Sliding Mode

Intelligent Sliding-Mode Control for
Five-DOF
Five DOF Active Magnetic Bearing
Control System
Prof. Faa-Jeng Lin
Electric Machinery and Control Laboratory

Department of Electrical Engineering
National Central University
Electric Machinery and Control Laboratory,

1 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
2 EE, NCU, Taiwan
Abstract
In this dissertation, a fully suspended five degree-of-freedom (DOF)
AMB control system,
y , which is composed
p of two radial AMBs
(RAMBs) and one thrust AMB (TAMB), is developed to fulfill the
requirements of the practical applications.
The motivations of this dissertation are:
Chapters 3-5
TAMB Control System
TAMB Positioning and Development of Tracking

System Tacking Controllers
Reference Trajectories and Performance

2×RAMB Measures Planning
Systems
Chapters 6-7
Five-DOF AMB Control System
Five-DOF AMB Regulating and Development of Decentralized

System Stabilizing Control Systems
Operating Conditions and Performance

Measures Planning
3 EE, NCU, Taiwan
Outline
Abstract
Introduction
y
TAMB System
Five-DOF AMB System
Sliding-Mode
Mode Control
Di
Discussions
i andd Conclusions
C l i
4 EE, NCU, Taiwan
Introduction
Magnetic suspension systems have attracted increasing attentions for
manyy applications
pp such as
Vibration isolation device
High-speed maglev train
Fast tool servo system
Fast-tool
Precision motion platform
Biomedical engineering
For several kinds of the magnetic suspension systems, magnetic

bearing (MB) is used to suspend and move the rotor to the
predefined positions functionally by the controlled electromagnetic
force without mechanical contact and friction.
friction

5 EE, NCU, Taiwan
Introduction
MB offers many features over conventional bearings such as
Advantages:
g Disadvantages:
g
Contact-free, frtionless High control complexity
Elimination of the lubrication High initial cost
Ideal for clean-room
clean room operation
Minimum Equipments for fully suspended AMB
Low losses and longer life Electromagnets 10
Higher rotational speed Position Sensors 5
Low vibration and noise Control Core (with 5 AD and 5 DA converters) 1
Tolerable against heat, cold, Power Amplifiers 5
vacuum Constant Current Sources 5
MB can be divided into the passive MB (PMB), i.e. permanent MB,

and the active MB (AMB), i.e., electromagnetic MB, mainly.
The term “passive” of PMB covers MBs that do not require any external current
feed to the bearing.
The term “active”
active of AMB implies that bearing forces are actively controlled by
means of electromagnets through a well-designed closed control loop.
6 EE, NCU, Taiwan
Introduction
What is AMB?
The elements of AMB: Electromagnet
Power Amplifier
Electromagnet
g
Rotor
Sensor
Controller Controller
Power Amplifier Rotor
Sensor
According to its structure and usage of AMB:

Radial AMB (RAMB) Thrust AMB (TAMB) Combination Disk
, i.e. Axial AMB
RAMB TAMB
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
Magnetic Bearing Stage

T b l l
Turbomolecular
Inlet
Guide Pumps
Vanes
Compressor
Motor and
Bearingless
Bearing Controls
Heart
Motor Hard Drive
PMB
Pump
High-Speed
g p Spindle
p

Vacuum Vessel Blood Pump Maglev Trains 8 EE, NCU, Taiwan
Introduction
In the AMB system:
A fixed current is supplied
pp to each electromagnet
g and can be referred as the
bias current ib. (ib=>x0)
A variable current is superimposed on the bias current, which is obtained by the
designed
g controller and can be referred as the control current ic. (ic=> △x)
The purpose of the bias setting is to improve the linearity of the
force-current relationships around the operating point (ib, x0).
The operating modes of the drive system which supply the currents
to the electromagnets can be classified into three classes:
Class-A
Cl A driving
d i i moded (differential
(diff i l driving
d i i mode,
d DDM);
DDM)
Class-B driving mode;
ib－ ic ib－ ic － ic
Class-C driving mode.
ib+ ic ib ic
Class-A Class-B Class-C

9 EE, NCU, Taiwan
Introduction
Comparisons of various developed AMB systems:

10 EE, NCU, Taiwan
Outline
Abstract
Introduction
y
TAMB System
Five-DOF AMB System
Sliding-Mode
Mode Control
Di
Discussions
i andd Conclusions
C l i
11 EE, NCU, Taiwan
System Dissection, Dynamic Analyses, and
Experimental
p Designs
g of Five-DOF AMB System
y
The fully suspended five-DOF AMB system is composed of four
radial DOF controlled by
y two identical RAMBs and one axial DOF
controlled by a TAMB.
y1 Position Sensors y2
Load Disk x1 x2
(C
(Case 2 Only)
O l )
Y
X
Rotor Z
z
Power Line
Platform
Left RAMB Induction Motor TAMB Right RAMB
Position
Signals Line Mass Disk Thrust Disk
PC-Based Control Core Inverter Drive System Five-DOFand

Electric Machinery AMB SystemLaboratory,
Control
12 EE, NCU, Taiwan
System Dissection, Dynamic Analyses, and
Experimental
p Designs
g of Five-DOF AMB System
y
In the case of an AMB, where iron is used in the stator and the rotor,
electromagnets
g cause a flux Φ to circulate a magnetic
g loop.
p
The relation of flux Φ and cross-sections As and A0 can be stated as
Φ = Bs As = B0 A0
Stator
One can obtain the following result
Φ Ni μ Ni
B0 = 0 = ≅ 0
A ⎛ ls 0 lr 0 2l0 ⎞ 2l0
⎜⎜ + + ⎟⎟ N As
⎝ μ s μr μ0 ⎠ i
Consider the Maxwell force l0

quadratic!
B02 A0 μ 0 A0 N 2i 2 i2 A0
f = = =k 2
μ0 4l02 l0 Ar
f Rotor f l s + l r + 2l0
μ0 A0 N 2 inversely quadratic! 2 2
where k = 4 Electric Machinery and Control Laboratory,
13 EE, NCU, Taiwan
Outline
Abstract
Introduction
y
TAMB System
Five-DOF AMB System
Sliding-Mode
Mode Control
Di
Discussions
i andd Conclusions
C l i
14 EE, NCU, Taiwan
TAMB System
A typically structure of the TAMB system is shown in Fig. 2.5. In
the TAMB system,
y , a thrust disk is embedded on the rotor and used
to carry out the rotor position control on the defined axial direction
practically.
How does it work?
Thrust Disk
Z
Axial Direction
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.
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
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
k aii ≡ z = 4k a 02 Nominal Air Gap z0 : 00.5mm

5mm
Sensor
∂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.
Control14 Bits Resolution i0 + iz

Signal
g v0 + v z
Coil
v0 − v z
Stator
DA Card
Power Position
Amplifier R
Rotor Sensor
Visual BASIC 6.0 i0 − iz Low-Pass Filter
Thrust
Disk
Rotor Position z Sensor Signal

AD Card TAMB System
16 Bits Resolution
Control subroutine with 5 kHz sampling rate
rate.
One controlled DOF requires 2 DACs. Electric Machinery and Control Laboratory,
18 EE, NCU, Taiwan
TAMB System
B. Reference Trajectories Planning
The designed periodic sinusoidal command can examine the smoothness of
tracking responses. Moreover, the designed non-periodic trapezoidal command
can verify the effectiveness of the instant and stable tracking responses.
Two test conditions are provided which are the nominal Case (Case 1) and the
parameter variation Case (Case 2). In Case 2, two stainless steel disks with
respective 0.38 kg weights are added to the left and right side of the rotor.
0 .1 m m
R e fe re n c e Thrust Disk
T ra je c to r y Z
S ta rt Axial Direction
0 .0 5 m m － 0 .1 m m 2 Sec
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
C. Performance Measures and Comparisons

