Professional Documents
Culture Documents
Sliding Mode
Sliding Mode
Five-DOF
Five DOF Active Magnetic Bearing
Control System
Power Amplifier
Electromagnet
g
Rotor
Sensor
Controller Controller
Sensor
RAMB TAMB
Electric Machinery and Control Laboratory,
7 EE, NCU, Taiwan
Source: Schweitzer, Waukesha
Introduction
Various applications of AMBs:
Ω Shaft Imbalance Load
θ Torque
AMB1
AMB2
AMB3
Aerodynamic
Loads
Helicopter and Tiltrotor Flywheels
2 Stage Direct
Inverter Drive Centrifugal
Impellers
Compressor
Motor and
Bearingless
Bearing Controls
Heart
Motor Hard Drive
PMB
Pump
High-Speed
g p Spindle
p
ib+ ic ib ic
Rotor Z
z
Power Line
Platform
Left RAMB Induction Motor TAMB Right RAMB
Position
Signals Line Mass Disk Thrust Disk
Rotor 16.6 mm
TAMB Position
Sensor
Nominal Air Gap z0 : 0.5mm
300 mm 200 mm
Source: S2M
Thus it’s important to develop tracking controllers for TAMB in practical applications
Thus, applications.
Electric Machinery and Control Laboratory,
15 EE, NCU, Taiwan
TAMB System
Drive System for TAMB System
A drive system
y includingg DDM and ppower amplifiers
p of TAMB is shown in
Fig. 2.6 in which the dotted line or solid line of the thrust disk means that the
thrust disk is centered in the air gap or deviated from the center in Z-axis,
respectively.
The total nonlinear attractive electromagnetic forces for Z-axis is modeled as:
Power Amplifier Coil i0 + iz
Fz = f 9 − f10
⎡ (i0 + iz )2 (i0 − iz )2 ⎤
Right
v0 + v z Z Fz Electromagnet
f9 = ka ⎢ − 2⎥
v0
++
Σ (
⎣ 0 z − z )2
( z + z ) ⎦
z0 z0 − z
Position 0
Sensor
vz
where:
Proposed
Control Laws
Thrust Disk z m k : electromagnet constant of TAMB
a
vz z +z i0: bias current
− Rotor z0 0
v0 + ib: control current
Σ f10
v0 − v z Left zo: nominal air gap
Electromagnet
z: rotor position in Z
Z-axis
axis
i0 − iz
Electric Machinery and Control Laboratory,
16 EE, NCU, Taiwan
TAMB System
Dynamic Model of TAMB System
Usingg the Newton’s law, the dynamic
y model of the TAMB control system
y as
shown in Fig. 2.5 can be described as follows:
m&z& + cz& − f dz = Fz
By taking the Taylor Taylor’ss expansion with respect to the nominal operating
position ( z = 0, iz = 0 ), the nonlinear electromagnetic force can be represented
by the following linearized electromagnetic force model:
Fz ≅ k ap z + k ai i z
where Thrust Disk
Z
Axial Direction
∂Fz ( z, iz ) i02
k ap ≡ = 4k a 3
∂z z = 0, iz = 0 z0 Rotor 16.6 mm
∂F ( z , i z ) i TAMB Position
∂i z z = 0, i z = 0 z0
Now, the dynamic model of the TAMB system using the linearized
electromagnetic force model can be described as follows:
m&z& = −cz& + k ap z + k ai i z + f dz Electric Machinery and Control Laboratory,
17 EE, NCU, Taiwan
TAMB System
Experimental Design for TAMB Control System
A. Experimental
p Setupp
The experimental setup for the TAMB control system including the TAMB
system, eddy-current position sensor, drive system with DDM and power
amplifiers,
p , and PC is shown in Fig.
g 2.10.
Rotor 16.6 mm
0 .0 5 m m
S ta rt
R e fe re n c e
T r aj ec t o ry
TAMB Position
0.38 kg 0.38 kg
0 .0 5 m m - 0 .1 m m 2Sec
Sensor
Nominal Air Gap z : 0.5mm 0
ib + i y1
Electromagnetic
Stator
130 mm x1 115 mm
ib − ix1 ib + ix1
Nominal Rotor
Air Gap ib − i y1
(0.4 mm)
100 mm
180 mm
+ v
y1
Position Sensor
Upper
( x
⎢⎣ b 1 − x )2
( x b + x 1 ) ⎥⎦
Proposed
Control Laws Left f2
Electromagnet
Right
Fy1 ≡ f 2 − f 4 = k r ⎢
(
⎡ ib + i y 2
−
) ib (
− i y )
2
⎤
2⎥
1 1
− v y1 Electromagnet Electromagnet
( y − y )2
( y + y )
+ Σ
vb − v y1 ⎢⎣ b 1 b 1 ⎥⎦
Y
f3 f1
vb vb − v x1 X where:
+ xb x1 kr: electromagnet constant of RAMB
Σ
− vx y1 ix , iy : bias currents in X-Y axes
1
ib − ix1 yb f 4 ib + ix1
b b
⎪ 1 2
Fz = f 9 − f10 θx =
y1 − y2 y1 − y2
=
a+b l
≅ k ap z + k ai iz
Electric Machinery and Control
x − x Laboratory,
x −x
θy = 2 1 = 2 1
23 aEE,
+ b NCU,
l Taiwan
Five-DOF AMB System
The dynamic model describing the motion of the rotor of the five-DOF AMB
