Artificial Intelligence in Engineering 12 (1998) 121-134

0 1997 Elsevier Science Limited

Printed in Great Britain. All rights reserved
ELSEVIER PII: SO954-1810(97)00011-3 0954-1810/98/$19.00

Artificial intelligence approaches to determination

of CNC machining parameters in manufacturing:
a review
Kyung Sam Park & Soung Hie Kim*
Graduate School of Management, Korea Advanced Institute of Science and Technology, 207-43 Choengryangri, Dongdaemun,
Seoul, Korea

(Received 1 March 1995; in revised version 13 May 1996; accepted 5 February 1997)

In Computer Numerical Control (CNC) machining, determining optimum or

appropriate cutting parameters can minimize machining errors such as tool
breakage, tool deflection and tool wear, thus yielding a high productivity or
minimum cost. There have been a number of attempts to determine the
machining parameters through off-line adjustment or on-line adaptive control.
These attempts use many different kinds of techniques: CAD-based approaches,
Operations Research approaches, and Artificial Intelligence (AI) approaches.
After describing an overview of these approaches, we will focus on reviewing
AI-based techniques for providing a better understanding of these techniques in
machining control. AI-based methods fall into three categories: knowledge-based
expert systems approach, neural networks approach and probabilistic inference
approach. In particular, recent research interests mainly tend to develop on-line
or real-time expert systems for adapting machining parameters. The use of AI
techniques would be valuable for the purpose. 0 1997 Elsevier Science Limited.

Key words: CNC machining, machining parameter, knowledge-based expert

system, neural network, influence diagram.

1 INTRODUCTION Consequently, the conservative cutting conditions

assuming a constant depth and width of cut do not
Computer Numerical Control (CNC) machining is perform high productivity. To overcome such a prob-
widely used in mold/die industries and airframe lem, the machining parameters should be adjusted
component manufacturing, because of its suitability according to the current in-process part geometry.
for high accuracy in machining complicated parts.’ In The objective of this paper is to review prior work on
the CNC machining, determining optimal cutting determining machining parameters in order to give a
conditions or parameters under the given machining better understanding to researchers and practitioners
situation is difficult in practice. Conventional way for in machining domain, since we have not found any
selecting these conditions such as cutting speed and publication on the survey. In particular, we will focus on
feedrate, has been based upon data from machining examining Artificial Intelligence (AI)-based methods. In
handbooks and/or on the experience and knowledge the next section, we describe an overview of related
on the part of programmer. The selected parameters, work that attempts to select an optimal machining
in most cases, are extremely conservative to protect parameters, and present AI-based approaches in the
excessive matching errors from tool failures such as tool subsequent sections.
deflection, wear, breakage, etc. As a result, the metal
removal rate is low because of the use of such
conservative machining parameters. 2 AN OVERVIEW
As frequently encountered in complex or free-formed
surface machining, the geometry of the part or work- Recently, there have been several attempts to deter-
piece prevents a constant depth and width of cut. mine the optimal machining parameters from off-line
adjustment or on-line adaptive control, thus the part
*Author to whom correspondence should be addressed. programmer does not have to spend time and effort to
127
128 Kyung Sam Park, Soung Hie Kim

