Open Access

New Dilated LMI Characterization for the Multiobjective Full-Order Dynamic Output Feedback Synthesis Problem

Journal of Inequalities and Applications20102010:608374

https://doi.org/10.1155/2010/608374

Received: 23 April 2010

Accepted: 17 September 2010

Published: 21 September 2010

Abstract

This paper introduces new dilated LMI conditions for continuous-time linear systems which not only characterize stability and performance specifications, but also, performance specifications. These new conditions offer, in addition to new analysis tools, synthesis procedures that have the advantages of keeping the controller parameters independent of the Lyapunov matrix and offering supplementary degrees of freedom. The impact of such advantages is great on the multiobjective full-order dynamic output feedback control problem as the obtained dilated LMI conditions always encompass the standard ones. It follows that much less conservatism is possible in comparison to the currently used standard LMI based synthesis procedures. A numerical simulation, based on an empirically abridged search procedure, is presented and shows the advantage of the proposed synthesis methods.

1. Introduction

The impact of linear matrix inequalities on the systems community has been so great that it dramatically changed forever the usually utilized tools for analyzing and synthesizing control systems.The standard LMI conditions benefited greatly from breakthrough advances in convex optimization theory and offered new solutions to many analysis and synthesis problems [13]. When necessary and sufficient LMI conditions are not possible, as it is the case for the static output control [4, 5], the multi-objective control [68], or the robust control [912] problems, sufficient conditions were provided, but were known to be overly conservative. Some dilated versions of LMI conditions have first appeared in the literature in [13], wherein some robust dilated LMI conditions were proposed for some class of matrices. Since then, a flurry of results has been proposed in both the continuous-time [6, 7, 10, 1417] and the discrete-time systems [5, 14, 1820]. These conditions offer, though, no particular advantages for monoobjective and precisely known systems, but were found to greatly reduce conservatism in the multi-objective [68, 19] and the robust control problems [9, 10, 1416, 18, 19]. In this respect, an interesting extension for the utilization of these dilated LMI conditions (as in, e.g., [2123]) provided solutions to the problem of robust root-clustering analysis in some nonconnected regions with respect to polytopic and norm-bounded uncertainties. Generally, the main feature of these LMI conditions, in their dilated versions, consists in the introduction of an instrumental variable giving a suitable structure, from the synthesis viewpoint, in which the controller parameterization is completely independent from the Lyapunov matrix. A particular difficulty though with these proposed dilated versions in the continuous-time case is the absence of dilated conditions as it is visible in [6, 15].

This paper introduces new dilated LMIs conditions for the design of full-order dynamic output feedback controllers in continuous-time linear systems, which not only characterize stability and performance specifications, but also, performance specifications as well. Similarly to the existing dilated versions, these new dilated LMI conditions carry the same properties wherein an instrumental variable is introduced giving a suitable structure in which the controller parameterization is completely independent from the Lyapunov matrix. In addition, scalar parameters are also introduced, within these dilated LMI, to provide a supplementary degree of freedom whose impact is to further reduce, in a significant way, the conservatism in sufficient standard LMI conditions. It is also shown, in this paper, that the obtained dilated LMI conditions always encompass the standard ones. As a result, the conservatism which results whenever the standard LMI conditions are used is expected to considerably reduce in many cases. This paper focuses on the multi-objective full-order dynamic output feedback controller design in continuous-time linear systems for which the constraining necessity of using a single Lyapunov matrix to test all the objectives and all the channels, which constitutes a major source of conservatism, is no longer a necessity as a different Lyapunov matrix is separately searched for every objective and for every channel. It is shown, in this paper, that despite constraining the instrumental variable, the new dilated LMI conditions are, at worst, as good as the standard ones, and, generally, much less conservative than the standard LMI conditions. The proposed solution is quite interesting, despite an inevitable increase in the number of decision variables in the involved LMIs and a multivariable search procedure that could be abridged through empirical observations. A numerical simulation is presented and shows the advantage of the proposed synthesis method.

