34 INTERNATIONAL CONFERENCE ON PRODUCTION ENGINEERING

randombroadAI and Robotics

Oct 15, 2013 (3 years and 7 months ago)

138 views


34
th

INTERNATIONAL CONFERENCE ON
PRODUCTION
ENGINEERING

29.
-

30. September 2011,
Niš, Serbia

University of Niš,
Faculty of Mechanical Engineering



TOWARD
S

A CONCEPTUAL DESIGN OF AN INTELLIGENT MATERIAL
TRANSPORT BASED ON MACHINE LEARNING AND AXIOMATIC DESIGN THEORY

Milica PETROVIĆ
1
, Zoran MILJKOVIĆ
1
, Bojan BABIĆ
1
, Najdan VUKOVIĆ
2
,
Nebojša Č
OVIĆ
3

1
University of Belgrade

=
䙡捵汴y=of=䵥捨an楣慬ibng
楮敥ringI=mrodu捴楯n=bngin敥ring=䑥p慲瑭en琬t䭲慬a楣攠䵡r楪攠NS=
NNNOM=B敬er慤攠3RI=o数ub汩挠of=卥pb楡
:
=
mmp整eov楣imasKbgK慣Krs

穭楬ikov楣]bgK慣Krs

bb慢楣im慳⹢gK慣Krs
=
=
2
University of Belgrade
-
Faculty of Mechanical Engineering, Innovation Center, Kraljice Marije 16 11120
Belgrade 35, Republic of Serbia:

nvukovic@mas.bg.ac.rs

3
Company FMP

d.o.o.
-

Belgrade, Lazarevački drum 6, 11030 Belgrade, Republic of Serbia
:
=
n敢ojs愮捯v楣ifmpK捯Krs
=

Abstract:

Reliable and efficient material transport is one of the basic requirements that affect
productivity

in sheet metal industry. This paper presents a methodology

for conceptual design of
intelligent material transport using mobile robot, based on axiomatic design theory, graph theory and
artificial intelligence. Developed control algorithm was implemented
and tested on the mobile robot
system Khepera

II

within the laboratory model of manufacturing environment. Matlab
©

software package
was used for
manufacturing

process simulation, implementation of search algorithms and neural network
training. Experimental

results clearly show that intelligent mobile robot can learn and predict optimal
material transport flows thanks to the use of artificial neural networks. Achieved positioning error of
mobile robot indicates that conceptual design approach can be used for

material transport and handling
tasks in intelligent manufacturing systems.


Keywords
: intelligent manufacturing systems, conceptual design, axiomatic design theory, neural
networks, mobile robot


1.

INTRODUCTION



For the last thirty years manufacture con
cepts have had
several redefinitions. In the eighties and nineties, the
concept of flexible manufacturing systems (FMS) was
introduced to develop a new family of products with
similar dimensions and constraints [1]. The
manufacturing enterprises of the 21
s
t

century are in an
environment where markets are frequently shifting, new
technologies are continuously emerging, and competition
is globally increasing. Rapid changes in pro
duct demand,
product design,

introduction of new products and
increasing global c
ompetition require manufacturing
systems to be highly flexible, adaptable and responsive
[1].

A methodology that includes the technological migration

[1]

from established flexible manufacturing systems
(FMS) to intelligent manufacturing system (IMS) is
pre
sented in this paper. For needs to
be addressed at the
design stage of new manufacturing system with
all
intelligent characteristics, this paper
would like to present

a methodology for conceptual design of manufacturing
systems using axiomatic design appro
ach.

Beside axiomatic design methodology, the

mentioned
requirements cannot be fulfilled without artificial
intelligence. According to the literature published by
CIRP and other manufacturing periodicals during the past
decade, nearly 34 modern manufacturin
g systems and
production modes have been proposed and 35
mathematical methods have been used for building
intelligent systems [2]. The wide application of these
intelligent mathematical methods or their combinations in
manufacturing will definitely enhance

the development of
manufacturing system
modelling

and provide the

new
solutions. Some of the

methods are
:

machine learning,
artificial neural networks,
heuristic search, and graph

theory, etc. Evolutionary computation (i.e. genetic
algorithms, genetic pro
gramming, evolutionary
programming, and evolutionary strategies) and artificial
neural network are the most widespread [3].
Intelligent

material
transport implies solving

path generation
problem
and
control movement of

an intelligent

agent

-

a
mobile

robot
. The graph algorithms are used to generate
path and artificial neural networks for prediction of
duration of manufacturing operations.
In

[
4]
different

graph
search algorithms
are
presented.



