You are on page 1of 13

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/334256859

Limit cycles in relay systems with fractional order plants

Article in Transactions of the Institute of Measurement and Control · July 2019


DOI: 10.1177/0142331219860302

CITATIONS READS

2 2,691

3 authors, including:

Ali Yüce Nusret Tan


Malatya Turgut Özal University Inonu University
28 PUBLICATIONS 204 CITATIONS 124 PUBLICATIONS 2,172 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Ali Yüce on 19 January 2020.

The user has requested enhancement of the downloaded file.


Article
Transactions of the Institute of
Measurement and Control
Limit cycles in relay systems with 2019, Vol. 41(15) 4424–4435
Ó The Author(s) 2019

fractional order plants Article reuse guidelines:


sagepub.com/journals-permissions
DOI: 10.1177/0142331219860302
journals.sagepub.com/home/tim

Ali Yüce1 , Nusret Tan1 and Derek P Atherton2

Abstract
In this paper, limit cycle frequency, pulse width and stability analysis are examined using different methods for relay feedback nonlinear control systems
with integer or fractional order plant transfer functions. The describing function (DF), A loci, a time domain method formulated in state space notation
and Matlab/Simulink simulations are used for the analysis. Comparisons of the results of using these methods are given in several examples. In addition,
the work has been extended to fractional order systems with time delay. Programs have been developed in the Matlab environment for all the theoreti-
cal methods. In particular, Matlab programs have been written to obtain a graphical solution for the A loci method, which can precisely calculate the
limit cycle frequency. The developed solution methods are shown in various examples. The major contribution is to look at finding limit cycles for relay
feedback systems having plants with a fractional order transfer function (FOTF). However, en route to this goal new assessments of limit cycle stability
have been done for a rational plant transfer function plus a time delay.

Keywords
Nonlinear systems, fractional order systems, relay systems, limit cycle, A loci

Introduction as a Padé one, has to be used. One contribution of this paper


is to look at this situation.
Relay feedback systems are unique in that the output from The major contribution however is to look at finding limit
the relay is not dependent on the continuous time variation of cycles for feedback systems having plants with a fractional
the input but only on where the input is equal to the relay order transfer function (FOTF). This situation has been
switching values. It is because of this that exact values for any looked at before, where previous studies (Atherton et al.,
possible limit cycles can be calculated as well as their stability. 2014a, 2014b) have considered the use of the describing func-
For typical feedback control system plants in a loop with a tion and compared the results with the A loci method, initially
relay having hysteresis and dead zone the normal limit cycles for relays with no dead zone and then with dead zone. Results
of interest have one positive and one negative pulse per period were compared with simulations where for FOTFs exact
and this paper is concerned with the evaluation of these. The results cannot be obtained due to the necessity of approximat-
methods presented can be extended to both the determination ing the FOTF by a TF approximation. This work confirmed
of asymmetric limit cycles and multipulse limit cycles, which that although the series involved with the A loci method can-
can occur for some rare plant transfer functions, but these will not be evaluated in closed form, as is the case for a TF with
not be considered here (Atherton, 2011; Choudhury and time delay, provided a sufficient number of terms was used in
Atherton, 1974). the summation good accuracy could be obtained. Here, we
There are two exact approaches for finding limit cycles in extend this work to study the use of the time domain method
feedback loops with rational transfer functions: one using a to obtain results for FOTFs with time delays, which requires
frequency response approach, often known as the Tsypkin or two approximations to get a plant state space representation,
A loci method (Atherton, 1966; Tsypkin, 1984), and the other namely the use of a TF to approximate the FOTF and a Padé
a time domain approach (Chung and Atherton, 1966; Hamel, approximation to approximate the time delay. The effect of
1950). The stability of a predicted limit cycle, when using the the order of the approximations used is examined. In the time
A loci method, can be assessed, apart from the situation domain method eigenvalues are calculated and for a stable
where multiple limit cycles may exist, using an intersection
criterion, similar to that for the describing function approach
1
(Tsypkin, 1984). An exact stability assessment of a limit cycle Engineering Faculty, Department of Electrical and Electronics
can be found using the time domain approach but this Engineering, Inonu University, Turkey
2
University of Sussex, UK
requires a state space description for the transfer function
both for the evaluation of the limit cycle and its stability Corresponding author:
(Balasubramanian, 1981). Thus, if one wishes to apply this Ali Yüce, Inonu University, Engineering Faculty, Department of Electrical
approach when the plant transfer function is a rational one and Electronics Engineering, Malatya 44280,Turkey.
plus a time delay, an approximation for the time delay, such Email: ali.yuce@inonu.edu.tr
Yüce et al. 4425

limit cycle one eigenvalue must have a unit value. For an


unstable limit cycle an eigenvalue must be greater than unity
and its closeness to unity reveals its relative stability. For,
example if one sets up a simulation with initial conditions
near to a limit cycle that has an eigenvalue just greater than
unity the dynamic movement away from the limit cycle will
be slow (Wadey and Atherton, 1987). Figure 1. A nonlinear feedback control system with relay.
The coverage of the paper is as follows. In the next section,
an outline of the calculations involved in using the A loci and
time domain methods for obtaining limit cycles in feedback
systems with a relay having dead zone and hysteresis is given.
In Section 3, the application of these methods to a system with
a plant modelled as a TF with time delay is examined, in par-
ticular with respect to the errors in having to approximate the
time delay by a Padé approximation to obtain a state space
description. Subsequent material is devoted to plants modelled
as FOTFs, which may in some cases provide a more accurate
model representation (Monje et al., 2010). In Section 4, the
application has been developed for FOTFs and some examples
are presented. Conclusions are given in Section 5.

