You are on page 1of 16

14

Parameter Sensitivity and State-Space Trajectories

14.1

INTRODUCTION

In Chap. 13 it is shown that an advantage of a state-feedback control system operating with high forward gain (hfg) is the insensitivity of the system output to gain variations in the forward path. The design method of Chap. 13 assumes that all states are accessible. This chapter investigates in depth the insensitive property of hfg operation of a state-feedback control system. This is followed by a treatment of inaccessible states. In order for the reader to develop a better feel for the kinds of transient responses of a system, this chapter includes an introduction to state-space trajectories. This includes the determination of the steady-state, or equilibrium, values of a system response. While linear time-invariant (LTI) systems have only one equilibrium value, a nonlinear system may have a number of equilibrium solutions. These can be determined from the state equations. They lead to the development of the Jacobian matrix, which is used to represent a nonlinear system by approximate linear equations in the region close to the singular points. 14.2 SENSITIVITY

The environmental conditions to which a control system is subjected affect the accuracy and stability of the system. The performance characteristics

of most components are affected by their environment and by aging.Thus, any change in the component characteristics causes a change in the transfer function and therefore in the controlled quantity. The effect of a parameter change on system performance can be expressed in terms of a sensitivity function. This M sensitivity function S is a measure of the sensitivity of the systems response to a system parameter variation and is given by
M S

dln M dln M d dln  d dln 

14:1

where lnlogarithm to base e, M systems output response (or its control ratio), and  system parameter that varies. Now dln M 1 dM dln  1 and 14:2 d M d d  Accordingly, Eq. (14.1) can be written    dM =Mo M S   M Mo d=o   o     dM  fractional change in output   M Mo fractional change in system parameter M d 
o

(14.3) where Mo and o represent the nominal values of M and .When M is a function of more than one parameter, say 1, 2 , . . . , k, the corresponding formulas for the sensitivity entail partial derivatives. For a small change in  from o , M changes from Mo , and the sensitivity can be written as    M =Mo M S  M M % 14:4 o =o  o To illustrate the effect of changes in the transfer function, four cases are considered for which the input signal r(t) and its transform R(s) are fixed. Although the responseY(s) is used in these four cases, the results are the same when M(s) is the control ratio. Case 1: Open-Loop System of Fig. 14.1a The effect of a change in G(s), for a fixed r(t) and thus a fixed R(s), can be determined by differentiating, with respect to G(s), the output expression Yo s RsGs 14:5

FIGURE 14.1

Control systems: (a) open loop; (b) closed loop.

giving dYo s Rs dGs Combining these two equations gives dYo s dGs dY s=Yo s Y s Yo s ! SGs s o 1 Gs dGs=Gs 14:7 14:6

A change in the transfer function G(s) therefore causes a proportional change in the transform of the output Yo(s). This requires that the performance specifications of G(s) be such that any variation still results in the degree of accuracy within the prescribed limits. In Eq. (14.7) the varying function in the system is the transfer function G(s). Case 2: Closed-Loop Unity-Feedback System of Fig. 14.1b Hs 1 Proceeding in the same manner as for case 1, for G(s) varying, leads to Yc s Rs dYc s Rs Gs 1 Gs dGs 1 Gs2 14:8 14:9

dYc s
Y s SGs

dGs 1 dGs Yc s Y s Gs1 Gs 1 Gs Gs c dYc s=Yc s 1 dGs=Gs 1 Gs 14:10

Comparing Eq. (14.10) with Eq. (14.7) readily reveals that the effect of changes of G(s) upon the transform of the output of the closed-loop control is reduced by the factor 1/j1 G(s)j compared to the open-loop control. This is an important reason why feedback systems are used.

Case 3: Closed-Loop Nonunity-Feedback System of Fig. 14.1b [Feedback Function Hs Fixed and Gs Variable] Proceeding in the same manner as for case 1, for G(s) varying, leads to Yc s Rs dYc s Rs Gs 1 GsH s dGs 1 GsH s2 14:11 14:12

dYc s
Y s SGs

dGs 1 dGs Y s Y s Gs1 GsH s c 1 GsH s Gs c dYc s=Yc s 1 dGs=Gs 1 GsH s 14:13

Comparing Eqs. (14.7) and (14.13) shows that the closed-loop variation is reduced by the factor 1/j1 G(s)H(s)j. In comparing Eqs. (14.10) and (14.13), if the term j1G(s)H(s)j is larger than the term j1 G(s)j, then there is an advantage to using a nonunity-feedback system. Further, H(s) may be introduced both to provide an improvement in system performance and to reduce the effect of parameter variations within G(s). Case 4: Closed-Loop Nonunity-Feedback System of Fig. 14.1b [Feedback Function Hs Variable and Gs Fixed] From Eq. (14.11), dYc s Rs Gs2 dH s 1 GsdH s2 14:14