calculate their optimal values. These attempts are The multiplicative model can be generated from
categorized as CAD-based approaches, Operations statistically planned machining tests (see Ref. 11). An
Research (OR) approaches, and AI approaches.‘3 advantage of this model is easily obtained even though
geometry of cutting tool is complicated such as ball-end
2.1 CAD-based approaches mill. Furthermore, it is reported the accuracy of the
models is quite good. ‘J~J’ The above models will play a
The off-line approach uses machining process models, role of constraints in optimizing machining parameters.
cutting force and tool wear models, based on a prior In addition, analytical process models for the prediction
knowledge gathered from off-line experiments. Based on of cutting force have been studied,7~8110V17 and their
the process models, cutting force and tool wear are application to on-line feedrate adjustment in end milling
calculated through computer machining simulation has been found in Ref. 6. However, a difficulty may exist
(CMS) using information on NC-code with initial in using such analytic models in practice because of their
machining parameters, tool shape and workpiece high computational complexity.
geometry. Using the results, an optimum machining
parameter for each tool motion is achieved by maxi- Computer machining simulation
mizing the metal removal rate (MRR) without violating The main objective of CMS for determining machining
machining constraints. parameters is to compute the maximum depth of cut (d)
The basic concept of optimizing machining param- and width of cut (w) for each tool motion from given
eters is that when the cutting force is too large at the part geometry, NC-code and tool configuration. Why
large depth and width of cut, either low feedrate or high compute the maximum point? The reason is machining
cutting speed, or both can be added to the NC-code. error from the tool failures is mostly occurred at the
However, note that too high cutting speed can not be maximum point. CMS of in-process workpiece can be
selected since the tool life is largely due to the cutting realized as a Boolean subtraction of the space occupied
speed.5 Most CAD-based approaches belong to the off- by the tool movement along the tool path from initial
line adjustment. Advantages of these methods are that part geometry. Hence, it is first needed to represent the
they are easy and effective in practical applications. part geometry for CMS.
Figure 1 shows a conceptual framework for simulation Solid modeling’8>‘912’or Z-buffer techniques22>23have
and optimization of machining. been used to model workpiece geometry for CMS. A
pape?’ has proposed a method of feedrate adjustment
Machining process models using a swept volume generation technique based on
The machining process models represent the rela- solid modeling. However, this method can calculate
tionships between the machining responses (i.e., cutting only average cutting force, thus it does not provide the
force and tool wear) and the machining conditions in a instantaneous cutting force that is necessary for
specific tool and workpiece. These models can be built estimating the tool failures. Z-buffer model is a
by prior knowledge obtained from field and laboratory form of discrete nonparametric representation in
experiments. An example of machining process model which the Z-values of the surface are given at grid
based on a multiplicative model is given by points on the XY-plane. More detailed description
on Z-buffer model and its application to control
Cutting Force (N), FC = ald2f a3dn4waS,
and monitoring of machining can be found in the
Tool Life (min.), TL = blvb2f b3db4wb5, literature.22’23

where V, f, d and w are, respectively, cutting speed

(mm/min), feedrate (mm/tooth), depth of cut (mm), Machining parameter optimization
and width of cut (mm); and oi and J3i are the model Based on the CMS and the machining process models,
parameters. feedrate (f ) and cutting speed (v) are determined in the

( /,+
Computer Machining Simulation chining Process Models I

Fig. 1. A framework for simulation and optimizaiton of CNC machining.

