You are on page 1of 11

Journal o f Intelligent Manufacturing (1991) 2,353-363

ITONUS: expert system for machining on a


lathe
T H O M A S L. W A R D l, P A T R I C I A A. S. R A L S T O N 2,
W A L D E M A R K A R W O W S K I 3 and W. D A R R I N H A L L l

Department of Industrial Engineering ~, Engineering Mathematics and Computer Science2 and


Center for Industrial Ergonomics-~, University of Louisville, Louisville, KY 40292, USA
Received October 1989 and accepted March 1991

A microcomputer-based expert system for the diagnosis of lathe machining operations has
been developed using the EXSYS shell on an IBM AT. Interviews with a journeyman
machinist resulted in an initial knowledge base of rules containing basic concepts and
relationships concerning set-up and control of lathe machining. The system provides the user
with analysis and advice regarding the diagnosis and certain set-up aspects of common lathe
operations, and has shown that it is feasible to elicit and mimic at least a portion of an expert
machinist's knowledge and special skills.

Keywords: expert system, machining, diagnosis, knowledge representation

1. Background
Computers are used to operate and organize the operation
of metal-cutting machines. In the case of computer
numerically controlled (CNC) machines, the servomechanisms that control speeds and feeds receive inputs
from a local computer located at the machine tool (Pressman and Williams, 1977). The computer memory stores
numerical control (NC) programs that determine the part
geometry, and machine feeds and speeds. In many factories, machine tools communicate with a central supervisory
computer. In the fully automated factory (one with zero or
low displaceable labor), the central computer may also
control the material-handling system that moves parts from
one machine to the next (Grover, 1980).
Widespread use of CNC machines has left a gap previously occupied by the intelligence of machinists. An
intermediate controller is required to close this gap, Fig. 1.
Such a controller should insure that operation is within the
constraints imposed by the machine, guard against forced
oscillations (chatter) of the tool, and provide for the
sensing and measurement required for automatic tool
replacement. Roughly, these are the intelligent functions
once supplied by the human machinist. A similar view of
the machinist has been described by Wright (1983) and
Wright and Bourne (1988).
Two attempts have been made to fill this gap: first, actual
0953-9875/91 $03.00 +. 129 Chapman& Hall

NC programs are written conservatively; speeds and feeds


are set low enough so that machine integrity is never
threatened and tool chatter never occurs. Also, tools are
replaced well before the end of their useful lives. Second,
adaptive control (AC) systems have been developed in the
research laboratory, but have not been applied in industry
(Schmenk, 1984; Wick, 1977 and Uhlman and Schmenk,
1981). More recently, emphasis has been on improving
total time utilization of machine tools and addressing such
issues as unattended operation and elimination of idle time
(Astrop, 1978; Stauffer, 1978 and Inaba, 1980).
Figure 1 shows two major functions of the journeyman
machinist: sensing and control. The research described is
concerned primarily with the control function. Sensing by
human machinists has been considered elsewhere (Wright
and Bourne, 1988; Dressman et al., 1987 and Ward et al.,
1988).
In a flexible manufacturing system (FMS) where production batches include sequences of differing parts, the AC
must recognize the appropriate constraints, and then
insure that machine operation is within those constraints
for each part. Similarly, tool wear sensing must estimate
the cumulative effect of wear that results from each
sequence of differing parts. Such performance has not yet
been achieved.
Exceptions invariably occur, situations that cannot be
handled by the central computer, and that must be cleared

354

W a r d et al.
Unexpected
mofefiol
variations

INPUT:
De~irea
low cost
high
quolit'/

product

Pad throughput
Tool life
Surface finish
Ch~pmanagement
Energy cor~umption

Fig. 1. Relation of the machinist's control and sensing functions to


global lathe machining control system.

by human intervention. These exceptions include failure of


parts to feed and/or be positioned in the machine or
material-handling system, undesirable chip flow, tool
crashes and breaks, and diminished quality due to tool
wear and/or chatter.
Two approaches which may provide performances
superior to AC are expert systems (ES) and fuzzy logic
control (FLC). These control approaches center on incorporating human experience rather than relying exclusively
on process modeling. They suggest themselves because of
the success human machinists have had in controlling
metal-cutting processes. ES originated from the study of
artificial intelligence (AI), whereas FLC had its origin in
control theory and fuzzy set theory. Unlike ES, most FLC
systems are formulated by the designer rather than being
extracted from a human expert. The ES is frequently used
for decision-making and may be accessed by a human
operator using a computer work station. The FLC is
frequently used for the control of technological systems
and has hard-wired sensor inputs, and outputs that control
hardware. But it is certainly the case that an ES that used
fuzzy logic (as many do) (Negoita, 1985 and Gupta and
Sanchez, 1982) and that controlled a technological system
would be functionally equivalent to an FLC that relied
upon a rule base extracted from a human expert.
This work is the result of a first step in closing the gap left
when CNC machines replaced the human machinist.
ITONUS (the system is named for Itonus, a figure from
Greek mythology who invented a method for polishing
metal) as a stand-alone expert system could be used for
troubleshooting NC programs. However, as it evolves,
ITONUS will be used to replace or redesign the rule base of
the FLC controller for a machine tool.

