2minte, Ke) (2099) where Ky = (0). Proof. Consider the models h = mine, Ko), consistent withthe null ouput Equation (10.99) follows immediately from the general results in Section 102 a ‘The above bound is independent of the duration of the experiment and the nature of the input sequence; therefore it represents an absolute lower limit on the error. Next, we obtain bounds for the impulse response experiment that depend on the number of samples taken (230) Lemma 10.14 The information diameter for the &, identification procedure with an impulse response experiment is given by: Dy(Z) =2[ nine #00) + $290] (10.100) Proof. Define the following: fay = {Bh ESA nt = me) Next, consider the pair of models and measurement noise vectors: {mer= ay and {rns =to nk) = —n{k) math) = n(k) = hymeS, mmen Hence the above produces zer0 outputs to an impulse excitation, that is, Ehy.m) = Ella, m2) = 0. Note that this isthe worst-case model (maximum f, norm) compatible with the experimental output and in the consistency set S(0). Equation (10.100) follows now from Lemma 103 and the fact that $ and N’ are symmetric and convex by considering d(S(0)] = Wy ~ fal, = 2. (A) and « being positive a The following comments are in order: # There is a significant quantitative difference between the bound in (10.100) and the global one in (10,99), the latter being a conservative lower bound for the identification error. TIMEDOMAIN IDENTIFICATION 365 + The diameter of information tends to zero as M — 90 and € -» 0. This ‘guarantees the existence of convergent algorithms, for example, inter- polation procedures. ‘© For a given noise bound level, there exists 4 maximum number of ex: perimental samples M,, which decreases the diameter of information. In other words, for M >'M,, the diameter remains constant, which means ‘hat any extra experimental sample will not increase the amount of useful information. This fact will be used in the formulation of an interpolatory convergent algorithm ((230)) in the next subsection, Under certain conditions ((163}) it can be proved that the impulse input sequence is optimal in terms of the diameter of information, for a given experiment duration M. These conditions are related to the fact that the 4@ priori information does not exclude models in the consistency set of the null experimental outcome. The optimality stems from the fact that this in- put produces outputs directly related to each time coefficient. Therefore all the uncertainties are decoupled, that is, each measurement noise element is related to only one model coefficient. For example, for the step response experiment, the reader may verify that the diameter of information satisfies Dy (Z) ~ minle, (0)]+2 > minle,@]+ S24) (10.101) ‘The proof is based on the same arguments as before, but using the following model: min{(0),€), k=0 i(k) = } 2(—1) minf(K),€], 1k 2. Since AC) is based only on the experimental information, it cannot distinguish between both plants, but should converge for all plants in S(O), Therefore if sy lim, AQ) ~ hol = 0366 ROBUST IDENTICATION = Mo, such that [AQ <1, M > Moy € <6 Asa consequence for M > My and € < y, WAG) — Ath > Wall ~ AQIIh 2 1 (10.108) which contradicts the assumption that A converges to a plant in S(0) (see also (160) a In [237] the following necessary and sufficient condition for the existence of a robustly convergent ¢; identification algorithm is established: 36 >O-such that lh «ull > Sli, Yh Eb (10.104) ‘here lulas < 1 without loss of generality. The above condition excludes non persistent inputs for which u(t) — 0 as f — oo. Specific untuned algorithms, based on experiments performed with inputs known as Galois sequences, have been proposed in (193). Furthermore, in [238 it is proved that there are no linear robust convergent algorithms in f, as was the case for Ha, identification. 10.4.2 Identification Procedures Next, several tuned convergent algorithms for f, robust identification are presented. These are based on the impulse response experiment although, with minor changes, can be applied to the step response experiment as well. ‘The fact that these algorithms produce a nominal model in the consistency set makes it unnecessary to test consistency, as long as such a model exists. ‘Untuned convergent algorithms for additional input signals, can be found in [141, 193, 194, 306} Based on the Firstn ine, (K)] (10.106) Proof. This proof follows a similar one in [146]. For every h € S,n € N and for any algorithm A, the following holds: We AQ) < eA) (10.107) where e(A) Is the worst-case local error defined in (10.108). Consider the following models and measurement noises {nw = min((k), ¢} {ne —hy(k) m(k) = -fu() am{k) = ~ha{k) => nS, mmeN, Eth m) = Elta, m) Now, applying equation (10.107), Wry AO), < eC) and Way + A(O)I, < (A) => [hall <5 lr ~ ADI, + $s + ACI, < CA) To obtain (10.106), simply take the infimum over A. a Lemma 10.17 For the identified model Higa) = yz’, nM the following ienifation error bound holds: Wh Hh) < ne + S01) (20108)368 ROBUST IDENTIFICATION The value of n, that minimizes the above bound is the smallest integer n achieving ®(n) © € Proof. The first part follows from: We Ah = SAG) = vel + S100) < nes 20k) (10.109) For the second pat, define a(n) =ne + 370K) ‘Therefore a(t +1) ~ a(n) = € ~ O(n) is nondecreasing because © is nonin creasing. As a consequence, the minimum of a(n) is achieved at the smallest rn, satisfying a(n+ 1) > a(n) a ‘An intuitwe interpretation of the above result is the following. Since VA(K)| < (k) and & is nonincreasing, for values of k beyond the point for which (K) < ¢, there is no useful information on the system which can be obtained from the data It is clear from (10.109) that the value of n should depend not only on M but aiso on ¢ in order for the algorithm to converge. This is supported by the fact that there are no untuned idemtfication algorithms based on impulse response data ((160, 193, 194)) as proved before. There are many values of ‘n that may be selected for the algorithm to converge. In the sequel, three of them are presented. Lemma 10.18 The algorithms obtained by selecting n as either m(M,«) = min(M ,¢p ~ cy Ine) or na(M ,€) = min(M, co + c1€-"),1 € (0,1) and ¢a,¢; real constants, are tuned to € and convergent. A third, asymptotically globally op- imal algorithm tuned 10 (@,€) can be obtained using ns = min(M,n.), Me defined in Lemma 10.17. Proof, For n= ni, a(n) < (ca ~ eyln€ + Ie + Dig, PCR). The fist term vanishes as ¢ -+0, and ny 00 as M 0 and-e'~~0. Since € &), this implies that the second term also goes to ze¥0. Convergence for n =n) can bbe proved in a similar way. For m= ns, when M— 00 then 1+, and therefore the upper bound in (10.108) achieves the optimal value in (10.106) Also, for any given 8 > 0, there is always aq such that S22 y,, (Kk) < 8/2. TIMEDOMAINIDENTIICATION 369 and € > 0 such that (g + 1}e < 8/2. Because the lower bound in (10.106) is alobal, the last procedure is asymptotically globally optimal. a ‘As a special case of the above results, consider again the set of models ‘Yoolp, K), which includes the exponentially stable discrete time linear sys- tems considered in the Ha. identification setup ((146)). For p > 1 and K > 0, Hoo(p, K) € A(D), the disk algebra of functions analytic in the open unit disk and continuous on the unit circle. It is well known that the impulse response ‘of this class of systems is bounded above by the sequence {Xp-*, k > 0} This leads to the following corollary. Corollary 10.1 For the identified model #1", defined in (10.108), the identifi cation error can be bounded as follows: Kp Wh Hay < nes SO (10.110) This bound is minimized by selecting the smallest integer n > ma, with pst Inp (20111) Proof. Follows immediately from Lemma 10.17 by noting that %x(), K) CS when selecting (k) = Kp™* o Notes 1. Depending on the resulting value of n, it may be necessary to balance and truncate the identified model H",, to obtain a smaller-order model ‘This is a standard procedure, which may be simplified, avoiding the computation of Gramians, using the algorithm in [45, 127]. The new error bounds are given now in terms of the |: oF the Hankel norm, ‘Similarly although computationally more expensive, an optimal Hankel ‘model reduction may be performed, 2. For the case of systems that have all positive impulse response values, the &; bound coincides with the Ha. bound. Therefore the asymptoti: cally &; optimal algorithm is also asymptotically Hx, optimal. 3. When the value of n, is very high, for example in cases where M is high and ¢ is very small, care must be taken when balancing, in particular, when the first singular vector of the Hankel matrix is computed ([45, 127)), to avoid numerical errors that may increase the approximation 4. The identified system can be represented by a family of models, in terms of the nominal H, and the identification error bound, This family could be unnecessarily Conservative if the convergence of the algorithm370 ROBUST IDENTIFICATION is not monotonically decreasing with € -+0, In this case there might exist ¢ > « for which the identification error is smaller than the one considered and still “covers” the class of models S and noise N= Gel€) C fu(€,). All the algorithms presented here have monotonically decreasing convergence. Based on all Samples Next we present, a strongly optimal convergent algorithm, that is, locally optimal for any set of experimental data. This al- gorithm is based on the one in {160} Consider the vector of experimental data y of length M and di ho (k) = min yy + €, (8) (wo.12) fi (k) = max ly, — €, -(K)] (10.13) Note that Jy(k) is the lowest value of h(k) compatible with the a priori information, therefore consistent. Similarly, hy(K) is the upper bound on. ‘h(k) for each k, so that model h belongs to the consistency set. In addition, the consistency interval {hy (k), hu (k)] has length 2min [e, (k)} The algorithm is as follows: siy.0,0), = {AMM KEM oar The above algorithm is called central because it selects the model in the center of the consistency set; therefore it is also strongly optimal. This lat property can be proved as follows. Lemma 1019 The worst-case error for algorithm (10.114) is e[Ay, ®,€)] = > min fe, (4)] + JO, - > G9] (10.115) and tends 10 zero in the limit M +00 and € ~+0. The algorithms is strongly optimal Proof. The bound can be obtained as in Lemma 10.17, using the fact that AY. ®, €), — A(R] < mine, P(K)], & Mad since the model is in the center of the interval iz(k),/u(&)]. The above bound can be achieved with a suitable choice of system and measurement noise, It is clear that the consistency set is symmetric with respect to the ‘model Hiy = Aly, , €); therefore the algorithm is also central. TIMEDOMAIN IDENTIFICATION 371 Note that the first term of the error is equivalent to ner [6] (10.116) with n, defined in Lemma 10.17. This error converges to zero as (1/e,M) — oo. Using the above equation, it is easily seen that the total error can be rewritten as the global lower bound in equation (10.106). It follows that the algorithm is globally optimal. o In the special case where (k) = Kp-*, the bound (10.115) reduces to ((160)): we ie e[Aly,®,€)] = Zemin Ol SH (10.117) ‘The above algorithm is clearly tuned to all a prior’ information. Next, an- other interpolatory algorithm, which is tuned only to the model information, that is, ©, will be presented. This algorithm is given by: sign (4) min|ly«{,(0)] Os ks M— 0, kom AQ.) = { (10.118) Lemma 10.20. The worst-case error for algorithm (10.118) is: 14(9,0)) = 5 ming, 20(0] + ¥ ee] (20.119) which tends 19 zero inthe limit M oo and € ~+0. Proof. The proof is similar to the one above, considering the fact that [Ay ~ A(k)| < min fe, 20(6)] when [yx| < (If, however, [| 2 (k), the following holds: LA, ®)e = A(K)] < 26(4) Since the experiment and the noise have the same sign, in this ease LAly. ®), — AC} < WH) ~ yu] < € In both cases this leads to the desired bound. °372 ROBUST IDENTIFICATION For the particular case when ¢(k) = Kp-*, the bound (10.119) simplifies 10 ({166)} elAly,)] ~ Y minfe, 2049) y 10.5 FURTHER RESEARCH TOPICS Due to the fact that robust identification is a currently active research area, there are yet many theoretical and computational aspects that have not been developed fully. In this section, several items that have not been included in the main part of this chapter will be commented on and referenced. An excellent recent survey of the area of robust identification can be found in {199}, which includes an extensive list of references 10.5.1, Unstable Systems ‘The methods developed in this chapter apply to stable LTI systems. For slowly divergent systems, that is, with slow unstable poles, care must be taken when carrying out the experiment, especially when considering its duration ({127}). In general, neither time nor frequency response experiments can be Performed over open-loop unstable systems. Rather, they should be carried ut after 2 stabilizing controller has been applied to the plant. In that ease, the closed-loop input/output data should be used to identify the open-loop plant. The identification error should also take into account the fact that both. ‘model and plant are unstable. Thus the 7, norm should be replaced by a suitable distance function. In [196] a complete framework for the robust identification of strongly'® stabilizable plants is presented. The measure of identification error is com- puted in terms of the gap, graph, or chordal metrics. 