Documente Academic
Documente Profesional
Documente Cultură
Abstract— The aim of this article is to investigate the trajec- on the control design to realize an active compensation
tory tracking problem of systems with uncertain models and of MUEPs. The adaptive approximation of uncertain
state restrictions using differential neural networks (DNNs). The functions employing a nonparametric modeling structure
adaptive control design considers the design of a nonparametric
identifier based on a class of continuous artificial neural networks [such as artificial neural networks (ANNs)] offers a reliable
(ANNs). The design of adaptive controllers used the estimated option to obtain the approximation of the MUEP based on
weights on the identifier structure yielding a compensating the states or output measurements [1]. Traditionally, ANNs
structure and a linear correction element on the tracking error. are key tools for developing pattern classification algorithms,
The stability of both the identification and tracking errors, image processing methods, nonparametric models of complex
considering the DNN, uses a barrier Lyapunov function (BLF)
that grow to infinity whenever its arguments approach some finite biological systems, and automatic adaptive controls [2]–[4].
limits for the state satisfying some predefined ellipsoid bounds. Nevertheless, the ANN ability to approximate complex
The analysis guarantees the semi-globally uniformly ultimately functional relations between time-dependent variables justify
bounded (SGUUB) solution for the tracking error, which implies their use within the design of adaptive disturbance rejection
the achievement of an invariant set. The suggested controller controllers.
produces closed-loop bounded signals. This article also presents
the comparison between the tracking states forced by the adaptive Usually, differential neural networks (DNNs) are the par-
controller estimated with the DNN based on BLF and quadratic ticular name given to ANNs, which may approximate MUEP
Lyapunov functions as well. The effectiveness of the proposal is affecting dynamic systems. DNNs have good approximation
demonstrated with a numerical example and an implementation (nonparametric identification) properties in the presence of
in a real plant (mass-spring system). This comparison confirmed bounded or locally Lispchitz MUEPs. In consequence, DNNs
the superiority of the suggested controller based on the BLF
using the estimates of the upper bounds for the system states. are useful instruments to solve diverse control problems for
a large class of uncertain real dynamic plants [5]–[7]. In the
Index Terms— Barrier Lyapunov functions (BLFs), differential context of systems theory, a main characteristic of DNN is the
neural network (DNNs), neuro-identification, state constraint
systems, trajectory tracking. possibility of using the Lyapunov stability theory to justify
the MUEPs approximation quality. Narendra et al. [8] pre-
sented one of the first stability analyses (Lyapunov based) for
I. I NTRODUCTION DNNs working as nonparametric identifiers for uncertain sys-
tems. Such analysis yields the adjustment rules for the DNN
M ODELING uncertainties and external perturba-
tions (MUEPs) provoke considerable troubles in
any closed-loop control realization for uncertain dynamic
weights (learning laws) based on the so-called dynamic back-
propagation. Subsequently, Rovithakis and Christodoulou [9]
systems. Indeed, such uncertainties are sources of instability used the Lyapunov’s second method to obtain the adaptive
or poor performance for the controlled plant. A feasible learning laws, which guaranteed the ultimate boundedness for
method for limiting the effect of MUEP is designing the identification error. Most of these Lyapunov functions
a feedback controller, which may reject them by direct are quadratic (QLFs). Eventually, diverse Lyapunov-based
compensation. Disturbance rejection is the common name approaches derived new adaptive learning laws for DNNs
for such controllers. An online disturbance estimator embeds applied as adaptive nonparametric identifiers, state observers,
or adaptive controllers [10]–[13]. Most of the Lyapunov func-
Manuscript received February 12, 2019; revised July 8, 2019 and tions were quadratic forms of the identification and the weight
September 30, 2019; accepted January 11, 2020. This work was supported by errors as well, which corresponds to the deviation between
the Instituto Politécnico Nacional under Grant sip 20201286. (Corresponding
author: Rita Q. Fuentes-Aguilar.) the optimal (in some sense) and the online time-dependent
Rita Q. Fuentes-Aguilar is with the Tecnológico de Monterrey at Campus estimated weights. Although DNNs are mature methods for
Guadalajara, Zapopan 45138, Mexico (e-mail: rita.fuentes@tec.mx). approximating MUEP in a large class of systems, their appli-
Isaac Chairez is with the Tecnológico de Monterrey at Campus Guadalajara,
Zapopan 45138, Mexico, and also with the Bio-processes Department, Unidad cations in adaptive control designs for real plants require some
Profesional Interdisciplinaria de Biotecnología (UPIBI)-Instituto Politécnico extra conditions. One of these is the necessity of including the
Nacional, Mexico City 07340, Mexico. natural state constraints, which appear because of the presence
Color versions of one or more of the figures in this article are available
online at http://ieeexplore.ieee.org. of physical limitations or the nature of the systems’ states (for
Digital Object Identifier 10.1109/TNNLS.2020.2966914 instance, movement range of articulations in robots, substance
2162-237X © 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: Shri Mata Vaishno Devi University. Downloaded on March 29,2020 at 06:58:51 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
concentrations in chemical or biotechnological systems, and nonlinear switched systems in lower triangular form. Among
some others) [3], [14]. the few studies on the BLF application for full-state con-
There are several evidences showing that the transient strained systems, Tee and Ge [16] developed a control design
period of the training process induces state oscillations of for nonlinear systems with state constraints using integral BLF.
large amplitude for the DNN identifier. Then, if these states Liu and Tong [23] introduced the adaptive controller design
appeared in the closed-loop realization of the adaptive con- (calculated with both symmetric and asymmetric BLFs) for
troller, there is a period while the controlled states may violate single-input-single-output systems with parameter uncertain-
the state restrictions. This aspect is relevant if the adaptive ties, and the entire set of states are constraint. A slightly dif-
controller based on the DNN estimated states must regulate ferent approach appears in [24] showing a class of distributed
real systems, which are highly susceptible to the control state constraints robotic device, which is controlled by the
action such a chemical, biotechnological, or biomedical plants. adaptive controller with gains estimated by a differential law.
