Professional Documents
Culture Documents
On-Line Intelligent Control of Submarine Periscopes: July 2007
On-Line Intelligent Control of Submarine Periscopes: July 2007
net/publication/4287418
CITATIONS READS
2 1,046
3 authors, including:
Ali Kazemy
Tafresh University
46 PUBLICATIONS 208 CITATIONS
SEE PROFILE
Some of the authors of this publication are also working on these related projects:
All content following this page was uploaded by Ali Kazemy on 22 March 2016.
a θ3 y3
⎡ 0 ⎤ ⎡⎢ bc1s 2 ⎤⎥
1
x 2 θ2 x3 ⎡x 2 ⎤ ⎢ ⎥
b⎥ ⎢
2
⎥
z2 ⎢ y ⎥ = 0T 1T ⎢ − ⎢ 1 bs s ⎥
⎢ ⎥ ⎢ 2 ⎥ = (4)
⎢ 0 ⎥ ⎢⎢ 2 ⎥
2 1 2 1 2
y2 ⎢⎣ z 2 ⎥⎦ ⎥
b ⎢ ⎥ ⎢1 ⎥
⎣⎢ 1 ⎦⎥ ⎢⎣ 2 bc 2 + d ⎥⎦
z1 d ⎡1 ⎤
y1 ⎢ 2 b ( −ω1s1s 2 + ω2c1c 2 ) ⎥
θ1 ⎡ x& 2 ⎤ ⎢ ⎥
⎢ y& ⎥ = ⎢ 1 b ω c s + ω s c ⎥
x1
⎢ 2⎥ ⎢ 2 ( 1 1 2 2 1 2 ) ⎥
(5)
⎣⎢ z& 2 ⎦⎥ ⎢ 1
⎥
⎢ b ω2 s 2 ⎥
Fig. 2. Link coordinate frames of the periscope, shown in Fig. 1.
⎢⎣ 2 ⎥⎦
where 0T1 is the transformation matrix from coordinate one to
The velocity of the upper piece is equal to the velocity of the
coordinate zero (the base coordinate), c1 = cos(θ1 ) ,
lower piece
s1 = sin(θ1 ) and so on.
246
Hence, the velocity of this link is equal to been depicted in Fig. 4. The modeling error has been shown in
Fig. 5. As this figure shows, the error between the
1 experimental data and proposed model is relatively very small.
v 32 = x& 32 + y& 32 + z&32 = a 2 (ω32 + ω22 + ω12c 22 − ω12c 22c 32
4 (11) The input signal, applied to the model and the real plant, id
+ 2ω1ω3c 2 + ω1 c 3 − 2ω1ω2 s 2 s 3c 3 − ω2 c 3 ).
2 2 2 2 shown in Fig. 6.
⎛1 ⎞ 1
k 3 = 2 ⎜ m3v 32 ⎟ + J 3ω32 , (12)
⎝2 ⎠ 2
1 1 1 1
p 3 = m3 gz 31 − m3 gz 32 = m3 gas 2 s 3 + m3 gd
2 2 4 2
(13)
⎛ 1 1 ⎞ 1
− ⎜ − m3 gas 2 s 3 + m3 gd ⎟ = m3 gas 2 s 3 .
⎝ 4 2 ⎠ 2
D. Lagrangian Equation
The Lagrangian equations for the periscope can be written as
⎛ 1 ⎞ ⎛ 1 ⎞
+ ⎜ m3a 2c 2 ⎟ θ&&3 + ⎜ − m3a 2 s 2 s 3c 3 ⎟ θ&&2 + m 2b 2ω1ω2 s 2c 2 60
⎝2 ⎠ ⎝ 2 ⎠ 40
1
+ m3a 2 (−2ω1ω2 s 2c 2 + 2ω1ω2c 2 s 2c 32 + 2ω1ω3c 3s 3c 22 20
2 0
− 2ω2ω3s 2 − 2ω1ω3c 3s 3 − ω22c 2c 3s 3 − ω2ω3s 2c 32 + ω2ω3s 2 s 32 ), -20
⎛ 1 ⎞ ⎛ 1 1 ⎞
τ 2 = ⎜ − m3a 2 s 2 s 3c 3 ⎟ θ&&1 + ⎜ J 2 + m 2b 2 + m3a 2 ⎟ θ&&2 -40
0 2 4 6 8 10
⎝ 2 ⎠ ⎝ 2 2 ⎠ Time(sec)
1 Fig. 4. Comparison of real and model response
+ m3a 2 (2ω1ω2c 2c 3s 3 − ω1ω3s 2c 32 + ω1ω3s 2 s 32
2 Modeling error
+ 2ω3c 3s 3 + ω12c 2 s 2 − ω12c 2 s 2c 32 + ω1ω3s 2 (15) 2
1 1 1.5
+ ω1ω2c 2c 3s 3 ) − m 2b 2ω12 s 2c 2 − m 2 gbs 2
2 2
1
1
+ m3 gac 2 s 3 ,
2 0.5
Degree
⎛1 ⎞ ⎛ 1 ⎞ 1
τ 3 = ⎜ m3a 2c 2 ⎟ θ&&1 + ⎜ J 3 + m3a 2 ⎟ θ&&3 + m3 gas 2c 3 0
⎝2 ⎠ ⎝ 2 ⎠ 2
-0.5
1
+ m3a 2 (−ω1ω2 s 2 − ω12c 3s 3c 22 + ω12c 3s 3 (16)
2 -1
+ ω1ω2 s 2c 32 − ω1ω2 s 2 s 32 − ω2c 3s 3 ).