10.5.2, Nonuniformly Spaced Experimental Points In the 2. identification two-stage procedures, it is assumed that the fre- quency experimental data points are equally spaced. Since this may not be the case in practical situations, the standard algorithm should be modified ac- cordingly. In (8] and (235] interpolation is used to perform identification with nonuniformly spaced experimental data points. In [127] a procedure that uses explicitly the unequally separated data points is presented, in the context of, model approximation, "A sronely sabiizable plant isthe one that can be nabilized hy a stable conte FURTHER RESEARCH TOPICS 373 10.5.3 Model Reduction In many of the algorithms presented in this chapter, the model order re- mains high. As a consequence, a further model reduction step needs to be performed and the model error bounds should be modified accordingly. This hhas been attempted in (142] for é identification and in (198, 253] for Hu. identification, 10.5.4 Continuous Time Plants ‘The robust identification framework presented so far considers both discrete and continuous time systems represented by H(z) in Section 10.3. Never- theless, in this last case, some extra considerations related to the frequency aliasing should be taken into account, Specifically, strictly proper stable LTT plants with an a priori upper bound on their roll-off frequencies should be considered. This problem has been treated in (9, 147) 10.5.5 Sample Complexity For application of the identification techniques to practical problems, it is important to know whether the experiments can be performed in a reasonable amount of time. Specifically, the following question may be asked. Which is the minimum time length of an experiment in order to arrive at a certain degree of accuracy, measured by the worst-case identification error? A general framework for this problem has been considered in [183] and specific results have been obtained in (163, 85, 243, 303]. These results are presented next Consider the FIR systems of length M, bounded experimental inputs [luli <1, and a (optimal over all algorithms and inputs) worst-case iden- tification error bounded by nf theca 20 as where € is the measurement noise uniform upper bound and the parame ter Q > 1 defines the level of uncertainty (Q = 1 corresponds to the optimal error bound 2e). For this case, the minimum length of the experiment dura- tion to achieve this bound is O@//2), with f() a strictly monotonically increasing function in (0,1), with f(0) = 0 and fim f(x) = 1 as x —» 1. Itean bbe shown that this isa tight bound, A consequence of this result is that the worst case clas of measurement noise (uniformly bounded by e) considered in time-domain experiments may be very conservative. Therefore, as suggested in [243], the class of measure~ ‘ment noise should be relaxed or noise should be considered in the average ‘On the other hand, the worst-case class of models does not need to be mod: ified. Tn the case of M., identification, the sample complexity is only quadratic in at ((199).374 ROBUST IDENTIFICATION 10.5.6 Mixed Time/Frequency Experiments ‘When applying robust identification techniques to either frequency- or time: domain experiments separately, it may happen that a model that efficiently fits the data in one domain may have poor performance in the other do- ‘main, For example, with only frequency experimental data points, a “good” frequency response fitting (small %,, error norm) may lead to a “poor” fit ting in the impulse response. Additionally, from an information theoretic viewpoint, more experiments produce a smaller consistency set of indistin- ‘guishable models, and as a consequence a smaller worst-case error, From a practical standpoint, in many cases robust identification algorithms are ap- plied to systems that may not be exactly LTI. Under these circumstances itis desirable to perform both time and frequency response experiments ((272)) to assess the validity of these assumptions. Thus, in these cases, an algorithm that combines both types of experimental data could take advantage of the additional data available Recent papers [77, 333] have proposed interpolatory algorithms that use data obtained from time-domain experiments and generate a nominal model together with an a, bound on the identification error. Therefore, although they combine time domain measurements with frequency domain error bounds, they do not use both types of experimental information simulta: neously, Alternatively, some recently propased algorithms explait hoth time and frequency domain experiments at the same time. In {273 necessary and suf- ficient conditions for the consistency of mixed time/ frequency experimental data for FIR systems are presented and an interpolatory identification algo- rithm is proposed, In [231] the above result was extended to IIR systems, based on a general interpolation theorem ([21}), which allows the simultan- cous interpolation of both the time and frequency experiments. The solution to this problem leads to a convex LMIls optimization, 10.5.7 Mixed Parametric/Nonparametric Models All ofthe results in this chapter address nonparametric identification of mod- els with a worst-case global bound. In many cases, part of the model has @ clear parametric structure. In these cases disregarding this information may lead to conservative results (see Example 10.4). This is usually the case in mechanical flexible structures, which have a well defined parametric model for the lower-frequency modes and an unknown higher-frequency behavior ‘that naturally leads to a nonparametric identification ((121)) In general, the parametric information does not appear explicitly in the 4 priori knowledge (K,p) usually considered. Therefore it is important to include it, if available, so that less conservative a priori estimates of (Kp) for the nonparametric portion can be derived. This is the case of systems ex- hibiting large peaks in the frequency response (for instance, a poorly damped FURTHER RESEARCH TOPICS 378 flexible structure), which must be “covered” by large values of K ({121)). This is a consequence of the fact that the usual a prion information characterizes only the smoothness and magnitude of the whole class of models but cannot distinguish among other properties, for example, low frequency model struc: ture. Instead, the a priori parametric knowledge provides more “structured” information, ‘There are several results that address this problem, for example, in [121. 175, 205, 232, 333]. In particular, the identification of affine parametric and nonparametric models using simultaneous time and frequency response ex- perimental data can be solved as a convex optimization problem ([232]). This includes cases of practical interest such as flexible mechanical structures, as well as models that can be described concisely in terms of a set of Laguerre ‘or Kautz functions of, more generally, any other basis of 2 ((148))APPLICATION EXAMPLES 11.1 SAC-C ATTITUDE CONTROL ANALYSIS. n Introduction The SAC-C satellite is a joint project between the Natioual Conunission of Space Activities (Argentina) and NASA (United States). It is essentially an Earth observation satellite that carries an Argentine CCD camera for Earth resources observation, a Global Positioning System (GPS) experiment from JPL (NASA), and a magnetic sensor from Denmark. The latter has a Star Imager and two Magnetometers mounted on an 8 meter boom, which will bbe deployed in orbit. The purpose of the boom is to magnetically isolate the ‘magnetometers from the spacecraft. The length of this boom makes it very flexible; therefore the model of the spacecraft cannot be considered as a rigid body (see Figure 11.1) The attitude objective is to point the camera toward the vertical with a precision of +1° around the roll and pitch axis; that is, the line of sight of the camera should be within a 1° square with respect to the vertical direction. Based on optical considerations, the jitter of the control should be on the order of 0.05%/s, Due to several practical limitations, it is important to have the simplest possible controller and the least amount of hardware, These limitations stem from the fact that the microprocessor onboard should not only implement the mission contro! mode but also several alternative safe-hold modes, as well as failure logics. Therefore a low-order controller is desirable. In addition, sensor and actuator hardware adds weight to the spacecraft, which in turn decreases the available payload. For this reason, only the minimum required rotations along the axes are sensed. Similarly only the minimum number 377378 APPUCATIONEXAMPLES Of torques are applied by the actuators. Thus only the roll and pitch axes are sensed by means of a scan wheel horizon sensor (SW) and the control torques are applied only in the pitch and yaw axes. These control torques act through magnetic torque coils (MTCs) and momentum wheels, the latter ‘built into the SW. In the design, the torque bounds of both actuators should bbe considered, due to limitations in the maximum magnetic moment of the MTC and the maximum size and acceleration of the momentum wheels ‘The desired attitude of the satellite in Mission Mode (MM) is the fol- lowing: roll along the velocity vector, yaw opposite to the vertical, and pitch perpendicular to the orbital plane (See Figure 11.1). Our goal is to design ‘@ minimal order controller to achieve this nominal attitude, This controller feeds back the roll and pitch angles and rates as well as the momentum wheel velocities, using static LOR feedback. The only dynamics involved in the controller arise from a yaw angle and rate observer (the only axis without sensor). Therefore the complete controller has only two states, is applied to only two of the three axes, and senses two out of three axes: as simple as it can be. The possibility of obtaining such a low-order controller and not hhaving to sense or actuate over all axes is due to the fact that the spacecraft and the orbital dynamics are coupled. ‘The attitude for the Earth Pointing Safe Hold Mode (EPSHM) is the same as for MM, except that the momentum wheels are kept at nominal constant speed. Here the control is applied using only the Earth magnetic held, through the MICs, The controller design is carried-out using a very simple procedure, that takes into consideration the limitations in the controller order, sensors, and actuators. This design is validated in Section 11.1.4 through an analysis involv: ing all the performance objectives as well as model uncertainties. Using the structured singular value, we assess the robustness of the controller against the unknown high-order dynamics of the boom, considered as global uncer- tainty. Robust performance is guaranteed using 2 for both modes: MM and EPSHM. Several time simulations of the complete system are presented in- volving random measurement noise and torque perturbations. Further details of this analysis are reported in [274] 11.1.2. Linear Model ‘The general nonlinear dynamic equation of the system is: Tut Teyt Ty + Ty = [pW 4 W x pW W xh tlle] SAC ATITUDE CONTROLANALYS'S 379. ‘ORBIT Xa FORWARD VELOCITY << A EARTH Figure 11.1, SAC-C Eorth observation satelite Here W is the angular rate of the spacecraft with respect to inertial space, represented as the rotation of a frame attached to the spacecraft with respect to the inertial reference frame, The inertia of the rigid part of the spacecraft is Fr and h represents the wheels momentum in the spacecraft frame. This frame is the usual roll (x), pitch (y), and yaw (z) axes mentioned before (see Figure 11.1), The external torques consist of the perturbation (7,), gravitational (T;), and magneticcontrol (Tx), that is, T = Ty + Ty + Tm The gravitational torque is due to the effect of the rigid body model of the boom and the magnetic control is produced by the MTCs. The perturbation torques are due to the aerodynamic drag (almost constant in pitch and yaw along the orbit) and the380 APPLICATION EXAMPLES magnetic and gravitational residual torques, both periodic and negligible as compared with the former: The wheels and magnetic torque col torques are used for control purposes the first one depends on the velocity change of the momentum exchange wheels and therefore does not change the total momentum of the spacecraft, the latter is an external torque that does change the total momentum. Note that these torques are applied only in the pitch and yaw axes, The reason is that a control torque inthe roll axis is unnecessary due to the orbital coupling between roll and yaw, as will be seen next. This reduces the number of actuators, therefore simplifying the design and minimizing costs The external perturbation torques are bounded but otherwise unknown Next, linearization is computed around the equilibrium conditions, cor responding to the nominal attitude of the satelite: the roll and yaw axes colinear with the velocity and local vertical vectors, respectively, and pitch perpendicular to the orbital plane, which completes the spacecraft frame (boom opposite to the local vertical). The roll, pitch, and yaw angles are de- noted as Op, Up, and Oy, respectively. The incremental values of the wheels momenta are 8h, in pitch and dh, in the yaw direction. For simplicity we can assume that Jy = diag [/, ly. Ie] Since the roll and yaw dynamics are coupled but independent from the pitch, we define the system state as xT = [x5 x7] ay ty = [Oe Oe Oy By dhe]", —xp=[ 6p Op Bhy|” ‘Adding the equations for the wheels torques Tey = —Siy and Tue = we obtain the state equations for both the {roll-yaw] and [pitch] dynamics: = Apt + Bry lies | +B.Tpe (12) pean «Bows [72 | + BTyy (13) { (0) = Ax(0) + Bult) + B,T, y= [22 Te] + aa xl + Dute) Oona Tent where x(1) has been defined in (11.1) and where: 1, wy 2 | SAC-C ATITUDE CONTROLANALYSS 381 Note that the system output consists only of the measured states, that is roll and pitch, Mattices Ay, Bry, Ap, Bpisn, By. and B, have been defined in [274 11.1.3: Design Constraints To simplify the controller as much as possible, a state feedback has been designed. The only controller dynamics will be the ones from the yaw angle and rate observer, due to the fact that only roll and pitch sensors are available (scan wheels) ‘A static LOR controller has been designed, taking into account the desired bounds for the different states and actuators when computing the weights to solve the minimization problem. These constraints are as follows Tue] [0012sing N-m d 2610- Nem Tr | | otzacose Nm aus) te) | foi Nm ‘This is due to the fact that the torque bound in each wheel is 12 mNm and that in pitch both add their torques projected through cosa (a < 10°). In yaw, on the other hand, they subtract both torques projected by sina. The bbound on the magnetic torque coils is computed as the average torque along the orbit. Each MTC has 15 Am? and the average magnetic field along the orbit is (V3/2) 0.4» 10-' Tesla, Therefore the actuator weight in the LOR problem is a diagonal matrix that has the inverse of these hounds as each element (Bryson's Rule, see Chapter 5). “The state weighting matrix is also diagonal with the inverses of the maxi- ‘mum allowable bound in each of the states. These are given by: bn oss Om r by o/s by r ah. |<] 0.06N-m-s co bp O1°/s op r ah] [006N-m-s The bounds in both 5h; and 8h, correspond to @ maximum deviation from the nominal wheel speed of 400 rpm, For a total momentum of 3.2 Nims. (2000 rpm, J = 0.0077) along the pitch axis, these bounds are 0.11 and 0.64 Noms, respectively. The observer involves only the [roll-yaw] dynamics, due to the fact that pitch is uncoupled, Furthermore, a reduced order observer will be used so382 APPLICATION EXAMPLES that the controller has the minimum possible order. The measured, estimated, ‘and input variables to this observer are, respectively, i a = af by 8 | Tw: melon 8 [#) 2 [Be] [* le rm Closed-Loop System The observer spstem is given by Ao | Bo Ao | Bor Bor fo] | au) Co | Do Co | Dor x2 with output fy and input [x2 7", In this case the observer gain can be computed analytically ([274). ‘The control signals, LOR gain matrix, and state feedback are oR Ts i. 4 | Tme| [Fe Fy Fie 03] | gy, a le Gir Oar Ons Fe |B] O18) h o ™ or a, where the feedback matrix has been partitioned according to the roll, yaw, {5h,, and pitch variables, From the above equation and the observer dynamics we compute the controller as follows: (Ao ~ Bo2FyCo)z + {Bor ~ Bor (Fy Dor + (Fe Fre )} %m Du O FrCo2~ (FrDoi + [Fe AD™[7 | =] or equivalently, a9) SAC-C ATITUDECONIROL ANALYSIS. 