Professional Documents
Culture Documents
discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/251885796
CITATIONS READS
7 43
3 authors, including:
Some of the authors of this publication are also working on these related projects:
Complex Systems Health Management (DX,PHM,CM) Using Artificial Intelligence View project
All content following this page was uploaded by Sina Tafazoli on 26 September 2017.
Abstract-- Neural system that controls movement and The next level is local feedback loops that perform basic
posture is a highly nonlinear complex system. Its reflex actions in spinal cord. There are several command
adaptability and easy accommodation to changes in centers in the brain that form the upper levels of
environment and task specifications make it an ideal hierarchy. They include stem, cerebellum, motor cortex
system. In this paper, the muscle control system from sensory cortex and many other that send intention for
spinal cord to muscle displacement has been studied. action to supraspinal controller. This hierarchy has shown
At first, a detailed nonlinear model is simulated in in Fig.1.
Simulink based on an already developed work. Then,
three system identification techniques are examined to
estimate the behavior of this complex system. The first
one is based on popular linear ARX model. Then, the
system is modeled by NARX neural network
(Nonlinear Autoregressive Network with Exogenous
Inputs) which has a powerful structural network in
modeling dynamical systems. Finally, a new method of
modeling using combined NARX and ARX structure
is proposed in which ARX gets the linear part of the
system and the NARX picks up the nonlinearities. The
simulation results demonstrate the superiority of the
latter method with respect to other examined
approaches.
I-INTRODUCTION
Effective treating the disease or injures affecting Fig..1. Structural representation of neuromuscular system [1].
posture and movement control is possible through better
knowledge of functional structure of controlling central In this work, first, we modeled a neuro-mascular
nervous system. Seeing how nervous system control structure from motoneuron to single muscle with details
complex movements and adapts to changing tasks and based on modeled proposed in He el al and then take this
environmental perturbations will help engineers inspire model as true model. Identification of ARX model and
controlling and modeling structures. A computational neural networks model and their combined NARX-ARX
approach based in simulation, signal processing and structure has been investigated mainly in this work.
estimation will help understanding fundamental principles
of this complex system. Computational models have been II- INTERACTIONS OF SENSORY FEEDBACK
proposed and implemented to this end. These model PATHWAYS IN THE SPINAL CORD
developments simultaneously helped understanding
function of CNS for motor control and contributed in motoneuron is command center that activates the
advance of new algorithms and controllers in robotics. single muscle in the spinal cord. It is also called extrafusal
Its the hierarchical structure that enables neural fiber. Several feedback loops (called reflexes) regulate
control system to achieve desired adaptation and output of the motoneuron. In addition, there are some
robustness. Musculoskeletal apparatus with sensory inputs from supra spinal centers and spinal pattern
organs with reflects the status of the system is at the generators which regulate its output. These loops are
bottom of the hierarchy [1]. negative feedback from Renshaw cells (R in Fig.2.),
feedback from spindles called recurrent inhibition (RI)
Sina Tafazoli is with the Tehran South Azad University, Tehran, Iran (intrafusal and la in Fig.2.), force feedback from Golgi
(corresponding author phone: 098-21-66920317; e-mail:
tafazolisina1@gmail.com). tendon organs (GTO) and feedback from skin sensors
Karim Salahshoor is with the Department of Instrumentation and [1],[2].
Automation, Petroleum University of Technology, Tehran, Iran (email:
salahshoor@put.ac.ir)
Mohammad Bagher Menhaj is with the Amirkabir University of
Technology, Tehran, Iran ( email: tmenhaj@ieee.org )
IV. SIMULATION EXPERIMENTS
6
Fig. 2. Detailed biological model represented as neuro- musculoskeletal
dynamics [1]. 5
4
Being able to change parameters and structures during
simulation is advantage of modeling approach. However, 3
10
Fig.3. Detailed and frequently used model structure for spinal control of 0
0 1000 2000 3000 4000 5000 6000 7000 8000 9000
movement [1].
Fig.6. Motoneuron firing rate
III. MODEL BUILDING
0.35
The Simulink is a useful platform for modeling
dynamical systems. We used Simulink Version 2006a to 0.3
as in Fig.4. 0.15
0.1
0.05
0
0 1000 2000 3000 4000 5000 6000 7000 8000 9000
90
The actual system is estimated with an ARX model.
80
The excitation input signal has been shown in Fig. 13.
70 7
60
6
50
40 5
30
4
20
10 3
0
0 1000 2000 3000 4000 5000 6000 7000 8000 9000 2
0
0 0.5 1 1.5 2 2.5
20 4
x 10
18 Fig.13. Identification input
16
4
16
2
14
0
0 1000 2000 3000 4000 5000 6000 7000 8000 9000
12
-0.2
10
-0.4
-0.6
5
-0.8
-1
0 1000 2000 3000 4000 5000 6000
0
0.8
-0.8
y (t ) = f ( y (t 1), y (t 2),... y (t n y ), -1
0 1000 2000 3000 4000 5000 6000
signal. 0.6
0.4
network 0
A. Introduction 0.6
y sp (t ) = spT (t ) -0.8
-1
0 1000 2000 3000 4000 5000 6000 7000 8000
where
Fig. 20.Identification input
sp = [1 2 .... n ] . T
1
0.8
combination of the two methods can result in a fit 0.2
that is better than the original linear model. In our 0
case the linear part has ARX model structure and
-0.2
nonlinear part has NARX based neural network
-0.4
structure.Thenonlinearpartpicksupanyadditional
nonlinearitiesandnoiseintheresiduals[4]. -0.6
-0.8
sp (t | sp ) = y (t ) y sp (t | sp ) (5) -1
0 1000 2000 3000 4000 5000 6000 7000 8000
Fig. 21. Thick: ARX output, Thin: Actual output
The architecture of this combination is presented in
Fig.26. Now, we add the NARX network to the model and let
it to pick up the nonlinearities and residues. At first step
we trained the NARX network with the system input and
the system error as target data. This process is shown in
Fig. 22.
1
0.8
0.6
0.4
Fig.19. Architecture of semi-physical models combined with 0.2
neural networks
0
linear part and the NARX - based neural network for -0.6
neural network part. -0.8
Estimation of the system with ARX 4 4 1 and error
-1
from actual data is shown in Fig. 21 (all data has been 0 1000 2000 3000 4000 5000 6000 7000 8000
0.5
-0.5
-1
-1.5
0 1000 2000 3000 4000 5000 6000 7000 8000
VIII.CONCLUSION
REFERENCES
[1] Jiping He, Maltenfort, M.G., Qingjun Wang, Hamm, Learning
from biological systems: modeling neural control, IEEE Control
Systems Magazine, Volume 21, Issue 4, Aug 2001 Pages: 55 69
[2] J.M. Nightingale, E.M Sedgewick, Control of movement via
skeletal muscles, IEE Control Engineering Series 11, Biological
systems, modeling and control, 1979, pages 84 - 108
[3] Lennart Ljung, System Identification - Theory For the User,
2nd ed, PTR Prentice Hall, Upper Saddle River, N.J., 1999
[4] Braun, M.W., Rivera, D.E., Stenman, A., Foslien, W, Comparison
of global nonlinear models and model-on-demand estimation
applied to identification of aRTP wafer reactor, Proceedings of
the 38th IEEE Conference on Decision and Control 1999, Volume
4, Page(s):3950 - 3955