.^
;\
HD28
.M414
no.
ALFRED
A
P.
WORKING PAPER
SLOAN SCHOOL OF MANAGEMENT
Case-based System to Support Electronic
Circuit Diagnosis
Thilo Semmelbauer,
Anantaram Balakrishnan,
and
Haruhiko Asada
WP#
3497-92-MSA
October, 1992
MASSACHUSETTS
INSTITUTE OF TECHNOLOGY
50 MEMORIAL DRIVE
CAMBRIDGE, MASSACHUSETTS 02139
A
Case-based System to Support Electronic
Circuit Diagnosis
Thilo Semmelbauer,
Anantaram Balakrishnan,
and
Haruhiko Asada
WP#
3497-92-MSA
October, 1992
'^
J 7 1992
'
A Case-based System
Support
Electronic Circuit Diagnosis ^
to
7/7/70 Semmelbauer
Leaders for Manufacturing
M. T., Cambridge, MA
I.
Anantaram Balakrishnan
Sloan Scfiool of Management
M.
T., Cambridge, MA
I.
Haruhiko Asada
Department
M.
I.
of
T.,
Mechanical Engineering
Cambndge,
MA
October 1992
Supported by MIT's Leaders for Manufacturing Program
Abstract
Diagnosis and repair operations have traditionally been major bottlenecks
circuit
in electronics
assembly operations because they are labor intensive and highly vanable. Rapid
technological advances, shrinking product lifecycles, and increasing competition in the
electronics industry have
made quick and
and
efficient diagnosis critical for cost control
quality improvement, while simultaneously increasing the difficulty of the diagnosis task.
This paper describes a case-based diagnosis support system to improve the effectiveness
and efficiency of circuit diagnosis by automatically updating
and
the set of candidate defects,
The case-based system
prioritizing tests during the sequential testing process.
stores
individual diagnostic instances rather than general rules and algorithmic procedures;
knowledge base grows as new
faults are detected
system provides distributed access
features that
make
it
quick
to
manufacturing applications.
approach
in
and diagnosed by the analyzers. The
to multiple users,
and incorporates on-line updating
adapt to changing circumstances. Because
update, and integrate with existing systems, this
We
its
method
is
it
is
easy to
install,
well-suited for real
have implemented a prototype version, and tested the
an actual electronics assembly environment.
We
describe the system's
underlying principles, discuss methods to improve diagnostic effectiveness through
principled test selection and sequencing, and discuss managerial implications for successful
implementation.
Introduction
1.
Overview
1.1
Electronic circuit diagnosis
the process of identifying the component(s) that
is
is
responsible for the malfunction of a circuit board so that corrective (repair) action can be
Subsequent exploration of the problem's root causes motivates process and product
taken.
improvement
efforts.
The diagnosis and
repair loop has traditionally been a major
bottleneck in circuit board assembly. Recent trends in the electronics industry such as the
rapid technological advances, increasing product complexity, and shon product lifecycles,
have considerably increased the imponance of quick and efficient diagnosis, while
simultaneously increasing the difficulty of the diagnosis task. Reducing diagnosis time and
variability,
strive
and improving diagnosis accuracy becomes
critical as electronics
towards lean manufacturing, quick response, and greater
companies
flexibility.
This paper describes a case-based diagnosis support system that
we developed
consultation with a high-volume consumer electronics assembly plant.
The
in
plant, faced
with a rapid turnover of products and an increasing shonage of skilled technicians, wanted
a
means
to
improve
workload on
the efficiency
and effectiveness of circuit diagnosis, reduce the
few expert analyzers, and develop
its
An
devoting significant system development resources.
systems and current practice revealed
based and model-based systems
to
that,
new
a tool to train
although
technicians, without
informal survey of available
the literature describes several rule-
support diagnosis, these systems were not widely used
by the electronics industry partly because they require considerable time and special
expertise to install
proposed
It
in this
(e.g., for
knowledge acquisition) and update. The case-based approach
paper offers a practical solution to the diagnosis support needs of industry.
accumulates knowledge on-line as analyzers detect and diagnose new
faults.
Using
this
information about prior diagnostic instances, the system eliminates candidate defects during
the sequential testing process,
and
prioritizes
prototype version of the system, and tested
environment
We describe the
it
subsequent
in
tests.
We
have implemented a
an actual electronics assembly
system's underlying principles, outline methods to improve
diagnostic effectiveness by reducing the
number
ot tests
needed
to
complete the diagnosis,
and discuss implementation aspects.
1.2
Why
is
circuit
Rapid changes
diagnosis important?
in electronics
product and process technology, such as smaller, surface-
mounted components, higher board
circuit
densities,
and more complex
circuitry,
diagnosis more challenging. With fewer accessible connections and
line functional testing
have made
test
pads, in-
can only identify the approximate area or block of components on the
- 1-
board that
is
responsible for the malfunction. Detailed diagnosis must be done off-line
through extensive probing and measurement.
Each component and connection on
site
or
a circuit board
is
a potential defect site,
and every
can have several possible defects such as an open or short circuit (due to poor soldering
component misalignment), component malfunction, inherent flaw
placement of the wrong component, or defect
rate of a
in the
board containing hundreds of components
in the circuit design,
The
board's internal wiring.
failure
therefore, several orders of
is,
magnitude higher than the probability of a single defective component or operation, making
testing
and diagnosis inevitable even with highly capable placement and soldering
operations. Isolating the root cause in such
complex and dense
consuming. Annual expenses directly related
hundreds of thousands of dollars even in a
to
circuits is
arduous and time
diagnosis and repair can exceed several
medium-sized
facility.
Work
process
in
diagnosis stage and delayed feedback for process control further increase this cost.
importandy, the time and effort required to diagnose a faulty board
a
few minutes
to
is
at the
More
highly variable (from
over 4 hours) depending on the type of defect, the analyzer's
skill,
the
effectiveness of diagnostic tools, and the incentive and performance evaluation systems.
Consequendy,
medium
The
testing, fault diagnosis
and repair operations are often major bottlenecks
or high volume printed circuit board assembly
cost and effort to train technicians
electronics companies.
