Evolving Systems

https://doi.org/10.1007/s12530-019-09281-w

ORIGINAL PAPER

General controllability and observability tests for Takagi‑Sugeno fuzzy

systems
J. A. Meda‑Campaña1 · J. de J. Rubio2   · C. Aguilar‑Ibañez3 · R. Tapia‑Herrera4 · R. Gonzalez‑Salazar1 ·
R. A. Rodriguez‑Manzanarez1 · G. Lopez‑Contreras1 · J. O. Hernandez‑Monterrosas1 · I. Elias2 · D. R. Cruz2

Received: 27 August 2018 / Accepted: 7 March 2019

© Springer-Verlag GmbH Germany, part of Springer Nature 2019

Abstract
An approach for investigating controllability and observability properties in Takagi-Sugeno (TS) fuzzy systems is given. The
proposed method is independent of the number of fuzzy rules acting at the same instant and independent of the number of
inputs and outputs included in the TS fuzzy model. Therefore, it can be applied to a wide class of fuzzy systems. The analysis
relies on the solution of a set of symbolic simultaneous equations with the fuzzy weights as the unknowns of such equations.

Keywords  Takagi-Sugeno fuzzy models · Fuzzy controllability · Fuzzy observability

1 Introduction Unfortunately, for nonlinear systems, such properties cannot

be easily analyzed, in general (Khalil 1996).
Controllability and observability properties have great On the bright side, some techniques allowing to effi-
importance in the monitoring and control of linear and non- ciently extend some linear results to the nonlinear field have
linear systems. For instance, their analysis allows determin- been developed. Among such techniques, the Takagi-Sugeno
ing the existence of feedback stabilizers and state estimators, (TS) fuzzy modelling stands out because of its capability
even before the design process is initialized (Kailath 1980; of describing complex nonlinear dynamics by the adequate
Khalil 1996). blending of linear local subsystems (Tanaka and Wang
For linear systems, controllability and observability 2001).
properties are determined by the rank of the controllabil- Furthermore, the relation between TS fuzzy models and
ity and observability matrices, respectively (Kailath 1980). evolving systems has been discussed in several works. For
instance in Angelov et al. (2004), the authors present some

* J. de J. Rubio I. Elias
[email protected] [email protected]
J. A. Meda‑Campaña D. R. Cruz
[email protected] [email protected]
C. Aguilar‑Ibañez 1
Sección de Estudios de Posgrado e Investigación de la
[email protected]
Escuela Superior de Ingenieria Mecanica y Electrica,
R. Tapia‑Herrera Campus Zacatenco del Instituto Politécnico Nacional,
[email protected] Mexico City, Mexico
2
R. Gonzalez‑Salazar Sección de Estudios de Posgrado e Investigación, Esime
[email protected] Azcapotzalco, Instituto Politécnico Nacional, Av. de las
Granjas no. 682, Col. Santa Catarina, 02250 Mexico City,
R. A. Rodriguez‑Manzanarez
Mexico
[email protected]
3
Centro de Investigación en Computación, Instituto
G. Lopez‑Contreras
Politécnico Nacional, Mexico City, Mexico
[email protected]
4
Universidad Tecnológica de la Mixteca, Huajuapan de León,
J. O. Hernandez‑Monterrosas
Oaxaca, Mexico
[email protected]

13
Vol.:(0123456789)
 Evolving Systems

techniques based on the recursive clustering and modified and Meda-Campana et al. (2018). Then, the general control-
recursive least squares (RLS) to perform on-line identifi- lability and observability tests are developed in Sect. 4. The
cation with evolving Takagi-Sugeno fuzzy models (eTS). usage of the approach is illustrated in Sect. 5. Finally, some
In Angelov and Yager (2013), a new data fusion operator concluding remarks are given in Sect. 6.
based on averaging that is weighted by the density of each
particular data sample is presented. In Angelov and Kasa- 2 Description of nonlinear systems
bov (2005), a new paradigm of the evolving computational by Takagi‑Sugeno fuzzy models
intelligence systems (ECIS) is introduced in a generic frame-
work of the knowledge and data integration (KDI). Finally, Nonlinear systems are present in almost every branch of engi-
in Baruah and Angelov (2014) and Baruah and Angelov neering and science, but the analysis of such systems is very
(2012), efficient on-line evolving clustering approaches for complex in general (Khalil 1996). On the other hand, the study
streaming data are proposed on the basis of fuzzy models. of linear systems (Kailath 1980) has reached a very mature
Although, the existence of feedback stabilizers and/or stage, at least at some important aspects as stability analysis,
observers can be directly determined from the controllability the design of stabilizers, state estimation, etc. In this scenario,
or observability property, respectively, most of the available the Takagi-Sugeno (TS) fuzzy modelling has been proposed
methods for TS fuzzy systems suggest the design of control- to simplify the analysis of nonlinear systems, by allowing to
lers, and observers, without any previous knowledge of their extend some linear results to the nonlinear field.
existence (Angelov et al. 2010, 2011, 2015; Leite et al. 2015; Basically, a nonlinear system given by
Meda-Campana et al. 2012; Mota et al. 2018; Precup et al.
2012, 2014).
̇ = f (x(t), u(t)),
x(t) (1)
On the other hand, recently in Meda-Campana et  al. y(t) = h(x(t)), (2)
(2015), an approach to verify the controllability property
with x(t) ∈ ℜ as the state, y(t) ∈ ℜ as the output, and
n m
for a class of TS fuzzy model is given. Nonetheless, such a
u(t) ∈ ℜp as the input, can be approximated, or even
method can only be applied on single-input TS fuzzy mod-
described in an exact way, by a TS fuzzy model, defined
els, with at most two fuzzy rules working at the same instant
by the suitable choice of linear local subsystems, which are
t≥0 . In that work, the controllability property is analyzed in
combined by means of some linguistic characterization of
every fuzzy interpolation region, and the overall fuzzy TS
the nonlinear properties of the system (Tanaka and Wang
fuzzy model is said to be controllable if it is controllable in
2001). For system (1), (2), the corresponding TS fuzzy
all of its regions.
model (Tanaka and Sugeno 1992) is defined by r rules of
Therefore, the main contribution of the present work is
the form:
to overcome such drawbacks by providing overall fuzzy
Rule i
controllability and observability tests for TS fuzzy models,
IF z1 (t) is Mi,1 and … and zq (t) is Mi,q THEN
which can be applied independently of the number of fuzzy
rules activated at the same time, and regardless the number ̇ = Ai x(t) + Bi u(t),
x(t) (3)
of inputs or outputs included in the fuzzy model.
At this point, it is important mentioning that this work y(t) = Ci x(t), (4)
is intended to give a first attempt to test controllability and where Mi,j are the fuzzy sets containing some knowl-
observability properties for TS fuzzy systems without the edge of the dynamics of the nonlinear system,
restriction of having at most two rules activated at the same z(t) = [z1 (t) z2 (t) ⋯ zq (t)] are the premise variables solely
time, in a practical way. The approach has been conceived dependent on x(t) (Tanaka and Wang 2001) and q is the
to be applied off-line and it would result very impractical to number of premise variables per rule, with i = 1, … , r  ,
apply the approach at every instant. In that sense, for evolv- j = 1, … , q , and r as the number of rules in the TS fuzzy
ing TS fuzzy models, it would be recommended applying model.
controllability and/or observability tests when the structure As expected, TS fuzzy models can also be defined by more
of the evolving system suffers a significative change, consid- typical mathematical expressions, i.e.,
ering that such properties will be hardly affected unless the
structure of the overall system has been modified. However, ∑
r
{ }
̇ =
x(t) hi (z(t)) Ai x(t) + Bi u(t) , (5)
the problem of determining when the structure of an evolv- i=1
ing TS fuzzy system suffers a significative change is beyond
the scope of current work.
The rest of the paper is structured as follows. A brief ∑
r
{ }
y(t) = hi (z(t)) Ci x(t) , (6)
reminder of TS fuzzy models is given in Sect. 2. Section 3 i=1
summarizes the test considered in Meda-Campana et al. (2015)

