You are on page 1of 11

Computers in Industry 43 Ž2000.

31–41
www.elsevier.nlrlocatercompind

Using inductive machine learning to support decision making in


machining processes
Bogdan Filipicˇ a,b,) , Mihael Junkar b
a
ˇ Stefan Institute, JamoÕa 39, SI-1000 Ljubljana, SloÕenia
Department of Intelligent Systems, Jozef
b
ˇ ˇ 6, SI-1000 Ljubljana, SloÕenia
Faculty of Mechanical Engineering, UniÕersity of Ljubljana, AskerceÕa
Received 8 August 1997; received in revised form 1 June 1999; accepted 1 February 2000

Abstract

In spite of their practical success, knowledge-based systems still suffer from considerable limitations. Specialized for
problem solving in a narrow domain, most systems possess very limited knowledge and are rather inflexible. Moreover,
building a knowledge base is the most critical phase in developing an expert system. In overcoming these limitations,
existing machine learning techniques, capable of deriving concepts from data, can be effectively applied. This work focuses
on machine learning from examples and its potential in discovering knowledge hidden in technological databases. Practically
oriented studies of automating two-decision procedures related to machining processes are presented: classification of
dielectric fluids used in electrical discharge machining, and tool selection in an industrial grinding process. The results show
the approach is beneficial in preventing poor process performance and improving product quality. It also allows for better
understanding of the processes at the shop floor level, and advances decision making at the technology planning level.
q 2000 Elsevier Science B.V. All rights reserved.

Keywords: Inductive machine learning; Technological databases; Decision support; Electrical discharge machining ŽEDM.; Grinding

1. Introduction ment of knowledge processing technology, play a


decisive role in managing resources, shortening de-
Knowledge processing technology has changed sign and production times, improving the quality of
computers from high-speed data manipulators to products and services, and reducing costs in numer-
highly specialized problem solvers. Knowledge-based ous settings w4x. Among the users are manufacturing
systems nowadays assist people in carrying out de- companies that employ expert systems to assist in
manding technical and business tasks in a variety of decision making and to release human operators
sectors. Expert systems, the most prominent achieve- from routine low-level tasks w11x.
In spite of its widespread use and commercial
success, knowledge processing technology still suf-
) fers from some considerable limitations. Most of the
Corresponding author. Department of Intelligent Systems,
systems to date have been intended for operation in a
ˇ Stefan Institute, Jamova 39, SI-1000 Ljubljana, Slovenia.
Jozef
Tel.: q386-1-477-33-52; fax: q386-1-425-1038. narrow domain and therefore possess a limited seg-
E-mail address: bogdan.filipic@ijs.si ŽB. Filipicˇ.. ment of expert knowledge. Consequently, they are

0166-3615r00r$ - see front matter q 2000 Elsevier Science B.V. All rights reserved.
PII: S 0 1 6 6 - 3 6 1 5 Ž 0 0 . 0 0 0 5 6 - 7
32 ˇ M. Junkarr Computers in Industry 43 (2000) 31–41
B. Filipic,