Indeed, it is necessary that the control realization considers the This is considered as an intermediate result of a controlled
state constraints since the design stage offers a safety plant BLF. An application of BLF and neural network techniques
operation [15]. This design strategy ensures the fulfilment of is presented in [25]. Here, the use of the integral BLF
the state restrictions permanently (the entire operation time technique aims to propose the distributed attitude coordinated
of the controlled plant), instead of asymptotically. The appli- tracking problem for multiple constrained spacecraft. The
cation of DNNs to approximate uncertain nonlinear systems reconstruction of the leader attitude and angular velocity is
with state restrictions is a relatively recent proposal [4], [16]. solved with the implementation of a sliding-mode estimator.
Two main approaches appeared to solve such problem: the He et al. [24] presented a tracking control for a uncertain
application of projection operators and the application of multilink robot with full-state constraints. The application of
stability analysis for state constraint systems. The first method the BLF guaranteed the uniform ultimate boundedness of the
establishes a limitation for the DNN state to satisfy the closed-loop system. Liu et al. [26] proposed the adaptive con-
predefined constraints for the variables of the uncertain plant. trol for nonlinear stochastic systems (in the Brunovskii form)
This method implies the introduction of integral learning laws with uncertain parameters using symmetric and asymmetric
as proposed by Garcia et al. [17]. Though this option offers BLFs yielding the semi-globally uniformly ultimately bounded
a reliable methodology, the technical difficulties in designing (SGUUB) for the identification error (in mean square sense).
the learning laws demotivated its application. Second, the sta- The results on the application of the BLF for the stochastic
bility analysis considering the state constraints appeared as system are even smaller. An application of a control designed
an efficient design technique for approximate models based using BLF can be revised in [27]. They considered the angle
on DNNs. This approach produces learning laws, which are of attack constraint caused by a scramjet, and the control laws
modulated by the distance between the current value of the were designed using BLF. A compensation for the unknown
DNN states and the known boundaries of the plant variables. actuator faults of the hypersonic flight vehicle is proposed
However, the regular QLF used in the stability analysis of the using back-stepping and composite learning.
approximation error is no longer valid because of the restricted The main contributions of this article include the application
domain of the states. Therefore, a new type of Lyapunov of BLF to deal with the adaptive control design for full-
function was introduced recently. These functions received state constraint systems that may not be converted to the
the name of barrier Lyapunov functions (BLFs) [18]–[20]. Brunovskii form, due to the uncertainty of the state relations.
BLFs are functions that grow to infinity when the states of In this case, the class of constraint states considers the ellipsoid
the system under analysis approach the domain boundaries. restriction [28], which is not a regular form of consider-
The characteristics of the BLF allow the application of the ing the constraints. Moreover, the SGUUB property for the
well-known controlled Lyapunov functions, which permit the identification error is formally confirmed via the application
calculation of the adaptive laws for the weights as a function of the second Lyapunov stability method. An experimental
of the approximation error. Moreover, the adequate implemen- evaluation justified the development of the proposed identifier
tation of BLF yields the design of adaptive controllers, which based on DNN for full-state restricted systems.
may guarantee that the state constraints are not transgressed. In some studies, there are evidences that the adaptive
To date, there are some studies showing the application of controllers designed by the application of BLF may not ensure
BLF for designing the learning laws of diverse ANNs [21]. the satisfaction of the state constraints and may demand long
Most of these approaches only consider the output restriction periods of gain tuning. Usually, the process of adjusting the
while the nonlinear system admits the Brunovskii canoni- gains implicates the recurrent (trial and error) modification of
cal representation, which is a restrictive assumption. Some the controller gains. This trial-and-error scheme may reduce
examples describing the application of BLF for the control the controller effectiveness, the actuators lifetime, and the
of output-constrained nonlinear systems appear in [16], where plant security. The second contribution of this article intro-
the advantages of the asymmetric and symmetric BLFs over duces a dynamic neuro-controller, which incorporates online
QLF are highlighted. Kim and Yoo [22] showed the design of gain control adjustment regulated by the system identification
an adaptive control for nonlinear pure-feedback systems with error [1], [4].
output constraints based on a class of integral barrier Lyapunov The application of a novel BLF allows proving the
functional approach. On the other hand, Niu and Zhao [20] closed-loop practical stability of the origin for both the
investigated the output tracking control problem of constrained identified as well as the trajectory tracking errors for a class of
Authorized licensed use limited to: Shri Mata Vaishno Devi University. Downloaded on March 29,2020 at 06:58:51 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
FUENTES-AGUILAR AND CHAIREZ: ADAPTIVE TRACKING CONTROL OF STATE CONSTRAINT SYSTEMS BASED ON DNNs 3
non-Brunovskii-form uncertain systems. The stability analysis A. Characteristics of the Uncertain System
considers the plant full-state restrictions and their effects on In this article, the nonlinear systems with uncertain structure
the trajectory tracking analysis. An adaptive controller emerges satisfy
from the BLF-based stability analysis. The controller gain
uses the adjusted weights of the DNN identifier. The solution d
x(t) = f (x(t), u(t)) + ξ(x(t), t). (1)
of both problems, the nonparametric identification and the dt
trajectory tracking, enforced the design of an exact expression Here, x is the state vector (x ∈ Rn ), x(0) = x 0 ,
x 0
< ∞.
