Manufacturing Process Planning Application

using Artificial Neural Networks

N.Ben Yahia:

F. Fnaiechl: Senior member IEEE, S . Abid' ;Sfudent member IEEE and B.Hadj Sassi *

( I ) : Ecole Superieure des Sciences et Technique de Tunis : 5 , AVTaha Husein 1008 Tunis c a d v a h i a 2 0 0 0 & ~
(FarhatFnaiech, Sab.Abid )@esstt.rnu.tn TCI :(216),71496066 Fax (216) 71 391 166
(2 ): Centre National de Formation des Formateurs et d'ingbnierie de Formation : 6, Rue de France Rades 2040 Tunis :
directeureenemhtIift2 Emai1.ati.Q
Topic number [901]

ABSTRACT
In this paper a Feedforward Neural Networks based
intelligent system for Computer Aided Process Planning
(CAPP) methodology is developed. This methodology
suggests the sequence of manufacturing operations to be
used, based on the attributes of a feature of the component.
By integrating this methodology Hith Computer Aided
Design (CAD), process planning can be generated, and
tested, which helps in realizing concurrent engineering.
Key words: Process Planning, CAD, Computer Aided
Manufacturing (CAM). CAPP, Neural Nehvorks.
manufacturingfeatures.

1. INTRODUCTION
Computer Aided Design and Manufacturing (CADICAM)
technologies have evolved over the last decade to automate
and integrate various stages of the product cycle to
improve the productivity of the conventional design and
manufacturing activities.
CAD relates to the activities like geometric/solid product
modelling, visualization, design analysis and drafting,
while CAM includes Computer Numerical Control (CNC)
equipements, robotics and flexible manufacturing system.
Since the CAD and CAM systems have been developed
independently, each one has its own method of
information representation which makes it difficult to
integrate them with each other [I].
Furthermore Process Planning is the critical bridge
between design and manufacturing [2]. Design information
can be translated into manufacturing language only
through process planning.
In this paper, we present an intelligent CAPP methodology
based on trained Neural Networks which help the designer
to choose the sequence of manufacturing operations based
on the attribute of feature components. In section 2, we
shall give a brief description of the problem formulation.
While in section 3 and 4, we present the development of
the CAD Model and the Neural Networks application. In
section 5; we give a case study and in section 6, we give a
final conclusion.

2. PROBLEM FORMULATION

translate the product engineering requirements into

detailed manufacturing [3]. [41.
Examining the recent developments in CAPP, it can be
observed that is now in a strategic position. Key research
issues of CAPP must include:
- Development of methodologies for complete
product definition that captures the design,
functional, and manufacturing aspects of the
part,
- Automation of process planning knowledge
acquisition with
artificial intelligence
pradigms,
- Development of intelligent interface between
CAD and CAM.
This paper reports the development of a CAPP system
which provides an integrated environment for part
synthesis and automated process planning to manufacture
prismatic features. This system is based on Neural
Networks application.

3. MODEL FOR DESIGN BY FEATURES

3.1. knowledge base for Process

planniog
The first part of thk paper presents a new methodology for
the description of the profile aimed job. This method is
calledhb kindin dynamicstudy, known as ETED [I].
The analysis of a knowledge planner consists to extract the
knowledge of the expert that it has acquired in general by
an initial training of professional experience displayed on
several years, perhaps of tens years . This task of
knowledge expertes extraction is the spring ofan analyst.
This analyst is capable to identify the different real
problem types well often complex that pose. It has to
describe in find methods of representation of real closest
environment.
In our case of study we are placed in the context of a
workshop producing in unit or very small batch size of
various pieces that are composed by several machining
features.
Flowing this, we have constituted a sample of persons to
inquire, that represent several activities :
the use of mechanical factories including the
operators on numerical machine-tool, the planner and
the expert in milling,
- mchanical teachers for, technological institutes,
professional training centers and study offices wich
are specialized in CAD/CAM solutions.
~

The traditional method of process planning is centered

around the 'process planner". The process planner
typically needs diverse applied knowledge in both design
and manufacturing and uses this knowledge, past
experience, handbooks and/or various databases to

0 Zoo2 IEEE SMC

WALK5

3.2. Feature Based Modeller

The results of Process Planing Knowledge base concerns
the creation and manipulation of manufacturing feartures
for each free types or interaction between them The
feature modeller is based on standard STEP (AP224) 141.
In this work, features are considered to he regions of a part
having some manufacturing significance. The particular
manufacturing context consider a mchining feature. Each
feature is formed by a set of attributes as follows:
-Tbe Attributed feature Pocket :
It is a negative volume, that intersects the preaxesting
shape at one end. It has a bottom face wich is connected to
all wall faces by concave edges (Figure I).
When used for detailed planning, most existing coding
schemes must carry enough informations with regard to
machined shapes, dimension, tolerence, surface finish and
detailed description.