13
Evolving Systems

where, as before, x(t) ∈ ℜn is the state, y(t) ∈ ℜm is the 3 Analysis of the fuzzy controllability
output, and u(t) ∈ ℜp is the control, with and observability property
for single‑input single‑output TS fuzzy
∏
q
models, with at most two fuzzy rules
𝜛i (z(t)) = Mi,j (zj (t)), (7)
j=1 activated at the same instant

and the fuzzy weights given by The linear system

𝜛 (z(t)) ̇ = Ax(t) + Bu(x(t)),

x(t) (13)
hi (z(t)) = ∑r i , (8)
i=1 𝜛i (z(t)) y(t) = Cx(t), (14)
with state as x(t) ∈ ℜn , input as u(t) ∈ ℜp , and output as
for all t ≥ 0 , and where Mi,j (zj (t)) is the membership value y(t) ∈ ℜm , is controllable if the rank of
for Mi,j at zj (t) . Also, because [ ]
 = B AB ⋯ An−1 B , (15)
∑
r
𝜛i (z(t)) > 0, 𝜛i (z(t)) ≥ 0, i = 1, … , r, (9) is n, where  is the controllability matrix, while (13), (14)
i=1 is observable if the rank of

the fuzzy weights satisfy: [ ]T

 = C CA … CAn−1 , (16)
∑
r
is n, where  is the observability matrix (Kailath 1980;
hi (z(t)) = 1, hi (z(t)) ≥ 0, i = 1, … , r, (10) Skogestad and Postlethwaite 2005). In a similar way, the
i=1
fuzzy controllability matrix for system (5), (6) or (11), (12)
for all t ≥ 0 . In the following, hi (x(t)) is used instead of is
hi (z(t)) because, as mentioned above, z(t) depends exclu- [
sively on x(t). (x(t)) = 𝜉0 (x(t)) 𝜉1 (x(t)) 𝜉2 (x(t))
] (17)
For the local subsystems in (5), (6), matrices Ai ∈ ℜn×n , ⋯ 𝜉n−1 (x(t)) ,
Bi ∈ ℜn×p , and Ci ∈ ℜm×n can be obtained by linearizing
the nonlinear system around some suitable operation points where 𝜉0 (x(t)) = B(x(t))
̃ and 𝜉𝓁 (x(t)) = A(x(t))𝜉
̃ 𝓁−1 (x(t))−
(x, u) = (xoi , uoi ) , i.e., 𝜉̇ 𝓁−1 (x(t)) ∀𝓁 = 1, … , n − 1 (Lee and Choi 2005; Meda-
Campana et al. 2015; Tsalakis and Ioannou 1993), which
𝜕f (x, u) 𝜕f (x, u)
Ai = ∣(xo ,uo ) , Bi = ∣(xo ,uo ) , becomes in
𝜕x i i 𝜕u i i

