0% found this document useful (0 votes)
19 views15 pages

L1 Adaptive Controller For A Class of Nonlinear Systems

This paper presents an extension of the L1 adaptive controller to a class of nonlinear systems where the con- trol effectiveness is time-varying and unknown, but with a known sign. Moreover, this class of nonlinear systems contains time-varying and unknown state-dependent nonlinearities. The proposed L1 adaptive controller consists of three components, a state predictor used to estimate real states, an adaptive law used to update the adaptive parameters in the state predictor, and a low-pass fil

Uploaded by

kyrillosrefaat18
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
19 views15 pages

L1 Adaptive Controller For A Class of Nonlinear Systems

This paper presents an extension of the L1 adaptive controller to a class of nonlinear systems where the con- trol effectiveness is time-varying and unknown, but with a known sign. Moreover, this class of nonlinear systems contains time-varying and unknown state-dependent nonlinearities. The proposed L1 adaptive controller consists of three components, a state predictor used to estimate real states, an adaptive law used to update the adaptive parameters in the state predictor, and a low-pass fil

Uploaded by

kyrillosrefaat18
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 15

Journal of Dynamic Systems, Measurement and Control.

Received April 21, 2011;


Accepted manuscript posted December 20, 2013. doi:10.1115/1.4026302
Copyright (c) 2013 by ASME

L1 Adaptive Controller for a Class of Nonlinear


Systems

d
Jie Luo

ite
Graduate Research Assistant
Department of Mechanical Engineering

ed
University of Connecticut
Storrs, CT 06269
Email: [email protected]

py
Chengyu Cao

Co
Assistant Professor
Department of Mechanical Engineering
University of Connecticut
Storrs, CT 06269
Email: [email protected]
ot
tN
rip

ABSTRACT
This paper presents an extension of the L1 adaptive controller to a class of nonlinear systems where the con-
trol effectiveness is time-varying and unknown, but with a known sign. Moreover, this class of nonlinear systems
sc

contains time-varying and unknown state-dependent nonlinearities. The proposed L1 adaptive controller consists
of three components, a state predictor used to estimate real states, an adaptive law used to update the adaptive
nu

parameters in the state predictor, and a low-pass filtered control law. First, the stable closed-loop reference system
is constructed. Then, the estimation errors between estimated states and real states are proved to be arbitrarily
small by increasing the adaptation rate. After that, we further prove that the adaptive controller ensures uniformly
Ma

bounded transient and asymptotical tracking of the reference system. The performance bounds can be systemati-
cally improved by increasing the adaptation rate. Simulation results on a single-link nonlinear robot arm verify the
theoretical findings.
ed

1 Introduction
pt

Adaptive control is a promising technique that can be applied to a wide variety of linear and nonlinear systems. Sig-
nificant theoretical results, such as the proof of stability and robustness [1–3] and the improvement of systems tracking
ce

and transient performance [4–6], have been developed. Since the 1980s, adaptive control of nonlinear systems has been
an important area of active research. Initially, adaptive control of nonlinear systems was stimulated by the development of
feedback linearization which is the feedback transformation of nonlinear systems into linear ones [7, 8]. The drawback of
Ac

feedback linearization is that it relies on exact cancelation of nonlinear terms in order to get linear input-output behavior.
Consequently, if the model of the nonlinear terms has some mismatch, the input-output behavior is no longer linear due to
the incomplete cancelation. In [7], it suggested using adaptive control to make cancelations of the nonlinear terms robust
when the uncertainty in the nonlinear terms is parametric. In the beginning of the 1990s, a new approach called ”adaptive
backstepping control” was proposed for controlling of nonlinear systems [9]. Backstepping is a recursive Lyapunov-based
scheme for a class of systems in strict-feedback form. For the class of systems transformable into the strict feedback form,
this approach guarantees global or regional regulation and tracking properties. The practical problem of this approach is that
the handling of nonsmooth nonlinearities has not been addressed. Alternatively, Neural Network (NN) based adaptive con-
trol methodology has been introduced as a promising candidate for controlling unknown nonlinear systems with guaranteed
closed-loop performance [10,11], due to its excellent approximation ability from universal function. However, the NN based
adaptive control is not currently applicable for time varying systems. In [12–15], by constructing a parallel reference system,
a novel architecture for the control of nonlinear single-input single-output (SISO) systems has been developed, named the L1

Downloaded From: http://dynamicsystems.asmedigitalcollection.asme.org/ on 02/04/2014 Terms of Use: http://asme.org/terms


Journal of Dynamic Systems, Measurement and Control. Received April 21, 2011;
Accepted manuscript posted December 20, 2013. doi:10.1115/1.4026302
Copyright (c) 2013 by ASME

adaptive controller. In the presence of unknown high frequency gain, time-varying unknown parameters and time-varying
bounded disturbances, fast adaptation and satisfactory transient response are delivered with stable tracking performance.
The prior results of the L1 adaptive control theory were limited to a class of nonlinear systems where the control
effectiveness is constant with a known sign [12–15]. In this paper, we consider a general class of nonlinear systems where the
control effectiveness is time-varying and unknown, but with a known sign. Moreover, this class of nonlinear systems contains
time-varying and unknown state-dependent nonlinearities. We prove that, subject to a set of mild assumptions, the nonlinear
system can be transformed into an equivalent linear system with time-varying unknown parameters and disturbances. Based
on the methodology in [15], the extension of the L1 adaptive controller in this paper permits fast adaptation and ensures
desired transient and steady state response for both the system’s input and output signals simultaneously. The norm bounds

d
for the error signals between the closed-loop adaptive system and the closed-loop reference system can be systematically

ite
reduced by increasing the adaptation rate.
The paper is organized as follows. Section 2 states some preliminary definitions. Section 3 gives the problem formula-
tion. Section 4 proposes an equivalent linear time-varying (LTV) system. In Section 5, the L1 adaptive control architecture

ed
is presented. Stability and uniform performance bounds are presented in Section 6. In Section 7, simulation results are
presented, while Section 8 concludes the paper.
Throughout this paper, I indicates the identity matrix of appropriate dimension, ||H(s)||L1 denotes the L1 -norm of H(s),

py
||x||L∞ denotes the L∞ -norm of x(t), ||xt ||L∞ denotes the truncated L∞ -norm of x(t) at the time instant t, and ∥x∥2 and ∥x∥∞
indicate the 2- and ∞- norms of the vector x respectively. Some of the proofs are included in the Appendix.