2. Fuzzy logic control


Fuzzy set theory was announced by Zadeh (1965) in 1965.
FLC is an application of that theory. Chang and Zadeh
(1972) first applied fuzzy set theory to control, but it was
Zadeh (1973) who formulated the basic approach to fuzzy

control design for complex processes. It was in the latter


reference that the concepts of a fuzzy relation and its
underlying inference rule were defined, and their application to fuzzy controller theory set out. Fuzzy industrial
controllers depend on linguistic control rules, which are a
collection of actions to be taken given certain conditions.
These statements, though imprecise, may still contain
relevant and important information. Fuzzy controller
algorithms are formed from the combination of such rules.
Fuzzy control of processes is an alternative when systems
cannot be well controlled by classical or modern control
techniques (Maiers and Sherif, 1985).
Reports of applications of FLC to control of machine
tools began with the work of Zhu et al. (1982). This work
produced a fuzzy logic controller for grinding. The singleinput, single-output loop controlled surface finish by
varying the feedrate.
Sakai etal. (1984) and Sakai and Ohkusa (1984 and 1985)
have considered the application of FLC to turning. They
assume a cutting condition space and a cutting state space.
The cutting condition space consists of speed, feed, and
depth of cut. The proposed cutting state space is defined to
control chip length, and to guard against chatter and tool
fracture. By assuming that there is an analytical mapping
from cutting condition to cutting state space (at least in a
local region) a preliminary framework for computing fuzzy
control statements is obtained. The framework is basically
rule base construction. However, they have also offered an
example that gives membership functions for chip length,
feed rate, and cutting speed, and that has rules for
adjusting feed and speed in response to chip length.
Ralston and Ward (1987) have examined the issue of FLC
of metal cutting on lathes within the context of computer
numerical control (CNC) and adaptive control (AC), and
with Dressman et al. (1987) they have considered the
related question of expert systems for metal cutting.
Ultimately, ITONUS will provide a base set of rules for
the intelligent (knowledge-based) control of a lathe using
FCC.

3. Expert systems technology


An expert system is a computer program that embodies the
expertise of one or more experts in some domain and that
applies this knowledge to make useful inferences for the
user (Waterman and Hayes-Roth, 1983). An ES program
uses symbolic reasoning rather than simple algorithm
execution. Knowledge is a collection of related facts,
beliefs, and heuristic rules (Hayes-Roth et al., 1983). In
expert system development, knowledge representation is
the establishment of a correspondence between a symbolic
reasoning system and the outside world.
This information constitutes the knowledge base. Knowledge acquisition is the extraction of knowledge from

I T O N U S : expert system f o r machining on a lathe

sources of expertise (human experts, books, documents,


computer files, etc.) and subsequent transfer to the knowledge base. The process of knowledge acquisition is
formally described by five stages (Hayes-Routh et al.,
1983): (1) identification - problem is identified; (2) conceptualization- knowledge representation scheme is tentatively selected; (3) formalization - structure to organize
knowledge is designed; (4) implementation - knowledge is
mapped into representational framework, and (5) testingprototype is evaluated by using a variety of example
problems or cases.
In the development of expert systems, three people or
groups of people are important. These are the domain
expert(s), the knowledge engineer, and the user(s). The
knowledge engineer elicits knowledge from the expert.
The ultimate use of an ES is that it allows the user to make
decisions as if he or she were the expert. The limitations of
expert systems are due to the fact that they fall short on
general intelligent behavior. For example, they are unable
to recognize problems for which their own knowledge base
is inapplicable or insufficient, they have no independent
means of checking conclusions, and they have weak
explanation skills.
If the ES makes use of a human expert, as most do,
extraction of knowledge from the expert is the most
time-consuming part of the ES development. Several
methods are often employed to extract knowledge from
human experts: interviews, protocol analysis, walkthroughs or observation of experts performing activities,
questionnaires, decision analysis, or use of induction rule
development by examining examples of problems the
expert solves (Hart, 1986; Greenwell, 1988; Gaines and
Boose, 1988 and Boose and Gaines, 1988).
The over-all construction of an expert system requires
that the ES must be built satisfactorily in a reasonable time
at a reasonable cost. Typically, seven steps are involved
(Turban, 1990): (i) specify problem and determine feasibility; (ii) select expert(s); (iii) conceptually design ES and
complete feasibility study; (iv) select hardware and software; (v) perform knowledge acquisition, representation,
and inferencing, (vi) build prototype, and (vii) evaluate
performance and make improvements. These steps were
followed in the development of ITONUS.

4. Knowledge acquisition
The task of gathering information, generally, from any
source, is called the knowledge acquisition, while the task
of gathering information from the domain expert is called
the knowledge elicitation (Shadbolt and Burton, 1990).
The main question in knowledge elicitation is how the
knowledge engineer can get the domain experts to tell what
they do. Knowledge acquisition (or elicitation) is the
scientific and engineering problem of formalizing a domain

355
expertise for the first time (Kodratoff et al., 1988). The
methods used to achieve this goal reflect the expert's ability
to explain his/her behavior. Kodratoff et al. (1988) divided
the knowledge elicitation process into three steps:
(1) Obtaining background knowledge through expert
interaction and literature on the domain;
(2) Learning full description of high-level and intermediary-level concepts (concept formation);
(3) Learning diagnostic rules and meta-level knowledge
(rule learning).
According to Buchanan (1982), the knowledge acquisition (or elicitation) is the transfer and transformation of
problem-solving expertise from some knowledge source to
a program, with potential sources of knowledge being
human experts, textbooks, databases, and even one's own
experience. In addition to such knowledge elicitation
techniques as expert interview, verbal protocol analysis
and observational studies (Welbank, 1983), a number of
psychological techniques, including the personal construct
theory (Shaw and Gains, 1984), and the concept of sorting
are being used. Recently, Belkin et al. (1988) have
proposed the discourse analysis method of knowledge
elicitation based on the collection of data consisting of
natural language human-human interactions. Table 1
shows a taxonomy of knowledge acquisition methods
modified from Santamarina and Salvendy (1989).
The main difficulty in knowledge acquisition is that in the
process of human thinking one attempts to understand and
the model is not subject to direct observation. Recently,
Johnson et al. (1988) developed a framework for representing expertise required to perform a given task. This
framework, based upon inferences made from a record of
problem-solving activities consists of the following:
(1) The expert can be viewed as a processor that has the
capability of producing a certain problem-solving behavior
using expertise. The task of knowledge acquisition is to
determine this expertise;
(2) The expert develops a set of actions and abilities that
are necessary to realize this expertise;
(3) Although one cannot observe the expertise directly,
the invocation of the expert's actions and abilities in a
record of problem-solving behavior can be observed;
(4) Since one can observe the invocation of actions and
abilities by the expert, some representation of the expertise
can be developed;
(5) A statement of the expertise required to perform a
task serves as a specification of the requirements for a
computer program that is designed to perform this task.
For example, the methodology for collecting and analyzing the protocol data leading to development of specification of expertise can be based upon identification of
operations, episodes, and data cues which are the basic
categories of behavior in the protocol records (Johnson et

Ward et al.