 Determination of CNC machining parameters 129

optimization module. For increasing the productivity, and unloading and tool changes to replace worn-out or
MRR has to be maximized while maintaining an damaged tools. Solving Model 2 is more complex
allowable load fluctuation on the cutting tool in spite because the model have multiple objectives and con-
of variations in depth of cut and width of cut. The MRR flicting between the objectives (a mathematical repre-
is expressed as MRR = kvfdw, where k = n/(rD), n is sentation and the solution method for Model 2 appear
the number of tooth, and D the diameter of the in Ref. 27).
tool.’ A mathematical model for such problem can be
formulated as follows: 2.3 Artificial intelligence approaches
Model 1: Maximize MRR = kvfdw
The on-line approach is an attempt to automatically
subject to v,in 5 v 5 vmax adapt and optimize the machining parameters based on
sensor information on machining responses in real time,
without CMS. Well-known sensor information is listed
FC I JG,,, as cutting force, tool wear, tool temperature and
acoustic emission. Note that the information of tem-
TLmin I TL 5 TL,,,
perature and acoustic emission can not be used in off-
HP 5 HPr,,,,, line methods using CMS. These information, however,
where HP represents the spindle horsepower can play very important role in machining control
(Nmm/min) as a constraint for the machine capacity, or adapting machining parameters, praticularly when
and I’min and Vmax, respectively, are minimum and occurring an abnormal machining due to unpredictable
maximum allowable values of V. HP is expressed based variables such as unknown material properties, tool
on the FC as HP = c - FC - v, where c is 0.041 as a conditions, etc.
constant.5 Once taking natural logarithms in Model 1, For on-line control, the following components or
it is converted into the standard linear programming techniques are required: (1) sensing devises, (2) repre-
(LP) form. Thus, the LP problem can be solved by senting the information from the sensor, and (3)
using a general algorithm referred to as the Simplex optimizing machining parameters. A description of
sensing devises is not presented in this paper. For the
method.24
description, refer to Refs 4, 16.
AI approaches offer a possible technique in order to
2.2 Operations research approaches handle the problems (2) and (3). One of the most
important factors for successive on-line control is the
Of course, the use of the above LP technique can be execution time with respect to machining control or
viewed as an OR approach. However, main research determining optimal machining parameters. Reaction to
interest of OR approaches is to minimize global machining conditions by tool wear, machine break-
machining cost by considering multiple criteria related downs and other failures must be carried out within
to machining, thus which problem is to solve a multiple seconds or milliseconds to guarantee the safety and
criteria optimization problem (for an overview of the reliability of the machining process. However, a
multiple criteria optimization problem, see Refs 25, simplistic adaptation of AI techniques to machining
26). These methods should be used for off-line control would be inadequate, because execution time of
adjustment because of the restriction of computational these systems are generally too long as compared with the
time. An advantage of these methods can provide a reaction time required for the machining control,
reference model, i.e., a general model because an particularly if the knowledge base becomes very complex.
exhaustive consideration on selecting machining There have been a number of studies on the
parameters is involved. A typical research is found in application of AI techniques to on-line control, which
Ref. 27. According to the research, the model without we categorize into knowledge-based expert systems
describing full mathematical form can be expressed as approach, neural networks approach and probabilistic
follows: inference approach. Each approach is described in the
Model 2: Maximize {MMR} next subsequent sections.

Minimize {surface roughness}

Minimize {machining cost}
3 KNOWLEDGE-BASED EXPERT SYSTEMS
subject to the constraints of Model 1. APPROACH
On surface roughness, there are two methods of its
measuring: root-to-crest roughness and roughness 3.1 An overview
average.2y3 The factors of measuring machining cost
per workpiece are cost of tool, cost of cutting, and costs Knowledge-Based Expert Systems (KBES) are intel-
associated with machine idle time, due to setup, loading ligent computer programs that capture the specific
130 Kyung Sam Park, Soung Hie Kim

