Professional Documents
Culture Documents
Abstract: In this paper, the stability of a discrete-time event-triggered control system over
noisy feedback channels is analyzed. The transmission between the controller and the actuator
is triggered by an event involving estimated control, desirable control, and a trigger threshold.
It is revealed that if the trigger threshold is less than a bound determined by the system
and controller matrices, then the closed-loop system is mean square stable. In addition, two
equivalent conditions based on algebraic Riccati inequality/equation (ARI/ARE) are proposed
to facilitate further analysis.
Keywords: Event-triggered control; Noisy channels; Bounded Real Lemma; ARI; ARE.
10494
19th IFAC World Congress
Cape Town, South Africa. August 24-29, 2014
The event-trigger strategy is then formulated in the fol- 3.1 Augmented Perturbed Linear Model
lowing algorithm.
Before proceeding, we introduce a lemma which will be
Algorithm 1:
used in the sequel.
Step 1. At the beginning of each updating process, m = tk , Lemma 1. (Special case of Bounded Real Lemma (BRL),
k ≥ 0, C(ξ(tk )) = kû(tk ) − ũ(tk )k2 − λ2 kũ(tk )k2 ≥ 0, see Ionescu et al. [1999].) For matrices A and B with
the event-triggering condition is satisfied, so the desirable compatible dimensions, let γ > 0 and G(z) , (zI −
control ũ(tk ) is transmitted as the channel input sk , and A)−1 B be given. Then, the following three statements are
received by the actuator as rk = sk + nk . Then, rk is equivalent:
adopted as the true control um for the plant. The next
estimated control û(tk + 1) is set as (BRLs1). A is Schur and kGk∞ < γ.
û(tk + 1) , −Kx(tk ) . (BRLs2). The discrete-time definite constrained algebraic
Riccati equation (ARE)
Step 2. If C(ξ(tk + 1)) ≥ 0, then set tk+1 , tk + 1.
AT XA−X +I −AT XB(−γ 2 I +B T XB)−1 B T XA = 0 (6)
Step 3. If C(ξm ) < 0 for m ≥ tk + 1 before the next
has a unique stabilizing solution X 0 under the definite
triggering time, then keep both the true control and the
constraint γ 2 I − B T XB 0, that is, (A − B(−γ 2 I +
estimated control unchanged,
B T XB)−1 B T XA) is Schur, and there exist V, W such that
um = um−1 = rk , ûm+1 , ûm = −Kx(tk ).
X − I − AT XA −AT XB
T
W
When for some m > tk + 1, C(ξm ) ≥ 0, the trigger event = [W V ] 0.
−B T XA γ 2 I − B T XB VT
occurs, and this m is detected as tk+1 .
(BRLs3). There exists X 0 solving the linear matrix
Step 4. A new updating cycle begins, go to Step 1. If finite
inequality (LMI)
tk+1 does not exist, we denote tk+1 = +∞.
X − I − AT XA −AT XB
0.
−B T XA γ 2 I − B T XB
2.2 “Perturbed Linear Model”
In the following theorem, we present a sufficient condition
Define the noise sequence {dm } as for the existence of quadratic Liapunov functions.
dm = nk , tk ≤ m < tk+1 ; E{kdm k2 } = nc σ 2 . (3) Theorem 2. Consider system (4) subject to Assumption 1.
For any K such that Ap , A − BK is Schur, if the trigger
Following the perturbed linear model in Heemels and threshold λ satisfies
Donkers [2013], which ignores the channel noise, we can 1
rephrase the event-trigger strategy for control over noisy 0<λ< , (7)
kKk · kGk∞
channels described in Section 2.1 as:
where G(z) , (zI − Ap )−1 B, then there exist parameters
xm+1 = (A − BK)xm + Bdm + Bwm , (4)
γ, θ > 0 such that
where 1 1 1
λθkKk < 1, + 2 < . (8)
dm−1 , if kûm − ũm k < λkũm k; θ2 γ kGk2∞
dm = Furthermore, there exist a positive definite matrix P and
a new random variable, otherwise.