A new
is
also
potential root causes in
training costs
its
facilities.
becoming
a significant concern for
months of
technician requires several
experience with a particular board to understand the
in
training
circuit's cause-effect
manufacturing process. As product
life
and
behavior and the
cycles shrink, the
must be amortized over smaller volumes for each board. Diagnosis
bottlenecks are especially acute during the critical production ramp-up period following a
new product
introduction. Capturing market share early
is
essential for competitive
survival in the electronics industry, but production facilities often cannot achieve the
necessary rapid ramp-up because yields are low, and analyzers are least productive since
they have not had significant experience with the
new
product.
Fast and accurate diagnosis not only reduces direct manufacturing cost and flow time,
but
is
also an essential prerequisite for successful quality
improved product design
(e.g.,
improvement
efforts tiirough
Soederberg [1992]), better process understanding and
characterization, tighter process control, and close supplier pannership (e.g.,
[1991]).
The
ability to
provide quick information feedback from the diagnosis stage to
previous stages (including vendors)
is
particularly important in
medium and
environments. For instance, a snick glue pin or a worn placement head,
early,
Wenstrup
can produce hundreds of defective boards.
-2-
if
high volume
not detected
Computer support
1.3
for circuit
diagnosis
Manual diagnosis of circuit boards can be both
inefficient
and error-prone. Empirical
evidence (Rouse and Morris [1985], Rouse and Hunt [1986]) suggests that humans are not
optimaJ diagnosticians since they have difficulty
data to prune the
list
in effectively utilizing all the
of candidate defects and systematically select the
other hand, building a completely automated system that
diagnosing
all
is
test
available test
sequence.
On
the
capable of anticipating and
possible defects can be prohibitively expensive; such a system might also
require extensive time and effort by specialists (design and process engineers or software
experts) to install and maintain. Therefore, a system that assists, rather than replaces,
human
diagnosticians
is
more
practicable.
To cope
with the dynamic electronics
environment, the system must be robust and flexible,
i.e., it
should easily adapt to design
and process changes. Practitioners prefer a flexible system that possibly diagnoses only
the
most
offers
common
defect types (that account for a vast majority of board failures) to one that
comprehensive diagnostic coverage but cannot easily adapt
The case-based approach
knowledge-based systems,
offers the potential to
the case-based
to changes.
meet these requirements. Unlike other
system accumulates knowledge on-line by
storing diagnostic instances (as analyzers identify
and diagnose new
faults) rather than
general rules or algorithmic procedures. Each diagnostic instance or case corresponds to a
distinct defect;
for that defect.
it
is
represented by the set of tests conducted and their observed outcomes
By comparing
the test results for the current board with previous instances,
the system can eliminate candidate defects and determine the effectiveness of subsequent
tests.
Recording diagnostic instances rather than general rules has the
p)otential
disadvantage of requiring a large database. However, our prototype implementation
experience suggests that the vast majority of faults encountered
very small fraction of the set of
all
in practice
correspond
to a
possible instances.
Because of its simplicity, the system can be maintained by
the analyzers themselves,
avoiding long knowledge-acquisition delays and the consequent system obsolescence.
Furthermore, the standard format of a case promotes learning and information sharing
among
technicians (from different shifts or production lines), and enables the system to
serve also as a training tool for
system viable and useful for
new
real
technicians.
These features make the case-based
manufacturing applications. Although case-based
reasoning has been applied to other domains such as legal argument and customer service,
we
are not
aware of
its
previous appHcation to electronic circuit diagnosis.
We enhance
the
basic case-based approach to incorporate historical data and other prior information for
systematically selecting tests and optimizing the search process.
approach,
we implemented and
To
validate the overall
tested a prototype version of the system in a
high-volume
printed circuit board (surface-mount) assembly line producing a hybrid (analog-Kiigital)
circuit
board for a state-of-the-art consumer electronics product. Our implementation
-3-
permits distributed usage via a local access network. Within five weeks of installation, the
system achieved
90%
90%
diagnose
to correctly
coverage,
i.e.,
of the defects occuring
we
implementation expenence,
we propose
week. Based on our
in the fifth
offer suggestions to
diagnosis function; in particular,
enough diagnostic instances
the system had acquired
improve the management of
a proactive
the
management approach and
incentive schemes that emphasize the analyzer's process improvement rather than fault-
finding role.
The
rest
of
this
paper
organized as follows. Section 2 describes the circuit diagnosis
is
process in more detail, and develops the goals for an ideal diagnosis suppon system.
Section 3 discusses the underlying structure and operation of a basic case-based system for
verifying defects based on test outcomes. Section 4 presents system enhancements to
optimize
different defects.
We
sequencing problem
heuristic
using historical data and other prior information on the likelihood of
test selection
methods
present a
to
dynamic programming formulation of
minimize the expected number of
tests
for on-line test selection. Section 5 briefly
implementation and the lessons we learned from
this
the optimal test
per board, and outline
descnbes our prototype
implementation exercise, and
identifies directions for further work.
2.
2.1
The Diagnosis Process
In-line testing:
Capabilities and tradeoffs
Circuit board assembly lines typically contain in-line inspection and testing stages, after
placement and soldering operations
(see, for instance.
Noble [1989]
for a description of the
processing steps in printed circuit board assembly). The main purpose of these in-line
screening operations
is
to identify defective boards , although they also
have limited
diagnostic capabilities. For instance, visual and X-ray inspection can identify missing or
misplaced components and possibly certain solder defects, but are not effective or reliable
for dense boards with small
components. Furthermore, complete manual or X-ray
inspection can be very time consuming and expensive.
Two types of in-line electrical
—
in-circuit testing
In-circuit tests verify the functionality of individual
testing.
automated
isolate
tests are available
test
equipment
(see, for instance.
components on
Maunder and Tulloss
and precisely identify certain types of defects
(e.g.,
[1990]).
open or short
passive components), but require hard fixturing and programming that
for short production runs.
is
the board using
They can
circuits, defective
difficult to justify
Moreover, the increasing use of surface-mount technology and
the high cost of board "real estate"
make
in-circuit testing infeasible or limit their diagnostic
capabilities (Garcia-Rubio [1988]). In-circuit tests are also prone to
(see, for
and functional
measurement
errors
example. Chevalier [1992]) especially for small, high performance components.