Co
2 Preliminaries
Consider a single-input single-output (SISO) LTV system ot
tN
ẋ(t) = A(t)x(t) + B(t)xin (t) , x(0) = x0
xout (t) = C(t)x(t) + D(t)xin (t) (1)
rip

where A(t) ∈ Rn×n , B(t) ∈ Rn×1 , C(t) ∈ R1×n , D(t) ∈ R. Let M define the input-output map of this system xout =
M {xin } ,where xin , xout are the system input and system output respectively.
sc

Definition 1. The system in Eqn. (1) is called uniformly bounded-input, bounded-output (BIBO) stable, if there exists
a positive constant a > 0 such that for any t0 and any bounded input signal xin (t) the corresponding response for x0 = 0
satisfies supt≥t0 ∥xout (t)∥∞ ≤ a supt≥t0 ∥xin (t)∥∞ .
nu

∫t
Definition 2. The L1 norm of the SISO LTV system in Eqn. (1) is defined as ∥M ∥L1 , supt≥τ, τ∈R+ τ |h(t, σ)|dσ , with
h(t, σ) as the impulse response of M .
Ma

Lemma 1. [16] Suppose the system state equation (1) is uniformly exponentially stable, and there exist finite positive
constants b, c, d > 0 such that for all t
ed

∥B(t)∥ ≤ b , ∥C(t)∥ ≤ c , |D(t)| ≤ d


pt

Then the state equation is also uniformly BIBO stable. Further, if the input xin (t) ∈ L∞ , then ∥xout ∥L∞ ≤ ∥M ∥L1 ∥xin ∥L∞ .
ce

Definition 3. The cascaded system of M and G is defined as M G , i.e.


Ac

(M G ){xin } = M {G {xin }} (2)

We further have ∥M G ∥L1 ≤ ∥M ∥L1 ∥G∥L1 , if ∥M ∥L1 and ∥G ∥L1 are finite. If the system is linear time-invariant (LTI), we
simply use its transfer function to indicate it.

3 Problem Formulation
Consider the following system dynamics

ẋ(t) = Am x(t) + b f (x, u,t)


y(t) = c⊤ x(t) , x(0) = x0 (3)

Downloaded From: http://dynamicsystems.asmedigitalcollection.asme.org/ on 02/04/2014 Terms of Use: http://asme.org/terms


Journal of Dynamic Systems, Measurement and Control. Received April 21, 2011;
Accepted manuscript posted December 20, 2013. doi:10.1115/1.4026302
Copyright (c) 2013 by ASME

where x(t) ∈ Rn is the system state vector (measurable), x0 is the initial state vector, u(t) ∈ R is the control signal, y(t) ∈ R
is the regulated output, b, c ∈ Rn are known constant vectors, Am is a known n × n Hurwitz matrix, and f : Rn × R × R → R
is an unknown nonlinear function.

Assumption 1. [Semiglobal Lipschitz condition on x] For any δ > 0, there exist Lδ > 0 and B > 0 such that | f (x, u,t) −
f (x̄, u,t)| ≤ Lδ ∥x − x̄∥∞ , | f (0, u,t)| ≤ B , for all ∥x∥∞ ≤ δ and ∥x̄∥∞ ≤ δ uniformly in u and t.

Assumption 2. For any δ > 0, there exist d fx (δ) > 0 and d ft (δ) > 0 such that for any u and ∥x∥∞ ≤ δ the partial derivatives
∂ f (x, u,t) ∂ f (x, u,t)
≤ d fx (δ) , ≤ d ft (δ) .

d
of f (x, u,t) with respect to x and t are piece-wise continuous and bounded
∂x ∂t

ite
Assumption 3. [Known sign of control effectiveness] There exist upper and lower bounds ωu > ωl > 0 such that ωl ≤
∂ f (x, u,t) ∂ f (x, u,t)
≤ ωu uniformly in t and x for all t ≥ 0. The partial derivative of with respect to t is bounded.

ed
∂u ∂u
The control objective is to design a full-state feedback adaptive controller to ensure that y(t) tracks a desired system’s

py
output ydes (t) with quantifiable transient and steady-state performance. The definition of ydes (t) is given as follows

Co
ẋdes (t) = Am xdes (t) + bkg r(t)
ydes (t) = c⊤ xdes (t) (4)

where r is the reference signal, and kg = −1/(c⊤ A−1


m b).
ot
tN

4 Definitions and Equivalent LTV System


We define two SISO LTV systems with input xin and output xout for the analysis of the control design
rip

{
ẋout1 (t) = −kω(t)xout1 (t) + kxin (t)
C: (5)
sc

xout (t) = ω(t)xout1 (t) , xout1 (0) = 0


{
ẋout1 (t) = −kω(t)xout1 (t) + kxin (t)
C̄ : (6)
xout (t) = xin (t) − ω(t)xout1 (t) , xout1 (0) = 0
nu

where k is a positive design parameter, and ω(t) is any time-varying signal subject to
Ma

ωl ≤ ω(t) ≤ ωu , |ω̇(t)| < ωd (7)


ed

Choose k in Eqn. (5) and Eqn. (6) to ensure that there exists ρr such that
pt

∥G ∥L1 σb + ∥H C ∥L1 ∥kg r∥L∞ + ∥r0 ∥L∞


< ρr (8)
ce

1 − Lρ ∥G ∥L1
Ac

where σb is the upper bound of f (0, 0,t), Lρ is the Lipschitz constant, r is the reference signal, r0 is the signal with its
Laplace transformation of (sI − Am )x0 , H is the map of H(s) = (sI − Am )−1 b, G = H C̄ is the cascaded system of the LTI
system H and the LTV system C̄ which was defined in Eqn. (6), and define

ρ = ρr + β (9)

with β being an arbitrary positive constant. ρ will be further used to characterize the bounds of real states. It follows from
Lemma 1 that ∥G ∥L1 and ∥H C ∥L1 are finite.

Lemma 2. If the derivative of the input exists, then the cascaded system C F is BIBO stable, where C is defined in Eqn. (5)
and F is s+b
b with b > 0.

Downloaded From: http://dynamicsystems.asmedigitalcollection.asme.org/ on 02/04/2014 Terms of Use: http://asme.org/terms


Journal of Dynamic Systems, Measurement and Control. Received April 21, 2011;
Accepted manuscript posted December 20, 2013. doi:10.1115/1.4026302
Copyright (c) 2013 by ASME

Proof. The proof is given in the Appendix. 

If ∥x∥∞ ≤ ρ and u ≥ 0, it follows from Assumptions 1 and 3 that


∫ u
∂ f (x(0), µ, 0)
f (x, u,t) = f (x, 0,t) + dµ
∂µ
µ=0
≥ −Lρ ∥x∥∞ − σb + ε + ωl u (10)

d
where σb = B + ε and ε is an arbitrarily positive constant. Repeating the above derivation, we could obtain the entire bounds
for f (x, u,t) when ∥x∥∞ ≤ ρ

ite
ωu u − Lρ ∥x∥∞ − σb + ε ≤ f (x, u,t) ≤ ωl u + Lρ ∥x∥∞ + σb − ε , if u < 0 (11)

ed
ωl u − Lρ ∥x∥∞ − σb + ε ≤ f (x, u,t) ≤ ωu u + Lρ ∥x∥∞ + σb − ε , if u ≥ 0 (12)

We further define ρur as

py
ρur = ∥C ∥L1 (ρr Lρr + σb + kg ∥r∥L∞ )/ωl (13)