parameters ε > 0,
ũm = −Kxm , û0 = 0, and c2 , λmax (P ), and c1 , λmin (P ) > 0, (9)
ũm , if kûm − ũm k ≥ λkũm k; such that for the Liapunov function V (x) , xT P x, the
ûm+1 =
ûm , if kûm − ũm k < λkũm k. following inequalities hold
V (xm+1 ) − V (xm ) ≤ −εkxm k2 + εθ2 kwm k2 + εγ 2 kdm k2 ,
and the perturbed variable wm is introduced as (10)
wm =
0, if kûm − ũm k ≥ λkũm k; c1 kxk2 ≤ V (x) ≤ c2 kxk2 . (11)
ûm − ũm , if kûm − ũm k < λkũm k.
Proof. First, the augmented perturbed linear model is
Thus, in perturbed linear model (4) of the event-triggered presented as follows. We introduce an auxiliary parameter
control, the perturbed term wm satisfies γp > 0, and denote
" γ #
kwm k ≤ λkũm k ≤ (λkKk)kxm k. (5) γp d m
γp γp
Ap , A−BK, Bp , B B , vm , θ . (12)
γ θ γp wm
3. STOCHASTIC STABILITY RESULTS
Then, system (4) of the perturbed linear model of the
In this section, the perturbed linear model is modified to discrete-time event-triggered control over noisy channels
an augmented perturbed linear model with some scalar can be modified as an augmented perturbed linear model
conditions, which plays a vital role in demonstrating the xm+1 = Ap xm + Bp vm . (13)
existence of Liapunov functions in Theorem 2. Based
on the result, the closed-loop stability is established in Next, we study kGp k∞ , which is the following finite H∞
Theorem 3. norm of Gp (z) , (zI − Ap )−1 Bp ,
10495
19th IFAC World Congress
Cape Town, South Africa. August 24-29, 2014
−1
kGp k∞ = kG∗p k∞ = sup
BpT e−jω I − ATp
which is equivalent to (16). When γ and θ are fixed, (19)
becomes an LMI, which can be solved efficiently and
ω
r
−1 reliably by existing algorithms (Boyd and Vandenberghe
= sup λ1 (ejω I − Ap )−1 Bp BpT e−jω I − ATp [2004]). As discussed in Heemels and Donkers [2013], there
ω may exist a Pareto optimal curve (Boyd and Vandenberghe
r
1 1 [2004]) of (γ, θ) for which LMI (19) is feasible, and a heuris-
= γp + 2 · kGk∞ , tic method may be used to find the curve. However, tuning
θ2 γ
where kGk∞ is the finite H∞ norm of G(z) = (zI − the two parameters γ and θ to ensure the feasibility of (19)
Ap )−1 B, and the last equality results from the fact that may not be an easy task. While in Theorem 2, a simple
scalar tuning rule for (γ, θ) is proposed in condition (8), or
γp γp
γp T equivalently (18), which can guide the design of the value
T B
Bp Bp = B B · γθp T of the trigger threshold λ.
θ γ γ B
1 1
3.2 Stochastic Stability
= γp2 + 2 BB T . (14)
θ2 γ
We are now in a position to present the main stability
Since result.
γ 2 kdm k2 + θ2 kwm k2 = γp2 kvm k2 , Theorem 3. Consider system (1) subject to Assumptions 1
inequality (10) is equivalent to and 2. For any K making (A − BK) Schur, if the trigger
threshold λ satisfies condition (7), then the closed-loop
ε γp2 kvm k2 − kxm k2 ≥ V (xm+1 ) − V (xm )
(15) system is stable in the sense of Definition 1, that is,
= (Ap xm + Bp vm )T P (Ap xm + Bp vm ) − xTm P xm , sup E{kxm k2 } < +∞.
m∈N
It is straightforward to verify that condition (15) is guar-
anteed by the following matrix inequality: Proof. By Theorem 2, we can choose parameters θ, γ > 0
satisfying condition (8), and let P 0 and ε > 0 such
X − I − ATp XAp −ATp XBp
0, (16) that inequalities (11), (10), and (16) hold. Combining
−BpT XAp γp2 I − BpT XBp inequalities (5), (11), and (10), we obtain that c1 kxm k2 ≤
where X = 1ε P 0. Furthermore, (16) can be ensured by V (xm ), and
the following strict inequality:
X − I − ATp XAp −ATp XBp
V (xm+1 )
0. (17)
−BpT XAp γp2 I − BpT XBp ≤ V (xm ) − ε 1 − λ2 θ2 kKk2 kxm k2 + εγ 2 kdm k2 ,
!