for the weights dynamics. Consider that X be the open set such that x ∈ X ⊂ Rn . Usu-
The organization of this article is as follows. Section II ally, most of the works considering state restrictions introduce
presents some mathematical preliminaries describing the BLF the restriction that each of the states in (1) is bounded, that is,
fundamentals, Section III describes the class of nonlinear each component of the state x i satisfies x i− < x i < x i+ ∀t ≥ 0.
systems with uncertain structure and full-state constraints, The scalar x i represents the i th component of the state
Section IV introduces the structure of the neuro-identifier vector x. The lower and upper limits for x i are x i− ∈ R and
considering the DNN formulation, the class of activation x i+ ∈ R, respectively. These limits are known in advance. This
functions, and the learning laws for the identifier weights, article considers a different class of state restriction defined by
Section V defines the tracing control design and the impact the set
of the learning laws on the controller gain, Section VI
= x ∈ X |
x
2P1 < x + . (2)
presents the numerical simulations evaluating the control
design on a simple academic example, Section VII describes Here, P1 = P1 , P1 ∈ Rn×n is a positive definite matrix,
the application of the adaptive controller on a real exper- and x + is a positive scalar describing the ellipsoid bounds for
imental system corresponding to a mass-spring device, the system state. The nonlinear vector Lipschitz function f :
Section VIII provides a general discussion on the benefits Rn+m → Rn is composed by n-uncertain nonlinear functions.
of the adaptive control approach attained in this article, and Locally, the effect of internal MUEPs is gathered in the
Section IX closes the study with some final remarks and future piecewise continuous and integrable term ξ(x(t), t). This term
trends. satisfies the following inequality:
ξ(x, t)
2ξ ≤ γ0 + γ1
x
2
(3)
II. M ATHEMATICAL F UNDAMENTALS ON BLF S γ0 , γ1 ∈ R+ 0 < ξ = ξ ∈R
n×n
.
The main tool to design the adaptive controller is the BLF. Under the assumption of the state restrictions, by substitut-
These functions are defined as follows [16]. ing (2) into (3), the following inequality holds:
Definition 1: A BLF is a scalar function V (z) (associ-
ξ(x, t)
2ξ ≤ γ0 + γ1 λ−1 +
max {P1 }{x }. (4)
ated with the dynamic system (d/dt)z(t) = f (z(t)) on
the open region containing the origin) that is continuous, Note that the class of internal MUEPs includes bounded
positive definite with continuous first-order partial derivatives perturbations as well as slow growing dynamics with respect
at every point of ⊆ Rn that satisfies the following to the system state. The control u ∈ Rm (m < n) is a piecewise
properties. continuous and integrable function. All the admissible controls
1) V (z) → ∞ as z approaches δ (boundary of ). belong to the following admissible set:
2) V (z) ≤ b ∀t ≥ 0 along the solution of (d/dt)z(t) = Uadm = {u:
u
2 ≤ u 0 + u 1
x
2 , u 0 , u 1 ∈ R+ }. (5)
f (z(t)) with z(0) ∈ and a given positive constant b.
Evidently, the adaptive control that should solve the problem
A BLF can be symmetric or asymmetric [16]. The following
statement must be included in Uadm .
definition formalizes the structure of barrier functions.
Definition 2: For any positive constants ka and kb , consider IV. N EURO -I DENTIFIER
the state constrain set defined by Z 1 = {z 1 ∈ R : −ka1 < z 1 <
Considering the DNN approximation capabilities described
kb1 } ⊂ R and N := Rl × Z 1 ⊂ Rl+1 be open sets.
in many classical texts on this field [1], the vector field in (1),
The use of BLF for controlled nonlinear systems with output
f (x, u), admits the representation composed by a nominal
constraint focused on the design of the negative semidefinite
system f0 (x, u | W0 ):Rn+m → Rn and the modeling error
function to ensure that the constraints are not transgressed.
f˜(x, u | W0 ):Rn+m → Rn . The Stone–Weierstrass and the
The use of BLF reveals that the stabilizing functions and
Kolmogorov theorems justify the approximation suggested
the control signal remain bounded [19]. Another important
here [29].
contribution of BLF is their capability to be less restrictive on
the initial conditions than QLF, [16], [20], [22]. A. Approximation Based on DNN
The approximation proposed here justifies that the dynamics
III. P ROBLEM S TATEMENT of the uncertain system (1) can also be described as follows:
d
This section presents the class of uncertain nonlinear x(t) = f 0 (x(t), u(t) | W0 )
systems with state restrictions, their characteristics, and dt
assumptions. + f˜(x(t), u(t) | W0 ) + ξ(x(t), t). (6)
Authorized licensed use limited to: Shri Mata Vaishno Devi University. Downloaded on March 29,2020 at 06:58:51 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
There are several alternatives to design the nominal dynam- B. Projected Identifier Structure
ics. This part of the modeling strategy based on DNN can Formally, x̂ ∈ Rn represents the state vector of the adaptive
be proposed using linear combinations of different nonlin- identifier. The learning law, W , should depend on the identi-
ear functions (polynomials, sinusoidal, and sigmoidal). These fication error, that is, the difference between the states of the
functions must form a basis of infinity-dimensional Hilbert uncertain system and the dynamic approximation, that is, x − x̂.