very sensitive and inflexible when faced with a fication procedure for dielectric fluids used in EDM.
problem not perfectly fitting into the predefined do- Section 5 reports how the selection of grinding tools
main. The technology also does not scale up linearly, at an industrial plant has been improved by learning
hence, extending the capabilities of a small-scale from technological data. Concluding remarks include
knowledge-based system may be a very demanding a summary of experiences gained in these studies,
task. Finally, due to insufficient automation, knowl- plans for future work, and a general view of the role
edge acquisition is the most expensive and time-con- of knowledge extraction techniques in connecting
suming stage in developing knowledge-based sys- activities at the shop floor and technology planning
tems. Most practical implementations have been levels.
accomplished by interviewing experts and reviewing
written sources, and manually encoding captured
knowledge in a form suitable for computer process- 2. Machine learning from examples
ing. On the other hand, automated knowledge syn- Machine learning is an information processing
thesis has been viewed as an emerging capability, technique concerned with deriving concepts from
and the knowledge acquisition bottleneck treated as a data. A well-explored machine learning approach is
crucial challenge for its further development w7x. learning from examples, also referred to as inductive
In our view, the existing knowledge acquisition machine learning. In this approach, examples of
techniques can assist in overcoming the limitations problem situations are submitted to a learning system
of current knowledge-based systems. A field offering that induces a general description of the underlying
valuable tools to knowledge engineers is machine concepts useful for problem solving. The resulting
learning. Methods developed in this field are suitable concept descriptions can have the form of if–then
for inducing the laws that govern domains of practi- rules, decision trees, or ready-to-use computer proce-
cal interest. As such, they can be employed to auto- dures.
mate the knowledge elicitation procedure. In this Learning examples can often be very naturally
way, more complex knowledge bases can be built in described with attributes and classes. Attributes rep-
shorter time with less engagement of experts. How- resent features from the considered domain, such as
ever, machine learning can be applied not only to temperature, the content of a chemical element, the
design knowledge-based systems, but also to identify shape of an object, etc., while class defines how an
or restructure expert knowledge relevant to solving a example with given attribute values is treated. Such
particular problem or improving an existent state. In descriptions may, for example, correspond to expert
practice it is very common that knowledge possessed decisions under given conditions, like setting process
by individual experts or dispersed in data collections parameters, given measurements and sensory data, or
is overlooked until extracted and presented in an diagnosing a fault from the observed symptoms.
explicit form. Among attribute-based learning systems, there is
These are our findings from the research of appli- a family of algorithms called TDIDT, which stands
cability of machine learning in steel product manu- for Top Down Induction of Decision Trees w14x.
facturing. The subject of our studies were machining Given the training examples in the form of attribute
processes, such as electrical discharge machining value vectors, these algorithms extract domain
ŽEDM. and grinding, and the objective was to auto-
knowledge in the form of a decision tree. A decision
mate decision making related to the processes. The tree is recursively constructed during the learning
machine learning approach used in the studies was process by splitting the training set into subsets until
learning from examples. all examples in the current subset belong to a single
The paper is further organized as follows. Section class. The basic tree construction procedure of a
2 presents machine learning from examples and the classical TDIDT learning algorithm, ID3 w15x, can be
learning system applied in the studies. Section 3 summarized as follows:
discusses the potential of machine learning in ma-
chining processes. In Section 4, a pilot learning Input: set of learning examples in attribute-based
experiment is described, conducted to derive a classi- notation.
ˇ M. Junkarr Computers in Industry 43 (2000) 31–41
B. Filipic, 33

Output: decision tree corresponding to learning approach to estimate the probabilities needed in these
examples. calculations is to statistically derive the correspond-
Tree construction algorithm: ing relative frequencies from the learning set.
In a decision tree obtained through the described
induction procedure, internal nodes denote tests on
IF all learning examples belong to a single class attributes, links correspond to attribute values, and
THEN create a leaf labeled with that class terminal nodes Žalso called leaves. are labeled with
ELSE classes. A set of if–then rules equivalent to the
1. using an information-theoretic measure, find decision tree can be obtained by following the paths
the most informative attribute in the learning from the root of the tree to each of the leaves. The
set; premise of a rule is the conjunction of attribute value
2. create the root of a tree labeled with that tests appearing in the visited internal nodes, and the
attribute; conclusion, the class appearing in the leaf. Either a
3. partition the learning set into subsets according decision tree or rules allow for efficient classification
to the values of the selected attribute; of new problem instances with unknown classes.
4. recursively apply the algorithm to each subset. Moreover, representing domain concepts in a com-
pact and transparent way, such a result is suitable for
The key step of the algorithm is determination of human inspection and often reveals a new insight
the most informative attribute on which to further into the considered domain.
partition the learning subset at each recursive step. Attribute-based learning is the most practically
This is carried in the following way. Let the learning successful form of machine learning. Commercially
set consist of examples from Nc classes and let the available learning programs exist, applicable in
distribution of classes be C. Then, according to knowledge acquisition for expert systems, and
information theory, the amount of information needed achieving impressive results in real domains w1x.
to classify an instance from this set is Among these tools is Assistant w3x, a successor of
ID3, which extends the basic tree induction algo-
Nc rithm in several ways. The extensions include learn-
EŽ C . s y Ý pi log 2 pi , Ž 1. ing from incompletely specified data, binarization of
is1 attributes, and tree pruning, which serves as a noise
where pi is the probability for the instance to belong handling mechanism. These upgrades have proved
to the ith class. If an attribute A with Nv possible useful in areas where learning has to be performed
values is chosen for partitioning the instances, the from incomplete and noisy data. Domains in which
amount of information needed for classification of an Assistant has been applied with success include engi-
instance from the set becomes neering, business, and medicine w17x. Its re-imple-
mentation, called Magnus Assistant w12x, includes
Nv Nc
pi j pi j mechanisms for better estimation of probabilities on
E Ž C N A. s y Ý pj Ý log 2 , Ž 2. small example sets and generating output in the form
js1 is1 pj pj
of directly applicable computer programs. These ex-
tensions allow for more accurate classification and
where pj is the probability for the instance to have
rapid application development.
the jth value of the attribute A, and pi j the probabil-
ity for the instance to both belong to the ith class
and have the jth value of the attribute A. The
informativity of the attribute is, therefore, 3. Potential of learning in machining processes
Inf Ž A . s E Ž C . y E Ž C < A . Ž 3.
As a key stage of steel product manufacturing,
and the most informative attribute is the one that machining in industrial settings is nowadays based
maximizes the value of Ž3.. The most straightforward on a number of new technologies and supported by
34 ˇ M. Junkarr Computers in Industry 43 (2000) 31–41
B. Filipic,

