Sie sind auf Seite 1von 9

ControlEng. Practice,Vol. 4, No. 5, pp.

625-633, 1996

~ )

Pergamon
PII:S0967-0661 (96)00044-5

Copyright 1996 Elsevier Science Ltd Printed in Great Britain. All fights reserved 0967-0661/96 $15.00 + 0.00

WEIGHTING FUNCTION SELECTION IN THE Hoo DESIGN PROCESS


R.W. Beaven*, M.T. Wright* and D.R. Seaward**
*Machine Controland Drives Research Group, Departmentof Mechanical and Electrical Engineering, Aston Unviersity, Birmingham, B4 7ET, UK (beavenrw@aston.ac.uk) **Molins Pie, Advanced Technology Unit, Coventry, CV4 7EZ, UK

(Received September 1995; in final form December 1995)

Abstract. The theory of Hoo-norm optimisation is well known, and the technique has been applied to numerous design examples. Although simple rules exist for the selection of weighting functions used in the controller synthesis, the process is still subjective in nature. This paper is concerned with two aspects of the determination of weighting functions: firstly, guidelines for the initial selection of weighting functions and secondly tuning the performance weighting function to achieve the required closed-loop performance. The theory is illustrated with reference to the design of a dc servo system controller. Keywords. Hoo controller synthesis, dc servo-systems.

1. INTRODUCTION Considerable progress has been made in the synthesis of Hoo controllers since the original work of Zames (1981). The technique has a number of theoretical advantages that make it appealing to practising control engineers, such as addressing both stability and performance criteria, and producing a system with almost exact loop shaping (Chiang and Safonov, 1992). The technique has been applied to numerous design examples, both theoretical and practical (Postlethwaite et al., 1986; Limbeer and Kasenally, 1986). In particular it has been shown, for a design example involving a de servo system used in incremental mode, that an Hoo controller can be designed that gives better closed-loop system performance than a traditional PID with velocity feedforward controller (Beaven, et al., 1994). Two broad criticisms, each consisting of many individual points, are levelled against the technique. Firstly, the selection of weighting functions, used in the controller determination, follows no objective set of criteria, although simple guidelines exist for the selection of weighting functions and general discussions on weighting function selection can be found in the standard texts, such as (Maciejowski, 1989). Secondly, the relationship between the frequency-dependent weighting functions and the actual closed-loop time response is difficult to reconcile.

In this paper a direct relationship between the nominal model, weighting functions and final Hoo controller is given (referred to as the "direct-form Hoo controller"). The initial selection and readjustment of the performance weighting function is discussed in terms of this direct-form Hoo controller. The discussion is limited to single-input single-output second-order systems (Liu and Liu (1990) demonstrated the use of this model in servo system applications). Many of the points made are high-righted with respect to de servo systems, but are considered valid in a wider control context. 2. RESUME OF H** CONTROLI.~R SYNTHESIS The Hoo norm of a transfer function, F, is its maximum value over the entire frequency range, and is denoted IIFII= sup a(F(jco))
co (1)

where 5(.) is the largest singular value of a transfer function. In the traditional Hoo controller synthesis (used in this paper) two transfer functions are used. The traditional two-sensitivity-function H ~ controller synthesis algorithm is used, since it allows a complex control problem to be split into two discrete sections, one dealing with stability, the other dealing with
625

626

R.W. Beaven et al. controller exists if positive semi-definite solutions can be determined to two Riccati equations, derived from the augmented state-space representation,and the spectral radius of the two solutions is strictly less than 7 (in this application 7 is assumed to be unity). A fuller description of the theory of H-- control is readily available (Pich6, et al., 1991; Francis, 1987). 3. PROBLEMS IN WEIGHTING FUNCTION SELECTION LundstrOm, et al. (1991) discussed two reasons for difficulty in weighting function selection: 1) In many real applications the performance specification is not detailed before the design starts, i.e. it is simply to achieve the best possible performance. 2) Alternative methods exist for weighting function selection since the problem can be formulated using different methods, and there are several physical intexpmtations of the Hoo-norm. Additional difficulties in weighting function selection include: 1) Under what criteria is the optimal controller to be selected, since different combinations of weighting functions produce different "optimal" controllers ? 2) The controller synthesis is such that the performance specification is achieved for the nominal model. Therefore the actual closed-loop performance may be so different from that required that the system is not viable. Guidelines on the adjustment of the weighting function parameters in such a situation are unavailable (to the authors'knowledge). 3) Non-linearities are difficult to include in the design process. 4) How are weighting functions to be selected for high-bandwidth demand profiles, such as the step and dwell protrfles commonly used in the process