Co
It follows from Lemma 1 that ∥C ∥L1 is finite.
It follows from Lemma 4 in [12] that there exists co ∈ Rn such that

c⊤
o H(s) = Nn (s)/Nd (s)
ot (14)
tN
where the order of Nd (s) is one more than the order of Nn (s), and both Nn (s) and Nd (s) are stable polynomials. Then, define
rip

∥H C c⊤1H c⊤
o ∥L1
γ1 , o
γ 0 + β1 < β (15)
1 − ∥G ∥L1 Lρ
sc

where β was introduced in Eqn. (9), and γ0 and β1 are any positive constant. As proved later, γ0 and β1 can be arbitrarily
small, and therefore there always exist γ0 and β1 to satisfy Eqn. (15). Notice that C is still a linear low-pass filter system even
nu

it is time-varying, and H C c⊤1H c⊤ ⊤


o L ≤ ∥H ∥L1 ∥C F ∥L1 ∥ F c⊤ H co ∥L1 . Since ∥C F ∥L1 is finite in Lemma 2, and F c⊤ H co
1 1 ⊤
o 1 o o
and H are stable and proper, we can conclude that the ∥H C c⊤1H c⊤ ⊤
o ∥L1 is finite. Similarly, we have finite ∥C c⊤ H co ∥L1 . We
1
o o
Ma

further define ρu and γ2 as

ρu = ρu r + γ 2 (16)
Lρ 1 1
γ2 = ( )∥C ∥L1 γ1 + ( ) C ⊤ c⊤
ed

γ0 (17)
ω ωl co H o L1
pt

In the next Lemma, we demonstrate that the nonlinear system in Eqn. (3) can be transformed into an equivalent linear
system with unknown time-varying parameters.
ce

Lemma 3. (i) If

∥xt ∥L∞ ≤ ρ
Ac

(18)

there exist ω(τ), θ(τ) and σ(τ) over [0, t] such that

ωl < ω(τ) < ωu (19)


|θ(τ)| < θb (20)
|σ(τ)| < σb (21)
f (x(τ), u(τ), τ) = ω(τ)u(τ) + θ(τ)∥x(τ)∥∞ + σ(τ) (22)

(ii) If in addition, u̇(τ) and ẋ(τ) are bounded, then ω(τ), θ(τ), and σ(τ) are differentiable with finite derivatives.
Proof. The proof is given in the Appendix. 

Downloaded From: http://dynamicsystems.asmedigitalcollection.asme.org/ on 02/04/2014 Terms of Use: http://asme.org/terms


Journal of Dynamic Systems, Measurement and Control. Received April 21, 2011;
Accepted manuscript posted December 20, 2013. doi:10.1115/1.4026302
Copyright (c) 2013 by ASME

5 L1 Adaptive Controller
Consider the following state predictor

( )
˙ = Am x̂(t) + b ω̂(t)u(t) + θ̂(t)∥x(t)∥∞ + σ̂(t)
x̂(t)
ŷ(t) = c⊤ x̂(t) , x̂(0) = x0 (23)

The adaptive estimates θ̂(t), σ̂(t), and ω̂(t) are defined as

d
ite
˙ = ΓProj(θ̂(t), −∥x(t)∥ x̃⊤ (t)Pb), θ̂(0) = θ̂
θ̂(t) ∞ 0
˙σ̂(t) = ΓProj(σ̂(t), −x̃⊤ (t)Pb), σ̂(0) = σ̂0

ed
(24)
ω̂(t)
˙ = ΓProj(ω̂(t), −x̃⊤ (t)Pbu(t)), ω̂(0) = ω̂0

py
where x̃(t) = x̂(t) − x(t), Γ ∈ R+ is the adaptation gain, P is the solution of the algebraic equation A⊤
m P + PAm = −Q, Q > 0,
and the projection operator ensures that the adaptive estimates ω̂(t), θ̂(t), and σ̂(t) remain inside the compact sets [ωl , ωu ],

Co
[−θb , θb ], and [−σb , σb ], respectively.
The control signal is generated through gain feedback of the following system

ot
χ̇(t) = ω̂(t)u(t) + θ̂(t)∥x(t)∥∞ + σ̂(t) − kg r(t)
u(t) = −kχ(t) (25)
tN

where k ∈ R+ is a positive design constant.


rip

The complete L1 adaptive controller consists of Eqn. (23)-(25) subject to the L1 -norm upper bound in Eqn. (8).
sc

6 Analysis of L1 Adaptive Controller


6.1 Closed-loop Reference System
nu

Given any ω(τ), θ(τ) and σ(τ) satisfying Eqn. (19)-(21) for all ∀ τ ∈ [0, t], we design a closed-loop reference system

( )
Ma

ẋre f (τ) = Am xre f (τ) + b ω(τ)ure f (τ) + θ(τ)∥xre f (τ)∥∞ + σ(τ) (26)
χ̇re f (τ) = ω(τ)ure f (τ) + θ(τ)∥xre f (τ)∥∞ + σ(τ) − kg r(τ) (27)
ure f (τ) = −kχre f (τ) (28)
ed


yre f (τ) = c xre f (τ) xre f (0) = x0 (29)
pt

with its property summarized in the next Lemma.


ce

Lemma 4. For the closed-loop reference system in Eqn. (26)-(29) subject to the L1 -norm upper bound in Eqn. (8), if
Ac

∥x0 ∥∞ ≤ ρr (30)

then

∥xre f ∥L∞ < ρr (31)


∥ure f ∥L∞ < ρur (32)

where ρr is introduced in Eqn. (8) and ρur is defined in Eqn. (13).

Proof. The proof is given in the Appendix. 

Downloaded From: http://dynamicsystems.asmedigitalcollection.asme.org/ on 02/04/2014 Terms of Use: http://asme.org/terms


Journal of Dynamic Systems, Measurement and Control. Received April 21, 2011;
Accepted manuscript posted December 20, 2013. doi:10.1115/1.4026302
Copyright (c) 2013 by ASME

6.2 Tracking Error Signal


Assume

∥xt ∥L∞ ≤ ρ (33)


∥ut ∥L∞ ≤ ρu (34)

If Eqn. (33) holds, Lemma 3 (i) implies that the system in Eqn. (3) can be rewritten over τ ∈ [0, t] as

d
ẋ(τ) = Am x(τ) + b (ω(τ)u(τ) + θ(τ)∥x(τ)∥∞ + σ(τ))

ite
y(τ) = c⊤ x(τ), x(0) = x0 (35)

where ω(τ), θ(τ), and σ(τ) are unknown time-varying signals subject to the bounds Eqn. (19) - (21) for all ∀ τ ∈ [0, t]. If

ed
x(τ) and u(τ) are bounded, it follows from Eqn. (25) and (35) that ẋ(τ) and u̇(τ) are bounded. Hence, Lemma 3 (ii) and Eqn.
(33)-(34) imply that the derivatives of ω(τ), θ(τ), and σ(τ) are bounded

py
|ω̇(τ)| ≤ dω (ρ, ρu ) < ∞ (36)
|θ̇(τ)| ≤ dθ (ρ, ρu ) < ∞ (37)