383, Note that the controller has the minumum number of states, given the avail able sensors (roll and pitch) and the minimum number of actuator outputs (yaw and pitch) given the the coupled dynamics. Using the above definitions for the controller, input, state, and output variables, the closed-loop system is given by: (A+ BD,C)x + BC yz + BpT» + BDyn 2 = Bre Az + Bin Ei Thus the state-space realization of the generalized plant G,(3) with states [27 27], disturbance inputs p 2 [ TF n? ]’, and e the output to be mini- mized is A+BD,C BC, | By BD, BC AL | Oe Be |_| Ac| & Teen Opn2 O54 Ones Ce | De DCG | Oe Dy 11.1.4 Robustness Anatysis The closed-loop dynamics computed in the last section correspond to the nominal system, which takes into account only the rigid body modes of both satellite and boom. Nevertheless, the flexible modes of the boom, should be taken into account explicitly in the robustness analysis, ‘These modes can be modeled as uncertainty in the high-order dynamics, using experimental data provided by the boom manufacturer ((295)), to com- pute the frequency distribution of the uncertainty. The boom dynamics is considered in this setup as follows: Hs = Grontsdrar v2 [ TH], ys 2[M] ‘The reason i that the diving input, which iis the Rexble dynamics Osilations the angla rotation of the spacecraft in rll and ich (he torsion around the yaw axis eligible) One other andthe pemtubation ofthe boom onthe stele sructre represented by the (intemal) trgues Applied nthe roland pth anc Gusto these esllation. The experiment in [295 determined lower requency yeamics ot! Hz. Therefore a convenient Weigh that “cover” the uncertain high fequency Ojai of the boom and assumes a 20% modeling eror at lower frequencies Seon Wa(s) = 04752 1,,5384 APPLICATION EXAMPLES The performance specifications call for minimizing the states and bounding the control torques of MWs and MTCs. Therefore a suitable performance ‘weight is a constant diagonal matrix with the inverse of the upper bounds in (115) and (11.6) as its elements. To perform the robustness analysis ofthe satelite the nominal and uncer. tain portions, and weights for robustness and performance are connected in the LFT form shown in Figure 11.2 ‘The closed-loop nominal system computed in Section 11.1.3 corresponds to the lower portion of the general system R(s) = Fe(P, K), connecting the inputs p to the outputs e, that is, Gy(s) = Ron(s). The total system, with- ‘out weights, connects the inputs (us,p) to the outputs (¥s,e) and has the following state-space realization: Ac Ris) = o10000000 000000100 G “feta vm To perform the robustness analysis the output weights must be added to the system as indicated in Figure 11.2. Furthermore, to normalize the anal- ysis, we scale the inputs in vector p so that the magnitude of the worst-case perturbations has unity norm. To this end we use a constant input weighting ‘matrix having the maximum perturbation torques and measurement noise as its elements. These perturbation torques are due mainly to the atmospheric mae va —Awnj2—1 Pe) ¢ fim} d y K(s) Figure 11.2. Unear toctional tanslormation setup for robustness analysis. SAC ATITUDECONTROL ANALYSIS 385, drag effect in yaw and pitch with a maximum of 10-° Nn, and tothe residual ‘magnetic torques in roll with a maximum of 10-® Nem. The angular measure iment error due to the scan wheels is 0.3°in roll and pitch, the only measured axes (recall yaw is estimated with an observer). The angular rate measure ment error in the same axes stems from the error in the LSB of the word that spans the total angle, this is 0.01". The rate is computed as the difference between two correlative angle samples, divided by the sampling time 7. For this reason, the rate “measurement” noise weight is similar to a notch fier centered al w = 2/T and scaled by 0.01*/s ols We (s+ 2/5)(s/207 +1) sels) au Finally, the rate measurement errors of the wheel velocity, which con- tribute to the error in their angular momentum in yaw (8h) and pitch (5h,), amount to a maximum of 2% of hyo. ‘The necessary and sufficient conditions to guarantee robust (internal) sta- bility, nominal performance, and robust performance are RS > [Ws(s)Ru (Wye <1 (a2) NP = [Wp(s)Ra2(5)Wir(5)I, SL Gusy RP e=> sup Ha [Wo(s)R(s)W)(s)] <1 ais) where Wo(s) = diag|Wa(s) Wp(s)], Wr(s) = diag[Wi(s) Wra(s)],and the ‘uncertainty structure A contains two full blocks, one for the boom high-order ‘dynamics, the second for the performance objectives. 5 Simulations ‘The following are simulated outputs of the main variables to be minimized, in Mission Mode (MM). Figures 11.3 to 11.6 show the angles, angular rates, and momentum wheel and magnetic control torques, all below the desired bounds. Figure 11.7 shows the robust stability and nominal and robust per- formance conditions, just below the value I ‘A different controller has been designed with the same methodology to control the spacecraft only with the MTCs, called Earth Pointing Safe-Hold Mode (EPSHM). In this case the wheels are kept at a constant speed and provide the necessary angular mmomentum to the system. In particular, the ‘weight on the MW control torques for the design is large enough so that the wheel torques are below 6 x 10° Nm, that is, almost constant speed, The robust analysis curves based on the infinity norm and the structured singular values are presented in Figure 11.8.386 ogress “3900828003605 aa09 000 000 ine) Figure 11.3. Rotation angus rors MM) otc) Yaw) Pa ete cons oo 0.009 t 009 oo oo A ! 0 360 20003004008 S003 tooo via) APPLICATION EXAMPLES Fo (i), Yam (Pach) ii Torque (nN) SAC ATITUDE CONTROL ANALYSIS. 387 i ores in Yaw (0) & Pa (-) sine 6) Figure 11.5. Momeniurn wheel contol foraues (MM), MIC toques in Yaw so) & Pe) Figure 11.4. Rotation ate erors (MM). sine (6) Figure 11.6. Magnetic contol lorques (MM).388 APPUCATIONEXAMPLES jj Nominal ana Robust Performance sis ah Robust Stabilty ESO RSEEE HEEL Av w (adie) 4 w (cadis) Figure 11.8, Robust stabilty and nominal and robust performance conaiions (EPSHM). CONTROLLER DESIGN FOR AD.O PLANT 389 11.2 CONTROLLER DESIGN FOR A D,O PLANT In this section we illustrate by step, the robust control methodology through an application (0 a D;O plant. The mathematical model of the plant is a linearized and simplified version of a nonlinear system with a delay. This particular application is a standard problem in the process control area. The Purpose of this example is to assess the incidence of model and disturbance luncertainty in this typical application. The model uncertainty arises from the time delay and the gains and poles of the transfer function, which depend on the particular linearization point. The usual control structure to handle systems with time delays is the Smith predictor ([289)). This classical method in process control has the advantage that, when the delay is known exactly, the controller designed to achieve (nominal) stability needs to take into account only the rational (nondelayed) part of the plant. In this example, based on the Smith predictor, we perform the robust analysis and design of a controller for this particular uncertain system, which exhibits all the relevant issues of process control problems. The following example is based on [268), and the interpretation of Smith predictors with uncertainty is described in [211] 11.2.1, Model of the Plant ‘The plant to be considered is an experimental process that obtains D;O by isotopic exchange between H;O and HyS (see (128, 206)). The plant has already been constructed but is not yet operational. ‘A general diagram of the process can be seen in Figure 11.9. This diagram is a simplification of the process flow sheet (there are more cold and hot towers in the actual plant) but retains the main issues of interest, The input water flow and liquid flows are marked with full lines and the gas flows with dashed ones. Two compressors have been included, as well as three heat exchangers that control the temperature in the process. The greatest D;O concentration in each stage is obtained between the hot and cold towers. A fraction of the partially enriched water of the first stage is sent to the second stage for further enrichment, The product is extracted from the second stage after the H3S stripper eliminates the gas. The isotopic exchange process that takes place in the cold and hot columns, respectively, is the following H,0+HDS “© HDO+H,S HDO+H,S C H,0+HDS The controlled variable is the ratio L/G, with L and G representing the liquid and gas operational flows, respectively. The value of G remains almost constant duc to the fact that it circulates on a closed circuit, as seen in the same figure; therefore the control is performed by changing the value of L390 APPLICATION EXAMPLES sho) Figure 11.9. Schematic dlagram of the D,0 plant. ‘A general nonlinear model of the whole process has been implemented ({206, 228}) for simulation purposes. This general model has been obtained from physical considerations and is too complicated for both analysis and control design. Therefore the controller analysis and synthesis are based on a linear model that takes differential H,O flow as an input and differential D;O ‘outlet concentration as the output. The linearization has been made at 0.01%, 0.3%, and 1% disturbances in the input water flow, under the restriction of 15% outlet concentration of D,O. The structure of the linear model remains the same for all three linearization points, but the parameters (gains and poles) differ from one point to another. Assuming continuity between the linearization points (water flow disturbance) and the model parameters, the linear process may be described by a set of models with the same linear structure but with parameters lying in real bounded intervals, ‘The delay is a consequence of the measurement procedure for the out- put variable. It consists of the collection of samples at each column, which takes around 15 minutes per column, with a transportation time between columns of another 15 minutes. All these samples are measured with a mass spectrometer, to obtain, after approximately a 4 hour test, the DO concen: LE CONTROLERDESIGNFOR ADO PLANT 391 tration. Depending on particular conditions, the elapsed time between the sample extraction and the final output value may differ significantly. There. fore for analysis purposes, we consider a nominal delay of 5 hours, obtained as an average value, with an uncertainty of £1 hour, In this linear model we also consider additive disturbances in the output. ‘These disturbances include several liquid and gas flows, which go through part of the process and which, for operational purposes, are bounded. If these bounds are exceeded, the required level of D;O extraction will not be ‘met (or we will have no extraction at all). Due to the fact that these flows circulate through most of the columns, we may assume that the dynamics of the disturbances is very slow (see Figure 11.9). For the above reasons, we ‘model the additive disturbances at the output as a set of bounded signals with unknown and slow dynamics, ‘The set of linear models representing the process is the following _— «fe, Gls) = 0 |e ot ae tye | Teme) Kee Miu 4 71 €lypwjh, vj >0, 4 ais) T=%+8, [8l<1, 8ER, where the nominal delay is x» = 5 hours. The uncertainty intervals for poles and gains are obtained from their values at the three linearization points mentioned before ([228)). The disturbances are represented as a weighted and bounded set, that is, {We(s)d{s)|ldl2 < 1). Here W,(s) takes care ofthe fact that the distur bances have a higher energy content in the lower frequencies (slow dynam ies). For the controller design step, a rational approximation of the nominal time delay needs to be made. Many different methods exiss, but @ standard simple procedure is the use of Padé approximants, Nevertheless, we will verify the robust performance condition with the exact values of the delay, without approximation errors, 11.2.2 Robustness Analysis In this section we obtain conditions on the controller such that stability, performance, and robustness are achieved. Due to the fact that the system hhas a delay, we work with the Smith predictor control structure, but including the uncertainty in the delay, poles, and gains. The classical Smith predictor structure ([28]) is presented in Figure 11.10, The stable plant G(s) consists of a rational term Gy(s,p) with a vector392 APPLICATION EXAMPLES of uncertain parameters p. p; € j.b,) (in this case the poles and gains in (11.15)) and an uncertain delay 7 = 7, + 8,[6] < 1. Define the nominal model 88 Gals) = € "*Gyy(s,po) which wil be used for the design. Next, we analyze step by step nominal and robust stability and performance Nominal Stability The Smith predictor operates by subtracting from the feedback signal the difference between the nominal model and the core. sponding minimal phase part (Figure 1110) This is equivalent tothe incl sion ofa prediction in the sytem since the feedback signal f() is exaely the output ata future time (1+), This procedure i elective only when theres Complete knowledge of the system (he parameters ake their nominal values +=, and p= pp) otherwise we will not be able to havea corect prediction oy From the same figure, we observe that, the nominal transfer function be- tween u and the feedback f, equals Gy, © Gy(s,p.). Therefore in terms of nominal stability, only a controller K(s) that stabilizes the nominal minimum phase part Gwo(s), regardless ofthe time delay 1», needs to be designed This isthe main advantage ofthe classical Smith predictor, but this esti to longer valid when considering performance of robustness issues, a5 will be seen nex Nominal Performance Based on the results in previous :chapters, nom- inal performance depends on the transfer function between the disturbance and the output error signal. Figure 11.11 yields the following transfer function in the nominal case, that is, G(s) = G(s) vals) = [10% Tu(s)] Walsdd(s) (1.16) AO Figure 11.10. Smith precictr. CONTROLER DESIGNFORAD.O PLANT 393 where we have defined the complementary sensitivity as Tals) © Gass) K(s) [1 + Gael )K(P The necessary and suffiient condition for nominal performance is Nominal performance => jlya(slz $1, ¥lld(s)I <1 si qt) Tu (s}] Wal) This leads to the following 21. control problem: Ip- sary hll “°° TaCO WAC oy An important point is that the nominal performance condition includes the nominal delay 7,.. As we mentioned before, the classical Smith predictor does ‘ot simplify the analysis step for performance, as it did with nominal stability. A controller that solves equation (11.18) can be obtained by a standard Ha. ‘optimal control algorithm, with a previous approximation of the time delay. The optimal solution without approximating the delay can be obtained from [116], but this approach cannot solve the robust performance prahlem Robust Stability In this section we analize the closed-loop stability for all Possible delays between 4 and 6 hours and all possible parameters p [poles and gains