system about the CG can be represented as follows
⎧m&x&c = Fx + Fx + f dx
⎪ 1 2
the rotor is rigid and the displacements are small
⎪m&y&c = Fy1 + Fy2 − mg + f dy bx1 + ax2 bx1 + ax2 y1 − y2 y1 − y2
⎪ xc = = θx = =
⎨m&z&c = Fz + f dz a+b l a+b l
⎪ && & by1 + ay2 by1 + ay2 x −x x −x
⎪ Jθ y − J zωθ x = − aFx1 + bFx2 + cf dx yc = = θy = 2 1 = 2 1
⎪ Jθ&& + J ωθ& = aF − bF − cf a+b l a+b l
⎩ x z y y1 y2 dy
Substituting
S b tit ti above
b equalities
liti into
i t the
th dynamic
d i model
d l off the
th five-DOF
fi DOF AMB
system, the dynamic model can be represented as:
⎧ aJ zω ⎛ 1 a2 ⎞
⎪ &x&1 + ( y&1 − y& 2 ) = ⎜⎜ + ⎟⎟ Fx1 + ⎛⎜ 1 − ab ⎟⎞ Fx2 ⎛ 1 ac ⎞
+ ⎜ − ⎟ f dx
⎪ Jl ⎝m J ⎠ ⎝m J ⎠ ⎝m J ⎠
⎪ ω ⎛ 1 b2 ⎞
bJ ⎛ 1 ab ⎞ ⎛ 1 bc ⎞
⎪ &x&2 − z
( y&1 − y& 2 ) = ⎜ − ⎟ Fx1 + ⎜⎜ + ⎟⎟ Fx2 + ⎜ + ⎟ f dx
⎪ Jl ⎝m J ⎠ ⎝m J ⎠ ⎝m J ⎠
⎪
⎪ aJJ zω ⎛ 1 a2 ⎞
⎨ &y&1 + (x&2 − x&1 ) = ⎜⎜ + ⎟⎟ Fy1 + ⎛⎜ 1 − abb ⎟⎞ Fy2 ⎛ 1 ac ⎞
+ ⎜ − ⎟ f dy − g
⎪ Jl ⎝m J ⎠ ⎝m J ⎠ ⎝m J ⎠
⎪
⎪ &y&2 − bJ zω ( x&2 − x&1 ) = ⎛⎜ 1 − ab ⎟⎞ Fy + ⎜ 1 + b ⎟ Fy
⎛ 2
⎞ ⎛ 1 bc ⎞
⎜ ⎟ + ⎜ + ⎟ f dyy − g
⎪ Jl ⎝m J ⎠ 1 ⎝m J ⎠ 2 ⎝m J ⎠
⎪
⎪&z& = 1 F + 1 f Electric Machinery and Control Laboratory,
⎪ m
z
m
dz
⎩ 24 EE, NCU, Taiwan
Five-DOF AMB System ⎡ Fx1 ⎤ ⎡ f1 − f 3 ⎤ ⎡ k rp x1 + k rii x1 ⎤
⎢F ⎥ ⎢ f − f ⎥ ⎢ k y + k i ⎥
⎢ x2 ⎥ ⎢ 2 ⎥ ⎢ rp 1 ri y1 ⎥
Power Line
Position Signals
Line Decentralized Power
AD DA
Control Systems Amplifiers
Card Card DDM
PC ×5 ×10
Position Signals PCI Bus PCI Bus Control Currents Drive System
vx1 vb+vx1
Control subroutine with 4 kHz sampling
︰ rate vb-vx1 Electric Machinery and Control Laboratory,
︰ 27 EE, NCU, Taiwan
Five-DOF AMB System
Five-DOF Control with Independent Loops:
It is noted that since the adopted rotor are relative large-scale size with longer
l th andd heavier
length h i weight i ht comparing
i withith the
th others,
th th rotor
the t speedd 4800RPM
is a relative high speed especially that five-DOF are controlled simultaneously.
C. Performance Measures and Comparisons
To measure the control performances, the root mean square (RMS) value TRΦ
and peak to peak value TPΦ of regulating errors are defined as follows:
n
∑ [φ ( N )]2
;
TRφ = N =1
n
TPφ = mφ − sφ where mφ = max(φ ( N )), sφ = min (φ ( N ));
N N
Φ can be x1, x2, y1, y2, and z.
Electric Machinery and Control Laboratory,
29 EE, NCU, Taiwan
Outline
Abstract
Introduction
System Dissection, Dynamic Analyses, and Experimental Designs
of Five-DOF AMB System
y
TAMB System
Five-DOF AMB System
Precise Tracking Control of TAMB System
Using Model-Free Control Methods
Using Adaptive Complementary Sliding
Sliding-Mode
Mode Control
Using Robust Non-Singular Terminal Sliding-Mode Control
Robust Control of Fully Suspended Five-DOF AMB System
Using Decentralized PID Neural Network Control
Using Decentralized Intelligent Double Integral Sliding-Mode Control
Di
Discussions
i andd Conclusions
C l i
Electric Machinery and Control Laboratory,
30 EE, NCU, Taiwan
Precise Tracking Control of TAMB System—
Using
g PID and RNN controls
According to the regulation of control currents, the rotor position on
the axial direction of a TAMB systemy can be controlled for the
tracking of various reference trajectories.
A. PID Control yo
τ
u (t ) = K P e(t ) + K I ∫ e(t )dτ + K D e&(t ) Σ Layer 3
0
W ( N + 1) = W ( N ) + ΔW
2 2 2 (Output Layer)
B. RNN Control j
W j2
j j
∂E ∂yo ∂net 3
ΔWij1 = −ηWR 1 o
Γj
∂yo ∂neto3 ∂W 1
ij
.. ... Layer 2
[ ]
z −1
S z −1
S z −1
S z −1
S
= η δ oW Γ j 1 − Γ j xi
R
W1 j
2
γ
Wijγ1 ( N + 1) =γ Wij1 ( N1 ) + ΔWij1
γ (Hidden Layer)
Wij
∂E ∂yo ∂neto3
ΔW j2 = −ηWR 2 1 1
∂yo ∂neto3 ∂W j2
xi
= ηWR 2 δ o Γ j S S Layer 1
(Input Layer)
δ o ≅ ( z m − z ) + ( z&m − z& ) = e + e& e e&
Supervised Control
Electric Machinery and Control Laboratory,
31 EE, NCU, Taiwan
Precise Tracking Control of TAMB System—
Using
g HPBRNN Control
C. HPBRNN Control
All the activation functions of hidden neurons use different orthonormal
Hermite polynomial basis function (OHPBF).
