experimental study of a random linear system

3
578 IEEE TRANSACTIONS ON AUTOXATIC CO-VTROL October Fig. 1-.I nonlinearcharacteristicwithathreshold, a linear. and a saturating region. represented by a product of two Gaussian-type functions e-Bs and (1 -CCDSZJ and a linear function Ax. Fig. ?-.A nonlinear characteristic xirh a linear and a saturating region represented by a product of a Gaussian iunction and a linear iunction Ax. Fig. 3--A nonlinearcharacteristicwithathreshold and a linear region represented by a product oi a function Ax. Gaussian-type function (1 -CcDS2j and a linear approximation to the nonlinear function y=f(r) leads to a simple solution with a sinusoidal input signal. In the following it will be shown that the same approximation leads also to a simple answer for stochastic noise input. A nonlinear characteristic with a thresh- old, linear and saturation region can be ap- proximated by the product of a linear and two Gaussian type functions (see Fig. 1). y = .~l~e-B“(1 - Ce-Dz’). (6) For C=O, this characteristic approximates a linear and saturationregion (see Fig. 2) and for B =O it approximates a threshold and a linear region (see Fig. 3). \\-hen (6j is inserted in (S), we find that For the case of a linenr and saturating re- gion (C=O), For the case of a threshold and linear region (B=O), The function coefficients A , B, C, and D can be evaluated graphically from a given characteristic y=,f(x), by trial and error, by successive approximation or with a com- puter. Simple analytical methods of calcu- lating these coefficients have been outlined in the previous communication on the sub- ject.l D. h,I. h‘k~ow Applied Physics Division R. A. Hum Radio and Electric Eng. Div. rational Research Council Ottawa, Ontario, Canada Experimental Study of a Random Linear System general theorem on the stability of linear systems with Gaussian random parameters has been published recently by N. J. Bershad’ (see -1ppendix). This theorem establishes sufficient and necessary condi- tions to insure that the impulse response of a system converges to zero, in the mean-square sense, after a sufficiently long time. From a practical point of view, however, one is more interestedindeterminingwhether the im- pulse response tends to converge in a rela- tively short time. This paper presents the re- sults of an analog simulation to determine the range of parameters for m-hich this “short-timestabilit>-”existsandcompares these regions with the results obtained from the theorem. A second-order linear system of the form :i: + 2rwt + w?[1 + a(t)]x = 0, rw > 0 (1) was simulated on a P-ACE TR-10 analog computer, as shown in Fig. 1. The noise source was Automation Laboratories’ ivIodel lOOB Gaussian Noise Generator, which has a “flat’ spectrum from 0 to 500 cps. The out- put of the low-pass filter has the exponential autocorrelation function The computer was run in repetitive opera- tion mode; M and 5 were fed into an oscillo- scope to display the phase plane trajectory of the system. -1 typicaldisplayisshownin Fig. 2. The system was deemed unstable if an>-impulse responses exceeded thelinear range of theoperational amplifiers of the computer. The results of the simulation are plotted in Fig. 3. The txo more or less straight lines which divide the plane into three regions are Manuscript received Ma? 8. 1964; revised July 6. 1964. Based on a thesis submitted by W. G. Tuel. Jr. to the iacultp oi Rensdaer PolytechnicInstitute in partial fulfillment oi the requirements for th; was sponsored jointly by Grant KO. AF-AFOSR M.E.E. degree. The research reported in this paper 2i9-63 irom the Air Force Office oi Scientific Re- search. and by Grant So. AF 19(628)-3859 from the .Air Force Cambridge Research Laboratory. Office of .Aerosoace Research. USAF.

Upload: p

Post on 01-Mar-2017

212 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Experimental study of a random linear system

578 I E E E T R A N S A C T I O N S ON A U T O X A T I C CO-VTROL October

Fig. 1-.I nonlinear characteristic with a threshold, a linear. and a saturating region. represented by a product of two Gaussian-type functions e - B s and (1 -CCDSZJ and a linear function Ax.

Fig. ?-.A nonlinear characteristic x i r h a linear and a saturating region represented by a product of a Gaussian iunction and a linear iunction A x .