356
Table 1. Taxonomy of knowledge acquisition tools (Santamarina and Salvendy, 1989).

(I) Problemsolving
(1) Protocol analysis
(2) Retrospective probing
(3) Procedural simulation or Problem analysis (Grover,
1983 and Waterman, 1986)
(4) On-site observation
(5) Forward scenario simulation (Grover, 1983)
(6) Introspective reports

Expert solves the problem aloud


Expert responds to specific probes after completion of the task
Expert solves real problem while being occasionally probed for the reasoning
process
Expert is watched solving a problem on site
Expert chooses a case and verbalizes the reasoning process in reaching the
goal
Experts try to explain their knowledge, skills, and decision process

(II) Domain dissection


Characteristics of a symptom are discovered by distinguishing the events that
could cause it from those that could not
Goals are successively broken down into sub-goals to the level of observable
facts
Grouping symptoms (Grover, 1980 and Hart, 1986) Symptoms are listed and successively grouped until the final goal is reached;
alternatively, the interview is driven towards the construction of rules which
help classify observations into more specific objects and activities
ETS- Repertory grid (Boose, 1986)
Elicitation is led to develop a grid of constructs- hierarchical breakdown
Path division (Kahn etal., 1985)
Seeks for a cause on the path linking a diagnosable event with an already
Path differentiation (Kahn et al., 1985)
reported symptom
Seeks to find whether an event is the result of overlapping causal paths or
Frequency conditionalization (Kahn etal., 1985)
non-overlapping ones
Differentiation (Kahn etal., 1985)
Seeks for conditions that will make a symptom more or less likely to occur
Finds symptoms that distinguish diagnosable events

(1) Distinguishing events (Hart, 1986andKahnetal.,


1985)
(2) Dividing goals (Grover, 1983 and Hart, 1986)
(3)
(4)
(5)
(6)
(7)
(8)

(I11) Domain description


Problem discussion (Waterman, 1986)
Characteristics and decisions (Hart, 1986)
Problem description
Critical incident technique (Hart, 1986)

(1)
(2)
(3)
(4)

(IV) System improvements


(1) System refinement (Waterman, 1986)
(2) System examination (Waterman, 1986)
(3) Performance feed-back (Gaines, 1986)
(4) Systemvatidation (Waterman, 1986)

Expert discusses information and procedures needed to solve problem


Expert lists characteristics and decisions; then he is asked to match sets
Expert describes characteristic problems for each type of answer
Expert describes interesting or difficult cases that he recalls
Expert provides problems to be solved with elicited knowledge
Expert examines system's knowledge and structure
Cases solved by expert and system are presented to other experts

(V) Groups of assessors


(1) Crawford slip method (Boose, 1986)
(2) Delphi method

Groups of individuals respond to questions on slips of paper


Structured form of communication, with controlled feed-back for
re-evaluation, oriented to facilitating consensus

(VI) Other methods


(1) Induction (Hart, 1986)
(2) Constraint propagation
(3) Brainstorming

Knowledge is extracted from a training set


Constraints defining cases in a training set are propagated to establish the
constraints that characterize the alternative solutions
Individuals list potential ideas which are then evaluated

al., 1988). Operations are primitive activities of problemsolving that do not depend on a particular context, like
collecting data or making computations. The patterns of
operations that are repeated within and across different
problems in the protocol data are called episodes. Data
cues are operands which comprise the data processed in the
problem-solving operations, and are indicated by domain
nouns.

According to Santamarina and Salvendy (1989), the


most common arrangement for knowledge acquisitioIi
scheme involves the interaction between the expert, knowledge engineer and the system under development. In this
framework, the knowledge acquisition involves three
steps: (1) the adaptation to the expert and his domain, and
to the computer structure; (2) the extraction of relevant,
correct and complete knowledge from the expert, and (3)

ITONUS: expert system for machining on a lathe


the implementation of this information in a computer
system.
As pointed out by Gaines and Boose (1988), the
knowledge acquisition bottleneck in development of expert systems is not only a problem of accessing and
translating what is already known, but the problem of
formalizing models for the first time. The main problems of
accessing the human expert's knowledge were outlined by
Gaines (1988) as follows:
(1) Expertise may be fortuitous. Results obtained may
be dependent on features of the situation which the expert
is not controlling;
(2) Expertise may not be available to awareness. An
expert may not be able to transmit the expertise by
critiquing the performance of others because he is not able
to evaluate;
(3) Expertise may not be expressible in language. An
expert may not be able to transmit the expertise explicitly
because he is unable to express it;
(4) Expertise may not be understandable when expressed in language. An apprentice may not be able to
understand the language in which the expertise is expressed;
(5) Expertise may not be applicable even when expressed in language. An apprentice may not be able to
convert verbal comprehension of the basis of a skill into
skilled performance;
(6) Expertise expressed may be irrelevant. Much of
what is learned, particularly under random reinforcement
schedules, is superstitious behavior that neither contributes nor detracts from performance;
(7) Expertise expressed may be incomplete. There will
usually be implicit situational dependencies that make
explicit expertise inadequate for performance;
(8) Expertise express may be incorrect. Experts may
make explicit statements which do not correspond to their
actual behavior and lead to incorrect performance.
The elicitation of knowledge from experts is time-consuming and usually lacks systematic conceptual design
methods (Cleaves, 1988). The requirement for knowledge
elicitation is to ensure that the expert's best judgement is
extracted. Cleaves (1988) points out that an important
prerequisite for selecting the known experts and in choosing the method of knowledge elicitation is defining what is
expertise.