knowledge of a particular domain and mimic the stop, and so on, it is necessary that an adaptive control
problem-solving strategies of human experts to provide algorithm that uses the recursive adaptive model and
recommendations?8-3o They represent a new problem- the constraint rules is developed. For example, the
solving paradigm that utilizes many techniques developed constraint rules can be expressed as shown in Fig. 3.
from AI research. The KBES can capture causal and In KBES, many techniques for knowledge representa-
inferential knowledge about machining processes to tion have been developed, for instance, production
provide expert-level recommendations during decision- rules, semantic nets, frames, etc. The type of knowledge
making processes and hence are valuable aids to representation that is appropriate in a given situation
machining operators who face increasingly complex tasks. depends on what sort of knowledge is being represented
With the KBES technique, machining control deci- and how it is to be applied. In time-critical machining
sions using the sensor information can be made to control applications, it is imperative the knowledge
maintain the machining parameters within critical representation scheme is efficient. Among the KBES
constraints. Strictly speaking, on-line control with the approaches to machining control, in Ref. 31 a frame-
KBES is an adaptive control of satisfying machining based scheme is used, and in Ref. 32 a production rule
constraints, simply stated ACC, rather than an adaptive representation is applied such as shown in the above
control with optimization, ACO. Whereas AC0 systems paragraph.
seek to adjust machining parameters in a direction that
optimize a predefined performance index, i.e., objective
such as MRR, the aim of ACC systems is that the 4 NEURAL NETWORKS APPROACH
machining parameters are adjusted to their maximum
possible values given the constraints of the machining 4.1 An overview
process.’ Recent research on machining control using
KBES techniques has been found in Refs 31, 32. In the Neural networks differ in various ways from con-
next subsection, we will describe a KBES framework for ventional expert systems to traditional computing. The
machining control and present an example of simple reasons are as follows. First, unlike traditional expert
production rules for the determination of machining systems where knowledge is made explicit, neural nets
parameters. generate their own knowledge by learning from domain
examples. This means that neural nets can easily
3.2 Structure of KBES for adaptive control make the knowledge base by learning, and they do not
require additional knowledge acquisition processes
A structure of the KBES approach for machining which require enormous time and efforts in the expert
control is shown in Fig. 2. It consists of three modules: systems. Supervised learning is achieved through the
a knowledge base, an inference engine, and a sensor data learning rule which adapts the connection weights of the
acquisition and processing module. The inference engine network in response to the inputs and the desired output
drives the system and interfaces with the knowledge pairs. Many other network learning rules have been
base and hence supplies advice to the user and an inverted also in Ref. 34.
explanation to justify the system’s line of reasoning. The Second, neural computing is both distributed and
knowledge base can provide near-optimal machining associative in knowledge representation.33 The dis-
control with experimental data. The methods for tributed and associative nature of neural net leads to
inference can be modeled as rules, e.g., IF (antecedent) a reasonable response even when presented with
THEN (consequence). incomplete or previously unseen input. In particular,
To achieve the near-optimal machining parameters multi-layer neural nets which register in their hidden
and machining control such as tool change, machining layers important features of the knowledge domain,

Knowledge Base

Fig. 2. A KBES structure for machining control.

 Determination of CNC machining parameters 131

feedforward neural net with one hidden layer is shown

Rulesfor Finding MachiningSituation: in Fig. 4. Each node or processing element (PE) in every
layer is fully connected to other PE in the proceeding
layer, and ever PE sums its weighted inputs and passes
through some kind of transfer function such as linear
or sigmoid functions. The learning parameters are the
Rule 1: IF (NOT (A&c,,?,,, ‘: Acstic c AcsticJ)
connection weights and the PE’s parameters, i.e.,
/* Acstic = Intensity of Acoustic etmss~?n from sensor */
threshold values.
THEN (Stop the machining operatmn AND Change tool If necessary) As a learning rule, Rumelhart and McClelland33 have
Rule 2: IF (NOT (Temp nil” < Temp - Temp,> ,,,”)) developed the generalized delta rule called back-
I* Temp : Temperature from sensor *I
propagation algorithm that is basically a gradient
method. This rule aims at minimizing the global error
THEN (Stop the machining operation AND Change tool if necessary)
of the system by adjusting the learning parameters. The
backpropagation algorithm does not always find global
Rule 5: IF ((FC[k] J FC[k]J AND (feedm” c feed))
minimum but may stop at a local minimum. However,
/* FC[k]-Cuttmg Force at axis k. k=x,y, from sensor. feed=Feedrate ‘1 in most cases, the system can usually be driven to the
THEN (feed = feed 0 01 until FC[k] L FC[kJ,J global minimum or to the desired accuracy with an
Rule 6: IF ((FC[k] -z FC[k],J AND (t‘eed I feed”,,,\) AND (speed speed,,,J)
appropriate choice of hidden PEs. The number of
THEN (speed = speed + 0 02 until FC[k] < FC’[k],,,,,b)
hidden PEs must be large enough to form a decision
region that is as complex as required by the given
Rule 7: IF ((FC[k] -’ F(‘[k],,,,\) AND (feed < feed,,,J)
problem, and on the other hand is small enough that the
THEN (feed = feed + 0 01 unti! FC[k] : FC[k]“J
generalization ability remains good.