Fig. 3--A nonlinear characteristic with a threshold and a linear region represented by a product oi a

function Ax. Gaussian-type function ( 1 -CcDS2j and a linear

approximation to the nonlinear function y = f ( r ) leads t o a simple solution with a sinusoidal input signal. I n the following it will be shown that the same approximation leads also to a simple answer for stochastic noise input.

A nonlinear characteristic with a thresh- old, linear and saturation region can be ap- proximated by the product of a linear and two Gaussian type functions (see Fig. 1).

y = . ~ l ~ e - B “ ( 1 - Ce-Dz’). (6)

For C=O, this characteristic approximates a linear and saturation region (see Fig. 2 ) and for B = O it approximates a threshold and a linear region (see Fig. 3) .

\\-hen ( 6 j is inserted in (S), we find that

For the case of a linenr and saturating re- gion (C=O),

For the case of a threshold and linear region (B=O),

The function coefficients A , B, C, and D can be evaluated graphically from a given characteristic y=,f(x), by trial and error, by successive approximation or with a com- puter. Simple analytical methods of calcu- lating these coefficients have been outlined in the previous communication on the sub- ject.l

D. h,I. h ‘ k ~ o w Applied Physics Division

R. A. Hum Radio and Electric Eng. Div.

ra t ional Research Council Ottawa, Ontario, Canada

Experimental Study of a Random Linear System

general theorem on the stability of linear systems with Gaussian random parameters has been published recently by N. J. Bershad’ (see -1ppendix). This theorem establishes sufficient and necessary condi- tions to insure that the impulse response of a system converges to zero, in the mean-square sense, after a sufficiently long time. From a practical point of view, however, one is more interested in determining whether the im- pulse response tends to converge in a rela- tively short time. This paper presents the re- sults of an analog simulation to determine the range of parameters for m-hich this “short-time stabilit>-” exists and compares these regions with the results obtained from the theorem.

A second-order linear system of the form

:i: + 2 r w t + w?[1 + a ( t ) ] x = 0, rw > 0 (1) was simulated on a P-ACE TR-10 analog computer, as shown in Fig. 1. The noise source was Automation Laboratories’ ivIodel lOOB Gaussian Noise Generator, which has a “flat’ spectrum from 0 to 500 cps. The out- put of the low-pass filter has the exponential autocorrelation function

The computer was run in repetitive opera- tion mode; M and 5 were fed into an oscillo- scope to display the phase plane trajectory of the system. -1 typical display is shown in Fig. 2. The system was deemed unstable if an>- impulse responses exceeded the linear range of the operational amplifiers of the computer.

The results of the simulation are plotted in Fig. 3. T h e t x o more or less straight lines which divide the plane into three regions are

Manuscript received Ma? 8. 1964; revised July 6. 1964. Based on a thesis submitted by W. G. Tuel. Jr. to the iacultp oi R e n s d a e r Polytechnic Institute in partial fulfillment oi the requirements for th; was sponsored jointly by Grant KO. AF-AFOSR M.E.E. degree. The research reported in this paper

2i9-63 irom the Air Force Office oi Scientific Re- search. and by Grant S o . AF 19(628)-3859 from the .Air Force Cambridge Research Laboratory. Office of .Aerosoace Research. USAF.

Page 2: Experimental study of a random linear system

I964 Correspondence 579

obtained from Theorems 1 and L 2 The region labeled “stable” is found from Theorem 1; the one labeled “unstable” is found from Theorem 2. The intermediate region is of undetermined character. I t is seen that the system is stable for a larger range of parameters than predicted by the theorem, especially for gain variations that are nar- row band with respect to the system. This is discussed below.

Fig. I-Simulation of random second-order system.

I

Fig. 2-Typical phase plane trajectories. * = I , r = O . S , G( s ) =l / ( s+l ) .