Figure 2. Relay with dead zone and hysteresis.


Limit cycles in relay systems
As mentioned in the introduction, relays are unique nonlinear
elements and because of this exact solutions can be found for
limit cycles in feedback loops containing relays. Here, to pres-
ent the approaches and to allow easy understanding of the
extension to FOTF systems the negative feedback loop shown
in Figure 1 is assumed. The plant has an integer transfer func-
tion G(s), or state space description (A, B, C, 0), and the relay,
Figure 2, has dead zone and hysteresis. The limit cycle is
assumed to be odd symmetric with the relay output, u(t),
which has pulse width t and period T , as shown in Figure 3.
This limit cycle with one positive and one negative pulse per
period, as mentioned in the introduction, is of the form typi-
cally found in feedback control systems. Figure 3. Assumed output of the relay.

X
Frequency domain approach y_ (t) = (2hv=p) gn ½cos (nvt + un )  cos (nvt  nvt + un )
nodd
The Fourier series for the relay output, u, is easily shown to ð4Þ
be expressible as
provided lims!‘ sG(s) = 0.
X 1
u(t) = (2h=p) ½sin nvt  sin (nvt  nvt) ð1Þ When lims!‘ G(s) 6¼ 0 and lims!‘ sG(s) 6¼ 0, then y(t) and
nodd
n y_ (t) are not continuous and well documented changes need to
be made to the final equations (12) and (13) below.
Writing for the transfer function Defining the A locus (Atherton, 1966) of the transfer func-
tion for odd terms n only by
G( jnv) = UG (nv) + jVG (nv) = gn ejun ð2Þ
AoG (u, v) = ReAoG (u, v) + jIm AoG (u, v) ð5Þ
then the output from the transfer function, y(t), is
where
X gn
y(t) = (2h=p) ½sin (nvt + un )  sin (nvt  nvt + un ) X
nodd
n ReAoG (u, v) = VG (nv) sin nu + UG (nv) cos nu ð6Þ
nodd
ð3Þ
X 1
provided lims!‘ G(s) = 0, and its derivative, y_ (t), is Im AoG (u, v) = ½VG (nv) cos nu  UG (nv) sin nu ð7Þ
nodd
n
4426 Transactions of the Institute of Measurement and Control 41(15)

then it is easy to show that (3) and (4) can be written Substituting for x(T =2) from (17) and x(t) from (15) in (16)
gives
y(t) = (2h=p)fIm AoG ( vt, v)  Im AoG ( vt + vt, v)g ð8Þ x(0) = eAf(T =2)tg ½eAt x(0) + A1 (eAt  I)Bh, which, on
using ekA = A1 ekA A can be shown to give
and
 x(0) = (I + eAT=2 )1 A1 (eAT=2  eAf(T=2)tg )Bh ð18Þ
y_ (t) = (2vh=p)fRe AoG ( vt, v)  Re AoG ( vt + vt, v)g ð9Þ

The two switching conditions require The relay input is  y =  Cx and it switches positive at time
0 when Cx(0) = d + D so that
 y(0) = d + D and  y_ (0) . 0 ð10Þ
C(I + eAT=2 )1 A1 (eAT=2  eAf(T =2)tg )Bh  (d + D) = 0 ð19Þ
and
Substituting for x(0) in (15) and using that when the relay
 y(t) = d  D and  y_ (t)\0 ð11Þ switches to zero output at time t, Cx(t) = d  D gives

Thus, the limit cycle solution is given from


C½eAt (I + eAT=2 )1 A1 (eAT =2  eAf(T =2)tg ) + A1 (I  eAt )
AoG (0, v)  AoG (vt, v) must have I:P: = p(d + D)=2h Bh  (d  D) = 0
ð12Þ ð20Þ
and R:P:\0

and These two nonlinear matrix equations are easily solved in


Matlab for the two unknowns t and T , using suitable initial
AoG (0, v)AoG ( vt, v) must have I:P:=p(dD)=2h and R:P:\0 values, estimated for example by a describing function analy-
ð13Þ sis. Now, the differential equation for the autonomous feed-
back loop can be written
Properties of the A loci can be found in reference (Atherton,
2011) and it has also been shown that closed form expressions q(D)x + p(D)n(x) = 0 ð21Þ
can be found for them, thus obviating the need to sum the
series term by term. It should also be noted that the solution where here x is now the input to the relay, with characteristic
can only exist if the computed result for the relay input wave- denoted by n, the transfer function G(s) = p(s)=d(s) and D
form does not cross the switching levels between the assumed denotes the differential operator d=dt. If x (t) is the limit cycle
times. This is often known as the continuity conditions and is solution and it is perturbed by a small amount Dx(t) then the
easily checked from a plot of the calculated waveform. differential equation for the perturbation is

q(D)Dx(t) + p(D)n0 (x (t))Dx(t) = 0 ð22Þ


Time domain approach
Using the state space approach first presented in reference where n0 (x) = dn(x)=dx. The limit cycle is stable if this equa-
(Chung and Atherton, 1966), denoting the plant state by x tion, which has periodically time varying coefficients, has a
and integrating the state equation over the time period from 0 stable solution. It can be written in the state space form
to t ł t with the relay output, h, one has D_x (t) = A(t)Dx (t) where the A matrix can be shown to be