advanced machining devices and CADrCAM facili- chining processes are rich sources of data. Depend-
ties w10x. There is, however, not enough awareness of ing on the type of the process and the need to be
the potential of the knowledge that accumulates satisfied, data can be acquired in different ways, e.g.
around such systems over time. through measurements, calculations, or observations.
In interaction with the customers of our laboratory The collected data helps technologists and operators
test facilities and steel product manufacturers seek- in making decisions related to a process. Examples
ing solutions to their problems, we are often faced of decisions that need to be made prior to or within
with situations where existing knowledge is ne- the process execution include setting parameter val-
glected or not effectively exploited. In most cases, ues, estimating process performance, carrying out
potential knowledge sources are human experts and control actions and many others. In view of machine
collections of technological data. Unfortunately, ex- learning, the available data are treated as attributes,
perts, such as experienced technologists and opera- and decisions as classes. Recorded over time and
tors, are unable to articulate their know-how in an possibly stored in a database, such situations can be
explicit, formal manner. Moreover, human expertise regarded as learning examples. To automatically in-
disappears from companies through fluctuation and duce decision procedures from examples of decisions
retirement. Automated knowledge acquisition from from the past, we use the Assistant learning system.
examples of expert decisions would therefore allevi- Our initial work on attribute-based learning re-
ate the problem of articulation and make the exper- lated to machining processes was the prediction of
tise applicable even when experts are not present. the grinding wheel performance from spectral char-
Similarly, technological databases may contain valu- acteristics of the emitted vibration signals w8x. In a
able knowledge on a machining process, which only similar manner, we assessed the workpiece surface
becomes transparent and operational when extracted roughness generated in a plunge–grinding process
from data. w5x. These laboratory experiments have been later
Using attribute-based machine learning, we han- followed by two practically oriented studies. One has
dle these situations as schematically shown in Fig. 1. been carried out in classification of dielectric fluids
We focus on two activities related to a machining used in EDM, and the other in the selection of tools
process, data acquisition and decision making. Ma- for an industrial grinding process.

4. Classification of dielectric fluids for EDM

4.1. Background

EDM is a machining technique through which the


surface of a metal workpiece is formed by discharges
occurring in the gap between the two electrodes, the
tool and the workpiece. The gap is flushed by the
third interface element, the dielectric fluid. The pro-
cess consists of numerous randomly ignited
monodischarges generating characteristic crater-tex-
tured surface. The quality of the obtained surface
depends on the stability of the gap conditions during
the process. The gap conditions are primarily influ-
enced by the type and state of the dielectric fluid.
The fluid properties affecting the discharge process
Fig. 1. From examples of decision situations to an automatically include viscosity, density, flash temperature, ion
generated decision procedure. number, and others. Incomplete deionization of the
ˇ M. Junkarr Computers in Industry 43 (2000) 31–41
B. Filipic, 35