-"he Attributed feature Gmove:

It is a negative volume too, that is depression with a
bottom face and two or more wall faces whose connection
chain is discontinuous and intersecting the object at four
faces (Figure 2).

.I

Figure 2 : Feature Groove

-The Attributed feature Step :

Is depression that usually intersects the object at four ends
(Figure 3).

14SI urface state

I
I

Ra3
HB3
Figure 1 : Feature Pocket

lShc atter (Hardness)

Figure 3 : Feature Step

The used model here is based on the design byfpntures
applied for grooves, pockets, steps and essentially their
intersections : steplpocket, steplstep, pocketlpocket,
steplgroove, pocketlgroove and groovelgroove. Each
feature is coded and considered in the part of this paper as
the Neuml Networks inputs.
Figure 4, presents three cases of PocketlGroove variation
features that have an influence for process choice.

GROUVE/POCKET
Case

:I

Case. 3

Case 2

5.IIH.Y

."

1.6Llr.S
r. 111,.

Figure 4 :Features Influenced dimensions for groovelpocket

4. NEURAL NETWORKS APPLICATION

4.1. Neural Network architectures
In this section, we present a method of process
organization of manufacturing by using a Neural Networks
system where the learning mechanism is
the
backpropagation algorithm [SI,[6].
The basic Neural Networks architecture includes an input
layer of neurons that receive the binary or continous
valued input signals, a number of hidden (intermediate)
layers that are highly interconnected and output layer with
one ormore neurons [7], [SI,[9].
The global system for process planing is composed by two
and " 2 ) with a parallel
related Neural Networks ("I
shucture (Figure 5).
- NNl is capable to select a machinig operations :
A machinig operation is attributed to a previously
recognised manuafcturing features. The choice is
made according to specific technological
knowledge and is based on rules following
geometric and tolerancing considerations, ,. ..
- NN2 capable to select machining toolsto be used:
The same logic is applied to each operation and
feature, and the appropriate machining tools are
selected. Outputs of the Neural Networks system
constitute- a matrix, which is represented by
binary form

Figure 5 : Global Neural Networks architecture

In this paper, we focus only on NNI while NN2 will be

presented in an other paper.
NNI architecture is composed by six elementary Neural
Networks ("11,
"12,
"13,
"14,
"15,
"16)
given in Figure 6.

.--.
.

GA/a

eoded

NNll

. .. .-

.: .: :.

-1
Figure 6 :Architecture for NNI

Table 1: Inputs for NNI 1

In figure 7, we present one exemple for local problem.

Here NNll can resolve dependently features for
groovelpocket. Inputs ofNNl1 are attributs of features and
outputs are local manufacturing process planning.

Proposed constraints linked to geometrical tolerances are

two types: constraints associated to free features and
constraints of relationship between pairs of features

Input layer

output layer

(fip:B).

Geometric tolerance

F c a m Ei

Relations
Ei/Ei+l

Relatioashio

Position

Tolerance
Dimensions

Location

,+F,

Grome~id
TD~C~OECI

Type i value i Reference

Figure 7: Architecture for NNI 1

Inputs of N N l l architecture are represents by a matrix
with 5x9 dimension wich value from 1.5 to 1000.
In table 1 we represent attributs of related groovelpocket
features with dimension and coded constraints.

Figure 8 :Features relationship

Outputs constitute also a matrix of 5x9 dimension which
represented by binary form. This matrix describes
proposed solutions given from expen planner (figure 9).
- Si=l: Process i is possible in this case E I E 2
- Si+ Process i is difficult or impossible in
this case EILE2.

Figure 9 :Proposal of manufacturing process for groovefpocket features.

4.2. Results for Training Neural Networks

"11.
The table2 shows the performance of the "11
system
depending on feature dimensions from lOmm to 400mm.
we show also the training time according to the number
of epochs (figure IO).