Multiplying and dividing Eq. (14.14) by H(s) and also dividing by Eq. (14.11) results in ! GsH s dH s dH s Y s dYc s Yc s % 1 GsH s H s H s c
Y s SH s

dYc s=Yc s % 1 dH s=H s

14:15

The approximation applies for those cases where jG(s)H(s)j>>1. When Eq. (14.15) is compared with Eq. (14.7), it is seen that a variation in the feedback function has approximately a direct effect upon the output, the same as for the open-loop case. Thus, the components of H(s) must be selected

as precision fixed elements in order to maintain the desired degree of accuracy and stability in the transform Y(s). The two situations of cases 3 and 4 serve to point out the advantage of feedback compensation from the standpoint of parameter changes. Because the use of fixed feedback compensation minimizes the effect of variations in the components of G(s), prime consideration can be given to obtaining the necessary power requirements in the forward loop rather than to accuracy and stability. H(s) can be designed as a precision device so that the transform Y(s), or the output y(t), has the desired accuracy and stability. In other words, by use of feedback compensation the performance of the system can be made to depend more on the feedback term than on the forward term. Applying the sensitivity equation (14.3) to each of the four cases, where M(s) Y(s)/R(s),where  G(s) (for cases 1, 2, and 3), and  H(s) (for case 4), M yields the results shown inTable 14.1.This table reveals that S never exceeds a magnitude of 1, and the smaller this value, the less sensitive the system is to a variation in the transfer function. For an increase in the variable function, a positive value of the sensitivity function means that the output increases from its nominal response. Similarly, a negative value of the sensitivity function means that the output decreases from its nominal response. The results presented in this table are based upon a functional analysis; i.e., the variationsconsidered are in G(s) and H(s). The results are easily interpreted when G(s) and H(s) are real numbers. When they are not real numbers, and where  represents the parameter that varies within G(s) or H(s), the interpretation can be made as a function of frequency. The analysis in this section so far has considered variations in the transfer functions G(s) and H(s). Take next the case when r(t) is sinusoidal. Then the input can be represented by the phasor R( jo) and the output by the phasor Y( jo). The system is now represented by the frequency transfer TABLE 14.1
Case 1 2 3 4

Sensitivity Functions System variable parameter G(s) G(s) G(s) H(s) SM  dYo =Yo 1 dG=G dYc =Yc 1 dG=G 1G dYc =Yc 1 dG=G 1 GH dYc =Yc % 1 dH=H

functions G( jo) and H( jo). All the formulas developed earlier in this section are the same in form, but the arguments are jo instead of s. As parameters vary within G( jo) and H( jo), the magnitude of the sensitivity function can be   plotted as a function of frequency. The magnitude SM jo does not have the  limits of 0 to 1 given in Table 14.1, but can vary from 0 to any large magnitude. An example of sensitivity to parameter variation is investigated in detail in the following section. That analysis shows that the sensitivity function can be considerably reduced with appropriate feedback included in the system structure.

14.3

SENSITIVITY ANALYSIS [3,4]

An important aspect in the design of a control system is the insensitivity of the system outputs to items such as: sensor noise, parameter uncertainty, crosscoupling effects, and external system disturbances.The analysis in this section is based upon the control system shown in Fig. 14.2 where T(s)Y(s)/R(s) and where F(s) represents a prefilter. The plant is described by P(s) and may include some parameter uncertainties (see Sec.14.4). In this system G(s) represents a compensator.The prefilter and compensator are designed to minimize the effect of the parameter uncertainties. The goal of the design is to satisfy the desired figures of merit (FOM). In this text it is assumed that F(s) 1. The effect of these items on system performance can be expressed in terms of the sensitivity function which is defined by
T S

 @T T @

! 14:16

where  represents the variable parameter in T. Figure 14.2 is used for the purpose of analyzing the sensitivity of a system to three of these items. Using the linear superposition theorem, where Y YR YC YN

FIGURE 14.2

An example of system sensitivity analysis.

