilize the repetitive controller, When the period changes, then the cutoff frequency must be chosen to work for all &, and this is equivalent to having it work for al true periods between 49.5 and 50.5 (whether or not the true period is in this range) shown on these plots. The nearest integer case corresponds to the pint at period 50 applied to all periods, Figure 7 varies the learning gain in the linear interpolation of the previous repetition law (3e), and the nearest integer case (3a) corresponds 10 the one point at period 50, applying to all the true periods plotted. For lower learning gains in Fig. 7 ‘we see thatthe cutoff frequency can reach the Nyquist frequency of 25 Hz as a result ofthe stabilizing effect of the interpolation filter, and for smaller gains more of the plot reaches this limit. Figure 8 considers going back two repetitions before applying linear interpolation (3k) for various constant values of the learning gain 9=@ and a y=7 lead. It is clear that by going back more than one repetition, the cutoff frequency has to be decreased substantially unless the learning gain is small. The left part of Fig. 9 shows corresponding plots for finear interpolation in the previous repetition, but with ‘various phase leads, and the right part treats the cubic interpolation version, ‘There will be a trade off between a lower cutoff needed for the cubie case compared to the linear interpolation case, and better final error levels or better particular solutions thet are reached after the transients have disappeared. Figure 10 displays the magnitude plots of Hy(e) and H,(e@"). Note that the plots for periods 50.4, 50.3, 50.2 and 50.1 are the same as the plots shown for 49.6, 40.7, 49.8 and 49.9, respectively. As expected the linear interpolation attenuates more throughout the range of frequencies ‘up to Nyquist. This contributes to improved particular sofutions for (5d) compared to (Se), and 2lso indicates that the cubic case has less ability to stabilize by interpolation a repetitive controller that is “otherwise unstable. CONCLUSIONS ‘The above developments have demonstrated the importance of addressing the issue of the period of the disturbance in repetitive control not coinciding with an integer number of sample times of the digital control system. Linear interpolation of the recursive repetitive control law terms for the previous repetition is studies as well as cubic interpolation. It is seen that such interpolation can actually stabilize an otherwise unstable repetitive controller, Higher order interpolation is shown to sive better steady state final error levels for low frequency etrors, but is likely to require a lower zero phase filter cutoff frequency to produce stabilization. What might be called higher order repetitive controllers (in analogy to higher order iterative learning control) that go back more than one repetition, are studied as a means to limit the spreading cffect of the interpolation going back in time, I is seen that going back multiple repetitions has a detrimental effect on the design process of stabilizing the learning, Hence, in most applications when using a frequency cut off, it appears that a higher order interpolation in the previous repetition is likely to be preferred to going back more repetitions. In most applications one would be interested in having the repetitive controller behave in a stable manner no matter what the period is of the periodic disturbance, Appropriate stability ‘conditions to satisty for this purpose are established here, and the design process is developed to pick the cutoff frequency to satisfy the stability conditions. ‘REFERENCES 1. R. W. Longman and YC, Huang, “Use of Unstable Repetitive Control for Improved Tracking Accuracy,’ Adaptive Sinactures and Composite Materials: Analysis and Applications, AD-Vol, 45/MD-Vol. 54, ASME, pp. 315-324, 2. R. W. Longman, R. A. Akogyeram, A. Hutton, and J.-N. Juang, “Stability of Matched Basis Function Repetitive Control,” Advances in the Astronautical Sciences, Vol. 105, 2000, 3. HJ. Chen, B, Agrawal, R. W. Longman, M. Q, Phan, and $. G, Edwards, “Rejection of Multiple Periodic Disturbances without Disturbance Correlated Measurements," this proceedings. 4, RW, Longman, “lerative Learning Control and Repetitive Control for Engineering Practice,” International Journal of Control, Special Issue on lerative Learning Control, 2000, 3. R. W, Longman, “Designing Iterative Learning and Repetitive Controllers,” chapter in Iterative Learning Control: Analysis, Design, Integration and Applications, Bien and Xu editors, Kluwer Academic Publishers, Boston, 1998, pp. 107-146. 