Co
|σ̇(τ)| ≤ dσ (ρ, ρu ) < ∞ (38)

Lemma 5. For the system in Eqn. (3) and the L1 adaptive controller in Eqn. (23)-(25), for any t such that Eqn. (33)-(34)
holds, we have
ot

tN
θm (ρ, ρu )
∥x̃t ∥L∞ ≤ (39)
λmin (P)Γ
rip

where

λmax (P)
sc

θm (ρ, ρu ) , 4θ2b + 4σ2b + (ωu − ωl )2 + 4 ∆


λmin (Q)
1
nu

∆ = (ωu − ωl )dω (ρ, ρu ) + θb dθ (ρ, ρu ) + σb dσ (ρ, ρu ) (40)


2

and dω (ρ, ρu ), dθ (ρ, ρu , and dσ (ρ, ρu ) are introduced in Eqn. (36) - (38) respectively.
Ma

Proof. The proof is given in the Appendix. 


ed

6.3 Transient and Steady-State Performance


Theorem 1. Consider the reference system in Eqn. (26)-(29) and the closed-loop L1 adaptive controller in Eqn. (23) - (25)
subject to Eqn. (8). If
pt

∥x0 ∥∞ ≤ ρr (41)
ce

and the adaptive gain is chosen to verify the lower bound


Ac

θm (ρ, ρu )
Γ> (42)
λmin (P)γ20

we have

∥x∥L∞ < ρ (43)


∥u∥L∞ < ρu (44)
∥x̃∥L∞ ≤ γ0 (45)
∥x − xre f ∥L∞ < γ1 (46)
∥u − ure f ∥L∞ < γ2 (47)

Downloaded From: http://dynamicsystems.asmedigitalcollection.asme.org/ on 02/04/2014 Terms of Use: http://asme.org/terms


Journal of Dynamic Systems, Measurement and Control. Received April 21, 2011;
Accepted manuscript posted December 20, 2013. doi:10.1115/1.4026302
Copyright (c) 2013 by ASME

where γ0 , γ1 and γ2 are introduced in Eqn. (15) and (17).


Proof. The proof is given in the Appendix. 

It follows from Eqn. (42) that arbitrarily small γ0 can be obtained by increasing the adaptive gain, which implies
that estimation errors can be arbitrarily small by increasing the adaptive gain. Moreover, the above analysis assumes zero
trajectory initialization error, i.e., x̂0 = x0 , which further improves the transient performance of estimation.
Note that the reference system is only for analysis purpose and not implementable since it uses unknown information
ω(t), θ(t) and σ(t). If adaptive gain Γ is large enough, γ0 can be arbitrarily small, as well as γ1 and γ2 . This implies that both

d
the control signal and system response of the adaptive and reference system could be arbitrarily close. Hence, performance

ite
of the closed-loop reference system provides complete characterization of the adaptive system. It follows from Eqn. (28)
that

ed
ω(t)ure f (t) + θ(t)∥xre f (t)∥∞ + σ(t) = C̄ {θ(t)∥x(t)∥∞ + σ(t)} + C {kg r(t)} (48)

py
Even kω(t) is time-varying, it still roughly indicates the bandwidth of the low-pass filter C . When k is large enough,
C {kg r(t)} ≈ kg r(t) and C̄ {θ(t)∥x(t)∥∞ + σ(t)} ≈ 0. Hence, the response of the reference system approximate desired xdes (t)

Co
defined in (4). As indicated in [12], a larger bandwidth leads to improved performance at the sacrifice of the time-delay
margin. Usually k is chosen to be comparable to the physical bandwidth of the control channel. Hence, only uncertainties
within its bandwidth will be canceled.

7 Simulations
ot
tN
Consider the dynamics of a single-link nonlinear robot arm rotating on a vertical plane

I q̈(t) + F(q(t), q̇(t), u(t),t) = 0 (49)


rip

where q(t) and q̇(t) are the measured angular position and velocity respectively, u(t) is the input torque, I is the unknown
sc

moment of inertia, and F(q(t), q̇(t), u(t),t) is an unknown nonlinear function that lumps the forces, torques and inputs due to
gravity, friction, disturbance and other external sources. Let x = [x1 x2 ]⊤ = [q q̇]⊤ . The system in Eqn. (49) can be presented
in the canonical form
nu

ẋ(t) = Am x(t) + b f (x(t), u(t),t) , y(t) = c⊤ x(t)


Ma

] [
0 1
where b = [0 1]⊤ ,
c = [1 Am = 0]⊤ , , and f (x, u,t)) = [1 1.4]x(t) − F(x2 , x1 , u,t) . The control objective is to
−1 −1.4
ed

design u(t) to achieve tracking of a bounded reference ydes (t) which is introduced in Eqn. (4).
pt

1 0.1 0.1
y
x1(t) − xh1(t)

des
y
ce

0
0.5 0.05
des
,y

−0.1
0 10 20 30 40
y−y
des

0 0 Time t
Ac
y

0.1
x2(t) − xh2(t)

−0.5 −0.05
0

−1 −0.1 −0.1
0 10 20 30 40 0 10 20 30 40 0 10 20 30 40
time t Time t Time t

(a) ydes (t) (dashed), and y(t)(dotted) (b) The error between y(t) and ydes (t) (c) Estimation errors

Fig. 1. Performance of L1 adaptive controller for Scenario 1

In ] of the L1 adaptive controller, we set k = 20, the adaptive gain Γc = 100000, Q = 2I, and hence
[ the implementation
1.4143 0.5000
P= .
0.5000 0.71430

Downloaded From: http://dynamicsystems.asmedigitalcollection.asme.org/ on 02/04/2014 Terms of Use: http://asme.org/terms


Journal of Dynamic Systems, Measurement and Control. Received April 21, 2011;
Accepted manuscript posted December 20, 2013. doi:10.1115/1.4026302
Copyright (c) 2013 by ASME

1 0.1 0.1
y
des

x1(t) − xh1(t)
0.05
y 0
0.5 0.05
−0.05

des
,y

−0.1
0 5 10 15 20 25 30 35 40

y−y
des

0 0 Time t
y

0.1

x2(t) − xh2(t)
0.05
−0.5 −0.05
0

−0.05
−1 −0.1 −0.1
0 10 20 30 40 0 10 20 30 40 0 5 10 15 20 25 30 35 40

d
time t Time t Time t

(a) ydes (t) (dashed), and y(t)(dotted) (b) The error between y(t) and ydes (t) (c) Estimation errors

ite
Fig. 2. Performance of L1 adaptive controller for Scenario 2

ed
Consider the first scenario where F(x2 , x1 , u,t) = (3 + 0.5 sin(2t))u(t) + eu(t) sin(x1 (t)) + x22 (t) + x12 (t) + sin(5t), and the

py
compact sets can be conservatively chosen as ωl = 0.5, ωu = 20, θb = 10, σb = 10. The simulation results of the L1 adaptive
controller are shown in Fig. 1(a)-1(c) for reference input ydes (t) with r = cos(0.5t). Figure 1(a) shows the transient and
tracking performance of output y(t). As shown in Fig. 1(b), the actual error between output y(t) and desired output ydes (t) is

