IASTED MIC 2000
1Automatic Control Lab. YNU, Korea
Precise Motion Control of Linear Pulse Motor based on Disturbance
Compensation using Neural Network
2000. 2. 14.
Young Gun Kwon, Jung-il Park
School of Electrical and Electronic Engineering,Yeungnam University, Korea
IASTED MIC 2000
2Automatic Control Lab. YNU, Korea
Purpose
• Propose a neural network-based learning controller
– Finding the current commands to reduce the ripple of error occurred when we control a linear pulse motor
IASTED MIC 2000
3Automatic Control Lab. YNU, Korea
Introduction
• Importance of a linear motor
• What is a linear motor?
• What is a linear pulse motor(LPM)?
– Merits of LPM
– Difficulties in controlling LPM using conventional control algorithm
IASTED MIC 2000
4Automatic Control Lab. YNU, Korea
• Excitation methods of LPM– Micro-step control to reduce force ripple
• Force ripple – Due to nonlinear factor or unmodeled
dynamics
• Studies for reducing force ripple – Finding the accurate model via either
experiments or theoretical studies for dynamics of a LPM
IASTED MIC 2000
5Automatic Control Lab. YNU, Korea
• Purpose of this paper :
– Getting a better performance by finding the current commands considering nonlinear factor and unmodeled dynamics
– Introducing the neural network-based current controller considering these factors at a time
• Validity of the proposed controller
– confirmed by motion control experiments
IASTED MIC 2000
6Automatic Control Lab. YNU, Korea
Principle of Hybrid type LPM
N S
Permanent magnet
Electromagnet
N S
Permanent magnet
Electromagnet
N S
Permanent magnet
Electromagnet
N S
Permanent magnet
Electromagnet
(a) (b)
(c) (d)
IASTED MIC 2000
7Automatic Control Lab. YNU, Korea
Neural network-based current controller
• Linear motor model
• Generating force
eL FFdtdxB
dtxdM =++2
2
(1)
( )
+
−= BABAe ix
pix
pKxiiF ππ 2cos2sin,, (2)
IASTED MIC 2000
8Automatic Control Lab. YNU, Korea
• Current commands
– If we use the current commands in equation (3), the force ripple is generated by unmodeled dynamics.
– Propose a neural network-based current control scheme
=
−= x
pix
pi BA
ππ 2cos,2sin(3)
IASTED MIC 2000
9Automatic Control Lab. YNU, Korea
• Structures of neural network applied in control fields
– Feedback error learning type
– Direct inverse model learning type
– Indirect model learning type
IASTED MIC 2000
10Automatic Control Lab. YNU, Korea
Structure of proposed controller
setpoint y
OutputPlantT H K
2HRM0205-1020L
err MotorDriver
xia
ib
++ A, /A
B, /B
PIDcontroller
CurrentCommand
+
-
++
NeuralNetwork1
++
uPID
uNN1
uIN
NeuralNetwork2
ibia
IASTED MIC 2000
11Automatic Control Lab. YNU, Korea
Structure of used neural networksetpoint(t)
output(t)
out_velocity(t)
velocity(t)
LPF
uNN1
(a) Neural network 1
IASTED MIC 2000
12Automatic Control Lab. YNU, Korea
setpoint(t)
setpoint(t-1)
Output(t)
Output(t-1)
IA
IB
IB
IA
(b) Neural network 2
IASTED MIC 2000
13Automatic Control Lab. YNU, Korea
Specifications of neural network
– Bipolar sigmoid activation function
– Momentum term
– Standard back-propagation algorithm
IASTED MIC 2000
14Automatic Control Lab. YNU, Korea
• The j-th hidden layer output
( ) ∑=
===N
kkjkjjj Qjxvgh
0,,2,1;; Lψψ (4)
• i-th output of output layer
( ) ∑=
===Q
jjijiii Mihwfy
0,,2,1;; Lψψ (5)
• Cost function
( )∑=
=M
iPIDUE
1
2
21 (6)
IASTED MIC 2000
15Automatic Control Lab. YNU, Korea
( ) ( )
( ) ( ) ( )nhnnwwEnwnw
jyiij
ijijij
ηδ
η
+=
∂∂
−=+1(7)
( ) ( )i
iPIDyi
fUwhereψψδ
∂∂
=,
( ) ( )
( ) ( ) ( )nxnnvvEnvnv
