You are on page 1of 4

FREQUENCY DOMAIN POLYNOMIAL ERROR NORM MODEL

ORDER REDUCTION TECHNIQUE: AN IMPROVEMENT

J.M. Araujo† , A.C. Castro† and E.T.F. Santos†

†Federal Institute in Education, Science and Technology of Bahia, Research Group on Signals and Systems
Rua Emdio dos Santos, S/N, Barbalho, Salvador - BA, Brazil
jomario@cefetba.br,castro@cefetba.br,eduardot@cefetba.br

Abstract— Model order reduction (MOR) is improvement of this method that does not require this
a highlight theme in systems and control the- restriction, solving the quadratic optimisation problem
ory. In this work, it is introduced a proposal in an unconstrained way, which means a great advan-
to overcome a drawback of the MOR technique tage for numerical purposes.
based on the minimization of the polynomial 2. The original method
numerator coefficients norm in error transfer The method proposed in [8] consists on, given an orig-
function, originally proposed for stable, mini- inal stable and minimum phase transfer function:
mum phase systems with the constraint that its
denominator must be Hurwitz. This restriction p(s) b1 sm + b2 sm−1 + . . . + bm−1 s + bm
leads to non-convexity for almost every case, G(s) = = (1)
q(s) sn + a1 sn−1 + . . . + an−1 s + an
but it seems to be needed in order to assure
the stability in the reduced model. It often It is obtained a reduced order transfer function:
degrades the method results due to the pos-
sible existence of a local minimum attractor.
pb(s) bb0 sl + bb1 sl−1 + . . . + bbl−1 s + bbl
The proposed method overcomes this restric- b
G(s) = = r (2)
tion and unstable poles and zeros swapping is qb(s) s +b a1 sr−1 + . . . + b ar−1 s + bar
given, preserving the magnitude response for
In which r < n, by coefficients norm minimization for
non-stable polynomials in the reduced model.
the polynomial numerator in the error function:
Successful tests were performed with the pro-
posed method, showing its effectiveness.
∧ ∧
Keywords— Order reduction, polynomial ∧ q (s)p(s) − p(s)q(s)
e(s) = G(s) − G(s) = (3)
norms, stability. ∧
q (s)q(s)

1. Introduction If unconstrained minimization problem is considered,


Model order reduction (MOR) is a well-known ap- the same is convex, and numerically robust optimiza-
proach applied to control theory in which, by some tion methods can be applied. However, it does not
criterion, the order n of an original model is reduced guarantee the stability of the final solution. Hence,

to r, where r < n. There is a wide range of MOR a constraint that q (s) must be Hurwitz is necessary.
techniques, and for linear models, are based on input- Notice that it is very difficult to establish conditions
output transfer functions and transfer matrices [1-8], on the original model that guarantees stability in re-
response-matching [9,10] and state-space descriptions duced model. Hence more, it is a very limiting issue
[11-13]. In [8], a very simple but efficient method for on the possible applications of this technique. In or-
order reduction of a SISO transfer function was intro- der to illustrate the pros and cons, the application of
duced. Unfortunately, the constraint that the polyno- this technique will be shown for two examples, one of
mial denominator must be Hurwitz to guarantee the 9th order and another of 8th order, G1 (s) and G2 (s)
stability of the reduced model is often strongly non- respectively from [10] and [2], both in the bottom of
convex. In this paper, it is introduced a procedure for this page.

s4 + 35s3 + 291s2 + 1093s + 1700


G1 (s) =
s9 + 9s8 + 66s7 + 294s6 + 1029s5 + 2541s4 + 4684s3 + 5856s2 + 4620s + 1700
35s7 + 1086s6 + 13285s5 + 82402s4 + 278376s3 + 511812s2 + 482964s + 194480
G2 (s) =
s8 + 33s7 + 437s6 + 3017s5 + 11870s4 + 27470s3 + 37942s2 + 28880s + 9600
Figure 2: Frequency responses of G2 (s) and the re-
Figure 1: Step responses of G1 (s) and the reduced duced models
models

Application of the technique in G1 (s) for 3rd order 2. If the reduced model is stable, then the solution
reduced model with two zeros leads to: was reached;
3. If not, it is necessary to swap the unstable poles,
0.1859s2 − 1.2533s + 2.686 zeros for the left half plane plus change nega-
Ĝ1a (s) = 3
s + 2.7923s2 + 4.3206s + 2.686 tive signals of gains, preserving this way the fre-
quency response magnitude.
The best result obtained in [6] for step response
match is: This procedure can be justified exploring the fact
1.286s2 − 4.667s + 7.7341 of magnitude response invariance of reduced model by
Ĝ1b (s) = 3 application of step three. Let be the reduced model
s + 5.9373s2 + 11.7414s + 7.7341
on pole-zero gain form:
Step responses for these two reduced models are Q
shown in Fig. 1. A stable, non-minimum phase model b (s − ηj )
G(s) = K Q (4)
is obtained for both cases. The unconstrained method (s − λi )
result is quite good when compared with the method The magnitude response of the reduced model is pre-
shown in [10]. Now, for the second example, the un- served by swapping the unstable poles/zeros around
constrained method application on G2 (s) and for a the imaginary axe or by change of gain signal, and, al-
5th order leads to the unstable, non-minimum phase though a non-optimal point for the minimization prob-
reduced model: lem is matched, the stability is guaranteed.
32.7602s4 + 27.3270s3 − 8.4655s2 Considering the steps above, one obtains the follow-
−8.6024s + 18.4519 ing reduced model for the second example:
Ĝ2b (s) = 5
s + 2.1065s + 1.2730s3 − 0.6774s2
4
32.7602s4 + 94.29s3 + 116.0217s2
+0.0567s + 0.9103 +68.5446s + 18.6591
Ĝ2b (s) = 5
3. The improvement s + 4.11498s4 + 7.6717s3 + 7.4745s2
+3.8678s + 0.9208
As shown above, the unconstrained method may lead
to unstable reduced transfer functions. The original Applying the Routh technique described in [2], the
proposition requires the constraint that qb(s) must be following reduced model is obtained:
Hurwitz. It introduces, in several cases, highly non-
linear inequalities that turns the optimization problem 55645.5s4 + 173419.1s3 + 463107.8s2
to be non-convex [8]. Here, the following is proposed +439546.0s + 194480
Ĝ2c (s) =
in order to avoid this drawback: 1963s + 6817.26s4 + 23973.4s3 + 31694s2
5

