Date post: | 07-Aug-2015 |
Category: |
Devices & Hardware |
Upload: | jun-steed-huang |
View: | 41 times |
Download: | 3 times |
Robotic Liquid Tension Identification with Particle Swarm Optimized Neural Network
By
Hong Xuan Qian, Jian Bing Wu, Yu Hui Shi and Jun Steed Huang
23 – 26 September 2012
Bandung, Indonesia
2Power Matters.
About the authors and the university
� The authors:• Hong was among the key engineers who set up the first automation line in China for Purina (now part of Nestle), her major is
industry automation and its applications, she is co-founder of GenieView Inc.
• Dr. Wu is a full time professor at Jiangsu University, who pioneered the research in mathematical models for sensor less motor, she collaborates well with industry partners.
• Prof. Shi is the adjunct professor at Jiangsu University, who co-authored the early books of Computational Intelligence ISBN9781558607590 and Swarm Intelligence ISBN9781558605954.
• Prof. Huang is the distinguished professor at Jiangsu University, who has initiated a number of multi-national industry-academic collaborations, he is co-founder of GenieView Inc.
� The university• Up to 8 digits 3.1415926 made by Wenyuan at AD461 remained the most accurate for following 900+ years, was from Jiangsu
where our University is exactly located!
22012 IEEE Symposium on Industrial Electronics and Applications
祖冲之祖冲之祖冲之祖冲之
3Power Matters.
Agenda
1. INTRODUCTION
2. TENSION MODEL FOR ROBOT MOTOR
3. SWARM NEURON OPTIMIZATION
4. MATLAB SIMULATION
5. CONCLUSION AND FUTURE WORK
32012 IEEE Symposium on Industrial Electronics and Applications
4Power Matters.
1. Problem: maintain a constant liquid tension
42012 IEEE Symposium on Industrial Electronics and Applications
A robot arm is tasked to apply the liquid glue on an
object with constant tension. There are a number of
ways to convert motor’s torque to linear tension:
�Single Sensor less Motor is simple and cheap, but
the tension may fluctuate.
�Single Motor with Sensors, the tension is under
control by the sensors but could be expensive.
�Dual Motor System without Sensors is relatively
simple, smooth and not too expensive.
Power Matters.
1. Solutions that we tried and failed
52012 IEEE Symposium on Industrial Electronics and Applications
a) The direct motor model calculation cannot follow the liquid
variation from the random interference and measurement errors
occurred in our factory;
b) Kalman filter can deal with random noise and measurement
errors, but it can not follow the variation of individual motor
parameter and its related aging factor;
c) The complicated adaptive model method can partially deal with
the individual variation, but it does not treat the random noise
and measurement error, as it may not converge properly due to
real time computation burden.
d) The error Back-Propagation (BP) algorithm is a popular training
method for feed-forward Neural Network (NN), but we found the
parameter trained for one motor does not fit the other.
Power Matters.
2. Liquid tension = Right tension – Left tension
62012 IEEE Symposium on Industrial Electronics and Applications
M1M0
kl kr
f1f2
f0
M2
ARM
2r�1r�
r r
lf rf
object
glue
motor inertia is analogous to mass, etc:
Motor belt Mass flow
Inertia Mass
Torque Disturbance
Rotation Line Speed
Tension Force
Power Matters.
2. Nonlinear map: current -> speed -> tension
72012 IEEE Symposium on Industrial Electronics and Applications
IsL IeL
“s” means starting, “e” means ending, “L” means left, “R” means right;
we want the liquid tension or tension difference between left belt and
right belt remains constant, from starting to the ending position during
the glue brushing process.
fsL = feL
current arm tension
rotation speed
- fsR - feR
LEFT
RIGHT
IsR IeR
Power Matters.
2. Nonlinear liquid boundary condition
82012 IEEE Symposium on Industrial Electronics and Applications
s
“s” means starting, “e” means ending, “o” means object;
we want the liquid tension has minimum variation,
it means optimum brushing speed.
feo
position
linear speed
fso
Object in between
eliquid tension
Power Matters.
3. Neurons to learn nonlinear relationships
92012 IEEE Symposium on Industrial Electronics and Applications
1�z )(ˆ kF
)2(1 �kis�1�z
1�z
1�z
)1(1 �kis�
)(1 kis�)1(1 �kis�
)2(2 �ki s�1�z
1�z
)2(2 �ki s�1�z
1�z
)(1 kis�
)2(1 �ki s�)(2 kis�
)1(2 �kis�
)(2 kis�)1(2 �ki s�
Tension
Control
Model
NN
Neural network s-Sigmoid function
Captures nonlinear from both liquid
and motor:
浦东
浦东
Power Matters.
4. PSO is smooth and quick!
1
02012 IEEE Symposium on Industrial Electronics and Applications
Merits
Optimized
STD of
Epoch
Average
Epoch
STD of Target
Value
BP 89 71 0.04
PSO 40 41 0.02
Power Matters.
5. Conclusion and future work
� The equation of liquid tension difference for robot motor driven arm system controlled by current is obtained.
� A neural network trained by PSO identifies the liquid tension from dual-motor system using laboratory data.
� The simulation shows the approach is a viable solution for high volume and precise tooling robot arms.
� New algorithm doubles the speed and the smoothness at the same cost.
� The disadvantage is that the amount of statistics needed could be huge to reduce mechanical vibration noise.
1
12012 IEEE Symposium on Industrial Electronics and Applications
12Power Matters.
Thank you!
� Questions are welcome:
1
22012 IEEE Symposium on Industrial Electronics and Applications