5. Knowledge representation
Human knowledge can be declarative, i.e. expressing the
state of the world through a set of specific statements, or
procedural, illustrating how to do things (Conway and
Wilson, 1988). The main difference between these two

357
types of knowledge is that procedural knowledge cannot be
as easily described or retrieved as in the case of declarative
knowledge. This difference has a direct bearing on the
knowledge representation schemes for expert systems. For
example, semantic nets and other schemata are used to
represent the easily describable declarative knowledge,
while frameworks and production systems are utilized for
representing the procedural knowledge,
According to Duce and Ringland (1988), the knowledge
representation problem can be described using three main
components. The first one is to find a knowledge
representation language or formal language in which the
knowledge domain can be described. The second component of the knowledge representation problem is the one
that can perform automatic inferences for the user. The
third component is how to develop a knowledge base that
accurately represents the understanding of the domain
area.
The main sub-problems of knowledge representation
were summarized by Duce and Ringland (1988) as follows:
(1) Expressive adequacy: is a particular knowledge
representation scheme sufficiently powerful? What knowledge can and cannot particular schemes represent?
(2) Reasoning efficiency: like all representation problems in computer science, a scheme that represents all
knowledge of interest and is sufficient to allow any fact of
interest to be inferred by no means guarantees that it will be
possible to perform the inference in an acceptable time.
There is generally a trade-off between expressive adequacy
and reasoning efficiency;
(3) Primitives: what are the primitives (if any) in knowledge representation? What primitives should be provided
in a system and at what level?
(4) Meta-representation: how do we structure the knowledge in a knowledge base and how do we represent
knowledge about this structure in the knowledge base?
(5) Incompleteness: what can be left unsaid about a
domain and how do you perform inferencing over incomplete knowledge and revise earlier inferences in the light of
later, more complete, knowledge?
(6) Real world knowledge: how can we deal with attitudes such as beliefs, desires and intentions? How do we
avoid the paradoxes that accompany self-reverential
propositions?
Many psychological studies investigated different knowledge representation schemes used to develop structures
for representing the human experts' knowledge. The forms
of knowledge representation can be described with respect
to their relevance to human information-processing paradigm (Conway and Wilson, 1988). In general, there are
three forms of knowledge representation, i.e. (1) procedural; (2) propositional; and (3) analogical. The procedural
representation focuses on the control of the representation
within the context of some general structure of knowledge.

358
The procedural representation may also combine the
knowledge control and representation into a single form.
The propositional representation uses abstract content
patterns to organize the propositional information (like
schemata and frames). Other propositional representations of concepts may use the semantic nets as representation schemes.

6. Applications of expert systems


There have been a number of attempts to aid or even'
supplant the human CNC part programmer. Preiss and
Kaplansky (1984) have used AI techniques to convert a
part drawing to a part program for CNC milling. Kamvar
and Melkanoff (1985) have addressed the same problem in
the case of turning. Altom and Houtzeel (1987) describe an
ES that generates NC instructions for turning, milling,
drilling, taping, and boring from manual input or directly
from a CAD database.
The selection of tools, speed, and feeds has also received
attention. Barkocy and Zdeblick (1984) described a knowledge-based system to select cutting tools, pass sizes, and
speeds and feeds for the detailed planning of machining
operations. Wang and Wysk (1986) developed an ES that
generates efficient machining parameters. Melkote and
Taylor (1988) developed an ES that consults an external
commercial database to select milling cutters, feed rates,
and spindle speed. Chang et al. (1988) have developed an
integrated design/manufacturing/inspection system that
includes a part programming system for a machining cell.
The real time control of metal cutting whether by a
human or a computer is a natural ES application. Koval
(1987) proposes a future production facility that includes
intelligent machine control together with expert systems,
expert databases, and large scale simulation via CRAYclass supercomputers. Dressman et al. (1987) describe a
framework for ES control of metal cutting.

7. Development of the ITONUS system


The system developed and described in this paper is a
rule-based ES which aids in diagnosing problems during
lathe operations and suggests set-up parameters based on
user input about material, type of operation, and part
description. In response to system prompts the user
supplies information about the operation, material, part
shape, occurrence of chatter, chip color, etc. The system
then provides suggestions to correct a problem or to avoid a
potential problem. The ES is a result of 15 intensive
discussions and walk-throughs with an expert machinist
who also had a keen interest in computers and automation.

W a r d et al.

7.1. Knowledge elicitation


Turban (1990) notes that potential sources of knowledge
include human experts, textbooks, databases, special research reports, and pictures, and that these sources can be
divided into two types: documented and undocumented.
Since the primary short-term goal of this research was to
capture the domain expertise of a journeyman machinist,
only sources of undocumented knowledge, the sort that
resides in people's minds, was considered. The knowledge
elicitation process and prototype development using EXSYS is summarized (Sprague and Ruth, 1988) in five steps:
(1) identify subject area and find expert; (2) extract
expert's knowledge; (3) transfer knowledge into facts
which can be used as decision rules in the knowledge base;
(4) build the knowledge base, and (5) test problem.
Once an experienced and knowledgeable journeyman
machinist was identified, an initial domain description was
extracted using a questionnaire. Questionnaire items (Fig.
2 as applied to lathe machining) were used to stimulate
problem discussion by the domain expert (Waterman,
1986). The responses to these questions were analyzed and
the first interview planned. The first few interviews were
used to identify the main areas in lathe machining to be
defined for later development of specific concepts and
relations. These were the most difficult interviews because
of the difficulty in defining and narrowing the domain.
Each interview was tape-recorded and transcribed for
analysis and follow-up questioning. A major difficulty
encountered was in guiding the course of the interview.
This task was easier as the expert learned the capabilities
and limitations of the knowledge representation system
and his role in developing it.
An excerpt from an early interview is shown in Fig. 3.
Some interesting problems were encountered when the
sessions were transcribed. The expert's comments contained many partial sentences, contradictions, omissions,
and repetitions. The absence of body language sometimes
made the recordings difficult to interpret. These problems
complicated the procedure of formalizing and categorizing
the information given in an interview. A future improvement would be to videotape the sessions.
The early interviews also suggested the type of knowledge representation scheme and inference mechanism to

1.
2.
3.
4.
5.
6.
7.