𝜕h(x) [
Ci = ∣ . (x(t)) = 𝜎0 (x(t)) 𝜎1 (x(t)) 𝜎2 (x(t))
𝜕x (xoi ,uoi ) ] (18)
⋯ 𝜎n−1 (x(t)) ,
When linearization around operation points is considered
to obtain the linear local subsystems, the membership func- with 𝜎0 (x(t)) = B(x(t))
̃ and 𝜎𝓁 (x(t)) = (A(x(t)))
̃ ̃
𝓁 B(x(t)) for
tions must be chosen such that the nonlinear dynamics (1), 𝓁 = 1, … , n − 1 , when the fuzzy system can be considered
(2) is sufficiently approximated by the overall TS fuzzy as a slow changing system (Palm et al. 1997).
model (5 ), (6). Sector nonlinearity presented in Tanaka and While, the fuzzy observability matrix for system (5), (6)
Wang (2001) can be applied when an exact description of or (11), (12) is
the nonlinear system is required. Such a method provides
⎡ 𝜂0 (x(t)) ⎤
both, matrices and membership functions for a TS fuzzy ⎢ 𝜂 (x(t)) ⎥
model capable of exactly describe, at least locally, the non- ⎢ 1 ⎥
linear system (1), (2). Also, notice that system (5), (6) can (x(t)) = ⎢ 𝜂2 (x(t)) ⎥, (19)
⎢ ⋮ ⎥
be rewritten as ⎢ 𝜂 (x(t)) ⎥
⎣ n−1 ⎦
̃
̇ = A(x(t))x(t)
x(t) ̃
+ B(x(t))u(x(t)), (11)
̃
y(t) = C(x(t))x(t), (12) where 𝜂0 (x(t)) = C(x(t))
̃ and 𝜂𝓁 (x(t)) = 𝜂𝓁−1 (x(t))A(x(t))−
̃
∑r ∑r
with A(x(t)) = i=1 hi (x(t))Ai , B(x(t)) = i=1 hi (x(t))Bi , and
̃ ̃ 𝜂̇ 𝓁−1 (x(t)) for 𝓁 = 1, … , n − 1 , becoming in
∑r
̃
C(x(t)) = i=1 hi (x(t))Ci (Tanaka and Wang 2001).

13
 Evolving Systems

⎡ 𝜁0 (x(t)) ⎤ (20) are square matrices. However, from (5), (6), or

⎢ 𝜁 (x(t)) ⎥ (11), (12), it can be deduced that (x(t)) ∈ ℜn×np and
⎢ 1 ⎥ (x(t)) ∈ ℜnm×n can be rectangular matrices, in gen-
(x(t)) = ⎢ 𝜁2 (x(t)) ⎥, (20)
⎢ ⋮ ⎥ eral. This problem can be sorted out by the iterative
⎢ 𝜁 (x(t)) ⎥ process of extracting n × n square sub-matrices from
⎣ n−1 ⎦
(x(t)) or (x(t)) , allowing the use of the determinant
during the testing of controllability, and/or observabil-
with 𝜁0 (x(t)) = C(x(t))
̃ and 𝜁𝓁 (x(t)) = C(x(t))(
̃ ̃
A(x(t)))𝓁 for ity properties. In the worst case, the number of n × n
𝓁 = 1, … , n − 1 , when the fuzzy system can be considered square sub-matrices, extracted from (x(t)) , that must
as a slow changing system. be investigated is:
( )
nm (nm)!
Remark 1  The computation of (17) and (19) can be difficult n
= , (21)
n!(nm − n)!
when the relative degree between the premise variable and
the input is less than n, or when non-smooth membership while
functions appear in the TS fuzzy model. ( )
np (np)!
n
= (22)
Notice that for single-input single-output (SISO) TS n!(np − n)!
fuzzy models, matrices (17)–(20) are square matrices. is the greater number of n × n square sub-matrices that
Therefore, in those cases, the analysis of controllability can be extracted from (x(t)).
and/or observability properties consists of determining if • Problem (2) Controllability and/or observability proper-
respective matrices are non singular, for any valid value ties can only be analyzed in terms of only one of fuzzy
of the membership functions. Thus, for TS fuzzy models weights, hi (x(t)) , with i = 1, … , r  , included in the TS
with at most two fuzzy rules activated at the same instant fuzzy model. This is because just one equation is avail-
t ≥ 0 , each one of the two fuzzy weights, hj (x(t) or hk (x(t)) , able to do so. Such equation is the one formed when the
involved in the interpolation region jk, between linear sub- determinant of (x(t)) or (x(t)) is equated to zero. For
systems j and k, can be expressed in terms of the other one, this reason, the aforementioned method requires that at
i.e., h (x(t)) = 1 − hj (x(t)) or hj (x(t)) = 1 − hk (x(t)) , because most two fuzzy rules are activated at the same time, and
∑r k
h (x(t)) = 1 , ∀t , where j = 1, … , r − 1 , k = j + 1, … , r ,
i=1 i as consequence, the test must be performed in each one
and hj ∩ hk ≠ � . As a result, in the interpolation region jk, of the fuzzy interpolation regions, separately. In fact, this
matrices (17)–(20) can be expressed in terms of one and is the main drawback of the approach discussed above.
only one fuzzy weight h∗ (x(t)) , where “ ∗ ” can be either j or In the next section, this problem is solved in a straight-
k, and the roots of det(j,k (x(t))) = 0 , or det(j,k (x(t))) = 0 , forward way.
can be easily found by means of mathematical software
as Matlab, Mathematica, Octave, etc., where j,k (x(t) and
j,k (x(t) are the controllability and observability matrices
for region jk, respectively. Clearly, the total number of 4 General controllability and observability
interpolation regions, in a TS fuzzy system with at most tests for TS fuzzy models
two rules activated, at the same instant is always r − 1.
So, from (7)–(10), the TS fuzzy model (5), (6) or (11), In this section, fuzzy controllability and fuzzy observability
(12) is fuzzy controllable or fuzzy observable in region jk analysis will be performed directly on the overall matrices
if the roots of the corresponding determinant, expressed (17)–(20 ), avoiding the region-by-region analysis presented
in terms of hj (x(t)) or hk (x(t)) , are outside of [0 1] . Finally, in Meda-Campana et al. (2015). To this end, the following
the overall TS fuzzy model is fuzzy controllable, or fuzzy definitions are needed:
observable, if it is controllable, or observable, in every
interpolation region. The detailed approach, from the con- Definition 1  A TS fuzzy system, described by Eqs. (5),
trollability point of view, can be found in Meda-Campana (6) or (11), (12), is fuzzy controllable if its associated con-
et al. (2015). trollability matrix, given by (17) or (18), has full rank for
At this point, two drawbacks of the previously discussed any valid value of the fuzzy weights included in the fuzzy
approach can be readily identified: system, i.e., a TS fuzzy system is fuzzy controllable if
rank(x(t)) = n , t ≥ 0.
• Problem (1) The approach is only valid for SISO TS
fuzzy models, i.e., it can only be applied when (17)– Definition 2  A TS fuzzy system, described by Eqs. (5),
(6) or (11), (12), is fuzzy observable if its associated

13
Evolving Systems

observability matrix, given by (19) or (20), has full rank equations in order to verify the rank of  and  for all t ≥ 0 ,
for any valid value of the fuzzy weights included in the whatever the case. The examples in the next section are
fuzzy system, i.e., a TS fuzzy system is fuzzy observable if intended to illustrate the use of the approach described so
rank(x(t)) = n , t ≥ 0. far.