The maximum tracking error TM, the average tracking error TA and the standard
deviation of the tracking error TS for the trajectory tracking
n
are adopted:
∑ (T ( N ))
(Te ( N ) − TA ) 2
e n
TM = max(Te ( N )), where Te ( N ) = z m ( N ) − z ( N ) TA = TS = ∑
N =1
n n N =1 n
19 EE, NCU, Taiwan
Outline
Abstract
Introduction
y
TAMB System
Five-DOF AMB System
Sliding-Mode
Mode Control
Di
Discussions
i andd Conclusions
C l i
20 EE, NCU, Taiwan
Five-DOF AMB System
Two pairs of the U-shaped drive coils are installed perpendicularly
on the stator and used to pproduce the pperpendicular
p attractive
electromagnetic forces.y1
110 mm
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

21 EE, NCU, Taiwan
Five-DOF AMB System
Drive System for Left RAMB
A drive system
y includingg DDM and ppower amplifiers
p of left RAMB is shown
in Fig. 2.8 in which the dotted line or solid line of the rotor means that the rotor
is centered in the aperture or deviated from the center in X-Y axes, respectively.
The total nonlinear attractive electromagnetic
g forces for X-axis and Y-axis are:
vb + v y1
DDM
+ ib + i y1
Fx1 ≡ f1 − f 3 = k r ⎢
(
⎡ ib + ix 2 )
−
ib (
− i x
2
) ⎤
2⎥
Σ Power Amplifier 1 1
+ 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
Proposed ix , iy : control currents in X-Y axes

1 1
Control Laws
xb, yb: nominal air gaps in X-Y axes
+ vx Lower
+ Σ 1 ib − i y1 Electromagnet x1, y1: rotor positions in XX-YY axes
vb + v x1
22 EE, NCU, Taiwan
Five-DOF AMB System
Dynamic Model for Five-DOF AMB System
Assumingg the rotor is a rigid
g and symmetric
y body,
y the relationshipp between the
center of gravity (CG) of the rotor and the five-DOF AMB is shown in Fig. 2.9.
The linearized electromagnetic force models and the relationships between the
rotor ppositions and the CG coordinates are approximated
pp as:
⎧m&x&c = Fx + Fx + f dx Fy2 = f 6 − f 8
無法顯示圖像。您的電腦可能沒有足夠的記憶體來開啟圖像，或圖像可能已毀損。請重新啟動您的電腦，並再次開啟檔案。如果仍然出現紅色 x，您可能必須刪除圖像，然後再次插入圖像。
⎪ 1 2
⎪m&y&c = Fy1 + Fy2 − mg + f dy

⎪ ≅ k rp y2 + k ri i y2
⎨m&z&c = Fz + f dz
⎪ && & Fx2 = f 5 − f 7
⎪ Jθ y − J zωθ x = − aFx1 + bFx2 + cf dx
⎪ Jθ&& + J ωθ& = aF − bF − cf
⎩ x z y y1 y2 dy ≅ k rp x2 + k ri ix2
Fy1 = f 2 − f 4 the rotor is rigid and the
displacements are small
≅ k rp y1 + k ri i y1
bx1 + ax2 bx1 + ax2
xc = =
Fx1 = f1 − f 3 a+b l
by1 + ay2 by1 + ay2
≅ k rp x1 + k ri ix1 yc =
a+b
=
l
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 ⎥
The dynamic model of the five-DOF AMB system ⎥ = ⎢be

F = ⎢ Fy1can f 7 ⎥ ≅ ⎢ k rp x2 using
f 5 −rearranged + k rii x2 ⎥the
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
matrix form as follows: ⎢ Fy2 ⎥ ⎢ f 6 − f 8 ⎥ ⎢k rp y 2 + k rii y2 ⎥
⎢⎣ Fz ⎥⎦ ⎢⎣ f 9 − f10 ⎥⎦ ⎢⎣ k ap z + k aii z ⎥⎦
M X&& + G X& = K F + ED + Cg
C
where X = [x1 x2 y1 y2 z ]T is the state vector; F = [Fx Fx Fy Fy Fz ] is
T
the electromagnetic force vector; D = [ f dx f dyy f dz ]T is the external disturbance

1 2 1 2
vector; C = [0 0 − 1 − 1 0]T is the gravity vector; M , G , K ,and E are the

mass, gyroscope, electromagnetic force, and external disturbance matrixes,
respectively, and defined as follows:
⎡1 0 0 0⎤0 ⎡ 0 0 α1 − α1 0⎤ ⎡ β1 β 2 0 0⎤
0 ⎡γ 1 0 0⎤
⎢0 0⎥⎥ ⎢γ
⎢ 1 0 0 ⎢ 0 0 − α 2 α 2 0⎥⎥ ⎢β β 0 0 0 ⎥⎥ ⎢ 2 0 0 ⎥⎥
⎢ ⎢ 2 3
M = ⎢0 0 1 0⎥ G = ⎢ − α 1
0 α1 0 0 0⎥ K =⎢0 0 β1 β 2 0 ⎥ E = ⎢ 0 γ 1 0⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢0 0 0 0⎥1
⎢ α2 − α2 0 0 0⎥ ⎢0 0 β 2 β3 0 ⎥ ⎢ 0 γ2 0⎥
⎣⎢0 0 0 0 1⎥⎦ ⎢⎣ 0 0 0 0 0⎥⎦ ⎢⎣ 0 0 0 0 β 4 ⎥⎦ ⎢⎣ 0 0 γ 3 ⎥⎦
where
h aJ zω , bJ zω , 1 a2 , β 2 = 1 − ab , 1 b2 , 1
α1 = α2 = β1 = + β3 = + β4 = ,
Jl Jl m J m J m J m
1 ac 1 bc
γ1 = − , γ 2 = + , γ 3 = 1 .
m J m J m
Observing the matrixes G and K , the coupling effects in the five-DOF AMB
system are serious apparently. Electric Machinery and Control Laboratory,
25 EE, NCU, Taiwan
Five-DOF AMB System
To decouple the original coupled dynamic model as five independent
subsystems for the purpose of decentralized control, the dynamic model can be
further rewritten as
M X&& = AX + BU + M H
where U = [ix ix i y i y iz ]T is the control current vector; H = [hx hx hy hy hz ]T
1 2 1 2 1 2 1 2
is the decoupled term vector; A and B are the stiffness,

stiffness and control gain
matrixes, respectively, and defined as follows:
⎡k rp β1 0 0 0 0 ⎤ ⎡k ri β1 0 0 0 0 ⎤
⎢ 0 krp β 3 0 0 0 ⎥⎥ ⎢ 0 k ri β 3 0 0 0 ⎥⎥
⎢ ⎢
A=⎢ 0 0 krp β1 0 0 ⎥ B=⎢ 0 0 k ri β1 0 0 ⎥
⎢ ⎥ ⎢ ⎥
⎢ 0 0 0 krp β 3 0 ⎥ ⎢ 0 0 0 k ri β 3 0 ⎥
⎢ 0 0 0 0 kap β 4 ⎥⎦ ⎢⎣ 0 0 0 0 k ai β 4 ⎥⎦
⎣
⎡ hx1 ⎤ ⎡ − α1 y&1 + α1 y& 2 + 2k rp β 2 x2 + 2k ri β 2i x2 + γ 1 f dx ⎤
⎢ h ⎥ ⎢ α y& − α y& + 2k β x + 2k β i + γ f ⎥
⎢ ⎥ ⎢
x 2 2 1 2 2 rp 2 1 ri 2 x1 2 dx ⎥
H = ⎢ hy1 ⎥ = ⎢ α1 x&1 − α1 x& 2 + 2k rp β 2 y 2 + 2k ri β 2i y2 + γ 1 f dy − g ⎥
⎢ ⎥ ⎢ ⎥
h
⎢ y2 ⎥ ⎢ 2 1− α x
& + α x
2 2
& + 2 k β
rp 2 1 y + 2 k β i
ri 2 y1 + γ f
2 dy − g ⎥
⎢⎣ hz ⎦⎥ ⎣⎢ γ 3 f dz ⎥⎦
Hence, the states of the dynamic model are completely decoupled. Therefore
the developed decentralized model-based control systems can be easily applied
26 EE, NCU, Taiwan
Five-DOF AMB System
Experimental Design for Five-DOF AMB Control System
R
A. Experimental
p Setupp
R
+ vb − v x1
Σ
− v x1 +
Inverter ADC 1 Proposed DAC 1 v x1 vb