in (11.15)] of the minimum phase part of the plant. Wale)d A lyautye Ks) [ber ou(s,p) Figure 11.11. Smim predictor wit: uncertainty.394 APPLICATIONEXAMPLES A practical! way to analyze models with parametric uncertainty is to “cover” them with dynamic uncertainty. In this case we have used a mult plicative dynamic description for the set of plants, as follows: ME {[1+A.Wa(s)Gmols), eC y (Sul <1} (1119) In this application, there are two different sources of parameter uncer. tainty, which in principle could be covered separately. This will produce a ‘smaller” set of plants than in the case where the coverage is performed over both parametric uncertainties together. As a consequence, the analysis results for the “smaller” set of models will always be less conservative than in the case where both parametric uncertainties are globally covered, For the delay, we transform the dashed block of Figure 11.1] into a nom: inal system with global multiplicative uncertainty, with p = py. To this end, consider the family of models GF Gs) — Gals) + Gols) (11.20) = eles -re out, BER, jl mate =| at2t) Figure 11.12 shows the above condition in a magnitude Bode plot. A rational " ama hoc analysis considering directly parametric uncertain could be made, bu iis ot the purpose of thin example. The resus on analysis of parametnc uncertainty srutures ae ee? Snslyieal methods that solve particular eases (and donot i hs spication) oF algorithmic ‘methods that solve general cases but ae computationally involved, infact « NP-complete prob Jem (jf), 24). There i 9 yet a sates method that supports a peneral parametric anal ‘Am cxcllent survey ena he found [35] CONTROLERDESIGNFORAD,O PANT 395 7 10 10 10% 10 107 10% wads) ‘Figure 11.12. Dynamic weight W(s} covering delay uncertainty. approximation Ws(s) for the weight can be computed from the same figure, Note that the set of models that covers the delay uncertainty is independent of the nominal delay 7, Following a similar procedure, a set of models with dynamic multiplicative uncertainty, which covers the parametric uncertainty in gains and poles can be obtained. For briefness it will not be repeated here, but the corresponding, ‘weighting function may be seen in Figure 11.13. ‘The necessary and sufficient condition to guarantee the stability of all ‘members of a family of models with multiplicative uncertainty” is IMs) Tus) <1 (122) From the above equation, we observe that the critical point in terms of robust stability i the frequency w. for which |W,(ja.)| > 1. From (11.22) itis clear that, for w > a, the values of |Tiy(jw)| should be less than one to achieve the robust stability condition. In some sense this frequency defines an upper bound for the closed-loop bandwidth, the latter interpreted in terms of Ty (5) 2 For the family M the condition is necessary and sufficient when replacing W with Wy, but since @'C AM forthe actual set of models 8 only sufeient396 APPLICATIONEXAMPLES Weights for uncertain delay & poles 10! Total uncertainty weight +04 : a Uncertain pote-gain weight 107 10 (ee ay w(radis) Figure 11.13. Dynamic weights covering pole gain and delay uncertainties, Tota Uncertainty wight is Wo(5) ‘The above condition could be useful in determining the relative impor- tance of both types of uncertainties for the robust stability of the plant. By comparing the weights in Figure 11.13 we observe that, for the frequency range of interest, the curve due to the pole and gain uncertainty remains be- low one. Therefore the limiting factor in the robust stability condition is the delay uncertainty, Nevertheless, in the lower-frequency range, the uncertainty in poles and gains is larger than the one due to the delay. Therefore, from a practical point of view, we should consider the total uncertainty weight W.(s), which covers both types of parametric uncertainty (see Figure 11.13). Finally, it should be noted that the Smith predictor structure allows this analysis to be independent of the nominal delay 7. Only the uncertainty is taken into account, which appears indirectly in the weights Wa(s) (see equation (11.21)) and W, (3) Robust Performance For robust performance we need to achieve the performance objective for all models in the set {[1 + &,Wu(s)] Ga(S), [Mul < 1). From Figure 11.11 we obtain the following transfer function between d and the output signal yals) = [10 Ty (s)] [1+ AaWals)Tu(s))! Walsddts) (11.23) ‘CONTROLLER DESIGN FOR AD, PLANT 397 where we have replaced G(s) = Gy(s) + 4uWal)Grols)- A necessary and sufficient condition for robust performance can be obtained from the above equation in the following way: | et Tue) + WaT)" Wa] SA, VS = Je, Nal <1 = |(l- Ty) Wal < inf 10+ WaT wl V5 = Jo ) Wal < inf 10 DI J ee |(1— Ty) Wel +1WuTul <1, Ys = jo (1.24) where we have selected a worst-case || < 1 at each frequency. Equation (11.24) is a combination of both the robust stability and nom- inal performance conditions of previous sections. The main difference with a standard mixed sensitivity problem is that both terms of this equation are applied to different plants—one to Gyzo(s) and the other to a combination Of Gyo(s) and Go(s). To recast it as a standard mixed sensitivity problem ‘we apply the following transformation Le Ty(s) = (Ie) +S u(s) (1125) where Sys) © 1+ Gols) )}" isthe sensitivity funtion, The goal is ow {o design a controler that solves: ani ggg Sm COWA)| + Wels) Ta] (1126) with Wy =e" Wa(s), The factor y not only provides relative weight be- tween the nominal performance and robust stability conditions, Dut also compensates forthe extra term (1-~€""*) in equation (11.28). At the de sign stage in the next section, the value of 7 is computed so tht the robust performance condition (11.24) is achieved inthe least conservative way. 11.2.3 Controller Design ‘The mixed sensitivity problem of equation (11.26) can be solved in two ways. ‘Hoe optimal control theory can be used to design a controller that solves: Wa(s)Ta(s) ll Wals)Suls) a that is, it minimizes the norm of the transfer matrix from the input v2 10 the outputs (z1,22) in Figure 11.14, Recall from Section 48 that the optimal value of (11.27) is at most twice the value of (11.26).398 APPLICATION EXAMPLES wile as wy cd K Lew. Figure 11.14. Mixed sonstivy probier, Alternatively synthesis can be used to design a controller that minimizes directly the value in (11.26). This is equivalent to cosy xe) —Wals)Tu(s)— Wals)Tu(s) {[ yWals)Su(s) yWe(s)Sw(s) } a the argument being the transfer function matrix between inputs (11,3) and outputs (21,22) of Figure 11.14. ‘The sensitivity Sy (s) and its complement Ty(s) are both functions of the ‘nominal minimum phase part of the plant Gy(s), which has the following state-space description: she 0 0 o 0.395 x 10° Oe nn 0 08 | ome ings) ORT gO nace 0 [*T gar xcao-e | 0 0 oo 0.166 x 10° we =[1 11 axe ‘The uncertainty weight, obtained from Figure 11.13, and the performance ‘weight from design considerations are, respectively, ) = 191 86% 1088 41 Wee 3x 10541 (129) ‘When equation (11.26) is solved using M4. optimal control, the resulting. ‘optimal solution does not necessarily achieve the optimal robust performance in (11.24), where we have a combination of minimum and nonminimum phase models. Therefore we have iterated over until the design achieves the robust performance condition (11.24) in the least conservative way. In fact, CONTROLER DESIGN FOR AD.O PLANT 399 for this case, the best performance has been obtained for the first iteration of a w-synthesis procedure, that is, the 74. optimal design.’ and a value of Youn = 1.58. For the design, we used a third order Padé approximation of the nominal, delay. Nevertheless, 10 verify robust performance in (11.24). the exact val ues of the delay have been taken into account. The plots of the robust and nominal performance and robust stability can be seen in Figure 11.15. In the same figure we can see the structured singular value 4 of the closed-loop system. We remark that, since the design has been made using the minimum phase part of the system Gyo(s), the value of is below the actual robust performance condition (11.24), which includes the delay. For the value Yaa the latter is very close to unity at w, = 2-10~ rad/s. In Figure 11.16 we can see the response of the system to an 11 hour sinusoidal disturbance of unit magnitude for 4,5. and 6 hour time delays. The frequency of the disturbance has been selected as a worst-case value according to the peak on robust performance in Figure 11.15 (w, = 2-10 rad/s, worst-case disturbance) Due to the fact that the present design optimizes condition (11.26), a 09 08| 07| o6| os| o4 03 02 oat ogee Te wiradts) Figure 11.18, Robust and nominal performance, robust stabilly, and y (forthe design sructute) “yma sandard problem, where the semsitvity and ts complement correspond tthe same mode! the performance peadoce hy synthesis design should have ben Petter than or east equal teste 0 optimal one400 APPLICATION EXAMPLES (Oe i0 eco eeeon] 40 60 6070 80 time (hours) Figure 11.16. Time response for worst.case disturbance (tobus disturbance rejection), possible improvement could be obtained by the use of methods that con- sider directly parametric uncertainty and the fact that the system is infinite dimensional. This is beyond the scope of this example, only intended to illus trate a complete step-by-step robust control analysis and design procedure ‘on a typical process control application. 11.3 X-29 PARAMETRIC ANALYSIS In this section we perform a robust stability analysis of the Jateral-directional controlled dynamics of NASAYs X-29 aircraft. This is a true multivariable control problem due to the coupling of the roll and yaw axes. The rigid body ‘model of the aircraft with a controller in place has been taken into considera- tion and the analysis consists of computing the multivariable stability margin (or equivalently j.) for the uncertainty in the aerodynamic coefficients. Fur- ther details of this example can be found in (265). 11.3.1 Linear Model First we' introduce the notation used throughout this section ({189]) in Table 11.1, corresponding to Figure 11.17. An illustration of the aircraft is shown in Figure 11.18. X29 PARAMEIRIC ANALYSIS 401 Table 11.1. Aiteraft parameters = Pitch rate ® + Lateral ‘acceleration () + Roll rate (p) + Longitudinal acceleration (a,) Figure 11.17, Lateral acceleration (@) Reference span (0) Roll and yaw moment coefficients Lateral force coefficient Gravity acceleration (322 fy/¢) Roll, piteh, yaw axes moments of inertia (slug) Roll and yaw axis cross inertia (slug) Altitude (ft) Mass (slugs) Roll, pitch, yaw rates (rad/s) Dynamic pressure (Ib/t) Reference area (ft) Sampling period (0.025 5) Velocity (ft/s) ‘Angle of attack (rad) Sideslip angle (rad) Aileron and rudder deflection (rad) Pitch and roll attitude (rad) Rudder Elevator * Normal ‘accelerator (@n) 4 ED + Yaw ae (0) Parameters of e generic aircraft.402 APPLICATIONEXAMPLES X29 PARAMETRIC ANALYSIS 403. 4 = pegtandsind +rtand cose (11.33) ee dy = 4 y Se (11.34) eam with the aerodynamic coefficients given by Cy = CyB +B Cr, 9, a3 : _ pb eth nse GB + C84 C1,8 + 6,20 + 2? 1136) oe eb E Cu = CyB Cag Ba+ Cn Bs CBP 4 Cy (37) The subscripts fi, ,, and 8, indicate the variables with respect to which the derivatives Cy,C), and C, are taken; for example, Cy, represents the deriva. tive of the lateral force with respect to the sideslip angle. These derivatives are computed by parameter identification procedures applied to the experi mental data. In addition, error bounds on these parameters are obtained as well ‘The linearized state-space equations, considering the input from the dif ferential aileron and flap w= (8, J” and the output y= |p r ¢ ay]? a= EArt el Bu (11.38) y = Cr+Du 2139) where the state is x=[ pr 6]!. Matrices (A,B) explicitly contain the uncertain aerodynamic coeficiems Ci, Cys Ch Cras Oy and Ca Gn, Sina cosa (4/¥)e0s0 Dace ae ane 20 7 an40) nT Figure 1.18, NASAXZ ocr Ini nr, B ly Cry M41. a GC, Cay, : y The lateral-iectional linear mode of the atralt (189) is presented next oD 4 ‘ additional matrices are B= Bey epsina—reosa+ £singeose (11.30) oe 7 v 10 0 0 Ply — Hes = GSbC) + rqlly ~ Ie) + pals 131) _|o Qo -he 0 fl a O -he hb 0 11.42) He bs ~ §80Ce+ path ~h) rah 0132) 00 04406 APPUICATIONEXAMPLES o 1 0 0 0) Or 8 i, to c o 0 014 (143) Cry ~Za/8 Xo le 0 o 0 0 0 lol 0 14a) Cr. Cr, Here we have defined mn = GS/mV, 4 =4S/mg, gy = 45d (1145) 4 = GSP LV, n= (gC, rane) (1146) dy = aC, +4(ly— 1), da =a, #g(le—4) (11.87) dx = = (Gn, ~ Glas), dae tan 6 (a148) ‘This model has four open-loop poles: the dutch roll, a complex conjugate pair that applies to the roll and yaw dynamics; the roll mode, a real stable pole; and the spiral mode, a low frequency unstable pole. ‘The controller structure for all flight conditions is the following: (5) = Ka(s) Kyls) (i149) where Ke(s) is: [ore Mins «Mua /2+ Xan] one (1150) and the elements of Ky(s) are given by (Ky) = Pils) Sp(s) [Ke —a-Blend-L(s)-K3] (1151) (Kyra = Ks S365) -Fo(5) Ps) L(s) (11.52) (Ks)is = —Ks-Po(s)-L(s)-g- Blend/V (113) (Kyu = Ke Fels) Pus) (assy (Kyla: = Puls) Sp(5) -[Kig ~ a Blend - L(s)] (iss) (Kolar = Kir Sis) Fe(3)- Pals) L(s) (1156) (Ky)zs = —Kir- Pals) L(3) -g- Blend/V (asa) (Ky)nu = Kun Fala) Pals) (1138) where Blend is a parameter. The elements of both matrices are shown in Table 11.2 X29 PARAMETRIC ANALYSIS 405 Table 11.2. Mattices Ko(s) ane Ko(3) Aye) Fourth-order differential ap actuator model AiG) Fourth-order rudder actuator model Zs) Padé approximation due to Sample & Hold delay Pus) Roll, yaw, and lateral acceleration prefiters Pus) Roll attitude prefiter 5,18) Second-order Roll rate sensor model Sis) Second-order yaw rate sensor model FAs) Second-order yaw notch fter Fis) Second-order lateral acceleration notch filter Lis) ‘Analog equivalence of digital filter L(z) 11.3.2 Results ‘The robust stability analysis was performed over two different linearization points. The first one corresponds to the nominal design flight conditions, the second one to a critical condition at sea level and with large angle of attack. These conditions are shown in Table 11.3 ‘The mathematical structure of the controller and nominal model of the plant remain unchanged, although the parameters have different values ac- cording to the fight condition. In the controller these parameters are Kz to Kis, Xepav Xin, Xzpas and Blend; in the nominal plant model the parameters are the aerodynamic coefficient derivatives. In this last case not only do the ‘nominal values of the parameters change but also their uncertainty bounds. Before a branch-and-bound analysis algorithm is applied to this example ((266, 281]), we proceed to compute the structure of the characteristic poly nomial f (5,4) as a function of the uncertain parameters mentioned before: this is G(sa) = D+C [se 40a] 8) 1159 f(s,A) = det [I + K(s)G(s,4)] (11.60) {eble 1.3. Unwotionconatons _ No Nea Mach No. 09 04 was som seni ° oe 0 ¢ sss als : Bat a A os 2a | ssn ar406 APPLICATIONEXAMPLES A = diag|5, 54) 161) with 6 the bounded