performance. Alternative methods such as the three design specification method (Engell, 1995) were not examined since the two-sensitivity-function approach is well documented and has been found to be adequate. The sensitivity function, S, and the complementary sensitivity function, T, are used in the controller synthesis and are given by: S= I (I+ GK) and T= GK (I + OK)
(2)

where K is the controller and G is the plant. It can be seen (Fig. 1) that S is the transfer function between the reference signal, r, and the tracking error, e, whereas T is the closed-loop transfer function. The sensitivity function has an associated frequency-dependent weighting function, WI. This weighting function is used to limit the magnitude of the sensitivity function within a particular frequency range (i.e., the H~-norm of the combined function WIS is less than a particular value, usually unity, IIWtSIIoo<l). This, WI, is the performance weighting function (measure) in the controller synthesis. In a similar fashion the complementary sensitivity function, T, is combined with a frequency-dependent weighting function, W2, which is used to limit the magnitude of T within a particular frequency range (i.e. II W2T Iloo< 1). This, W2, is the robustness weighting function (stability constraint) in the controller synthesis. The two sensitivity functions are combined, with their respective weighting functions, into a single augmented state-space representation of the design problem (as standard (Chiang and Safonov, 1992)). The problem then becomes one of finding a stabilising controller, if it exists, that satisfies both the performance and robusmess criteria. Glover and Doyle (1988) showed that a stabilising

d Y
K

where r = reference input u = control signal y =plant output K ffi Controller


Fig. I. General Block Diagram

d = disturbance input n = sensor noise e = tracking error G = Plant

H**Design Process industry ? 5) Many practical systems require controllers with limited complexity (due to the limited time/computing facilities available). In general an increase in the update time of a process leads to a decrease in performance (Morari and Zafiriou, 1989). 6) Undefined time delays, both in the implementation of the controller and inherent in the system, are difficult to incorporate into the design process. 7) Controllers are usually implemented digitally, so the effects of digitisation (i.e. time delays and numerical roundoff) are difficult to incorporate into the design process. 8) Controller reduction, which is often required to produce implementable conm)llers, is difficult to quantify, and hence to incorlxn'ate into the design process. 4. DIRECT-FORM Hoe CONTROLLERS The relatively simple state-space H=o method of Glover and Doyle (1988) with the initial transformation detailed by Safonov, et al. (1989) was used to determine the direct-form Hoe controller. The direct-form Hoo controller was determined using the symbolic manipulators Mathematica (Wolfram, 1991) and MapleV (Char, et al., 1991). These symbolic manipulators allow a general solution of an equation to be formed. In the case of the direct-form H ~ controller a symbolic version of the nominal plant model and the weighting functions were defined and the steps detailed by Safonov et al. (1988), and used in the MatlabTM script HINF.M, for determination of a Hoo controller, were completed on the symbolic model and weighting functions. At each stage in the construction of the direct-form Hoe controller a numerical check was completed by comparing the numeric solution of the symbolic form, with numerical values assigned to each symbol, to that obtained using the MaflabTM script HINF.M. Once the controller had been determined extensive numerical analysis (based on a comparison with the results from the MatlabTM script HINF.M ) was undertaken to ensure the validity of the controller. If the nominal model is restricted to a simple stable second-order system (appropriate for a de servo system (Liu and Liu, 1990)) of the form
Go(s) =