Co
very small. Also, as shown in Fig. 1(c), the estimation errors between estimated states from state predictor and real states are
very small. In Fig. 1(c), xh1 represents the estimated state for x1 , and xh2 represents the estimated state for x2 . For the second
scenario, we change the nonlinear function to: F(x2 , x1 , u,t) = (3 + 0.5 sin(2t))u(t) + sat(u(t)) sin(x1 (t)) + ln(1 + |x1 (t)|) +
sat(x2 (t))x1 (t) + sin(5t). sat(·) represents the saturation function with the upper bound 1.5 and lower bound −1.5. Without
ot
retuing the L1 adaptive controller, the simulation results are shown in Fig. 2(a)-2(c). Figure 2 also shows a good transient
and tracking performance of the output with very small estimation errors. In short, the L1 adaptive controller ensures smooth
tN
and uniform transient and tracking performance in the presence of different unknown nonlinearities without requiring any
retuning.
rip

8 Conclusion
sc

A novel L1 adaptive control architecture is presented in this paper, which ensures desired transient response in addition
to stable tracking for general uncertain systems in the presence of time-varying control effectiveness with a known sign, as
well as unknown state-dependent and time-varying nonlinearities. First, a stable closed-loop system is constructed. Subject
nu

to a set of mild assumptions, the nonlinear system can be transformed into an equivalent linear system with time-varying
unknown parameters and disturbances. Then, the estimation errors between estimated states from state predictor and real
Ma

states are proved to be arbitrarily small by increasing the adaptation rate. Finally, we prove that the control signal and the
system output response can arbitrarily approximate the counterparts of the closed-loop reference system.
ed

References
[1] Sastry, S., and Bodson, M., 1989. Adaptive control: stability, convergence, and robustness. Prentice-Hall, Englewood
pt

Cliffs, NJ.
[2] Kreisselmeier, G., and Anderson, B., 1986. “Robust model reference adaptive control”. IEEE Transactions on Auto-
ce

matic Control, 31(2), pp. 127–133.


[3] Ioannou, P., and Tsakalis, K., 1986. “A robust direct adaptive controller”. IEEE Transactions on Automatic Control,
31(11), pp. 1033–1043.
Ac

[4] Sun, J., 1993. “A modified model reference adaptive control scheme for improved transient performance”. IEEE
Transactions on Automatic Control, 38(8), pp. 1255–1259.
[5] Krstic, M., Kokotovic, P., and Kanellakopoulos, I., 1993. “Transient-performance improvement with a new class of
adaptive controllers”. Systems & Control Letters, 21(6), pp. 451–461.
[6] Miller, D., and Davison, E., 1991. “An adaptive controller which provides an arbitrarily good transient and steady-state
response”. IEEE Transactions on Automatic Control, 36(1), pp. 68–81.
[7] Sastry, S., and Isidori, A., 1989. “Adaptive control of linearizable systems”. IEEE Transactions on Automatic Control,
34(11), pp. 1123–1131.
[8] Kanellakopoulos, I., Kokotovic, P., and Morse, A., 1991. “Systematic design of adaptive controllers for feedback
linearizable systems”. IEEE Transactions on Automatic Control, 36(11), pp. 1241–1253.
[9] Krstic, M., Kanellakopoulos, I., and Kokotovic, P., 1995. Nonlinear and adaptive control design. John Wiley and Sons,
New York, NY.

Downloaded From: http://dynamicsystems.asmedigitalcollection.asme.org/ on 02/04/2014 Terms of Use: http://asme.org/terms


Journal of Dynamic Systems, Measurement and Control. Received April 21, 2011;
Accepted manuscript posted December 20, 2013. doi:10.1115/1.4026302
Copyright (c) 2013 by ASME

[10] Cao, C., and Hovakimyan, N., 2007. “Novel L1 neural network adaptive control architecture with guaranteed transient
performance”. IEEE Transactions on Neural Networks, 18(4), pp. 1160 –1171.
[11] Calise, A., Hovakimyan, N., and Idan, M., 2001. “Adaptive output feedback control of nonlinear systems using neural
networks”. Automatica, 37(8), pp. 1201–1211.
[12] Cao, C., and Hovakimyan, N., 2008. “Design and analysis of a novel L1 adaptive control architecture with guaranteed
transient performance”. IEEE Transactions on Automatic Control, 53(2), pp. 586–591.
[13] Cao, C., and Hovakimyan, N., 2008. “L1 adaptive controller for systems with unknown time-varying parameters and
disturbances in the presence of non-zero trajectory initialization error”. IEEE Transactions on Automatic Control,
81(7), pp. 1147–1161.

d
[14] Cao, C., and Hovakimyan, N., 2007. “Guaranteed transient performance with L1 adaptive controller for systems with

ite
unknown time-varying parameters and bounded disturbances: Part i”. In American Control Conference, pp. 3925–3930.
[15] Cao, C., and Hovakimyan, N., 2008. “L1 adaptive controller for a class of systems with unknown nonlinearities: Part
i”. In American Control Conference, 2008, pp. 4093–4098.

ed
[16] Rugh, W., 1996. Linear system theory. Prentice Hall, Upper Saddle River, NJ.

py
Appendix
Proof of Lemma 2. Following from Definition 3, the cascaded system C F can be described by

Co
k
ẋout1 (t) + kω(t)xout1 (t) = (ẋin (t) + bxin (t)) (50)
b
ot
xout (t) = ω(t)xout1 (t) (51)
tN
Since Eqn. (50) is linear, xout1 = x1 + x2 where

ẋ1 (t) + kω(t)x1 (t) = kxin (t) (52)


rip

k
ẋ2 (t) + kω(t)x2 (t) = ẋin (t) , x1 (0) = x2 (0) = 0 (53)
b
sc

It follows from Eqn. (7) and (52) that


nu

1
∥x1 ∥L∞ ≤ ∥xin ∥L∞ (54)
ωl
Ma

The solution of Eqn. (53) is given as

∫t ∫ t ∫τ
ed

k
x2 (t) = e− 0 (kω(τ))dτ ẋin (τ)e 0 (kω(τ1 ))dτ1 dτ (55)
0 b
∫ t ∫t
k
ẋin (τ)e− kω(τ1 )dτ1
pt

= τ dτ (56)
0 b
∫ t
k ( − ∫τt kω(τ1 )dτ1 ∫t )
xin (τ)]t0 − xin (τ)de− τ kω(τ1 )dτ1
ce