and TR TN TC YR FL 1L R YN L 1L N YC P 1L C 14:17a 14:17b 14:17c

and L GP (the loop transmission function), the following transfer functions and sensitivity functions (where  P) are obtained, respectively, as:
T SP R

FG 1L G 1L 1 1L

14:18a 14:18b 14:18c

T SP N

T SP C

Since sensitivity is a function of frequency, it is necessary to achieve a slope for Lm Lo( jo) that minimizes the effect on the system due to sensor noise. This is the most important case, since the minimum BW of Eq. (14.17b) tends to be greater than the BW of Eq. (14.17a), as illustrated in Fig. 14.3. Based on the magnitude characteristic of Lo for low- and highfrequency ranges, then: For the low-frequency range, where jL( jo)j>>1, from Eq. (14.18b),    1  T  SP N %  P jo     L jo T  SP N % G jo   P jo  14:19

For the high-frequency range, where jL( jo)j<<1, from Eq. (14.18b), 14:20

The BW characteristics of the open-loop function L( jo), with respect to sensitivity, are illustrated in Fig. 14.4. As seen from this figure, the low-frequency sensitivity given by Eq. (14.19) is satisfactory but the highfrequency sensitivity given by Eq. (14.20) is unsatisfactory since it can present a serious noise rejection problem. Based upon the analysis of Fig. 14.4, it is necessary to try to make the phase margin frequency of (the loop transmission BW), small enough in

FIGURE 14.3

Frequency response characteristics for the system of Fig. 14.2.

FIGURE 14.4

Bandwidth characteristics of Fig. 14.2.

order to minimize the sensor noise effect on the systems output. For most practical systems n ! w 2 and Z1 14:21 log ST do 0 P  T  SP N  < 1 ! log ST N < 0 P  T  SP N  > 1 ! log ST N > 0 P
0

14:22a 14:22b

Thus, the designer must try to locate the condition of Eq. (14.22b) in the highfrequency range where the system performance is not of concern; i.e., the noise effect on the output is negligible. The analysis for external disturbance effect (see Ref. 4) on the system output is identical to that for cross-coupling effects. For either case, low sensitivity is conducive to their rejection. 14.4 SENSITIVITY ANALYSIS [3,4] EXAMPLES

This section shows how the use of complete state-variable or output feedback minimizes the sensitivity of the output response to a parameter variation.This feature of state feedback is in contrast to conventional control-system design and is illustrated in the example of Sec. 13.9 for the case of system gain variation. In order to illustrate the advantage of state-variable feedback over the conventional method of achieving the desired system performance, a system sensitivity analysis is made for both designs. The basic plant Gx(s) of Fig.13.9a is used for this comparison. The conventional control system is shown in Fig. 14.5a. The control-system design can be implemented by the complete state-feedback configuration of Fig.14.5b or by the output feedback representation of Fig.14.5c,which uses Heq as a fixed feedback compensator.The latter configuration is shown to be best for minimizing the sensitivity of the output response with a parameter variation that occurs in the forward path, between the state xn and the output y. For both the state-feedback configuration and its output feedback equivalent, the coefficients ki are determined for the nominal values of the parameters and are assumed to be invariant.The analysis is based upon determining the value of the passband frequency ob at jMo( job)j Mo 0.707 for the nominal system values of the conventional and statevariable-feedback control systems. The system sensitivity function of Eq. (14.3), repeated here, is determined for each system of Fig. 14.5 and evaluated for the frequency range 0 o ob:        dM  M  14:23 S s M M  M d  M Mo  o o  o

FIGURE 14.5 Control: (a) unity feedback; (b) state feedback; (c) state feedback, H-equivalent. The nominal values are p1 1, p2 5 and A 10.

Note that M is the control ratio. The passband frequency is ob 0.642 for the system of Fig. 14.5a with nominal values, and ob 1.0 for the state-feedback system of Figs. 14.5b and 14.5c with nominal values. The control ratios for each of the three cases to be analyzed are Figure 14.5a: M s Figure 14.5b: M s Figure 14.5c: M s 10A s 3 2Ak3 p1 p2 s2 p1 p2 10Ak3 k2 s 10Ak1 14:26 10A s 3 2Ak3 p1 p2 s2 p1 p2 2A5k2 k3 p2 s 10Ak1 14:25 10A s 3 p1 p2 s 2 p1 p2 s 10A 14:24

FIGURE 14.6

Sensitivity due to pole and gain variation.