fluid, for example, results in the occurrence of sta-


tionary-located discharges, which are harmful to the
workpiece surface.
In practice, a dielectric fluid to be used for a
particular EDM operation is chosen by an operator.
Although this decision may seem trivial at first sight,
it actually involves selecting from a growing number
of products available in the market today. In addi-
tion, an EDM system can operate in different regimes,
such as roughing and finishing, which also require
different dielectric fluids. There are certain general
rules prescribing appropriate dielectric fluids for in- Fig. 2. A decision tree for classifying dielectric fluids used in
electrical discharge machining.
dividual regimes, but most of the operators’ deci-
sions are based on their experience and intuition.
To assist the consumers, fluid manufacturers spec- with known properties and recommended regimes
ify relevant physical and chemical properties of their w9x. In addition to the fluids tested in our laboratory,
products and recommend process regimes in which products of some other manufacturers were also
they should be used. These guidelines are obtained considered. A fragment from the learning data set,
through time-consuming laboratory experiments. At showing the considered fluid properties and recom-
the Faculty of Mechanical Engineering, University of mended machining regimes, is shown in Table 1.
Ljubljana, we have been performing laboratory anal- In the learning experiment, the fluid properties
yses of dielectric fluids produced by a Slovenian were treated as attributes and machining regimes as
chemical company for years. For each product, a classes. The result of learning was the decision tree
number of physical and chemical properties were for classifying dielectric fluids, presented in Fig. 2.
determined through measurement and experimenta- The two internal nodes indicate tests on the values of
tion, and the fluid classified into one of the three flash temperature and viscosity, and the leaves sug-
categories: suitable for roughing, suitable for finish- gest appropriate machining regimes. Following the
ing, or universal, i.e. suitable for both machining paths from the root to the leaves, we obtain three
regimes. if–then rules:

4.2. The learning experiment Rule 1: IF flash temperatureF 85 THEN finish-


ing.
In the course of time, the manufacturer of dielec- Rule 2: IF flash temperature) 85 AND viscosity
tric fluids and our operators became more interested F 3.9 THEN universal.
in the correlation between the fluid properties and Rule 3: IF flash temperature) 85 AND viscosity
the machining process performance. To explore ) 3.9 THEN roughing.
whether the appropriate machining regime could be
predicted from the fluid properties, we applied the The obtained rules partition the considered set of
Assistant learning program to a database of products dielectric fluids into three subsets corresponding to

Table 1
A fragment from the dielectric fluid database used in the machine learning analysis
Product name Viscosity Density Aromates Boiling Flash Solidif. Ion number Machining
wmm2 rsx wgrcm3 x w%x temp. w8Cx temp. w8Cx temp. w8Cx w–x regime
CCL 51 5.7 0.823 6.0 268 122 y15 1.07 roughing
Erozol BP180 1.7 0.749 0.1 195 70 y14 NrA finishing
PED 4549 2.0 0.797 4.0 196 68 y6 0.35 finishing
36 ˇ M. Junkarr Computers in Industry 43 (2000) 31–41
B. Filipic,