If the gain variation is narrow band with respect to the deterministic part of the system, ie., if the parameter variation is slow, then the random process may be ap- proximated by a random variable a with the same probability distribution. The system (1) then becomes a time-invariant system which is stable if a> -1. For a true Gaussian distribution, a < - 1 with some nonzero probability; however, all physical noise sources are amplitude limited and non-Gaussian to some extent. Hence, in a practical case, one can say that (1) is stable if prob [a < - 1 ] <E, \\-here e is some small number.

For a particular noise source used, the value ~ = 0 . 0 1 represents the amplitude limi- tation well. Then (1) is stable if prob [a< -11 <0.01. In terms of the Gaussian distribution

2 3 9 4 5

(a) (C)

Fig. 3-(a) Experimental results: x-unstab!e. &-stable; 0=1000; G(s) =lOO/(s+100). (b) Experimental results: x-unstable, &stable; = loo; G(s) =lO/(s+lO). ( c ) Esperimental results: x-unstable, 0-stable; w=lOOO; G(s) =lO/(s+lO).

16 SO0

12

8

4

0 I 2 3 4 3 . 5

I6 H w SO0

12

!stable I ,

I 2 3 4 3 . 5

(a) b)

Fig. 4-(a) Experimental results, clipped noise: x-unstable. &stable: w =1000; G(s) =lOO/(s+lOO). (b) Experimental results. clipped noise: s-unstable. 0-stable; 0=1000; G(s) =lOO/(s+lOO).

this implies that Z<O.lM. Since u* is the autocorrelation value b(O), if

oso = 2uZ (:) < 0.37 (t), the system should be stable. This condition is plotted as the line labeled “1 per cent probability” in Fig. 3. I t is seen that this provides a good explanation for the larger stability regions.

To further justify this conclusion, a di- ode limiter was inserted between the lox- pass filter and the system. The results of experiments with this configuration are

from the input So and the degree of limit- i r ~ g . ~ T h e results confirm the conclusion that the larger stability regions are due to amplitude limiting of the random process.

In summary, the experimentally deter- mined parameter ranges for which the im- pulse response of a second-order system con- verges in a short time generally agree with the values obtained by application of Rershad’s theorem. The difference between the experiment and the theorem, which is a larger stable region than predicted for nar- row-band parameter variations, is due to the non-Gaussian nature of any physical noise source. This explanation is confirmed by experiments conducted with artificially limited noise.

Page 3: Experimental study of a random linear system

580 IEEE T R A X S A C T I O X S O N A U T O M A T I C C O X T I Z O L October

APPENDIX

Theorems on the stability of random linear systems:'

Theorem I : Given the differential equation

L,[v(91 + ao(t!u(l) = W ) , E'"'(0) = 0, i = 0, 1, . . ' , q - 1

where L , is a constant parameter gth-order differential operator, and ao( t ) is a sta- tionary, zero-mean. continuous in the mean- Gaussian process, then

lim (u?(t), < cze-?(-c280)r - 1-

where LqF4t)] = 6( t ) , and ce+ =f(t) 2 I ~ ( t ) ' , for c, I, t >O.

S O = J-1 I {a(l)a(t + r ) ) I dr

with the autocorrelation (a( t )a( t+~)\ hav- ing a rational power spectrum.

Theorem 2:

1 where m(t )2O, and For the differential equation in Theorem

with b;, SG>O, and P

so = so;, i=l

then

where 6, C-' denote Laplace and inverse Laplace transforms, respectively.

\V. G. TUEL, JR. P. li. DERUSSO

Dept. of Elec. Eng. Rensselaer Polytechnic Institute

Troy, X. Y .

Analysis of Approximation of Linear and Time-Invariant Systems Pulse Response by Means of Laguerre Finite Term Expansion

A simple calculation method for correla- tion functions through orthonormal series expansions was suggested by Lampard.' hIishkin and Braun refer to orthonormal series expansions for the determination of a mathematical model of nonlinear systems.? Also, Schetzen has studied the problem of optimum Laguerre finite term expansion of

1 D. G . Lampard, aA new method of determi:ing Manuscript received April 27. 1961.

correlation functions oi stationary time series, J. I E E (London); September. 1954.