The system sensitivity function for any parameter variation is readily obtained for each of the three configurations of Fig. 14.5  usingthe respecby   tive control ratios of Eqs. (14.24) to (14.26). The plots of SM jo, SM jo, p1 p2  M  and SA jo vs. o, shown in Fig. 14.6, are drawn for the nominal values of p1 1, p2 5, and A 10 for the state-variable-feedback system. For the conventional system, the value of loop sensitivity equal to 2.1 corresponds to  0.7076, as shown on the root locus in Fig.13.10a.The same damping ratio is used in the state-variable-feedback system developed in the example in Sec.13.9. Table 14.2 summarizes the values obtained.

TABLE 14.2
Figure   M job    %M job 

System Sensitivity Analysis  M  Sp job  1  M  Sp job  2  M  SA job 

Control system

Conventional system design 14.5a 0.707 (ob 0.642) 1.09 1.28 1.29

1. Nominal plant (A 0.21, p1 1, p2 5): 2:1 Gs ss 1s 5 14.5a 0.338 52.4

2. P1 2:

2:1 Gs ss 2s 5 14.5a 0.658 6.9

3. P2 10:

2:1 Gs ss 1s 10 14.5a 1.230 74.0

4. A 0.42:

Gs

4:2 ss 1s 5

State-variable feedback system design [M is given in Eqs. (13.111), (14.25), and (14.26), respectively] 14.5b,c 0.036 0.707 (ob 1.0) 0.05 for X3 inaccessible 3.40 for X3 accessible 0.051

5. Nominal plant: 95:4s2 1:443s 1:0481 GsHeq s ss 1s 5

6. p1 2: 14.5b,c 0.683 3.39

GsHeq s

95:4s2 1:443s 1:0481 ss 2s 5 14.5b 0.160 77.4

7a. p2 10, X3 accessible: 95:4s2 6:443s 1:0481 GsHeq s ss 1s 10 14.5c 0.682 3.54

7b. p2 10, X3 inaccessible: 95:4s2 1:443s 1:0481 GsHeq s ss 1s 10 14.5b,c 0.717 1.41

8. A 20:

GsHeq s

190:8s2 1:443s 1:0481 ss 1s 5

TABLE 14.3

Time-Response Data %Mp t   Mp Mpo  Mpo 100

Control system 1 2 3 4 5 6 7a 7b 8

Mp(t) 1.048 1.000 1.000 1.189 1.044 1.035 1.000 1.046 1.046

tp, s 7.2 4.1 4.45 4.5 4.6 4.4

ts , s 9.8 16.4 15.0 10.15 6 6 24 6.5 5.6

4.8 4.8 13.45 0.86 4.4 0.181 0.192

The sensitivity function for each system, at o ob, is determined for a 100 percent variation of the plant poles p1 1, p2 5, and of the forward gain A, respectively. The term %jMj is defined as     M job  Mo job    %jM j 100 14:27 Mo job  where jMo( job)j is the value of the control ratio with nominal values of the plant parameters, and jM( job)j is the value with one parameter changed to the extreme value of its possible variation.The time-response data for systems 1 to 8 are given inTable 14.3 for a unit step input. Note that the response for the state-variable-feedback system (5 to 8, except 7a) is essentially unaffected by parameter variations. M The sensitivity function S represents a measure of the change to be expected in the system performance for a change in a system parameter. This measure is borne out by comparing Tables 14.2 and 14.3. That is, in the   frequency-domain plots of Fig. 14.6, a value SM jo 1.29 at o ob 0.642 A is indicated for the conventional system, compared with 0.051 for the statevariable-feedback system. Also,the percent change %jM( jo)j for a doubling of the forward gain is 74 percent for the conventional system and only 1.41 percent for the state-variable-feedback system. These are consistent with the changes in the peak overshoot in the time domain of %Mp(t) 13.45 and 0.192 percent, respectively, as shown in Table 14.3. Thus, the magnitude of %M( jo) and %Mp(t) are consistent with the relative magnitudes of  %SM job . Similar results are obtained for variations of the pole p1. A An interesting result occurs when the pole p2 5 is subject to variation. If the state X3 (see Fig.14.5b) is accessible and is fed back directly through k3,the   sensitivity SM jo is much larger than for the conventional system, having a p2