627

1990). The direct form of the controll~, if it exists, is given by

-(b 4- ~8 4- ~) Xeel3P 2 "2 K(s)--~tt4-S)(~2 S2+(Xeel 1p2+lq/~2)s4-0t2Q +Xeol2P 2)

(5)
where X=on, Xoe~2and Xoel3 ~ elements of the first row of the solution of the single Riccati equation. The Riccati equation solution is determined using the following Hamiltonian matrix -
1

-~
0

0
0

_p2 0 ~2
0 0

0
0

0
LO0

~ E2.i 0 ct2e2 1-e2


1-e 2

~-eO 82-1 O ea(~-0) 1.e2


e2.1

0 fl

0 _L. 1.e 2 -1 0 0 "-.:g-e2.1

0 0

0 e= (P.~.I.-O) (P.~-0) 2
_

0
e2_ 1 _

(6)

The eigenvalues and hence eigenvectors of Lop are determined. Because of the form of the matrix this can be done symbolically (this is not shown here due to its complexity). The eigenvectors are sectioned as

v:[V,1v12]
V21 V22
x~ = v21v~.

(7)

with Vn and V21 forming the stable eigenspace of L~. The Riccati equation solution is given by
(8)

The direct-form H-= controller allows two closedloop forms of the system to be determined, one for the initial determination of the weighting functions and the second for alteration of the weighting functions, after the controller has been implemented, in an attempt to achieve the specified closed-loop performance.
If the actual plant is equivalent to the identified

model then the closed-loop system response is T = CL1 CL2s 3 + CL3 s2 + CL4s + CL5 (9)

s2+~s+8

(3)

and if the performance weighting function W l and the robustness weighting function W2 are given by Wl=es+0 s+tt
, W 2 = s2 + y s + t ~ , (4)

where CL1 = a X ~ 3 p 2 , CL2 ffi - a 2 , CL3 = ( - a 2 - a2tt - X.~I p2) CL4 = (-a2tt - a2t~ - ttXoqlp2 - X ~ n p 2) and CL5 -- (-~2~1~ - ~Xool2P2 + (xXool3p2). If the actual plant is

p A s2+Bs+C

then the Hop controller can be determined by the solution, Xoo, of a single Riccati equation (Hvostov,

(1o)

628

R.W. Beaven et al.

the closed-loop system response is

V (s2+~s+8)
l"s3+(-lff'-BI'+V)s2+(V~-I.IB1P-I')s+VS-p_r'C

(11)

with r = {X2S24"(~l/Ot2+XOOll p2)S+(0t2L'~+Xoo12P2) and V = X**13ct2A. A more complex real model could have been used, but for simplicity a second-order model is used. A more complex nominal model could be used and the direct form Hoo controller could be redetermined for use with higher order systems (for example third or fourth order systems). The only difference in the determination of the direct-form Ho* controller for use with higher order nominal models would be that the Riccati Equation would not have a symbolic solution (obtained by the eigenvector method) due to the nature of the L,o matrix. The controller would also be of greater complexity, so that the direct-form Hoo controller could not be so rc.~lily understood and used. Additionally it is common practice to reduce the complexity of a controller (in order that it can be implemented)~ meaning that any benefits of using a more complex model may be eliminated.

5. NOMINAL MODEL AND ROBUSTNESS


WEIGHTING FUNCTION SELECTION The nominal model can be determined experimentally using a swept sine or step response test, or by a more complex identification method such as least-squares parameter identification. Experience has shown that these techniques are adequate for dc servo systems (Beaven, et al., 1994). Alternatively, the nominal model can be constructed theoretically, using readily available information such as torque constants, rotor inertia, etc. The nominal model should be restricted to its simplest form, since the controller complexity is related to that of the nominal model, via the augmented state~ space representation (Chiang and Safonov, 1992). The robustness weighting function is constructed to allow for two types of disparity between the nominal model and actual system: firstly, that due to inaccuracies in the nominal model, such as unmodelled dynamics, and secondly, the disparity due to changes in the dynamics of the system. Both can be estimated by a comparison of the actual and theoretical response of the system over a period of time. The robustness weighting function is constructed such that
GOt) -I