The feedback of the output layer with self-connections is added to improve the
convergence
g speed
p and pprecision.
yo
H1 (t ) = 1 t2
1 − Output Layer Feedback 1
h j (t ) = e 2
H j (t ) ∑ Layer 3
H 2 (t ) = 2t 2 j j! π W2
Z −1
(Output Layer)
W 1 yr
M j
z −1
S
H n (t ) = 2tH n −1 (t ) − 2(n − 1) H n − 2 (t ), n ≥ 3 Γj β
z −1
z −1 .. z
− 1 ... z -1
Layer 2
h1 h2 h3 h4 (Hidden Layer)
α α α α
universal
1 approximator
1
xi
h5 h6 h7 h8 Layer 1
(Input Layer)
e e&
Electric Machinery and Control Laboratory,
32 EE, NCU, Taiwan
Precise Tracking Control of TAMB System—
Using
g HPBRNN Control
To obtain the leaning algorithm for the HPBRNN, the BP learning rule is also
adopted here.
∂E H ∂E ∂yo ∂net o
3
ΔW = −η
1
j
H
W1
= −ηW 1 = ηWH1δ o Γ j
∂W j1
∂yo ∂neto ∂W j
3 1
E ( N + 1) < E ( N )
∂E H ∂E ∂yo ∂nett
3
ΔW 2 = −ηWH2 = −η 2 =
o
η H
δ y
W2 o r
yo
∂W 2 W
∂yo ∂neto ∂W
3 2
W j1 ( N + 1) = W j1 ( N ) + ΔW j1 S z −1
E(N ) E( N )
ηWH
1 = η H
= xi
⎧ n ⎡ ∂E ( N ) ⎤ 2
⎫ W2
⎧⎡ ∂E ( N ) ⎤ 2
⎫ Layer 1
⎪ ⎪ ⎪ ⎪
2⎨∑ ⎢
(Input Layer)
⎥ + ε ⎬ 2 ⎨ ⎢ ⎥ + ε ⎬
⎪⎩ j =1 ⎢⎣ ∂W 1
j ⎥⎦ ⎪⎭ ⎪⎩⎢⎣ ∂
∂W
W 2
⎥⎦ ⎪⎭ e e&
Electric Machinery and Control Laboratory,
33 EE, NCU, Taiwan
Precise Tracking Control of TAMB System—
Using
g HPBRNN Control
D. Comparison of RNN and HPBRNN
The pproposed
p HPBRNN can pperform better function approximation
pp due to its
activation functions with different OHPBFs.
Γ1 h1
1
00.88
88 00.31
0.7
06
0.6
0.5
× 0.48 sets and different strengths. × 0.48 sets (OHPBFs) and strengths.
0.4
0.3
0.2
0.1
0. 4224 0.1488
0
2.5696
-5 -4 -3 -2 -1 0 1 2 3 4 5
1
Γ2 h2 1.5401
ΔW j2 = ηWR 2 δ o Γ j ΔW j1 = ηWH1δ o Γ j
0.9
0.8
0.7
yo
0.6
0.5
0.4
0.88 yo 0.62
0.3
0.2
0.1
× 1.58 × 1.58 Output Layer Feedback 1
∑
Σ
0
1.3904 0.9796
-5 -4 -3 -2 -1 0 1 2 3 4 5
Γ3 h3 W2
Z −1
2 W 1 yr
0.48 1.58 0.11
W j 0.75 0.48 1.58 0.11 0.75 j
0.9
0.8
0.7 z −1
S
0.88 0.47
0.5
j j
.. ...
0.4
.. ...
0.3
× 0.11
z −1
z − 1
z
× 0.11 α
z −1
S
-1
0.2
z −1
S z −1
z −1
S z −1
S
α α α
0.1
0
-5 -4 -3 -2 -1 0 1 2 3 4 5
γ γ γ γ
︰ 0 0968
0.0968 W 1
ij
︰ 0 0517
0.0517 1 1
1 1
Γ8 h8
1
xi
0.9
xi
0.88 0.48
0.8
S S
0.7
0.6
× 0.75
0 75 × 0.75
0 75
0.5
0.4
03
0.3
0.2
e e& e e&
0.66
0.1
0
-5 -4 -3 -2 -1 0 1 2 3 4 5
0.36
2 2
Electric Machinery and Control Laboratory,
34 EE, NCU, Taiwan
Precise Tracking Control of TAMB System—
Using
g HPBRNN Control
The configuration of the TAMB control system using HPBRNN with on-line
learning algorithm and adaptive learning rates is shown as follows:
e e& e e&
ηWH
ε On-line Learning
1
Varied Learning
η H
Algorithms
Rates (3.29)&(3.30) W2
(3.22)-(3.25)
Supervised Γj yr
ΔW j1 , ΔW 2
Control
z
−1
e vz + i0 + iz
zm + z
−1
∑
Power Ri ht Coil
Right C il
∑ + _
..
e& Amplifier
z
− 1
_
z d dt ∑
i0 − iz
Left Coil z
...
+
z
-1
TAMB system
−1
v0 DDM
HPBRNN
0A 0A 0A 0A
0.5A Total Current i0 + iZ 2Sec 0.5A Total Current i0 + iZ 2Sec 0.5A Total Current i0 + iZ 2Sec 0.5A Total Current i0 + iZ 2Sec
0A 0A 0A 0A
( )
(c) (f) ( )
(c) (f)
0A 0A 0A 0A
0.5A Total Current i0 + iZ 2Sec 0.5A Total Current i0 + iZ 2Sec 0.5A Total Current i0 + iZ 2Sec 0.5A Total Current i0 + iZ 2Sec
0A 0A 0A 0A
( )
(c) (f) (c) (f)
0A 0A 0A 0A
0.5A Total Current i0 + iZ 2Sec 0.5A Total Current i0 + iZ 2Sec 0.5A Total Current i0 + iZ 2Sec 0.5A Total Current i0 + iZ 2Sec
0A 0A 0A 0A
(c) (f) (c) (f)
⎛d ⎞⎛ d ⎞
s c (t ) = ⎜ + λ ⎟⎜ − λ ⎟Λ(t )
⎝ dt ⎠⎝ dt ⎠
Then, define the sliding surfaces summation variable σ as follows:
σ (t ) = s g (t ) + sc (t )
= 2[e&(t ) + λe(t )]
= 2[s(t )] Electric Machinery and Control Laboratory,
44 EE, NCU, Taiwan
Precise Tracking Control of TAMB System—
Using
g CSMC
Theorem 4.1: Considering the TAMB system represented by (4.2), if the
proposed CSMC control law UCSMC, which is composed of an equivalent
control
t l law
l Ueq and d a hitting
hitti control
t l law
l Uhit, is
i adopted
d t d as the
th control
t l effort
ff t ,
then the GUUB stability of the proposed CSMC system can be guaranteed.