different machining regimes. It is interesting to see for each grinding operation. At the process planning
that only two fluid characteristics, i.e. its flash tem- stage, only general instructions about tool selection
perature and viscosity, are sufficient to classify the are given, while the main burden of determining a
fluids. While the impact of viscosity was known tool for every individual task falls on the operators.
before, our study showed that the flash temperature To bridge the information gap between the shop
of a fluid is also an important indicator of the EDM floor level and the production planning department,
regime in which the fluid should be used. the state of the existing grinding technology was
Further tests were carried out to check the classi- recorded by interviewing the operators. Stored in a
fication accuracy of the obtained decision procedure. technological database, the acquired data included
The original data set was randomly split into two the workpiece and process characteristics and the
subsets: 70% of examples were used for learning and related grinding wheel specifications. This work was
the remaining 30% for testing. A decision tree was initiated by Roethel and Junkar w16x. The objective of
induced from the learning instances and applied to further cooperation with the company has been to
the test instances. The predicted classes were com- induce general rules for tool selection from the tech-
pared with known machining regimes to obtain the nological database. In analyzing the existing state
proportion of correctly classified test instances. This and suggesting steps towards its improvement, ma-
test was performed 10 times, each time using differ- chine learning played an essential role w6x.
ent split of the data into learning and test subsets.
The average proportion of correctly classified test 5.1. The tool selection task
examples was then assumed as an estimate of the
classification accuracy. The resulting value was Grinding of an industrial knife involves six opera-
97.4%, which shows that, given the properties of a tions related to the knife surfaces Žsee Fig. 3.. The
dielectric fluids, it is possible to predetermine the cutting surface is first pre-ground and then ground
recommended machining regime with high accuracy. with a different grinding wheel. The two grinding
Although elementary, this case study illustrates operations are called face shaping and face dimen-
how a practical decision procedure can be extracted sioning, respectively. Grinding of upper and lower
from examples of problem instances. We see that the surfaces, and length and width sides requires a single
approach is beneficial for both the manufacturers and grinding operation for each surface. Each of these
consumers of dielectric fluids. For a manufacturer, it operations also has to be carried out with a specific
has a potential of providing guidelines for the design grinding wheel.
of new fluids and shortening the test procedure, The workpiece material, steel, can be of different
while a consumer can automate fluid selection by chemical structure. Its characteristics are determined
learning from decisions made in the past. by the content of six elements: carbon ŽC., chromium
ŽCr., molybdenum ŽMo., silicon ŽSi., vanadium ŽV.,

5. Selection of grinding wheels in industrial knife


manufacturing

Another machine learning exploration of techno-


logical data has been conducted for a large steel
company. One of its departments produces knives
used in industry for cutting materials, such as metal,
wood, paper, and leather. Crucial for the quality of
knives is grinding which is the final phase of knife
manufacturing. To meet the desired quality stan-
dards, it is necessary to select an appropriate grind-
ing tool for every type of the workpiece material and Fig. 3. Knife surfaces to be machined by grinding.
ˇ M. Junkarr Computers in Industry 43 (2000) 31–41
B. Filipic, 37

and tungsten ŽW.. Important for the material proper- Table 2


ties is also the heat treatment of a workpiece. Steel Example prescriptions for grinding wheel selection Žmachine:
BVP-40; steel composition: 0.9% C, 4% Cr, 5% Mo, 1.9% V,
of three different treatments is used at the depart- 6.5% W; operations: grinding lower surface and length side of an
ment of industrial knives, annealed and quenched at industrial knife.
two different temperatures. Heat treatment Grinding wheel specification
Grinding wheels are characterized by three fea-
Hardness Grain size Porosity
tures: hardness, grain size, and porosity, which is
also referred to as structure. To denote values of Annealing G 24 10
Quenching 48–54 HRc K 46 5
these characteristics, standardized scales are used by Quenching 55–63 HRc H 46 8
wheel manufacturers. For practical purposes, wheel
characteristics are often treated qualitatively. For
example, wheels of hardness E, F and G are consid-
ered soft, grain size below 16 denotes very rough
wheels, wheels with structure designation higher than wheel characteristics separately. In other words,
10 are highly porous, etc. learning was performed three times, every time con-
Given the workpiece chemical structure, its heat sidering a different wheel characteristic as class. In
treatment and the grinding operation to be per- the learning procedure, qualitative values of charac-
formed, the task is to determine the appropriate teristics were treated. Class values for hardness rep-
grinding wheel in terms of its hardness, grain size, resented in the learning set were medium, soft and
and porosity. very soft, grain size was rough or medium, and
structure medium, open or highly porous. This strat-
egy resulted in three decision trees, one for each
characteristic of a grinding wheel.
5.2. Learning tool selection rules
The other strategy was learning complete wheel
specifications. Here, a single learning task was con-
A machine learning analysis of tool selection sidered where class values were elements of the
prescriptions acquired from operators and stored in a Cartesian product of hardness, grain size and poros-
technological database was done with the Assistant ity classes used in the first strategy. Values of the
learning program. The learning set consisted of 190 compound wheel specification class were, for in-
regularly used prescriptions, covering the entire pro- stance, soft–medium grain size–open structure, very
gram of industrial knives manufactured at the depart- soft–rough grain size–highly porous, etc. The result
ment. Learning instances were described by eight of learning was a single decision tree covering com-
attributes Žthe content of six chemical elements, heat plete specifications of grinding wheels.
treatment of the workpiece, and the grinding opera-
tion. and the grinding wheel specification, consisting 5.3. Results
of wheel hardness, grain size and porosity. Examples
of prescriptions for wheel selection are shown in Results of the machine learning experiments for
Table 2. They are intended for grinding lower sur- the entire set of grinding machines at the industrial
face and length side of a knife made of steel with knife manufacturing department are summarized in
given chemical structure. It can be seen, for exam- Table 3. The classification accuracy was obtained
ple, that for the annealed workpiece the prescribed empirically, using the test procedure presented in
hardness of the grinding wheel is G, grain size 24 Section 4.2. Surprisingly, the accuracy of the ob-
and porosity 10. In the same manner, the wheels are tained decision trees was quite low. It was only
prescribed for quenched steel. slightly higher than the relative frequency of the
In machine learning analysis of prescriptions, two majority class instances in the learning set, which is
different learning strategies were applied. They dif- the lower bound for the classification accuracy of
fered in treating the wheel specifications. The first any reasonable classifier. To explain this outcome,
strategy was learning to determine each of the three the tool selection prescriptions used for learning and
38 ˇ M. Junkarr Computers in Industry 43 (2000) 31–41
B. Filipic,