-1.5
247
Input signal ∂E ∂y ∂u ( k )
8 = −e ( k ) = −e ( k ) y u ( k ) , (19)
∂W ∂W ∂W
6
where y u ( k ) = ∂y ( k ) / ∂u ( k ) represents the sensitivity of the
4
plant with respect to its input. If the term y u ( k ) in equation
2
(19) is eliminated, then the original backpropagation is
0 obtained. The sign of this term is an important factor for
Voltage
⎛ ⎞ I&a =
1
(
− Ra I a − k b θ&M + v a )
y = ρ ⎜ ∑w Oj Q j + b O ⎟ , Q j = ρ (q j ) , (16) La . (18)
⎝ j ⎠ τ = −J θ&& − B θ& + k I
M M M M M a a
1
q j = ∑w ij u i + b j , ρ ( ⋅) =
I I
, (17)
1+ e ( )
− ⋅
i where I a , Ra and La are the armature current, resistance and
inductance respectively, k b and k a are motor torque constant
and back-emf constant, respectively, θ M ,τ M , B M and J M are the
where w ijI and w Oj are the input and output weights,
motor rotational angle, output torque, damping coefficient and
respectively, b jI and b O
are the input and output bias weights, rotor inertia, respectively, and v a is the input voltage to the
respectively, and u i are the inputs to the neural network, which motor.
Motor and plant parameters are written in table 1 and 2,
is shown in Fig. 7. respectively. Fig. 9 depicts the desired and the actual response
Error backpropagation training algorithm has been for θ5 , when the neural network is trained off-line using 400
employed. The error function for the training cycle can be
samples generated from real plat. Is this case, there will be no
defined as on-line training. The tracking error is shown in Fig. 10. As it
can be expected, a non-adaptive controller cannot be robust
1 1
( y r ( k ) − y ( k ) ) = 2 e ( k )2 , against changes in system parameters. Figs. 11 and 12 show
2
E = (18)
2 the tracking and the error while friction parameter, that is
modeled in motor dynamics, is increased by 300% at t = 2 sec.
where y r ( k ) and y ( k ) are the desired and actual responses of Figs. 12-15 show the same results when the NN is trained
on-line. As Figs. 14 and 15 show, the adaptive neuro controller
the plant. The gradient of error in (18) with respect to an
is very robust against changes in system parameters and can
arbitrary weight vector can be written as quickly adapt itself to these changes.
248
VI. CONCLUSION
Tracking with Off-Line MLP
In this paper, first, dynamic equations of a submarine
periscope was derived and verified with real data acquired 10 Reference
from experimental setup. Then, a neuro controller was 8 Tracking
designed to control one link of the system. The neural network 6
controller was trained off-line and on-line to show the
4
robustness of the controller against changes in system
parameters. the system on off-line and on-line case. At the end, 2
Degree
advantage and defect of the proposed controllers had been 0
inspected. Continuation of this research includes application of -2
the neuro controller to the experimental periscope and
-4
comparing the results with the simulations.
-6
TABLE I -8
MOTOR PARAMETERS -10
Ra ( Ω ) La ( mH ) K b ( mV/rpm ) K a ( mNm/A ) 0 2 4
Time(sec)
6 8 10
40.2 0.8 3.2 35.1 Fig. 9. Desired and actual response, when the NN in trained off-line.
Tracking error
8
TABLE II
PLANT PARAMETERS 7
m1 ( kg ) m2 ( kg ) m3 ( kg ) a (m ) b (m ) d (m ) 6
4
Degree
3
ρ
I
wij O
wl 2
u1 ρ
u2 y (k ) 1
ρ
0
O
ui b -1
0 2 4 6 8 10
ρ I
Time(sec)
6
4
TD +- 2
e θ3
Degree
θ3r u Plant 0
N. N. dynamic
-2
equations
Controller -4
-6
TD
-8
TD -10
0 2 4 6 8 10
Time(sec)
Fig. 8. Control block diagram Fig. 11. Reference tracking with off-line MLP control and parameters
changing
249
Tracking error Tracking with On-Line MLP with Changing parameters
8
10 Reference
7
Tracking
6
8
5
4 6
Degree
Degree
3
4
2
1 2
0
0
-1
-2
0 2 4 6 8 10 -2
0 0.5 1 1.5 2 2.5 3
Time(sec)
Time(sec)
Fig. 11. Tracking error with off-line MLP control and parameters changing Fig. 14. Reference tracking with on-line MLP control and parameters
changing
Tracking with On-Line MLP
Tracking error
10 8
8
7
6
6
4
2 5
Degree
0 4
Degree
-2
3
-4
2
-6
-8 1
-10 0
0 1 2 3 4 5
Time(sec) -1
0 0.5 1 1.5 2 2.5 3
Fig. 12 Desired and plant response in on-line control Time(sec)
Fig. 15. Tracking error with on-line MLP control and parameters changing
Tracking error
8 REFERENCES
7 [1] Y.M. Liang, H.R. Tyan, H.Y.M. Liao and S.W. Chen, "Stabilizing image
sequence taken by the camcorder mounted on a moving vehicle,"
6 Proceedings of IEEE International Conference on Intelligent
Transportation Systems, pp. 90-95, Shanghai, China, 2003.
5 [2] Ping-Ho Chen, "Periscope using common optical path to have stabilized
panoramic view," US patent No. US 6,347,010B1, Feb. 12, 2002.
4 [3] Ping-Ho Chen, "A panoramic stabilized periscope with common optical path,"
Degree
250