Now, in order to analyze controllability and/or observ- Remark 2  In the adaptive systems, the parameters are
ability properties in terms of the r fuzzy weights at once, adapted while the structure is not adapted, i.e. the param-
an equal number of equations must be available. In this eters of the rules change while the rules number is constant.
work, the inclusion of (10) is proposed during the for- Consequently, the adaptive systems have the same rules
mulation of the desired set of simultaneous equations. number as the TS fuzzy model, and the Definitions 1 and 2
Such an approach is described below, by considering are applied for the controllability and observability tests of
hi ≡ hi (x(t)), ∀i = 1, 2, … , r ,  ≡ (x(t)) , and  ≡ (x(t)). adaptive systems.
The following relations can be readily obtained from (10):
Remark 3  In the evolving systems, the parameters are
1 − h1 − h2 − ⋯ − hr = 0, (23) adapted while the structure also is adapted, i.e. the param-
1 − h2 − h3 − ⋯ − hr = h1 , (24) eters of the rules change while the rules number also change.
Consequently, the evolving system must be divided in sev-
1 − h1 − h3 − ⋯ − hr = h2 , (25) eral evolving subsystems for which in this time interval they
have a constant rules number, and the Definitions 1 and 2
⋮
(26) must be applied for the controllability and observability tests
1 − h1 − h2 − ⋯ − hr−1 = hr . of each evolving subsystem.
On the other hand, from (17)–(20), it results obvious that
det() and det() are functions of hi , ∀i = 1, 2, … , r . But, if 5 Examples
h1 is substituted by expression (24) in (17)–(20), then deter-
minants det() and det() will depend on h2 , h3 , … , hr−1 Consider the system given in Tanaka and Wang (2001) (pp.
and hr  . In a similar way, the substitution of h2 by (25) 93):
in (17)–(20) causes that det() and det() depend on ( )
ẋ 1 (t) = x2 (t) + sin(x3 (t)) + x12 (t) + 1 u(t),
h1 , h3 , … , hr−1 and hr , and so on, until reaching the substitu-
tion of hr by (26) in (17)–(20), such that determinants det() ẋ 2 (t) = x1 (t) + 2x2 (t),
and det() turn into functions of h1 , h2 , … , hr−2 and hr−1 . ẋ 3 (t) = x12 (t)x2 (t) + x1 (t),
With that in mind, consider det1 (∗) as the determinant where ẋ 4 (t) = sin(x3 (t)),
the weight h1 has been removed by using (24), det2 (∗) as ( )
the determinant where the weight h2 has been removed by y1 (t) = x12 (t) + 1 x4 (t) + x2 (t)
using (25), and so forth, where * is  or  , as the case may y2 (t) = x2 (t) + x3 (t),
be, with det() and det() depending on the complete set of
fuzzy weights h1 , h2 , … , hr.
Now, r + 1 equations are available to find r unknowns where x1 (t) ∈ [−a a] , x3 (t) ∈ [−b b] , with a and b as posi-
(the fuzzy weights) and to analyze the controllability and/ tive values. Using sector nonlinearity, the nonlinear terms
or observability properties of the TS fuzzy model. They are: can be expressed as:
det(∗) = 0, (27) x12 (t) = N1,1 (x1 (t)) ⋅ a2 + N1,2 (x1 (t)) ⋅ 0,
det1 (∗) = 0, (28) sin(x3 (t)) = N2,1 ⋅ 1 ⋅ x3 (t)
sin(b)
det2 (∗) = 0, (29) + N2,2 (x3 (t)) ⋅
b
⋅ x3 (t),

⋮
(30) where
detr (∗) = 0.

In this work, it is suggested to use (27) with a combination N1,1 (x1 (t)), N1,2 (x1 (t)), N2,1 (x3 (t)), N2,2 (x3 (t)) ∈ [0 1],
of r − 1 of the remaining equations. The rationale behind N1,1 (x1 (t)) + N1,2 (x1 (t)) = 1,
this suggestion is that (27) involves all of the fuzzy weights. N2,1 (x3 (t)) + N2,2 (x3 (t)) = 1.
Then, mathematical software as Matlab, Mathematica,
Octave, etc., can be used to solve the set of simultaneous