= [e (57)
b 0
∫ t
k ( − ∫τt kω(τ1 )dτ1 ∫t )
= [e xin (τ)]0 − xin (τ)e− τ kω(τ1 )dτ1 (kω(τ))dτ
t
(58)
Ac

b 0

∫t
From Eqn. (7), we know that [e− τ kω(τ1 )dτ1 ]t
0 ≤ (1 − e−kωl t ). Then we obtain

k k2
∥x2 ∥L∞ ≤ (∥xin ∥L∞ + ∥xin (0)∥L∞ ) + ωu ∥xin ∥L∞ (59)
b b

Therefore,

(1 k k2 )
∥xout ∥L∞ ≤ ωu ∥xin ∥L∞ + (∥xin ∥L∞ + ∥xin (0)∥L∞ ) + ωu ∥xin ∥L∞ (60)
ωl b b

Downloaded From: http://dynamicsystems.asmedigitalcollection.asme.org/ on 02/04/2014 Terms of Use: http://asme.org/terms


Journal of Dynamic Systems, Measurement and Control. Received April 21, 2011;
Accepted manuscript posted December 20, 2013. doi:10.1115/1.4026302
Copyright (c) 2013 by ASME

The xout is bounded for bounded xin , and Lemma 2 is proved. 

Proof of Lemma 3. From Assumption 3 and by the mean value theorem, we can obtain

∂ f (x, u, τ)
f (x, u, τ) = f (x, 0, τ) + |ui u
∂u
= g(τ) + ω(τ)u(τ) (61)

d
for any τ ∈ [0, t], where ui is a point among the segment connecting u to 0. From Assumption 3, we also know that

ite
ωl ≤ ω(τ) ≤ ωu (62)

ed
for any τ ∈ [0, t]. It follows from Assumption 1 that

py
|g(τ)| ≤ Lρ ∥x∥∞ + B < Lρ ∥x∥∞ + σb (63)

for ∥x∥∞ ≤ ρ and any τ ∈ [0, t].

Co
Next, we prove that there exist θ(τ) and σ(τ) over τ ∈ [0, t] such that

|θ(τ)| < θb (64)


|σ(τ)| < σb
ot (65)
g(τ) = θ(τ)∥x(τ)∥∞ + σ(τ) (66)
tN

where θb is defined as θb = Lρ . It follows from Eqn. (63) that


rip

g(0) < θb ∥x(0)∥∞ + σb (67)


sc

which implies that there exist θ(0) and σ(0) such that
nu

|θ(0)| < θb , |σ(0)| < σb (68)

and
Ma

g(0) = θ(0)∥x(0)∥∞ + σ(0) (69)


ed

We construct the trajectories of θ(τ) and σ(τ) according to the following dynamics
[ ] [ dg(τ) ]
pt

θ̇(τ) dτ − θ(τ)
d∥x(τ)∥∞
= A−1
η
dτ (70)
σ̇(τ) 0
ce

where
Ac

[ ]
∥x(τ)∥∞ 1
Aη = (71)
−(σb − |σ(τ)|) θb − |θ(τ)|

with the initial values being bounded according to Eqn. (68). Using derivation similar to Lemma 2 in [15], we can prove
Eqn. (64)-(66) while Lemma 3(i) follows from Eqn. (61) and Eqn. (64)-(66).
In the above analysis, g(t) is continuous and |g(t)| never approaches the bound Lρ ρ + σb . ω(t) also never approaches its
bounds. It follows from Assumption 3 that ω̇(τ) is bounded for any τ ∈ [0,t]. In light of Assumption 2, ∂ f (x,0,τ)
∂x and d f (x,0,τ)

are bounded, which implies ġ(τ) is bounded. Further, from Eqn. (61) and (66), bounded x(τ) and u(τ) imply bounded ẋ(τ)
over [0,t] for the system dynamics in Eqn. (3), and hence bounded d∥x(τ)∥ dτ

, although the derivative d∥x(τ)∥


may not be
continuous. It follows from Eqn. (71) that det(Aη ) ̸= 0, and therefore we conclude from Eqn. (70) that θ̇(τ) and σ̇(τ) are
bounded. This concludes the proof. 

Downloaded From: http://dynamicsystems.asmedigitalcollection.asme.org/ on 02/04/2014 Terms of Use: http://asme.org/terms


Journal of Dynamic Systems, Measurement and Control. Received April 21, 2011;
Accepted manuscript posted December 20, 2013. doi:10.1115/1.4026302
Copyright (c) 2013 by ASME

Remark 1. Note that although Lemma 3 proves the existence and boundedness of θ̇(τ), σ̇(τ), their continuity is not guar-
anteed. The reason is that d∥x(τ)∥


can be piece-wise continuous due to the definition of the ∞ norm. Thus, in Eqn. (70) the
right-hand side can be a piece-wise continuous functions of t.

Proof of Lemma 4. If Eqn. (31) is not true, since ∥xre f (0)∥∞ = ∥x0 ∥ < ρr and xre f (t) is continuous, there exists t such
that the truncated norm

∥xre ft ∥L∞ ≤ ρr (72)

d
and xre f (t) = ρr (73)

ite
Substituting Eqn. (28) into Eqn. (26), we have xre f (τ) = G {rre f1 (τ)} + (H C ){kg r(τ)} + r0 (τ) , where rre f1 (τ) =

ed
θ(τ)∥xre f (τ)∥∞ + σ(τ) . Using Assumption 1 and the upper bound in Eqn. (72), we arrive at the following upper bound
∥rre f1 ∥L∞ ≤ Lρr ∥xre ft ∥L∞ + ∥σ∥L∞ . Hence,

py
∥xre ft ∥L∞ ≤ ∥G ∥L1 (Lρr ∥xre ft ∥L∞ + ∥σ∥L∞ ) + ∥H C ∥L1 ∥kg r∥L∞ + ∥r0 ∥L∞

Co
Equation (74) implies

∥G ∥L1 σb + ∥H C ∥L1 ∥kg r∥L∞ + ∥r0 ∥L∞


ot
∥xre ft ∥L∞ ≤ (74)
1 − Lρr ∥G ∥L1
tN
and combining Eqn. (8) implies that ∥xre ft ∥L∞ < ρr , and contradicts Eqn. (73). This proves Eqn. (31).
It follows from Eqn. (28) that
rip

ure f (τ) = C {kg r(τ) − rre f1 (τ)}/ω(τ) (75)


sc

Hence,
nu

∥ure f ∥L∞ < ∥C ∥L1 (Lρr ρr + σb + kg ∥r∥L∞ )/ωl (76)


Ma

Combining the definition of ρur in Eqn. (13) and (76), we have


ed

∥ure f ∥L∞ < ρur (77)


pt

which proves Eqn. (32), and concludes the proof. 


ce

Proof of Lemma 5. It follows from Eqn. (33) - (34) that Eqn. (35)-(38) hold for any τ ∈ [0,t]. Consider the following
Lyapunov function candidate
Ac