6, H, Elei, R. W. Longman, M, Phan, J-N. Juang, and R. Ugolett, "Discrete Frequency Based Learning Control for Precision Motion Control," Proceedings of the 1994 TEBE International Conference on Systems, Man, and Cybemeties, San Antonio, TX, Oct, 1994, pp. 2767-2713, 7. ¥-C. Huang and R. W, Longman, “The Source of the Often Observed Property of Initial Convergence Followed by Divergence in Learning and Repetitive Control," Advances in the Astronautical Sciences, Vol. 90, 1996, pp. 555-572, 8. T, Songehon and R. W. Longman, “Comparison of the Stability Boundary and the Frequency Response Stability Condition in Learning and Repetitive Control,” Proceedings of the Second International Workshop on ‘Multidimensional (nD) Systems, Technical University Press, Zielona Gora, Poland, June, 2000, Czocha Castle, Lower Silesia, Poland, pp.121-126. . 5 a g 2 = | \reos=ae ; 25} | esi — soar 05 a 05 1 16 a Tealaxe ' se 5 > g9 j : E | [resce= sor j = |\rem=ar as} |fewot se — Ure eepeion =—== ta ar 05 a ry 1 16 0.5 ie bad 7 2 ea aes Rec aie 1 S j 2 —E fs 05 0 05 1 15 18 Rec aes Fig. | The effect of different true periods, 49.5, 49.7 and 49.9 on the Nyquist stability condition using pure integral control based learning (y = 1,¢ = 0.5) with nearest integer, linear interpolation, cubic interpolation, and linear interpolation two repetitions beck. Imaginary axis Imaginary axis ered = 504 Nearest ining near ineepelaton 2 sac 5 0s 0 08 4 _lnaginary axis ost Imaginary axis Om eeeetiatea, Real axis Fig. 2 The same as Fig. 6 7 05 0 08 1 Real axis | for periods 50.1, 50.3. ost / imaginary axis og Fig. 3 Stabilization by interpolation, y = 4 / § = | [setae : Nowest reper — = Liar inniaten = Z Cubinerlatn ~~ O50 08 4 18 5 Fe afOLtrirO SHEE Real axis Real axis repetitions back, y=7. 15 Fig. 4 Loops when interpolation two Pot EQ@)=| ——+ —" __. — a) © [waar oracas a produces a frequency transfer function, The magnitude of the transfer function tells what attenuation is made for each frequency «@ in the command and disturbance, to form the steady state response after all repetitive control learning is completed In order to adjust the repetitive control law to produce a stable process, we take the following approach, This approach is suggested as a means of obtaining good learning transients in (6), and is a sufficient condition for stability for the repetitive control problem [4,6]. Stability is a property of the homogeneous difference equation (Sa), so we can set the right hand side to zero, What remains ean be ‘written to look like a transfer function from the error at one period to the error a the next period, 2PB(2)=[1-@(@)e" GIB) ~0(2)2"G(z) <1 V @toNyquist.z= (6a) In the second of these equation, we ask that the transfer fonction in the frequency domain be less than ‘one for all frequencies. This indicates that, in so far as the response during a repetition can be characterized by steady state frequency response, there will be monotonic decay of the tracking error amplitudes for all frequency components of the error as the periods progress. In the following sections the above three equations will be successively modified by various filters and interpolators. To keep the relationship to these equations evident, we will maintain the equation numbers (4), (5), and (6) and keep incrementing the letter that follow. Introducing Zero Phase Low Pass Filtering Its difficutt to satisfy (7a) forall frequencies up to Nyquist. As a result we suggest adjusting the parameters in the repetitive control iaw until itis satisfied to an acceptably high frequency, and then cutting off the learning above that frequency using a zero phase low pass filter Hap(2). Here we consider that this filter is applied to the total command to the feedback controller (see [4] for alternatives), so that lap 2UG@) + OG@)z” Be)) Gd), ‘This results is a modified set of equations (4,5,6) Ue? — Hey (2)) + Hep @)O(2)2" GE) = [e? — Hep (D142) - VO) (4b) 2? H(z) E rt - 01 (Sb) O-aR arash onder ” 2P EQ) = Heyl - Oz” GWE) [Pal @(2)27 Ge] <1 ¥ @roNyquistz =e (6b) In the sequel, the repetitive control laws will be modified to address the issue of the true period PT not being an integer number of sample times, i.e. 5 is not an integer. We adopt the notation that (> is the nearest integer larger than p, and 7 is the nearest integer smaller than. (and similarly Gp)* and (j)" for any integer j multiple of ). For each of the modified repetitive control laws we will generate the corresponding equations (4,56) that address the question of final error levels and stability, or the required cutoff frequency for a zero phase filter in order to produce stability POSSIBLE APPROACHES TO THE PROBLEM OF PERIOD NOT EQUAL TO AN INTEGER MULTIPLE OF THE SAMPLE TIME "There are various possible approaches to the problem under consideration: 1, The simplest thing to do is to simply pick the nearest integer number top and use that as the p in any one of the learning faws (3). 