By Lemma 1, the feasibility of (17) is equivalent to that ε 1 − λ2 θ2 kKk2
≤ 1− · V (xm ) + εγ 2 kdm k2 .
c2
r
1 1
γp > kG̃k∞ = γp + 2 · kGk∞ ,
θ2 γ
Taking expectation, and using Assumption 2 as well as (3)
which is the scalar condition and (8), we obtain that E{kxm k2 } ≤ E{V (xm )}/c1 , and
1 1 1 E{V (xm+1 )} ≤ αE{V (xm )} + nc σ 2 εγ 2 , ∀m ≥ 0,
+ 2 < .
θ2 γ kGk2∞
where α , 1 − ε 1 − λ2 θ2 kKk2 /c2 < 1.
Next, by condition (7), we can choose any θ and γ such By (16) and (9), P εI + ATp P Ap εI, c2 ≥ c1 ≥ ε.
that Therefore, α ≥ ελ2 θ2 kKk2 /c2 > 0. Defining a sequence
1 θkGk∞
θ ∈ kGk∞ , , γ>p . (18) {Em } as
λkKk θ2 − kGk2∞
Em+1 = α · Em + nc σ 2 εγ 2 , E0 = E{V (x0 )},
Thus, we have verified the existence of γ, θ > 0 satisfying and applying the comparison principle (Gil’ [2007]), we
condition (8), which guarantees the existence of the matrix obtain that
P such that inequalities (10) and (11) hold. The proof of 1 − αm
Theorem 2 is completed. E{V (xm )} ≤ Em = αm E0 + · nc σ 2 εγ 2 .
1−α
Remark 1. (i) From (8), it follows that a small γ, which
intuitively means that the influence of channel noise on Consequently, for ∀ m ≥ 0,
system trajectories is weak, will lead to a large θ, which
in turn results in small value of λ and, consequently, more E{V (xm )}
E{kxm k2 } ≤
frequent updating. c1
(ii) To find desirable P , λ, θ, and ε satisfying the conditions c2 m c2 nc σ 2 γ 2
≤ α E{kx0 k2 } + , (20)
in Theorem 2, one may resort to solving the following c1 c1 (1 − λ2 θ2 kKk2 )
matrix inequality
and supm≥0 E{kxm k2 } is upper-bounded by
P − εI − ATp P Ap −ATp P B −ATp P B
−B T P Ap εγ 2 I − B T P B −B T P B 0, c2 c2 n c σ 2 γ 2
E{kx0 k2 } + < +∞,
T
−B P Ap T
−B P B εθ I − B T P B
2 c1 c1 (1 − λ2 θ2 kKk2 )
(19) which completes the proof.
10496
19th IFAC World Congress
Cape Town, South Africa. August 24-29, 2014
γp T
Remark 2. (i) ISS stability (Sontag [2008]) has been es- γ B
γp γp
tablished in (20). If the ISS gain (sup E{kxm k2 }/(nc σ 2 )) γp2 I − γp T · P · B B
θ B γ θ
is considered as the event-triggered control performance, 1 T 1 T
I − γ 2 B P B γθ B P B
the important role of the extremal eigenvalues c1 and c2 = γp2 · 0. (24)
1
of matrix P has also been demonstrated. Although the − γθ B T P B I − θ12 B T P B
desirable parameters θ and γ can be found from condi-
tion (18), they are not sufficient to determine the control Then, applying the Schur complement, we obtain that
performance. 1
I − 2 BT P B
(ii) It is noted from (20) that the role of tuning rule (8) γ
or (18) is not only for theoretical analysis but also for −1 (25)
1 T 1 T T
implementation. When designing the event-triggered con- 2 2B PB I − 2B PB B P B.