13
 Evolving Systems

Solving for N1,1 (x1 (t)) , N1,2 (x1 (t)) , N2,1 (x3 (t)) , and N2,2 (x3 (t)) , with i = 1, … , r  , x(t) ∈ ℜ4 , y(t) ∈ ℜ2 , u(t) ∈ ℜ1 , a = 0.8 ,
one has b = 0.6 and
2
x2 (t) x2 (t) ⎡0 1 0⎤1 ⎡1+a ⎤
N1,1 (x1 (t)) = 1 2 , N1,2 (x1 (t)) = 1 − 1 2 , ⎢1 2 0 ⎥
0 ⎢ 0 ⎥
{ a a A1 =⎢ ⎥, B1 = ⎢ ⎥,
b sin(x3 (t))−sin(b)x3 (t) ⎢1 a2 0⎥0 ⎢ 0 ⎥
, x3 (t) ≠ 0 ⎣0 0 0⎦1 ⎣ 0 ⎦
N2,1 (x3 (t)) = x3 (t)(b−sin(b))
1, x3 (t) = 0, � 2
�
0 1 0 1+a
C1 = ,
0 11 0
and ⎡0 1 0⎤
sin(b) 2
⎡1+a ⎤
{ ⎢ b ⎥ ⎢ 0 ⎥
1 2 0 0⎥
A2 =⎢
b(x3 (t)−sin(x3 (t)))
0 0 ⎥ 2 ⎢⎢ 0 ⎥⎥
, x3 (t) ≠ 0 ,B = ,
N2,2 (x3 (t)) = x3 (t)(b−sin(b)) ⎢1 a2
0, x3 (t) = 0. ⎢0 0 sin(b)
0 ⎥⎦ ⎣ 0 ⎦
⎣ b
� �
Thus, by defining the fuzzy sets 0 1 0 1 + a2
C2 = ,
0 11 0
M1,1 (x1 (t)) = M2,1 (x1 (t)) = N1,1 (x1 (t)),
⎡0 1 0⎤ 1 ⎡1⎤
M3,1 (x1 (t)) = M4,1 (x1 (t)) = N1,2 (x1 (t)), ⎢1 2 0⎥ 0 ⎢0⎥
0 ⎥⎥ 3 ⎢⎢ 0 ⎥⎥
A3 =⎢ ,B = ,
M1,2 (x3 (t)) = M3,2 (x3 (t)) = N2,1 (x3 (t)),
⎢1 0 0
M2,2 (x3 (t)) = M4,2 (x3 (t)) = N2,2 (x3 (t)), ⎣0 0 0⎦ 1 ⎣0⎦
� �
0 101
C3 = ,
which are depited in Fig. 1, the nonlinear system can be 0 110
described by the following TS fuzzy model: ⎡0 1 sin(b)
0⎤ ⎡1⎤
Rule i ⎢ b ⎥ ⎢0⎥
1 2 0 0⎥
IF x1 (t) is Mi,1 and x3 (t) is Mi,2 THEN A4 =⎢ , B4 = ⎢ ⎥,
⎢1 0 0 0⎥ ⎢0⎥
⎢0 0 sin(b)
0 ⎥⎦ ⎣0⎦
̇ = Ai x(t) + Bi u(t),
x(t) ⎣ b
� �
y(t) = Ci x(t), 0 101
C4 = .
0 110

The fuzzy weights h1 (x(t)) , h2 (x(t)) , h3 (x(t)) and h4 (x(t)) ,

∑4
with i hi (x(t)) = 1 , can be readily obtained from (7)–(10).
They are individually depicted in Fig. 2, while their com-
bined work can be found in Fig. 3.
From Figs. 2 and 3, it can be concluded that more than
two fuzzy rules are activated at the same time. Besides, local
matrices Ci ∈ ℜ2×4 , with i = 1, … , 4 , correspond to multiple
output linear systems. So, a non-square fuzzy observability
matrix will be obtained from (19), (20). Consequently, the
approach given in Meda-Campana et al. (2015) can be used
to analyze neither controllability nor observability of the TS
fuzzy model. Next, such properties will be investigated by
means of the approach given in the previous section.

5.1 Controllability analysis

In order to simplify the analysis, TS fuzzy model is

assumed as a slow changing system. Therefore, the con-
trollability matrix, (x(t)) can be obtained from (18). Then,
Fig. 1  Fuzzy sets

13
Evolving Systems

controllability property can be analyzed by considering the

following set of simultaneous equations:

det((x(t))) =0,
det1 ((x(t))) =0,
(31)
det2 ((x(t))) =0,
det3 ((x(t))) = 0,

with det((x(t))) , det1 ((x(t))) , det2 ((x(t))) , det3 ((x(t))) as

in (27)–(30), and where the unknowns are h1 (x(t)) , h2 (x(t)) ,
∑4
h3 (x(t)) , h4 (x(t)) ∈ [0 1] , with i=1 hi (x(t)) = 1 , ∀t ≥ 0 . Mat-
lab is used to find the solution of (31), subject to (10).
Because no solution, hi (x(t)) ∈ [0 1] with i = 1, … , 4 ,
can be found for (31), it can be concluded that matrix (18),
(x(t)) , has full rank for all t ≥ 0 , and consequently, by Defi-
nition 1, the TS fuzzy model is fuzzy controllable.
Consider
[ a similar
]T TS fuzzy model but now with Fig. 3  Fuzzy weights working together
B3 = −1 0 0 0  . In this case, at least two solutions for
(31) can be found. They are:
reason, it can be concluded that, at least at that point, the
1. h1 (x(t)) ≈ 0.378   , h2 (x(t)) = 0   , h3 (x(t)) ≈ 0.621  , TS fuzzy model is noncontrollable, and the neighborhood
h4 (x(t)) = 0 , and around such fuzzy weights should be carefully addressed
2. h1 (x(t)) = 0 , h2 (x(t)) = 0 , h3 (x(t)) = 0.5 , h4 (x(t)) = 0.5. during the control design.

Clearly, both of them satisfy hi (x(t)) ∈ [0 1] with i = 1, … , 4 , 5.2 Observability analysis

∑4
and i=1 hi (x(t)) = 1 , ∀t ≥ 0 . However, in the former, the
value of h3 (x(t)) lies outside the valid ranges shown in Again, the TS fuzzy model is considered as a slow changing
Figs. 2 and 3. Therefore, this situation will not occur. system, which allows to obtain the fuzzy observability matrix,
On the other hand, the latter solution lies within the (x(t)) , from (20). Notice that (x(t)) ∈ ℜ8×4 is a rectangu-
region of valid weights depicted in Figs. 2 and 3. For that lar matrix. Therefore, the determinant approach can not be
directly applied to analyze the rank of (x(t)) . Instead, n × n
square sub-matrices from (x(t)) must be considered during
the observability analysis. As mentioned before, the worst case
is the one where the total number of combinations given by
(21) is reached. Thus, consider 1,2,3,4 (x(t)) as the square sub-
matrix formed by rows 1, 2 3, and 4 of (x(t)) . Now, the same
procedure used previously to check the controllability property
can be used to analyze the rank of 1,2,3,4 (x(t)) . To this end,
the following set of simultaneous equations can be considered:
det(1,2,3,4 (x(t))) =0,
det1 (1,2,3,4 (x(t))) =0,
(32)
det2 (1,2,3,4 (x(t))) =0,
det3 (1,2,3,4 (x(t))) =0,