Control Laws
Sensor 1 + vb + vx
+ Σ
1
Left RAMB Inductor Right RAMB

.
.
.
. R
.
Position .
Rotor R
Sensors + v0 − v z
Σ
− v0 +
Total Total ADC 5 Proposed DAC 5 v0

v0
Currents Currents Sensor 5
Control Laws
+
Induction + Σ v0 + vz
Motor PC-Based Digital Control System
TAMB Analogy Circuit
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:
Source: S2M Electric Machinery and Control Laboratory,

28 EE, NCU, Taiwan
Five-DOF AMB System
B. Operating Conditions Planning
Rotating Speed (RPM) Load (kg)
Case 1: General Operating Condition 2400 none
Case 2: Parameter Variation Condition 2400 0.38
C
Case 3 High
3: Hi h Rotating
R i Speed
S d Condition
C di i 4800 none
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.
29 EE, NCU, Taiwan
Outline
Abstract
Introduction
y
TAMB System
Five-DOF AMB System
Sliding-Mode
Mode Control
Di
Discussions
i andd Conclusions
C l i
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
31 EE, NCU, Taiwan
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&
32 EE, NCU, Taiwan
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
The varied learning rates, ∑

Output Layer Feedback 1
Layer 3
which guarantee the W 2 ( N + 1) = W 2 ( N ) + Δ WW
2 2 Z −1
(Output Layer)
1 yr
convergence of the tracking W j
W j1 ( N + 1) = W j1 ( N ) + ΔW j1 S z −1
error based on the analyses Γj β

off a discrete-type
di L
Lyapunov z −1
z −1.. z
− 1 ... z -1
Layer 2
(Hidden Layer)
function, are derived as α α α α
follows: 1 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&
33 EE, NCU, Taiwan
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
0.9 Fuzzy point of view: Fuzzy point of view:

8 distinct rules with the same fuzzy 31 8 distinct rules with different fuzzy
0.8
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 0.88 Γ 0.88 0.88 0.31 0.62 Γ 0.47 0.48β

0.6
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
34 EE, NCU, Taiwan
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

35 EE, NCU, Taiwan
Using
g Model-Free Control Methods
A. Experimental results of TAMB control system using PID
controller
From the experimental results as shown in Figs. 3.5(a), (d) and 3.6(a), (d), poor
tracking responses are obtained under both conditions at Case 1 and Case 2
owingg to the fixed-gains
g PID controller is unable of dealingg with the
uncertainties in practical TAMB control system.
0.1mm 0.1mm 0.05mm 0.05mm

Reference Reference Rotor Position Rotor Position
Trajectory Trajectory
Start Start Start Start
Reference Reference
Rotor Position Rotor Position Trajectory Trajectory
0.05mm －0.1mm 0.05mm －0.1mm 2Sec 0.05mm －0.1mm 2Sec 0.05mm －0.1mm 2Sec
2Sec
(a) (d) (a) (d)
Tracking Error Tracking Error Tracking Error Tracking Error
0.01mm 0.01mm 2Sec 0.01mm 2Sec 0.01mm 2Sec

2Sec
(b) (e) (b) (e)
0 5A
0.5A T t l Current
Total C t i0 – iZ 2S
2Sec 0 5A
0.5A Total Current i0 – iZ 2Sec 0 5A
0.5A Total Current i0 – iZ 2Sec 0 5A
0.5A Total Current i0 – iZ 2Sec
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)
Case 1 Case2 Case 1 Case 2

36 EE, NCU, Taiwan
Using
B. Experimental results of TAMB control system using RNN
controller
Though the tracking responses have been improved, the tracking errors are
unavoidable due to the limitation of the approximated ability of RNN.
The possible solutions may increase the number of hidden neurons or change
the learning rates. However, the increased neurons will result in heavy
computational burden and inappropriate rates can cause the system unstable.
0.1mm 0.1mm 0.05mm 0.05mm
Reference Reference Rotor Position Rotor Position
Reference Reference
Rotor Position Trajectory Trajectory
0.05mm －0.1mm 2Sec 0.05mm －0.1mm Rotor Position 2Sec 0.05mm －0.1mm 2Sec 0.05mm －0.1mm 2Sec
(a) (d) (a) (d)
0.01mm 2Sec 0.01mm 2Sec 0.01mm 2Sec 0.01mm 2Sec

(b) (e) (b) (e)
0 5A
0.5A T t l Current
2Sec 0 5A
0.5A T t l Current
2Sec 0 5A
0.5A T t l Current
Total C t i0 – iZ 2Sec 0 5A
0.5A T t l Current
2Sec
0A 0A 0A 0A
0A 0A 0A 0A
( )
(c) (f) (c) (f)

37 EE, NCU, Taiwan
Using
C. Experimental results of TAMB control system using HPBRNN
controller
From the experimental results, the tracking performances are much improved,
and the required total currents are also successfully reduced.
Consequently the robustness of the TAMB control system using HPBRNN
Consequently,
under the occurrence of uncertainties for both test conditions can be clearly
observed.
0.1mm 0.1mm 0.05mm 0.05mm
Reference Reference Rotor Position
Rotor Position
Reference Reference
Rotor Position Rotor Position Trajectory Trajectory
0.05mm －0.1mm 2Sec 0.05mm －0.1mm 2Sec 0.05mm －0.1mm 0.05mm －0.1mm 2Sec
2Sec
((a)) (d) (a) ((d))
0.01mm 2Sec 0.01mm 2Sec 0.01mm 0.01mm 2Sec

2Sec
(b) (e) ( )
(b) (e)
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
0A 0A 0A 0A
(c) (f) (c) (f)

38 EE, NCU, Taiwan
Using
The performance measures of the PID, RNN and HPBRNN
controllers are shown in Table 3.1 and Table 3.2,, respectively.
p y
Obviously, the proposed HPBRNN possesses the best robust control
performance and can control the TAMB system effectively.
Table 3.1 Performance measures of TAMB control system using model-free control methods at Case 1.
PID RNN HPBRNN
Tracking Errors, μm
Sinusoid Trapezoid
p Sinusoid Trapezoid
p
Maximum 17.277 9.264 14.787 15.614 7.633 8.489
Average 4.625 2.428 2.074 2.153 1.874 1.678

Standard Deviation 3.138 2.134 1.804 1.909 1.323 1.263
Table 3.2 Performance measures of TAMB control system using model-free control methods at Case 2.
PID RNN HPBRNN
T ki Errors,
Tracking E μm
Sinusoid Trapezoid Sinusoid Trapezoid Sinusoid Trapezoid
Maximum 18.655 10.042 17.816 17.134 8.077 9.975
Average 4.962 2.808 2.401 2.451 2.043 1.860

Standard Deviation 3.606 2.530 1.979
Electric1.898
Machinery1.402 1.332
and Control Laboratory,
39 EE, NCU, Taiwan
Outline
Abstract
Introduction
y
TAMB System
Five-DOF AMB System
Sliding-Mode
Mode Control
Di
Discussions
i andd Conclusions
C l i
40 EE, NCU, Taiwan
Using
g Adaptive
p Complementary
p y SMC
To obtain a standard second order state equation for the controller
design,
g , the dynamic
y model of the TAMB system
y shown in ((2.19)) is
rewritten as follows:
&z&(t ) = A(z; t )z& (t ) + B( z; t ) z (t ) + G( z; t )U (t ) + d ( z; t ) m&z& = −cz& + k ap z + k ai i z + f dz
The above dynamic model is rewritten considering the separation of

nominal system parameters and system parameter variations as
&z&(t ) = [An + ΔA(z; t )]z& (t ) + [Bn + ΔB ( z; t )]z (t ) + [Gn + ΔG ( z; t )]U (t ) + d ( z; t )
= An z& (t ) + Bn z (t ) + GnU (t ) + L( z; t )
where L( z; t ) is called the lumped uncertainty and defined as