Lathe machiningquestion
What are the inputs or problems?
What are the outputs or solutions?
Whichtypes of inputs cause difficultiesfor the expert?
How are the problems characterized?
How are the solutions characterized?
What sort of knowledgeis used?
How are problems or methods broken into smaller units?

Fig. 2. Initial questionnaire items as applied to lathe machining.

ITONUS: expert system for machining on a lathe


Questions (Q) by the knowledge engineer and answers (A) by the domain expert
Q. What kinds of lathe operations are especially difficult and what do you have to
pay special attention to when performing them?
A. Length of part always has an effect on how the machine operation must proceed.
The longer the material the greater chances are for chatter problems. Speeds,
feeds, tool height, and cutting depth also must be considered. Also, tolerances
that are very close add to the machining time.
Q. What are some common symptoms and conditions that you look for?
A. Chatter is a common problem in turning operations. It can be caused by a
number of conditions - flexibility of materials, tool set at wrong height, wrong
tool, feeds too fast or slow to match toohng, or a combinations of many things.
Q. What kinds of decisions do you have to make regarding these symptoms?
A. Chatter can be reduced in a lot of ways. A tool can be selected that has a smaller
radius on the point, or could be shaped differently. Cutting speeds may have
to be reduced or feeds could be increased, or decreased. Tooling height could
be raised or lowered.
Q. What are the consequences of the actions you take when a condition arises that
requires actions?
A. The consequences are that you begin to make well machined parts at a consistent
rate with very. few bad parts and with very little wear and tear on equipment.

359
protocol analysis sessions, it was possible to apply the
concepts and relations developed and construct a set of
rules to place in the shell.

7.2. Expert system shell


Analysis of eight PC-based shells resulted in selection of
EXSYS, version 3.2.5 (EXSYS, 1985). EXSYS uses
production rules for its knowledge representation. A
certainty factor can be associated with each rule outcome.
The rule base may be searched by backward or limited
forward chaining. It has a menu-driven knowledge-base
editor that provides some control over screen formatting.
Why/how questioning is available to the developer and
user, and there are related explanation and on-line help
facilities. Examples may be saved for use in sensitivity
analyses.

Fig. 3. Excerpt from an early interview.


7.3. Knowledge base

use. This resulted in the appropriate shell selection. A


rule-based, backward chaining system was indicated since
the expert's knowledge most often assumed the form of
rules-of-thumb and IF-THEN-ELSE statements. The
most common methods and actions performed involved
monitoring, diagnosis, and control.
In the final sessions before actual rules were developed,
and the knowledge base entered, specific areas in lathe
machining were identified for application to the expert
system. The prototype incorporated the areas shown in
Fig. 4. These machining aspects are dependent on several
factors, including type of operation, material, workpiece
shape, type of cut, and other variables. Protocol analysis
(step-by-step explanation of problem solving by the expert)
was used in the last interviews to produce specific concepts
and relations for actual rule development. The expert
explained his problem-solving approach to a variety of test
cases in lathe machining. Over the course of several

Area
Planning

Diagnosis

Aspect
tool selection
speed selection
feed rate selection
chatter recognition
chatter problem correction
avoidance of excessive tool wear
maintenance of finish and tolerance requirements
avoidance of heating problems

Fig. 4. Specific areas identified for application to the expert


system.

The knowledge base of ITONUS consists of three components: rules, qualifiers, and choices. Rules are the
IF-THEN statements which contain the relationships and
heuristics to be manipulated by the inference mechanism.
Qualifiers are conditions which comprise the IF portion of
the rules, and consist of a statement ending in a verb.
Several values may be assigned to each qualifier corresponding to different states of that qualifier. For example:

Qualifier

Value

the operation is

turning
facing
boring
parting

EXSYS permits values to be negated by preceding them


with NOT, and it permits the selection of two or more
values connected by the logical OR. Choices are the
conclusions which form the THEN portion of the rules.
The appropriate choices are printed at the end of a
consultation as a list of recommended actions to be taken
by the user. For example:

Choice
use a steady or follow rest
inspect tool
increase speed
chatter is possible

EXSYS rules are frequently of the form, 'IF qualifier value


THEN choice.' So that, based on the above examples, 'IF

Ward et al.