achievable by Wl. (This was noted by Chiang and Safonov (1992)). Therefore, the inverse robustness weighting function should be a high-bandwidth transfer function so that it acts as an unstructured broadband stability function. The robustness weighting function zeros are transformed, with some modification, to the poles of the controner. Therefore if the W'21 poles are high-frequency they will be transformed into high-frequency poles in the controller, and can be removed in limited-bandwidth applications. The controller simplification (removal of high frequency poles / zeros) can be completed by an analysis of the controller frequency response and a consideration of the system bandwidth. For sampled control systems (such as those used to control dc servo systems) sampling places two limitations on the achievable system bandwidth: one due to the lime delay (phase shift) introduced into the system (Seaward, 1989) and the other due to the sampling effect (i.e. Shannon Sampling Theorem (Lynn and Fuerst, 1993)). This means that the achievable system bandwidth can be estimated and any controller zero and poles that are outside this range (with an appropriate safety margin) could be removed. Additionally the controller complexity can be reduced (if necessary) by use of standard reduction techniques such as zero / pole cancellation or Hankel-norm minimisation. A number of simple rules are available for the determination of the robustness weighting function:1) Make the robustness weighting function improper, such that GW 2 is proper (since GW 2 is used in the controller synthesis): this will limit the complexity of the final controller. 2) W2 can be determined using Eq. 12 and the actual and nominal model system response. 3) If possible, the weighting function should be selected such that W'2 is a high-bandwidth transfer t function for two reasons. Firstly, this will reduce any unnecessary restriction on the performance specification. Secondly, it will allow a possible reduction in the complexity of the controller (in limited-bandwidth systems, i.e. sampled systems) by removing high-frequency controller poles,

inducedby We.
6. PERFORMANCE WEIGHTING FUNCTION SELECTION Two general strategies for deriving performance weighting functions are discussed. Lundstr0m, et al. (1991) use a classical frequency domain specification using the following criteria 1) Steady-state offset less than O; 2) Closed-loop bandwidth higher than rOB; 3) Amplification of high-frequency noise less than a factor M; Then the performance weighting function Wl is of the form

Oo0O))

[ ~ ]W20o)[

vo

(12)

where G is the actual system response (or expected response) at each frequency and Go is the model response at each frequency. Analysis of the direct-form H,o controller shows that the 0dB crossover frequency of W'21 limits that

HooDesign Process

629

Wl .... s + Moo B M(s+OoB)

(13)

The second form of performance weighting function (Pich~, et al., 1991) is given by
Wl = 11s + Z s + ~X (14)

This can be used to re-determine the value of ~t. Note that ~t is not set to zero, since this would place a controller zero on the imaginary axis, causing a 90 phase shift to occur at low frequencies, which could cause stability problems. 7. DESIGN EXAMPLES 7.1 Simple Design Example The points made in the previous section are illustrated using the following design example. The plant and nominal model are given as follows: Actual Plant = 4070 s2 + 145s + 545 40~0 s2 + 125s + 525

The parameter lq, which is limited to 0 < 11< 1, is the high-frequency limiting value of Wl, and serves to constrain the maximum resonance peak in the ideal frequency response. The parameter Z > 0 gives the 0dB crossing of W1 (and thus of S) when ~ is small. The parameter k was suggested to be closely related to the steady-state error of the system.