2. Background

Consider the linear time-invariant continuous-time and input-free system
(2.1)

where the state vector , the perturbation vector , and the performance vector . All the matrices , , , and have appropriate dimensions. Let be the system transfer matrix from input to output . The following two lemmas are well known (see, e.g., [1, 3]) and provide necessary and sufficient conditions for System (2.1) to be asymptotically stable under an performance constraint and a performance constraint, respectively. These standard results will be extensively used in this paper.

Lemma 2.1.

System (2.1) with is asymptotically stable and if and only if there exist symmetric matrices and such that
(2.2)

Lemma 2.2.

System (2.1) is asymptotically stable and if and only if there exists a symmetric matrix in such that
(2.3)

3. Multiobjective Control Synthesis

Consider the continuous-time time-invariant linear system with input
(3.1)
where the state vector ,the perturbation vector , the input command vector , the performance vector , and the controlled output vector , and all the matrices , , , , , , , and have the appropriate dimensions. In the dynamic output feedback case, the control law is given by the state equations
(3.2)
As this controller is supposed to be of a full order , , , and . The closed-loop system is then described by the augmented state equations
(3.3)
where
(3.4)
The closed loop system transfer matrix from input to output then becomes
(3.5)
It is supposed that this system is of a multichannel type meaning that the perturbation vector is partitioned into a given number (say ) of block components,
(3.6)
and the performance vector is partitioned into a given number (say ) of block components,
(3.7)
It is supposed that some performance specifications are defined with respect to a particular channel (a path relating input component to output component ) or a combination of channels. It is also supposed that, for a given control law strategy, these performance specifications can always be expressed in terms of an and/or a transfer matrix norm of the corresponding channel, namely, , where the matrices and are set to select the desired input/output channel from the system closed-loop transfer matrix . In fact, is a -block row matrix of dimension such that only the th block is nonzero and is the identity matrix in . Similarly, is an -block column vector of dimension such that only the th block is nonzero and is the identity matrix in . The problem of the multi-objective controller synthesis is to construct a controller that stabilizes the closed loop system and, simultaneously, achieves all the prescribed specifications. It is easy to see that, for each channel , the closed loop transfer matrix is given by
(3.8)
On the channel basis, the closed-loop system is then described by
(3.9)
where
(3.10)

The dynamic output feedback synthesis multi-objective problem consists of looking for a dynamic controller that stabilizes the closed loop system and, in the same time, achieves the desired and/or performance specifications for every single system channel. More specifically, the dynamic output feedback synthesis multi-objective problem aims at making System (3.1) possess the following propriety.

Propriety P

System (3.1) is stabilizable by a dynamic output feedback law (3.2) such that, for every channel , either or both of the following two conditions hold:

with ;

.

Theorem 3.1 (the standard sufficient conditions).

If there exist symmetric matrices and , general matrices , , and and, for every channel ij, there exists a symmetric matrix such that either or both of the following two conditions are satisfied:

[StdH2]
(3.11)
[StdH∞]
(3.12)
then, Propriety holds, and furthermore, a set of the controller parameters defined in (3.2) is given by
(3.13)
where the nonsingular matrices and are obtained via the equation
(3.14)

Proof.

If either or both of conditions [StdH2] and [StdH∞] are satisfied, let and let be a nonsingular transformation matrix, with and selected from (3.14) (among infinitely many possibilities) via the singular value decomposition of . In view of (3.13) and (3.14), the following useful identities are easily derived:
(3.15)
As either or both of conditions [StdH2] and [StdH∞] are satisfied, by the congruence lemma applied to each LMI and in view of the identities listed just above, either or both of the following conditions are also satisfied, respectively,
  1. (i)
    (3.16)
     
  1. (ii)
    (3.17)
     
(3.18)

According to Lemmas 2.1 and 2.2, these are precisely the sufficient standard LMI conditions, expressed on a channel basis, for Propriety to hold.