-4-
leading to false acceptance or rejection of boards. Functional tests measure the circuit's
response
signals.
meet
board's output
(at the
pons or edge connectors)
to different
combinations of input
Functional tests can confirm whether the interactions between various components
the specifications; typically, they can trace a board's failure to the responsible circuit,
but not to a particular com]X)nent or connection in this circuit.
Selecting the testing strategy for a product involves addressing tradeoffs between
higher assembly cycle times (for
boards) to conduct extensive in-line diagnostic tests
all
versus higher off-tine diagnosis effon (for defective boards)
As
have high resolution.
the board defect rate decreases and
this tradeoff increasingly favors the strategy
We
the in-line tests
component density
increases,
is tiien
performed off-line by
next describe this off-line diagnosis process, and identify
opportunities to support this process using a computer-based system. For brevity,
use the
word "diagnosis"
to refer to the detailed, off-line diagnosis
has failed in-line (functional)
circuit
of the malfunction
procedure
corrective action
(e.g.,
for off-line
board that has failed functional
(e.g.,
we
will
procedure after a board
tests.
2.2 Sequential testing
Diagnosing a
do not
of using in-line testing only to identify
defective boards; detailed diagnosis and defect verification
skilled technicians.
when
diagnosis
tests entails identifying tiie
source
a compxDnent or connection on the board) so that appropriate
replacing or resoldering the component) can be taken to repair the
board Diagnostic information
also serves as the input for root cause analysis
and
continuous improvement of vendors, processes, and product design.
We define a dfifg^l as the finest grain
identified by the analyzer.
defects might require the
Each defect has an associated corrective action
same
problem
In
are different; therefore,
defects even though both require the
resisitor).
Each defect has a
by applying one or more
set
tests
.
;
different
corrective action. For instance, a defective resistor can act
as either a short circuit or an open circuit.
identify the
source of malfunction that can be unambiguously
same
each case, the symptoms and the means
we view
test is
two
possibilities as separate
corrective action (namely, replacing the
of associated
A
these
to
symptoms
that the analyzer
must discover
an action that must be performed either by a
technician or by automated test equipment, resulting in an observation about the behavior of
the board.
A test might consist of applying a specified set of electrical
certain circuit elements (e.g., tuning capacitors),
circuit
test
and probing selected locations on the
board to measure the voltage or view the electrical signal on an oscilloscope. Each
has two or more possible results or outcomes
or less than 4.5 volts); observe that
we
inputs, adjusting
discretize
it
if
the output
(e.g..
voltage
measurement
is
is
greater than 4.5 volts
continuous
(e.g.,
voltage)
by defining the range of continuous values associated with each outcome.
For convenience, and without loss of generality,
-5-
all
of our subsequent discussions assume
has two possible outcomes which
that
every
test
depends on the defect
test
We
in the board.
we denote
or
as
1
.
The outcome of a
permit defects to have indeterminate outcomes
for certain tests, e.g., for a certain defect, the voltage at a particular location
or
1
depending on, say, the
state
of a tlip-flop device.
We
may
be either
also recognize that the analyzer
might have only partial prior information regarding which defect produces what outcome
for each test; however,
we assume
that the available information
adequate to uniquely
is
identify each defect.
To
identify the true defect in a malfunctioning board, the analyzer uses a sequential
testing
procedure
the board's failure
between these
.
The process
mode
during functional testing) and a set of
The analyzer successively
defects.
observes the outcome
with a candidate set of defects (which depends on
starts
at
(i.e.,
each stage, and eliminates (from the candidate defect
the set contains
comprehensive
(i.e.,
can distinguish
and applies various
selects
defects that cannot produce the observed outcomes.
complete
tests that
all
If
tests,
set) all
the initial candidate defect set
possible defects), and the set of available tests
some combination of known
test
outcomes uniquely
is
is
identifies
each
candidate defect), the procedure terminates with the true defect as the only remaining
member of the
candidate defect
In practice, analyzers
might not always follow a systematic procedure
eliminate defects. Diagnosis
some
intuition,
circuit
set.
is
to select tests
and
often considered an "art", and analyzers rely largely on
knowledge, memory, and experience as they perform diagnosis.
Often, they do not formally maintain and update the set of candidate defects by comparing
the defects' expected test
outcomes with
the observed outcomes.
may
candidate defects and their expected outcomes
vary from technician to technician.
A common
not even be documented, and might
"myopic approach for troubleshooting
"
consists of hypothesizing a particular defect, attempting to verify
appropriate
tests,
revising the guess
if
Indeed, the set of
it
by performing the
the tests disprove the first hypothesis,
and so on.
Alternatively, the analyzer might follow a predetermined test sequence represented, for
instance, as a fault tree.
A purely manual diagnosis
when
it
is
process can be both inefficient and inaccurate especially
not supported by clear procedures and documentation. For instance, using a
static fault tree
can result
in
unnecessary
approach that uses recent defect history
tests
to
and unproductive search
(relative to an
guide the search). Likewise, diagnosis without
adequate documentation and formal recording procedures can lead to replication of
false identification. Also, these informal
transfering
knowledge
let
us
first
and
approaches are not conducive to learning,
to other analyzers, or generating appropriate information for quality
improvement. To identify the desirable characteristics of a computer-based system
analyzers,
tests
to assist
explore the available opponunities to improve diagnosis productivity.
-6
Diagnosis decision support requirements
2.3
Consider an intermediate stage
several tests and observed the results of each
questions,
(1)
all
Given
the
specific
(2)
Does
(3) If
aimed
at
when
diagnosis process
in the
the analyzer has applied
The analyzer now
test.
faces four related
deciding what to do next:
outcomes of all previous
tests,
can
we conclude
that the
board contains a
(known) defect?
the board contain a previously
unknown
more than one (known) defect remain
defect?
which available
as candidates,
test(s)
can
distinguish between these defects?
(4)
We
What
test to
apply next?
two questions
will refer to the first
as Test
as Defect Identification
and the remaining two
,
Selection.
Consider the information needs and actions for the defect identification questions
An
.
affirmative answer to either question takes the board out of the regular diagnosis iteration.
