4.4 Adaptive NN Control Design
4.4.2 Output Feedback Control
74 4 Altitude Control of Helicopters with Unknown Dynamics forj D1; 2; :::; , where
AD 2 66 66 64
0 1 0 0 0 0 1 0
0 0 0 1 1 1 2 1
3 77 77 75
; bD 2 66 66 64 0 0 ::: 0 1
3 77 77 75
(4.56)
Since belongs to a compact set, and u is bounded, we know that jy.j /j Yj. Therefore, there exists a constantt > 0such that for t > t,j.j /.t /j Dj, whereDjis a positive constant independent of. This leads toj.j /j hj WDBDj,
whereBis the normofŒ1 12::: 1T. ut
Remark 4.16. Note thatk1k converges to a neighborhood ofk, provided thatyand its derivatives up to the-th order are bounded. Hence, kC1
k is a suitable observer to estimate thekth order output derivative.
To prevent peaking [52], saturation functions are employed on the observer signals whenever they are outside the domain of interest:
is D Ni sat i
N i
; Ni max
.z;W ;Q V /2˝Q .i/; sat.a/D 8<
:
1fora <1 a forjaj 1 1 fora > 1
(4.57) foriD1; 2; :::; , whereQDŒQ1; :::;QT, and the compact set˝WD˝z˝W˝V, where˝z,˝W, and˝V are defined in (4.47), (4.48), and (4.49) respectively, denotes the domain of interest.
Now, we revisit the control law (4.24)–(4.26) and adaptation laws (4.30)–(4.31) for the full-state feedback case. Via the certainty equivalence approach, we modify them by replacing the unavailable quantities zi andZ with their estimates,zOi WD
is
i1 O˛i1 andZO WD Œ1;2s; :::;
s
1; ;Oz;˛PO1T respectively, fori D 2; :::; . Therefore, the control laws are given by
O
˛1D k1z1C Pyd; O
˛i D kiOzi Ozi1C PO˛i1; (4.58)
unnD OWTS.VOTZ/;O (4.59)
ub D kzO Oz1kb
ZOWOTSOo02
F CSOoCSOo0VOTZO 2
O z
(4.60)
Due to the fact that the actual NN is in terms ofZO while the ideal NN is in terms of Z, the following Lemma is needed.
Lemma 4.17. [35] The error between the actual and ideal NN output can be written as
WOTS.VOTZ/O WTS.VTZ/D QWT.SOo OSo0VOTZ/O C OWTSOo0VQTZO Cdu
(4.61) whereSOo WDS.VOTZ/;O SOo0 WDdiagfOs0o1; :::;sO0olgwith
O
s0oi Ds0.vOTiZ/O D ds.za/ dza
ˇˇˇzaDOvTiZO; iD1; 2; :::; l; (4.62)
and the residual termduis bounded by
du kWkSOo0VOTZOCSOo C kVkFZOWOTSOo0
F (4.63)
Accordingly, the adaptation laws are designed as WPO D W
h
.SOo OSo0VOTZ/O OzCWWO i
(4.64) VPO D V
hZOWOTSOo0zOCVVO i
(4.65) Using the backstepping procedure similar to Sect.4.4.1, and substituting (4.59), (4.64), and (4.65) into the derivative ofValong the closed loop trajectories, it can be shown that
VP
1X
jD1
kjz2j
g0C gP
2g
z2 g
X1 jD2
kjzjQzjC X1 jD2
zj.˛POj1 P˛j1/
C
1
X
jD3
zjQzj1z" QWT h