2.

AXIOMATIC DESIGN THEORY


Axiomatic

design

theory

is

an attempt

at synthesis of

the
basic

principles

of design

in

various

engineering

fields

and

in
all

phases

of design
.
This

design

methodology

is

based

on

identifying

customer needs

and their
transform
ation
into

correspondent

functional

requirements

in the physical

do
main. According to

[
5], going from one
domain to another is calle
d mapping and it happens in the
e
ach

design
phase
: conceptual
,
product and

process
design phase, respectively
.
Further
m
ore,

the design
process

is done

through the

iterative

mapping

between

th
e
functional requirements

(
FR
s
)
in

the functional

domain
,

and

the design

parameters

(
DP
s
)
in the

physical

domain
,
for

each

hierarchical

level (Fig.1)
.






Fig.1.
Concept of domain, mapping and a
xio
matic decomposition


In mathematical terms, t
he relationship between the FRs
and DPs is expressed as

[5]
:


{
FR
}
=
|
A
|


{
DP
}

(
1)


where {
FR
}
denotes the functional requirement vector,
{
DP
}

denotes the design parameter vector, and
|
A
|
denotes the
design matrix that characterizes the design
process. The structure of the matrix
|
A
|

defines the type
of
design being considered and for the three hierarchical
levels particular design matrices
|
A
|
are presented in the
Table 1. It can be concluded that
|
A
|

matrix

in the second
hierarchical level

is

triangular and for that reason
we can
change some
DP
s to s
et some other FRs without affecting
the rest of FRs [5].

Such a design is called a decoupled
design.
In the third hierarchical level
|
A
|
matrix is
diagonal

and
each of the FRs can be
satisfied
independently by means of one DP. Such a design is
called an uncoupled design.



Table
1. List

of the functional requirements,
correspondent design parameters

and correspondent
|
A
|

matrices



X

Impact

0

No impact

DP1
: Mobile robot







DP1
1: Odometry motion

model




DP12: Path planning
module




DP13:
Manufacturing

process


simulation




DP1
4:
N
eural
n
etwor
ks






DP111:
Sensory information

from


encoders




DP112:
Sensory

information

from

the



camera





DP121:
Pat
h planning a
lgorithms

DP122:
Control a
lgorithms

DP1
41: Parameters for neural



networks

training

FR1: Intelligent material transport

X











FR11:
Determining m
obile robot position and orientation


X

X

X

X







FR12: Path plan
ning


0

X

X

X







FR13: Prediction of
manufacturing

process parameters


0

0

X

X







FR14: Machine learning of material transport flows


0

0

0

X







FR111:
Determining p
arameters in motion model






X

0

0

0

0


FR112:
Determining p
osition
and orientation of the



characteristic objects in the environment






0

X

0

0

0


FR121: Generating path nodes






0

0

X

0

0


FR122: Path following






0

0

0

X

0


FR141:
Getting expected p
erformance of IMS






0

0

0

0

X












F
R
1



F
R
11



F
R
1
2



F
R
1
3



F
R
1
4



F
R
1
11



F
R
1
12



F
R
1
21



F
R
1
22



F
R
1
41



DP112



DP121



DP111



DP141



DP122



DP11



DP12



DP14



DP13



DP1

Functional

Requirements

{FRs}



Customer

Attributes

{CAs}



Design Parameters

{DPs}


Customer

Domain

Functional

Domain

Physical Domain


Needs

specification


3.
MOBILE ROBOT IN
A

MANUFACTURING

ENVIROMENT


To explain mobile robot motion

and actions
in
manufacturing environment,
five

modules are developed.


3.1.