γ θ θ
trol, one cannot make the trigger threshold λ approach
the upper bound in (7), otherwise the upper bound of
Note that B T P B = θ2 I − θ2 I − B T P B , then
E{kxm k2 } in (20) will increase to infinity and the closed-
loop stability may not be guaranteed. −1
1 T 1 T
B PB I − 2B PB BT P B
θ2 θ
4. EQUIVALENT ARI/ARE −1
1 1
= I − I − 2 BT P B I − 2 BT P B BT P B
In this section, the matrix P in the quadratic Liapunov θ θ
function will be obtained by means of ARI/ARE, which −1
will be simplified to equivalent ARI/ARE in Theorem 6. 1 T
θ2 I − θ2 I − B T P B − B T P B
= I − 2B PB
These results can facilitate further analysis of control θ
performance. = θ θ I − BT P B
4 2
−1
− θ2 I − B T P B. (26)
From the analysis in Section 3, the matrix P and the
parameter ε can be scaled. Without loss of generality, Combining (25), (26), and that θ2 I − B T P B 0 resulting
we set ε = 1. Then, under the constraints that X = from (24), we obtain that
1 2 T
ε P = P 0 and γp I − Bp XBp 0, applying the Schur
−1
complement (Boyd and Vandenberghe [2004]), we obtain γ 2 I − B T P B θ4 θ2 I − B T P B − θ2 I − B T P B;
that (16) as well as (19) can be guaranteed by the following
algebraic Riccati inequality (ARI) with definite constraint θ4
θ2 I − B T P B I ⇒ β 2 I − B T P B 0. (27)
(Damm [2004]): γ 2 + θ2
R(P ) , P − I − ATp P Ap − Conversely, if β 2 I − B T P B 0, then as the implication
−1
ATp P Bp γp2 I − BpT P Bp BpT P Ap 0, (21) in (27) is in fact an equivalence relation, we obtain that
matrix inequality (25) holds. And similarly, I− θ12 B T P B
γp2 I − BpT P Bp 0, 1 T 1 T −1 T
γ 2 θ 2 B P B(I − γ 2 B P B) B P B.
where R(·) is the Riccati operator (Damm [2004]).
Applying the Schur complement, we obtain the posi-
By Section 3.1 and Lemma 1, if condition (8) is satisfied, tive definiteness of γp2 I − BpT P Bp , which completes the
there exists P being the unique positive definite stabilizing proof.
solution to the following ARE with definite constraint:
Let γp2 I − BpT P Bp 0, and denote Mp , Bp (γp2 I −
P − I − ATp P Ap =
−1 BpT P Bp )−1 BpT , M , B(β 2 I −B T P B)−1 B T . The following
ATp P Bp γp2 I − BpT P Bp BpT P Ap , (22) proposition will be obtained.
γp2 I − BpT P Bp 0. Proposition 5. The matrices Mp and M are equal.
Although ARI (21) and ARE (22) have provided P , they Proof. By the Matrix Inversion Formulas (Zhou et al.
involve the auxiliary parameter γp and the auxiliary ma- [1996]), (β 2 I − B T P B)−1 = β12 I + β12 B T P (β 2 I −
trix Bp . We will derive equivalent ARI/ARE independent BB T P )−1 B. Then,
of γp and Bp as follows. Denote
1 1 1 −1
β,q > kGk∞ , βλkKk < 1. (23) M= 2
BB T + 2 BB T P β 2 I − BB T P BB T
1 1
β β
θ2 + γ 2 −1
1 1 1 1
= 2 BB T + 2 BB T P −1 − 2 BB T BB T .
We have the following proposition. β β β β2
Proposition 4. The following equivalence of positive defi- Similarly,
niteness holds
1
γp2 I − BpT P Bp 0 ⇔ β 2 I − B T P B 0. Mp = Bp BpT +
γp2
−1
1 T −1 1 T 1
B B
p p P − BB Bp BpT .
Proof. If γp2 I − BpT P Bp 0, γp2 γp2 γp2
10497
19th IFAC World Congress
Cape Town, South Africa. August 24-29, 2014
10498