360

the operation is turning T H E N chatter is possible,' would be


a valid (though not particularly helpful) rule.
Two categories of machinist control: (1) planning or
set-up, and (2) diagnosis or troubleshooting, governed rule
development. Planning was considered to involve the
choice of initial feeds and/or speeds based on material type.
There are 46 rules9 Fourteen rules are related to speed
and feed selection for set-up based on material type and
operation. Fifteen rules relate to set-up conditions to avoid
chatter or what to do if chatter occurs. The remaining 17
rules pertain to control and diagnosis of certain problems 9
Advice to increase or decrease feeds and/or speed to avoid
wear, tool failure, and part quality is given9
A list of c o m m o n work materials was divided into
categories according to their effects on the tool. The hard
materials suggested a slower set-up speed, while less
demanding materials suggested a faster speed and/or feed.
Within these groups, specific recommendations are given
for each material. For different operations such as parting
or boring, or for workpieces with other than cylindrical
shapes, rules were made regarding tool position, possibility
of chatter, and need for extra workpiece support.
When rules pertaining to troubleshooting or diagnosis
fire early in an EXSYS execution, the values of qualifiers in
the IF part of the rules may not be known 9 When this
occurs, the user is prompted for the required value. For
example if a rule of the form ' I F Material type is Brass
T H E N . . . ' fires and the type of material is not known, the
user will be p r o m p t e d with 'Material type is' as shown at the
top of the consultation in Table 2. At present EXSYS has
19 qualifiers and each qualifier can take on from 2 to 13

Table 2. The following sample consultation is based on a steelboring operation9 The part is 41 cm long and 10 cm in diameter.
The workpiece is cylindrical, chatter is not present, and chips are
flowing freely. However, the load on the machine is increasing
and the chips are beginning to turn blue. Tolerances are still being
held and the status of tool wear is unknown. The bore diameter is
3.18 cm and the length of the bore is 15 cm. A rough cut is being
taken. Ellipses ( . . . ) indicate that the session has been edited by
deleting values.
Material type is
1 Cast iron
2 Bronze
3 Aluminum
4 Stainless steel
11 Forging
4
Operation is
1 Parting
2 Turning
3 Facing
4 Boring

Table 2--contd.
4
Please input part length
: 16
Please input part diameter
:4
Chatter is
1 Present
2 Not present
3 No machining has been performed yet
2
Workpiece is
1 Square or hexagonal
2 Cylindrical
2
Chips are
1 Flowing freely and consistently
2 Flowing poorly (long, 'stringy', non-uniform)
1

Ampere reading is
1 Increasing
2 Decreasing
1

Tool wear is
1 Present
2 Not present
3 Not known
4 No machining has been performed yet
3
Chip color is
1 Straw
2 Blue coming off the workpiece
3 Same as workpiece
4 Changing from straw to blue
5 No machining has been performed yet
4
Part tolerance is
1 Incorrect
2 Correct
2
Please input bore diameter
: 1.25
Please input bore length
:6
Cut is
1 Finish cut
2 Rough cut

1
2
3

Values based on 0 - 10 system


Inspect tool
For setup, select a speed that is slower than normal
Tool wear is likely

10
9
7

I T O N U S : expert system f o r machining on a lathe


values. The consultation in Table 2 asked for values for 10
of the 19 qualifiers.
Rules containing recommended corrective or preventive
action were developed considering the various input conditions. Possible reasons for problems such as tool wear are
given based on user input. Appropriate actions that are
recommended appear as choices in the THEN part of rules.
At present EXSYS has 27 available choices. Three of the 27
choices ('Inspect tool,' etc.) appear at the end of the
consultation in Table 2.
The recommendations given in the EXSYS execution
have values assigned from zero to ten. This numerical
confidence factor, a feature of EXSYS, is an average of the
confidence factors which appear in all the rules that fire.
The factor for each rule is selected by the domain expert.
These confidence factor values represent, on a scale of zero
to ten, the relative importance of a recommendation or the
likelihood of a problem occurring, depending on the
particular rule category.

7.4. Verification and validation


Verification and validation is a continuous process that is
applied during development for system improvement and
toward the end of development for evaluation of a particular prototype or improved version of the system (Turban,
1990). Difficulties that have been pointed out by Assad and
Golden (1986) are centered around these questions:
(1)
(2)
(3)
(4)

What characteristics should be evaluated?


How should performance be evaluated?
How should test problems be selected?
How should one evaluate the system's mistakes?

Because of the imprecise nature of the answers, the


solutions adopted may be a compromise between formal
methods and pragmatics (Brul6 and Blont, 1989). Santamarina and Salvendy (1989) have suggested solutions that
can be arranged in a hierarchy:
(1) System examination in which the expert examines
the system's knowledge and structure (Waterman, 1986).
(2) System refinement in which the expert provides
problems to be solved with the elicited knowledge (Waterman, 1986).
(3) System validation in which cases solved by the expert
and system are presented to other experts (Waterman,
1986).
(4) Performance feed-back in which the system is monitored in a controlled and limited environment (Gaines,
1986 and Brul6 and Blont, 1989).
ITONUS was subjected to system evaluation (item 1
above) during the last five interviews and system refinement (item 2 above) during the last two interviews.
EXSYS has a facility for testing which allows comparison
of one consultation (EXSYS execution) with a new one.

361
Old and new results are displayed together showing
changes in values and recommendations. This facility is
helpful in rule development. Development of the rules
evolved over the course of the interviews. Rules were
progressively modified, added or deleted. During the last
two interviews four test cases (Hall, 1988) were evaluated
and resulted in further modification of the rule base.
System validation (item 3 above) and performance feedback (item 4 above) can be expected to produce further
improvement.

8. Conclusions
ITONUS is an ES for lathe machining that represents a first
step in capturing and eliciting the unique knowledge of an
expert machinist. Basic concepts about planning and
control of lathe operations have been applied in an ES
environment. ITONUS gives practical advice on a limited
number of lathe operation scenarios. A non-expert, having
only a limited knowledge of lathe machining, can receive
suggestions for set-up, operation, and diagnosis of certain
common problems. Obvious limitations are the lack of
robustness of the knowledge base and the fact that only one
expert was consulted. Future work will aim to involve more
experts, especially those familiar with CNC machines;
rules will be developed to deal with specific types of sensor
feed-back and CNC programming concepts; the knowledge base will be expanded to include specific tool
selection, material, and speed/feed relationships for setup; testing will be done in an industrial setting; and further
attention will be given to human perception of heat effects.
Ultimately, the expert machinist as embodied in the expert
system will be used to redesign or replace the rule base and
controller of an FLC machine tool. The ultimate goal of
this research is to develop a metal-cutting process control
that will permit unattended but reliable operation of a
lathe.