ðt A(t) = A + Bn0 (x (t))C ð23Þ


At At
x(t) = e x(0) + e Bhdt ð14Þ
For the specific case of switching nonlinear elements the sec-
0
ond term in equation (23) is zero at all times apart from the
Thus, at time t switching instants so that A(t) is piecewise continuous. It is
then known for this situation that equation (22) is stable if
ðt and only if the eigenvalues of the matrix
At
x(t) = e x(0) + eAt Bhdt = eAt x(0) + A1 (eAt  I)Bh: ð15Þ
Y
m
0
Q= exp (Ai ti ) ð24Þ
i=1
Similarly integrating from t to T =2 when the relay output is 0
gives are all less than unity, apart from one which will have unit
magnitude (Willems and Pandit, 1971). Here, Ai are the con-
x(T =2) = eAf(T =2)tg x(t) ð16Þ stant values of the matrix A in the periods ti and the period T
is given by
Since the limit cycle has odd symmetry
X
m
x(T =2) =  x(0) ð17Þ T= ti ð25Þ
i=1
Yüce et al. 4427

Table 1. Results from simulation, A locus and state space methods.

K Method v rad/sec t sec u = vt Stability Eigenvalues

8 Simulation 1.402 1.813 2.542 Stable For v1 =1.402 For v2 =1.013


A locus 1.402 1.813 2.542 Stable 0.0027 0.0000
1.007 0.314 0.316 Unstable 1.0000 1.0002
Time domain 1.402 1.813 2.542 Stable 0.4524 4.1925
1.013 0.311 0.315 Unstable
4.5 Simulation 1.367 1.057 1.445 Stable For v1 =1.367 For v2 =1.340
A locus 1.368 1.062 1.453 Stable 0.0011 0.0008
1.339 0.935 1.252 Unstable 0.9264 1.0000
Time domain 1.367 1.057 1.445 Stable 1.0000 1.0828
1.340 0.940 1.260 Unstable
4.488 Simulation 1.355 1.003 1.359 Stable For v1 =1.355 For v2 =1.353
A locus 1.360 1.023 1.391 Stable 0.0010 0.0009
1.348 0.970 1.308 Unstable 0.9944 1.0001
Time domain 1.355 1.002 1.358 Stable 0.9999 1.0055
1.353 0.994 1.345 Unstable

Thus, for the case of an odd symmetrical limit cycle in the The switching conditions are  Cx(0) = d + D and
above system Cx(t) = d  D, as before.
    The frequency domain solution is easy to implement for a
hBC hBC time delay in the plant as it is easily shown that (Atherton,
Q = exp½Af(T=2)  tg exp exp½At exp ð26Þ
j_x(t)j j_x(0)j 1981) if G1 (s) = G(s)est , then

and the limit cycle will be stable if all the eigenvalues of Q are AG1 (u, v) = AG (u + vt, v) ð32Þ
less than or equal to unity. One eigenvalue is always unity cor-
responding to the limit cycle solution, that is the limit cycle Example 1: In this example a feedback loop with the follow-
amplitude neither increases or decreases. ing time delay transfer function and relay parameters is con-
sidered, namely

Extension to plant with time delay K


G(s) = eLs , d = 1, h = 1, D = 0 ð33Þ
Many plant transfer functions include a time delay and it is s(s + 1)(0:5s + 1)
easy to extend the above results to this case. For the state
In order to determine the number of terms to take in the A
space analysis, assuming as in section 2 that the relay switches
loci, results were first obtained for no delay, that is L = 0.
positive at time zero, then if the time delay is u the input to
To find possible limit cycles the A loci of equations (12)
the plant dynamics (A, B, C, 0) is the delayed relay output
and (13) were plotted and the values of v and t that satis-
u(t  u). One then has
fied them were found and from these relationships the val-
ues of v and t, which satisfied both. Values were found to
x(u) = eAu x(0) ð27Þ
be accurate to four significant figures using 101 terms when
x(t + u) = eAt x(u) + A1 (eA(t + u)  eAu )Bh = eA(t + u) x(0) compared with the exact answers using the closed forms.
Results from the time domain method and simulations
+ A1 (eA(t + u)  eAu )Bh
were also obtained. A sample is shown in Table 1 starting
ð28Þ from K = 8, which was then reduced until the limit cycles
disappeared.
Since the input to G(s) is zero from t + u to T =2 then Results were then found for L = 0:2 using 101 terms in the
x(T =2) = eA½(T =2)tu x(t + u) so that A loci series. They are given in Table 2 from the A loci, simu-
lation and time domain methods for different values of K.
x(T =2) = eAT=2 x(0) + A1 (eAT=2  eA½(T =2)t )Bh ð29Þ Padé approximations of 2/2 and 4/4 were used for the time
delay term in the time domain method as shown.
Since x(T =2) =  x(0) one has
From Table 2 it can be seen that the system is now stable
as expected for the lower gain of K ł 2:82, compared with
 x(0) = (I + eAT=2 )1 A1 (eAT=2  eAf(T=2)tg )Bh ð30Þ
K ł 4:486 for no time delay. The eigenvalues for K = 8, 6, 4,
and 3 and 2.82 using 2/2 and 4/4 Padé approximations are given
in Table 3. From Table 3, it can be observed that the results
x(t) = eAt x(0) + A1 (eAt  eAu )Bh ð31Þ obtained from 2/2 and 4/4 Padé approximations in this case
are almost identical. Therefore, it would appear to be satisfac-
where (30) is the same as for no time delay tory to use a 2/2 Padé approximation. For the case of K = 4
4428 Transactions of the Institute of Measurement and Control 41(15)

Table 2. The limit cycle frequencies obtained from A locus, simulation and state space for L=0.2.