L( z; t ) = ΔA(z; t )z&(t ) + ΔB(z; t )z (t ) + ΔG(z; t )U (t ) + d ( z; t )
Here, the bound of the lumped uncertainty is necessary to be known

in advance and satisfies the inequality as follows:
δ = L( z; t ) + ξ
41 EE, NCU, Taiwan
Using
g SMC
A. Sliding-Mode Control (SMC)
SMC has ggained a lot of attentions in recent yyears for its insensitivityy to system
y
parameter variations and external disturbance once the system trajectory
reaches and stays on the sliding surface.
First,, a conventional slidingg surface is defined as follows:
⎛d ⎞
s (t ) = ⎜ + λ ⎟e(t )
⎝ dt ⎠
Diff
Differentiating
ti ti S(t) with ith respectt to
t time:
ti
s&(t ) = e&&(t ) + λe&(t )
= &z&m (t ) − An z& (t ) − Bn z (t ) − GnU (t ) − L( z; t ) + λe&(t )
The globally uniformly ultimate boundedness (GUUB) stability is guaranteed if
the control effort uses the SMC control law USMC designed as:
−1 ⎡ ⎛ s (t ) ⎞⎤
U SMC (t ) = Gn ⎢&z&m (t ) − An z& (t ) − Bn z (t ) + λe&(t ) + δ sat ⎜ ⎟⎥
⎣ ⎝ Φ ⎠⎦
where sat(‧) and Φ are the saturation function and the boundary layer
thickness respectively.
thickness, respectively

42 EE, NCU, Taiwan
Usingg SMC
The saturation function is adopted in SMC to e& s(e, e&)
reduce the chattering as follows:
⎛ s (t ) ⎞ ⎧⎪sgn(( s (t )),
)) s (t ) > Φ
sat ⎜ =
⎟ ⎨
⎝ Φ ⎠ ⎪⎩ s / Φ , s (t ) ≤ Φ e
Therefore,, the SMC control law ensures that any y
initial states outside the boundary layer (s(0)>Φ)
will reach the boundary layer in finite time and Φ
remain inside thereafter (s(t)≦Φ).
Hitting control gain δ
Good insensitivity to uncertainty, but cause system unstable easily.
St bl control
Stable t l effort,
ff t but
b t weakk robustness.
b t
Boundary layer thickness Φ

Reduce chattering, but increase steady-state errors.
Precise tracking response, but high chattering control effort.
Therefore, a trade-off problem between chattering and control accuracy arises.

43 EE, NCU, Taiwan
Using
g CSMC
B. Complementary Sliding-Mode Control (CSMC)
The CSMC can efficiently y reduce the g guaranteed ultimate bounds by y half
while using saturation function compared with SMC.
Moreover, faster transient response of the tracking error during the reaching
phase also can be obtained.
p
τ
The control object of the CSMC is to let e& , e and ∫ edτ move toward the
0
neighborhood of the intersection of the integral and complementary sliding
surfaces.
Define the integral sliding surface sg and complementary sliding surface sc as:
2
⎛d ⎞ t
s g (t ) = ⎜ + λ ⎟ Λ (t ) where Λ (t ) = ∫ e(τ )dτ
⎝ dt ⎠ 0
⎛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
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
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
For ease of discussion, the outputs of the Σ Σ Layer 3

(Output Layer)
RHNN are rewritten itt as follows:
f ll α∗ β∗
y1 = Pˆ (e | αˆ ) = αˆ T Γ Γj
w jo
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
αˆ = [ w11 w21 L wn1 ]T , βˆ = [ w12 w22 L wn 2 ] .

1
T
xi
By universal
B i l approximation
i ti theorem,
th Layer 1
∗T
P = P (e | α ) + ε p = α Γ + ε p
∗ ∗ (Input Layer)
T e1 e2
Q = Q ∗ (e | β ∗ ) + ε q = β ∗ Γ + ε q
optimal weighting vectors minimum approximated errors
46 EE, NCU, Taiwan
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
Q(t ) = Gn Ξ(z; t ) , Ξ(z; t ) = 2e&(t ) + λe(t ) − sc (t ) + Gn sc (t ) .

−1
The MIMO RHNN estimator is proposed to estimate the two complicated

dynamic functions P(t) and Q(t) directly including the lumped uncertainty to
further improve the control performance.

47 EE, NCU, Taiwan
Using
g ACSMC
Theorem 4.2: Considering the TAMB system represented by (4.2), if the
proposed ACSMC control law UACSMC, which is composed of the MIMO
RHNN estimator ti t designed
d i d as Ue with ith estimation
ti ti lawsl f the
for th two
t d
dynamic
i
functions P(t) and Q(t), the robust compensator designed as Uc with adaptive
laws, is adopted as the control effort U, then the asymptotic stability of the
proposed ACSMC system can be guaranteed. guaranteed
U ACSMC (t ) = U e (t ) + U c (t )
αˆ& = ηα σ (t ) Γ
U e (t ) = Pˆ (e | αˆ ) + λQˆ (e | βˆ ) using g estimation laws &
βˆ = η β λσ (t ) Γ
&ˆ = η σ (t )
[ ]
U c (t ) = εˆ p + λ εˆq + s g (t ) using adaptive laws
ε p p
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 →∞
48 EE, NCU, Taiwan
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

49 EE, NCU, Taiwan
Using
g Adaptive
p Complementary
p y SMC
A. Experimental results of TAMB control system using SMC system
Poor trackingg responses
p are obtained due to the saturation function is used.
Though the tracking response can be improved with larger δ or smaller Φ, the
chattering will be more serious, and the control effort will be saturated easily.
0.1mm 0.1mm
0.05mm Rotor Position 0.05mm Rotor Position
Reference Reference
Start Trajectory Start Trajectory Start Start
Reference Reference
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)
0.01mm 2Sec 0.01mm 2Sec 0.01mm 0.01mm 2Sec

2Sec
(b) (f) (b) (f)
Sliding Surface Sliding Surface Sliding Surface Sliding Surface
0V 0V 0V 0V
500mV 2Sec 500mV 2Sec 500mV 2Sec 500mV 2Sec

(c) (g) (c) (g)
0A 0A 0A 0A
0A 0A 0A 0A
(d) (h) (d) (h)
Case 1 Case2 Case 1Machinery and Control

Electric Case 2
Laboratory,
50 EE, NCU, Taiwan
Using
g Adaptive
p Complementary
p y SMC
B.Experimental results of TAMB control system using CSMC system
Both the integral
g and complementary
p y sliding
g surfaces are approach
pp zero nearly.
y
The tracking responses at Case 2 are unacceptable due to the CSMC with small
δ and large Φ is not robust enough to deal with the uncertainty in terms of the
additional load disk.
0.1mm 0.1mm Reference Reference
Reference Reference Trajectory 0.05mm Trajectory 0.05mm
Rotor Position Rotor Position
(a) (f) (a) (f)

(b) (g) (b) (g)
Integral Sliding Surface Integral Sliding Surface Integral Sliding Surface Integral Sliding Surface
0V 0V 0V 0V

(c) (h) (c) (h)
Complementary Sliding Surface Complementary Sliding Surface Complementary Sliding Surface Complementary Sliding Surface
0V 0V 0V 0V

(d) (i) (d) (i)
0A 0A 0A 0A
0A 0A 0A 0A
(e) (j) (e) (j)

Electric Case 2
Laboratory,
51 EE, NCU, Taiwan
Using
g Adaptive
p Complementary
p y SMC
C.Experimental results of TAMB control system using ACSMC system
From the experimental
p results,, the tracking
gpperformances are improved
p greatly
g y
due to the on-line tuning of the control parameters.
Since the control efforts are obtained by the MIMO RHNN majorly and
q y, the required
exquisitely, q total currents are also reduced effectively.
y
0.1mm 0.1mm Reference Reference
Reference Reference
Trajectory 0.05mm Trajectory 0.05mm
Rotor Position Rotor Position
Sec
(a)
( ) (f) (a) (f)
0.01mm 2Sec 0.01mm 2Sec 0.01mm 0.01mm

2Sec 2Sec
(b) (g) (b) (g)
Integral
g Sliding
g Surface Integral
g Sliding
g Surface Integral Sliding Surface Integral Sliding Surface
0V 0V 0V 0V
500mV 2Sec 500mV 2Sec 500mV 500mV

2Sec 2Sec
(c) (h) (c) (h)
Complementary Sliding Surface Complementary Sliding Surface Complementary Sliding Surface Complementary Sliding Surface
0V 0V 0V 0V
500mV 2Sec 500mV 2Sec 500mV 500mV