.SOo OSo0VOTZ/QO zCWWO i
trn VQT
hZOWOTSOo0QzCVVO
ioC jzjh
kWkSOo0VOTZOCSOo CkVkFZOWOTSOo0
F
iCzub (4.66)
76 4 Altitude Control of Helicopters with Unknown Dynamics From the inequalities in (4.33) and (4.34), we know that
VP
1X
jD1
kjz2j
g0C gP
2g
z2 g
X1 jD2
kjzjQzjC X1 jD2
zj.˛POj1 P˛j1/
C
1
X
jD3
zjQzj1z" QWT.SOo OSo0VOTZ/QO z OWTSOo0VQTZOzQ
CW
2 .k QWk2C kWk2/CV
2 .k QVk2FC kVk2F/C1 2kWk2 C1
2kVk2F C1 2z2
SOo0VOTZOCSOo 2CZOWOTSOo02
F
Czub
(4.67) Substituting the bounding control (4.60) into (4.66) yields
VP
X1 jD1
kjz2j
k1 2
z2
g0C gP
2g
z2 g W
2 k QWk2V
2 k QVk2F
QWT.SOo OSo0VOTZ/O zQ OWTSOo0VQTZOzQ X jD2
kjzjzQj C X jD2
zj.˛POj1
P˛j1/C X jD3
zjzQj1C W C1
2 kWk2C V C1
2 kVk2F C1 2"N2
kb1
2 SOo0VOTZOCSOo 2CZOWOTSOo02
F
.z2CzQz/ (4.68) The following lemma is useful for handling the terms containing the estimation errorszQj, forj D1; 2; :::; .
Lemma 4.18. There exist positive constantsFiandGi, which are independent of, such that, fort > t,i D1; 2; :::; 1, the estimates˛POi andzOisatisfy the following inequalities:
j PO˛i P˛ij Fi; jQzij WD jOzizij Gi: (4.69)
Proof. Starting fromiD1, we know that PO
˛1 P˛1 D k1.POz1 Pz1/D k1Q2: (4.70) Subsequently, it is easy to obtain that
PO
˛2 P˛2 D k2.Q3.˛PO1 P˛1// Q2k1Q3
D .1Ck1k2/Q2.k1Ck2/Q3DWa2T2 (4.71) wherea2WDŒ.1Ck1k2/ .k1Ck2/Tand2WDŒQ2Q3T. Suppose that
PO
˛i2 P˛i2DaTi2i2; PO
˛i1 P˛i1DaTi1i1; (4.72)
wherej DŒQ2:::QjC1Tandaj is a vector of constants, forj Di2; i1. Then, by induction, it can be shown that
PO
˛i P˛i D ki
QiC1.˛POi1 P˛i1/
Qi .˛POi2 P˛i2/ C RO˛i1 R˛i1; D ki
QiC1aiT1i1
QiaTi2i2 CaiT1Pi1DaiTi (4.73) fori D 3; :::; 1. From Lemma4.15, we know thatkik Hi, whereHi WD Œh2; :::; hiC1T, which leads to the fact thatj PO˛i P˛ij kaikHi DWFi, and thus (4.69) is proven.
To prove (4.69), note that Q
zi D Qi.˛Oi1˛i1/
D Qi.ki1zQi1 Qzi2C PO˛i1 P˛i1/ (4.74) By following a similar inductive procedure, starting fromQz1 D Q1 andQz2 D Q2 .˛O1˛1/D Q2, it can be shown thatzQi DaTzii, whereaziis a constant vector. Using the property in (4.69), it is straightforward to see thatjQzij kazikHi DWGi. The
proof is now complete. ut
Using Lemma4.18, it is clear that, fort > t, the following inequalities hold:
QWT.SOo OSo0VOTZ/O zQ
2k QWk2C 2
k OSok C k OSo0VOTZOk 2G2; (4.75)
OWTSOo0VQTZOzQ
2k QVk2FC 2
ZOWOTSOo02
FG2; (4.76)
78 4 Altitude Control of Helicopters with Unknown Dynamics
X jD2
kjzjzQj X jD2
kj 2
z2jC2Gj2 ; (4.77)
1
X
jD2
zj.˛POj1 P˛j1/ X jD2
1 2
z2j C2Fj2 ; (4.78)
X jD3
zjQzj1 X jD3
1 2