Motion model


The position of

the mobile

robot

is determined by the

system state

vector

x
t

=

(
x,
y
,
θ
), where
x

and y

are

the
components

that

define the

position

vector
,
and

θ

is

the
angle

orientation
.
Mathematical formulation

for
mobile
robot

odometry is given by

(2
):



(
2)


where

x'
,
y
'
and

θ'
are the
components

of
the
state vector

at

time

t
',
x
,
y

and

θ

components

at time t
;
Δs

the
incremental path lengths [6]
.


3.2.

Material

flow
analysis


Material transport analysis in manufacturing
environment

was recognized as the first task in a path planning
module. F
irst of all, flow line layout design is adopted.
After that,
the
data

about machines, parts and time
duration of operations should be gathered and

analyzed.
Table 2

presents a list of machines, and Table 3 presents

a

list of parts.


Table
2
. List
of machin
es

in

manufacturing plant


Machine

Machine type

M#1

Shearing machine

M#2

CNC punch press for punching
and blanking

M#3

Hydraulic punch press

M#4

Punch press for punching and
blanking

M#5 and M#6

Pillar drill (bench drill)

M#7

Circular saw

M#8

Whetti
ng machine

M#9

Line for machining parts made
of cooper



Table
3
. List
of representative parts

in

manufacturing
plant


Part

Description

P#1

Transport fuse

P#2

Mainbusbar support

P#3

Support d800

P#4

Busbar 2 L1


After defining number of parts and ma
chines, we need to
define quantitative
relations

between them. In general,
this dependence can be presented with matrix M
DM,
which
is written using matrix M (matrix of machines) and D
(matrix of parts) [7].



(
3)


If we need time dependence between machines and parts,
we put the time duration of machine operation to
a
correspondent machine instead of parameter p
ij
. At the
end, using graph theory, we define matrix of distances
between machines (R) [6].



3.3.

Path planning
algorithms



Three

algorithms

are
developed and implemented for the

mobile

robot

path

planning

task
. The first one is A*
search algorithm,
that

is

used

for

finding

the shortest

path

between

start and

goal

points. It combines
Dijkstra
algorith
m
and
bread
-
first

search algorithms. Using M
DM

matrix, the second algorithm
determinates sequence of
machines

for each representative part and chooses
machine the robot should visit,
according to a minimal
distance

criteria. Finally, the third algorithm is

used for
determining the

order of

machines

in accordance to
manufacturing

process.

This

algorithm

generates
characteristic

time

parameters

of the manufacturing
process

(
the duration of
the operation

on

the machine
)
and

time

parameters

related

to
part
tran
sport

to the

machine

(time needed for mobile robot to travel between

machines
).


3.4.

Prediction of manufacturing process
parameters


It is known

that

engineering

processes

generally do not
have

deterministic

nature.

The
processes

that are
important

for

the ma
terial

transport

task

in terms of

duration

are the

machining

process

and the

process of

robot

movement

between

the defined

nodes

(
machines).
Considering the fact that these processes have stochastic

nature
,
we can conclude

that

nominal time duration of
ope
rations
,
as well as

time of

transport

from one node

to
another
,

are different for each part
. For that reason,
uniform distribution is chosen to model stochasticity of
the
nominal time duration.


3.5.

Neural Networks for prediction of duration
of manufacturing
operations


Implementation of the neural networks (NN) to model
various problems in production engineering goes back to
the 20
th

century. According to [8], there are three basic
categories of their use: classification, prediction and
functional approximati
on. Prediction of the next node
(machine) in the path, where robot needs to go and deliver
the part, is based on past values
of
the
system state (in this
case the time parameters of the process and
the
time of
robot movement between the machines) and the

current
values
of
the
system state (the node where the robot is
currently located).
For NN

training the

Matlab

Neural

Network

Toolbox is

used
,
with

supervised
learning
algorithm
(
Levenberg
-
Marquardt)

[
8]
and

the sigmoid

activation

function
.


4.

EXPERIMENT
AL RE
S
ULTS


The experimental model of manufacturing environment is
static and positions of machines are
a priori

known.
Experimental model and the
Khepera

II

mobile robot are
shown
in Fig. 2
. The first goal is test accuracy of path
following. During track
ing the trajectory, the robot has to
deliver part to machines, according to manufacturing
process, defined by matrix M
DM.