uncertainty of parameter p, around its nominal value Pov hati, p= Pa +5, [8 < 8, A'symbolic manipulation solver has been used to determine the mathematical structure of this equation, in particular the exponents of the uncertain parameters, to determine which are repeated ‘ones. The mapping theorem can be applied to the parameters that appear linearly. On the other hand the ones raised to powers larger than one appear in a polynomial fashion in the characteristic equation and therefore should be restricted by the equality contrain's mentioned in Chapter 7, reducing them to a moltiinear dependence. in this case, the characteristic polynomial /() 's multilinear in all parameters but to, in this latter case with a quadiatie dependence. The new set of parameters generated by the adition of equality restrictions (see Chapter 7) Pi=P2 = Cip~ Poi = Peat (11.62) >= Pa = Cay = Pos + 8s = Poa + 5s (11.3) Ps = C= Post 8s (11.64) Po = Eng = Poo + 8 (11.65) b= C= Path (166) Ps = Cny = Pow + 8x (11.67), ‘The bounds on the uncertainties are 8) = 5) = 333%, 5 =3,=137,% K= 128% By = 58%, 5, = 189%, By = 14.3% (11.68) {or the fist linearization point and Bb =39%, == 111%, H=DO% iggy A= 636%, = 685%, | = 226% for the second. ‘The multivariable stability margin Km(jw) [1a,(Jw)] Was computed for a certain range of frequencies. The minimum values for each flight condition were km(Ja) = 3.15 [ia,(s0) = 0.317) at w = 3.5 rad/s in the fist case, and Kyg(Jo) = 2.89 [19, (Jen) = 0.346] at « = 10 rad/s for the second case. Thus the design is robustly stable. ‘The worst-case parameter combination, from the stability viewpoint, can also be obtained from the algorithm. This is the vector p* whose image frst reaches the origin of the complex plane for the smallest scaling factor k. This is a very important element in the analysis and redesign (if necessary) of the controller CONTROLOF A DCTOE RESONANTCONVERTER 407 This worst-case combination of parameters for flight conditions | and 2 were, inthis case, Ai = diag[) By Be 086-5) - i &) a1) ds = diag [ds 8 8-8-8. (O18-]) Be ] LTH The design of K(s) was made at both linearization points for the set of ‘nominal parameters py = py + 4y. with Ay as follows: ay = diag(—5 -& 5B -3 -& ] -&] C1172) {As observed, this disagrees with the values obtained in (11.70) and (11.71) for both fight conditions. This is a consequence of the fact that itis difficult to develop a physical iuurion on how the parameter variations will influence the stability of the system, especially in MIMO plants with a large number of uncertain parameters, As stated previously (see Chapter 7). the exact computation of ky (ja) or 443,(70) in the cases of parameter uncertainty has an exponential time compttation with the number of parameters, Nevertheless. in many cases, as the one presented here, it can be computed in a practical way by branch-and: bound procedures. Further details of the algorithm used in this ease ean be found in (266, 281). and this particular example has been developed in (265) under a NASA grant. 11.4 CONTROL OF A DC-TO-DC RESONANT CONVERTER 11.4.1 Introduction DC-to-DC resonant converters have been the object of much attention late since they have the potential to provide high-performance conversion. al- lowing for smaller, lighter power supplies. However. they require using a control circuit capable of maintaining the desired output voltage under dif- ferent operating conditions. In this section (based on the recent paper [65)) ‘we illustrate the z-synthesis technique by designing a controller for a paral- lel resonant converter. In addition to guaranteeing stability for a wide range of load conditions, this controller rejects disturbances at the converter input while keeping the control input and the settling time within values compatible with a practical implementation, 11.4.2. The Conventional Parallel Resonant Converter Figure 11.19 shows a diagram of a conventional second-order parallel res: nant converter. The combinations of the diodes and transistors form bi- directional switches operating at 50% duty ratio, Thus. in cach switching period. the resonant circuit L-C is alternatively excited by +Vy and — Vy408 APPLICATIONEXAMPLES [ts rtp EY Figure 11 Conventional second-order PRC cireutchagram, ‘The large output inductor L,, and capacitor C, are used to minimize the load effect on the resonant capacitor voltage and to ensure the constant out put voltage through the output circuit (37). As for notation, the resistor Ry and the voltages V_ and V,, represent the load, the line (input), and the output, respectively. Throughout this section we use as nominal parameters the following val- vves, taken from the design example in Chapter 2 of (37) L = 4.18 aH C = 11300F Ry = 208.330 = 10mH C= 47 nF V, = 100V Vv, = 250V f, > 200kHz VEC = 60.390 CONTROL OF A DC-TO.DC RESONANICONVERTER 409 For convenience, we introduce the following normalized variables: v, where the resonant frequency f, 11.4.3 Small Signal Mode! Under steady-state conditions it can be shown that, for a PRC operating in the continuous conduction mode [37], there are four circuit modes in each ‘witching period. Thus the converter is a nonlinear, variable structure system, with its steady-state trajectory uniquely determined by the normalized switch: ing frequency F,, and the load condition Q,. For a given operating point, a discrete time, small signal model of the converter can be obtained by using a perturbation method. The sampling time for this discrete time model is equal to T,/2, where T, = 1/f, is the switching period. Therefore it follows that this model is correct under small-signal perturbations with frequencies up to the operating switching frequency 1, = 2nf, = 1.26 x 10®radjs ‘The discrete time model from the normalized switching frequency Fyy and the normalized line Vag to the normalized output Vp (10 simplify the nota- tion, we use the same variables for both the steady state and its perturbation) at the nominal operating point is given by the following state-space rcaliza- tion [37|: X(k +1) = AX(k) + BUR) Volk) = CX(K) where 0.2767 0.6108 0.6644 0.0053 0.0075 0.9387 [ 0.8219 0.5504 -2 “| 6.4684 0.4834 B= | 106774 1.9499 = 0.0002. 0.0162410 _-APPUCATION EXAMPLES C= |0 0 345] The state variables and inputs are defined as X= [iath) vad) tal)" UCR) = [Fal Vag] where iy, 3nd Jy, are the normalized resonant inductor current, capacitor voltage, and output current, respectively. 11.4.4 Control Objectives Figure 11.20 illustrates the diagram used for control design. In the small signal model of the converter there ate two inputs: line voltage Vy and switching frequency F,,. The objective is to synthesize a controller having 4 input the error signal (obtained by comparing the output voltage versus the reference input r) and as output the switching frequency, F,, such that the output voltage is kept at a prescribed level (in our case V,, ~ 250 Vie. Vo = 2.5) at all operating points, This problem can further be divided into four parts ‘© Line Regulation (Nominal Performance). The line voltage is often unreg, tlated and could have a substantial range of variation, typically around + 20%. This variation will be modeled as an external disturbance, thus leading to a disturbance rejection problem. Performance specifications for this type of problem are usually given in terms of time-domain quan- tities, such as: 1. Zero steady-state error 2, Small overshoot at output (usually less than 10% for reference input step response). 3. Appropriate settling times for both fine and reference input step responses (5 ms at most in our case). Line Volage 7 CONVERTER ouput Switching Frequency Reference. ——] CONTROLLER Figure 11.20. the dlogrom tor contol design CONTROL OF ADCTODC RESONANT CONVERTER 411 4. A closed-loop bandwidth of at least 360 Hz. in order 1o successfully suppress line ripple. © Load Regulation (Robust Stability). Ow the other hand, the load con- dition could also vary over a wide range. Since the load R,, enters the ‘dynamics of the model, load variations will appear as model uncertainty and could possibly lead to stability problems. Normally the load changes from 10% at low load to 90% at full load condition, Other model uncer tainties, such as unmodeled high-frequency dynamics and uncertainties in the resonant inductor £. and capacitor C, will also be considered, Robust Performance. Since the converter operates over a wide range of load conditions, the performance requirements must be satisfied at all operating points. This is equivalent to requiring satisfactory response under both line and load variations. 11.4.5 Analysis of the Plant Control Characteristics For a PRC converter operating under steady- state conditions, the input-output relationships can be represented by the control characteristics curves, relating the output voltage to the load and switching frequency. Given any two variables among the normalized output Vo, Switching frequency ratio Fy,, and output load Q,. the third variable can be determined from the curves, Thus these curves allow one to casily visualize the effects of the switching frequency and load on the converter ‘output. From a control point of view, the control characteristics curves allow us to make an initial estimate of the load change that can be tolerated and to see some of the difficulties inherent in the load regulation problem. Figure 11.21 shows the control characteristics curves for various output loads Qp, obtained analytically from the steady-state analysis. To maintain the output voltage constant in the presence of perturbations, the controller should adjust the switching frequency to keep the converter operating along the dashed line indicated in Figure 11.21. As the converter is perturbed away from the nominal operating point (marked with an asterisk in the figure) the plant dynamics may vary significantly, resulting ina difficult control problem, Remark Note that from Figure 11.21 it follows that at lighter loads (higher Ry, larger Qp, and lower J,), a small frequency change will result in larger output changes. Thus we should expect that the control problem will become more difficult at larger Q, values, In the next section we will show. through a frequency-domain analysis, that this is exaclly the case. Frequency Responses From the discrete time state-space model, we can easily get the z-transfer functions from the normalized switching, frequency412 APPUCATIONEXAMPLES 19-— O5 055 06 065 07 075 08 085 09 095 1 Fre Figure 11.21. The conventional PRC contol characterises curves Fax and the normalized line input Vrg to the normalized output Vn Vno(2) Vnol2) Fus(2) Vrg(2) | ~ [o02) Ge(2) Ct-ap"B (11.73) Following a common approach, we will carry out the analysis of the plant and the synthesis of a digital controller using a w-plane approach. To this ‘effect, the bilinear transformation _Lash/4 ran (7) is used to obtain the transfer functions in the frequency domain s. These transfer functions, still denoted as G(s) and G,(), are given by (5+ 798041)(5 ~ 792481)(s — 800003) 2.652 x 10-2 + 7950415 = 792431)(s — 800003) ot (5 +29167)(s + 83363 + 202487i) Coe) (s = 800003)(s + 484950 + 2902534) Gals) = 136710 poreny(s R365 20D8T|) LT) as) ‘The above transfer functions correspond to the nominal load R, = 208.330. AS stated before, since the load enters the dynamics of the converter, load variations result in different transfer functions, Figures 11.22 and 11.23 show ‘CONTROL OF A OC-TO-OC RESONANT CONVERTER 10 104 NOE ak at aa at Al yor 1020 10* tPF C10? wirads) Figure 11.22, Frequency responses G(s) a ferent load conditions. ior igs i iv wieadls) Figure 11.28. Frequency responses G(s) at diferent load conditions. as.414 APPLICATIONEXAMPLES the frequency responses of G(s) and G,(s) corresponding to several different load conditions, respectively. These figures show that as the load becomes lighter (larger R,), the over shoot inereases, leading to a more difficult control problem. This conclusion is consistent with the conclusion drawn in the last section from the study of the control characteristics. On the other hand, the control characteristics require that Q» be greater than Vqp in order to get the prescribed output voltage. Since in our design example the value of Vn» is chosen to be 2.5, it follows that R, should be greater than 151. Therefore, in the sequel, we will assume that R, varies within the range 151 to 1200 0. 11.4.6 Control Design ‘As mentioned before, our goal is to design a controller that satisfies the performance specifications listed in Section 11.4.4 for all load conditions in the range 151.0 < R, < 12000, assuming that the values of the components of the resonant tank are known within a 10% tolerance. In the sequel we solve this problem by recasting it into a robust performance synthesis form and using synthesis. To this effect we need first to describe the family of plants corresponding to different values of the load as a nominal plant subject to uncertainty. Plant Description and Uncertainty Weight Selection In this exam- ple we choose to model the uncertainty caused by load variations by using a single, norm-bounded, multiplicative uncertainty that covers all possible plants. Let G¥(s) and G(s) denote the transfer functions from the control input and line input to the output at operating points other than the nominal point (Ry 208.332), respectively: In the sequel we represent these transfer functions as GMs) = Gls) [t+ 8/(5)Wi)] au) GPe(s) = Gas) [1 + Bg(5)65)] (11.78) where G(s) and G;(s) are the nominal transfer functions given in equations (11.75) and (11.76), respectively, W;(s) and We(s) are fixed weighting fune- tions containing all the information available about the frequency distribution of the uncertainty, and 4y(s) and A,(s) are stable transfer functions represent- ing model uncertainty. Furthermore, without loss of generality (by absorbing any scaling factor into W,(s) and W(s) if necessary), it can be assumed that [Sr(s)lx <1 and liMe(s)[le <1. Thus W;(s) and W,(s) are such that their respective magnitude Bode plots cover the Bode plots of all possible plants. Some sample uncertainties corresponding to different values of the load R, are shown in Figures 11.24 and 11.25. We can see that in both figures the ‘multiplicative uncertainties have a peak around the resonant frequency. This Peak becomes larger and steeper as the load resistance R, increases. Magnitude Figure 11.24, Multiplicative uncertainties (contol output and weight Magnitude 108 ‘CONTROL OF A DC-IO-DC RESONANT CONVERTER 10 104 107! 