2, The next level of sophistication is to apply linear interpolation to the right hand side of a epetitive ‘control law such as (3a) in order to approximate an unknown such as 1u((k— JT) in terms of the stored quantities u((k—p")P) and w((k— p-)P), ‘One might expect to improve the result in Item 2 by making use of a higher order than linear interpolation 4, When interpolating the accumulated signal term such as u((k—p)T) in a recursive repetitive control law, the dependence on errors gets spread over a lager set of time steps as one goes back further in time. To partially handle this matter, one can consider using a higher order repetitive control law that picks out the errors or interpolated errors of interest for a number of repetitions back in time, and then applies the recursive accumulation term with associated interpolation that number of repetitions back. Also, it could be possible to pick a number of repetitions such that no interpolation is necessary on this term, ‘The following sections treat each of these possible approaches. There are two classes of problems that wwe can pose: 1, We can suppose that the period is know, is constant, but just does not happen to be an integer number of sample times. And then we design our repetitive control law making use of this knowledge. 2. A more common situation might be that we want the repetitive controt law to work over a range of periods for the disturbance. Using the zero phase filtering, one can ask for a frequency cutoff that will stabilize the repetitive control process for all periods in some range. The results can then handle fluctuations or drifts in the speed of wheels on board a spacecraft, provided they are sufficiently slow that the quasi static approach to the analysis is justified. In other words, they must be slow compared to the setting time of the repetitive control process. USING THE NEAREST INTEGER ‘When a repetitive control law such as (3a) is applied using a value p that is the closest integer to p, then equations (4a,5a,6a) immediately apply to the problem without any change. ‘The question of stability is unrelated to whether p is actually the true period or not. What is new is that ¥q(z) and V(z) in (Aa) are samples of functions that are periodic with petiod , and hence the 2” ~1 is not taking the difference of periodic function one period apart. There can be as much as one half time step error in going back one period in the difference. Hence, there is a nonzero forcing function in (4a). Some insight into the nature of the forcing function can be obtained from the following considerations, Picking the closes integer means that the maximum error in the representation of a period is one half time step. At Nyguist frequency there are two time steps per period. Hence the maximum error is one fourth of a period. So the z?=1 can be the difference of a sinusoid and the same sinusoid three Arnpltude Neatestineger ersten praca ena Nearest eger veralaion Urea ep) cutie 068) = near een) == 06 1 46 2 28 8 Frequenoy/Sample rate (rsd) o 05 i %8 @ 28 3 Frequency/Sample rate (rad) Prose oa y= 2 Nearest nager eters near (trea) cube (trap) Unser (28) Paseo 7= 5 Nesestinwener eerste Urea ep) cubis 1188) ~ naar @eep) = ost is 2 28 3 FrequencyiSample rate (rad) o 05 4 ts 2 28 3 FrequencyiSample rate (rad) Phase wad y= Neacet nagar ‘teroteters Ursa (1 9p) Amplitude Pass ad y= 8 Nearest integer, Interpetatis Linear 140) cote 11) cube (Ire) — Lines (205) ~ 10 06 1 48 2 25 8 o 05 1 15 2 25 3 Frequency/Sample rate (rad) Frequency/Sample rate (rad) Fig, 5 Magnitude plots for the frequency transfer functions from command or disturbance to error, for phase leads 7 from one to six ($= 0.5) os 1) 18 28 FrequencyiSamole nearest integer Linear interpoiation ~ Gubi nteroolaion os 1 18 228 oS 1S 2 8 Frequency/Sample os 1) 15) 2 28 FrequencyiSample Linear interpolation [Cubic intorotation — o 05 1 «18 2 26 Froquency/Sample 0 05 1 15 2 28 FreauencyiSample Fig, 6 Replot of Fig. 5 showing all intermediate frequencies on a linear scale.

You might also like