K Stability A loci Simulation Time domain method

2/2 Padé 4/4 Padé

v rad/sec t sec v rad/sec t sec v rad/sec t sec v rad/sec t sec

8 Stable 1.093 2.537 1.096 2.530 1.093 2.537 1.093 2.537


Unstable 0.381 0.250 - - 0.372 0.250 0.386 0.250
6 Stable 1.099 2.400 1.101 2.392 1.099 2.399 1.099 2.399
Unstable 0.537 0.339 - - 0.539 0.339 0.537 0.339
4 Stable 1.110 2.080 1.112 2.069 1.110 2.080 1.110 2.080
Unstable 0.753 0.554 - - 0.754 0.555 0.753 0.554
3 Stable 1.099 1.606 1.100 1.576 1.099 1.604 1.099 1.604
Unstable 0.955 0.921 - - 0.956 0.923 0.956 0.922
2.82 Stable No Solution - No Solution - 1.051 1.249 1.051 1.249
Unstable No Solution - No Solution - 1.051 1.249 1.051 1.249
2.81 Stable No Solution - No Solution - No Solution - No Solution -
Unstable No Solution - No Solution - No Solution - No Solution -

when v = 0:754 rad= sec some of the corresponding eigenva- shown in Figure 5. Similar results were obtained for K=3
lues are greater than one, which means this frequency is the with L varied as shown in Table 5 and Figure 6.
frequency of an unstable limit cycle, on the other hand, when
v = 1:110 rad= sec one eigenvalue is equal to one and the oth-
ers are smaller than one since this frequency is the stable limit Limit cycles in systems with a FOTF
cycle frequency. For K = 2:82 the frequency of the stable The application of the time domain limit cycle method to sys-
limit cycle is 1.051 rad/sec and the frequency of the unstable tems with a fractional order plant is not straight forward as
limit cycle is 1.051 rad/sec which are equal. Therefore, we can to obtain a state space description for the plant requires the
conclude that K = 2:82 is the critical value for stability. This fractional order one to be replaced by an approximate integer
also can be seen from Figure 4, which shows graphs of the one. Then for good accuracy the order of this transfer func-
variation of the frequencies of the stable and unstable limit tion needs to be quite high. On the other hand, since fre-
cycles with K. quency responses of fractional order plants are exact the
It will also be noted from Tables 1 and 3 that the unstable frequency domain approach can be done exactly but the A
eigenvalue for the unstable limit cycle approaches unity as K loci have to be found computational by summation of the
is decreased towards the point where no limit cycle exists. A series to a finite number of terms.
practical aspect of this is that in a simulation if one can start
on the unstable limit cycle divergence from it will be slower as Example 2: Consider now the system of Figure 1 with the
this value of K is approached. For two fixed values of K FOTF given in (34).
and varying L, the results are given in Tables 4 and 5. From
Table 4 it can be seen that for K=4 and L=0.03 the time K
G(s) = , d = 1, h = 1, D = 0 ð34Þ
domain method predicts a critical limit cycle frequency value s1:2 (s + 1)(0:5s + 1)
of 1.271 rad/sec, that is the stability limit, whereas the A loci
The procedure given above is applied to this transfer function
and simulation methods show the system to have no limit
and the results obtained from the A loci method and simula-
cycle, that is, stable, for these values. This error is due to the
tion are given in Table 6. Here, for the simulation we
used Oustaloup’s seventh order integer approximate transfer
function given in (35) for the frequency band
(vl , vh ) = (0:01 rad= sec , 100 rad= sec ).

1 s7 + 80:73s6 + 1378s5 + 5967s4 + 6806s3 + 2045s2 + 155:9s + 2:512


= ð35Þ
s0:2 2:512s7 + 155:9s6 + 2045s5 + 6806s4 + 5967s3 + 1378s2 + 80:73s + 1
Padé approximation used to get the state space representation From Table 6, it can be seen that the system becomes sta-
and the convergence to the stability limit as L is varied is ble according to the A locus method for K ł 2.3739. The
simulation results were slightly in error due to the 7th order
Yüce et al. 4429

Table 3. The eigenvalues for L=0.2 and different values of K.

K Eigenvalues

2/2 Padé 4/4 Padé

8 v = 1:093 rad/sec v = 0:372 rad/sec v = 1:093 rad/sec v = 0:386 rad/sec


1.0000 2707.1757 1.0001 212.7602
0.2834 0.1235 0.2834 0.3310
0.0006 0.0000 0.0006 0.0000
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000
0.0000 0.0000
6 v = 1:099 rad/sec v = 0:539 rad/sec v = 1:099 rad/sec v = 0:537 rad/sec
1.0000 15.2775 1.0001 16.6029
0.3097 1.2145 0.3097 1.1217
0.0005 0.0000 0.0005 0.0000
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000
0.0000 0.0000
4 v = 1:110 rad/sec v = 0:754 rad/sec v = 1:110 rad/sec v = 0:753 rad/sec
1.0000 4.6842 1.0001 4.6712
0.3975 1.0059 0.3974 1.0097
0.0004 0.0000 0.0000 0.0000
0.0000 0.0000 0.0004 0.0000
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000
0.0000 0.0000
3 v = 1:099 rad/sec v = 0:956 rad/sec v = 1:099 rad/sec v = 0:956 rad/sec
1.0000 1.7299 1.0001 1.7298
0.6369 1.0008 0.6366 1.0016
0.0002 0.0000 0.0002 0.0000
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000
0.0000 0.0000
2.82 v = 1:051 rad/sec v = 1:051 rad/sec v = 1:051 rad/sec v = 1:051 rad/sec
1.0000 1.0001 1.0002 1.0001
1.0000 1.0001 1.0002 1.0001
0.0001 0.0001 0.0001 0.0001
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000
0.0000 0.0000