In the first case, the action consists of repairing the identified defect, possibly after
confirming the diagnosis.
outcomes
(i.e.,
the
answer
If
no known defect causes
to question 2 is affirmative),
board contains a new defect, or
inaccurate, or
(iii)
the
(ii)
the actual test
the
assumed
test
observed cumulative
we have
outcomes
outcomes were measured
test
three possibilities:
for the
known
incorrectly.
the
(i)
defects are
In all tiiree
simations, the board requires extraordinary actions such as confirming previous test
outcomes, consulting with design and process engineers, checking with suppliers, and
updating current knowledge.
If
multiple (known) defects remain in the candidate
seek to identify relevant
tests,
no such
test, if that test
test is available, the
to
first
this
process corresponds to determining, for
has different expected outcomes for the candidate defects.
analyzer must develop and apply one or more
distinguish the defects. Otherwise,
perform next. As we
the test selection questions
not yet performed, that can distinguish between the
remaining possible defects. Conceptually,
each available
set,
we must decide which among
shall see later, this decision
new
If
tests to
the available tests to
impacts the number of iterations needed
complete the diagnosis.
Computer suppon
logical reasoning
for the defect identification questions
system that automatically compares the observed
expected outcomes for
outcomes. Similarly,
relevant tests
at
might consist of a database or
all
test
outcomes with
the
candidate defects, and eliminates defects with inconsistent
to support test selection, a
computer-based system can identify the
each stage, and apply a systematic method to select the next
effective, the diagnosis support
test.
To
be
system must meet cenain key requirements such as ease of
7-
initializing
features of a computer-based system to
•
Development
a system that
•
We
and updating, and quick response.
effort:
suppon
list
below these and other desirable
electronic circuit diagnosis:
the short product lifecycles of electronic products,
Given
becomes operational quickly and without
The system must be easy
Maintainability/ robustness:
to update, preferably
(e.g.,
by the users
software expens
engineers) often introduces delays, causing inconsistencies and system
obsolescence. Another important requirement
to the frequent design
•
require
significant expense.
themselves; relying on external resources to update the system
and knowledge
we
and process changes
is
the system's ability to adapt gracefully
that characterize the electronics industry.
Compatibility: Technicians often view structured tools and methodologies as
impediments
to creativity; compatibility with current practices
organization
is,
therefore, critical to the success
ergonomics, ease of use,
flexibility,
and manual
and the culture of the
and effectiveness of the system. System
ovemde
options are
some of
the
ways
to
improve compatibility.
•
Response
Since each board requires several
time/efficiency:
candidate defect
set,
the system
does not become the botdeneck
•
Effectiveness:
must respond rapidly
test iterations to
to these questions so that diagnosis
activity.
The main purpose of a diagnosis support system
time per board. Part of
this
prune the
savings
comes from eliminating
of the routine bookkeeping and lookup functions.
Two
is to
errors
reduce the diagnosis
and automating some
other features can reduce
diagnosis time:
•
reducing the (expected) number of
tests
through systematic search and principled
test
selection and sequencing; and,
•
providing good "coverage",
i.e.,
when
the system has stabilized,
it
must correctly
diagnose a large percentage of the defects.
Other desirable system features include
prepare periodic
summary
abilities to: interface
with existing databases,
reports for process improvement, highlight inconsistencies and
faciUtate learning, diagnose both digital and analog (e.g., radio frequency) circuits, and
operate in multi-user environments.
The next
section uses these performance criteria to
motivate the case-based method.
3.
3.1
The Case-based Approach
Knowledge base
for Circuit Diagnosis
for diagnostic
systems
Circuit diagnosis requires three broad classes of
knowledge
—
historic, heuristic,
and
fundamental knowledge. Historic and other prior information on the likelihood of defects
can help the analyzer prioritize candidate defects, thus enabling an effective choice of
sequence. For example,
if
a
placement machine frequentiy misplaces a particular
8-
test
part, the
analyzer might
first
examine
that
pan before pert'orming other
knowledge, sometimes termed "shallow" knowledge, refers
tests.
to
Heuristic or empirical
an empincal association
between observed symptoms and diagnostic conclusions. This category might include
rules
and procedures, as well as
effective after
shoncuts that the analyzer has found
tricks or
to
be
some troubleshooting expenence. Fundamental or "deep" knowledge
implies an understanding of the underlying physics of circuit elements which
we
use to
predict circuit response. Design and test engineers rely on such fundamental knowledge,
often encoded in a circuit simulator, to design and troubleshoot electronic circuits.
Depending on
the primary source of diagnostic
knowledge, we can distiguish between
experience- based (empirical knowledge) and model-based (fundamental knowledge)
systems to support circuit diagnosis. The model-based approach uses an analytical model
or a circuit simulator to dynamically identify the potential causes for the observed
symptoms during
(i.e.,
the sequential testing process.
Since diagnosis entails "backtracking"
given ceruin observed symptoms or outputs, what values of circuit elements can
produce these outputs), we need a translator
(that characterize the output for a
conven
to
traditional circuit design
models
given circuit configuration) into corresponding diagnostic
models. Davis and Hamscher [1988]
illustrate the three basic steps
generation, hypothesis testing, and hypothesis discrimination
—
in
— hypothesis
model-based diagnosis.
Gensereth [1984], deKleer and Williams [1987], and Hamscher [1988] describe various
model-based systems
a pproach
is
to
diagnose digital
circuits.
The most common experience -based
a rule-based (expen) system that relies on the experience of
develop if-then rules that enable the system
to
human
reason about the behavior of
malfunctioning board (see Semmelbauer [1992] for an illustration of
this
experts to
tiie
approach).
of the pioneering apphcations of the rule-based approach was for medical diagnosis
see
Harmon and King
(e.g.,
[1985]). Unlike the field of medicine, however, electronic circuits
and manufacturing processes vary widely, and
radical changes.
One
In general, the literature
the technologies undergo frequent and
on knowledge-based diagnostic systems focusses
on knowledge representation issues and inference mechanisms
to identify the
underlying
source of the problem, and does not adequately address issues of optimizing the sequential
search process.
For electronic
circuit diagnosis, both the
certain advantages, but also
framework
that
impose
overcome
By
limitations.
combines both approaches
respective strengths and
model-based and rule-based approaches offer
in a
considering a decision suppxsrt
complemetary fashion, we can exploit
the disadvantages.