z2j C2Gj21 : (4.79) By substitution of the inequalities (4.75)–(4.79) into (4.68), it is straightforward to obtain the following expression:
VP
1
X
jD1
kjz2j
k1 2
z2
g0C gP
2g
z2
g .W / 2 k QWk2 .V /
2 k QVk2FC
k OSok C k OSo0VOTZkO 2CZOWOTSOo02
F
2G2 C
X jD2
kj
2 .z2j C2Gj2/C X jD2
1
2.z2j C2Fj2/C X jD3
1
2.z2j C2Gj21/ CW C1
2 kWk2C V C1
2 kVk2F C 1 2"N2 1
2
kb1
2 SOo0VOTZOCSOo 2CZOWOTSOo02
F z22G2 (4.80) The RHS terms can be rearranged into a more convenient form for analysis:
VP k1z211
2.k21/z22
1
X
jD3
1
2.kj 2/z2j1
2.k3/z2
g0C gP
2g z2
g .W /
2 k QWk2.V /
2 k QVk2FC X jD2
kj
2 2Gj2C X jD2
1 22Fj2 C
X jD3
1
22Gj12 C W C1
2 kWk2C V C1
2 kVk2F C 1 2"N2 1
2
kb1
2 SOo0VOTZOCSOo 2CZOWOTSOo02
F z2
2C 2 2kb1
G2
: (4.81)
Finally, by appropriately choosing the control parameterskandkbas follows:
k2 > 1; k3; :::; k1> 2; k> 3; kb> 1
2; W; V > ; (4.82) it can be shown that
VP c1VCc2K
z2c3 ; (4.83)
where c1WDmin
2k1; .k21/; .k32/; :::; .k12/; g.k3/; .W / max.W1/; .V /
max.W1/
; (4.84)
c2WD 1
2"N2C W C1
2 kWk2CV C1 2 kVk2F C1
22 0
@X
jD2
Fj2C
1
X
jD2
.kjC1/Gj2CG2 1
A; (4.85)
c3WD
C 2
2kb1
G2; (4.86)
KWD 1 2
kb1
2 SOo0VOTZOCSOo 2CZOWOTSOo02
F
: (4.87)
It can be shown that
VP.t / co1V.t /Cco2; t t co1WD min
(
2min.K1/; min
K232I
max.M / ; miniD1;2;3f4ik Qik2g max.1/
)
(4.88) co2WD
X3 iD1
1C i
2 kik2C 1
2k"k2C2max.KN2KN2TC/2 (4.89) To ensure that > 0, the control gainsK1 and K2 are chosen to satisfy the following conditions:
min.K1/ > 0; min
K23 2I
> 0: (4.90)
80 4 Altitude Control of Helicopters with Unknown Dynamics We are ready to summarize our results for the output feedback case under the following theorem.
Theorem 4.19. Consider the helicopter dynamics (4.1) under Assumptions 4.1–
4.5, with output feedback control laws (4.59)–(4.60), adaptation laws (4.64)–(4.65), and high gain observer (4.53) which is turned on at timetin advance. For initial conditions .0/, .0/, W .0/,Q V .0/Q starting in any compact set ˝0, all closed loop signals are SGUUB, and the tracking error z1 converges to the steady state compact set:
˝z1 WD (
z12R ˇˇˇˇ ˇjz1j
s 2cN2
c1
)
; (4.91)
wherecN2WDc2Cc3K, andN c1is as defined in (4.84).
Proof. We consider the following two cases for the stability analysis:
Case 1:jzj>p c3
For this case, the last term of (4.83) is negative, thus yielding
VP c1VCc2; (4.92)
which straightforwardly implies that all closed loop signals are SGUUB, according to Lemma4.9. However, whenjzj p
c3, the last term of (4.83) may not be negative, leading to a more complicated analysis, as shown in the subsequent case.