Rule n: IF ((FC[k] :’ FC[k],J AND (feed 1 feed”,,,\) AND (speed ? speed,,>J) 4.3 Optimal control phase
THEN (Stop machinmg operatmn)
Assuming that a neural network has been trained by the
procedure mentioned previously, then the objective of
Fig. 3. An example of rules for adapting machining parameters. optimal control phase is to determine an appropriate
machining parameter that optimizes a performance
can use this hidden knowledge to generate non-trivial index, given machining constraints on the network
generalizations. outputs.
In machining domain, neural nets can possess abilities Let us consider a trained neural net with n input PEs
to learn from experience and to use the knowledge and m output PEs. Let ai and dip respectively, be actual
gathered during the learning process to optimize the and desired output of the ith output PE. For k of the m
machining control. Experience is represented by input- output nodes, the di represent the desired outputs of the
output data, where input variables are machining PEs, whereas for the remaining m - k PEs, the di is the
parameters such as feedrate and cutting speed and maximum allowable outputs of the PEs (e.g., cutting
output variables are signals from sensors such as cutting force and horsepower). Then, a performance index PI
force, tool wear, temperature and acoustic emission. The is defined by PI = w1 - ERR - w2. MRR, where MRR
aim of learning is to establish a generalized mapping represents metal removal rate (see Section 2.1)
between the input and output, where note that this
statement is a supervised learning. This section deals ERR = 5 (di - ai)2/2,
with a supervised learning approach (see Refs 35,36). For i=l
unsupervised learning approaches refer to Refs 37, 38.
In addition, there are several techniques for repre-
senting the input-output relationship: multiple output
regression (see Section 2.1), the group method of data layer

handling, and the neural network (see next subsection).

A study4 has reported that the neural network is the
most effective method for tool wear identification Hidden
layer
through a comparative analysis among the above
techniques. Furthermore, neural net knowledge-base
acquired during the learning phase can subsequently be
Input
used for determining optimal machining parameters. layer

4.2 Learning phase

Fig. 4. A neural net structure for representing machining
For building machining knowledge base, a two-layer knowledge.
132 Kyung Sam Park, Soung Hie Kim

and w1 and w2 are constants that represent the relative information. The nature of the influences is specified
importance of ERR and MRR. at the functional level and further quantified at the
Machining optimization problem is to find the n numerical level.
inputs, denoted by pi, that minimize PI subject to the From the discussion in the preceding two paragraphs,
following constraints: influence diagram is defined by an acyclic directed-graph
G = (N, A) with A c N x N: It contains three types of
Pmini I Pi I Pmaqr i= 1 ,...,n,
nodes in the node set N. The chance node, which is
di - ai > 0, i=k+l,...,m. circular shape, represents uncertain or certain states
(e.g., cutting force, tool wear, acoustic emission), the
The solution method of this constrained minimization
rectangular-shaped decision node (e.g., feedrate, cutting
problem can be found in Ref. 35.
speed) reveals a variables whose value is chosen by the
decision maker, and the diamond-shaped value node
(e.g., metal removal rate) represents the objective to be
5 PROBABILISTIC INFERENCE APPROACH
maximized in expectation by the decision analysis. It
should be noted that influence diagrams on the
Agogino et a1.43 have proposed an influence diagram
topological level do not need a mathematical or
as a framework for integrating machining operator’s
probabilistic basis to justify themselves. Their influences
expertise, first-principle knowledge and experimental
are justified by mathematical or probabilistic repre-
data for the wide range of sensors possible for in-process
sentation at the functional level. At the final level,
monitoring and control. The use of multiple sensors
numerical level, utilities of the decision maker, and
reduces the sensitivity of the system to any specific
probability distributions from prior information by
sensor’s drawbacks. The non-deterministic or prob-
experiments are assessed numerically for each node.
abilistic nature of the inference problem and noisy
Shown in Fig. 5 is a simple example of an influence
sensor data is handled by operations with Bayesian
diagram for machining optimization.
probability.
Once a complete influence diagram is generated, the
Influence diagram has been developed for represent-
diagram is manipulated and evaluated for determining
ing complex decision problems based on incomplete and
the optimal decision strategy. A direct solution proce-
uncertain information from a variety of sources.39~44
dure to automate influence diagrams has been proposed
Knowledge of the interrelationships between variables is in Refs 40-42. This algorithm consists of the value-
represented in a compact graphical and numerical preserving translations, node removal and arc reversals,
framework which identifies the critical variables and which correspond to the rollback procedure in deci-
explicitly reveals any conditional independence between
sion tree models.45 For more detailed description on
them. applied influence diagrams to machining monitoring
The knowledge representation using influence dia-
and control, see Ref. 43.
grams can be viewed from three hierarchical levels:
topological, functional and numerical level.39 At the
topological or relational level, the nodes in the diagram
represent the key variables in the system being modeled, 6 CONCLUDING REMARKS
and the arcs or arrows identifies conditional influences
or functional relations between the nodes. In the CNC This paper presented a survey of prior studies on
machining, examples of key variables are machining determining an optimal machining parameter and
parameters and machining responses from sensor machining control. We particularly focused on reviewing