Acknowledgements
The initial concepts and planning for the project was done
with support from the National Science Foundation research award R118610671. The actual research was funded
by a grant from the University of Kentucky Center for
Robotics and Manufacturing Systems. Special thanks are
due to journeyman machinist James Hartlage, the domain
expert.

References
Altom, B. and Houtzeel, A. (1987) M-TES-an expert system for
N.C. manufacturing, in Twenty Fourth Annual Meeting and
Technical Conference Proceedings - AIM Tech '87: Back to
Basics, Beloit, WI, pp. 73--4.

W a r d et al.

362
Assad, A. A. and Golden, B. L. (1986) Expert systems,
microcomputers, and operations research. Computers and
Operations Research, 13.
Astrop, A. (1978) Swedes show the way with unmanned line.
Machinery and Production Engineering, 132, 17-19.
Barkocy, B. E. and Zdeblick, W. J. (1984) Knowledge-based
system for machining operation planning, in Conference
Proceedings of the Sixth A UTO FA CT, Detroit, MI.
Belkin, N. J., Brooks, H. M. and Daniels, P. J. (1988) Knowledge
elicitation using discourse analysis, in Knowledge Acquisition
for Knowledge-Based Systems, Gaines, B. R. and Boose, J.
H. (eds), Academic Press, London, pp. 107-23.
Boose, J. H. (1986) Expertise Transferfor Expert System Design.
Elsevier, New York.
Boose, J. H. and Gaines, B. R. (eds) (1988) Knowledge
Acquisition Tools for Expert Systems. Vol. 2, Academic
Press, California.
Brul6, J. F. and Blont, A. (1989) Knowledge Acquisition,
McGraw-Hill, New York.
Buchanan, B. G. (1982) New research in expert systems, in
Machine Intelligence 10, Hayes, J. E., Michie, D. and Pao,
Y. H. (eds), Edinburgh University Press, Edinburgh, pp.
127-68.
Chang, S. and Zadeh, L. (1972) On fuzzy mapping and control.

1EEE Transactions on Systems, Man, and Cybernetics,


SCM-2, 30-42.
Chang, T. C., Anderson, D. C. and Mitchell, O. R. (1988) Q T C an integrated design/manufacturing/inspection system for
prismatic parts, in 1988 ASME Computer Engineering Proceedings, San Francisco, CA, pp. 417-26.
Cleaves, D. A. (1988) Cognitive biases and corrective techniques
for improving elicitation procedures for knowledge-based
systems, in Knowledge Acquisition for Knowledge-Based
Systems, Gaines, B. R. and Boose, J. H. (eds), Academic
Press, London, pp. 23-35.
Conway, T. and Wilson, M. (1988) Psychological studies of
knowledge representation, in Approaches to Knowledge
Representation, Ringland, G. A. and Duce, D. A. (eds),
Wiley, New York, pp. 117-60.
Dressman, J. B., Karwowski, W., Ralston, P. A. S. and Ward, T.
L. (1987) Framework for expert system for control of metal
cutting, in Proceedings of the IXth International Conference
on Production Research, Mital, A. (ed.), University of
Cincinnati, pp. 2579-92.
Duce, D. and Ringland, G. (1988) Background and introduction,
in Approaches to Knowledge Representation: An Introduction, Ringland, G. A. and Duce, D. D. (eds), Wiley, New
York, pp. 1-11.
EXSYS (1985) EXSYS expert system development package, in
User Manual, EXSYS, Inc., Albuquerque.
Gaines, B. (1986) An overview of knowledge acquisition and
transfer, in Proceedings of the Knowledge Acquisition for
Knowledge-Based Systems Workshop, Canada.
Gaines, B. R. (1988) An overview of knowledge-acquisition and
transfer, in Knowledge Acquisition for Knowledge-Based
Systems, Gaines, B. R. and Boose, J. H. (eds), Academic
Press, London, pp. 3-21.
Gaines, B. R. and Boose, J. H. (1988) Knowledge acquisition for
knowledge-based systems, in Knowledge Acquisition for
Knowledge-Based Systems, Gaines, B. R. and Boose, J. H.
(eds) Academic Press, London, pp. 8-19.

Greenwell, M. (1988) Knowledge Engineering for Expert Systems, John Wiley, New York.
Grover, M. D. (1983) A pragmatic knowledge acquisition
methodology, in Proceedings of the Eight International Joint
Conference of Artificial Intelligence, Germany, pp. 436-8.
Grover, M. P. (1980) Automation, Production Systems, and
Computer-Aided Manufacturing, Prentice Hall, New Jersey.
Gupta, M. M. and Sanchez, E. (eds) (1982) Approximate
Reasoning in Decision Analysis. North-Holland, Amsterdam.
Hall, W. D. (1988) ITONUS: Expert System for Metal Cutting on
a Lathe, Master's thesis, University of Louisville.
Hart, A. (1986) Knowledge Acquisition for Expert Systems.
McGraw-Hill, New York.
Hayes-Roth, F., Waterman, D. A. and Lenat, D. B. (1983) An
overview of expert systems, in Building Expert Systems,
Hayes-Roth, F., Waterman, D. A. and Lenat, D. B. (eds)
Addison-Wesley, Mass., pp. 3-29.
Inaba, H. (1980) Concept and practice of unmanned machines.
Metalworking Engineering and Marketing, 2, 54-60.
Johnson, P. E., Zualkernan, I. and Garber, S. (1988) Specification of expertise, in Knowledge Acquisition for KnowledgeBased Systems, Gaines, B. R. and Boose, J. H. (eds)
Academic Press, London, pp. 125-45.
Kahn, G., Nowlan, S. and McDermott, J. (1985) Strategies in
knowledge acquisition. IEEE Transactionson Pattern Analysis and Machine Intelligence, PAMI-7,511-22.
Kamvar, E. and Melkanoff, M. A. (1985) Automatic generation
of GT codes for rotational parts from CAD drawings, in