2Sec 2Sec
(d) (i) (d) (i)
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
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)

Electric Case 2
Laboratory,
52 EE, NCU, Taiwan
Using
g Adaptive
p Complementary
p y SMC
The performance measures of the SMC, CSMC and ACSMC
systems
y are shown in Table 4.1 and Table 4.2,, respectively.
p y
Obviously, the proposed ACSMC possesses the best robust control
Table 4.1 Performance measures of TAMB control system using SMC, CSMC, and ACSMC systems at Case 1.
SMC CSMC ACSMC
Sinusoid Trapezoid
p
Maximum 10.081 8.115 3.957 4.174 3.324 3.563
Average 4.388 2.819 1.634 1.556 1.054 1.110

Table 4.2 Performance measures of TAMB control system using SMC, CSMC, and ACSMC systems at Case 2.
SMC CSMC ACSMC
T ki Errors,
Tracking E μm
Maximum 11.161 9.259 4.421 4.759 3.971 3.655
Average 4.812 3.095 1.702 1.786 1.204 1.319

Electric1.395
53 EE, NCU, Taiwan
Outline
Abstract
Introduction
y
TAMB System
Five-DOF AMB System
Sliding-Mode
Mode Control
Di
Discussions
i andd Conclusions
C l i
54 EE, NCU, Taiwan
Using
g TSMC
A. Terminal Sliding-Mode Control (TSMC)
First,, define the terminal sliding
g surface sm to achieve ffinite time tracking
g
control (FTTC) as follows
q
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
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
A finite time ts, which is taken to travel from e(t

( r)≠ ( r+ts))=0 , can be
)≠0 to e(t
obtained
e&
q q
1 0 − p 1−
λ ∫e (t
ts = − e p
de = e(t r ) p sm (0)
r ) λ ( p − q)
Therefore, both the tracking error e

and its derivative e& will converge to
zero in finite time by using the e
sm (t s )
y
TSMC system as shown in Fig.
g 5.1.
sm (t r )
sm = 0
56 EE, NCU, Taiwan
Using
g TSMC
q
−1
Now, considering the TSMC control law, the forth term containing e(t ) e&(t ) p
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
57 EE, NCU, Taiwan
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.
Moreover,, when sn=0,, the system

y dynamic
y e&(t ) = −λe(t ) q / p is equivalent
q to the
one of TSMC. Therefore, the finite time ts taken to reach the equilibrium point
e=0 of the NTSMC system is the same as the one of the TSMC system.

58 EE, NCU, Taiwan
Using
g RNTSMC
C. MISO RHNN Estimator
In the pproposed
p RNTSMC,, the MISO RHNN with two inputs
p and one outputp is
proposed as the uncertainty estimator.
For ease of discussion, the output of the RHNN are rewritten as follows:
yˆ (e | Wˆ ) = Wˆ T Γ
where Γ = [ Γ1 Γ 2 L Γ n ]T Wˆ = [ w1 w2 L wn ]T
y y
By universal approximation theorem, 1 2
T
L = y * (e | W ∗ ) + ε = W ∗ Γ + ε Σ Σ Layer 3
(Output Layer)
optimal weighting vectors

w jo
minimum approximated
pp errors Γj
.. ...
The MISO RHNN uncertainty estimator z Layer 2
−1
z −1
z −1
z −1

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)
lumped uncertainty is unnecessary. e1 e2

59 EE, NCU, Taiwan
Using
g RNTSMC
D. Robust Non-Singular Terminal Sliding-Mode Control (RNTSMC)
Comparing
p g with the TSMC and NTSMC systems, y , the RNTSMC system y
possesses no chattering and maintains asymptotic stability due to the switching
function and the saturation function are not used, respectively.
Theorem 5.1: Considering g the TAMB systemy represented
p by
y ((4.2),
), if the
proposed RNTSMC control law URNTSMC(t), which is composed of the
equivalent control law Ueq(t), the MISO RHNN uncertainty estimator URHNN(t)
with estimation law and the robust controller Ur(t) with adaptive law, is
adopted as the control effort , then the asymptotic stability of the proposed
RNTSMC system can be guaranteed.
U RNTSMC (t ) = U eq (t ) + U RHNN (t ) + U r (t )
⎡
−1 q 2− ⎤
p
U eq (t ) = Gn ⎢ &z&m (t ) − An z&(t ) − Bn z (t ) + λ e&(t ) ⎥ q
⎢⎣ 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
60 EE, NCU, Taiwan
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

61 EE, NCU, Taiwan
Using
g RNTSMC
The solution of the actual complex problem in the programming of
the implementation.
p
The terminal sliding surface sm and the non-singular terminal sliding surface sn:
q p
1
sm (t ) = e&(t ) + λe(t ) p Complex value sn (t ) = e(t ) +
e&(t ) Complex value q
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
62 EE, NCU, Taiwan
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
Reference Reference
Reference Reference
(a) (e) (a) (e)

(b) (f) (b) (f)
Terminal Sliding Surface Terminal Sliding Surface Terminal Sliding Surface Terminal Sliding Surface
0V 0V 0V 0V

(c) (g) (c) (g)
0A 0A 0A 0A
0A 0A 0A 0A
(d) (h) (d) (h)

Electric Case 2
Laboratory,
63 EE, NCU, Taiwan
Using
g Terminal SMC
B.Experimental results of TAMB control system using NTSMC system
One can observe that the trackingg responses
p obtained from the NTSMC system
y
are much degraded compared with the ones obtained from the TSMC system.
Though NTSMC solve the singular problem of TSMC, the steady-state errors
become worse ((the p
proof will be discussed in conclusion).
)
0.1mm 0.1mm
Reference Reference
Reference Reference
(a) (e) (a) (e)

(b) (f) (b) (f)
Nonsingular Terminal Sliding Surface Nonsingular Terminal Sliding Surface Nonsingular Terminal Sliding Surface Nonsingular Terminal Sliding Surface
0V 0V 0V 0V

(c) (g) (c) (g)
0A 0A 0A 0A
0A 0A 0A 0A
(d) (h) (d) (h)

Electric Case 2
Laboratory,
64 EE, NCU, Taiwan
Using
g Terminal SMC
The reasons of the unbounded control efforts for TSMC and
NTSMC systems
y are:
The control law of TSMC system:
−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 ⎥⎦
The control law of NTSMC system:

−1
⎡ q
2q− p
⎤
U NTSMC (t ) = Gn ⎢&z&m (t ) − An z& (t ) − Bn z (t ) + λ e&(t ) q
+ δsgn ( sn (t ))⎥
⎣⎢ p ⎦⎥
If p=5, q=3, then

q
−1 1 singular
i l problem
bl
e(t ) p
e&(t ) = e&(t )
2
5
e (t )
2 q− p 1
enlarged control effort More frequently
e&(t ) q
= e&(t ) 3
Ex: 0.02^(1/3)=0.2714
65 EE, NCU, Taiwan
Using
g Terminal SMC
C.Experimental results of TAMB control system using RNTSMC system
Accordingg to the on-line lumped p uncertaintyy estimation via MISO RHNN
estimator, the best control performance and robustness are obtained.
0.1mm 0.1mm
Reference Reference
Start Start Trajectory Start Start
Trajectory Reference Reference
－0.1mm
01 R t Position
Rotor P iti 2Sec 00
0.05mm 01
－0.1mm Rotor Position 2S
2Sec 0.05mm 01
－0.1mm 2Sec 00
0.05mm －0.1mm
01 2S
2Sec
0.05mm
(a) (f) (a) (f)

(b) (g) (b) (g)
Nonsingular Terminal Sliding Surface Nonsingular Terminal Sliding Surface Nonsingular Terminal Sliding Surface Nonsingular Terminal Sliding Surface
0V 0V 0V 0V

(c) (h) (c) (h)
Estimated Value of Lumped Uncertainty Estimated Value of Lumped Uncertainty Estimated Value of Lumped Uncertainty Estimated Value of Lumped Uncertainty
0V 0V 0V 0V
1V 2Sec 1V 2Sec 1V 2Sec 1V 2Sec

(d) (i) (d) (i)
0A 0A 0A 0A
0A 0A 0A 0A
(e) (j) (e) (j)