Theorem 3.1 provides sufficient conditions for the existence of a single multi-objective dynamic output controller in terms of LMI conditions in which common Lyapunov matrices are enforced for convexity. This is known to produce, in general, overly conservative results. The following theorem attempts at reducing the effect of this limitation.

Theorem 3.2 (the dilated sufficient conditions).

If there exist general matrices , , , , , and and for every channel ij, for some scalars and , there exist symmetric matrices , , , , , general matrices and such that either or both of the following two conditions are satisfied:

(i)[DilH2]

(3.19)
(ii)[DilH∞]
(3.20)
Then, Propriety P holds, and furthermore, a set of the controller parameters defined in (3.2) is given by
(3.21)
where the nonsingular matrices and are obtained via the equation
(3.22)

Proof.

If either or both of conditions [DilH2] and [DilH∞] are satisfied, let and let be a nonsingular transformation matrix with and selected from (3.22) (among infinitely many possibilities) via the singular value decomposition of . In view of (3.21) and (3.22), the following useful identities are easily derived:
(3.23)
On the other hand, let us introduce
(3.24)
As either or both of conditions [DilH2] and [DilH∞] are satisfied, by the congruence Lemma applied to each LMI and in view of the identities listed just above, either or both of the following conditions are also satisfied, respectively.
  1. (i)
    (3.25)
     
  1. (ii)
    (3.26)
     
To summarize, we have proven that if either or both conditions [DilH2] and [DilH∞] are satisfied, then either or both of the following conditions are also satisfied:
  1. (i)
    (3.27)
     
  1. (ii)
    (3.28)
     
The third LMI of the first item condition is equivalent to
(3.29)
which, according to the elimination lemma [3], leads to
(3.30)

The two previous LMIs are equivalent to and , that is, for any , .

Similarly, the LMI of the second item condition is equivalent to
(3.31)
According to the Elimination lemma, this leads to
(3.32)
The previous two matrix inequalities are equivalent to
(3.33)
Via the Schur lemma, the latter inequality is equivalent to and
(3.34)

Clearly, as , there always exists a sufficiently large which satisfies this LMI. According to Lemmas 2.1 and 2.2, these are precisely the sufficient standard LMI conditions, expressed on a channel basis, for Propriety to hold.

Theorem 3.2 also provides sufficient conditions for the existence of a single multi-objective dynamic output controller in terms of LMI conditions in which the constraint of a common Lyapunov matrix is no longer needed. Convexity is rather insured by constraining the instrumental variables to be common. This is known to produce, in general, less conservative results than those obtained with the standard conditions of Theorem 3.1. Reducing further this conservatism is also possible through the positive scalar parameters and . A simple multidimensional search procedure can be carried out in the space of these parameters in order to obtain the values of these parameters for which LMI (3.19) and/or LMI (3.20) are feasible and produce the best multi-objective dynamic output controller with optimal performance levels. This multidimensional search procedure can, however, be overly expensive if the number of channel gets larger. A solution to this rather annoying limitation will be proposed in the next section. Yet, the important results of Theorem 3.2 constitute a significant contribution to the multi-objective control problem.

Next, the important question on whether or not the standard conditions could possibly be recovered by the dilated conditions will be addressed in the following section.

4. Recovery Condition

In the following theorem, it will be shown that our proposed dilated LMI conditions for the design of multiobjective full-order dynamic output feedback controllers do indeed encompass the standard conditions. This situation will be of great importance, as it will guarantee that conservatism will be almost always reduced. Similar results do exist in the literature in both the discrete-time [19] and the continuous-time case [6, 7]. The continuous-time results were, however, strictly concerned with the multi-channel synthesis problem and only in [7] that the recovery of the standard approach is proven. In view of this, the following theorem extends the discrete-time results to the continuous-time case. This point constitutes the major contribution of this paper.