+27963.3s + 9600
1. Unconstrained minimization of the polynomial
numerator coefficients norm in error transfer The frequency response and the step response for the
function ; original and reduced models, by proposed method and
[2] V. KRISHNAMURTHY, V. SESHADRI, Model
Reduction Using the Routh Stability Criterion, IEEE
Transaction on Automatic Control, 1978, Vol. AC-2-3,
No. 4, pp. 729-731.
[3] T.N. LUCAS, Scaled Impulse Energy Approxima-
tion for Model Reduction, IEEE Transactions on Au-
tomatic Control, 1988, Vol. 33, No. 8, pp. 791-793.
[4] C.-S. HSIEH, C. HWANG, Model reduction of
continuous-time systems using a modified Routh ap-
proximation method, IEE Proceedings D, Vol. 136,
No. 4, 1989, pp. 151-156.
[5] L.A. AGUIRRE, Design of controllers by means of
model reduction techniques, Electronics Letters, Vol.
29, No. 4, 1993, pp. 389-390.
[6] T.N. LUCAS, A.R. MUNRO, Model reduction by
generalised least-squares method, Electronics Letters,
Vol. 27, No. 15, 1991, pp. 1383-1384.
Figure 3: Step responses of G2 (s) and the reduced [7] R. LULUS, Comparison of Approaches for Model
models Reduction, In Procedings of 9th IASTED Conference
on Intelligent Systems and Control, 2006.
by Routh method [2] are plotted respectively in figure [8] J.M. ARAUJO, A.C. CASTRO, E.T.F. SANTOS,
2 and figure 3, and again, the results of the proposed Reduo de ordem no domnio da freqncia baseada na
unconstrained improved method are quite better than minimizao da norma dos coeficientes polinomiais do
the original Routh method. erro, SBA Controle e Automao, 2008, Vol. 19, issue 3,
Let be another example from [14]. The original pp. 1-5.
transfer function is: [9] R.A. EL-ATTAR and M. VIDYASAGAR, Order re-
duction by l1 and l∞ norm minimisation, IEEE Trans-
s5 + 1014s4 + 14069s3 + 69140s2
action on Automatic Control, 1978, Vol. 23, No. 04,
+140100s + 100000
G(s) = 6 pp. 731-734.
s + 222s5 + 14541s4 + 248420s3
+2220000s + 1000000 [10] S. MUKHERJEE, SATAKSHI, R.C. MITTAL,
Model order reduction using response-matching tech-
Tab. 1 gives the comparison between the proposal im- nique, Journal of the Franklin Institute, 2005, Vol.
proved technique and the methods of Routh and the 342, pp. 503–519.
described in [14], by using relative integral of squared [11] B.C. MOORE, Principal component analysis in
error (ISE) metric. It is clear the best quality on re- linear systems: ?controllability, observability and
sults obtained by the proposed technique. model reduction, IEEE Transaction on Automatic
Control, 1981, Vol. 26, No. 01, pp. 17-32.
4. Conclusion
[12] L. PERNEBO, L. SILVERMAN, Model reduction
An improvement in the method for model reduction
via balanced state-space representations, IEEE Trans-
considering the polynomial norm coefficients mini-
action on Automatic Control, 1982, Vol. 27, n. No, pp.
mization of error transfer function numerator was pro-
382-387.
posed. The good results obtained show the effec-
tiveness of the proposed constrained method, which [13] G. MUSCATO, Parametric generalized singular
means that there is no need to consider complicated perturbation approximation for model order reduc-
constraints. It has a great advantage from numerical tion. IEEE Transactions On Automatic Control, Vol.
point view, and it helps to consolidate the originally 45, n. 02, 2000, pp. 339-343.
proposed methodology. The proposed improvement [14] T.N. LUCAS, Model reduction by condensed
assures its applicability. The limitation of method re- continued-fraction method, Electronics Letters, Vol.
mains the same of the original, that is, only minimum 21, No. 16, 1985, pp. 680-681.
phase, stable systems are considered.
REFERENCES

[1] T.C. HSIA, On the simplification of linear systems,


IEEE Transaction on Automatic Control, 1972, Vol.
AC-17, No. 04, pp 372-374.
Tab. 1: Comparison between proposed methodology against condensed continued-fractions and Routh
technique.

You might also like