Electric Case 2
Laboratory,
66 EE, NCU, Taiwan
Using
g Terminal SMC
The performance measures of the TSMC, NTSMC and RNTSMC
systems
y are shown in Table 5.1 and Table 5.2,, respectively.
p y
Obviously, the proposed RNTSMC possesses the best robust control
Table 5.1 Performance measures of TAMB control system using TSMC, NTSMC, and RNTSMC systems at Case 1.
TSMC NTSMC RNTSMC
Sinusoid Trapezoid
p
Maximum 7.697 6.997 8.448 7.834 5.303 5.963
Average 2.086 1.984 4.161 2.376 0.826 0.561

Table 5.2 Performance measures of TAMB control system using TSMC, NTSMC, and RNTSMC systems at Case 2.
TSMC NTSMC RNTSMC
T ki Errors,
Tracking E μm
Maximum 7.260 9.140 9.541 9.355 5.666 7.372
Average 2.866 2.127 4.114 2.533 0.908 0.626

Electric2.203
67 EE, NCU, Taiwan
Outline
Abstract
Introduction
y
TAMB System
Five-DOF AMB System
Sliding-Mode
Mode Control
Di
Discussions
i andd Conclusions
C l i
68 EE, NCU, Taiwan
Robust Control of Fully Suspended Five-DOF AMB
System
y Using
g Decentralized PIDNN Control
The network structure of the PIDNN controller used for φ-axis of
the five-DOF AMB system
y is shown as:
P control
u1 o1 (
o1 ( N ) = f P u1 ( N ) )
Σ fP
wij1 2
= ∑ wi11 ( N )φi ( N )
2
w j
eφ Z −1 1
i =1
Z −1
u2 o2
Σ
I control
Σ fI vφ
(
o2 ( N ) = f I u2 ( N ) + o2 ( N − 1) )
e&φ 2
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

69 EE, NCU, Taiwan
System
y Using
Te on-line learning algorithm of the PIDNN using supervised
ggradient decent method: Connective Weight Between
Hidden and Output Layers
∂E ( N )
Δ w 2j ( N ) = −η W 2 = η W 2 δ φ ( N )o j ( N )
∂ w 2j ( N )
u1 o1 w 2j ( N + 1) = w 2j ( N ) + Δ w 2j ( N )
wij1 Σ fP Optimal Learning Rate

2
w j E( N )
eφ ηW = 2
Z −1 1 ⎧ 3 ⎡ ∂E ( N ) ⎤ 2 ⎫
Z −1 ⎪ ⎪
u2 o2 2⎨∑ ⎢ 2 ⎥
+ ε ⎬
Σ fI Σ vφ ⎪⎩ j =1 ⎢⎣ ∂w j ⎦⎥ ⎪⎭
Connective Weight Between

e&φ Input and Hidden Layers
u3 o3 ∂E ( N )
Δwij1 ( N ) = −ηW 1 = ηW 1δ φ ( N ) w2j ( N )φi ( N )
Σ fD ∂wij ( N )
1
−1 wij ( N + 1) = wij1 ( N ) + Δwij1 ( N )

1
Z −1
Optimal Learning Rate
Layer 1 Layer 2 Layer 3 E(N )
ηW =
Input Layer Hidden Layer Output Layer 1
⎧ 2 2
⎫
⎪ 3 ⎡ ∂E ( N ) ⎤ ⎪
2⎨∑ ∑ ⎢ ⎥ +ε⎬
1 2 ⎢⎣ ∂ w ij ⎥⎦
1
⎪⎩ i =1 j =1 ⎪⎭
E ( N + 1) = eφ ( N + 1) < E ( N )
2 Electric Machinery and Control Laboratory,
70 EE, NCU, Taiwan
System
y Using
The configuration of the ψ-axis AMB control subsystem using PIDNN
controller with on-line learning algorithm and adaptive learning rates is shown
as follows:
f ll
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

71 EE, NCU, Taiwan
System
y Using
A. Experimental results of fully suspended five-DOF AMB control
system
y usingg decentralized PID controller
Five conventional PID controllers are also used to construct a decentralized
PID controller for the comparison of control performances.
t
uφ = K Pφ eφ + K Iφ ∫ eφ dt + K Dφ e&φ
0
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)
Rotor Position in z-Axis Rotor Position in z-Axis Rotor Position in z-Axis

z (mm)
z (mm)
z (mm)
0 0 0
0.1mm 0.05sec 0.1mm 0.05sec 0.1mm 0.05sec

Time (sec) Time (sec) Time (sec)
Case 1 Case2 Case 3

2400RPM 2400RPM with Load 4800RPM
72 EE, NCU, Taiwan
System
y Using
system
y usingg decentralized PIDNN controller
Five proposed PIDNN controllers are used to construct a decentralized PIDNN
controller to achieve the regulating and stabilizing purposes for the fully
suspended
p five-DOF AMB control system.
y
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)
z (mm)
z (mm)
0 0 0

Case 1 Case2 Case 3

73 EE, NCU, Taiwan
System
y Using
The performance measures of the decentralized PID controller,
decentralized PIDNN controller are shown in the following.
g
Obviously, the proposed decentralized PIDNN controller possesses
the better robust control performance.
Regulating Errors Case 1 Case 2 Case 3
Axes
(μm)
PID PIDNN PID PIDNN PID PIDNN
RMS Values 79.02 60.68 92.94 77.64 82.72 73.81
x1
Peak to Peak Values 359.1 278.7 390.5 373.3 356.0 286.5
y1 RMS Values 88.39 50.85 114.09 72.38 130.60 87.62
RMS Values 90.23 59.95 102.98 70.86 125.87 104.42

x2
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
RMS Values 14.42 9.45 17.01 11.30 21.45 12.34

z

74 EE, NCU, Taiwan
Outline
Abstract
Introduction
y
TAMB System
Five-DOF AMB System
Sliding-Mode
Mode Control
Di
Discussions
i andd Conclusions
C l i
75 EE, NCU, Taiwan
System
y Using
g Decentralized IDISMC System
y
From the developed decoupled dynamic model, the standard second
order state equation
q of the one-axis AMB subsystem
y is:
M X&& = AX + BU + M H
&x&(t ) = Ax (t ) + BU (t ) + h(t )
Considering the separation of the nominal parameters and parameter

variations as follows: ⎡ h ⎤ ⎡ − α y& + α y& + 2k β x + 2k β i + γ f
⎢ ⎥ ⎢
x1 1 1 1 2 rp 2 2 ri 2 x2 1 dx ⎤
⎥
&x&(t ) = [ An + ΔA( x; t )]x(t ) + [Bn + ΔB ( x; t )]U (t ) + h(t ) ⎢ hx2 ⎥ ⎢ α 2 y1 − α 2 y2 + 2k rp β 2 x1 + 2k ri β 2ix1 + γ 2 f dx ⎥
& &
⎢ hy1 ⎥ = ⎢ α1 x&1 − α1 x&2 + 2k rp β 2 y2 + 2k ri β 2i y2 + γ 1 f dy − g ⎥
⎢ ⎥ ⎢ ⎥
= An x(t ) + BnU (t ) + L( x; t ) ⎢hy2 ⎥ ⎢− α 2 x&1 + α 2 x&2 + 2k rp β 2 y1 + 2k ri β 2i y1 + γ 2 f dy − g ⎥
⎢⎣ hz ⎦⎥ ⎢⎣ γ 3 f dz ⎥⎦
L(x;t) is called the lumped uncertainty and defined as

L( x; t ) = ΔA( x; t )x(t ) + ΔB( x; t )U (t ) + h(t )
The bound of the lumped uncertainty is necessary to be known in

advance and satisfies the inequality as follows:
L ( x; t ) < δ
76 EE, NCU, Taiwan
System
y Using
g Decentralized ISMC System
y
A. Decentralized Integral Sliding-Mode Control (ISMC)
First,, the regulating
g g error e≡xd－x is defined where e represents
p the reference
position and is designed as zero in the AMB applications.
Then, a conventional integral sliding surface s1 is defined as follows:
t
s1 (t ) = e&(t ) + c1e(t ) + c2 ∫ e(τ ) dτ
0
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.

