Difference between revisions of "Robust Performance"
(→Chapter Summary) 

Line 28:  Line 28:  
== Chapter Summary ==  == Chapter Summary ==  
+  This chapter focuses on the analysis of ''robustness'' of feedback systems:  
+  <ol>  
+  <li><p>  
+  Uncertainty can enter a model in many forms. ''Parametric uncertainty'' occurs when the values of the parameters in the model are not precisely known or may vary. ''Unmodeled dynamics'' are a more general class of uncertainty in which some portions of the systems behavior are not included in the model, either due to lack of knowledge or simplicity. Unmodeled dynamics can be taken into consideration by incorporating an uncertainty block with bounded input/output response. Common types of unmodeled dynamics include ''additive uncertainty'', ''multiplicative uncertainty'' and ''feedback uncertainty''.  
+  </p></li>  
+  
+  <li><p>  
+  The ''Vinnicombe metric'' (or <math>\nu</math>gap metric) provides a measure of the distance between two transfer functions. It is defined as  
+  <center><amsmath>  
+  \delta_\nu(P_1, P_2) = \begin{cases} d(P_1,P_2), &{\rm if\ }  
+  (P_1,P_2)\in\mathcal C\\1, &{\rm otherwise,}\end{cases}  
+  </amsmath></center>  
+  where <math>d(P_1, P_2)</math> is a distance measure between the two transfer function  
+  <center><amsmath>  
+  d(P_1, P_2) = \sup_{\omega} \dfrac{P_1(i\omega)   
+  P_2(i\omega)}{\sqrt{(1 + P_1(i\omega)^2)(1 + P_2(i\omega)^2)}},  
+  </amsmath></center>  
+  and <amsmath>\mathcal C</amsmath> is the set of all pairs <math>(P_1, P_2)</math> such that the functions <math>f_1=1+P_1(s)P_1(s)</math> and <math>f_2=1+P_2(s)P_1(s)</math> have the same number of zeros in the right halfplane  
+  </p></li>  
+  
+  <li><p>  
+  Robust stability can be determined through the use of the Nyquist plot. The ''stability margin'' <math>s_m</math>, defined as the shortest distanced from 1 to the Nyquist curve, provides a measure of robustness. For an additive perturbation <math>\Delta(s)</math>, the system is robustly stable if  
+  <center><amsmath>  
+  \Delta<\Big\frac{1+PC}{C}\Big  
+  \qquad\text{or}\qquad  
+  \delta=\Big \frac{\Delta}{P} \Big < \frac{1}{T}.  
+  </amsmath></center>  
+  </p></li>  
+  This condition can be derived using the ''small gain theorem'' and allows us to reason about uncertainty without exact knowledge of the process perturbations.  
+  <li><p>  
+  </p></li>  
+  
+  <li><p>  
+  The ''Youla parameterization'' provides a description of all controllers that can stabilize a given process. For a stable system, stabilizing compensators are of the form  
+  <center><amsmath>  
+  C=\frac{Q}{1PQ},  
+  </amsmath></center>  
+  where <math>P</math> is the process transfer function and <math>Q</math> is an arbitrary stable transfer function. For an unstable process, we write the process dynamics as  
+  <center><amsmath>  
+  P(s) = \frac{a(s)}{b(s)} = \frac{A(s)}{B(s)},  
+  </amsmath></center>  
+  where <math>A(s)</math> and <math>B(s)</math> are stable transfer functions. Stabilizing compensators have the form  
+  <center><amsmath>  
+  C=\frac{G_0+QA}{F_0QB}.  
+  </amsmath></center>  
+  where <math>F_0(s)</math>, <math>G_0(s)</math> and <math>Q(s)</math> are all stable transfer functions.  
+  </p></li>  
+  
+  <li><p>  
+  In addition to stability, uncertainty can also affect the performance of a system. For additive uncertainty, the load response satisfies  
+  <center><amsmath>  
+  \frac{dG_{yd}}{G_{yd}}=S\frac{dP}{P}.  
+  </amsmath></center>  
+  The response to load disturbances is thus insensitive to process variations for frequencies where the magnitude of the sensitivity function <math>S(i\omega)</math> is small. Similarly, the response of the controller to noise in the presence of additive uncertainty satisfies  
+  <center><amsmath>  
+  \frac{dG_{un}}{G_{un}}=T\frac{dP}{P},  
+  </amsmath></center>  
+  indicating that the controller is insensitive to noise when the complementary sensitivity is small. Control design in the presence of uncertainty can be done by using the Gang of Four to insure that the appropriate sensitivity functions are all well behaved.  
+  </p></li>  
+  
+  <li><p>  
+  State space design, using eigenvalue (or pole) placement, can also be analyzed using sensitivity functions. State space design is complicated in the presence of right half plane poles and zeros, which limit the achievable performance. Design rules for robust pole placement include:  
+  * Slow stable process zeros should be matched by slow closed loop poles  
+  * Fast stable process poles should be matched by fast closed loop poles  
+  * Slow unstable process zeros and fast unstable process poles impose performance limitations  
+  </p></li>  
+  
+  <li><p>  
+  Design methods for robust stability performance include ''quantitative feedback theory'' (QFT), ''linear quadratic control'' (LQG) and and <amsmath>H_\infty</amsmath> control. Strong robustness results are available for <amsmath>H_\infty</amsmath>control. Letting <math>z=H(P(s),C(s))w</math> represent the input/output response between a set of generalized disturbances <math>w</math> and the generalized error <math>z</math>, it can be shown that  
+  <center><amsmath>  
+  \H(P,C)\_\infty=\frac{1}{\delta_\nu(P,1/C)}.  
+  </amsmath></center>  
+  The Vinnecombe metric can thus be considered as a generalized stability margin (comparable to <math>s_m</math>) and it follows that if we find a controller <math>C</math> with <amsmath>\H(P,C)\_\infty<\gamma</amsmath>, then this controller will stabilize any process <math>P_*</math> such that <math>\delta_\nu(P,P_*)<1/\gamma</math>.  
+  </p></li>  
+  </ol>  
== Exercises ==  == Exercises == 
Revision as of 16:22, 18 May 2008
Prev: Frequency Domain Design  Chapter 12  Robust Performance  Next: Bibliography 
This chapter focuses on the analysis of robustness of feedback systems. We consider the stability and performance of systems who process dynamics are uncertain and derive fundamental limits for robust stability and performance. We also discuss how to design controllers to achieve robust performance.
Textbook ContentsRobust Performance (pdf, 30Jan08)