space to get an exact approximation of the uncertain function The DNN-based identifier satisfies the following dynamic
f (x(t), u(t) [29]. When the number of terms used in the representation:
approximation is finite, the approximation error f˜(x(t), u(t))
must be considered. The set of parameters W0 , which are d
x̄(t) = f0 (x̂(t), u(t) | W (t))
bounded, should be adjusted to obtain the most accurate dt
match between the nominal f 0 (x(t), u(t) | W0 ) and the x̂(t) = Pr{x̄(t)}
nonlinear dynamics f (x(t), u(t)), in other words, to reduce d
W (t) = (W (t), x̂ (t), x(t), u(t)) (11)
f˜(x(t), u(t) | W0 )
as much as possible. Considering that dt
the nonlinear dynamics f (x(t), u(t)) is locally Lipschitz and where the projection operator satisfies
under the class of admissible controls u(t), the following
upper bound for the error modeling f˜(x(t), u(t) | W0 ) can z̄, if z̄ ∈
Pr{z̄} = (12)
be obtained [4]: ẑ ∂ , if z̄ ∈
/
f˜(x(t), u(t) | W0 )
2 f ≤ f˜1
x(t)
2 where ẑ ∂ ∈ ω, such that ẑ ∂ = minẑ∈ω
z̄ − ẑ
2 . Note
that ẑ ∂ can be calculated online. The introduction of the
+ f˜0 f˜1 , f˜0 ∈ R+ , f = f ∈ Rn×n . (7)
projection algorithm forces the identifier states to fulfill the
According to the approximation approach [1], the nominal state restrictions for the uncertain system (1). The class of
dynamics of the DNN are selected as follows: state restrictions can also consider the box-type constraints
analyzed in some other studies regarding adaptive controllers
f 0 (x(t), u(t) | W0 ) = Ax(t) + W0 (x(t), u(t)) (8) based on ANN and BLF. Under the given box-type restrictions,
it is possible finding the maximum inscribed ellipsoid, that is,
W0,1 1 (x)
W0 = (x, u) = . (9) the corresponding matrix P1 . A robust control design based
W0,2 2 (x)u(t)
on this argument can be consulted in [28]. The adjustment
The matrices in (8) are A ∈ R , W0,1 ∈ R , of the DNN structure is defined by the function , which
n×n n×l
W0,2 ∈ Rn×s , 1 (·) ∈ Rl and 2 (·) ∈ Rs×m . is known as the learning algorithm. The second method of
Remark 1: A neural approximation presented in (8) has Lyapunov establishes this learning using a particular BLF. The
only one single neuron layer. With more layers, the number set of nonlinear functions f 0 (x̂(t), u(t) | W (t)), in (8), may
of neurons in the single layer increases. In this scenario, there be defined as an affine system composed by the following
is no necessity to include hidden layers in the structure of sections: the first one approximates the control-independent
the NN [4]. What is more relevant, additional complexities dynamics part through the combination of a Hurwitz fixed
may appear in the learning algorithm with the modification of matrix A ∈ R
n×n (selected by the designer) and the nonlinear
the neural structure. In addition, the trajectory tracking is not part which is approximated by some variable parameters W1 (t)
severely improved [1]. with sigmoidal multipliers. The second one defines the vector
The initial adaptive parameters (weights) W0,1 and W0,2 are field associated with the control action using the second set
unknown. However, we consider that they are bounded with of adaptable parameters W2 (t). Indeed, the complete structure
known bounds, that is of the approximation system is
+ + f 0 (x̂(t), u(t) | W ) = A x̂(t)+W1 (t)σ (x̂)+W2 (t)φ(x̂ )u(t). (13)
W0,1 0,1 W0,1 ≤ W0,1 W0,2 0,2 W0,2 ≤ W0,2 (10)
Here, W1 ∈ Rn×l , W2 ∈ Rn×s , σ ∈ Rl , and φ ∈ Rs×m . The
+ +
with W0,1 and W0,2 positive definite matrices of appropriate sigmoid activation vectors, σ and φ, are
dimensions. Also, the matrices 0,1 and 0,2 are positive
−1
definite and symmetric. In [1], the structure in (8) was tested σr (x(t)) = ar 1 + br e −cr x
to design adaptive identifiers, state estimators, and adaptive
−1
controllers. φr, p (x(t)) = ar, p 1 + br, p e −cr, p x . (14)
In (8), the adaptive parameters W0 are assumed to be
unknown but bounded and constant. Consequently, adjusting The parameters ar , ar, p , br , and br, p are positive scalars, while
the structure of f 0 (x(t), u(t) | W0 ) is the approximation cr ∈ R and cr, p ∈ R are eligible vectors of gains. Many
n n
problem formally, which usually implies introducing some methods can be used to adjust these parameters; however, for
time-dependent weights W such that they can track W0 . This the sake of simplicity, we adjusted them by trial and error.