with det(1,2,3,4 (x(t))) , det1 (1,2,3,4 (x(t))) , det2 (1,2,3,4 (x(t))) ,

det3 (1,2,3,4 (x(t))) as in (27)–(30), and where the unknowns
are h (x(t)) , h2 (x(t)) , h3 (x(t)) , h4 (x(t)) ∈ [0 1] , with
∑4 1
h (x(t)) = 1 , ∀t . As before, Matlab is used to solve (32)
i=1 i
subject to (10). In this case, no solution, hi (x(t)) ∈ [0 1] with
Fig. 2  Fuzzy weights i = 1, … , 4 , can be found for (32), and it can be already

13
 Evolving Systems

concluded that matrix (20), (x(t)) , has full rank for all in radians per second). The state variables, x1 , x3 , and x5
t ≥ 0 , and consequently, by Definition 2, the TS fuzzy model are in meters and they represent the linear displacements
is fuzzy observable, and an approach as the one presented along axes x, y and z, respectively. While x7 , x9 , and x11
in Meda-Campana et al. (2018) can be used to construct an are in radians and they describe the angular displacements.
observer for the considered fuzzy system. The parameters considered are: b = 54.2 × 10−6 N rms2 as
the thrust factor, d = 1.1 × 10−6 N m s2 as the drag factor,
5.3 Case 2 l = 0.24  m as the distance from the center of the quad-
rotor to the center of the rotors, m = 1 kg as the mass of
Now, consider the model of the quadrotor given in Meda- the quadrotor, g = 9.8 sm2 as the acceleration of gravity,
Campana (2018): Jtp = 104 × 10−6 N m s2 as the total moment of inertia for the
rotors, and Ixx = 8.1 × 10−3 N m s2 , Iyy = 8.1 × 10−3 N m s2 ,
ẋ = f (x, u), (33) and Izz = 14.2 × 10−3 N m s2 , as the moments of inertia
y = g(x), (34) respect to axes x, y and z, respectively.
Obviously, the procedure used in the previous example
where [ ]T to obtain the TS fuzzy model results impractical in this case
x = x1 … x12  , u = [u1 … u4 ]T ,
because of the great number of nonlinearities involved in
f (x, u) =
the model. Therefore, a simplified TS fuzzy model based on
⎡ x2 , ⎤ the linearization around some operation points is advised.
⎢ (sin(x ) sin(x ) + cos(x ) sin(x ) cos(x )) 𝛽1 , ⎥ Two linearizations around the following operation points:
⎢ 11 7 11 9 7 m ⎥
⎢ x4 , ⎥
⎢ (− cos(x ) sin(x ) + sin(x ) sin(x ) cos(x )) 𝛽1 , ⎥ ⎡ 0 ⎤ ⎡ 0 ⎤
⎢ 11 7 11 9 7 m ⎥ ⎢ 0 ⎥ ⎢ 0 ⎥
⎢ x6 , ⎥ ⎢ 0 ⎥ ⎢ 0 ⎥
⎢ 𝛽 ⎥ ⎢ ⎥ ⎢ ⎥
⎢
−g + (cos(x9 ) cos(x7 ) m1 ,
⎥ ⎢ 0 ⎥ ⎢ 0 ⎥
⎢ x8 , ⎥, (35) ⎢ 0 ⎥ ⎢ 0 ⎥
⎢ Iyy −Izz Jtp l𝛽2 ⎥ ⎢ ⎥ ⎢ ⎥
⎢ x10 x12 I − I x10 Ω + I , ⎥ xop1 =⎢
0 ⎥ and xop2 = ⎢ 0 ⎥, where
⎢
xx xx xx
⎥ ⎢ x7− ⎥ ⎢ x7+ ⎥
x10 , ⎢ 0 ⎥ ⎢ 0 ⎥
⎢ ⎥
⎢
Izz −Ixx Jtp
x8 x12 I + I x8 Ω + I ,
l𝛽3
⎥ ⎢ ⎥ ⎢ ⎥
⎢ yy yy yy
⎥ ⎢ 0 ⎥ ⎢ 0 ⎥
⎢ x 12 , ⎥ ⎢ 0 ⎥ ⎢ 0 ⎥
⎢ Ixx −Iyy 𝛽4 ⎥ ⎢ 0 ⎥ ⎢ 0 ⎥
⎣ x x
8 10 I + Izz ⎦ ⎢ ⎥ ⎢ ⎥
zz
⎣ 0 ⎦ ⎣ 0 ⎦

with x7− = −10𝜋

180
radians and x7+ = 10𝜋 180
radians. The proposed
fuzzy weights are: ( )
𝛽1 = b(u21 + u22 + u23 + u24 ), x7 −x7
h1 (x7 ) = 21 + 12 cos x −x 1 and h2 (x7 ) = 1 − h1 (x7 ) ,
7+ 71
𝛽2 = b(u24 + u23 − u21 − u22 ),
which are depicted in Fig. 4.
𝛽3 = b(u22 + u23 − u21 − u24 ), (36) As consequence, the nonlinear system (33)–(37) is approxi-
𝛽4 = d(u21 + u23 − u22 − u24 ), mated by the following TS fuzzy model:
Ω = u1 − u2 + u3 − u4 ,
∑
2
{ }
ẋ = Ai x + Bi u , (38)
i=1
and
[ ]T ∑
2
g(x) = x1 x3 x5 x7 x9 x11 , (37) y= Ci x, (39)
where the effective control inputs u1 , u2 , u3 , and u4 are
i=1