V (x̃(τ), ω̃(τ), θ̃(τ), σ̃(τ)) = x̃⊤ (τ)Px̃(τ) + Γ−1 (ω̃2 (τ) + θ̃2 (τ) + σ̃2 (τ)) (78)

where

ω̃(τ) , ω̂(τ) − ω(τ), θ̃(τ) , θ̂(τ) − θ(τ), σ̃(τ) , σ̂(τ) − σ(τ) (79)

It follows from Eqn. (23) and (35) that over [0,t]

x̃˙ = Am x̃(τ) + b(ω̃(τ)u(τ) + θ̃(τ)∥x(τ)∥∞ + σ̃(τ)) (80)

Downloaded From: http://dynamicsystems.asmedigitalcollection.asme.org/ on 02/04/2014 Terms of Use: http://asme.org/terms


Journal of Dynamic Systems, Measurement and Control. Received April 21, 2011;
Accepted manuscript posted December 20, 2013. doi:10.1115/1.4026302
Copyright (c) 2013 by ASME

where x̃(0) = 0. The projection algorithm in adaptive law ensures that

ωl ≤ ω̂(τ) ≤ ωu , |θ̂(τ)| ≤ θb , |σ̂(τ)| ≤ σb (81)

and from Eqn. (19)-(21), we can obtain

max (Γ−1 (ω̃2 + θ̃2 + σ̃2 )) ≤ (4θ2b + 4σ2b + (ωu − ωl )2 )/Γ (82)

d
τ∈[0,t)

ite
and therefore

ed
θm (ρ, ρu )
V (0) ≤ (4θ2b + 4σ2b + (ωu − ωl )2 )/Γ ≤ (83)
Γ

py
Let t1 ∈ (0,t] be the first time-instant of the discontinuity of any of the derivatives, ω̂(t),
˙ ˙
θ̂(t), σ̂(t),
˙ θ̇(t), or σ̇(t). Next

Co
we prove that

θm (ρ, ρu )
V (τ) ≤ , ∀τ ∈ [0,t1 ]
ot (84)
Γ
tN
The derivative of V (τ) is

V̇ (τ) = x̃˙⊤ (τ)Px̃(τ) + x̃(τ)⊤ Px̃˙(τ) + 2Γ−1 (ω̃(τ)ω̃


rip

˙ (τ) + θ̃(τ)θ̃˙ (τ) + σ̃(τ)σ̃˙ (τ)) (85)

for any τ ∈ [0,t1 ]. Using the projection based adaptation laws, Eqn. (24), and from Eqn. (79) and (80), we further get
sc

V̇ (τ) ≤ −x̃⊤ (τ)Qx̃(τ) − 2Γ−1 (ω̃(τ)ω̇(τ) + θ̃(τ)θ̇(τ) + σ̃(τ)σ̇(τ))


nu

(86)

for any τ ∈ [0,t1 ). If at any τ ∈ [0,t1 )


Ma

θm (ρ, ρu )
V (τ) ≥ (87)
Γ
ed

then it follows from Eqn. (78), (82) and the definition of θm (ρ, ρu ) in Eqn. (40) that
pt

λmax (P) 1
ce

x̃⊤ (τ)Px̃(τ) ≥ 4 ( (ωu − ωl )dω (ρ, ρu ) + θb dθ (ρ, ρu ) + σb dσ (ρ, ρu )) (88)


λmin (Q) 2
Ac

and hence

λmin (Q) ⊤
x̃⊤ (τ)Qx̃(τ) ≥ x̃ (τ)Px̃(τ)
λmax (P)
1
(ωu − ωl )dω (ρ, ρu ) + θb dθ (ρ, ρu ) + σb dσ (ρ, ρu ))
≥42 (89)
Γ

It follows from Eqn. (19)-(21) and Eqn. (81) that

|ω̃(τ)| ≤ ωu − ωl , |θ̃(τ)| ≤ 2θb , |σ̃(τ)| ≤ 2σb (90)

Downloaded From: http://dynamicsystems.asmedigitalcollection.asme.org/ on 02/04/2014 Terms of Use: http://asme.org/terms


Journal of Dynamic Systems, Measurement and Control. Received April 21, 2011;
Accepted manuscript posted December 20, 2013. doi:10.1115/1.4026302
Copyright (c) 2013 by ASME

for all τ ∈ [0,t1 ). Since ω̇(τ), θ̇(τ) and σ̇(τ) are continuous over [0,t1 ), the upper bounds in Eqn. (19)-(21) and Eqn. (36)-(38)
lead to the following upper bound

ω̃(τ)ω̇(τ) + θ̃(τ)θ̇(τ) + σ̃(τ) (ωu − ωl )dω (ρ, ρu ) + 2θb dθ (ρ, ρu ) + 2σb dσ (ρ, ρu )
≤ (91)
Γ Γ

θm (ρ,ρu )
Hence, if V (τ) ≥ Γ , then from Eqn. (86), (89) and (91), we have

d
V̇ (τ) ≤ 0 (92)

ite
θm (ρ,ρu )
It follows from Eqn. (92) that V (τ) ≤ for any τ ∈ [0,t1 ). Since λmin (P)∥x̃∥2 ≤ x̃⊤ (τ)Px̃(τ) ≤ V (τ), then for any

ed
Γ
τ ∈ [0,t1 )

py
θm (ρ, ρu )
∥x̃∥2 ≤ (93)
λmin (P)Γ

Co
Since V (τ) is continuous, we further have


∥x̃∥∞ ≤
θm (ρ, ρu ) ot
, τ ∈ [0,t1 ] (94)
λmin (P)Γ
tN

Let t2 ∈ (t1 ,t] be the next time-instant such that it is the discontinuity of any of the derivatives, ω̂(t),
˙ ˙
θ̂(t), σ̂(t),
˙ θ̇(t), or
rip

σ̇(t). Using similar derivations as above, we can prove that


sc

θm (ρ, ρu )
∥x̃(τ)∥∞ ≤ , τ ∈ [t1 ,t2 ] (95)
λmin (P)Γ
nu

Iterating the process until the time instant t, we get


Ma


θm (ρ, ρu )
∥x̃(τ)∥L∞ ≤ (96)
λmin (P)Γ
ed

which concludes the proof. 


pt

Proof of Main Theorem. The proof will be done by contradiction. Assume that Eqn. (43)-(44) are not true. Then,
since ∥x(0)|∞ < ρ, |u(0)| = 0 < ρu , and x(τ) and u(τ) are continuous, there exists t¯ ≥ 0 such that
ce

∥x(t¯)∥∞ = ρ (97)
Ac

or ∥u(t¯)∥∞ = ρu (98)

while

∥xt¯∥L∞ ≤ ρ, ∥ut¯∥L∞ ≤ ρu (99)

In what follows, we prove Eqn. (99) ensures

∥(x − xre f )t¯∥L∞ < γ1 (100)


∥(u − ure f )t¯∥L∞ < γ2 (101)

Downloaded From: http://dynamicsystems.asmedigitalcollection.asme.org/ on 02/04/2014 Terms of Use: http://asme.org/terms