Table 3
Learning grinding wheel specifications for the entire group of machines
Specification element Classes Majority Classification
class w%x accuracy w%x
Hardness 3 44.2 58.5
Grain size 2 52.6 65.2
Structure 3 51.0 57.7
Complete specification 13 27.4 36.3

the resulting misclassifications were examined in chines on which the majority of the prescriptions
detail. It turned out that in a number of cases a Žabout 30 per machine. are applied. The classifica-
variety of different grinding wheels were used for tion accuracy tests confirmed compatibility between
the same operation, workpiece chemical structure the resulting decision trees and the tool selection
and heat treatment. Disagreements were up to two prescriptions. Moreover, according to the involved
classes per wheel characteristic. This was due to operators, the induced decision trees were in accor-
different preferences of operator teams on individual dance with their way of selecting the tools. It also is
machines. As a matter of a fact, these discrepancies worth to mention that the operators found decision
led to inconsistent quality of the products and re- trees a compact and easy-to-understand representa-
quired an excessive stock of grinding wheels. The tion of the decision making procedure.
machine learning analysis clearly pointed out the Fig. 4 shows the decision tree obtained for tool
reason for these drawbacks and showed the need to selection on the BVP40 machine where five different
adjust the prescriptions. types of wheels are used. Giving an explicit rule for
As the operators’ decisions applied on a single each of the wheel classes, it effectively substitutes
machine were expected to be more consistent, further the underlying 28 prescriptions for tool selection. It
exploration was done in this direction. Tool selection also reveals an important simplification: an appropri-
prescriptions were divided into subsets with respect ate wheel for grinding operations to be performed on
to grinding machines on which they were in use. this machine can be determined by only considering
Machine learning was then applied separately to the the content of carbon in the workpiece and its heat
subsets to generate decision trees for tool selection treatment.
on individual machines. Again, the classification ac- Partial results obtained for individual machines
curacy of the resulting decision trees was determined represent a base for unifying the tool selection pre-
empirically. Table 4 shows the results of learning scriptions for the entire department. Through com-
complete wheel specifications for three grinding ma- pact description of the existent state, differences can

Table 4
Learning grinding wheel specifications for individual machines
Grinding machine Classes Majority Classification
class w%x accuracy w%x
BVP-40 5 32.2 95.3
BVP-60a 3 66.7 96.0
SIELEMANN-a 2 66.7 100.0
ˇ M. Junkarr Computers in Industry 43 (2000) 31–41
B. Filipic, 39

Fig. 4. A decision tree for grinding wheel determination on machine BVP40.