structure working together with the adaptive law W such These functions satisfy the sector conditions
that x is asymptotically reproduced by the identifier state x̂,
σ̃ (z , z )
2 ≤ L
z − z
2
1 2 σ σ 1 2
and under the corresponding persistent excitation condition,
W approximates W0 simultaneously. ×
φ̃(z 1 , z 2 )u
2φ ≤ L φ
z 1 −z 2
2
u
2 (15)
Authorized licensed use limited to: Shri Mata Vaishno Devi University. Downloaded on March 29,2020 at 06:58:51 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
FUENTES-AGUILAR AND CHAIREZ: ADAPTIVE TRACKING CONTROL OF STATE CONSTRAINT SYSTEMS BASED ON DNNs 5
where σ̃ (z 1 , z 2 ) = σ (z 1 ) − σ (ẑ 2 ), φ̃(z 1 , z 2 ) = φ(z 1 ) − φ(zˆ2 ), external signals considered as noises and without the mod-
and L σ and L φ are positive and constant scalars, while σ eling inaccuracies, the error must converge to the origin
and φ are positive matrices. Here, z 1 , z 2 ∈ X ⊆ Rn . These asymptotically. The differential laws are
functions are globally bounded in Rn , that is,
+ −2(1+κ)K i−1 P1 ψi (x̂, u)
¯
σ (z 1 )
2σ ≤ L +
σ ,
φ(z 1 )u
φ ≤ L φ
u
.
2
(16) ¯ x̂, u) =
i (, , i = 1, 2. (19)
1 1 + ¯
x −2(1 + κ) P1 ¯
In (11), a Hurwitz matrix A as well as the time-varying
parameters Wi (t), i = 1, 2 should be found to guarantee the The matrices K 1 , K 2 are constants that chosen during the
state approximation of x(t) by x̂(t), that is, offline training stage [4], and the matrix P1 is the solution
of the Riccati-like matrix inequality [31] given by
lim sup
x(t) − x̂(t)
2 ≤ β
t →∞
P1 A1 + (A1 ) P1 + P1 R1 P1 + Q 1 < 0 (20)
where β > 0 is proportional to the power of noises and
modeling error. Section IV-C introduces the formalization of + +
with A1 = A + 0.5α1 In , R1 = W0,1 + W0,2 and Q 1 =
the adaptive identifier using BLF, that is, the design of the
−1
learning laws for W . L σ + L φ u 0 + u 1 λmin {P1 } .
Lemma 1: Consider the DNN-based approximation (6) of
the uncertain system (1) with the modeling error f˜ satisfying
C. Convergence Analysis for the Projected the restriction (7) and the uncertainties fulfilling (3); then,
DNN Identification Error if there exists a positive scalar α1 > 0 such that (20) has
Let us consider the identifier structure in (11) with the at least a positive definite and symmetric solution P1 , then
representation of the nominal dynamics given in (13); the the origin is an SGUUB for the identification error with an
dynamics of the nonparametric DNN identifier are governed ultimate bound 1 given by
by
d lim sup
2 ≤ 1 (21)
x̄(t) = A x̂(t) + W1 (t)σ (x̂(t)) + W2 (t)φ(x̂(t))u(t) t →∞
dt
x̂(t) = Pr{x̄(t)} . (17) 1 = (γ0 + f˜0 + λ−1 ˜ +
with min {P1 }(γ1 + f 1 )x )/
The objective of applying the identifier (17) can be stated as (2λmin {P1 }(1 + κ)α1 ).
follows: for the class of uncertain systems (1), to select the Proof: Let us consider the energetic-type function for the
constant matrix A and to design a set of nonlinear adaptive identification error
algorithms r (Wr (t), (t), u(t)), r = 1, 2 with = x − x̂ to
n, p
x+
adjust the time-varying parameters W1 (t) and W2 (t) by V1 , W̃1 , W̃2 = ln
x + − 2(1 + κ) P1
d
Wi (t) = i ((t), x̂(t), u(t)), i = 1, 2 2
dt + ki tr W̃i W̃i (22)
in such a way that the estimation of the upper bound of the i=1
identification error is SGUUB with a bound equal to β,
or equivalently, the origin is a practically stable equilibrium with κ > 0 a scalar. This proposed energetic func-
point [30]. Based on the projection of the state on ¯ = x − x̄, tion admits the fundamental properties of BLF. Note that
¯
the dynamics of satisfies
n, p ¯ W̃1 , W̃2 ) where
V1 (, W̃1 , W̃2 ) ≤ V1 (,
d ¯
(t) = A(t) + W̃1 (t)σ (x̂(t)) + W0,1 σ̃ (x(t), x̂(t))
¯ W̃1 , W̃2 = ln x+
dt V1 ,
¯ P1
x + − 2(1 + κ) ¯
+W̃2 (t)φ(x̂ (t))u(t) + W0,2 φ̃(x(t), x̂(t))u(t)
2
+ f˜(x(t), u(t)) + ξ(x(t), t). (18) (23) + ki tr W̃i W̃i .
¯
Note that
≤
. The stability analysis derived in this i=1
article claims the existence of an SGUUB equilibrium point
Now, take its full-time derivative, which corresponds to
for both the identification and the tracking errors. The formal
definition of SGUUB is as follows.
4(1 + κ) ¯
¯ (t)P1 d (t)
Definition 3: The equilibrium point x eq of the identification d V1 (t) = dt
error with the state-feedback controller regulated by a suit- dt x + − 2(1 + κ) ¯ (t)P1 (t)
¯
able integrable controller is SGUUB, if for any , a compact 2
d
subset of Rn , and all x(0) ∈ , there exists an > 0 and a +2 K i tr W̃i (t) W̃i (t) . (24)
time T = T (, x(0)) such that
x(t) − x eq
< for all t ≥ T . dt
i=1
The scalar is named the ultimate bound.