integer approximation used for the fractional order of 1 s0:2 K=2.38 the stable and unstable limit cycle frequencies
and according to the results of simulation the system becomes obtained from the seventh order approximation are equal to
stable for K ł 2.38. each other at 0.952 rad/sec. so that the critical gain K for sys-
The results obtained from the time domain method for tem stability is 2.38 compared with the A locus value of
K=8 and K=2.38 using Oustaloup first, third, fifth and 2.3739. Further investigations of higher order Oustaloup
seventh order approximations are given in Tables 7 and 8. approximations showed that a 13th order one was required to
For K=8 it can be seen from Table 7 that the stable limit yield a critical gain and frequency to three figures in agree-
cycle is computed from the first, third, fifth and seventh order ment with the A locus result.
approximations at frequencies of 0.579 rad/sec, 0.938 rad/sec, Another approximation that can be used in obtaining
0.985 rad/sec and 0.990 rad/sec respectively, and the unstable results is to take fewer terms in the A loci series although
limit cycle is computed from first, third, fifth and seventh there is no reason to do this with modern computation facili-
order approximations at 0.257 rad/sec, 0.431 rad/sec, 0.342 ties. Table 9 shows the results for values of n equal to 1, 3, 5,
rad/sec and 0.339 rad/sec respectively. Table 8 shows that for 21 and 101. This shows as expected that errors in the results
4430 Transactions of the Institute of Measurement and Control 41(15)

Example 3: Consider the system of Figure 1 with a transfer


function including both fractional order and time delay as
given in (36).

K
G(s) = eLs , d = 1, D = 0, h = 1 ð36Þ
sa (s + 1)(0:5s + 1)

In this example, three different cases are examined:

1. a and L are constant and K is varied;


2. K and L are constant and a is varied;
3. K and a are constant and L is varied.

In both the time domain method and simulation,


Oustaloup’s, seventh order integer approximation has been
used for a. The frequency bandwidth for Oustaloup method
is selected as (vl , vh ) = (0:01 rad= sec , 100 rad= sec ). Padé
approximations of order 2/2 and 4/4 have been used for L in
Figure 4. Stable and unstable limit cycle frequencies using the time both the time domain method and the simulation. The only
domain method for L=0.2 and different values of K. approximation in the A loci method is of course taking a
finite number of terms, namely 101, used in the summation.
Here, for the first case, a = 1:2 and L = 0:2 are selected. The
limit cycle frequency and pulse width results of the A loci,
are worse for lower values of n when the limit cycle has nar- simulation and time domain methods are given for varying K
rower pulses. Interestingly, however, around the gain value in Table 10. It can be seen in Table 10 that the results
where no limit cycle exists the accuracy for n = 5, that is three obtained from the different methods are in agreement to the
terms only in the series, is comparable with the simulation accuracy expected with the computation methods used and
results having the approximation for the fractional integrator. the system is stable for K ł 1:6466.
For the second case, K = 6 and L = 0:2 are selected. The
same procedure is applied for this case. The limit cycle

Table 4. K=4 and varying L.

L Stability A loci Simulation Time domain method

2/2 Padé 4/4 Padé

v rad/sec t sec v rad/sec t sec v rad/sec t sec v rad/sec t sec

0.02 Stable No solution – No solution – No solution – No solution –


Unstable No solution – No solution – No solution – No solution –
0.03 Stable No solution – No solution – 1.271 1.026 1.271 1.026
Unstable No solution – No solution – 1.271 1.026 1.271 1.026
0.04 Stable 1.285 1.109 1.296 1.189 1.286 1.110 1.286 1.110
Unstable 1.259 0.994 – – 1.258 0.990 1.258 0.990
0.05 Stable 1.291 1.273 1.293 1.288 1.291 1.271 1.291 1.272
Unstable 1.197 0.862 – – 1.198 0.864 1.197 0.862
0.10 Stable 1.233 1.631 1.235 1.602 1.233 1.630 1.233 1.630
Unstable 1.010 0.674 – – 1.010 0.675 1.010 0.675
0.15 Stable 1.168 1.875 1.170 1.855 1.169 1.874 1.169 1.874
Unstable 0.869 0.597 – – 0.870 0.598 0.869 0.598
0.20 Stable 1.110 2.080 1.112 2.063 1.110 2.080 1.110 2.080
Unstable 0.753 0.554 – – 0.754 0.555 0.753 0.554
0.25 Stable 1.058 2.264 1.059 2.251 1.058 2.263 1.058 2.264
Unstable 0.654 0.529 – – 0.656 0.529 0.654 0.529
0.30 Stable 1.011 2.433 1.012 2.422 1.011 2.432 1.011 2.433
Unstable 0.515 0.570 – – 0.569 0.514 0.566 0.514
0.35 Stable 0.969 2.592 0.970 2.582 0.969 2.591 0.969 2.592
Unstable 0.484 0.505 – – 0.488 0.506 0.486 0.506
0.40 Stable 0.931 2.744 0.932 2.734 0.931 2.742 0.931 2.743
Unstable 0.407 0.502 – – 0.404 0.502 0.409 0.502
Yüce et al. 4431

Table 5. K=3 and varying L.