be more easily analyzed and uniform solutions found ple, a review in Ref. w13x.. In machining, grinding
where possible. The goal of the ongoing work is to appears to be particularly well suited for employ-
form a consistent knowledge base of an expert sys- ment of intelligent systems, including machine learn-
tem for tool selection in manufacturing industrial ing techniques. Their applications range from param-
knives. The system is meant to support regular pro- eter setting and tool selection to adaptive systems
duction as well as experimentation in developing integrated with machine controllers w2x. Applications
new products. These enhancements are undertaken as in machining domains presented in this paper partic-
a step towards higher flexibility in responding to ularly demonstrate how applicable rules can be effec-
customer demands. tively extracted from technological data.
Encouraged by the described pilot experiment in
classifying dielectric fluids for EDM, we now con-
tinue with facilitating the interactions within the
6. Conclusions customer–supplier chain. The issue that has recently
brought the two sides closer to each other is ecology.
In our experience, knowledge extraction from ex- Not only the impact of dielectric fluids on the quality
amples in real-world domains is an iterative process. and effectiveness of machining, but also their envi-
Cooperation of both domain experts and knowledge ronmental effects are becoming more and more im-
engineers in providing and refining the training set portant. Dielectric fluid manufacturers are forced to
and evaluating the results in consecutive iterations is also consider noxiousness and organoleptic proper-
vital for its success. However, when properly used, ties of the products, their dissolution characteristics,
machine learning is a powerful tool in building etc. In a new investigation, these effects are planned
knowledge bases and automating decision making. A to be monitored at several customers, while the fluid
number of recent reports confirm that this informa- manufacturers will provide additional physical and
tion processing technique is gaining interest in a chemical characteristics for their products. Using the
wide range of manufacturing sectors Žsee, for exam- inductive machine learning approach to correlating
40 ˇ M. Junkarr Computers in Industry 43 (2000) 31–41
B. Filipic,

data from both sides, we expect to contribute to their enable designers to make proper decisions at early
common goal: to produce, at minimal costs, dielec- planning stages.
tric fluids that would satisfy both performance re-
quirements and ecological standards.
In selecting tools for grinding industrial knives,
the machine learning analysis has proved beneficial References
in detecting inconsistencies and suggesting correc-
tions of existing prescriptions. When implemented, w1x I. Bratko, Applications of machine learning: towards knowl-
the expert system for tool selection with consistent edge synthesis, New Gener. Comput. 11 Ž3. Ž1993. 343–360.
w2x W.B. Rowe, Y. Li, B. Mills, D.R. Allanson, Application of
rules for the entire industrial knives department will
intelligent CNC in grinding, Comput. Ind. 31 Ž1. Ž1996.
allow for significant reduction of production costs in 45–60.
the factory. In particular, an analysis showed that w3x B. Cestnik, I. Kononenko, I. Bratko, Assistant-86: a knowl-
elimination of inconsistent tool selection prescrip- edge-elicitation tool for sophisticated users, in: I. Bratko, N.
tions could decrease costs of man and machine time Lavracˇ ŽEds.., Progress in Machine Learning, Sigma Press,
and tool wear by up to 30%. Since product orders Wilmslow, UK, 1987, pp. 31–45.
w4x E. Feigenbaum, P. McCorduck, H.P. Nii, The Rise of the
from various customers do not vary much in dimen- Expert Company, Times Books, New York, 1988.
sions and geometry, the new system will contribute w5x B. Filipic,
ˇ M. Junkar, I. Bratko, A. Karalic,
ˇ An application of
to savings in both the operating time and the invest- machine learning to a metal-working process, Proceedings of
ments into wheel stock. Finally, the new approach the 13th International Conference Information Technology
will offer general instructions for tool selection in Interfaces ITI ’91, Cavtat, Croatia Ž1991. 167–172.
w6x B. Filipic,
ˇ M. Junkar, Machine learning refinement of opera-
grinding new materials with chemical structure and tor decisions in a manufacturing process, Proceedings of the
heat treatment similar to that of already encountered 2nd World Congress on Intelligent Manufacturing, Processes
examples. Exploiting the similarities, a large part of and Systems, Budapest, Hungary Ž1997. 42–47.
time-consuming experiments now performed on trial w7x F. Hayes-Roth, N. Jacobstein, The State of knowledge-based
and error basis will then be unnecessary, and labora- systems, Commun. ACM 37 Ž3. Ž1994. 27–39.
w8x M. Junkar, B. Filipic, ˇ I. Bratko, Identifying the grinding
tory experimentation on new materials will be more process by means of inductive machine learning, Comput.
systematic and cheaper. At the same time, equipped Ind. 17 Ž2–3. Ž1991. 147–153.
with better knowledge about the technology, the w9x M. Junkar, B. Filipic, ˇ
ˇ M. Znidarsic,
ˇ ˇ An AI approach to the
designers of new products are expected to be more selection of dielectricum in electrical discharge machining,
efficient in their work. in: Proceedings of the 3rd International Conference Ad-
vanced Manufacturing Systems and Technology AMST ’93,
In conclusion, since the application of appropriate Udine, Italy, 1993, vol. 2, pp. 11–16.
technology for a specific manufacturing task remains w10x M. Junkar, P. Levy, Nonconventional processes integrated
the basic challenge for both technologists and de- into manufacturing systems, Manuf. Syst. 23 Ž2. Ž1994.
signers, decision support in finding appropriate tech- 151–159.
w11x R. Maus, J. Keyes ŽEds.., Handbook of Expert Systems in
nological solutions is of great importance. Novel
Manufacturing, McGraw-Hill, New York, 1991.
concepts, such as concurrent engineering, require w12x D. Mladenic, ´ Magnus Assistant, Technical report IJS DP-
comprehensible and reliable information in order to ˇ Stefan Institute, Ljubljana, 1994 Žin Slovenian..
6938, Jozef
effectively implement design for manufacturing. w13x L. Monostori, A. Markus, H. Van Brussel, E. Westkamper, ¨
Much of the needed information can be extracted Machine learning approaches to manufacturing, Ann. CIRP
from data and experience gained on production lines 45 Ž2. Ž1996. 675–712.
w14x J.R. Quinlan, Induction of decision trees, Mach. Learn. 1 Ž1.
in the shop floor. In our view, machine learning Ž1986. 81–106.
methods are an appropriate tool for incorporating w15x J.R. Quinlan, Discovering rules by induction from large
expert knowledge into decision making procedures collection of examples, in: D. Michie ŽEd.., Expert Systems
for machining. They help in clarifying complex in- in the Microelectronic Age, Edinburgh Univ. Press, 1979, pp.
terrelations among parameters involved in machining 168–201.
w16x F. Roethel, M. Junkar, Development of a technological
process and enhance process performance prediction database for industrial knives, Technical report FS 8605,
and control. They also support transfer of relevant Faculty of Mechanical Engineering, University of Ljubljana,
information to the technology planning level, and 1986 Žin Slovenian..
ˇ M. Junkarr Computers in Industry 43 (2000) 31–41
B. Filipic, 41