Model-based systems offer
their
the
advantage of direcdy integrating design and diagnosis: when the designer changes the
circuit
is
schematic (and adds relevant sub-models for
new components),
the diagnosis
model
automatically updated, permitting instantaneous troubleshooting capability without
requiring any
human experience
with the
new
circuit
-9-
(Davis [1988]). However, because
they are very computationally intensive, model-based diagnostic systems are not wellsuited for on-line applications. Furthermore, technical difficulties
suitable diagnostic
models
Sugaya [1986], Tong
for analog circuits (see, for
et al.
still
remain
building
in
example, Bennetts [1981], Kritz and
[1989]) due to bi-directional signal propagation, and the impact
of physical layout and tolerances on the performance of high speed radio-frequency
circuits.
Dealing with bridge defects and improper (open) connections also poses problems
since the basic circuit structure itself changes (Priest and
Welham
[
On
1990]).
the other
hand, the rule-based approach can incorporate heuristic knowledge to deal with these
complexities, and provide quick on-line response during diagnosis; the infrastructure of
tools to
implement a rule-based system
also quite well-established.
is
expert systems might require long development times (Priest and
especially for
initial
knowledge acquisition
(e.g.,
However,
Welham
traditional
[1990]),
Newstead and Pettipher [1986]).
Furthermore, since each component and connection on the board can cause malfunction,
anticipating and encoding rules for every possible defect
instance, the initial
covered
less than
knowledge
20%
of
we
that
weeks of system operation.
almost impossible. For
acquired for one board to build a prototype system
the defect types that
all
is
were encountered during the
Finally, a system that relies
engineers to update the knowledge base (for instance,
first five
on software experts or knowledge
when
the circuit design or
manufacturing process changes) can introduce expensive delays.
To cope
with the rapid pace of change
in the electronics industry,
we
require a
diagnostic system that can both adapt easily to the changing environment, and reduce
we
diagnosis time in production by providing quick on-line response. Next,
case-based approach that
is
similar to rule-based systems in
its
describe a
on-line operation, but can
exploit the capabilities of model-based systems to adapt to changing circuit designs and
new
defects.
Elements of case-based reasoning
3.2
The goal of our
project
was
to
develop and
test a
viable in a practical manufacturing environment.
diagnostic support system that
Two
is
important observations regarding
practice motivated our proposed case-based approach:
1.
The ubiquitous 80-20 law
of
all
applies also to electronic circuit diagnosis,
possible defects occur in over
80%
i.e.,
of the defective boards. Figure
less than
1,
20%
which shows
the cumulative relative frequencies of various defects (arranged in decreasing order of
occurence) for our prototype system, clearly
20
characteristic, a
illustrates this
phenomenon. Given
this 80-
system that can operate with incomplete knowledge progressively
,
adding defects as they are encountered,
is
preferable to enumerating, a priori,
all
possible
defects. This strategy has four important advantages: (a) the developmental effort and
time
is likely to
be significantly lower;
(b)
focusing on the relatively few defects that have
-10-
occured
in the past
can significantly reduce the response time;
(c) since
it
can tolerate
incomplete knowledge and can be updated on-line, the system can easily adapt to changes
human expens
process and product design; and (d) by relying on
in
defects, the system
diagnose new
non-threatening and hence likely to be well-accepted.
is
experience (reported
to
in
Our
Section 5) indicates that the strategy of progressively adding
defects as they are encountered provides over
90%
new
coverage within just a few weeks of
operation.
2.
A
simple
way
performed and
defect.
from an expert analyzer
to "learn"
their respective
outcomes
The observed outcomes
board characterize the
new
for
all
—
system maintainability, and as a means
to
Case-based reasoning (CBR) originated
An
"case").
the
early system,
HYPO,
operated
system combines reasoning about the
it
a convenient
to clever searching
area of legal argument (hence, the term
in the
common
law domain of trade-secret law;
statutes (rules) with the precedents (cases) that
to the current situation
of the case base on a number of significant
example, Rissland and Skalak [1988]). Applications of
CBR
Lee and Liu [1988] apply
Tokumaru [1990]
for
in the
attributes.
to robotic
assembly
CBR to
The
by the researcher.
Legal argument continues to dominate the most recent case-based reasoning
limited.
new
communicate among expens.
"closest" cases with respect to these attributes are selected for perusal
for
the tests
knowledge representation
concern them. The problem of finding the case(s) appropriate
amounts
—
as the analyzer explores and identifies a
the tests that the analyzer applied to diagnose the
making
defect,
by recording the actions
is
literature (see,
the fault diagnosis are
cell diagnosis,
and Ishida and
present a generalized approach to case-based diagnosis using the frame
theory. For electronic circuit diagnosis, the case-based approach offers the attractive
features of compatibility and incremental, on-hne
tolerate
incomplete knowledge, and
its
knowledge
acquisition; the system can
coverage increases with time.
We enhance the
conventional case-based approach by adding features to incorporate:
(i)
robust knowledge acquisition: the system identifies incomplete and inconsistent
(ii)
knowledge
(e.g.,
verification
and updating steps
distributed usage:
due
to
measurement
errors or design changes), and incorporates
to correct these errors;
by implementing
the system in a distributed client-server
environment with a shared database, several users can simultaneously access and use
the system; and,
(iii)
effective test sequencing: the
system reduces the number of
tests
by applying
test
selection heuristics that account for prior probabilities of different defects.
In the electronic circuit diagnosis context, a
experience, represented by
its test
case
outcomes. (A
is
rule,
an instance of a diagnostic
by contrast,
is
a generalization about
diagnostic experience gathered over time.) Each case corresponds to a particular defect.
-
11
-
The case base
consists of the collection of instances corresponding to different defects
Recall that each defect has either a deterministic or indeterminate
observed
in the past.
outcome
for every test; funher, the value of a deterministic
1)
we
or unknown. For a given defect,
signature of
We
visualize the case base as a n x
that defect.
diagnosed case or defect. The
the
i.e.,
expected outcome of
test
j
of expected outcomes for
m
as
all tests
available tests.
in the case
this
for defect
i
row has
base represents the signature of the
a 0,
1, 1,
U
or
indeterminate, or unknown.
is 0, 1,
knowledge
are detected and diagnosed (incremental
i
depending on whether the
We
user to dynamically add rows (cases) and columns (tests) to the case base as
3.3
(0 or
m matrix, with each row corresponding to a previously
row
i
element of
j
outcome might be known
Suppose we have n defect types and
the
defect type,
^r to its set
r.
permit the
new
defects
acquisition).