Theorem 4.1.

For, the multi-objective dynamic output feedback synthesis problem, if the standard LMI conditions of Theorem 3.1 are satisfied and achieve, with a given controller, the upper bounds and , then the dilated inequality conditions of Theorem 3.2 are also satisfied with the same controller and with the upper bounds and .

Proof.

If the standard LMI conditions of Theorem 3.1 are satisfied for a given controller and achieve, for every channel, the upper bounds and , then there exist symmetric matrices and such that
(4.1)
and/or
(4.2)
Let us prove that these standard LMI conditions imply that the dilated inequality conditions of Theorem 3.2 are also satisfied with the same controller. When expressed in terms of the system closed-loop parameters, the right-hand sides of the dilated LMI conditions of Theorem 3.2 take the following form:
(4.3)
and/or
(4.4)
Let, in these matrices, , , , , and , these right-hand sides become
(4.5)
and/or
(4.6)
Let us prove, for these four matrices above, that the second matrix is positive definite while the third and/or the fourth matrices are both negative definite. Clearly, the standard conditions imply that
(4.7)
By virtue of the Schur complement lemma, the third matrix and/or the fourth matrix will be negative definite if and only if ,
(4.8)
and/or
(4.9)
As, from the standard and conditions,
(4.10)

there always exists an which achieves, simultaneously, these two conditions. As a result, the dilated inequality conditions of Theorem 3.2 are also satisfied. This proves that the dilated LMI multi-objective conditions always encompass the standard ones. Clearly, this means that the dilated-based approach yields upper bounds that are always and .

Theorem 4.1 has proven that the dilated LMI conditions of Theorem 3.2 do indeed encompass the standard ones of Theorem 3.1. The multidimensional search procedure carried out in the space of the scalars being exhaustive, up to a given discretization step that could be made as small as needed, does indeed cover every region, and in particular, the region where the standard conditions are recovered and which is defined by , where is greater than a minimum value defined by the two LMIs just in the proof above. In practice, the value of can be easily computed through a simple one dimensional line search procedure over these two LMIs.

On the other hand, at the light of the results of Theorem 3.2, a controller which achieves the best global performance level can be obtained through the minimization of the global objective function .Under this setting, it appears that optimality is always achieved very close to where all the and all the coincide. This purely empirical rule,observed with many examples we have tried, fits nicely to where the recovery of the standard conditions can be proved. In order to achieve optimality, it is then reasonable to abridge the costly multi-dimensional search procedure to a much cheaper one-dimensional search in the line for all channels. In this way, this proposed simple line search procedure not only provides a near optimal solution, but achieves the recovery condition which guarantees that this solution is, at least, as good as the one provided by the standard conditions.

5. An Example

Consider the LTI unstable third-order plant
(5.1)
The system is supposed to be consisting of two channels. Channel 1 connects the perturbation vector to the performance component , while Channel 2 connects the perturbation vector to the performance component . The objective here is to find a stabilizing full-order (i.e., a third order) dynamic output feedback controller which achieves simultaneously and optimally the performance specifications and , relatively to Channel 2 and Channel 1, respectively. Optimality is here defined as the minimization of , giving equal importance to the two channels. The use of the dilated LMI conditions can be carried out through a search procedure in the plane . Figure 1 is a three-dimensional plot which depicts the waveform of in that plane. This figure clearly shows that optimality is achieved close to the direction where . In this example, it is found that the minimum value of which guarantees recovery is . The abridged search procedure along the line produced a near optimal global performance of and when . Clearly, in this example,improvement is being made in the region below where recovery is not necessarily there. Table 1 lists the simulation results obtained with the sufficient standard LMI conditions of Theorem 3.1 and with the sufficient dilated LMI conditions of Theorem 3.2.
Table 1

Simulation results, with representing the LMI produced full-order dynamic output feedback controller.