77 EE, NCU, Taiwan
System
y Using
g Decentralized Ideal IDISMC System
y
B. Decentralized Intelligent Double Integral Sliding-Mode Control
(a) Ideal IDISMC
To improve the steady-state control performance, a double integral sliding
surface s2 is defined as follows:
t t t
s2 (t ) = e&(t ) + c1e(t ) + c2 ∫ e(τ )dτ + c3 ∫ ∫ e(τ )dτ dτ
0 0 0
Differentiating s2(t) with respect to time, one can obtain

t
s&2 (t ) = e&&(t ) + c1e&(t ) + c2e(t ) + c3 ∫ e(τ )dτ
0
t
= &x&d (t ) − An x(t ) − BnU (t ) − L( x; t ) + c1e&(t ) + c2 e(t ) + c3 ∫ e(τ )dτ
0
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 ⎥⎦

78 EE, NCU, Taiwan
System
y Using
g Decentralized Ideal IDISMC System
y
Observing the ISMC and ideal IDISMC, the integral term is embedded in both
the integral and double integral sliding surfaces but reflected only in the
IDISMC control
t l law.
l
t
s1 (t ) = e&(t ) + c1e(t ) + c2 ∫ e(τ ) dτ
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
−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
79 EE, NCU, Taiwan
System
y Using
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 )
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 ⎦
Layer 1 Layer 2 Layer 3 Electric Machinery and Control Laboratory,

80 EE, NCU, Taiwan
Input Layer Hidden Layer Output Layer
System
y Using
y
(c) IDISMC system
Theorem 7.1: Considering g the AMB subsystem y represented
p by
y ((7.2),
), if the
proposed control law UIDISMC of IDISMC, which is composed of a robust
controller Ur with adaptive law, and the MPIDNN observer Uy with learning
algorithm, is adopted as the control effort , then the asymptotic stability of the
AMB subsystem is guaranteed.
U IDISMC (t ) = U r (t ) + U y (t )
U r (t ) = Bn [&x&d (t ) − An x(t ) + εˆ + ξs2 (t )] using adaptive law ε&ˆ = ηε s2 (t )
−1
&
−1
U y (t ) = Bn yˆ (t ) using estimation law Wˆ = η s (t )O
W 2
Proof: Choose the Lyapunov function as

~ 1 2 B ~ ~ B h W~ = W ∗ − Wˆ and
V ( s2 (t ), W , ε~ ) = s2 (t ) + n W T W + n ε~ 2 where d ε~ = ε − εˆ
2 2ηW 2ηε
Taking the time derivative of the Lyapunov function, one can obtain
~ 1 ~ T &ˆ 1 ~ &
V& ( s (t ), W , ε~ ) = s (t ) s& (t ) − W W − ε εˆ = −ξs (t )
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.
81 EE, NCU, Taiwan
System
y Using
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
Algorithm Z−1 ŷ MPIDNN ib + ic ib − ic

fI
Σ Observer
(7.25)
Σ
Σ fD
(7.23) One-Axis AMB
d
Z −1
Subsystem
& (7.2)
dt Σ fH
Ŵ
IDISMC

82 EE, NCU, Taiwan
System
y Using
y
system
y usingg decentralized ISMC system
y
Five ISMC systems are used to construct a decentralized ISMC system for the
comparison of control performances.
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
x 1 －Axis (mm) x2 －Axis (mm) x1 －Axis (mm) x2 －Axis (mm) x 1 －Axis (mm) x2 －Axis (mm)

z (mm)
z (mm)
z (mm)
0 0 0

Case 1 Case2 Case 3

83 EE, NCU, Taiwan
System
y Using
y
system
y usingg decentralized IDISMC system
y
Five proposed IDISMC systems are used to construct a decentralized IDISMC
system to achieve the regulating and stabilizing purposes for the fully
suspended
p five-DOF AMB control system.
y
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)
z (mm)
0 0 0

Case 1 Case2 Case 3

84 EE, NCU, Taiwan
System
y Using
y
The performance measures of the decentralized ISMC system,
decentralized IDISMC system
y are shown in the following.
g
Obviously, the proposed decentralized IDISMC system possesses
the better robust control performance.
Regulating Errors Case 1 Case 2 Case 3
Axes
(μm)
ISMC IDISMC ISMC IDISMC ISMC IDISMC
RMS Values 56.27 45.04 69.93 45.55 76.90 45.81
x1
y1 RMS Values 59.87 45.29 82.39 62.94 102.34 77.54
RMS Values 66.32 58.48 79.50 61.28 125.34 97.27

x2
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
RMS Values 10.43 8.93 11.70 10.43 14.06 10.93

z

85 EE, NCU, Taiwan
Outline
Abstract
Introduction
y
TAMB System
Five-DOF AMB System
Sliding-Mode
Mode Control
Di
Discussions
i andd Conclusions
C l i
86 EE, NCU, Taiwan
Discussions
This dissertation had developed various PC-based control systems to
control the rotor pposition in the axial direction of the TAMB system
y
for the tracking of various reference trajectories.
Therefore, the moveable controlled rotor of the TAMB system can be applied to
different operating demands and environments.
environments
On the other hand,

hand several PC
PC-based
based decentralized control systems
were developed based on the proposed decoupled dynamic model to
regulate and stabilize the five-axes of the rotor of the five-DOF
AMB system in the centers considering the existences of the serious
uncertainties.
According to the decoupled procedure and decentralized concept, the
controller design was simplified and the computational burden was reduced.

87 EE, NCU, Taiwan
Discussions
Table 8.1 shows the comparison of different ultimate bounds of
various SMC approaches
pp includingg SMC,, CSMC,, ACSMC,, TSMC,,
NTSMC, and RNTSMC.
Since SMC, CSMC, TSMC, and NTSMC use saturation functions for
chattering reduction,
reduction only GUUB stabilities can be guaranteed.
guaranteed
Since the saturation functions are not used in the ACSMC and RNTSMC, both
of them can guarantee the asymptotic stabilities through Lyapunov theorem and
Barbalat’ss Lemma.
Barbalat Lemma
SMC CSMC ACSMC TSMC NTSMC RNTSMC
2 2
⎛d ⎞ ⎛d ⎞
s g (t ) = ⎜ + λ ⎟ Λ (t ) s g (t ) = ⎜ + λ ⎟ Λ (t )
⎝ dt ⎠ ⎝ dt ⎠
p p
1 1
Sliding ⎛d ⎞ ⎛d ⎞⎛ d ⎞ ⎛d ⎞⎛ d ⎞ q s n (t ) = e (t ) + e&(t ) q sn (t ) = e(t ) + e&(t ) q
s( t ) = ⎜ + λ ⎟e(t ) sc (t ) = ⎜ + λ ⎟⎜ − λ ⎟ Λ(t ) sc (t ) = ⎜ + λ ⎟⎜ − λ ⎟Λ (t ) sm (t ) = e&(t ) + λe(t ) p λ λ
Surfaces ⎝ dt ⎠ ⎝ dt ⎠⎝ dt ⎠ ⎝ dt ⎠⎝ dt ⎠
where p > q where p > q
t t
where Λ ( t ) = ∫ e(τ )dτ where Λ (t ) = ∫ e(τ )dτ
0 0
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 →∞
GUUB stabilities asymptotic stabilities Electric

GUUBMachinery
stabilitiesand Control Laboratory,
asymptotic stabilities
88 EE, NCU, Taiwan
Discussions
The ultimate bounds of tracking error e can be ordered as
TSMC<CSMC<SMC<NTSMC.
The ultimate bounds of tracking error’s derivative e& can be ordered as
CSMC<TSMC=SMC<NTSMC.
Obviously, CSMC and TSMC own the two best steady-state control
performances.
SMC CSMC ACSMC TSMC NTSMC RNTSMC