Diagnosing a defective board: Successive refinement through
case matching
Diagnosing a defective board corresponds
performing
and observing
tests
their
to gathering
enough information (by
outcomes) about the board's signature, while
simultaneously searching the case base, to identify a matching case.
found, the board contains a
new
If
no matching case
is
defect or an inaccurate case. In either situation, the case
base must be updated to reflect this
latest
experience.
We
remark
that the general case-
based approach goes beyond case matching and database updating functions to possibly
include reasoning and extrapolation.
capabilities. Figure 2
shows
support system (CDSS).
the
We
Our implementation does
not exploit these additional
complete flow chart for our case-based
first
explian
to update the case base (Section 3.4).
how
circuit diagnosis
the system operates before describing
how
Section 4 descnbes enhancements to improve the
system's test sequencing capabilities, and Section 5 discusses
how
to integrate this
system
with a model-based approach.
At each stage of
consisting of
tests
the search process, the system maintains:
known
all
performed thus
defects
far,
and
(ii)
whose
(i)
a candidate defect set
signatures match the observed outcomes for
the available test set containing
all tests that
all
the
have not yet
been perfonned but whose outcome differentiates one or more candidate defects from the
remaining defects
in the
analyzer chooses a
outcome of
next
test).
test
candidate
from
In the basic version
of the case- based system, the
the available test set, performs this test, and observes the
the test. (In Section
When
set.
4 we describe system enhancements
the analyzer enters the observed
outcome
to optimally select the
into the system, the candidate
defect set and available test set are updated as follows:
•
remove from
test differs
for this test
the candidate defect set
all
defects
from the observed outcome. Note
is
unknown
or indeterminate;
12-
whose expected outcome
that
we
retain all defects
for the latest
whose outcome
•
delete the test that
whose outcome
is
performed from the available
was
just
the
same
(or
in the updated candidate defect
If the initial
case base
is
unknown
or indeterminate) for
all
remove any
test
the remaining defects
set.
comprehensive
(i.e.,
contains
enough information regarding expected outcomes
accurate, the process
test set; also,
must terminate with
possible defect types, and has
all
to uniquely identify each defect)
and
the candidate defect set containing only the true
defect.
To
illustrate this process,
Figure
consider the simple 5-component, series network shown
Suppose we have a board whose abnormal output signal for
3.
indicates malfunction in
board are good); for
i
=
one of
1,2
the 5
5,
components
defect
i
the functional test
the intermediate connections
(all
refers to a malfunction in
component
on
the
We
i.
have
four available tests that apply a stimulus at the input terminal, and measure the voltage
each of the four intermediate connections (see Figure
abnormal voltage, indicates malfunction of one of
to
An outcome
3).
in
at
of 0, corresponding
we
upstream components. Thus,
the
can represent the signature for each defect as a vector containing 4 binary elements. Defect
1
has the signature
shows
{0000},
defect 2 has signature {I
the case base containing the complete signatures for
does not have any indeterminate outcomes, and
(i.e.,
he
test 3,
3
first
applies test 4
and so on
(i.e.,
component
answer the defect
At
the
that the technician
(i.e.,
initially
Table
5 defects. This
1
example
assume complete information
end of
always selects the
test
3 is defective).
identification
and
test set at
test selection
each stage; the columns of
questions of Section
defects
{
1, 2,
i.e.,
this table
2.
the third iteration, the candidate defect set contains a single defect (defect
sequence eliminates from the available
(for this
then
5),
Table 2 traces the successive states of the system,
and the available
just applied) at each step
outcome
sequence,
conclusively identified. Suppose a board contains defect
proving conclusively that component 3 on the board
chosen
tests in reverse
measures the voltage between components 4 and
until the defect is
the candidate defect set
3),
we
all
so on.
no "unknown" outcomes).
For simplicity, assume
i.e.,
000), and
However,
if
we had
test set
first
same outcome
only one
applied
example) automatically eliminates
3} have the
defective. For this example, the
is
test
test 3
4 since
test (the test that
instead of test 4.
all
was
its
"0"
remaining candidate
"0" for test 4.
This example motivates the following three imponant observations:
1.
Exploiting prior information:
Suppose past history or a
priori information (e.g.,
problems) suggests that the likelihood of defect 3
defects.
Since
tests 2
and 3 are adequate
knowledge about vendor quality
is
significantly higher than the other
to isolate defect 3, the analyzer
-13-
might choose
to
perform these two
tests first.
would require only two
For our sample board containing defect
3, this
complete the diagnosis (compared to the
iterations to
strategy
3 iterations
required for the static reverse-order sequence). For complex circuits, this savings in
number of
tests
how
can be significant. Section 4 discusses
other prior knowledge to dynamically select an effective
2.
to exploit historical
and
sequence.
test
Incomplete knowledg e:
Our example assumes
know
the expected
UU
{0
U}
outcome)
(U
is
if
U}
1
outcome of each
we know
the complete signature for each defect,
3.
(the
1
element "U" denotes an unknown
produce a "1" outcome for
case base and our original reverse
outcome
not eliminate defect
1
same
(shown
as before
in
Table
we have one
outcome of
test
2 for defect
available test (test
1
(in this case,
the expected
1) that
we
outcomes
process and reduce the number of
the partial signatures,
particular,
However,
two defects (defects
can eliminate defect
will observe a "1"
defect
Also note
tests.
I's
1.
We
1
at
and
must
outcome, eliminating defect
Thus, complete
3.
for each defect can accelerate the pruning
that,
we can improve our knowledge
we can change
unknown). Therefore,
is
1
order to conclusively prove that the board contains defect
knowledge about
2).
for test 2 during the third iteration eliminates defect 2 but does
(since the
necessarily perform test
1) in
similarly,
sequence, the set of candidate defects
test
the start of the fourth iteration, the candidate defect set contains
and
1;
between the 5 defects.
for the first 3 iterations remain the
3),
test
Table 3 shows an "incomplete" case base containing adequate
to distinguish
observing a "1"
we
2 and 3 are the only essential tests to uniquely
tests
i.e.,
i.e.,
every defect. However, the partial signature
test for
that all other defects
isolates defect 3,
knowledge
this
we know
sufficient to isolate defect
identify defect
Using
that
even though we staned with only
after this diagnostic exercise.
expected outcome
In
for test 2 from "U" to "0" in the
case base, potentially improving future performance.