Coordinates of start

and
goal point is known.
While executing the transport task
,

the robot optimizes the path between the m
achines us
ing
A* algorithm

[
6]
.

The mean position errors during the
first experiment in x and y directions are Δx=0.5598 [cm]
and Δy=1.4624 [cm].



Fig.2. Mobile robot motion in
laboratory model of
manufacturing environment


The next experiment is conducted i
n same conditions,
but

the coordinates of

the goal

point

are

not known

at the

beginning. This parameter depends on

the

time robot
needes to travel from one machine to another. When

the

robot

finishes
transport

of
the last

representative part

to
machine

for

the

first

operation
,
its current pose

is passed
to
NN
.
Based on this information,

NN

predict

the nearest

machine where manufacturing

operations are completed

and

generate

information

about the

future

robot actions.



5.

CONCLUSION


This paper presents a meth
od for conceptual design of
mobile robot material transport in intelligent
manufacturing system. Intelligent mobile robot, with
a
priori

known static obstacles in the environment, has the
ability to generate an optimal motion path in accordance
with the re
quirements of the manufacturing process and
priority servicing of machine tools. Mobile robot learns
the optimal transport routes and sequ
ence of manipulation
by using neural network

[
6]
. Neural network was
developed to predict the parameters of manufactur
ing
process and to learn characteristic time parameters of the
process. For the purposes of the simulation we used the
nominal time parameters (estimated using empirical data)
of the manufacturing process, and its stochastic nature is
modeled according to
uniform distribution

[
6]
. Search
algorithms and neural network models are developed in
Matlab environment and implemented on a
Khepera II

mobile robot.
Achieved positioning error of mobile robot
indicates that conceptual design approach

based
on
axiomatic design theory and neural networks

can be used
for material transport and handling tasks in intelligent
manufacturing systems.


ACKNOWLEDGMENT
S


This paper is part of the project:
An innovative,
ecologically based approach to implementation of
intelligent manufacturing systems for production of sheet
metal parts
, financed by the Ministry of Education and
Science of the Serbian Government, Grant
TR
-
35004
.


REFERENCES


[1]

REVILLA, J., CADENA, M. (2008)
Intelligent
Manufacturing Systems: a methodology

for
technological migration
, Proceedings of the

World
Congress on Engineering,
Vol II
, London U.K, pp.
1257
-
1262.

[2]

QIAO, B., ZHU, J.
(2000)
Agent
-
Based Intelligent
Manufacturing System for the 21st Century
,
International Forum for Graduates and Young
Resea
rches of EYPO, Hannover, The World
Exposition in German.

[3]

BREZOCNIK, M., BALIC, J., BREZOCNIK, Z.
(2003)
Emergence of Intelligence in Next
-
generation
Manufacturing Systems
, Journal

Robotics &
Computer
-
Integrated Manufacturing, Vol. 19, pp. 55

63.

[4]

SIEGWART
, R., NOURBAKHSH
, I. R., (2004)
Introduction to Autonomous Mobile Robots
, MIT
Press, Cambridge, Massachusetts.

[5]

SUH, N. P. (1990)

The

Principles of Design
, Oxford
University Press, New York.

[6]

PETROVI
Ć
, М., MILJKOVI
Ć
, Z., BABI
Ć
, B.,
Č
OVI
Ć
, N. (2011)
Artifici
al neural networks and
axiomatic design theory in conceptual design of
intelligent material transport
, Proceedings of the
37
th
JUPITER CONFERENCE with foreign participants,
Belgrade,
pp. 3.72
-
3.79, (in Serbian).

[7]

MILJKOVIĆ, Z., BABIĆ, B. (2005)
Machine
-
Par
t
Family Formation by Using ART
-
1 Simulator and
FLEXY
, FME Transactions, New Series, Vol.33,
No.3, pp. 157
-
162.

[8]

MILJKOVIĆ, Z., ALEKSENDRIĆ, D. (2009)
Artificial Neural Networks


Solved Examples With
Short Theory Background
, Faculty of Mechanical
Engineeri
ng, University of Belgrade, Belgrade, (in
Serbian).

x

y

S
tart

Goal