(17)

Note that the performance weighting function given in Eq. 4 is of the same form as those used by both Lundstr0m, et al. 0991) and Pich6, et al. (1991), with the weighting function parameters simply defined differently.
The authors have developed a strategy for determining the performance weighting function that is closely related to that used by Pich6, et al. (1991). The closed-loop form in Eq. 9 (used for the initial estimate of the weighting functions), shows that the steady-state response for a step input (a de servo system used in incremental mode experiences a series of step demands) is ~L1 CL5
=

Nominal Model =

The performance specification is for an endpoint error of less than 0.1%, after not more than 45 ms, for a step input. The nominal model was determined from the actual plant using a swept sine wave test. A comparison of the actual and nominal responses at set frequencies yields the robustness specification given below W2 = ~2 + 9.5x104~ + 1.9x108 2x10s
(18)

00('131)2 -x21a~- ~tXoot2p2 + aXoox3p2

(15)

Next, an initial value of tt should be determined. From Eq. 12 it can be seen that all the required parameters are known except for X*012 and X**~3; initial estimates of these can be obtained using a standard performance weighting function of Wl(standard) - 0,998 + 1 s+O (19)

Thus, the higher the value of tt the greater the steadystate error of the system. An analysis of the direct form of the controller indicates that an increase in the value of 0 leads to an increase in the frequency of the poles in the closed-loop system and an increase in the controller gain, increasing the system bandwidth and therefore increasing the speed of response (Raven, 1987). Note that this ignores the effect of high-frequency noise amplification. This concurs with classical control theory, since an increase in the value of 0 is equivalent to a decrease in the tracking error over a particular frequency. The allowable values of 0 and ~t are interrelated with an increase in tt leading to a possible increase in 0. Therefore there exists a trade-off between the two parameters. The initial design specification of a system usually includes an acceptable level of endpoint error which can be used to determine an initial value of ~t. Analysis of Eq. II yields the steady state response of XooI3 p2A 8 Xeo13p2AS- I.t (ot2~ + XooI2p2)C "

This performance weighting function has been found to produce a controller with virtually all plants used. Experience has shown that the value of ~ should be set to 0.99 to stop the closed-loop system from having an excessive resonant peak. During the determination of the direct-form Hoo controller it was found that e cannot be greater than unity (for the particular controller synthesis set-up used in this paper). Using this performance weighting function, the value of X**12 is 5.1006x10q and X**13 is -4.6538x10-4; these values are taken from the single Riccati equation solution given in Eqs 6-8. These lead to a steady-state value of -7.5392x1016 -2.3355x1016tt - 7.3592x1016 "
(20)

(16)

A value of tt = 0.003 would yield a system with a steady-state error within tolerance; however the X** values in the final controller will be different from those used in the estimate, since they are dependent

630

R.W. Beaven et al. Additionally the controller has a lower low frequency gain (50% less at 0.1 Hz). When this controller is applied to the actual plant the performance specification is achieved in 44 ms, as shown in Fig. 4. 1.002

on the performance weighting function parameters selected. Therefore a Ix value of 0.0001 is selected, giving margin of error of 30. The corresponding optimal 0 value is 82.2. The performance weighting function is Wl = 0.99s + 82.2 s + 0.0001 (21)

The weighting functions (plus the nominal model) were used to produce the Hoo controller 3,9465X1011s2 + 4.933|g10136 + 2.0719x1014 s3 + 5.6759x107s 2 + 6.0716x1012s + 6.0716x10 g" (22) The controller has zeros at-120.6485 and-4.3515 and poles at -5.6652x107, -1.0717x10s and -lxl0 "4. The response in Fig 2. shows the nominal closedloop response of the system, determined assuming that the actual and identified models are equivalent. The required performance specification is met in 25 ms. The step response of the system if the controller is implemented with the actual plant is shown in Fig.3. The response is outside the specification, since the system overshoots by 0.158% and requires 189ms to settle to within 0.1% of the step input. The endpoint error (after approximately 500ms) is zero, indicating that the nominal model parameters c~and 8 are close to the actual plant parameters. The simplest method available to overcome the problem of overshoot (in this example) is to increase g, effectively shifting the entire response downwards. The required shiftis between -0.06 and -0,09%, The redetermination of ~t can be completed using either Eq. 12 or Eq, 13. For the sake of simplicity Eq, 12 is used. The steady-stateresponse, using the Xoo determined in the previous controller synthesis (since the performance parameters are close to their final values, i.e. the X~ values are closer to their final values) becomes "2.6217x1022 .9.9589x1019g. 2.6217x1022 (23) 1.002 ~ 1.000