2 2
⎛d ⎞ ⎛d ⎞
s g (t ) = ⎜ + λ ⎟ Λ (t ) s g (t ) = ⎜ + λ ⎟ Λ (t )
⎝ dt ⎠ ⎝ dt ⎠
p p
1 1
Sliding ⎛d ⎞ ⎛d ⎞⎛ d ⎞ ⎛d ⎞⎛ d ⎞ q s n (t ) = e (t ) + e&(t ) q sn (t ) = e(t ) + e&(t ) q
s( t ) = ⎜ + λ ⎟e(t ) sc (t ) = ⎜ + λ ⎟⎜ − λ ⎟ Λ(t ) sc (t ) = ⎜ + λ ⎟⎜ − λ ⎟Λ (t ) sm (t ) = e&(t ) + λe(t ) p λ λ
Surfaces ⎝ dt ⎠ ⎝ dt ⎠⎝ dt ⎠ ⎝ dt ⎠⎝ dt ⎠
where p > q where p > q
t t
where Λ ( t ) = ∫ e(τ )dτ where Λ (t ) = ∫ e(τ )dτ
0 0
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 →∞

89 EE, NCU, Taiwan
Discussions
The different ultimate bounds of SMC, CSMC, TSMC, and NTSMC
systems
y due to various values of λ are depicted
p in Fig.
g 8.1.
The larger values of λ result in the smaller ultimate bounds of tracking error e
for SMC, and CSMC, and TSMC systems;
The larger values of λ result in the larger ultimate bounds of tracking error’s
error s
derivative e& for NTSMC system.
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.
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

91 EE, NCU, Taiwan
Discussions
The comparison of control performance measures including RMS
and ppeak to ppeak values of respective
p axes at Cases 1-3:
The control performance behaviors in accordance with the descending order are
decentralized IDISMC system, decentralized PIDNN controller, decentralized
ISMC system,
y , and decentralized PID controller
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
RMS values at Case 1 RMS values at Case 2 RMS values at Case 3

PID PIDNN ISMC IDISMC PID PIDNN ISMC IDISMC PID PIDNN ISMC IDISMC
(μm) (μm) (μm)
500 500 500
450 450 450
400 400 400
350 350 350
300 300 300
250 250 250
200 200 200
150 150 150
100 100 100
50 50 50
0 0 0
x1x1 y1x2 x2y1 y2y2 zz x1x1 y1x2 x2y1 y2y2 zz x1x y1x x2y y2y zz
1 2 1 2
Peak-to-peak values at Case 1 Peak-to-peak values at Case 2 Peak-to-peak values at Case 3

92 EE, NCU, Taiwan
Conclusions
The major contributions of this dissertation are concluded as
follows:
1) the complete analyses of the dynamics and the detailed discussions of the
drive systems for TAMB and RAMB systems;
2) the successful developments of the various tracking controllers for the TAMB
control system;
3) the complete comparison of guaranteed ultimate bounds of various SMC
approaches while using saturation functions;
4) the successful implementations of the TAMB control system using the
developed tracking controllers;
5) the
h successful
f l derivation
d i i off the h decoupled
d l d dynamic
d i model
d l off the
h five-DOF
fi DOF
AMB system for the purpose of decentralized control;
6) the successful developments of the various decentralized control systems for
the regulation and stabilization of the fully suspended five-DOF AMB control
system;
7) the successful implementations of the five-DOF AMB control system using
the developed decentralized control systems.
93 EE, NCU, Taiwan
Thank You for Your
Attention!
Questions?

94 EE, NCU, Taiwan
Using
g TSMC
How to calculate ts ? q
After the terminal sliding y e&(t ) = −λe(t )

g surface is reached,, the equality
q p
holds.
Thereafter,
q q
de(t ) dt 1 −
= −λe(t ) ⇒
p
= − e(t ) p
dt de(t ) λ
q q
1 − 1 −
de(t ) ⇒ ∫ dt = −
λ∫
dt = − e(t ) p
e(t ) p
de(t )
λ
0
⎡ ⎤
1⎢ 1 1− ⎥
q q
1 0 −
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 ⎥⎦
95 EE, NCU, Taiwan

Sliding Mode

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Sliding Mode

Uploaded by

Copyright:

Available Formats

Intelligent Sliding-Mode Control for

Prof. Faa-Jeng Lin

Electric Machinery and Control Laboratory

Electric Machinery and Control Laboratory,

TAMB Positioning and Development of Tracking

Reference Trajectories and Performance

Five-DOF AMB Regulating and Development of Decentralized

Operating Conditions and Performance

 For several kinds of the magnetic suspension systems, magnetic

Electric Machinery and Control Laboratory,

 MB can be divided into the passive MB (PMB), i.e. permanent MB,

 Power Amplifier Rotor

 According to its structure and usage of AMB:

Magnetic Bearing Stage

Electric Machinery and Control Laboratory,

Class-A Class-B Class-C

Electric Machinery and Control Laboratory,

PC-Based Control Core Inverter Drive System Five-DOFand

 Consider the Maxwell force l0

k aii ≡ z = 4k a 02 Nominal Air Gap z0 : 00.5mm

Control14 Bits Resolution i0 + iz

Rotor Position z Sensor Signal

C. Performance Measures and Comparisons

Electric Machinery and Control Laboratory,

Proposed ix , iy : control currents in X-Y axes

⎪m&y&c = Fy1 + Fy2 − mg + f dy

 The dynamic model of the five-DOF AMB system ⎥ = ⎢be

the electromagnetic force vector; D = [ f dx f dyy f dz ]T is the external disturbance

vector; C = [0 0 − 1 − 1 0]T is the gravity vector; M , G , K ,and E are the

is the decoupled term vector; A and B are the stiffness,

Inverter ADC 1 Proposed DAC 1 v x1 vb

Left RAMB Inductor Right RAMB

Total Total ADC 5 Proposed DAC 5 v0

Source: S2M Electric Machinery and Control Laboratory,

 The varied learning rates, ∑

error based on the analyses Γj β

0.9 Fuzzy point of view: Fuzzy point of view:

0.88 0.88 Γ 0.88 0.88 0.31 0.62 Γ 0.47 0.48β

Electric Machinery and Control Laboratory,

0.1mm 0.1mm 0.05mm 0.05mm

Tracking Error Tracking Error Tracking Error Tracking Error

Start Start Start Start

0.01mm 0.01mm 2Sec 0.01mm 2Sec 0.01mm 2Sec

Case 1 Case2 Case 1 Case 2

Start Start Start Start

0.01mm 2Sec 0.01mm 2Sec 0.01mm 2Sec 0.01mm 2Sec

Case 1 Case2 Case 1 Case 2

Start Start Start Start

0.01mm 2Sec 0.01mm 2Sec 0.01mm 0.01mm 2Sec

Case 1 Case2 Case 1 Case 2

Average 4.625 2.428 2.074 2.153 1.874 1.678

Average 4.962 2.808 2.401 2.451 2.043 1.860

 The above dynamic model is rewritten considering the separation of

where L( z; t ) is called the lumped uncertainty and defined as

 Here, the bound of the lumped uncertainty is necessary to be known

Electric Machinery and Control Laboratory,

Boundary layer thickness Φ

Therefore, a trade-off problem between chattering and control accuracy arises.

 For ease of discussion, the outputs of the Σ Σ Layer 3

αˆ = [ w11 w21 L wn1 ]T , βˆ = [ w12 w22 L wn 2 ] .

Q(t ) = Gn Ξ(z; t ) , Ξ(z; t ) = 2e&(t ) + λe(t ) − sc (t ) + Gn sc (t ) .

 The MIMO RHNN estimator is proposed to estimate the two complicated

Electric Machinery and Control Laboratory,

For several kinds of the magnetic suspension systems, magnetic

MB can be divided into the passive MB (PMB), i.e. permanent MB,

Power Amplifier Rotor

According to its structure and usage of AMB:

Consider the Maxwell force l0

The dynamic model of the five-DOF AMB system ⎥ = ⎢be

The varied learning rates, ∑

The above dynamic model is rewritten considering the separation of

Here, the bound of the lumped uncertainty is necessary to be known

For ease of discussion, the outputs of the Σ Σ Layer 3

The MIMO RHNN estimator is proposed to estimate the two complicated

A finite time ts, which is taken to travel from e(t

Therefore, both the tracking error e

Moreover,, when sn=0,, the system

The control law of NTSMC system:

If p=5, q=3, then

Considering the separation of the nominal parameters and parameter

L(x;t) is called the lumped uncertainty and defined as

The bound of the lumped uncertainty is necessary to be known in

Differentiating s2(t) with respect to time, one can obtain

Proof: Choose the Lyapunov function as