U CSMC (t ) = U eq (t ) + U hit (t )
−1
[
U eq (t ) = Gn × &z&m (t ) − An z& (t ) − Bn z (t ) + 2λe&(t ) + λ2e(t ) + λsg (t ) ] sg=sc=0
−1 ⎡ ⎛ σ (t ) ⎞⎤ sc=0
U hit (t ) = Gn × ⎢δ sat⎜ ⎟⎥ sg=0
⎣ ⎝ Φ ⎠⎦ SC=0
Proof: Choose the Lyapunov function candidate VCSMC as 5
[
1
]
VCSMC ( s g (t ), sc (t )) = s g2 (t ) + sc2 (t )
2 ee& 00
Taking the time derivative of the first Lyapunov –5
–5
function, and using above equations, then 5 00 e
00 e&
V&CSMC ( s g (t ),
) sc (t )) = s g (t ) s&g (t ) + sc (t ) s&c (t ) 5
∫∫ tedt
edtdτ –5
0
…
SMC CSMC
≤ −ξ σ (t ) ≤ 0
s(t)≦Φ σ(t)≦Φ
=>2 s(t)≦Φ
whenever σ (t ) > Φ .
CSMC owns half guaranteed ultimate
Only( sthe
VCSMC g (t GUUB
), s c (t )) = [
1 2
stability
2
s g (t ) +
is s 2
]
guaranteed.
c (t )
Electric bounds compared
Machinery andwith SMC. Laboratory,
Control
45 EE, NCU, Taiwan
Precise Tracking Control of TAMB System—
Using
g ACSMC
C. MIMO RHNN Estimator
In the ACSMC system,y , two complicated
p dynamic
y functions are estimated by
y
two outputs of the MIMO recurrent Hermite neural network (RHNN) estimator
directly.
Here,, the output
p vector of the MIMO RHNN is P = α ∗T Γ + ε Q = β ∗T Γ + ε
p q
ˆ ˆ
[ y1 y2 ] = [ P Q ] y y 1 2
y = Qˆ (e | βˆ ) = βˆ T Γ
2
z −1
z −1 .. z −1 ...
α
z −1
Layer 2
α α α (Hidden Layer)
where Γ = [ Γ1 Γ 2 L Γ n ]T,
1
T e1 e2
Q = Q ∗ (e | β ∗ ) + ε q = β ∗ Γ + ε q
optimal weighting vectors minimum approximated errors
Electric Machinery and Control Laboratory,
46 EE, NCU, Taiwan
Precise Tracking Control of TAMB System—
Using
g ACSMC
D. Adaptive Complementary Sliding-Mode Control (ACSMC)
The p proposed
p ACSMC with the MIMO RHNN estimator p promotes the control
system from GUUB to asymptotic stability without the uncertainty bound.
First, reformulate the term sg-λsc as follows:
s&g (t ) − λsc (t ) = &z&m (t ) − An z& (t ) − Bn z (t ) − GnU (t ) − L( z; t ) + 2λe&(t ) + λ2e(t ) − λsc (t )
= H (z; t ) − GnU (t ) + λΞ(z; t ) − λGn sc (t )
[
= Gn Gn H (z; t ) − U (t ) + Gn λΞ(z; t ) − λsc (t )
−1 −1
]
= Gn [P(t ) − U (t ) + λQ(t ) − λsc (t )]
y1 y2
where H (z; t ) = &z&m (t ) − An z& (t ) − Bn z (t ) − L( z; t ) , P(t ) = Gn H (z; t ) ,
−1
ε&ˆq = η q λσ (t(t )
Proof: Choose the Lyapunov function candidate VACSMC as
VACSMC
CS C ( s g (t ), s c (t ), α p q
2
[ g ]
~, β~, ε~ , ε~ ) = 1 s 2 (t ) + s 2 (t ) + Gn α~ T α~ + Gn β~ T β~ + Gn ε~ 2 + Gn ε~ 2
c
2ηα 2η β 2η p
p
2η q
q
~
where α~ = α ∗ − αˆ , β = β ∗ − βˆ , ε~p = ε p − εˆ p , and ε~q = ε q − εˆq .
Taking the time derivative of VACSMC, and using Barbalat’s Lemma. Then,
V& = −λG σ 2 (t ) ≤ 0 , lim σ 2 (t ) = 0 . asymptotic stability is guaranteed
ACSMC n
t →∞
Electric Machinery and Control Laboratory,
48 EE, NCU, Taiwan
Precise Tracking Control of TAMB System—
Using
g ACSMC
The configuration of the TAMB control system using ACSMC system with
MIMO RHNN estimator is shown as follows:
d e& v0
zm + e
dt
Integral sg
∑ Sliding Surface DDM
z − ∫e e1
(4.15)
+ σ +
∫ + ∑
∑ +
+ U e+ vz −
Complementary MIMO RHNN ∑ ∑
Sliding Surface Estimator +
((4.17)) sc (4.38) εˆ& p , ε&ˆq Uc
Robust Power
Pˆ (e | αˆ ) Qˆ (e | βˆ ) Compensator Amplifier
(4.39)
Σ Σ
εˆ p εˆq λ i0 − iz i0 + iz
..