Fig. 5. An influence diagram for determining machining parameters.

 Determination of CNC machining parameters 133

AI-based methods for on-line adaptive control, i.e.,

KBES, neural networks and probabilistic inference
approaches. These approaches are commonly based on 1. Balakrishnan, P. and Dervries, M. F., Sequential estima-
tion of machinability parameters for adaptive optimization
the sensor information of machining responses in real
of machinability data base systems. ASME Journal of
time, and based upon the prior knowledge from Engineering for Industry, 1985, 107, 159- 166.
machining experiments in field and laboratory, such 2. Bhattacharya, A. R., Faria-Gonzalez and Ham, 1. Y.,
as machining constraints, machinability data (i.e., Regression analysis for predicting surface finish and its
mechanical and material data of cutting tool and application in the determination of optimum machining
workpiece), etc. The prior knowledge can be elicited conditions. ASME Journal of Engineering for Industry,
1970,92,711-714.
from data of machining handbooks (e.g., see Ref. 14) 3. Chandiramani, K. L. and Cook, N. H., Investigation on
or the experience and knowledge of the part the nature of surface finish and its variation with cutting
programmer. speed. ASME Jounal of Engineering for Industry, 1964,86,
Observe that the difference that underlies AI 134-140.
approaches is only the style of knowledge base being 4. Chryssolouris, G. and Domroese, M., An empirical study
of strategies for integrating sensor information in
represented in the system and its reasoning (or inference) machining. Annals of CZRP, 1989, St?(l), 425-428.
for machining optimization and control. That is: The 5. DeGarmo, E. P., Black, J. T. and Kohser, R. A., Materials
KBES approach uses a knowledge base such as and Processes in Manufacturing. MacMillan Publishing
production rules (see Ref. 32) and frame-based tech- Co., New York, 1984, pp. 429-629.
niques (see Ref. 31), with an inference engine. The 6. Fusell, B. K. and Srinivasan, K., An investigation of the
end milling process under varying machining conditions.
neural networks approach uses a feedforward network ASME Journal of Engineering for Industry, 1989, 111,
model (see Refs 35, 36) based on supervised learning 27-36.
rules or self-organization network model (see Refs 37, 7. Kline, W. A., DeVor, R. E. and Lindberg, J. R., The
38) based on unsupervised or competitive learning prediction of cutting forces in end milling with application
rules, without an additional inference engine. The prob- to cornering cut. International Journal of Machine Tool
Design and Research, 1982, 22(l), 7-22.
abilistic approach uses an influence diagram model (see 8. Kline, W. A., DeVor, R. E. and Shareef, I. E., The
Ref. 43) with a probabilistic reasoning engine. prediction of surface accuracy in end milling. ASME
In building an intelligent machining control system, a Journal of Engineering for Zndustry, 1982, 104,272-278.
principle problem is how to represent knowledge in a 9. Koren, Y., Computer Control of Manufacturing Systems.
way that is suitable for a particular machining domain. McGraw-Hill Co., New York, 1983, pp. 193-220.