w17x T. Urbancic, Mihael Junkar obtained his PhD de-


ˇ ˇ I. Kononenko, V. Krizman,
ˇ Review of applica-
gree from the Faculty of Mechanical
tions by Ljubljana Artificial Intelligence Laboratories. Tech-
Engineering, University of Ljubljana,
nical report IJS DP-6218, Jozefˇ Stefan Institute, Ljubljana,
Slovenia. There, he is now an associated
1991.
professor at the Department of Manufac-
Bogdan Filipicˇ received his PhD in turing Technologies, and the head of the
Computer Science from the University Laboratory for Alternative Technologies
of Ljubljana, Slovenia, in 1993. He is and Engineering Metrology. He is also a
currently a research associate at the De- visiting professor at the University
partment of Intelligent Systems of the Joseph Fourier in Grenoble, France. His
ˇ Stefan Institute, Ljubljana, and an
Jozef research areas are machining process
assistant professor of Computer and In- identification, modeling and control,
formation Science at the University of management of innovative technologies and engineering metrol-
Ljubljana. His research interests include ogy. He has published over 50 research papers and led several
evolutionary computation, machine research projects for industry.
learning and knowledge-based systems.
He is active in promoting these tech-
niques in practical problem solving in engineering, manufacturing,
and industrial process optimization. He has published in several
scientific journals, including Computers in Industry, Engineering
Applications of Artificial Intelligence, IEEE Transactions on Sys-
tems, Man and Cybernetics, and ReÕiew of Scientific Instruments.
He is a member of IEEE Computer Society and Slovenian Artifi-
cial Intelligence Society.

You might also like