--i/'i
-

0.998 0

20

40 60 80 Time in ms

100

Fig. 2. Ideal closed-loop response 1.002

f-l I l

~ 1.000

0.998 0

20

40 60 80 Time in ms

100

Fig. 3. Actual closed-loop response

A ~t value of 0.2 is selected, to get a -0.076% steadystate offset, which allows for a margin of error, caused by the expected change in the Xoo values. The corresponding optimal 0 value is 82.8. The performance weighting function becomes WI = 0.99s + 82.8 s+0.2 (24)

~ 1.000

This can be used to re-determine the controller as 0.998 1.3255x101t~ 2 + 1.6568xlO13s + 6.9587x1013 s3 + 1.9134x107s2 + 2.0401x1012s + 4.0802x1011 (25) The controller has zeros at -120.6485 and -4.3515 and poles at -1.9027x10% -1.0722x10 s and -0.2. 0 , 20 40 60 80 Time in ms 100

Fig. 4. Actual closed-loop response adjusted H~o controller

H~ Design Process 7.2 More Complex Design Example An Advanced Continuous Simulation Language (ACSL) simulation was constructed for an actual design example. The simulation consists of a nonlinear continuous fourth-order brushless de motor model with a digital PID controller operating with velocity feedback. The velocity controller is nested within a digital positional control loop. Both control loops have an update time of 1 ms. The models used for the simulation of the motor model and basic controller have been previously demonstrated to give results which correlate well with those achieved experimentally (Beaven, et al., 1994). The design problem is as follows: an Electro-Craft BRU200 DM-25 drive module controlling a S-3016 brnshless de motor (rotor polar moment of inertia 8.3x10 -5 kgm 2) is required to rotate an axis with a total referred inertial load of 1.718x10 "4 kgm 2 forwards through 105 deg in 20ms and then return the load to its original position in a further 20 ms. The endpoint error for each directional motion must not be greater than +1.5 deg. The motor has an 8000 pulse per revolution optical eneoder, and therefore each directional motion corresponds to moving through 2334 pulses in 20 ms with a maximum endpoint error of -+33 pulses. Note that for the sake of simplicity only one directional motion is studied. The Bode plot of the system response (Fig. 5) was determined using a swept sine wave test (dotted line phase, and solid line gain). This was used to construct the nominal model Go(s) = 4x107 s2 + 1001 s + 100 (26) This produces the following controller 0.705 ls 2 + 705.8235s + 70.5115 . (28) s2 + 1.0714x105s + 107.1439

631

The controller has zeros at -1.0009x103 and -0.0999 and poles at -1.0714x105 and -0.001. The controller was transformed, using a Tustin bilinear transform, to the discrete-time controller u(0
=

0.01938*e(t) - 0.02584*e(t- 1) + 0.06454*e(t-2) + 0.03665*u(t-1) 0.96335*u(t-2) (29)

e(0 = error signal at time t. u(0 = control signal at time t. Note that the sampling rate of the system (which is a hardware limit) is such that the high frequency controller pole is outside the controller bandwidth and will not be directly translated to the appropriate discrete time location. This demonstrates one of the problems in determining controllers for practical applications. Problems in controller implementation are one of the reasons that there is often a requirement to redetermine the performance weighting function, and hence the controller, to achieve the desired closed-loop system performance. When this controller was implemented in the position control loop around the motor simulation, the results in Fig. 6 were obtained. The system response is not within the tolerances, since the system overshoots to 2514 pulses and requires 30ms to settle within _+33 of 2334 pulses. The response shows no appreciable steady-state offset, indicating, using Eq. 13, that the identified model parameters and 8 were of the correct order. However the transient response indicates that the value identified for the I~ parameter was incorrect, since the system displays an excess overshoot and settling time (the expected system response, if the identified model were equivalent to the actual system would be the same as that displayed in Fig 2.). Using Eq. 11 and the MatlabTM function STEP.M (Grace et al., 1992) the value of B was determined as 420 (i.e., the actual model is more lightly damped than the identified model). This approach (the estimation of the underlying plant parameters, for use in the redetermination of the performance weighting functions) is used since it utilises all available information and is intuitively appealing. The performance weighting function parameter 0 is used to alter the transient response. This approach can be justified in two ways. Firstly from traditional control theory a decrease in 0 is effectively a reduction in the required level of system error, which would correspond to a decrease in controller gain and hence of system overshoot. Secondly, from an analysis of the direct-form H.* controller, a reduction in 0 by a factor P has, to a first approximation, the following effects on the X** values. The value of X*-11 is decreased proportionally, the value of X*-12 is decreased by a factor of (P + P/2) and the value of X*-13 is decreased by a factor of (P - P/2). Therefore