10. Martellotti, M. E., An analysis of the milling process.
There have not been a specified guideline for the Transactions of the American Society of Mechanical
choice of a knowledge representation scheme. The Engineers, 1941, 63, 677-700.
general or conceptual guidelines are efficiency and 11. Neter, J., Wasserman, W. and Kutner, M. H., Applied
sufficiency, where sufficiency means that the precision Linear Statistical Models: Regression, Analysis of Variance,
of the predictions we obtain meets our requirements, and Experimental Design. Irwin Inc., Homewood, Illinois,
1985.
and efficiency refers to their practical applicability. All 12. Park, K. S., Analysis of Cutting Force in Ball-End Milling
AI approaches mentioned in this paper may be sufficient Process (in Korean). M.E. thesis, Dept. of Industrial
for representing a machining knowledge, although their Engineering, Korea Advanced Institute of Science and
knowledge representation styles are different. If the Technology, Korea, 1991.
above statement is true, the most important criteria 13. Park, K. S. and Kim, S. H., Use of AI-based approaches
for determining CNC machining parameters in manu-
becomes efficiency.
facturing. In Encyclopedia of Microcomputers, eds A. Kent
There may be some factors for measuring efficiency in and J. G. Williams. Marcel Dekker, Inc., New York,
quantitative or qualitative manner, for example, easy for 1995.
naive user and reaction time. Among them, reaction 14. Thompson, R. W., Tipnis, V. A. and Kegg, R. L. (eds),
time would be the most important factor in real-time Machinability Testing and Utilization of Machining Data.
Materials/Metalworking Technology Series, American
machining control systems. Namely, reaction to
Society for Metals, Metals Park, Ohio, 1979.
machining failures and unfavorable machining has to 15. Tipnis, V. A., Buescher, S. C. and Garrison, R. C.,
be made within seconds or milliseconds to guarantee the Mathematically modeled machining data for adaptive
safety and reliability of the machining process. In this control of end milling operations. Proceedings of the
point of view, we prefer a knowledge representation Fourth NAMRC, 1976, 279-286.
16. Tlusty, J. and Andrews, G., A critical review of sensors for
scheme in which reaction time is less sensitive as the size
unmanded machining, Annals of the CZRP, 1983, 32(2),
of the knowledge base increases. 563-572.
Apart from the sufhciency, what kind of knowledge 17. Tlusty, J. and MacNeil, P., Dynamics of cutting forces in
representation technique is better than the others in a end milling. Annals of the CZRP, 1975, 24(l).
time-critical machining control system? This paper can 18. Mantyla, M. An Introduction to Solid Modeling, Computer
Science Press Inc., 1988.
not give an answer for this question, because there
19. Wang, W. P. and Wang, K. K., Geometric modeling for
have not been a comparative analysis among the AI swept volume of moving solids. ZEEE Computer Graphics
approaches. The problem in order to address this and Applications, 1986, 6, 8- 17.
question is a promising further research issue. 20. Wang, W. P., Solid modeling for optimizing metal
 Kyung Sam Park, Soung Hie Kim