3.
Multiple defects:
The
signature of a defect represents the anticipated test outcomes assuming that the board
contains only that defect.
in
What happens
if
a board contains multiple defects? Suppose,
our example, the board contained defects
2, is still correct, i.e.,
replace
component
3,
component
we must
functioning of the board.
3
is
3
and
1.
The
However,
defective.
again apply the functional
If the
diagnosis, illustrated in Table
we
after
test to
verify defect 3 and
confum proper
board malfunctions, the diagnosis process must be
repeated; in this example, the second round will detect defect
Although our example used a very simple
circuit,
it
has served to illustrate several generic
concepts relating to effective diagnosis. Real circuits contain
feedback loops, bidirectional signals,
tests that
- 14
1.
many
complexities such as
probe the same location but with multiple
measurements, stimuli, and component
to these
complex
and so on. Our methodology applies even
settings,
circuits.
knowledge and
3.4 Updating the case base: Incomplete
inaccuracies
We
have already seen one opportunity
performance: suppose
we
identify defect
i
to
update the case base and improve subsequent
end of the diagnosis process, and suppose
at the
the current case base contains only the partial signature for defect
performed during the current diagnosis whose outcome
unknown, we can replace
outcome
defect
the
unknown
We
updating step
j
known
is
to
is
augment defect
i's
if
j
was previously
have a deterministic outcome for
will refer to this updating step as signature
not valid
i
test
entry in the case base with the actual observed
for the current board (assuming test
i).
for defect
Then, for every
i.
augmentation
the board contains other defects besides defect
signature only after repairing
it
Note
.
i;
that this
hence,
we can
and verifying the proper functioning of
the board.
Our discussion
covers
all
in
Section 3.4 assumed that the case base
possible defect types) and accurate.
If
is
the case base
comprehensive
is
(i.e., it
incomplete or inaccurate,
the matching and successive defect/test elimination process might not necessarily terminate
with a single (truej -efect in the candidate defect
(i)
The process terminates because
of observed
test
set.
There are two
the candidate defect set
outcomes does not match with
is
empty
the signature of
possibilities:
,
i.e.,
the combination
any known defect.
This situation could arise either because:
(a) the
board contains a new defect
(b) at least
that
one of the defect signatures recorded
data entry errors, design changes,
(c) the
had not occured previously;
in the case base is inaccurate (due to
etc.); or,
observation and measurement of the acaial
test results
was erroneous.
In all three cases, the board requires "extraordinary" actions to identify the true
symptoms and cause of the problem;
these actions might entail detailed probing and
experimentation by the analyzer as well as consultations with design and
production
staff,
and component vendors.
defect, its signature in the case base
If
the board
new
defect,
we must add
refer to this step as case addition
for that defect,
.
found
a
new
.
On
if
engineers,
known
to contain a
must be checked and corrected,
refer to this ujxiating step as signature correction
contains a
is
test
necessary.
the other hand,
if
the board
case (row) corresponding to that defect.
The new case contains
the expected test
who diagnosed
defect,
the defect or revised the signature, the
corrective (repair) action required, the defect category, and
who
to notify (or
how
to
use) for quality improvement. Although the existing tests suffice to distinguish this
-15-
We
outcomes
and possibly some auxiliary information such as the name of the
the analyzer or engineer
We
new
the
to
(ii)
defect from the previously
empty defect
set
known
defects (the observed test
provide adequate "proof for
add one or more new
The diagnosis process
this defect), the
can quickly isolate
tests that
outcomes
that led to
analyzer might wish
this defect.
terminates because the available
test set is
empty but
the
,
candidate defect set contains more than one possible defect. This occurrence indicates
erroneous
test
"Inadequate
measurements, inaccurate defect signatures, or inadequate
tests" refers to the situation
between the candidate defects
tests are required.
perform
that test or if
its
delete a test
expected outcome
indeterminate or unknown) for
all
is
from
test
the
outcomes
are
unknown
the available test set
same (same known
in the
or
new
when we
value,
the remaining candidate defects. Again, this
stopping condition (with multiple candidate defects and no remaining
extraordinary actions to verify the
tests.
the current tests cannot distinguish
many
either because
we
Recall that
when
tests)
requires
measurements, check the defect signatures stored
test
case base, complete the partial signatures of the remaining candidate defects, or
add new
tests that distinguish
between these defects.
We
refer to this process as test
addition.
Figure 2 contains a flow chart summarizing the case matching and updating process. This
process includes a defect verification step
at the
end of a successful diagnosis exercise.
Verification entails confirming previous observed outcomes and applying any tests that
remain in the available
defect, a
new
test set; if
one or more outcomes conflict with the hypothesized
case must be added. This optional step
is
especially important after design
changes and during production ramp-up when many new cases are added
During normal operation, defects are automatically verified
defect, the board passes the functional test.
A
to the case base.
after repairing the identified
if,
board failing the
defect verification to determine whether the original diagnosis
test
again requires explicit
was erroneous or
if
the board
contains multiple defects.
In
summary,
robust,
the case-based approach to support circuit diagnosis permits incremental,
and distributed knowledge
the case-based system
knowledge
symptoms
defects.
update.
acquisition
.