10" Cn a a nT) wiradis) 10 10 108 10F 10 wiradis) Figure 11.25, Mutipicatve uncertainties (ne fo output) and weight as416 APPUCATION EXAMPLES Based on these plots, the following multiplicative uncertainty weights were chosen for control design: 4x 10-45 +065 Wy) = AXIO 06 (1179) 104s Wo) = ee (1180) The magnitude frequency responses of W,(s) and We(s) are also shown in Figures 11.24 and 11.25, respectively. These figures clearly show that attempt- ing to cover the sharp peak around the resonant frequency will result in large gaps between the weight and the uncertainty at high frequencies, introducing conservatism at that frequency range. On the other hand, a tighter fit at high frequencies using higher-order functions will result in high-order controller. The weights (11.79) and (11.80) used in our design provide a good trade-off between robustness and controller complexity. ‘We turn our attention now to the effects of changes in the values of L and CC, the resonant tank components. Since these changes affect the location of the resonant peak, they could conceivably destabilize any controller design based on its nominal location. Figure 11.26 shows the changes in the trans: fer functions due to +10% changes in the values of L and/or C. It is worth noticing that our choice of weighting functions W; and W, will also cover this family of plants, even at the extreme load conditions Ry = 12000 and 151.0. Thus a robust controller designed using these weighting functions will be able to accommodate both changes in the load condition and uncertainty in L and C. Figure 11.27 shows a block diagram of the converter, taking into account the uncertainty. Here A; and Ay are scalar blocks, representing the model uncertainty perturbations from the control and line inputs, respectively, and W; and We are the corresponding uncertainty weights. Performance Weight Selection As discussed in Chapter 7, in order to guarantee robust performance we need to add to the structure shown in Figure 11.27 an additional (fictitious) uncertainty block 4p, along with the corresponding performance weights W, and W,, associated with the track ing/regulation error and the control effort, respectively, resulting in the block diagram shown in Figure 11.28. The selection of the performance weights W, and W, entails a trade-off among good regulation versus peak control action. The weight on the control error W,(s) is usually selected to be very large at low frequencies in order to get good tracking and regulation, Additionally, the order of the weights should be kept low in order to keep the overall controller complexity low. A g00d compromise between performance and compleaity is given by weighting CONTROL OF ADC-TODC RESONANTCONVERTER 417 Control to Output fe eee ee 107 10? 109 10% 10° 108 107 10% wads) Line 1 Output 10? us —— 10! ww fo 5 F151, L-9THH,C=t00F 20? 10 xo! 10) = = “107 10 70 “6 79 10! wired) Figure 11.26. Uncertoinies due to 410% changes of L ond/or C of extreme loos conditions Re = 120010 and 155,418 APPUCATIONEXAMPLES Figure 11.27. the block diagram of the converter including the uncertainty due to 00d and component variations. functions of the form W.(s) (181) BSA where A is the desired steady-state error (A will be zero if zero steady state is required); T, approximately determines the bandwidth (w, = 1/T) and Figure 11.28. The block diagtam for synthe ‘CONTROL OF A DC-IO‘OC RESONANTCONVERER 419 hence the rising time and settling time; where the ratio 73/7; is associated with performance requirements against high-frequency noise ((188]). Note that there is no exact relationship between the parameters 7; and T> and time-domain performance specifications given in terms of rise-time, setling: time, and overshoot. The design of multiobjective robust controllers subject to both time- and frequency-domain specifications is, to a large extent, an ‘open problem, although some progress has been made recently [301] In this particular design example we selected the following weights: w= ae ns 104s we) = cay The weight on the control input W,(s) was chosen close to a differentiator to penalize fast changes and large overshoot in the control input, The weight W(s) guarantees zero steady-state error. The frequency responses of W.(s) and W,(s) are shown in Figure 11.29. The closed-loop bandwidth correspond: ing to this choice of weights is approximately 1/0.0006 = 1700rad/s. Log Magnitude 10° cr 104 Frequency (radiansisec) Figure 11.28. Performance weights We(s) and Wu(s) 10420 APPUICATIONEXAMPLES 11.4.7, Controller Synthesis Combining the uncertainty description with the performance weights yields aan uncertainty structure 4 consisting of two scalar blocks (corresponding to the robust stability requirements) and a 2x2 block (corresponding to the ro- bust performance requirements). Since the A structure has only three blocks, the upper bound of 1, inf pep IDM D> |, coincides with its exact value. The robust controller was synthesized using the . Analysis and Synthesis Toolbox (20), applied to the block diagram shown in Figure 11.28. After four D-K itera: tions with third-order D-scalings, we obtained a 18th-order controller yielding veep = 0.9721. Finally, Hankel norm model reduction yielded a sixih-order controller with virtually no performance degradation (size = 0.9753 <1). ‘The state-space description of this reduced-order controller is given by K=Oyst~A)"'By + De (11.84) where 025 1008 -1as¢ nats nis 1.296, 1708 1320045 998e+4 3190045 1460044 —2213045 Aga | HIS 9980044 —16%e+4 S2MMeS 1200004 1.983665 w=] chats 34mess “320805 “2erdews 172905 4053045, FUN6L 14600 +4 1200644 hm8645 -2664e83 LOIBer S 1296 2213e05 “LBBerS “408305 “101965 “RORSe rs BY ~(-2338 7.983 5.345 —6.610 0.543 6.060) 10-2 G.=(-0.935 3.193 -2.138 2.644 0.217 2.424) x 10 De=0 In order to benchmark the performance of the robust controller, we also designed a phase-lag controller using classical design tools, based on the plant frequency responses at the various operating points shown in Figure 11.22. To improve performance, this controller was further tuned by trial and error. ‘The transfer function of the final controller is given by 0.025 + 200, 3402 The frequency responses of both the 4. and the phase-lag controllers are shown in Figure 11.30, Both controllers have similar responses at low fre- quencies, while at high frequencies the gain of the u controller decays faster im order to accommodate the model uncertainties at high frequencies. Figure 11.31 shows the closed-loop frequency responses for the nominal plant and for the lightest load considered in the design. Note that in both cases the y controller provides lower gain and better roll-off at high frequencies. Moreover, while the response corresponding to the phase-lag controller is acceptable for the nominal plant, it exhibits a large peak at the resonant frequency at light loads. As we show next, this peak results in significant performance deterioration at these loads. K, = (115) CONTROL OF A DC-TODC RESONANTCONVERIER 421 Log Magnitude 10 1 oe 10 rg 108 Frequency radianssec) 9 7 = & so é 09 50nd 10 10 10 10 6 10 Froquency (rasianaee) Figure 11.20. Frequency responses o 1 (solid) and phase-og (dashed) controls. 11.4.8 Simulation Results Linear Simulations The closed-loop system corresponding to the j. con- troller was simulated at the nominal operating point R, = 2080) and at two ‘extreme cases R, = 151 and 12000, using the corresponding linear model of the plant. The time responses 10 20% step change in line voltage Vag and reference input r are shown in Figure 11.32, For the nominal case R, = 2080, the settling time is about 2.5 ms for both line voltage change and reference input change. The output responses are satisfactory since the settling time is smaller than the required 5 ms, with ro overshoot. The control action in these responses is also adequate, without overshoot or abrupt change. This is due to the choice of the weight Wi, penalizing fast changes and overshoots in the control action ‘When the operating point moves to R, = 151.0, the setting times for both step changes are about 4 ms, This increase is due mainly to the significant decrease in plant static gain (see Figure 11.22). The u controllers undertuned at this operating point in order to achieve robust performance. When the operating, point moves toward lighter loads, the responses are almost the same as for the nominal plant, except that for the case R, = 12000 (note422 APPLICATIONEXAMPLES 10° — _ : ~ ~~ Phase-iag contalier Gain 104 p-controlle oe 108 10 10° Frequency (rad) ° Phase-ag controller /\ 1ot w-conteoller 108 10° 10" 10 10° Frequency (ad) Figure 11.31. Closedoop trequency rospontes for x (sold) ond phaselag (dashed) controllers: (4) Ry = 20881 and (0) Re ~ 1200, CONTROL OF ADCO-0C RESONANTCONVERIER 423 409) aa x10 085 oe 0G Time (582) 10? Time (see) 492 @) (b) Figure 11.82. Linear simulation resus with 4. conkolier at diferent operating points Fo = 208 1 (oid), 181 12 (dashed), and 1200 1) (does): (a) relerence input stop change (20%) and (bine vollage step change (20%) that this load is the lightest Joad considered in our design). some chattering in both the output and control input starts to show up at the beginning of the responses. The occurrence of the chattering is linked to the large peak in the plant frequency response at lighter loads, barely covered by the uncertainty weights W; and Wy From the simulation results it follows that the controller achieves robust, performance, since all performance specifications are satisfied at all operat: ing points of interest. However, significant variation of performance is also observed. This isa direct result of the large variation in the plant dynamics. and any fixed linear controller can do very litle in this respect, To reduce ‘this variation will require using a nonlinear, gain scheduling controller. The same simulation was performed for the closed-loop system corre sponding to the phase-lag controller. The time responses to 20% step change im line voltage Vy and feference input r at three different operating points — R, = 1510, 2080, and 1200—are shown in Figure 11.33. They are424 APPLICATIONEXAMPLES om 07 aso ano _ 5 5 aso | asl 0 0) aay 0 1 2 3 4 rece) gs : —— om a 2 34 2 3 Time (660) y Tere 600) yo! (a) (b) Figure 11.38. Linea simulation resus wih phaselag controler at diferent operating pins Re = 208 0 (soli), 181.0 (dashed), ond 1200 £1 (coted): (a) reference Input step. ‘change (20%) ond (b) ne volage step change (20%), similar to the responses with the 1 controller except that the performance is far worse for Ry = 12000. This is due to the phase-lag controllers inability to provide enough attenuation to counteract the increment in the magnitude of the resonant peak of the plant at heavy loads, as shown in Figure 11.31. Furthermore, as shown in Figure 11.34, at R, = 240010 the phase-lag con- troller fails to stabilize the system, while the controller can still produce acceptable performance. Nonlinear Simulations and Validation While simulations using the linearized model of the plant corresponding to different load conditions can usually provide an approximate evaluation of load regulation performance, this is usually insufficient to assess the performance of a highly nonlinear system such as the converter. Thus to further validate the controller, a non- linear simulation of the PRC circuit was performed using P-Spice. Figure 11.35, shows the responses due to reference input and line voltage step changes. Note that these results are similar to those obtained using a linear simulation CONTROL OF ADCIO.DC RESONANICONVERIER 425, 1 har om; Oem oa 0 foo 3 od aa rr os Cectmenianate aes ee nes) yg? me ees) yg? Input 2 Opt Int 22 4 eos 0s é q $4 $004 4 i a Eo q ons| oi se 8% Os se Tine 08 49? Time ane ag? ® 4 4 co a 2] = a ce lteeas Te see9} 40? “we sees) ap? 10" _Iput2 vet ‘to? _ hpi 2 Ou? «poe meen) , E os| il i a} 2 2 os ay Tre sos) 9? Figure 11.34, Stop responses of he closecioop systom at = 2400.2: (6) 4 contoler (slab) and (©) phoseag controller (unsiabi),426 APPLICATIONEXAMPLES 220) 200] S280 260) 2405 SO nce ses) Time (560) 97 095; og toss 03] 88 Time ($20) 49? Time (see) 30? (a) (by Figure 11.38. Nonlinear simulation resus wih. contoler a ferent operating points p= 208 1 (sold), 181 11 (dashed) and 1200 11 (coted) s: (a) reference input sep ‘change (20%) and (bine vollage sep change (20%). as shown in Figure 11.32. In the responses to line voltage step change and reference step change, settling times are slightly larger than those in the li car simulation, The chattering observed in the output voltage is substantial, due to the periodic switching behavior of the converter (periodic charge and discharge of capacitor). i ro i ‘6 fu bi BIBLIOGRAPHY ‘Abdallah C. "On the robust control of uncertain nonlinear systems,” Proceed ings American Control Conference, San Diego, 1990. Abdallah C., Dawson D., Dorato P, Jamshidi M, “Survey of robust control for rigid rob0W," IEEE Control Sysiems Magazine, Vol 11, NO. 2, 1991 ‘Abedor J., Nagpal K., Poola K.,"A linear matrix inequality approach to peak to peak gain minimization,” International Journal of Robust and Nonlinear Con: ‘rol, Yo. 6, pp. 99-927, 1996. ‘Ackermann J, "Parameter space design of robust control systems,” IEEE Trans actions on Auiomatic Control, Vol 25, pp. 1058-1072, 1980. ‘Ackermann J., Robust Control: Systems with Uncertain Physical Parameters Springer-Verlag, New York, 1993, ‘Adamjan VM,, Arov DZ., Krein MG., "Analytic properties of Schmidt pairs for a Hankel operator and the generalized Schur-Takagi problem,” Math USSR Sbornik, Vol. 15, pp. 31-73, 1971 Adamjan VM. Arov D.Z., Krein M.G., “Infinite Hankel block matrices and related extension problems,” AMS Translations, Vol. 111, No. 133, 1978, Akcay H., Gu G., Khargonekar P, “Identification in ., with nonuniformly spaced frequency response measurements," Proceedings American Control Con- ference, Chicago, 1992 ‘Akcay H., Gu G., Khargonekar P, “A class of algorithms for identification in Ex: continuous-time case,” JEEE Transactions on Automatic Control, Vol. 38, No. 2, 1993, ‘Anderson B.D O., VongpanitlerdS., Network Analysis and Synthesis: A Modern Systems Theory Approach, Prentice-Hall, Englewood Clif, NI, 1973. ‘Anderson B., Kraus F, Mansour M., Dasgupta S., “Easily testable sufficient conditions for the robust stability of systems with muliaffine parameter de- pendence,” in Robustness of Dynamic Systems with Parameter Uncertainties, M Mansour, S. Baler, W. Trudl, Eds. Birkhauser, Berlin, 192 427428 BIBLIOGRAPHY 12] Anderson B.D.O., Moore. JB. Optimal Conrot: Linear Quadratic Methods, Prentice-Hall, Englewood Cis, NI, 1990. 113) Astrom K., Eykhoft P, “System identication: survey.” Automatica, Vl 7, pp. 123-162, 197. [14) Astrom K. Witenmark B.. Computer Controled Systems: Theory and Design, Prentice-Hall, Englewood Cli, NJ, 1984 115) Athans M., Falb P, primal Control: An Inroduction 1 the Theory and Is ‘Applications, MeGraw Hil, New York, 1966, [16] Athans M, KapasourisP, Kappos E, Spang H.A. IM, “Linear quadratic gaus- Sian with Toop transfer recovery methodology for the F100 engine.” AIAA Journal of Guidance, Control and Dynamic, Vo. 9, No. 1, 1986 17] Balakrishnan VS., Boyd S., Balem S, “Branch and bound algorithm for com- Puting the minimum stability degree of parameter dependent linea systems” ‘Technical Report. Information Systems Laboratory Stanford University, 1991 (18) Balarishaan VS.. Boyd .P, “On computing the worst case peak gain of linear systems," Systems and Controls Leters, Vo. 19,192 119) Balas G., Doyle J.C, "Identiication of flexible structures for robust contol,” IEEE Comiro Systems Magazine, Vo. 10, No.4 199. [20) Balas G., Doyle .C. Glover K, Packard A, Smith R, p-Analysis and Syuess Toolbox. The MathWorks Ine, Natick, MA, 1991 [21] Ball.A., Gotberg I, Rodman L, Interpolation of Rational Matrix Functions, Operator Theory: Advances and Applications, Vol. 4, Birkhauser, Beri, 1990 [22] Ball 1A, Helton J.W, Walker MLL, "7, control for nonlinear systems wth ‘output feeaback.” IEEE Transactions on Automatic Control, Vo. 38, No. 4 1998, [23] Bancroft A.R, “Heavy water G-S process research and development achiewe- ments,” Atomic Energy Canada Limited (AECL) Report No. 6215, 197. 