Adaptive Laws
(4.42) & (4.43) Lower Upper
RHNN
d Coil Coil
dt
& λ
α&ˆ , βˆ Estimation Laws TAMB System
(4.40) & (4.41)
z
λ
ACSMC System
0.1mm 0.1mm
0.05mm Rotor Position 0.05mm Rotor Position
Reference Reference
Start Trajectory Start Trajectory Start Start
Reference Reference
Trajectory Trajectory
0.05mm -0.1mm Rotor Position 2Sec 0.05mm -0.1mm Rotor Position 2Sec 0.05mm -0.1mm 2Sec 0.05mm -0.1mm 2Sec
(a) (e) (a) (e)
Tracking Error Tracking Error Tracking Error Tracking Error
0V 0V 0V 0V
0A 0A 0A 0A
0.5A Total Current i0 + iZ 2Sec 0.5A Total Current i0 + iZ 2Sec 0.5A Total Current i0 + iZ 2Sec 0.5A Total Current i0 + iZ 2Sec
0A 0A 0A 0A
(d) (h) (d) (h)
Integral Sliding Surface Integral Sliding Surface Integral Sliding Surface Integral Sliding Surface
0V 0V 0V 0V
Complementary Sliding Surface Complementary Sliding Surface Complementary Sliding Surface Complementary Sliding Surface
0V 0V 0V 0V
0A 0A 0A 0A
0.5A Total Current i0 + iZ 2Sec 0.5A Total Current i0 + iZ 2Sec 0.5A Total Current i0 + iZ 2Sec 0.5A Total Current i0 + iZ 2Sec
0A 0A 0A 0A
(e) (j) (e) (j)
0V 0V 0V 0V
0V 0V 0V 0V
0A 0A 0A 0A
0.5A Total Current i0 + iZ 2Sec 0.5A Total Current i0 + iZ 2Sec 0.5A Total Current i0 + iZ 2Sec Total Current i0 + iZ
0.5A 2Sec
0A 0A 0A 0A
(e) (j) (e) (j)
sm (t ) = e&(t ) + λe(t ) p
where λ is a designed positive constant; p and q are both positive odd integers
which should satisfy the following condition p>q.
If the
th control
t l effort
ff t U(t) off the th TSMC control t l law
l UTSMC(t) is
i designed
d i d as:
−1
⎡ q
q
−1 ⎤
U TSMC (t ) = Gn ⎢&z&m (t ) − An z& (t ) − Bn z (t ) + λ e(t ) e&(t ) + δ sgn( sm (t ))⎥
p
⎣⎢ p ⎥⎦
Then, the following stability result can be obtained:
1 d
sm (t ) 2 < −ξ sm (t )
2 dt
Thus, the tracking error e will reach the terminal sliding surface sm in finite
time tr as shown in the following
sm (0)
tr ≤
ξ Electric Machinery and Control Laboratory,
55 EE, NCU, Taiwan
Precise Tracking Control of TAMB System—
Using
g TSMC
On the other hand, when the terminal sliding surface sm=0 is reached, the
system dynamic e&(t ) is
q q
e&(t ) = −λe(t ) p
sm (t ) = e&(t ) + λe(t ) = 0
p
λ ∫e (t
ts = − e p
de = e(t r ) p sm (0)
r ) λ ( p − q)
may occur a singularity problem if e&(t ) ≠ 0 when e(t ) = 0 due to the power of e(t )
ex: p=5,
5 q=33
is negative.
⎡ q
q
−1 ⎤ 1
U TSMC (t ) = Gn ⎢ &z&m (t ) − An z& (t ) − Bn z (t ) + λ e(t ) e&(t ) + δsgn ( sm (t ))⎥ 5 2 e&(t )
−1 p
⎢⎣ p ⎥⎦ e (t )
Therefore, the TSMC can not guarantee a bounded control effort if e&(t ) ≠ 0
when e(t ) = 0. e&
sm (0)
Th h the
Though h singularity
i l i problem bl only
l
occurs in the reaching phase when
the condition q<p<2q holds, it is singularity problem
noted
t d that
th t the th singularity
i l it problem
bl
may also occur even after the e
terminal sliding surface sm(tr+ts) is sm (t s )
reached due to the computational non-singularity
i l it problembl
errors and uncertain factors. sm (t r )
∵ e&(t ) = −λe(t )
q/ p >0
2q− p
singularity
g y problem
p
∴ e(t ) q / p −1
e&(t ) = −λe(t ) p sm = 0
Electric Machinery and Control Laboratory,
57 EE, NCU, Taiwan
Precise Tracking Control of TAMB System—
Using
g NTSMC
B. Non-Singular Terminal Sliding-Mode Control (NTSMC)
Now,, the NTSMC is further pproposed
p to overcome the singularity
g yp
problem of
the TSMC. Define the non-singular terminal sliding surface sn as
p
1
sn (t ) = e(t ) + e&(t ) q
λ
Considering the non-singular terminal sliding surface sn, the TAMB system is
stable if the control effort U(t) uses NTSMC control law UNTSMC(t) designed as:
−1
⎡ q
2q− p
⎤
U NTSMC (t ) = Gn ⎢&z&m (t ) − An z& (t ) − Bn z (t ) + λ e&(t ) q
+ δ sgn( sn (t ))⎥
⎢⎣ p ⎥⎦
It is obvious that the forth term of the NTSMC control law UNTSMC will not
result in the singularity problem as lone as the condition q<p<2q holds.
T
L = y * (e | W ∗ ) + ε = W ∗ Γ + ε Σ Σ Layer 3
(Output Layer)
α α α α (Hidden Layer)
with superior approximated ability is
1 1
employed to directly estimate the lumped
uncertainty of the TAMB on-line. xi
Layer 1
Thus, the exact value of the bound of the (Input Layer)
⎢⎣ p ⎥⎦ p
&ˆ sn (t ) p −1
using estimation law W = −ηW
−1 q
U RHNN (t ) = −Gn yˆ (t ) e&(t ) Γ
λ q p
sn (t ) p −1
U r (t ) = −Gn [εˆ − ζsn (t )] &
−1
using adaptive law εˆ = −ηε e&(t ) q
λ q
~
V&RNTSMC ( s (t ), W , ε~) = −ζs (t ) 2
n n
lim ξs (t ) 2 = 0 asymptotic stability is guaranteed
t →∞
n
Electric Machinery and Control Laboratory,
60 EE, NCU, Taiwan
Precise Tracking Control of TAMB System—
Using
g RNTSMC
The configuration of the TAMB control system using RNTSMC system with
MISO RHNN estimator is shown as follows:
RNTSMC System Drive System
v0
d e& sn RHNN Uncertainty U RHNN DDM
Non-Singular
zm + e
dt
Estimator +
∑
Sliding Surface
(5.14)
(5.26) + ∑
z − λ ŷ & U eq + U RNTSMC +
Wˆ ∑
−
∑
+ vz
d
dt
⎛ d ⎞2 Equivalent RHNN Σ
+
⎜ ⎟
⎝ dt ⎠ Control Law Ur ε&ˆ
(5.25) Power
Robust Amplifier
Controller
(5 27)
(5.27)
i0 − iz i0 + iz
εˆ ζ
Left Right
S l t d
Selected Coil Coil
Parameters Estimation Law Adaptive Law
p, q, λ (5.28) (5.29)
TAMB System
2 p=5,
ex: e=-2, 5 q=33 λ ex: e& =-2,
2 p=5,
5 q=33
The powers q/p of terminal sliding surface when e<0, and p/q of non-singular
terminal sliding surface when e& <0 may lead the terms to the complex values.