Lecture Materials
Supplemental Information

Chapter Summary
This chapter focuses on the analysis of robustness of feedback systems:
Uncertainty can enter a model in many forms. Parametric uncertainty occurs when the values of the parameters in the model are not precisely known or may vary. Unmodeled dynamics are a more general class of uncertainty in which some portions of the systems behavior are not included in the model, either due to lack of knowledge or simplicity. Unmodeled dynamics can be taken into consideration by incorporating an uncertainty block with bounded input/output response. Common types of unmodeled dynamics include additive uncertainty, multiplicative uncertainty and feedback uncertainty.
The Vinnicombe metric (or gap metric) provides a measure of the distance between two transfer functions. It is defined as
where is a distance measure between the two transfer function
and is the set of all pairs such that the functions and have the same number of zeros in the right halfplane
Robust stability can be determined through the use of the Nyquist plot. The stability margin , defined as the shortest distanced from 1 to the Nyquist curve, provides a measure of robustness. For an additive perturbation , the system is robustly stable if
The Youla parameterization provides a description of all controllers that can stabilize a given process. For a stable system, stabilizing compensators are of the form
where is the process transfer function and is an arbitrary stable transfer function. For an unstable process, we write the process dynamics as
where and are stable transfer functions. Stabilizing compensators have the form
where , and are all stable transfer functions.
In addition to stability, uncertainty can also affect the performance of a system. For additive uncertainty, the load response satisfies
The response to load disturbances is thus insensitive to process variations for frequencies where the magnitude of the sensitivity function is small. Similarly, the response of the controller to noise in the presence of additive uncertainty satisfies
indicating that the controller is insensitive to noise when the complementary sensitivity is small. Control design in the presence of uncertainty can be done by using the Gang of Four to insure that the appropriate sensitivity functions are all well behaved.
State space design, using eigenvalue (or pole) placement, can also be analyzed using sensitivity functions. State space design is complicated in the presence of right half plane poles and zeros, which limit the achievable performance. Design rules for robust pole placement include:
 Slow stable process zeros should be matched by slow closed loop poles
 Fast stable process poles should be matched by fast closed loop poles
 Slow unstable process zeros and fast unstable process poles impose performance limitations
Design methods for robust stability performance include quantitative feedback theory (QFT), linear quadratic control (LQG) and and control. Strong robustness results are available for control. Letting represent the input/output response between a set of generalized disturbances and the generalized error , it can be shown that
The Vinnecombe metric can thus be considered as a generalized stability margin (comparable to ) and it follows that if we find a controller with , then this controller will stabilize any process such that .
This condition can be derived using the small gain theorem and allows us to reason about uncertainty without exact knowledge of the process perturbations.
Exercises
Frequently Asked Questions
Errata
 Errata: Extraneous text "!design" in Chapter 12
 Errata: Maximum sensitivity occurs at frequency omega ms not omega sc
 Errata: Caption for Figure 12.4 doesn't quite match figure
 Errata: d(P 1, P 2) is the longest chordal distance, not shortest
 Errata: Riemann sphere has radius (not diameter) 1
 Errata: In Example 12.6, n = 1 should be n = 1
 Errata: For Youla parameterization, P = b(s)/a(s) = B(s)/A(s) and C0 = G0(s)/F0(s)
 Errata: In Figure 12.8, the signs of A and B are reversed
 Errata: In equation (12.13) and the displayed equation above it, T should be T
 Errata: In Example 12.8, controller zero is at s = 3.5
 Errata: In Example 12.8, process zero is at s = 2
 Errata: In Example 12.9, poles are at a, p1, p2
 Errata: In equation (12.16), a is not defined and a factor of a is missing in numerator the last term
 Errata: After equation (12.23), generalized error is z, not w