Software for Discrete Manufacturing, Proceedings of the Sixth


International IFIP/IFA C Conference, North-Holland, Amsterdam, pp. 259-72.
Kodratoff, Y., Manago, M. and Blytie, J. (1988) Generation and
noise, in Knowledge Acquisition for Knowledge-Based Systems, Gaines, B. R. and Boose, J. H. (eds), Academic Press,
London, pp. 301-24.
Koval, A. (1987) Expert systems in the workplace: the application
of expert systems technology to the industrial production
process. ASME Pressure Vessels and Piping Division, 126,
7-10.
Maiers, J. and Sherif, Y. S. (1985) Applications of fuzzy set
theory. IEEE Transactions on Systems, Man, and Cybernetics, SMC-15, San Francisco, CA, 175-86.
Melkote, R. and Taylor, D. L. (1988) Implementation of rule
based selection of milling cutters, feed rates and spindle
speed, in 1988 ASME Computer Engineering Proceedings,
pp. 427-32.
Negoita, C. V. (1985) Expert Systems and Fuzzy Systems.
Benjamin/Cummings, California.
Preiss, K. and Kaplansky, E. (1984) Automated CNC milling by
artificial intelligence, in Conference Proceedings of the Sixth
A UTOFACT, Detroit, MI.
Pressman, R. S. and Williams, J. E. (1977) Numerical Control and
Computer-Aided Manufacturing, Wiley, New York.
Ralston, P. A. S. and Ward, T. L. (1987) Fuzzy logic control of
metal cutting on lathes, in Proceedings of the 1Xth International Conference on Production Research, Mital, A. (ed.),
University of Cincinnati, pp. 2586-92.
Sakai, Y., Nakato, T. and Ohkusa, K. (1984) Identification and
failure-detection system for cutting process, in Proceedings

I T O N U S : expert system f o r machining on a lathe


of the IMEKO TK-14 International Symposium on Metrology
for Quality Control in Production, Tokyo, pp. 62-7.
Sakai, Y. and Ohkusa, K. (1984) On a control system for cutting
process, in Proceedings of the 16th CIRP International
Seminar on Manufacturing Systems, Tokyo, pp. 188-95.
Sakai, Y. and Ohkusa, K. (1985) A fuzzy controller in turning
process automation, in Industrial Applications of Fuzzy
Control, Sugeno, M. (ed.), North-Holland, Amsterdam, pp.
139-51.
Santamarina, C. and Salvendy, G. (1989) Knowledge elicitation
for fuzzy set based systems, unpublished technical report.
Schmenk, M. J. (1984) Acoustic emission and the mechanics of
metal cutting, unpublished ASME conference paper.
Shadbolt, N. and Burton, M. (1990) Knowledge elicitation, in
Evaluation of Human Work: A Practical Ergonomics Methodology, Wilson, J. and Corlett, E. N. (eds), Taylor &
Francis, London, pp. 321--45.
Shaw, M. L. G. and Gaines, B. R. (1984) An interactive
knowledge elicitation technique using personal construct
technology, in Knowledge Acquisition for Expert Systems: A
Practical Handbook, Kidd, A. L. (ed.), Plenum Press, New
York.
Sprague, K. G. and Ruth, S. R. (1988) Developing Expert Systems
Using EXSYS. Mitchell Publishing, California.
Stauffer, R. N. (1978) A new concept in flexible automation.
Manufacturing Engineering, 80, 66--8.
Turban, E. (1990) Decision Support and Expert Systems: Management Support Systems. 2nd edn, Macmillan, New York.
Uhlman, W. T. and Schmenk, M. J. (1981) Torque controlled
machining for numerical control machining centers, unpublished paper presented at IEEE conference.
Wang, H. and Wysk, R. A. (1986) Expert system for machining
data selection. Computers and Industrial Engineering, 10,
99-107.

363
Ward, T. L., Bronner, J. E., Ralston, P. A. S., Dressman, J. B.
and Hnat, W. P. (1988) Sensors for expert control of metal
cutting on a lathe, in Proceedings of the 1988 Symposium on
Advanced Manufacturing, University of Kentucky, pp. 3740.
Waterman, D. A. (1986) A Guide to Expert Systems. AddisonWesley, Mass.
Waterman, D. A. and Hayes-Roth, F. (1983) An investigation of
tools for building expert systems, in Building Expert Systems,
Hayes-Roth, F., Waterman, D. A. and Lenat, D. B. (eds),
Addison-Wesley, Mass., pp. 169-215.
Welbank, M. (1983) A review of knowledge acquisition techniques
for expert systems. British Telecom Research Laboratories
report.
Wick, C. (1977) Automatic adaptive control of machine tools.
Manufacturing Engineering, September, 38-45.
Wright, P. K. (1983) Physical models for tool wear for adaptive
control, in Computer Integrated Manufacturing.
Wright, P. K. (1983) Physical models for tool wear for adaptive
control. Computer Integrated Manufacturing.
Wright, P. K. and Bourne, D. A. (1988) Manufacturing Intelligence, Addison-Wesley, Mass.
Zadeh, L. A. (1965) Fuzzy sets. Information and Control, 8,
338-53.
Zadeh, L. A. (1973) Outline of a new approach to the analysis of
complex systems and decisions processes. IEEE Transactions on Systems, Man, and Cybernetics, 3, 28-44.
Zhu, J. Y., Shumsheruddin, A. A. and Bollinger, J. G. (1982)
Control of machine tools using the fuzzy control technique.
Annals of CRIP, 31,347-52.

You might also like