1 Dimensions1

Fnnrh

M9F

Trarnine

0.116 I Sigmol de
Table 2 :Performance of NNI 1

.le3

Ea

25

10

?p

30

40

I R~monsc1

II s

YI

Figure. 10: Training N N I 1 system evolution

(for groovelpocket)

5. CASE STUDY
For the present system, Neural Netwodcs training data is
developed keeping in view the principales of design for
manufacturing, for each manufacturing operation. Based
on the particular values (Inputs and Outputs) of a feature's
attributes. the Neural Networks system can identifv the
sequence of operations necessary to produce the feature.
The classification of process according to precision is
performed using tolerance and surface finish values that
are normally achieved for the process.
Thus, the Neural Networks model has to be trained to
differentiate between this situations. This differentiation is
built into the second stage of the neural network. This
second stage is able to propose a suitable tool for machinig
feature.
In the case of experimentation study, we have chosen a
prismatic part of piece that formed ,by three families. .
.-eai+ues;i-goove/pocket, . steplstep- and groovelstep
(figurell)

-.

Figure 11 : Experimentation case for prismatic piece

Inputs

outputs

1 1 0 0 0 0 0 0 0 0 ( s 1 , s 2.....)
0 0 1 0 0 0 1 0 0 0 (S3,. ..S7,. ...)
0 0 1 0 0 0 1 0 0 0 ( S 3,...S7,....)
0 0 0 0 1 1 0 0 0 0 ( . . S 5 , S 6 ,....)
0 1 0 0 1 1 0 0 0 O(...S2,...S5,S6, ...)

Elm2 : 0.300 0.160 0.187 04.330 OOOOOOlO

E3E5 : 0.120 0.220 0.100 0.340 0.200 0.340 010
E4E5::0.120 0.220 0.100 0.340 0.2000.340010
E5/E7:0.320 0.800 0.340 0.130 0.106 0.615 001
E7B8 : 0.640 0.100 0.300 0.800 0.375 001

6. CONCLUSION
The avantage of this new method of process planning is
the ability to generate an optimal sequence which is
difficult iu a real manufacturing environment. The
sequence generated is near optimal when it is successful in
minimising the number of set-ups and minimising the
number ofmanufacturing tool changes.
The development of powerful and flexible CAPP systems
will be useful to the manufachning engineering specialists
working in concurrent engineering teams. In the present
work, use of Artificial Neural Networks approach has
enabled to develop a very flexible intelligent CAPP
methodology that can be easily trained to handle new types
of components. A components with prismatic features has
been analyzed by this procedure. The methodology
presented here can substantially reduce the time taken to
generate process plans and the results will be of consistent
quality.

[3] D.Chandra, Reddy and K. Ghosb

Inteligent Process Planning for coucurent engineering
p
1998.
[4] M. Radouani, A. Saka and M. Carrard
Le Dimensionnemeut et le toltrancement vectoriels : outil
de C A O Etude comparative entre le dimensionnement et
le tolerancement geomttriques et vectoriels Analyse des

CP199 Tanger Maroc-2526 Nov. 1999

Cil
S. Abid. F. Fnaiech and M. Naiim
._
A Fast FeedForward Training Algorithm Using A
Modified Form of Standard Back-Propagation Algorithm
IEEE Transactions on Neural Networks, pp. 424-430,
March 2001.
[qS. Abid, F. Fnaiecb and M. Najim
Evaluation of the feed forward neural network covariance
mahix e m r
P m . IEEE ICASP 2000, June 5-9 2000, Istanbul, Turkey.
[7] K. Nezis and G. Vosniakon
Recognizing 2 R)shape features using a neural network
7. REFERENCES
and heuristics
[I] N. Ben Yabia, F. Fnaiecb and B. Hadj Sassi.
Computer Aided Design, Vol29 No 7 pp. 523-539,1997,
Application des rdseaux de neurones multicouches pour
Elsevier Science Ltd.
Ielaboration de phases dusinage
[SI J. Mou
Conftrence International Francophonie dAutomatique
A Method of using neural networks and inverse kinematics
CIFA pp 640 645,5-7 juillet ,2000.
-~.formacMn_etoolserror estimation adcorrection
[21 A. LEFEBRE, L SABOIJIUN,J. GOGU
Journal of manufacturing science and engineering Vol 119
and J. RENAUD
May IY97
Design Process Modeling of Process Planning for flexible
Thomas
I.
[9]
Lines Based on Conceptual Graphs and Design Rules.
Gentration de sCquences daltsage par approche neuronale
Applied to Cylinder Head Machining
IDMME 2000, 3rd International Conference on integrated
IDMME 2000, 3rd International Conference on htegrated
Design and Manifacturing for Mecanical ingineering,
Design and Manifacturing for Mecanical iugineering,
Canadien Society for Mechanical Engineering 2000.
Canadien Society for Mechanical Engineering 2000.

~~

~~~