L Stability A loci Simulation Time domain method

2/2 Padé 4/4 Padé

v rad/sec t sec v rad/sec t sec v rad/sec t sec v rad/sec t sec

0.03 Stable No solution - No solution - No solution - No solution -


Unstable No solution - No solution - No solution - No solution -
0.04 Stable No solution - No solution - No solution - No solution -
Unstable No solution - No solution - No solution - No solution -
0.05 Stable No solution - No solution - No solution - No solution -
Unstable No solution - No solution - No solution - No solution -
0.10 Stable No solution - No solution - No solution - No solution -
Unstable No solution - No solution - No solution - No solution -
0.15 Stable No solution - No solution - 1.084 1.174 1.084 1.174
Unstable No solution - No solution - 1.084 1.174 1.084 1.174
0.20 Stable 1.099 1.605 1.102 1.558 1.100 1.605 1.100 1.605
Unstable 0.954 0.924 - - 0.956 0.922 0.956 0.921
0.25 Stable 1.058 1.877 1.065 1.857 1.059 1.878 1.059 1.878
Unstable 0.835 0.810 - - 0.838 0.810 0.837 0.809
0.30 Stable 1.016 2.088 1.013 2.077 1.017 2.093 1.017 2.093
Unstable 0.741 0.751 - - 0.743 0.752 0.742 0.751
0.35 Stable 0.976 2.274 0.967 2.270 0.976 2.281 0.976 2.282
Unstable 0.663 0.721 - - 0.663 0.717 0.660 0.716
0.40 Stable 0.938 2.453 0.938 2.444 0.939 2.454 0.939 2.455
Unstable 0.592 0.697 - - 0.593 0.696 0.588 0.694

Figure 5. Stable and unstable limit cycle frequencies using the time Figure 6. Stable and unstable limit cycle frequencies using the time
domain method for K=4 and different values of L. domain method for K=3 and different values of L.

frequency and pulse width results are given for varying a in that there is a really significant difference between the 2/2 and
Table 11. It can be seen in Table 11 that while the simulation 4/4 Padé approximation results for the case of an unstable
result is stable for a ł 0:7, the A loci and time domain meth- limit cycle, because the pulse width of the unstable limit cycles
ods indicate limit cycle frequencies. Also, the A loci and time is narrower so that the limit cycle waveform at the relay out-
domain methods show a stable system for a ł 0:6. put contains more higher frequencies.
Finally, K = 2 and a = 1:2 are selected. The results for
limit cycle and pulse width are given in Table 12. It can be
seen from the table that the system is stable for no time delay
case (L = 0). Furthermore, it is shown in Tables 10, 11 and 12
4432 Transactions of the Institute of Measurement and Control 41(15)

Table 6. The limit cycle frequencies from the A locus method and theoretical methods are covered, the approximate describing
simulation. function approach, the A loci method that involves a summa-
tion of an infinite series, where the DF approximation just
K A loci Simulation takes the first term, and a time domain approach formulated
v rad/sec t sec v rad/sec t sec
in the state space representation. Results are compared with
those from simulations, which like all the theoretical methods
8 0.987 2.884 0.991 2.871 involves approximations for the FOTF.
0.309 0.168 - - To develop the work in a logical and understandable man-
3 1.004 2.150 1.008 2.131 ner it was felt necessary to first give some results for a rational
0.755 0.680 - - plant transfer function. Here, for a plant with no time delay it
2.5 0.991 1.738 0.993 1.704 is concluded that all the methods, apart from the approximate
0.877 1.017 - -
describing function approach, give accurate results for obtain-
2.4 0.973 1.535 0.968 1.455
ing the limit cycle frequency and pulse width. The A loci pro-
0.922 1.201 - -
2.38 0.963 1.447 No Solution No Solution vides a nice graphical interpretation and the limit cycle
0.938 1.286 stability can be assessed from an intersection criterion. The
2.375 0.956 1.400 No Solution No Solution time domain method has the advantage that the eigenvalues it
0.945 1.330 gives provide information on the limit cycle dynamics. When
2.3739 0.951 1.365 No Solution No Solution extending the approaches to a rational plant transfer function
0.951 1.365 with time delay, as in example 1, the previous comments
apply apart from for the time domain method. Here, since a
rational approximation is required for the time delay an
approximate solution is obtained. It is not easy to find how
Conclusions good the approximation should be to obtain a given accuracy
but it is clear from the results, particularly so for the results
The objective of this paper has been to look at methods for
for example 3 in Table 10, that they are worse for a given
determining the frequency and stability of limit cycles in relay
order time delay approximation for the unstable limit cycle.
feedback systems with fractional order plant transfer func-
This is to be expected as the unstable limit cycle has narrower
tions. The theoretical techniques can be extended to more
pulse widths, which indicates a larger content of higher har-
complicated forms of limit cycle but here the concentration
monics, so the time delay approximation for better results has
has been on the normal form found in control systems with
two pulses per period for a relay with dead zone. Three

Table 7. Results from the time domain method for K=8.

K v t sec u = vt Eigenvalues Stability


rad/sec
1.Order Oustaloup 3.Order Oustaloup 5.Order Oustaloup 7.Order Oustaloup

8 1.Order Oustaloup 1.0000 1.0000 1.0000 1.0000 Stable


0.579 5.274 3.054 0.3663 0.8053 0.9077 0.9324
3.Order Oustaloup 0.0012 0.3974 0.5296 0.7683
0.938 3.069 2.879 0.0000 0.0160 0.3816 0.3751
5.Order Oustaloup 0.0006 0.0249 0.3613
0.985 2.892 2.849 0.0000 0.0006 0.0289
7.Order Oustaloup 0.0000 0.0004
0.990 2.871 2.842 0.0000 0.0000
0.0000
0.0000
1.Order Oustaloup 26.5001 7.5179 4.5189 4.9450 Unstable
0.257 0.102 0.026 1.3485 0.9020 0.7020 0.7472
3.Order Oustaloup 0.0000 0.6013 0.7514 0.8135
0.431 0.180 0.078 0.0000 0.0002 0.0684 0.4392
5.Order Oustaloup 0.0000 0.0000 0.0172
0.342 0.174 0.060 0.0000 0.0000 0.0000
7.Order Oustaloup 0.0000 0.0000
0.339 0.173 0.059 0.0000 0.0000
0.0000
0.0000
Yüce et al. 4433

Table 8. Results from the time domain method for K=2.38.