100 80 60 "" ____.

0
-100

-200 ~ 40 20
0 "~

-300 i 400

.1

10 100 Frequency in Hz

1000

Fig. 5. Frequency response of the system. The following weighting functions, determined as in the first design example, were used: Wl ffi ~ , W 2
s + 0.001

= s2 + 9"SxlO4s + l'9xlOg 2xlOs (27)

632

R.W. Beaven et al.

the direct-form H** controller, together with a function such as STEP.M in MatlabTM,can be used to find, numerically, a first approximation of the required value of 0. Initially the value of tt is readjusted (increased to 1) using Eq. 16, to assist in altering the transient response of the system by shifting the entire response downwards, by 1.12% (26 pulses) and by increasing the frequency of the low-frequency controller pole. Then Eq. 11 is used to determine, using the approximate X,* values relationship, a first approximation of the value of 0 to he used. This yields a 0 value of 60, which (using Eq. 11 and the function STEP.M) corresponds to a closed-loop system with a 0.5% (12 pulse) overshoot. Thus, the performance weighting function

3OO0
m | m | m m

becomes

10

20 30 40 Time in ms

50

Wl - 0.99s + 60 s+l This produces the following controller

(30)

Fig. 7. Final closed-loop response Note that Eq. 11 usually only determines the approximate value of 0 to he used, and a number of iterations are required to determine the final performance weighting function. 8. CONCLUSIONS This paper has addressed the problem of selecting weighting functions in the design of using the H** techniques so that this, presently ad hoe, portion of the controller design process can be carried out in a more structured way (although a certain degree of trial and error is still required). Two different parts of the design problem have been distinguished - designing the initial controller, and refining the controller to match the desired closedloop response for the system. A direct transformation between the parameters of the weighting functions and plant and the resultant controller, has been used in both cases. An indication of the method used in the determination of the direct-form Hoo controller has been given.

0.4651s 2 + 465.5847s + 46.3302 .(31) s2+ 1.0173x105s + 1.0173x104 The controller has zeros at -1.0009x103 and -0.0995 and poles at -1.0173x105 and -1. Additionally the controller gain is lower (70% of the original value at 1 Hz). The controller is then digitised to u(t) = 0.01345"e(0 - 0.01793*e(t-1) + 0.00448*e(t-2) - 0.03756*u(t-1) 0.96048*u(t-2) (32) e(0 = position error at time t. u(0 = control signal at lime t. When this controller was implemented it produced the closed-loop system response shown in Fig. 7, where the system settles to within .+.33of 2334 pulses in 20 ms.

controllers

3000 I 2OOO I000 0


0
I I I I

9. ACKNOWLEDGEMENTS The financial support of the Engineering and Physical Sciences Research Council and of Molins Plc for this work is gratefully acknowledged. In addition the authors would wish to record the invaluable advice and guidance of the staff of the Molins Plc Advanced Technology Unit in establishing benchmark standards. 10. REFERENCES 10 20 30 40 Time in ms

50

Fig. 6. Initial closed-loop system response

Beaven, R.W., Wilkes, L.D., Wright, M.T., Garvey, S.D. and Friswell, M.I., 1994, "An Adaptive Control System Incorporating H** Concepts for High Speed Machinery with Independent Drives" Proc. LMech.E., Part I, Journal of Systems and Control Engineering, Vo1208, pp43-52