value of 3.40 at ob 1.0. However, if the equivalent feedback is obtained from X1 through the fixed transfer function [Heq(s)]0 p2 , Fig. 14.5c, the   sensitivity is considerably reduced to SM j1 0.050, compared with1.28 for p2 the conventional system. If the components of the feedback unit are selected so that the transfer function [Heq(s)] is invariant, then the time-response characteristic for output feedback through [Heq(s)] is essentially unchanged when p2 doubles in value (seeTable14.3) compared with the conventional system. In analyzing Fig. 14.6, the values of the sensitivities must be considered over the entire passband (0 o ob). In this passband the state-variablefeedback system has a much lower value of sensitivity than the conventional system. The performance of the state-variable-feedback systems is essentially unaffected by any variation in A, p1, or p2 (with restrictions) when the feedback coefficients remain fixed, provided that the forward gain satisfies the condition K ! K min (see Probs. 14.2 and 14.3). The low-sensitivity state-variablefeedback systems considered in this section satisfy the following conditions: the systems characteristic equation must have (1) b dominant roots (< n) which are located close to zeros of Heq(s) and (2) at least one nondominant root. One approach for determining the nominal value for the variable parameter  for a state-variable-feedback design is to determine k for the minimum,   midrange, and maximum value of . Then evaluate and plot SM jo for the  range of 0 o ob for each of these values of . The value of  that yields  the minimum area under the curve of SM jo vs. ois chosen as the value for o.  The analysis of this section reveals that for the state-variable-feedback system any parameter variation between the control U and the state variable Xn has minimal effect on system performance. In order to minimize the effect on system performance for a parameter variation that occurs between the states Xn and X1, the feedback signals should be moved to the right of the block containing the variable parameter (see Sec. 14.5), even if all the states are accessible. The output feedback system that incorporates an invariant Heq(s) has a response that is more invariant to parameter variations than a conventional system. 14.5 PARAMETER SENSITIVITY EXAMPLES

Example 1. The (Y/R)T , Eq. (13.118), for Design Example 1 of Chap. 13 contains no nondominant pole. As seen from the root locus of Fig. 13.16 for this example, the system is highly  sensitive to gain variations. Using  Eq. (13.117) in Eq. (14.3) with  A yields SM jo 1.84 at ob 1.41.  Example 2. The (Y/R)T of Design Example1is modified to contain one nondominant pole; see Eq. (13.121). Using Eq. (13.121), with p 50 and  KG,   in Eq. (14.3) yields SM job  0.064 at ob 1.41. Thus, the presence of a 

TABLE 14.4
Example 1 2 3

Time-Response and Gain-Sensitivity Data (Y/R)D 2s 2 s2 2s 2s 2 MP tp, s ts, s K1, s1 4.2 4.2 1.0 0.98 0.98  M  SK j!b  G ob 1.41 1.84 0.064 0.087

1.043 3.1 1.043 3.2

100s 2 s2 2s 2s 2s 50

100s 1:4s 2 1.008 3.92 2.94 s2 2s 2s 1s 2s 70

  nondominant pole drastically reduces SM job , which is indicative of the  reduction of the sensitivity magnitude for the frequency range 0 o ob. Example 3. The (Y/R)T of Design Example 2 is modified to improve the value of ts; see Eq. (13.127). Using Eq. (13.127), with p 70 and  KG, in   Eq. (14.3) yields SM job  0.087 at ob 1.41.  The data inTable 14.4 summarize the results of these three examples and show that the presence of at least one nondominant pole in (Y/R)T drastically reduces the systems sensitivity to gain variation while achieving the desired time-response characteristics. 14.6 INACCESSIBLE STATES [5]

In order to achieve the full benefit of a state-variable-feedback system, all the states must be accessible. Although, in general, this requirement may not be satisfied, the design procedures presented in this chapter are still valid. That is, on the basis that all states are accessible, the required value of the state-feedback gain vector k which achieves the desired Y(s)/R(s) is computed. Then, for states that are not accessible, the corresponding ki blocks are moved by block-diagram manipulation techniques to states that are accessible. As a result of these manipulations the full benefits of state-variable feedback may M not be achieved (low sensitivity S and a completely stable system), as described in Sec. 14.4. For the extreme case when the output y x1 is the only accessible state, the block-diagram manipulation to the G-equivalent (see Fig. 14.8c) reduces to the Guillemin-Truxal design. More sophisticated methods for reconstructing (estimating) the values of the missing states do exist (Luenberger observer theory). They permit the accessible and the reconstructed (inaccessible) states all to be fed back through the feedback coefficients, thus eliminating the need for block-diagram manipulations and maintaining the full benefits of the state-variable feedback-designed

You might also like