Problem

Synthesis method

 

Standard/controller

Dilated/controller

Two-dimensional search procedure   with

  

  

One-dimensional search procedure with

  

 

Decision variable number = 30

Decision variable number = 87

Figure 1

3D-plot of the waveform in the plane   .

The advantage of using the dilated rather than the standard LMI conditions is quite visible with this example. Indeed, around a 30% improvement on and a 25% improvement on performance levels were possible. However, this improvement comes at the expense of almost tripling the number of decision variables involved in the proposed dilated LMI conditions (see Table 1).

6. Conclusion

This paper has presented new dilated LMI conditions for the design of multiobjective full-order dynamic output controllers in continuous-time systems that are able to cope not only with stability analysis and performance specifications, but also, with performance specifications as well. The paper developed new controller synthesis procedures which offer no particular advantage for precisely known monoobjective systems, but significantly reduce conservatism in the multi-objective control problem, as the main property of these new dilated LMI conditions, besides the fact thattheyallow a complete independence between the standard Lyapunov matrix and the controller parametersis that they alwaysencompass the standard ones.A numerical simulation is presented which supports these claims. The extension of these results to other control issues such as the robust controller, model predictive controller, and filter design problems is rather straightforward and yet very useful.

Authors’ Affiliations

(1)
Département de Génie Electrique, Ecole Supérieure des Sciences et Techniques de Tunis
(2)
Unité de Recherche SICISI, Ecole Supérieure des Sciences et Techniques de Tunis