Note that usually depends on the modeling quality and the Here, the notation (d/dt)V1 ((t), ¯ W̃1 (t), W̃2 (t)) =
class of uncertainties affecting the dynamics of the system (1). (d/dt)V1 (t) simplified the style of this article. A relative
Therefore, it should be proven that without the presence of straightforward analysis [1] for the numerator of (24)
Authorized licensed use limited to: Shri Mata Vaishno Devi University. Downloaded on March 29,2020 at 06:58:51 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
leads to states x̂ such that the states of system (1) x track the reference
d ¯ trajectory x ∗ ∈ Rn asymptotically, that is
¯ (t)P1
2 (t)
dt lim
x(t) − x ∗ (t)
→ 0.
¯ (t)P1 (t)
≤ −α1 ¯ + β1 t →∞
¯ ¯
+(t) P1 A1 + A1 P1 + P1 R1 P1 + Q 1 (t) The control design aims to solve the tracking problem that
implies the necessity of introducing the reference trajectory,
2
+2 ¯
tr W̃i (t)P1 ψi (x(t), u(t))(t) (25) which must be the solution to the initial value problem for the
i=1 so-called reference model. Such reference satisfies
with ψ1 (x, u) = σ (x) and ψ2 (x, u) = φ(x)u. This result d ∗
x (t) = f ∗ (x ∗ (t), t) (29)
outcomes from the application of the matrix inequality X Y + dt
(X Y ) ≤ X −1 X + Y Y , which is valid for any with x 0∗ fixed. The variable x ∗ ∈ Rn defines the reference
X, Y ∈ Rnxk and for any positive defined matrix 0 ≤ = trajectory to be tracked. The design of the reference function
∈ Rn×n . This inequality is an extension of the Young f ∗ can be directly solved by the differentiation of the proposed
inequality [31]. The substitution of (25) on (24) yields reference trajectory. Here, f ∗ ∈ R × R → R is a locally
d Lipschitz function with respect to its first argument. To get
V1 (t) the control design, define the associated averaged performance
dt
¯ ¯ index describing the tracking quality as
α1 (t)P 1 (t) − β1
≤ −2(1 + κ) t
x + − 2(1 + κ) ¯ (t)P1 (t)
¯ −1
J (t) := lim sup(t + ν) (H (τ ))dτ (30)
P1 A1 + A1 P1 + P1 R1 P1 + Q 1 t →∞ τ =0
+2(1 + κ)
x + − 2(1 + κ) ¯ (t)P1 (t)
¯ with H (τ ) :=
x(τ ) − x ∗ (τ )
2Q +
u(τ )
2R . Let us define
2 δ = x̂ −x ∗ ; then,
x −x ∗
2 ≤
2 +
δ
2 . The application of
d W̃i (t)
+2tr K i W̃i (t) ¯
+ i ((t), x̂(t), u(t)) . the Young inequality [31] implies that the performance index
dt could be presented as follows:
i=1
(26)
J (t)
t
Based on the existence of the positive solution P1 for the
≤ 1 + −1 lim sup(t + ν)−1 (τ )2 dτ
Riccati-like inequality (20) and the learning laws for the t →∞
Q
τt=0
identification problem (19), the differential inclusion (26)
admits the following upper bound: + 1 + lim sup(t + ν)−1 δ(τ )2 + u(τ )2 dτ.
Q R
t →∞ τ =0
d ¯
2(1 + κ) (t)P ¯
1 (t) − α1−1 β1 (31)
V1 (t) ≤ −α1 .(27)
dt x + − 2(1 + κ)¯ (t)P1 (t)
¯
This method separates the designs of the adaptive state iden-
¯
Note that if (t)P ¯ tification problem from the closed-loop adaptive controller
1 (t) ≤ 1 given in (21), then
construction. Even when this strategy is useful to solve the
d problems independently, there is not an implicit solution of
V1 (t) ≤ 0. (28)
dt the separation principle. Therefore, a joint stability analysis
Note that the second derivative of V1 can be straightforwardly for both the identification and the control design must be
estimated. This second derivative is bounded within ; then, proposed to justify the solution of the trajectory tracking
the application of the Barbalat’s lemma and the argument problem. To obtain the estimated state, the identifier developed
yielding to (28) establishes the SGUUB for the identification in Section IV can be used. The control action should be
error. Based on the comparison principles and using the argu- applied simultaneously to the identifier evolution. The exact
n, p
ment, V1 (, W̃1 , W̃2 ) ≤ V1 (, ¯ W̃1 , W̃2 ); then, the result solution of the tracking control problem can be obtained by
n, p
in (28) can be extended to V1 . the application of the local optimal control technique. This
method overcomes the complexities of the mathematical and
V. C ONTROL D ESIGN AND C ONVERGENCE A NALYSIS FOR practical points of views raised by the necessity of minimizing
THE T RACKING E RROR BASED ON THE P ROJECTED
the second term in (31). Then, by the local optimal concept,
I DENTIFIER A PPROXIMATION instead of minimizing the averaged performance index, let us
minimize the local losses
δ(t)
2 +
u(t)
2 within the given
Note that according to the problem statement of this article, class of controllers in Uadm , namely, feedback controllers
the set can be either defined by the own ellipsoid restrictions with time-varying gain parameters. This part of the control
or formed by the inscribed ellipsoid within the intersection design problem considers an extension to the BLF proposed
of the sets characterizing the box-type state restrictions i , in Section IV. According to the published results on adaptive
that is, = ∩ni=1
r
i , where nr is the number of restrictions. control solution based on DNN, the control model based on
Usually, nr = n. The closed-loop control problem solved in the identifier states can be designed as follows:
this article can be formally presented as follows: to design a
closed-loop control structure u = u(x̂) based on the estimated u(t) = [W2 (t)φ(x̂(t))]+ [u 1 (t) + u 2 (t)] (32)
Authorized licensed use limited to: Shri Mata Vaishno Devi University. Downloaded on March 29,2020 at 06:58:51 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
FUENTES-AGUILAR AND CHAIREZ: ADAPTIVE TRACKING CONTROL OF STATE CONSTRAINT SYSTEMS BASED ON DNNs 7
lim sup
+ δ
2 ≤ c (37)
t →∞
n, p
x+
Fig. 1. Block diagram of the identification and control implementations. V2 , δ, W̃1 , W̃2 = ln
x + −2(1 + κ) P1
where [D]+ is the Moore–Penrose inverse of D. The first term
2
u 1 is a compensatory element for the nonlinear DNN structure, + ln 1+δ P2 δ + K ic tr W̃i W̃i .