the frequency of rotors 1, 2, 3, and 4 respectively (given with

13
Evolving Systems

⎡0 1.0 0 0 0 0 0 0 0 0 0 0 ⎤
⎢0 0 0 0 0 0 0 0 9.8 0 − 1.7 0 ⎥
⎢0 0 0 1.0 0 0 0 0 0 0 0 0 ⎥⎥
⎢
⎢0 0 0 0 0 0 − 9.8 0 0 0 0 0 ⎥
⎢0 0 0 0 0 1.0 0 0 0 0 0 0 ⎥
⎢ ⎥
0 0 0 0 0 0 1.7 0 0 0 0 0 ⎥
A1 =⎢
⎢0 0 0 0 0 0 0 1.0 0 0 0 0 ⎥
⎢0 0 0 0 0 0 0 0 0 0 0 0 ⎥
⎢ ⎥
⎢0 0 0 0 0 0 0 0 0 1.0 0 0 ⎥
⎢0 0 0 0 0 0 0 0 0 0 0 0 ⎥
⎢0 0 0 0 0 0 0 0 0 0 0 1.0 ⎥
⎢ ⎥
⎣0 0 0 0 0 0 0 0 0 0 0 0 ⎦
⎡0 1.0 0 0 0 0 0 0 0 0 0 0 ⎤
⎢0 0 0 0 0 0 0 0 9.8 0 1.7 0 ⎥
⎢0
⎢ 0 0 1.0 0 0 0 0 0 0 0 0 ⎥⎥
⎢0 0 0 0 0 0 − 9.8 0 0 0 0 0 ⎥
⎢0 0 0 0 0 1.0 0 0 0 0 0 0 ⎥
⎢ ⎥
0 0 0 0 0 0 − 1.7 0 0 0 0 0 ⎥
A2 =⎢ , Fig. 4  Fuzzy weights
⎢0 0 0 0 0 0 0 1.0 0 0 0 0 ⎥
⎢0 0 0 0 0 0 0 0 0 0 0 0 ⎥ and
⎢ ⎥
⎢0 0 0 0 0 0 0 0 0 1.0 0 0 ⎥
⎡1 0 0 0 0 0 0 0 0 0 0 0⎤
⎢0 0 0 0 0 0 0 0 0 0 0 0 ⎥ ⎢0 0⎥
⎢0 1.0 ⎥
0 1 0 0 0 0 0 0 0 0
⎢
0 0 0 0 0 0 0 0 0 0
⎥ ⎢ ⎥
0 0 0 0 1 0 0 0 0 0 0 0⎥
⎣0 0 0 0 0 0 0 0 0 0 0 0 ⎦ C1 = C2 = ⎢ .
⎢0 0 0 0 0 0 1 0 0 0 0 0⎥
⎡ 0 0 0 0 ⎤ ⎢0 0 0 0 0 0 0 0 1 0 0 0⎥
⎢ 0 0 0 0 ⎥ ⎢ ⎥
⎢ ⎥ ⎣0 0 0 0 0 0 0 0 0 0 1 0⎦
⎢ 0 0 0 0 ⎥
−3 −3 −3 −3
⎢ 4.0 ⋅ 10 4.0 ⋅ 10 4.0 ⋅ 10 4.0 ⋅ 10 ⎥
⎢ 0 0 0 0 ⎥ 5.3.1 Observability test
⎢ ⎥
0.023 0.023 0.023 0.023 ⎥
B1 =⎢ ,
⎢ 0 0 0 0 ⎥ Because it is well-known that this model is controllable only
⎢ −0.69 − 0.69 0.69 0.69 ⎥ the observability test will be performed. So, from matrices
⎢ ⎥
⎢ 0 0 0 0 ⎥ involved in the TS fuzzy model of the quadrotor, it can be
⎢ −0.69 0.69 0.69 − 0.69 ⎥ readily obtained that the first 12 rows of the observability
⎢ ⎥
⎢
0 0 0 0
⎥ matrix  ∈ ℜ72×12 , ∀t form the square matrix:
⎣ 0.033 − 0.033 0.033 − 0.033 ⎦
⎡1 0 0 0 0 0 0 0 0 0 0 0⎤
⎡ 0 0 0 0 ⎤ ⎢0 0 1 0 0 0 0 0 0 0 0 0⎥
⎢ ⎥ ⎢0 0 ⎥⎥
0 0 0 0
⎢ ⎥ ⎢ 0 0 0 1 0 0 0 0 0 0
⎢ 0 0 0 0 ⎥ ⎢0 0 0 0 0 0 1 0 0 0 0 0⎥
⎢ −4.0 ⋅ 10−3 − 4.0 ⋅ 10−3 − 4.0 ⋅ 10−3 − 4.0 ⋅ 10−3 ⎥ ⎢0 0 0 0 0 0 0 0 1 0 0 0⎥
⎢ 0 0 0 0 ⎥ ⎢ ⎥
⎢ ⎥ ⎢0 0 0 0 0 0 0 0 0 0 1 0⎥
0.023 0.023 0.023 0.023 ,
B2 =⎢ ⎥, ⎢0 1 0 0 0 0 0 0 0 0 0 0⎥
⎢ 0 0 0 0 ⎥ ⎢0 ⎥
⎢ −0.69 ⎥ 0 0 1 0 0 0 0 0 0 0 0
− 0.69 0.69 0.69 ⎢ ⎥
⎢ ⎥ ⎢0 0 0 0 0 1 0 0 0 0 0 0⎥
⎢ 0 0 0 0 ⎥ ⎢0 0 0 0 0 0 0 1 0 0 0 0⎥
⎢ −0.69 0.69 0.69 − 0.69 ⎥
⎢ ⎥
⎢0 0 0 0 0 0 0 0 0 1 0 0⎥
⎢
0 0 0 0
⎥ ⎢ ⎥
⎣ 0.033 − 0.033 0.033 − 0.033 ⎦
⎣0 0 0 0 0 0 0 0 0 0 0 1⎦