Systems." McGraw-Hill Book Co.. Inc.. S e w York. 2 E. Mishkin and L. Braun, '.Adaptive Control

X. Y . ; 1961.

a generical function with the condition of minimizing mean-square error.3

The problem of approximation obtain- able, stopping a Laguerre series expansion to a fixed term, is therefore a n object of in- terest. This problem was examined by the author4 under the hypothesis that the func- tion to be expanded is known and has the form of a linear, time-invariant. and lumped parameters-)-stem pulse response

with z; = x ; + j ! ; and x , > 0. (1)

The generical Laguerre function is

and its Laplace transform is

For the function (1) we can consider the expansion con\-ergent in the mean,

f ( t ) = 2 kAdk&it) 0

being its generical coefficient

At = J 0 * j ( l W k ( l ) d l . (3)

If we decompose this expansion in the partial series in regard to each exponential term of the function ( 1 ), we obtain

where, clearly, Axi is the Kth coefficient of the expansion of i th exponential of f(t). Therefore, it is

oc

R,e-z;l = Ek .4i<;Sk(l) . 0

.kcording to (3) . we have

Keeping ( 2 ) and (5) in mind, we can con- sider the Fourier transform of (4)

P - 7nRi

i X i - ? j1 /2 ) ( jw - 111/2) t ] . (6)

It can now he seen that F ( j u ) is ex- pressed through the sum of p complex geo-

(2; + 111/2!(jw + m / 2 )

3 M . Scherzen." Optimum Laguerre Finite Term

Quart. Prop. Rept. S o . 71: October. 1963. Expansion of Function% M.I.T.. Cambridge. Mass.

mediante lo sviluppo in serie di funzioni di Lawerr: C. Bruni. *Xnalisi dell'approseimazione ottenibile

della risposta impulsiva di sistemi lineari e normali. Ric. Sci. R e d . A,, Rome. Italy. pp. 377-408; IIav. 1964.

metric series, and the generic modulus of ratio is

(z; - 111/2)(jw - n1/2) I I zi - 111/2 ~

(zi + 1 ~ / 2 ) ( j u + m/2) I 1

= J (Ti - m/2)' j (Xi + ?12/2)' + y;'

I n the simple h>-pothesis that every ex- ponential term has the same importance in the constitution of f(t), it is reasonable to think that the \vhole series converges as rapidll- to f( t ) as each geometric series converges to its own limit. Therefore, for even- geometric series relative to each ex- ponential term, we should minimize the quantity X<. I t is possible to get the same result for ever\- partial series, minimizing the figure of merit

as being ~ ( t ) the error due to the arrest of expansion to the generic nth term

Ei(f) = R;e-*,t - A a ; S k ( t ) = EL. Aki&(t ) .

.According to Parseval's theorem, it is 0 n+l

Ei = J o * ~ ; 2 ( t ) d t = & I I?, ,l+l

and, keeping ( 5 ) in mind,

Hence, in order to reduce E; t o a mini- mum, for every n, it is necessary to mini- mize, as noted, the quantity

dl; = j z i - m / 2 (x; - 1n/2)2 + y i 2

2; + n1/2 I ( x i + t i ~ / 2 ) ~ + ?'i2 ___ = J

The value of ]?I which satisfies this con- dition is

m ; O = 2d/si2+yi* = 2 I zi , and with this selection, the result is

I t can, therefore, be seen that for every pole or pair of conjugate complex poles, there is for nz, an optimum value nz0 which provides the maximum convergence of rela- tive partial series. Thus, if the function f(t) (1) is reduced simply to one or two ex- ponential terms with conjugate complex ex- ponents, the problem of optimal selection of 172 would be solved, and the relati\-e Value 11; min would give an idea of the rapidity of the expansion convergence. Function (1) is actually given by the superposition of several exponential terms; but, if we sup- pose, as dread>- pointed out, that the im- portance of different terms is approximately equal, it might be reasonable to choose the value of nz which minimizes the greatest of M i , and consider it for estimating con- vergence rapidity of the whole series.

For this procedure it is possible to give simple and convenient graphical representa- tions deriving from the considered formulas