K v tsec u = vt Eigenvalues Stability


rad/sec
1.Order Oustaloup 3.Order Oustaloup 5.Order Oustaloup 7.Order Oustaloup

2.38 1.Order Oustaloup 1.0000 0.8086 1.0000 1.0026 Stable


0.582 4.864 2.831 0.3885 1.0000 0.8894 0.9976
3.Order Oustaloup 0.0000 0.6648 0.9063 0.9298
0.949 1.989 1.888 0.0011 0.0149 0.5308 0.7614
5.Order Oustaloup 0.0005 0.0217 0.3567
0.967 1.497 1.448 0.0000 0.0002 0.0240
7.Order Oustaloup 0.0000 0.0001
0.952 1.358 1.293 0.0000 0.0001
0.0000
0.0000
1.Order Oustaloup 9.3129 1.6769 1.1312 1.0025 Unstable
0.345 0.371 0.128 1.0016 0.9999 0.9999 0.9975
3.Order Oustaloup 0.0000 0.7834 0.9026 0.9298
0.816 0.951 0.776 0.0000 0.0067 0.5189 0.7614
5.Order Oustaloup 0.0000 0.0183 0.3567
0.928 1.242 1.153 0.0001 0.0001 0.0240
7.Order Oustaloup 0.0000 0.0001
0.952 1.358 1.293 0.0000 0.0001
0.0000
0.0000

Table 9. Results from A loci for different values of n.

K n = 1 (DF) n=3 n=5 n = 21 n = 101

A loci A loci A loci A loci A loci

v rad/s t sec v rad/s t sec v rad/s tsec v rad/s t sec v rad/s t sec

8 1.008 2.797 0.989 2.874 0.987 2.883 0.987 2.884 0.987 2.884
0.997 0.318 0.810 0.253 0.705 0.230 0.446 0.187 0.309 0.168
3 1.009 2.122 1.008 2.129 1.005 2.143 1.004 2.150 1.004 2.150
1.008 0.993 0.857 0.769 0.7921 0.709 0.756 0.681 0.755 0.680
2.5 No Solution 0.994 1.682 0.9916 1.730 0.991 1.738 0.991 1.738
No Solution 0.905 1.071 0.8838 1.030 0.877 1.017 0.877 1.017
2.4 No Solution 0.967 1.447 0.9712 1.490 0.973 1.534 0.973 1.535
No Solution 0.938 1.263 0.9365 1.256 0.922 1.201 0.922 1.201
2.38 No Solution No Solution No Solution 0.963 1.445 0.963 1.447
No Solution No Solution No Solution 0.939 1.290 0.938 1.286
2.375 No Solution No Solution No Solution 0.956 1.400 0.956 1.400
No Solution No Solution No Solution 0.945 1.330 0.945 1.330
2.3739 No Solution No Solution No Solution No Solution 0.951 1.365
No Solution No Solution No Solution No Solution 0.951 1.365

to be good at higher frequencies, which is the case when a has to be found. In the case of the A loci method, however,
higher order approximation is used. the only approximation is in taking a finite number of terms
In extending the methods to an FOTF first without time in an infinite series, as unlike the case of a rational transfer
delay all the methods are approximate including the simula- function a closed form solution for the series is not obtain-
tion results for which a rational approximate transfer function able. The error, however, is controllable by any of several
4434 Transactions of the Institute of Measurement and Control 41(15)

Table 10. The limit cycle frequencies obtained from A loci, simulation and the time domain method for a = 1:2 and L=0.2.

K Stability A loci Simulation Time domain method

2/2 Padé 4/4 Padé

v rad/sec t sec v rad/sec t sec v rad/sec t sec v rad/sec t sec

8 Stable 0.810 3.620 0.812 3.605 0.813 3.605 0.813 3.605


Unstable 0.041 0.102 - - 0.043 0.113 0.049 0.071
6 Stable 0.812 3.520 0.813 3.505 0.815 3.505 0.815 3.505
Unstable 0.047 0.145 - - 0.053 0.156 0.051 0.100
4 Stable 0.818 3.307 0.820 3.291 0.821 3.291 0.821 3.291
Unstable 0.048 0.220 - - 0.046 0.228 0.051 0.190
3 Stable 0.824 3.071 0.826 3.054 0.827 3.055 0.827 3.055
Unstable 0.053 0.300 - - 0.057 0.315 0.050 0.194
2 Stable 0.830 2.481 0.832 2.460 0.833 2.460 0.833 2.460
Unstable 0.513 0.827 - - 0.535 0.847 0.535 0.848
1.8 Stable 0.822 2.199 0.824 2.169 0.824 2.171 0.824 2.171
Unstable 0.625 1.059 - - 0.641 1.083 0.641 1.083
1.7 Stable 0.807 1.957 0.808 1.908 0.807 1.913 0.807 1.913
Unstable 0.693 1.268 - - 0.709 1.306 0.709 1.306
1.6466 Stable No solution No solution No solution No solution
Unstable No solution No solution No solution No solution