24} Banks HLT, Burns).A., “Hereditary contol problems: numerical methods based ‘on averaging approximation,” STAM Journal of Conirol and Optimization, No. 16, pp. 169-208, 1978 [25] Banks FLT, Kappel F, “Spline approximations fo functional differential equa tions,” Journal of Differential Equations, Vo. 34, pp. 496-822, 1979. [26] Barabsnov AE. Granichin OLN. “Optimal controler fo afinear plant with bounded noise,” Aulomation and Remote Control, Vol. $, pp. 39-46, 1984 [27] Barmish BR., “Invariance of the strict Hurwite property for polynomials with perturbed cosficients," IEEE Transactions on Automatic Control, Vo. 29,1984 {First presented atthe 1983 Conference on Decision and Control, Tex.) Barmish B. Khargonckar PP. “Robust stability of feedback control systems sth uncertain parameters and unmodelled dynamics.” Proceedings American Control Conference, Georgia, 1988 (29) Barmish B R, “New tools fr robustness analysis," Plenary Session, Proceedings Conference on Decision and Control, Austin, Texas, 1988. Barmish B, Ackermann JE, Hu HZ. “The tee structured decomposition: a new approach to robust stability analysis,” Proceedings Confrence on Informa ton Setences and Systems, Johns Hopkins University, Baltimore, 1989 es} [p9) BRUOGRAPKY 429 [31] Barmish B.R., "A generalization of Kharitonov’ four polynomial concept for ‘robust stability problems with linearly dependent coefficient perturbations,” IEEE Transactions on Automatic Control, Vol. 34, No, 2, 1989, [22] Barmish B., Khargonekar PP, Shi Z.C., Tempo R., "Robustness margin need not be a continuous function of the problem data,” System and Control Letters, Vol. 15, 196. [53] Barmish B.R,, ShiZ., "Robust stability of class of polynomials with coefficient depending mulilinearly on perturbations,” JEEE Transactions on Automatic Conirol, Vol. 35, No.9, 1990, [34] Barmish B., Kang H., “A survey of extreme point results for robustness of control systems,” Automatica, Vol. 29, No, 1, 1993, [35] Barmish BR. New Tools for Robusiness Analysis, Memillan, London, 1994, [36] Bartiet A.C, HollotC, Lin H., “Root locations of an entire polytope of poly ‘nomials it suffices to check the edges.” Mathematics of Control, Signals and Systems, Vol. 1, No.1, 1988. [57] Batarseh L., Steady Site Analysis and Design of High Order Resonant Convert rs, PhD. Thesis, University of Illinois, Chicago, June 1990 [38] Beck C., Doyle JC. “Mixed s2 upper bound computation,” Proceedings Con: {ference on Decision and Control, Tucson, 1992 [89] Bertsekas DR, Rhodes LB., “On the minmax reachability of target set and target tubes,” Automatica, Vo. 7, pp. 233-247, 1971 (40) Bersekas DP, “Infinie-cime reachability of state-space reglons by using feed: back control," JEBE Transactions on Automatic Control, Vol. AC-17, No. 4, PP. 604-613, 1972 [a1] Blanchini F, Sanaier M., “Rational £” suboptimal compensators fr continuous- time systems.” JEEE Transactions on Automatic Control, Vol. 39, No. 7, pp. 1487-1492, 1994, [42] Blanchini, F, “Non-quadratic Lyapunov funetion for robust contra” Automat ‘ca, Vol 31, No. 3, 451-461, 1994 (43) Blanchini F, Sznaier M. “Persistent disturbance rejection via static state feed back," IEEE Transactions on Automatic Control, Vol. 4, No. 6, pp. 1127-1131, une 1995, [44] Becker N., Grimm WM. “On £; and £,, stability approaches for the robust control of robot manipulators,” IEEE Transactions on Auomatic Control, Nol 33, No.1, 1988. (45) BetczynskiB., Kale 1, Cain G.D., “Approximation of FIR by IR digital filters: an algorithm based on balanced model reduction,” IEEE Transactions on Signal Processing, Vol. 40, No. 3, 992. {46] Bernstein, D., Wie, B., Organizers, Invited Session on “A Benchmark Prob- lem for Robust Ms / 7. Control Design,” Session WP-1S, Proceedings of the American Control Conference, San Diego, pp. 961-973, 1950. [47] Bhattacharyya SP, Chapellat H., Keel LH, Robust Convol: The Parametric Approach, Prentice-Hall, Englewood Cliff, NJ, 1995 (48) BialasS.,“A necessary and sufficient condition for stability of interval matrices, International Journal of Control, Vol. 37, pp. 117-122, 1983,430 19 150) 154] [ssi 6 1371 Iss 9 (60) (on (62) (63) (4), (6s) (6s) 67) BisLIOGRAPHY DBialas S. Garloff J, “Stability of polynomials under coefficient perturbations,” IEEE Transactions on Automatic Control, Vol. 30, 1985. Biernacki RM., Hwang H., BhattacharyyaS-P, "Robust stability with structured real parameter perturbations,” JEEE Transactions on Automatic Control, Vol. 32, 1987 Blondel V. and Gevers, M.. "Simultaneous stabibzation of three linear systems is rationally undecidable,” Mathematics of Control, Signals and Systems, Vol. 6, pp. 135-145, 1993, Bode H.W, Network Analysis and Feedback Amplifier Design, Van Nostrand, Princeton, NJ, 1945, Bodson M., Chiasson J.N., Novotnak RT, Rekowski R.B., “High-performance nonlinear feedback control ofa permanent magnet stepper motor,” IEEE Trans. actions on Control Systems Technology, Vol. 1, No. 1, 1993, Bose NK, “A system theoretic approach to stability of sets of polynomials,” Contemporary Mathematice, Vol. $7, pp. 25-38, 1985, Boyd S., Desoer C.A., “Subharmonic functions and performance bounds in linear time invariant feedback systems,” IMA Jounal of Mathematics, Control and Information, Vol 2, pp. 153-170, 1985. Boyd SP, Doyle J.C, "Comparison of peak and RMS gains for discrete-time systems,” Systems and Controls Lentrs, Vol 21, 1983. Boyd S. Balakrishnan V., Kabamba P, “A bisection method for computing the ‘Yo norm of a transfer matrix and related problems.” Mathematics of Control, Signals and Systems, No, 2, 1989. Boyd S,,Barrat CH., Linear Controller Design Limits of Performance, Prentice: Hall Information and System Sciences Series, Prentice-Hall, Englewood Clif, NJ, 1991, Boyd S., EI Ghaoui L., Feron E., Balakrishnan V, Linear Matrix Inequalities in Spstems and Control Theory, SIAM Studies in Applied Mathematics, Vol. SIAM, Philadelphia, 1994 Braatz R.D., Young PM., Doyle .C., Morari M. "Computational complexity of ju calevlation,” IEEE Transactions on Automatic Control, Vol. 38, No, 5, 1998. Brockett RW, Finite Dimensional Linear Systems, Wiley, New York, 1970. Bruck J, Blaum M., “Neural networks, errr-correctng codes and polynomials over the binary r-cube,” IEEE Transactions on Information Theory,No.9, 1989. Bryson AE,, Ho ¥.C,, Applied Optimal Control, Hemisphere Publishing, New York, 1975. Bu J., Sznaier M,, Holmes M., "A linear matrix inequality approach or synthe- sizing low order ! controllers,” Proceedings 38th IEEE Conference on Decision and Cantrl, Kobe, Japan, pp- 1875-1880, 1996. Bu J., Senaier M. Batarseh 1, Wang Z.Q., “Robust controller design for @ parallel resonant converter using, w-synthesis.” [EEE Transactions on Power Electronics, Vol. 12, No. 5, pp- 837-853, September 1997 Bu J., MultObjectve Robust Control: Design of Mixed €°/P,. Controllers, PhD. Dissertation, The Pennsylvania State University, 1997. Burns J, Peichl G., “On robustness of controllability for nite dimensional approximations of distributed parameter systems,” Technical Report. Dept. of (63) (69 (70 m 7) 73) (74) i (76) m (78) 9) (80) isi (82) 83] aauocRarHy 431 Mathematics. Virginia Polytechnic Institute and State University. Blacksburg, 1985, Chapeliat H., Bhattacharyya SP. °A generalization of Khatitonov’s theorem: robust stability of interval plants." IEEE Transactions on Automatic Control. Vol 34, No. 3 1989 CChapeliat H.. Bhattacharyya S.P, “An alternative proof of Kharitonov’s theo- rem." IEEE Transactions on Automatic Conwrol, Vol. 4. No. 4.1989. CChapellat H, Dahleh M. Bhattacharyya SP, “Robust stability under structured ‘and unstructured perturbations." IEEE Transactions on Automatic Con. Vol 35, No. 10, 1990 Chapellat H., Dahleh M, Bhattacharyya SP, “Robust stability manifolds for ‘multilinear interval systems.” IEEE Transactions on Automatic Conirol. Vol. 38 No. 2, 1993, Chen CT. Linear System Theory and Design, Holt, Rinehart & Winston. New York, 1984, Chen J., Fan M. Nett C, “The structured singular value and stability of un- certain polynomials: @ missing link,” ASME Annual Wier Meeiig, Atlanta, 1991, Chen J. Nett C. Fan M, “Worst-case system identification in W.: validation of priori information, essentially optimal algorithms. and ertor bounds.” Proceed: ings American Control Conference, Chicago. 1992; also in IEEE Transactions ‘on Automatic Control. Vol. 40, No. 7. 1995 Chen J. Nett C, Fan M.. “Optimal non-parametric system identification from arbitrary corrupt finite time series: a control-oriented approach.” Proceedings “American Conirol Conference, Chicago, 1992. Chen J, Farrell, Nett Zhou K.,"% identification of multivariable systems by tangential interpolation methods.” Proceedings Conference on Decision and Control, Florida, 1994 Chen J., Nett C., “The Carathéodory-Fejér problem and the 1. /t! identi cation: 9 time domain approach,” JEEE Transactions on Automatic Control Vol. 40, No. 4, 1995 Chiang, R, and Safonov, MG., “Ha. Robust control synthesis for an undamped. hon-colocatedspring-mass system.” Praceedings American Control Conference. PP. 966-967, 1990. Curtain R., Glover K., “Balanced realisations for infinite dimensional systems.” Workshop on Operator Theory and lis Applications. Amsterdam. 1985, Dahleh MA., Pearson J../-Optimal controllers for discrete time systems, Proceedings American Control Conference. Seattle. pp. 1964-1968. 1986, Dahlen M.A.. Pearson J. "f"-Optimal controllers for MIMO discrete time sys tems.” IEEE Transactions on Automatic Control, Vol 32. No. 4.1987 Dahlen M.A,, Boyd Pearson J. °'-Optimal compensators fr continuous time systems.” [EEE Transactions on Automatic Conirol, Nol. 32, No. 10. pp. 88— 895, 1987, Dahlch MA.. Pearson J.. ‘Optimal rejection of persistent disturbances. robust stability and mixed sensitivity minimization.” FEEE Transactions om Autonnatic Control. Nol 33, No. 8, 1988.432 BIBLIOGRAPHY WN) Dableh M.A.. Ohta Y.,"A necessary and suflicient condition for robust BIBO. stability” Systems and Control Lens. Vel, 11. pp. 271-275, 198 {Ks} Dalen M.A... TheodosopoulosT. Tsitsiklis J, ~The sample complexity of worst ease [dentcation of FLR. linear systems,” Systems and Control Leter. Vo. 20. Dp. 157-166, 1993 'Ko) Dahleh M.A.. Khammash M.. “Controller design for plants with structured un certainty.” Automatica, Vol 29. No. 1. pp. 37-56. 193. {87| Dahleh M.A.. Diaz-Bobillo LJ... Control of Uncertain Svstems: A Linear Pro gramming Approach, Prentice-Hall, Englewood Cis. NI 1995 (88) Fu M.. Dasgupta S.. Blondel V.. “Robust stability under a class of nonlinear parametric perturbations.” IEEE Transactions on Auwomatic Consrol, Vol. 4. No, 2.1985, [89] D’Auetis CE. ouroduccién a los Sistemas Nolineales de Control y sus Aplica (a&b) € F. Operation & is associative, that is, (@& b)&e= a&(b&C) = akb&e for a,bceF Operation & is commutative, that i, a & b = b &a for a,b F. Set F contains the neutral element ng with respect 10 &, that is, there exists ng © F such that (a &ng) =a for all ac F 44s446 MATHEMATICAL BACKGROUND 5. Set F contains the inverse element al, with respect to &, that is, for all ae F, there exists aly € F such that (a8 al.) = ne Set F is closed with respect to Operation + is associative Set F contains the neutral element n, with respect to * Set F contains the inverse element al with respect t0 Operation + is distributive with respect to &, that is, (a& b) we = (a & (x0) for a,b,c F. Take, for example, set R equipped with operations (+, x) as (8, +) respec: tively, then mg =0.,= 1, ah = ~a, and al = a7! (a0), A.1.2. Linear Vector Space Definition A2._A set V is a linear vector space over the field (F,+,*) if and only if the following properties are satisfied: (Im the sequel the elements of F and V will be called scalars and vectors, respectively) Set V is closed with respect to + Operation + is associative in V. Operation + is commutative nV. Set V contains the neutral element with respect 10 + Set V contains the inverse element with respect 10 + V is closed with respect to operation x between scalars and vectors. Operation x among scalars and vectors is associative in the scalars, that is, (ax b) xv =ax (b xv) =ax bx for abe F and vey. 8. Disiributive I: (a+b) x v = (a xv) +(b xv) for a,b € F and ve ¥. 9. Distributive 2: (u+v} x a= (ux a) +(v xa) fora € F and uv € V. 10. Field F contains the neutral element of operation x between vectors and scalars, that in x0 =v form. € F and v€ V. In the sequel, equivalently. inear vector space, linear space, or vector space will be used Definition A.3_ The set of vectors {04,...,9} in the linear space V, defined over the field F, is linearly independent if and only if the linear combination ev bo ay = a = 0, with a, € Fi =1,...,m. Other wise they are called linearly dependent, Definition A.4 A set of linearly independent vectors {Uy,..-,tn} of @ linear space V constinue a base of V if and only if any v € V can be represented ‘as a linear combination of these vectors. The dimension of V is the minimum number of vectors that constitute a base, in this case n. ALGEBRAIC STRUCTURES 447 Linear spaces may have finite or infinite dimension, as will be shown next bby means of several examples. Example Al The marrices in ""' or C°!, constirue finite-dimensional lin ear spaces. The representations of position in physical space with respect 10 a certain reference frame also constitute a finite-dimensional linear space. The sequences of real numibers Sz. oF the set of continuous functions defined in a real interval Cig, form infinite-dimensional linear spaces. These last two cases may represent the pulse sequences in digital systems or certain classes of signals in analog circuits, respectively A.1.3. Metric, Norm, and Inner Products Definition A.S A metric space (V,d\-,.)) is defined in terms of a linear vec~ tor space V and a real function d{-,»):V x V = R,, satisfying the following conditions: 1 diy) 20, Vay ev. 2dyy)=0 =e xray: 3. dry) = dla), Yayev. 