Th f
Therefore, th new form
the f i adopted
is d t d as the
th following
f ll i new form:
f
q p
1
sm (t ) = e&(t ) + λ e(t ) sign (e(t ))
p sn (t ) = e(t ) + q
e&(t ) sign (e&(t ))
λ
Both of the above forms are continuous and differentiable although the
absolute value and signum
g operators
p are involved. Moreover,, their derivatives
can be expressed as
q q
−1 1 p −1
p
s&m (t ) = e&&(t ) + λ e(t ) p e&(t ) s&n (t ) = e&(t ) + q
e&(t ) &e&(t )
p λ q
Electric Machinery and Control Laboratory,
62 EE, NCU, Taiwan
Precise Tracking Control of TAMB System—
Using
g Robust Non-Singular
g Terminal SMC
A. Experimental results of TAMB control system using TSMC system
From the experimental
p results,, the rotor p
position can not track the reference
commands precisely and immediately.
Moreover, the unstable control efforts which are cause by the singular problem,
result in the vibratory
y trackinggpperformance and terminal sliding g surface.
0.1mm 0.1mm
0.05mm Rotor Position 0.05mm Rotor Position
Reference Reference
Start Trajectory Start Trajectory Start Start
Reference Reference
Trajectory Trajectory
0.05mm -0.1mm Rotor Position 2Sec 0.05mm -0.1mm Rotor Position 2Sec 0.05mm -0.1mm 2Sec 0.05mm -0.1mm 2Sec
(a) (e) (a) (e)
Tracking Error Tracking Error Tracking Error Tracking Error
0V 0V 0V 0V
0A 0A 0A 0A
0.5A Total Current i0 + iZ 2Sec 0.5A Total Current i0 + iZ 2Sec 0.5A Total Current i0 + iZ 2Sec 0.5A Total Current i0 + iZ 2Sec
0A 0A 0A 0A
(d) (h) (d) (h)
0V 0V 0V 0V
0A 0A 0A 0A
0.5A Total Current i0 + iZ 2Sec 0.5A Total Current i0 + iZ 2Sec 0.5A Total Current i0 + iZ 2Sec 0.5A Total Current i0 + iZ 2Sec
0A 0A 0A 0A
(d) (h) (d) (h)
−1
⎡ q
q
−1 ⎤
U TSMC (t ) = Gn ⎢ &z&m (t ) − An z& (t ) − Bn z (t ) + λ e(t ) e(t ) + δsgn ( sm (t ))⎥
p
&
⎢⎣ p ⎥⎦
Nonsingular Terminal Sliding Surface Nonsingular Terminal Sliding Surface Nonsingular Terminal Sliding Surface Nonsingular Terminal Sliding Surface
0V 0V 0V 0V
Estimated Value of Lumped Uncertainty Estimated Value of Lumped Uncertainty Estimated Value of Lumped Uncertainty Estimated Value of Lumped Uncertainty
0V 0V 0V 0V
0A 0A 0A 0A
0.5A Total Current i0 + iZ 2Sec 0.5A Total Current i0 + iZ 2Sec 0.5A Total Current i0 + iZ 2Sec 0.5A Total Current i0 + iZ 2Sec
0A 0A 0A 0A
(e) (j) (e) (j)
u3 o3 = ∑ wi12 ( N )φi ( N ) + o2 ( N − 1)
Σ fD i =1
−1
Z −1
D control
Layer 1 Layer 2 Layer 3 (
o3 ( N ) = f D u3 ( N ) − u3 ( N − 1) )
Input Layer Hidden Layer Output Layer 2 2
= ∑ w ( N )φi ( N ) − ∑ wi13 ( N − 1)φi ( N − 1)
1
i3
i =1 i =1
Z −1 1 ⎧ 3 ⎡ ∂E ( N ) ⎤ 2 ⎫
Z −1 ⎪ ⎪
u2 o2 2⎨∑ ⎢ 2 ⎥
+ ε ⎬
Σ fI Σ vφ ⎪⎩ j =1 ⎢⎣ ∂w j ⎦⎥ ⎪⎭
eφ e&φ eφ e&φ
ηW
ε Varied Learning On-line Learning
1
Rates ηW 2 Algorithms
(6.16) & (6.17) (6.8)-(6.11)
oj w 2j
ΔWij1 , ΔW j2
eφ Σ fP
i +i
φd + Z −1 Z −1 vφ Drive System with b φ One-Axis
∑ e&φ Σ fI Σ DDM and Power
_
φ d dt Amplifiers
p ib − iφ
AMB φ
Σ fD
Z −1
PIDNN Controller
vb
maximum run outs: 0.2 mm maximum run outs: 0.238 mm maximum run outs: 0.282 mm
Left Rotor Orbit Right Rotor Orbit Left Rotor Orbit Right Rotor Orbit Left Rotor Orbit Right Rotor Orbit
y2 -Axis (mm)
y2 -Axis (mm)
y2 -Axis (mm)
y1 -Axis (mm)
y1 -Axis (mm)
y1 -Axis (mm)
0.1mm 0.1mm 0.1mm 0.1mm 0.1mm 0.1mm 0.1mm 0.1mm 0.1mm 0.1mm 0.1mm 0.1mm
x1 -Axis (mm) x2 -Axis (mm) x 1 -Axis (mm) x2 -Axis (mm) x 1 -Axis (mm) x2 -Axis (mm)
z (mm)
z (mm)
0 0 0
maximum run outs: 0.16 mm maximum run outs: 0.19 mm maximum run outs: 0.22 mm
Left Rotor Orbit Right Rotor Orbit Left Rotor Orbit Right Rotor Orbit Left Rotor Orbit Right Rotor Orbit
y2 -Axis (mm)
y 2 -Axis (mm)
y 2 -Axis (mm)
y1 -Axis (mm)
y1 -Axis (mm)
y1 -Axis (mm)
0.1mm 0.1mm 0.1mm 0.1mm 0.1mm 0.1mm 0.1mm
0.1mm 0.1mm 0.1mm 0.1mm 0.1mm
x1 -Axis (mm) x2 -Axis (mm) x 1 -Axis (mm) x2 -Axis (mm) x 1 -Axis (mm) x2 -Axis (mm)
z (mm)
0 0 0
y2 RMS Values 66 54
66.54 52 00
52.00 73 29
73.29 46 71
46.71 148 03
148.03 74 75
74.75
The GUUB stability is guaranteed if the control effort U(t) is using the control
l UISMC(t) off ISMC designed
law d i d as:
U ISMC (t ) = Bn
−1
[&x&d (t ) − An x(t ) + c1e&(t ) + c2e(t ) + δsat(s1 (t ))]
Therefore, the control law of ISMC ensures that any initial states outside the
boundary layer (s1(0)>Φ) will reach the boundary layer in finite time and
remain inside thereafter (s1(t) ≦Φ).