i=1
and it is defined as follows:
(38)
u 1 (t) = Ax∗ (t) − f ∗ (x ∗ (t), t). (33)
This proposed energetic function admits the fundamental prop-
The second term u 2 is used for the reference trajectory
erties of BLF for the identification result and the normalizing
tracking x ∗ , which is defined as follows: n, p
result for the tracking error. Note that V2 (, δ, W̃1 , W̃2 ) ≤
¯ δ, W̃1 , W̃2 ) where
u 2 (t) = − 1 + δ (t)P2 δ(t) D P2 δ(t) (34) V2 (,
with Ac1 = A + 0.5α1 In , Ac2 = A + 0.5α2 In , R1c = Here, the simplified notation ((d V2 (t))/dt) goes for
+ + + +
W0,1 + W0,2 , R2c = W0,1 + W0,2 , Q c1 = L σ + L φ (u 0 + (d/dt)V2 ((t), δ(t), W̃1 (t), W̃2 (t)) contributed to the writing
−1
u 1 λmin {P1 }), and Q 2 = Q 0 + σ1 . The following lemma
c style of this article. The straightforward analysis for the numer-
introduces the result which justifies the convergence of the ator of (40) and taking into account the dynamic trajectories
identification and the tracking errors simultaneously. of (1) modeled by the BLF-DNN as adaptive identifier (17)
Lemma 2: Consider the DNN-based approximation (6) of and the reference trajectories, the tracking error obeys the
the uncertain system (1) with the modeling error f˜ satisfying following equation:
the restriction (7) and the uncertainties fulfilling (3). Then,
if there exists positive scalars α1 > 0 and α2 > 0 with α2 = d
P −1/2 Q 0 P −1/2 δ(t) = Aδ(t) + W̃1 (t)σ (x̂) + W0,1 σ (x̂)
λmin such that both matrix inequalities (36) have dt
the positive definite and symmetric solutions P1 and P2 ; then, +W2 (t)φ(x̂)u(t) − (Ax ∗ + f ∗ (x ∗ (t), t)).
Authorized licensed use limited to: Shri Mata Vaishno Devi University. Downloaded on March 29,2020 at 06:58:51 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
The substitution of the dynamics for the identification and weight adjustment laws (35) yields
tracking errors on (40) yields
d2 θ (t)a θ (t)
2(1 + κ)¯ (t)P1 V2 (t) ≤
d
V2 (t) = + dt 2 (1 + θ Pθ )2
dt x − 2(1 + κ) ¯ (t)P1 (t)
¯
where is a positive-definite matrix. As all the elements that
× A(t) + W̃1 (t)σ (x̂(t)) + W0,1 σ̃ (x(t), x̂(t))
form θ := [ , σ (x̂) , (φ(x̂)u) , f˜ , ξ , δ ] are bounded
+W̃2 (t)φ(x̂ (t))u(t) + W0,2 φ̃(x(t), x̂(t))u(t) or piecewise continuous, then, (d 2 V2 (t)/dt 2 ) is bounded
+ f˜(x(t), u(t)) + ξ(x(t), t) almost everywhere. Hence, equivalently, (d/dt)V2 (t) is uni-
formly continuous, combined with the fact that V2 (t) is
2δ (t)P2 Aδ(t) + W̃1 (t) + W0,1 σ (x̂)
+ bounded from below, and (d/dt)V2 (t) is a negative semidefi-
1 + δ (t)P2 δ(t) nite if δ(t) ∈ .
W2 (t)φ(x̂)u(t)− Ax ∗ + f ∗ (x ∗ (t), t) Remark 2: The feasibility of matrix inequalities (20) and
+2δ (t)P2
1 + δ (t)P2 δ(t) (36) has been studied for many years. The seminal study
2
d proposed in [1], Chapter 4, has proposed sufficient conditions
+2 K ic tr W̃i W̃i . (41) to justify the existence of a positive-definite solutions for the
dt
i=1 Riccati-like matrix inequalities. Then, for the first case studied
1/2 1/2
in this article, if the pairs A1 , R1 are controllable, Q 1 , A1
The numerator in the first term in (41) satisfies (25). Based on
is observable, and
this condition, the derivative of the BLF in (41) can be upper
bounded by Q 1 A 1 >0 (44)
¯ ¯ A 1 R1
d (1 + κ)(−α1 (t)P 1 (t) + 1 )
V2 (t) ≤
dt x + − 2(1 + κ) ¯ (t)P1 (t)
¯ then (20) has at least one positive-definite solution. A similar
(1 + κ) ¯
¯ (t) P1 A1 + A P1 + P1 R1 P1 + Q 1 (t) set of conditions can be attained for the corresponding matrix
1
+ inequality (36).