Table 11. The limit cycle frequencies obtained from A loci, simulation and time domain for K=6 and L=0.2.

a Stability A loci Simulation Time domain method

2/2 Padé 4/4 Padé

v rad/sec t sec v rad/sec t sec v rad/sec t sec v rad/sec t sec

1.2 Stable 0.812 3.520 0.813 3.505 0.815 3.505 0.815 3.505
Unstable 0.047 0.145 - - 0.053 0.156 0.051 0.100
1.1 Stable 0.949 2.908 0.950 2.902 0.950 2.902 0.950 2.902
Unstable 0.062 0.224 - - 0.061 0.229 0.066 0.182
1 Stable 1.099 2.400 1.100 2.393 1.099 2.399 1.099 2.399
Unstable 0.537 0.339 - - 0.545 0.339 0.543 0.339
0.9 Stable 1.265 1.964 1.263 1.968 1.263 1.968 1.263 1.968
Unstable 0.828 0.399 - - 0.826 0.399 0.824 0.399
0.8 Stable 1.446 1.577 1.441 1.587 1.442 1.587 1.441 1.587
Unstable 1.086 0.461 - - 1.082 0.460 1.080 0.460
0.7 Stable 1.638 1.208 No solution 1.631 1.226 1.631 1.227
Unstable 1.379 0.546 - - 1.366 0.541 1.364 0.540
0.6 Stable No solution No solution No solution No solution
Unstable No solution - - No solution No solution
Yüce et al. 4435

Table 12. The limit cycle frequencies obtained from A loci, simulation and time domain for K= 2 and a = 1:2.

K Stability A loci Simulation Time domain method

2/2 Padé 4/4 Padé

v rad/sec t sec v rad/sec t sec v rad/sec t sec v rad/sec t sec

1 Stable 0.509 5.309 0.512 5.275 0.512 5.266 0.512 5.274


Unstable 0.055 0.455 - - 0.186 0.612 0.041 0.428
0.8 Stable 0.561 4.701 0.559 4.671 0.564 4.666 0.564 4.671
Unstable 0.051 0.450 - - 0.111 0.540 0.042 0.430
0.6 Stable 0.627 4.057 0.631 4.030 0.631 4.029 0.631 4.031
Unstable 0.024 0.380 - - 0.059 0.464 0.042 0.430
0.4 Stable 0.714 3.347 0.718 3.324 0.718 3.325 0.718 3.325
Unstable 0.029 0.390 - - 0.039 0.423 0.039 0.423
0.2 Stable 0.829 2.479 0.833 2.460 0.833 2.460 0.833 2.460
Unstable 0.499 0.822 - - 0.535 0.847 0.536 0.848
0 Stable No solution No solution No solution No solution
Unstable No solution - No solution No solution

methods one may use to decide where to terminate an infinite Atherton DP (2011) An Introduction to Nonlinearity in Control Sys-
series. It is shown in example 2 that good accuracy can be tems. Derek Atherton & Ventus Publishing ApS, ebook at
obtained with a relatively small number of terms. When the bookboon.com.
FOTF also has a time delay, as in example 3, using the time Atherton DP, Tan N, Yeroglu C, et al. (2014a) Computation of limit
domain approach requires in addition to the transfer function cycles in nonlinear feedback loops with fractional order plants. In:
2014 International Conference on Fractional Differentiation and Its
approximation an additional approximation for the time delay.
Applications (ICFDA 2014). Catania, Italy, 23–25 June 2014, pp.
1–6. IEEE.
Declaration of conflicting interests Atherton DP, Tan N, Yeroglu C, et al. (2014b) Limit cycles in non-
linear systems with fractional order plants. Machines 2(3):
The author(s) declared no potential conflict of interests with 176–201.
respect to the research, authorship and/or publication of this Balasubramanian R (1981) Stability of limit cycles in feedback sys-
article. tems containing a relay. IEEE Proceedings D (Control Theory and
Applications) 128(1): 24–29.
Choudhury S and Atherton D (1974) Limit cycles in high-order non-
Funding linear systems. Proceedings of the Institution of Electrical Engineers
The author(s) disclosed receipt of the following financial sup- 121(7): 717–724.
port for the research, authorship, and/or publication of this Chung J and Atherton D (1966) The determination of periodic modes
in relay systems using the state space approach. International Jour-
article: This work is supported by the Scientific and Research
_ nal of Control 4(2): 105–126.
Council of Turkey (TÜBITAK) under Grant no. EEEAG-
Hamel B (1950) Étude mathématique des systèmes à plusieurs degres
115E388. de liberte décrits par des équations linéaires avec un terme de com-
mande discontinu. Proceedings of the Journe´es d’Ètudes des Vibra-
ORCID iD tions AERA. Paris.
Monje CA, Chen Y, Vinagre BM, et al. (2010) Fractional-order Sys-
Ali Yüce https://orcid.org/0000-0002-4642-6272 tems and Controls: Fundamentals and Applications. London:
Springer Science & Business Media.
Tsypkin IZ (1984) Relay Control Systems. Cambridge: CUP Archive.
References Wadey M and Atherton D (1987) A simulation study of unstable limit
Atherton DP (1966) Conditions for periodicity in control systems con- cycles. IFAC Proceedings Volumes 20(12): 149–154.
taining several relays. 3rd IFAC Congress, Paper 28E. London. Willems J and Pandit M (1971) Stability theory of dynamical systems.
Atherton DP (1981) Stability of Nonlinear Systems. New York, NY, IEEE Transactions on Systems, Man, and Cybernetics 1(4):
USA: John Wiley & Sons. 408–408.

View publication stats

You might also like