Five ISMC systems are used to construct a decentralized ISMC system for the
comparison of control performances.
In order to achieve the GUUB stability of the AMB subsystem, the control law
of the ideal IDISMC is designed as follows:
U IDISMC (t ) = Bn ⎡ &x&d (t ) − An x(t ) + c1e&(t ) + c2 e(t ) + c3 ∫ e(τ )dτ + δsat ( s2 (t ))⎤
−1 t
⎢⎣ 0 ⎥⎦
U ISMC (t ) = Bn
−1
[&x&d (t ) − An x(t ) + c1e&(t ) + c2e(t ) + δsat(s1 (t ))]
t t t
s2 (t ) = e&(t ) + c1e(t ) + c2 ∫ e(τ )dτ + c3 ∫ ∫ e(τ )dτ dτ
0 0 0 PID Controller
U IDISMC (t ) = Bn ⎡ &x&d (t ) − An x(t ) + c1e&(t ) + c2 e(t ) + c3 ∫ e(τ )dτ + δsat ( s2 (t ))⎤
−1 t
⎢⎣ 0 ⎥⎦
Therefore, the ideal IDISMC with integral control features results in improved
steady-state error performance compared with the ISMC.
However, it will cause saturated control effort easily and the control parameters
c1, c2, c3, δ, and Φ are very difficult to design overall.
Hence, the optimal output of the MPIDNN observer also named as the
observation target, is defined as follows:
t
y (t ) = c1e&(t ) + c2 e(t ) + c3 ∫ e(τ )dτ − L( x; t )
0
Electric Machinery and Control Laboratory,
79 EE, NCU, Taiwan
Robust Control of Fully Suspended Five-DOF AMB
System
y Using
g Decentralized IDISMC System
y
(b) MPIDNN observer
The well-designed
g MPIDNN observer results in ingenious
g mapping
pp g between
the network architecture and observation target.
According to the on-line tuning, not only the PID control gains KP, KI, and KD
are adaptive
p real time but also the lumped
p uncertainty
y L((x;;t)) is observed on-line.
yˆ (t )
U IDISMC (t ) = Bn ⎡ &x&d (t ) − An x(t ) + c1e&(t ) + c2 e(t ) + c3 ∫ e(τ )dτ + δsat ( s2 (t ))⎤
t
−1 yˆ (t ) = K P (t )o1 + K I (t )o2
⎢⎣ 0 ⎥⎦
+ K D (t )o3 + K L (t )o4
o1 ( N ) = f P (u1 ( N ) )
u1 o1
wij1 Σ fP (
o2 ( N ) = f I u2 ( N ) + o2 ( N − 1) ) 4
yˆ ( N ) = ∑ w 2j ( N )o j ( N )
w2j
e u2
Z −1
1 o3 ( N ) = f D (u3 ( N ) − u3 ( N − 1) ) j =1
o2
( )
Σ fI
o4 ( N ) = f L u4 ( N )
u3 o3 Σ ŷ
yˆ (e | Wˆ ) = Wˆ T O
⎡ 2
⎤
Σ fD e p ⎢− ∑ wi14 ( N )ei ( N )⎥
1 − exp
−1 ⎣ i =1 ⎦ y = y * (e | W ∗ ) + ε
e& Z −1
=
⎡ 2 ⎤
u4 o4 1 + exp ⎢− ∑ wi14 ( N )ei ( N )⎥ =W ∗ O +ε
T
Σ fL ⎣ i =1 ⎦
&
−1
U y (t ) = Bn yˆ (t ) using estimation law Wˆ = η s (t )O
W 2
ηW ηε
2 2 2 2
By using Barbalat’s
Barbalat s Lemma
Lemma, it can be shown that s2(t) will converge to zero as
t→∞, therefore, asymptotic stability is guaranteed.