Case 2:jzj p c3
For this case, we want to show that, as a result of z being bounded, the function Kin (4.87) is also bounded, for which (4.83) can be expressed in the form of (4.9), convenient for establishing SGUUB property. To this end, note that the derivative of V1is given by
VP1
1
X
iD1
kiz2i
1
X
iD2
kizizQiC
1
X
iD2
zi.˛POi1 O˛i1/C
1
X
iD3
ziQzi1Cz1z
(4.93) According to Lemma4.18, we can show that
VP1
1
X
iD1
kiz2i C
1
X
iD2
kijzijGi C
1
X
iD2
jzijFiC
1
X
iD3
jzijGi1C jz1jp c3
k1z211
2.k21/z22 X2 iD3
1
2.ki 2/z231
2.k3/z21
C1 2c3C1
2 X1 iD3
2Gi21C 1 2
X1 iD2
2Fi2C1 2
X1 iD2
ki2Gi2
c4V1Cc5; (4.94)
where the positive constantsc4andc5are defined by
c4WDminf2k1; k21; k32; :::; k12; k3g; (4.95) c5WD
2
"
c3C
1
X
iD2
Fi2C
2
X
iD2
.1Cki/Gi2Ck1G12
!#
: (4.96)
This implies that z.t /satisfies the inequality kz.t /k
s 2
V1.0/C c5
c4
Cc3 DW Nz: (4.97)
According to Lemma4.11, it follows from the boundedness of z.t /that the internal states.t /are also bounded, i.e.,
k.t /k a1.zNC Nd/Ca2DW N; (4.98) wherekd.t /k Nd for constantNd > 0, based on Assumption4.1. Thus, the vector of NN inputsZO is also bounded as follows
ZO
1d C Nz;N2
; :::; N
1;;N p
c3CG;˛NP1CF1 T
DW NOZ (4.99) where the constant ˛NP1 > 0 is an upper bound for ˛P1
z; 1d; 1d.1/; :::; 1d./ . Exploiting the properties of sigmoidal NNs [30], it can be shown that
SOo0VOTZOCSOo
F 1:224p
l: (4.100)
As a result, from the adaptation law (4.64), the dynamics of the neural weights WO DŒWO1; :::;WOi; :::;WOlTcan be shown to satisfy the inequality
WPO min.W/WWO C1:224p
lmin.W/.p
c3CG/; (4.101) which results in
k OW .t /k k OW .0/k C 1:224p l.p
c3CG/
W DW NOW; (4.102)
82 4 Altitude Control of Helicopters with Unknown Dynamics
whereWNO is a positive constant. Accordingly, from (4.87), and the fact thatkSo0kF 0:25 p
l, we can show thatKis bounded as follows:
K l 2
kb1
2 1:498C0:0625
ZNOW NO 2
DW NK; (4.103)
whereKN is a positive constant. From (4.83) and (4.103), we obtain that
VP c1VC Nc2; (4.104)
wherecN2WDc2Cc3KNis a positive constant.
Having obtained (4.104) for Case 2, we can compare it with (4.92) of Case 1 to see that (4.104) describes a larger compact set in which the closed loop signals remain, by virtue of the fact that cN2 c2. Hence, the performance bounds can be analyzed from (4.104), as a conservative approach. A nice property is that as diminishes to zero, we havecN2 ! c2, and the performance can be analyzed from (4.92) instead, albeit conservatively.
Based on (4.104), we can directly invoke Lemma4.9to conclude SGUUB for all closed loop signals. Since it is straightforward to prove that the tracking error z1 D11d converges to the compact set˝z1, by following the steps outlined in the proof of Theorem4.12, we have omitted the proof. ut Remark 4.20. It follows from Theorem4.19that the size of the steady state compact set˝z1, to which the tracking error converges, depends on the ratioccN2
1, which contain tunable parameters. Thus, we can reduce the size of˝z1 by appropriately choosing the parameters. For instance, by choosing the control gainsk1; :::; k large and the observer parametersufficiently small, the ratio ccN2
1 can be decreased, to the effect that˝z1 diminishes.