13
 Evolving Systems

which has full rank, and it is independent of the fuzzy tracking in video streams based on recursive density estimation
weights. In other words, the TS fuzzy model is fuzzy and evolving Takagi-Sugeno fuzzy systems. Int J Intell Syst
26:189–205
observable. Angelov P, Skrjanc I, Blazic S (2015) A robust evolving cloud-based
controller. In: Kacprzyk J, Pedrycz W (eds) Springer handbook of
computational Intelligence. Springer, pp 1435–1449
Baruah RD, Angelov P (2012) Evolving local means method for clus-
6 Conclusion tering of streaming data. In: 2012 IEEE International Conference
on Fuzzy Systems, pp 1–8
Baruah RD, Angelov P (2014) Dynamically evolving clustering and its
A general approach to investigate controllability and observ- application to structure identification of evolving fuzzy models.
ability properties for TS fuzzy systems has been developed IEEE Trans Cybern 44(9):1619–1631
on the basis of a set of symbolic simultaneous equations, Kailath T (1980) Linear systems. Prentice-Hall Inc, Upper Saddle River
where the fuzzy weights are the unknowns of such equa- Khalil HK (1996) Nonlinear systems. Prentice Hall, Upper Saddle
River
tions. The main advantages of the proposed approach are: Lee HC, Choi JW (2005) Ackermann-like eigenvalues assigment for-
(1) the analysis can be carried out regardless the number mulae for linear time-varying systems. IEE Proc Control Theory
of fuzzy weights acting at the same instant, (2) the dimen- Appl 152(4):427–434
sions of input and output vectors are not relevant for the Leite D, Palhares RM, Campos VCS, Gomide F (2015) Evolving gran-
ular fuzzy model-based control of nonlinear dynamic systems.
proposed approach. However, as in most of the approaches IEEE Trans Fuzzy Syst 23(4):923–938
for TS fuzzy models, a large number of fuzzy rules may Meda-Campana JA (2018) On the estimation and control of nonlinear
increase the complexity of the problem. The applicability of systems with parametric uncertainties and noisy outputs. IEEE
the approach has been illustrated by means of an interesting Access 6:31968–31973
Meda-Campana JA, Gomez-Mancilla JC, Castillo-Toledo B (2012)
TS fuzzy model. Finally, the use of the presented tests may Exact output regulation for nonlinear systems described by Tak-
help to develop feedback stabilizers and observers for TS agi-Sugeno fuzzy models. IEEE Trans Fuzzy Syst 20(2):235–247
fuzzy models in a less conservative way. Meda-Campana JA, Rodriguez-Valdez J, Hernandez-Cortes T, Tapia-
Herrera R, Nosov V (2015) Analysis of the fuzzy controllability
Acknowledgements  Authors are grateful with the Editor-in-Chief, property and stabilization for a class of t-s fuzzy models. IEEE
Associate Editor, and Reviewers for their valuable comments and Trans Fuzzy Syst 23(2):291–301
insightful suggestions, which helped to improve this research signifi- Meda-Campana JA, Araceli G, Rubio JJ, Tapia-Herrera R, Hernandez-
cantly. Authors thank the Instituto Politécnico Nacional, the Secretaría Cortes T, Curtidor-Lopez AV, Paramo-Carranza LA, Cazares-
de Investigación y Posgrado, the Comisión de Operación y Fomento Ramirez IO (2018) Design of stabilizers and observers for a class
de Actividades Académicas, and the Consejo Nacional de Ciencia y of multivariable ts fuzzy models on the basis of new interpolation
Tecnología for their help in this research. functions. IEEE Trans Fuzzy Syst 26(5):2649–2662
Mota VC, Damasceno FA, Leite DF (2018) Fuzzy clustering and fuzzy
validity measures for knowledge discovery and decision making
Compliance with ethical standards  in agricultural engineering. Comput Electron Agric 150:118–124
Palm R, Driankov D, Hellendoorn H (1997) Model based fuzzy con-
Ethical approval  This article does not contain any studies with human trol: fuzzy gain schedulers and sliding mode fuzzy controllers.
participants or animals performed by any of the authors. Springer, Berlin
Precup R-E, Tomescu ML, Radac M-B, Petriu EM, Preitl S, Dragos
C-A (2012) Iterative performance improvement of fuzzy control
systems for three tank systems. Expert Syst Appl 39:8288–8299
Precup R-E, Filip H-I, Radac M-B, Petriu EM, Preitl S, Dragos C-A
(2014) Online identification of evolving Takagi-Sugeno-Kang
References fuzzy modelsfor crane systems. Appl Soft Comput 24:1155–1163
Skogestad S, Postlethwaite I (2005) Multivariable feedback control,
Angelov P, Kasabov NK (2005) Evolving computational intelligence analisis and design. Wiley, New York
systems. In: Proceedings of the 1st international workshop on Tanaka K, Sugeno M (1992) Stability analysis and design of fuzzy
genetic fuzzy systems, pp 76–82 control systems. Fuzzy Sets Syst 45(2):135–156
Angelov P, Yager R (2013) Density-based averaging—a new operator Tanaka K, Wang HO (2001) Fuzzy control systems design and analysis.
for data fusion. Inf Sci 222:163–174 (Including Special Section A linear matrix inequality approach. Wiley, New York
on New Trends in Ambient Intelligence and Bio-inspired Systems) Tsalakis KS, Ioannou PA (1993) Linear time-varying systems, control
Angelov P, Victor J, Dourado A, Filev D (2004) On-line evolution and adaptation. Prentice Hall, Englewwod Cliffs
of Takagi-Sugeno fuzzy models. In: 2nd IFAC Workshop on
Advanced Fuzzy/Neural Control (AFNC04) IFAC Proceedings Publisher’s Note Springer Nature remains neutral with regard to
vol 37, no (16), Oulu, Finland, September, 16-17, pp 67–72 jurisdictional claims in published maps and institutional affiliations.
Angelov P, Filev D, Kasabov N (2010) Evolving intelligent systems-
methodology and applications. Wiley, New York
Angelov P, Sadeghi-Tehran P, Ramezani R (2011) An approach to
automatic real-time novelty detection, object identification, and

13