khjjk
jkjkjk
ηδ
η
+=
∂∂
−=+1(8)
( )j
jM
iijyihj
gwwhere
ϕϕ
δδ∂
∂=∑
=1,
IASTED MIC 2000
16Automatic Control Lab. YNU, Korea
Experimental environments• Methods of experiments
– Using DSP board to implement control algorithm• TMS320C31
– Motion speed • High speed : 4.5mm/sec
• Low speed : 0.2mm/sec
IASTED MIC 2000
17Automatic Control Lab. YNU, Korea
• Specifications of experimental system
– LPM(2HRM0205-1020L) • Pitch length : 0.8mm
• Stroke : 1m
– PWM current driver
– Linear scale resolution : 20㎛
– DSP board : TMS320C31 • Sampling time : 2.5msec
IASTED MIC 2000
18Automatic Control Lab. YNU, Korea
Structure of experimental system
PCDSP
DAC
DAC
Counter
Driver LPM scale A B/A /B
IASTED MIC 2000
19Automatic Control Lab. YNU, Korea
Block diagram of current driver
LPMPWM
L298NStep
motor driver
LA25-NPCurrentdetector
inputIa, Ib
+
-
PI
IASTED MIC 2000
20Automatic Control Lab. YNU, Korea
Experiment results• Initial conditions of NN1 :
– learning rate =0.3
– momentum coefficients =0.9
• Initial conditions of NN2
– learning rate =0.1
– momentum coefficients =0.9
• Initial weights :– chosen as values from –0.1 to +0.1
IASTED MIC 2000
21Automatic Control Lab. YNU, Korea
Desired velocity profile
0 1 2 3 4 5 6 7 8 9 100
1
2
3
4
5
Velo
city
[mm
/sec
]Time [sec]
0 4 8 12 16 20 24 28 32 36 400.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8 x10-3
Velo
city
[m/s
]
Time [sec]
Low speed High speed
IASTED MIC 2000
22Automatic Control Lab. YNU, Korea
Results at low speed
0 4 8 12 16 20 24 28 32 36 40-0.10
-0.08
-0.06
-0.04
-0.02
0.00
0.02
0.04
0.06
0.08
0.10 x10-3
Erro
r [m
]
Time [sec]0 4 8 12 16 20 24 28 32 36 40
-0.10
-0.08
-0.06
-0.04
-0.02
0.00
0.02
0.04
0.06
0.08
0.10 x10-3
Erro
r [m
]
Time [sec]
Open loop controller only FF+PI controller
IASTED MIC 2000
23Automatic Control Lab. YNU, Korea
Results at low speed
0 4 8 12 16 20 24 28 32 36 40-0.10
-0.08
-0.06
-0.04
-0.02
0.00
0.02
0.04
0.06
0.08
0.10 x10-3
Erro
r [m
]
Time [sec]0 4 8 12 16 20 24 28 32 36 40
-0.10
-0.08
-0.06
-0.04
-0.02
0.00
0.02
0.04
0.06
0.08
0.10 x10-3
Erro
r [m
]
Time [sec]
NN2+PI+FF controllerNN1+PI+FF controller
IASTED MIC 2000
24Automatic Control Lab. YNU, Korea
Results at low speed
0 5 10 15 200.00
0.02
0.04
0.06
0.08
0.10
0.12x10-3
MSE
Learning iteration
0.0050.050PI+FF+NN2
0.0100.075PI+FF+NN1
0.0200.075PI + FF
0.0400.095Open loop
Offset error
(mm)
Peak to peak
error(mm)Controller
Mean squared error, PI+FF+NN2 controller
Performance comparison of each controller at low speed
IASTED MIC 2000
25Automatic Control Lab. YNU, Korea
Results at high speed
0 1 2 3 4 5 6 7 8 9 10
-0.1
0.0
0.1
0.2
0.3 x10-3
Erro
r [m
]
Time [sec]0 1 2 3 4 5 6 7 8 9 10
-0.1
0.0
0.1
0.2
0.3 x10-3
Erro
r [m
]
Time [sec]
Open loop controller only PI+FF controller
IASTED MIC 2000
26Automatic Control Lab. YNU, Korea
Results at high speed
0 1 2 3 4 5 6 7 8 9 10
-0.1
0.0
0.1
0.2
0.3 x10-3
Erro
r [m
]
Time [sec]
PI+FF+NN2 controller
IASTED MIC 2000
27Automatic Control Lab. YNU, Korea
Performance comparison of each controller at high velocity
Controller
Peak to
peak
error(mm)
Offset
error(mm)
Open loop 0.17 0.100
PI+FF 0.13 0.030
PI+FF+NN2 0.10 0.025
IASTED MIC 2000
28Automatic Control Lab. YNU, Korea
Conclusions• Introduced neural network of feedback
error learning type changes a current command to improve position accuracy.
• Neural network works efficiently for reducing the position ripple error.
• We must develop better neural network to operate for periodic error.
Top Related