removal of three-dimensional NC end milling. Journal of optimization for intelligent machining. Journal of
Manufacturing Systems, 1988, 7(l), 57-65. Manufacturing Systems, 1991, 10(6), 464475.
21. Hook, T. V., Real-time shaded NC milling display. 33. Rumelhart, D. and McClelland, J., Parallel Distributed
Computer Graphics, 1986, 20(4), 15-20. Processing, Vol. 1. MIT Press, Cambridge, 1986.
22. Jerard, R. B. and Drysdale, R. L., Geometric simulation 34. Lippmann, R., An introduction to computing with neural
of numerical control machining. Proceedings of ASME nets. IEEE Transactions on ASSP, 1987, 4(2), 4-22.
International Computers in Engineering Conference, 1988, 35. Rangwala, S. S. and Dornfeld, D. A., Learning and
129-136. optimization of machining operations using computing
23. Takata, S., Tsai, M. D., Inui, M. and Sata, T., A cutting abilities of neural networks. IEEE Transactions on
simulation system for machinability evaluation using a Systems, Man, and Cybernetics, 1989, 19(2), 299-314.
workpiece model. Annuls of the CZRP, 1989, 38(l), 36. Dornfeld, D. A., Neural network sensor fusion for tool
417-420. condition monitoring. Annuls of CZRP, 1990, 39(l),
24. Taha, H. A., An Introduction to Operations Research. 101-105.
MacMillan Publishing Co., New York, 1982, 15-157. 37. Burke, L. I., Competitive learning based approaches to
25. Chankong, V. and Haimes, Y. Y., Multiobjective Decision tool-wear identification. IEEE Transactions on Systems,
Making: Theory and Methodology. North Holland, New Man, and Cybernetics, 1992, 22(3), 559-563.
York, 1983. 38. Burke, L. I., An unsupervised approach to tool wear
26. Steuer, R. E., Multiple Criteria Optimization: Theory, identification. ZIE Transactions, 1993.
Computation, and Application. John Wiley & Sons, New 39. Howard, R. A. and Matheson, J. E., Influence diagrams.
York, 1986. In The Principles and Applications of Decision Analysis,
27. Malakooti, B. and Deviprasad, J., An interactive multiple Vol. 2, eds R. A. Howard and J. E. Matheson. Menlo
criteria approach for parameter selection in metal cutting. Park, CA, Strategic Decision Group, 1984.
Operations Research, 37(5), 805-8 18. 40. Shachter, R., Evaluating influence diagrams. Operations
28. Hayes-Roth, F., Waterman, D. A. and Lenat, T. J., Research, 1986, 34, 871-882.
Building Expert Systems. Addison-Wesley, Reading MA, 41. Shachter, R., Probabilistic inference and influence
1983. diagrams. Operations Research, 1988, 36, 589-604.
29. Buchanan, B. G. and Shortliffe, E. H., Rule-Based Expert 42. Rege, A. and Agogino, A. M., Topological framework for
Systems: The MYCZN Experiments of the Stanford representing and solving probabilistic inference problems
Heuristic Programming Project. Addison-Wesley, Reading in expert systems. IEEE Transactions on Systems, Man,
MA, 1984. and Cybernetics, 1988, 18(3), 402-414.
30. Lu, S. C-Y. and Komanduri, R. (eds), Knowledge-Based 43. Agogino, A. M., Srinivas, S. and Schneider, K. M.,
Expert Systems for Manufacturing. The American Society Multiple sensor expert system for diagnostic reasoning,
of Mechanical Engineers, New York, 1986. monitoring and control of mechanical systems. Mechanical
31. Lingarkar, R., Liu, L., Elbestawi, M. A. and Sinha, N. K., Systems and Signal Processing, 1988, 2(2), 165-185.
Knowledge-based adaptive computer control in 44. Oliver, R. M. and Smith, J. Q. (eds), Influence Diagrams,
manufacturing systems: a case study. IEEE Belief Nets and Decision Analysis. John Wiley & Sons, New
Transactions on Systems, Man, and Cybernetics, 1990, York, 1990.
20(3), 606-618. 45. Bunn, D. W., Applied Decision Analysis. McGraw-Hill
32. Billatos, S. B. and Tseng, P. C., Knowledge-based Inc., New York, 1984.