HooDesignProcess Char, B.W., Geddes, K.O., Gonnet, G.H., Leong, B.L., Monagar, M.B. and Watt, S.M., 1991, MapleV Library Reference Manual, SpringerVerlag, New York, U.S.A. Chiang, R.Y. and Safonov, M.G., 1992, Robust Control Toolbox, The Maths Work Inc, Massachusetts, U.S.A. Engell, S,, 1995, "Design of Robust Control Systems with Time-Domain Specifications", Control Engineering Practice, Vol. 13, No. 3, pp365-372. Francis, B.A., 1987, A Course in H** Control Theory Springer-Verlag, Berlin. Glover, K. and Doyle, J.C., 1988, "State-Space Formulae for all Stabilizing Controllers that Satisfy an Hoo-normBound and Relations to Risk Sensitivity", Systems and Control Letters, Vol 11, pp167-172. Grace, A., Laub, A.J., Little, J.N. and Thompson, C.M., 1992, Control Toolbox - For Use with Matlab TM, The MathWorks, Inc., Cochituate Place, 24 Prime Park Way, Natick, Mass. 01760, USA. Hvostov, H.S., 1990, "Simplifying H~ Controller Synthesis Via Classical Feedback System Structure", IEEE Transactions on Automatic Control, Vol. 35, No. 4, April 1990, pp485-488. Limbeer, D.J.N. and Kasenally, E., 1986, "H.* Optimal Control of a Synchronous Turbogenerator" in Proc. IEEE 25th Conf. Decision Contr. (Athens, Greece), Dec. 1986, pp62-65. Liu Tian-Hua and Liu Chang-Huan, 1990, "Implementation of AC Servo Controllers Employing Frequency-Domain Optimisation Techniques", IEEE Transaction on Industrial Electronics, Vol. 37, No 4, August 1990. LundstrOm, P., Skogestad, S. and Wang, Z-Q., 1991, "Performance Weight Selection For H-infinity and tt-eontml Methods", Transaction Institute of Measurement and Control, Vol. 13 No 5, pp241252.

633

Lynn, P.A. and Fuerst, W., 1993, Introductory


Digital Signal Processing with Computer Applications, John Wiley & Sons Ltd., Coventry,

U.K. Maciejowski, J.M., 1989, Multivariable Feedback Design, Addison-Wesley Publishing Company, Wokingham,England. Morari, M. and Zafwiou, E., "i989, Robust Process Control, Prentice-Hall International Editions. Pith6, R., Pohlolalnen, S. and Virvalo, T., 1991, "Design of Robust Controllers For Position Servos Using H-infinity Theory", l.Mech.E. Journal of Systems and Control Engineering, Vol 4, No. 1, pp299-307. Postlethwaite, I., D-W. Gu., S.D.Young and M.S. Tombs, "Industrial Control System Design Using H** optimisation", Proc. IEEE 25th Conf. Decision Contr. (Athens, Greece), Dee. 1986, 131312-13. Raven, F.H., 1987, Automatic Control Engineering: Fourth Edition, McGraw-Hill International Editions, New York. Safonov, M.G., Limbeer, DJ.N. and Chiang, R.Y., 1989, "Simplifying the H** Theory Via Loopshifting, Matrix-Pencil and Descriptor Concepts.", Int. J. Control, Voi 50, No. 6. pp2467-2488. Seaward, D.R., 1989, Continuous Phase
Synchronised Drives (For A Rod Making Machine), PhD Thesis, The University of Aston

in Birmingham, Birmingham, U.K. Wolfram, S, 1991, Mathematica: A system for doing mathematics by computer, Addision-Wesley Publishing Company, Reading, Massachusetts, U.S.A. Zames, G., 1981, "Feedback of Optimal Sensitivity: Model Reference Transformations, Multiplicative Semi-norms, and Approximate Inverses", IEEE Trans. AC, Vol. AC-26, 1981, pp301-320.

Das könnte Ihnen auch gefallen