References

  1. Boyd S, El Ghaoui L, Feron E, Balakrishnan V: Linear Matrix Inequalities in System and Control Theory, SIAM Studies in Applied Mathematics. Volume 15. SIAM, Philadelphia, Pa, USA; 1994:xii+193.View ArticleMATHGoogle Scholar
  2. Nesterov Y, Nemirovskii A: Interior-Point Polynomial Algorithms in Convex Programming, SIAM Studies in Applied Mathematics. Volume 13. SIAM, Philadelphia, Pa, USA; 1994:x+405.View ArticleMATHGoogle Scholar
  3. Skelton RE, Iwazaki T, Grigoriadis G: A Unified Approach to Linear Control Design, Taylor and Francis Series in Systems and Control. Taylor and Francis, London, UK; 1997.Google Scholar
  4. Crusius CAR, Trofino A: Sufficient LMI conditions for output feedback control problems. IEEE Transactions on Automatic Control 1999, 44(5):1053–1057. 10.1109/9.763227MathSciNetView ArticleMATHGoogle Scholar
  5. Lee KH, Lee JH, Kwon WH: Sufficient LMI conditions for output feedback stabilization of linear discrete-time systems. IEEE Transactions on Automatic Control 2006, 51(4):675–680. 10.1109/TAC.2006.872766MathSciNetView ArticleGoogle Scholar
  6. Apkarian P, Tuan HD, Bernussou J: Continuous-time analysis, eigenstructure assignment, and synthesis with enhanced linear matrix inequalities (LMI) characterizations. IEEE Transactions on Automatic Control 2001, 46(12):1941–1946. 10.1109/9.975496MathSciNetView ArticleMATHGoogle Scholar
  7. Ebihara Y, Hagiwara T: New dilated LMI characterizations for continuous-time multiobjective controller synthesis. Automatica 2004, 40(11):2003–2009. 10.1016/j.automatica.2004.06.009MathSciNetView ArticleMATHGoogle Scholar
  8. Scherer C, Gahinet P, Chilali M: Multiobjective output-feedback control via LMI optimization. IEEE Transactions on Automatic Control 1997, 42(7):896–911. 10.1109/9.599969MathSciNetView ArticleMATHGoogle Scholar
  9. Dettori M, Scherer CW: New robust stability and performance conditions based on parameter dependent multipliers. Proceedings of the 36th IEEE Conference on Decision and Control (CDC '00), 2000 5: 4187–4192.Google Scholar
  10. Ebihara Y, Hagiwara T: A dilated LMI approach to robust performance analysis of linear time-invariant uncertain systems. Automatica 2005, 41(11):1933–1941. 10.1016/j.automatica.2005.05.023MathSciNetView ArticleMATHGoogle Scholar
  11. Feron E, Apkarian P, Gahinet P: Analysis and synthesis of robust control systems via parameter-dependent Lyapunov functions. IEEE Transactions on Automatic Control 1996, 41(7):1041–1046. 10.1109/9.508913MathSciNetView ArticleMATHGoogle Scholar
  12. Gahinet P, Apkarian P, Chilali M: Affine parameter-dependent Lyapunov functions and real parametric uncertainty. IEEE Transactions on Automatic Control 1996, 41(3):436–442. 10.1109/9.486646MathSciNetView ArticleMATHGoogle Scholar
  13. Geromel JC, de Oliveira MC, Hsu L: LMI characterization of structural and robust stability. Linear Algebra and Its Applications 1998, 285(1–3):69–80.MathSciNetView ArticleMATHGoogle Scholar
  14. Duan Z, Zhang J, Zhang C, Mosca E: Robust and filtering for uncertain linear systems. Automatica 2006, 42(11):1919–1926. 10.1016/j.automatica.2006.06.004MathSciNetView ArticleMATHGoogle Scholar
  15. Peaucelle D, Arzelier D, Bachelier O, Bernussou J: A new robust D-stability condition for real convex polytopic uncertainty. Systems & Control Letters 2000, 40(1):21–30. 10.1016/S0167-6911(99)00119-XMathSciNetView ArticleMATHGoogle Scholar
  16. Shaked U: Improved LMI representations for the analysis and the design of continuous-time systems with polytopic type uncertainty. IEEE Transactions on Automatic Control 2001, 46(4):652–656. 10.1109/9.917671MathSciNetView ArticleMATHGoogle Scholar
  17. Xie W: An equivalent LMI representation of bounded real lemma for continuous-time systems. Journal of Inequalities and Applications 2008, 2008:-8.Google Scholar
  18. de Oliveira MC, Bernussou J, Geromel JC: A new discrete-time robust stability condition. Systems & Control Letters 1999, 37(4):261–265. 10.1016/S0167-6911(99)00035-3MathSciNetView ArticleMATHGoogle Scholar
  19. de Oliveira MC, Geromel JC, Bernussou J: Extended and norm characterizations and controller parametrizations for discrete-time systems. International Journal of Control 2002, 75(9):666–679. 10.1080/00207170210140212MathSciNetView ArticleMATHGoogle Scholar
  20. Farges C, Peaucelle D, Arzelier D, Daafouz J: Robust performance analysis and synthesis of linear polytopic discrete-time periodic systems via LMIs. Systems & Control Letters 2007, 56(2):159–166. 10.1016/j.sysconle.2006.08.006MathSciNetView ArticleMATHGoogle Scholar
  21. Arzelier D, Henrion D, Peaucelle D: Robust D-stabilization of a polytope of matrices. International Journal of Control 2002, 75(10):744–752. 10.1080/00207170210141824MathSciNetView ArticleMATHGoogle Scholar
  22. Bachelier O, Peaucelle D, Arzelier D, Bernussou J: A precise robust matrix root-clustering analysis with respect to polytopic uncertainty. Proceedings of the American Control Conference, July 2000, Chicago, Ill, USA 5: 3331–3335.Google Scholar
  23. Bosche J, Bachelier O, Mehdi D: An approach for robust matrix root-clustering analysis in a union of regions. IMA Journal of Mathematical Control and Information 2005, 22(3):227–239. 10.1093/imamci/dni007MathSciNetView ArticleMATHGoogle Scholar

Copyright

© J. Zrida and K. Dabboussi. 2010

This article is published under license to BioMed Central Ltd. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.