Journal of Dynamic Systems, Measurement and Control. Received April 21, 2011;
Accepted manuscript posted December 20, 2013. doi:10.1115/1.4026302
Copyright (c) 2013 by ASME

The proof of Eqn. (100)-(101) will be done by contradiction, too. Assume that Eqn. (100)-(101) are not true. Then, since
∥x(0) − xre f (0)∥∞ = 0 ≤ γ1 , ∥u(0) − ure f (0)∥∞ = 0 ≤ γ2 , and x(τ), xre f (τ), u(τ), ure f (τ) are continuous, there exists t ∈ [0, t¯]
such that

∥x(t) − xre f (t)∥∞ = γ1


∥u(t) − ure f (t)∥∞ = γ2 (102)

d
while

ite
∥(x − xre f )t ∥L∞ ≤ γ1 , ∥(u − ure f )t ∥L∞ ≤ γ2 (103)

ed
Note that Eqn. (99) implies ∥xt ∥L∞ ≤ ρ and ∥ut ∥L∞ ≤ ρu . Hence, it follows from Lemma 3 that there exist ω(t), θ(t) and
σ(t) with bounded derivatives such that the system in Eqn. (3) is equivalent to the LTV system in Eqn. (35) over [0,t], which

py
will be used in the following analysis. It follows from Eqn. (42) and Lemma 5 that

Co
∥x̃t ∥L∞ ≤ γ0 (104)

Let r̃(τ) = ω̃(τ)u(τ)+ θ̃(τ)∥x(τ)∥∞ + σ̃(τ) , r1 (τ) = θ(τ)∥x(τ)∥∞ +σ(τ) . It follows from Eqn. (25) that χ̇(t) = ω(t)u(t)+
r1 (t) − kg r(t) + r̃(t) . Consequently ot
tN
χ̇(t) + kω(t)χ(t) = r1 (t) − kg r(t) + r̃(t) (105)
rip

Using the definition of C from Eqn. (5), we can write

ω(t)u(t) = C {kg r(t) − r1 (t) − r̃(t)}


sc

(106)
nu

and the system in Eqn. (3) consequently takes the form

x(t) = (H C̄ ){r1 (t)} + (H C ){kg r(t) − r̃(t)} + r0 (t)


Ma

(107)

Using the same ω(t), θ(t) and σ(t) as in Eqn. (35) and applying the reference control algorithm as in Eqn. (26)-(28), we
have
ed

χ̇re f (t) + kω(t)χre f (t) = rre f 1 − kg r(t)


pt

ure f (t) = −kχre f (t) (108)


ce

where rre f 1 = θ(t)∥xre f (t)∥∞ + σ(t). Similarly, we have


Ac

ω(t)ure f (t) = C {kg r(t) − rre f 1 (t)} (109)

and

xre f (t) = (H C̄ ){rre f 1 (t)} + (H C ){kg r(t)} + r0 (t) (110)

Let e(τ) = x(τ) − xre f (τ). It follows from Eqn. (107) and (110) that

e(τ) = (H C̄ ){r2 (τ)} − (H C ){r̃(τ)} (111)

Downloaded From: http://dynamicsystems.asmedigitalcollection.asme.org/ on 02/04/2014 Terms of Use: http://asme.org/terms


Journal of Dynamic Systems, Measurement and Control. Received April 21, 2011;
Accepted manuscript posted December 20, 2013. doi:10.1115/1.4026302
Copyright (c) 2013 by ASME

over [0,t] where

r2 (τ) = θ(τ)(∥x(τ)∥∞ − ∥xre f (τ)∥∞ ) (112)

We have x̃(s) = H(s)r̃(s) , which leads to (H C ){r̃(t)} = (H C c⊤1H c⊤ 1 ⊤


o H ){r̃(t)} = (H C c⊤ H co ){x̃(t)}. Hence, we have the
o o
following upper bound

d
1
∥et ∥L∞ ≤ ∥G ∥L1 ∥r2t ∥L∞ + ∥H C c⊤
o ∥L1 ∥x̃t ∥L∞ (113)
c⊤

ite
oH

Since ∥x(τ)∥∞ − ∥xre f (τ)∥∞ ≤ ∥x(τ) − xre f (τ)∥∞ ≤ ∥et ∥L∞ for any τ ∈ [0,t], it follows from Eqn. (20) and (112) that

ed
∥r2t ∥L∞ ≤ Lρ ∥et ∥L∞ (114)

py
From Eqn. (113) we have ∥et ∥L∞ ≤ ∥G ∥L1 Lρ ∥et ∥L∞ + ∥H C c⊤1H c⊤
o ∥L1 ∥x̃t ∥L∞ . Eqn. (104) and the L1 -norm upper bound

Co
o
∥H C ⊤1 c⊤o ∥L1
from Eqn. (8) leads to the following upper bound ∥et ∥L∞ ≤ 1−∥G ∥L1 Lρ γ0 ,
co H
which along with Eqn. (15) leads to

∥et ∥L∞ ≤ γ1 − β1 < γ1


ot (115)
tN
Note that from Eqn. (106) and (109) one can derive ω(t)(u(t) − ure f (t)) = −C {r2 (t)} − C {r̃(t)}. Since C {r̃(t)} =
(C c⊤1H c⊤ 1 ⊤
o H ){r̃(t)} = (C c⊤ H co ){x̃(t)}, it follows from Eqn. (114) that
o o
rip

Lρ 1
∥(u − ure f )t ∥L∞ ≤ ∥C ∥L1 ∥(x − xre f )t ∥L∞ + (1/ωl ) C ⊤ c⊤ ∥x̃t ∥L∞ (116)
ωl co H o L1
sc

It follows from Eqn. (104), (116), and the definition of γ2 in Eqn. (17) that
nu

1 ⊤
∥(u − ure f )t ∥L∞ ≤ (Lρ /ω)∥C ∥L1 (γ1 − β1 ) + (1/ωl ) C γ0 < γ2
Ma

co (117)
c⊤
oH L1

The upper bounds in Eqn. (115) and (117) contradict the equality in Eqn. (102), which proves Eqn. (100)-(101).
ed

Since ω(t), θ(t) and σ(t) are bounded, it follows from Eqn. (8) and Lemma 4 that

∥xre ft¯ ∥L∞ ≤ ρr , ∥ure ft¯ ∥L∞ ≤ ρur


pt

(118)
ce

Combining Eqn. (100)-(101) and Eqn. (118), we have


Ac

∥xt¯∥L∞ < ρ , ∥ut¯∥L∞ < ρu (119)

which contradicts the equalities in Eqn. (97) and (98). Hence, Eqn. (43)-(44) are proved.
Since Eqn. (99) holds for any t¯, Eqn. (46)-(47) follow from Eqn. (100)-(101) immediately. The upper bound in Eqn.
(45) follows from Eqn. (104) directly. 

Downloaded From: http://dynamicsystems.asmedigitalcollection.asme.org/ on 02/04/2014 Terms of Use: http://asme.org/terms

You might also like