Electric Machinery and Control Laboratory,
81 EE, NCU, Taiwan
Robust Control of Fully Suspended Five-DOF AMB
System
y Using
g Decentralized IDISMC System
y
The configuration of the one-axis AMB control subsystem using IDISMC
system with MISO MPIDNN RHNN observer is shown as follows:
xd + e d
Ur
∑
dt
Double Integral s2 Robust U IDISMC
Adaptive Law Controller ∑
x − ∫ Sliding Surface
(7.24) + vc vb
(7.8) (7.22) +U
∫∫∫∫
ε&ˆ
y
s2 Di S
Drive System
t
⎛ d ⎞2 with DDM
⎜ ⎟
⎝ dt ⎠ εˆ ξ and Power
Amplifiers
Learning Σ fP
Σ fD
(7.23) One-Axis AMB
d
Z −1
Subsystem
& (7.2)
dt Σ fH
Ŵ
IDISMC
maximum run outs: 0.16 mm maximum run outs: 0.18 mm maximum run outs: 0.25 mm
Left Rotor Orbit Right Rotor Orbit Left Rotor Orbit Right Rotor Orbit Left Rotor Orbit Right Rotor Orbit
y 2 -Axis (mm)
y 2 -Axis (mm)
y 2 -Axis (mm)
y1 -Axis (mm)
y1 -Axis (mm)
y1 -Axis (mm)
z (mm)
z (mm)
0 0 0
maximum run outs: 0.14 mm maximum run outs: 0.15 mm maximum run outs: 0.19 mm
Left Rotor Orbit Right Rotor Orbit Left Rotor Orbit Right Rotor Orbit Left Rotor Orbit Right Rotor Orbit
y 2 -Axis (mm)
y 2 -Axis (mm)
y 2 -Axis (mm)
y1 -Axis (mm)
y1 -Axis (mm)
y1 -Axis (mm)
0.1mm 0.1mm 0.1mm 0.1mm 0.1mm 0.1mm 0.1mm 0.1mm 0.1mm 0.1mm 0.1mm 0.1mm
x1 -Axis (mm) x2 -Axis (mm) x 1 -Axis (mm) x2 -Axis (mm) x1 -Axis (mm) x2 -Axis (mm)
z (mm)
z (mm)
0 0 0
y2 RMS Values 50 94
50.94 44 20
44.20 59 20
59.20 45 27
45.27 95 10
95.10 68 53
68.53
p
Φ Φ lim e(t ) = 0 ⎛ Φ⎞q lim e(t ) ≤ Φ lim e(t ) = 0
lim e(t ) ≤ lim e(t ) ≤ lim e(t ) ≤ ⎜ ⎟
Ultimate t →∞ λ t →∞ 2λ t →∞
t →∞ ⎝λ⎠
t →∞ t →∞
Bounds q
li e& (t ) ≤ 2Φ
lim
t →∞
lim e& (t ) ≤ Φ
t →∞
lim e& (t ) = 0
t →∞
li e& (t ) ≤ 2Φ
lim
t →∞
lim e&(t ) ≤ (2 λΦ ) p lim e&(t ) = 0
t →∞
t →∞
p
Φ Φ lim e(t ) = 0 ⎛ Φ⎞q lim e(t ) ≤ Φ lim e(t ) = 0
lim e(t ) ≤ lim e(t ) ≤ lim e(t ) ≤ ⎜ ⎟
Ultimate t →∞ λ t →∞ 2λ t →∞
t →∞ ⎝λ⎠
t →∞ t →∞
Bounds q
li e& (t ) ≤ 2Φ
lim
t →∞
lim e& (t ) ≤ Φ
t →∞
lim e& (t ) = 0
t →∞
li e& (t ) ≤ 2Φ
lim
t →∞
lim e&(t ) ≤ (2 λΦ ) p lim e&(t ) = 0
t →∞
t →∞
e e&
λ λ
(a) (b)
Fig. 8.1 Different ultimate bounds of SMC, CSMC, TSMC, and NTSMC systems due to Φ=0.001, p=5, q=3, and various values of λ.
(a) Ultimate bounds of tracking error. (b) Ultimate bounds of tracking error’s derivative.
Electric Machinery and Control Laboratory,
90 EE, NCU, Taiwan
Discussions
Comparisons of different control methods for TAMB control system
From the bar charts,, the ACSMC and RNTSMC systems
y result in better control
performances comparing with the others.
(μm) (μm)
20 20
PID RNN HPBRNN SMC CSMC PID RNN HPBRNN SMC CSMC
18 18
ACSMC TSMC NTSMC RNTSMC ACSMC TSMC NTSMC RNTSMC
16 16
14 14
12 12
10 10
8 8
6 6
4 4
2 The top four SMC approaches are RNTSMC, ACSMC, CSMC, and TSMC. This
2
0 0
result
Maximumis consistent
Average with theDeviation
Standard analysis of ultimate bounds from
Maximum AverageTableStandard
8.1.Deviation
periodic sinusoidal command at Case 1 periodic sinusoidal command at Case 2
(μm) (μm)
20 20
PID RNN HPBRNN SMC CSMC PID RNN HPBRNN SMC CSMC
18 18
ACSMC TSMC NTSMC RNTSMC ACSMC TSMC NTSMC RNTSMC
16 16
14 14
12 12
10 10
8 8
6 6
4 4
2 2
0 0
Maximum Average Standard Deviation Maximum Average Standard Deviation
i di trapezoidal
non-periodic t id l commandd att Case
C 1 non-periodic
i di trapezoidal
id l commandd at Case 2
PID PIDNN ISMC IDISMC PID PIDNN ISMC IDISMC PID PIDNN ISMC IDISMC
(μm) (μm) (μm)
160 160 160
140 140 140
120 120 120
100 100 100
80 80 80
60 60 60
40 40 40
20 20 20
0 0 0
x1x1 y1x2 x2y1 y2y2 zz x1x1 y1x2 x2y1 y2y2 zz x1x1 y1x2 x2y1 y2y2 zz
λ ∫e (t
ts = − e(t ) p
de = − ⎢ e(t ) p ⎥
r ) λ ⎢1 − q ⎥
⎣⎢ p ⎥⎦
e (tr )
⎡ ⎤ ⎡ ⎤
1⎢ 1− ⎥ 1 ⎢ −1 1− ⎥
q q q
1 p 1−
= − ⎢0 − e(t r ) ⎥ = − ⎢
p
e(t r ) ⎥ =
p
e(t r ) p
λ ⎢ 1− q ⎥ λ ⎢ p−q ⎥ λ ( p − q)
⎢⎣ p ⎥⎦ ⎢⎣ p ⎥⎦
Electric Machinery and Control Laboratory,
95 EE, NCU, Taiwan