4 dena) x=0 3. lex||=lal-ix], Vx Via F. 4, xe yll) is defined in terms of 4 linear vector space V and a function <_,- >: x V+ F that satisfies the following conditions ¥ x,y,z €V and a ¢ F: kxtye>acne>+ FS. 20, Vr 60 i 2, 3. a. 4 S=0 oox=0. Here @ represents the complex conjugate of element a, Example A.3- The three normed spaces from the last example can be made inner product spaces with the following definitions: 1, Take vectors in C* and define t [Foire (ayy ‘An inner product space generates a normed vector space when using its inner product (o define a norm, as follows: |x\|= =X. >. In addition, a normed vector space generates a metric space when using its norm to define the distance function, as follows: d(x, y) = lx ~ yl Definition A.B Two vectors x,y in (V,< +>) are said 10 be orthogonal if and only if =0. If, in addition, =1 and < y,y >= 1, they are called orthonormal ‘A.2 FUNCTION SPACES 4.2.1 Introduction Next, some important properties of infinite-dimensional vector spaces are briefly reviewed. Definition A.9 A sequence {x,} in a metric space (V, 4) is said to be a Cauchy sequence if and only if Ve > 0 AN such that nm > N > dn, %m) < € Definition AO A sequence {9} in a metric space (V,d) converges to x, €V if and only if ¥€ > 0 AN such that n > N => d(tq%.) < € IL is easy to prove that any convergent sequence is a Cauchy sequence. (On the other hand, the converse, is not necessarily true. Example Ad Consider among the rational numbers Q the sequence {3 3.14 34415. This isa Cauchy sequence, although it does not con: verge because = ¢ Definition A.11 A metric space (V,d) is said 10 be complete if and only if all Cauchy sequences are convergent in V. A.2.2 Banach and Hilbert spaces jon A.12 A complete normed vector space (V, | -|\) is ealled a Banach ition A.I3 A complete inner product space (V.<-,->) is called a Hilbert space450. MATHEMATICAL BACKGROUND x(t) Figure A.1. Nonconvergent sequence in Cra Example AS The infinite-dimensional vector space of continuous functions defined in the unit interval C3, #8 not complete when using the metries (x,y) = |x = ylh oF daxyy ilo, where these norms have been de: {fined in (AS). ‘Note in Figure A.J thatthe sequence X(t) of continuous functions converges ww a discontinuous one. Therefore the Space of continuous functions is not complete, under these metrics 2.3. Operator and Signal Spaces Many ofthe useful function spaces, which willbe described inthis subsection, include other than continuous functions. In these cases integration should be understood inthe Lebesgue integral sense (Riemann integral is valid ony for continuous functions). In addition, the concepts of minimum and maximum of a function (min, max) are replaced by their generalizations: infimum and supremum (int, sup). Finally, essential supremum, denoted as ess sup, refers to the supremum of a function for al but sets of measure zero in its domain These functional analysis concepts can be found in greater detail in 174) Time Domain ‘* C\(R,R) is the space of absolutely integrable scalar functions, equipped with the norm (a0) FUNCHIONSPACES 451 © £\(R,R"™*) isthe space of absolutely integrable matrix functions, equip. ped with the norm fil mol wn (Any 1* £:(R,R™") is the space of square integrable matrix functions of 1 € R, with inner product - + 74(B, RY") isthe subspace of £(R, R°*") formed by causal functions, that is, zero for: <0, The notation £,((0, 20), R”™) is sed equivalently + 715 (8, RP") isthe subspace of £3(R,R”") consisting of anricausalfonc- tions, that is, zero for 1 > 0. The notation £((—o0,0],R"*™) is used equivalently. Next, we show that for a causal stable system G, ||Gilz can be intexpreted as the RMS value ofits output z = G+ u when the input u isa white noise signal with unit covariance. To this effect, given a wide sense stationary random vector signal u(r) define its autocorrelation as ate) =€ ote] cary where £ denotes expectation. The Fourier transform of Ru(1) is called the spectral density Suu(jw), that is, Saja) = J Rane ar (a4) follows that €(u w) = trace [E(uu)] = trace [Ry(0)] Ibis a standard result that if z= Gu then Sj) = G(jo)Su(ja)G"Gw). Thus aeto-meettool= 2 ["wowtutwide Ly eS 52 = ay [Trace 610)" a) = 16 where we used the facts that Sya(jiw) = I and that trace(GG") = trace(G").452 MATHEMATICAL BACKGROUND Frequency Domain # C)R,C"™") is the space of square integrable matrix functions of jo, @ € R, with inner product, = a | se (FUG y0y"}de (Ais) and norm ||F\|2° V. Here we have defined G'(s) £ GT-s), + 7(UR,C™") is the subspace of £(jR, C""") composed of analytic func- tions in Re(s) > 0 such that spf trace|Fjw+o)Fyw+o)'|de 0 = 1,UR,C™™) is the subspace of £..()R,C™") of analytic and (essen- tially) bounded functions in Re(s) <0. £{R,-) and £,GR,-) are Hilbert spaces; £,.(JR,-) is a Banach space. In many cases, when these spaces are restricted to the set of rational transfer functions or matrices, an R is added to the notation, that is, RL3, RLao. RH, RM. In the main part of the book, the arguments of these spaces will be dropped for simplicity, these being clear from the context A.2.4 lsomorphism Definition A.14 Two vector spaces V and W are said to be isomorphic if and only if there exists a linear transformation L.: V —+ W, satisfying the following conditions: 1. L(x) = LO) > x= y EV, that is, N(L) = {Q}. This property is defined as monomorphism or one to one. 2. w= L(x) for all w € W with x € V, that is, R(L) = W. This property is called epimorphism or onto. FUNCIIONSPACES 453 Here N() and R(.) represent the kernel and range spaces of an operator, respectively. Definition AS Two normed vector spaces (V, |v) and (W,||- lw) are said to be isomette if and only if they are isomorphic and, in addition, the @ linear transformation L satisfies lly = WL(0)|w for any vector v € V. Example A The normed linear vector spaces £3(R,R) and £3(R,©) are isomorphic and isomeric under the linear transformation F : g(t) Ga) defined as the Fourier transform: Vf oper J 16003? ao This is also known as Parseval’s theorem. Furthermore, the inner product ‘among vectors in both spaces is also preserved, isle Giz (A.19) y — (A21) As a.consequence, the “algebraic structure” of both spaces is similar. In addi- tion, orthonormal vectors in one space will remain orthonormal in the trans- formed space, that is, “geometry” and “proportion” are preserved from one vector space to the other. Due to the isometry that the Fourier transform establishes between £(R) and £,()R), their subspaces can be related in the same way. Therefore %9(R) and Ha()R) are isometric, as well as 43 (R) and 13 ()R). Hence causal signals (or systems) ate isometric with stable transfer matrices and anticausal signals (or systems) with antistable ones. A.2.5 Induced Norms Definition A16 Consider two normed linear vector spaces (V,\\- Iv) and (WV, Iv) and a linear transformation L : V — W. The induced norm of the transformation is defined as follows: A IL()lw Wy 2 sup WOOl (a2) vive Telly It can be proved that equivalently the induced norm can be computed as SUP yy, 1 HEC)lw.454 MATHEMATICAL BACKGROUND Example A.7 1. Consider the operator A € R"”" and use the Euclidean vector norm js in both R" and R” A. ieTATAX Malia © 9B Tid sel V aT = YEE) | aaa) (ax) The latter will be defined in the next section as the maximum singular value of A, 2. If the norm |) is used instead in RY and R, the following induced norm is obtained: Ath x = max Yai (A2a 3. When || - oo is used in R" and R" the following induced norm is obtained: All ow = max lanl (a2s) i ‘The induced norm quantifies the maximum “amplification” of a linear operator, 8 generalization of the concept of gain in SISO transfer functions. ‘Take, for example, the following matrix: f [om 10° 0 0.001 (a26) which can be interpreted as a linear operator A :R? ~ R?. Its eigenvalues Ay 0.001 do not measure the maximum amplification of vectors in RE when multiplied by the matrix. On the other hand, its induced norm (Alla 2 = (A) ~ 10? does. For example the input w= 0. 1)", produces an ‘output having Aullp = a(A) = 108 ‘An important property of induced norms is the fact that they are sub- smulipicative, that is, [> Lill w < [!Lalv—wlLi yw Proof. From (A.22) we have that jLyllw < hLllw-—wlyllw- Ht follows that MaLayllw < Waliw av hLaylw < Law ovliLallw-awllyllw— (A27) Equivalently, Mataylh Se < Uallw_wllLslww A28) ly = Weal wile (A28) ‘The proof follows now directly from the definition of induced norms, © FUNCTION SPACES 458 A.2.6 Some Important Induced System Norms In this section we consider the £* and £? induced system norms and we show that they coincide with the €, and H.. norms introduced in Section ALB The £*-+£ Case For simplicity consider the SISO case. For any in put u(t) € £*, lull = 1, we have that the corresponding output y = g + w satisfies boots fee sandr s [enue s [nian =i ‘This shows that lglle.c. < ligll. To show that the equality holds, assume that lg = Ho. Then, given € > 0, there exist T such that (A29) bom es fella sto (a30) ider che following signal aoe {omar osrsT 0 otherwi (ey) I is easily seen that the corresponding output y(¢) satisfies jy > |y(T)] > Bo ~ €. Thus sup gulls 2 Ho ~€ (ax) Halles ex ‘The proof follows now from the fact that ¢ is arbitrary, The CC Case Consider any input w € £2,;Iuj2 = 1 and let z = Gu Using the frequency-domain definition of |- >. we have that Hells = WGu(s)3 (A33) =¢ [ L2G Yay yaya) ds ve Eff stove genie de 1p (Ge 3 fw e)W10) deo GON. (As)456 MATHEMATICAL BACKGROUND ‘This shows that the Mq norm is an upper bound of the £> induced norm, To show equality we need to find some w. € £2, |Iu.ll2 = 1, such that [22 = UiGllx. Assume that sup, @ [Gla)] is achieved at the frequency w.. Consider the input 1a(t) = k(e}e-* £05 (wal)man, WHETE Vas iS the Unitary right sin gular vector that corresponds to 4 (GUw,)] and k(e) is such that |4.(3)h for € > 0. As ¢ +0 we have that leallS = Tim G(s)u.(5)IH 5 i [GUoomek(e)[6(w +0.) + 5(w— a)? do + HG, )omex|” = 2 (Gye)] =I), (35) The 4 Norm ‘The Hz norm does not admit an interpretation as a norm induced by a single space (and thus it does not have the submuliplicaive property). Nevertheless, it i instructive to show that it can be considered fs an induced system norm, if the input and ouput spaces are allowed to have different norms. Given a signal u(t) we can define its autocorrelation and spectral density proceeding exactly as in Section A23. If Ryi(2) and Sy(jw) exist and Ry, is bounded, then w is said to be a power signal, with power (semi)norm given by |lull = trace [R,,(0)]. A power signal v is said to have bounded spectral density if ||Si«(J0}l|~ <0. In this case we can define its spectral density norm |jlls = |Sve(Jw)|}ac- Proceeding as in Section ‘A.23 it can easly be shown (by considering a signal u(?) having Sy, = 1) that Gl = IIIs Finally, in the SISO case, |j.\|2 can be interpreted as the £; — £o. norm. ‘To establish this fact consider € Cx, Illy 1. The corresponding output satisies viol = [f° ete ~ nara « ([Lat-otae)” (fiwerae)" = Isl (A36) where the second line follows from the Cauchy-Schwartz inequality, To com: plete the proof we need to show that there exists at least one signal u(t) € Lays lhully = 1 such that lly > gla To this effect take a-d Osh (a37) DUALITY AND DUAL SPACES 487 Obviously falls = 1 and Ae lla > Iy(OW TED Wolke 2 DOl= Tei Fa A.3 DUALITY AND DUAL SPACES (Aan) 4.3.1 The Dual Space Consider a linear vector space X. A functional on X is a function f:X'— R, Definition AAT Consider a normed linear space X. Its dual space X" is defined as the space of all bounded linear functionals on X, equipped with the norm | fl = stPjx4<1 If) It can be shown (187 that X" isa Banach space. In the sequel, given x € X and r © X", (x,r) denotes the value of the linear functional r at the point x. Example A.7: The Dual of f°, 1< p< oo Consider the space #” of all real sequences h = {,} for which >", /hi)? < 90, equipped with the norm Welle = COR Mul?" (i p =e then [Ase = supy !hu|)- Given 1 < p < 00, define the conjugate index g by the equation! (A39) Next, we show that every linear functional fon €? has a unique representation of the form A) = Show (A40) where = {n} € €*. Assume first that 1

0, there exists my € M such that ie ml) QandU ecm, are unitary matrices. ‘crow pan "Thun npone ocelot ast of rbonomategemesr satay DE o viacaw = [8 8] ass Let = YEA. Without tos of generaiy we ssume thatthe eigen vale of WPA are ode. that 2 Dene Vand te itr anv -n columnsof Vesecvey Repiang m (A00) min {\jrl>, Ax =b, Ae C™™, n < mh (as) The solution to these problems can be obtained by means of the matrix pseudoinverse or generalized inverse A*. The latter can efficiently be com puted with the use of the singular value decomposition as follows: =z 0 a-u[% 8 : : EO] y. Je eer fs Qo ay ere (Ae) ‘The input vector x = (1 1/2 L2"" yields Ax = —2""1 ay that tends to the zero vector as n —+ 90. Hence for a large enough m this ‘matrix is computationally singular, although itis theoretically not. ‘The eigenvalues are Ay = --- = An = —1 and the determinant is det(A) = (-1)", none of which provide a clear measure ofthe computational singularity of this matrix. In this last case, the reason is that det(A) = []} A,, which therefore depends on all eigenvalues, not only on the smallest one. In the former case, the measure provided by the smallest magnitude eigenvalue only quantifies the effect of the matrix on the eigenvectors, but not on any arbitrary vector in the input space ‘On the other hand, by considering the minimum value [Axl can take over all possible input vectors x (normalized by {illp 4.0 10 avoid trivial solutions), the measure of singularity can be computed as Axi RTATAR int Mele ing iehzo [ie] neo (EPID way twat which is directly the minimum singular value oA) = 2!" SINGULAR VALUES 463, Condition Number Consider the linear matsix equation Ax = b. The ‘maximum relative error 6b due to an error 5x, can be found as follows 86H yp Adele ner [Ole ~ Se ART =1 with axils = lal sup IAB inf, [Axl with [Sxl = [lela = 1 A) 2 7 wa!) (Ass) where 5x and x are independent vectors. The problem is equivalent to the one of finding the maximum relative error 8x of the solution for @ given error 8b in the data vector b. The condition number x(:) provides this measure and quantifies the relation between the maximum and mini- mum “gains” of A (according to different input vector “directions”). For ‘example, orthogonal (rotation) or unitary matrices are “spherical” because U'U = 1 = oy == oy = 1s hence in this ease «(U) = 1. On the other hhand, for singular matrices « —+ oo.APPENDIX B SYSTEM COMPUTATIONS In the following sections we present the state-space representations of the ‘most usual interconnections between linear systems. These representations are used to manipulate systems in a computer code. There is standard soft ware that implements these interconnections in state-space form, as in the case of the j-Toolbox and the Robust Toolbox of Matlab. Consider the following linear systems ous) = 54] (Ba) 2 | Be | (62) Gals) = |-p with states, inputs, and outputs (x1,4,y1) and (x, u2,92), Fespe series connection can be seen in Figure B.1, where uu). yam ‘The objective is to compute a state-space realization (not necessarily min- imal) of the product system. To this end, consider the state of Ga(s}Gu(s) as ely. The and x= [xf 21)’, and proceed as follows: AU = Arante) + Bini) (B3) 465