x + − 2(1 + κ) ¯ (t)P1 (t)
¯
2
(1 + κ) 2 ¯
tr W̃i (t)P1 ψi (x(t), u(t))(t) VI. N UMERICAL S IMULATIONS
i=1
+
x + − 2(1 + κ) ¯
¯ (t)P1 (t) Consider a second-order output feedback system (45) taken
δ (t) P2 A + A P2 + P2 R2 P2 + Q 2 δ(t) from [18]
+
1 + δ (t)P2 δ(t) d x 2 (t) − x 1 (t)
α2
δ(t)
P2 − 2
2
2δ (t)P2 W̃1 (t)σ (x̂) x 1 (t) = x 2 (t) + 1 + 0.1 sin(0.1t)
− + dt 1 + x 14 (t)
1 + δ (t)P2 δ(t) 1 + δ (t)P2 δ(t) d
2 x 2 (t) = x 12 (t) + 0.1 cos(0.1t) + u(t) (45)
d dt
+2 K ic tr W̃i W̃i
dt
i=1
where x = [x 1 , x 2 ], x ∈ R2 , x 1 , x 2 are system states, and
W2 (t)φ(x̂ )u(t)−Ax ∗ + f ∗ (x ∗ (t), t) u is the input. The objective is for x to track the desired
+2δ (t)P2 . (42)
1 + δ (t)P2 δ(t) trajectory f ∗ (x ∗ , t), which is generated by a second-order
filter f ∗ (x ∗ , t) = [ωn2 /(s 2 + 2ηωn s + ωn2 ]y ∗ (t) with ωn = 1.5,
If the learning laws proposed in (35), considering the η = 0.8, and for y ∗ (t) defined to be a square wave of
closed-loop controller in Fig. 5(a) and by the assumption amplitude Y0 = 0.5, period T = 20 s.
on the existence of the positive-definite solution for the
nonstandard Riccati matrix inequalities (36), the differential
inclusion (42) can be upper bounded as follows: A. Identification Results
2 − 2α −1
d α1 2(1 + κ)
(t)
¯ A training procedure was developed before the identification
P1 1 1
V2 (t) ≤ − is carried out. To solve this problem, the system (45) is con-
dt 2 x + −2(1 + κ) ¯ (t)P1 (t)
¯
−1
sidered with the constants K 1 , K 2 until the Barrier DNN fits
δ(t)
P2 −α2 2
2
δ(t)
2P2 D P2 the model. When the training was finished, the identification
− α2 − . (43)
1 + δ (t)P2 δ(t) 1 + δ (t)P2 δ(t) procedure using the adaptive adjustment for the weights (19) is
complete. This training process includes the adjustment of the
¯
Note that (1 + κ)
(t)
2 ≤ α −1 and
δ(t)
2 ≤ α −1
parameters characterizing the activation functions of the DNN.
P1 1 1 P2 2 2
then (d/dt)V2 (t) ≤ 0. The application of the Barbalat’s A recurrent method yields the selection of the adequate set
lemma justifies the asymptotic properties of functions and of parameters, verified by the size of the ultimately bounded
their derivatives like in this case. Calculating the second zone for the identification error. For this simulation, the initial
derivative of (39) on the trajectories of the identifier and conditions were set as follows: x 1 (0) = 10−2 , x 2 (0) = 10−3 ,
control errors (within the region where (d/dt)V2 (t)) and the x̂ 1 (0) = 3, and x̂ 2 (0) = 1.2.
Authorized licensed use limited to: Shri Mata Vaishno Devi University. Downloaded on March 29,2020 at 06:58:51 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
FUENTES-AGUILAR AND CHAIREZ: ADAPTIVE TRACKING CONTROL OF STATE CONSTRAINT SYSTEMS BASED ON DNNs 9
Fig. 2. Comparison of the system states (black solid line) and the estimated Fig. 4. Comparison of the system states x1 , x2 , the BLF, and the QLF neural
trajectories obtained with the DNN identifiers with weights estimated as by network states x̂1 , x̂2 with control action in a 35-s interval in contrast with
the application of the QLF (red dashed line) and BLF (blue dotted line). the reference signal x1∗ (t), x2∗ . (a) Trajectory tracking state x1 . (b) Trajectory
(a) System state x1 . (b) System state x2 . tracking state x2 .
Authorized licensed use limited to: Shri Mata Vaishno Devi University. Downloaded on March 29,2020 at 06:58:51 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
TABLE I
C UMULATIVE E RROR FOR THE I DENTIFICATION AND C ONTROL P ROCESS
U SING THE W EIGHTS E STIMATED BY THE A PPLICATION OF THE
C ONTROLLED BLF AND T YPICAL QLF
Authorized licensed use limited to: Shri Mata Vaishno Devi University. Downloaded on March 29,2020 at 06:58:51 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
FUENTES-AGUILAR AND CHAIREZ: ADAPTIVE TRACKING CONTROL OF STATE CONSTRAINT SYSTEMS BASED ON DNNs 11
Authorized licensed use limited to: Shri Mata Vaishno Devi University. Downloaded on March 29,2020 at 06:58:51 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
Authorized licensed use limited to: Shri Mata Vaishno Devi University. Downloaded on March 29,2020 at 06:58:51 UTC from IEEE Xplore. Restrictions apply.