Unhke
other knowledge- based approaches,
employs on-line and incremental,
acquisition. All cases have the
rather than batch-oriented,
same format; a case contains information on
that the user identifies to be relevant in defining
Because of the standardized format,
the case base
the
and distinguishing between
is
easy to understand and
The system begins with incomplete knowledge and poor
defect coverage, but each
unsuccessful diagnosis initiates a process of case-building to ensure subsequent coverage
of the same defect. Our experience suggests that the case coverage increases very rapidly
during the
first
few weeks. By only maintaining
the diagnostic
knowledge
that
is
needed,
the system can provide quick resix)nse. Incompleteness and inconsistencies in the case
16
base are easy to detect: they result
in
unusual termination of the diagnosis process (empty
candidate defect set or empty available
—
steps
make
The various correction and updating
test set).
signature augmentation, signature correction, case addition, and test addition
the
knowledge acquisition process robust and adaptive. The case-based system can
be implemented using a shared relational database
in a
multi-user computing environment,
thus providing wide access to analyzers at different locations as well as design, test and
process engineers and production supervisors. This implementation also allows the case
base to be integrated with other factory information systems, conveying current vendor,
design, and process problems to the diagnosis operation, and feeding back defect
information for quality improvement. The system can also keep track of the cumulative
relative frequencies
and trends of different defects; we next describe how
relative frequencies to guide
4.
and optimize the sequential
to use these
testing process.
Optimal Test Selection
By automating
each
the search and updating functions, the basic
iteration of the diagnosis process.
another means to significantiy reduce the
Section 3 illustrates, the
CDSS
reduces the time for
Decreasing the number of iterations provides
diagnosis time per board.
total
number of iterations depends on
As our example of
the test sequence.
We,
therefore,
require an effective test selection policy or rule: at each stage of the diagnosis process, the
method helps us decide which
thus
We refer to
far).
available test to apply next (based on the observed outcomes
minimizes the expected number of
a test selection rule that
the optimal test selection policy. This section presents a
tests as
dynamic programming formulation
for the optimal test selection problems, and describes on-line heuristic rules that the basic
CDSS
can incorporate
to
reduce the expected number of
tests.
These
rules rely
on
prior
information concerning the likelihood of different defect types, estimated from historical
data on defects (relative frequencies and trends), knowledge about specific problems at the
vendor
site
changes,
4.1
or in the assembly process, inputs from design engineers on
new
engineering
etc.
Test selection example
We first consider a simple example
sequencing on the number of
to better
understand the impact of
that defect,
i.e., test
outcome
for all other defects (or vice versa).
test (test
1
in
Figure 3
is
and
prove the defect. Suppose a board contains
tests required to
one of n potential defects, and each defect has a single associated
and sufficient to "prove"
test structure
i
test that is
has a "1" outcome for defect
We
i,
and a "0"
will refer to this type of test as n focused
a focused test). Thus, the
complete case base for n focused
corresponds to a n x n identity matrix. Suppose n =
5,
and the
five possible defects
the following probabilities of occurence: 0.1, 0.1, 0.2, 0.2, and 0.4.
the tests in order of decreasing defect probabilities
17
both necessary
tests
have
Intuitively, applying
minimizes the expected number of
tests.
For our example, the "decreasing probability sequence" corresponds
reverse index order
tests,
say Ej, using this strategy
Ej
On
1 )
until the defect is identified.
to
applying the
tests in
The expected number of
is:
=
1x0.4 + 2x0.2 + 3x0.2 + 4x0.1+5x0.1
=
2.3 tests per board.
the other hand, for a static policy that applies tests in the forward sequence 1-2-3-4-5.
the expected
number of tests E2
E2
The optimal
a
5-4-3-2-
(i.e.,
43%
1x0.1+2x0.1+3x0.2 + 4x0.2 + 5x0.4
=
3.7 tests per board.
strategy saves,
savings.
Thus, optimal
is:
=
on average,
(A more skewed probability
test
for quality
Because we assumed
that
optimal testing policy that
easy to implement,
each defect has a focused
is static (i.e.,
tests in
measurement,
etc.)
and eliminate only one defect
tests
jointly identify
Joint tests reduce the
test,
our example has a simple
(because of constraints on probing,
at a
analyzers rely on a combination of focused a.nd joint
3).
reduce diagnosis cost, and
decreasing order of defect probabilities.
tests are infeasible for certain defects
combined outcomes can
to
does not vary with the observed outcomes) and
Since focused
both necessary and sufficient, joint
second sequence,
improvement.
sequence the
i.e.,
to the
distribution can give even greater savings.)
sequencing has considerable potential
provide quick feedback
Figure
board relative
1.4 tests per
time (except
tests.
at the final step),
Unlike a focused
cannot isolate defects individually but their
and prove the defect
number of required
under certain circumstances (depending on
tests to
their structure
(e.g., tests
diagnose
2 through 4 in
all
defects,
The
configuration problem" of deciding what combination of focused and joint
paper, assuming instead that the available tests are predetermined.
set contains joint tests, the
test to
previous
4.2
test
outcomes;
test selection
optimal
we
test selection
next formulate this
tests
this tradeoff in this
When
the available test
policy might be dynamic,
i.e.,
the decision
problem.
problem seeks an optimal or near-optimal
tests
between dynamic and
regarding previous
test selection
Problem formulation and optimal solution
minimizes the expected number of
We distinguish
include
apply next depends on the current "state" of the system defined by the
Test selection:
The
"test
tests to
poses challenging tradeoffs between the number of available
and the expected number of iterations per board; we do not address
on what
and can,
and the relative likelihood of
various defects), reduce the expected number of iterations per board.
in the available test set
test that is
test
outcomes
same sequence regardless of the
or iterations to diagnose a malfunctioning board.
static policies; a
dynamic policy uses information
to select the next test,
test
testing policy that
while a
static
policy selects the
outcomes. Dynamic policies perform bener
(i.e.,
require fewer tests on average) since they use more information. The optimization of
-18-
they
sequenrial search processes has been previously modeled (see, for example, Whinston and
Moore
[1986]), and applied, for instance, to computer
Whinston [1988]). The
search (Moore, Richmond, and
file
has also addressed optimal
reliability literature
sequencing
test
Binney
issues for certain special circuit and test structures (e.g., Nachlas, Loney, and
[1990], Butler and Lieberman [1984] study series systems with focused
unlike our dynamic knowledge acquisition context,
that all defect types
and
without provisions for
their probabilities are
unknown
this
However,
tests).
stream of literature often assumes
known, and
the requisite tests are available,
or indeterminate outcomes. Furthermore, the knowledge-
based diagnosis systems described
in the literature
do not appear
to incorporate optimal
search principles.
Dynamic Programming Formulation
4.2.1
This section presents a standard dynamic programming formulation to clarify the
ingredients and tradeoffs in test selection during circuit diagnosis. First,
some
We
notation.
m available
types and
contains
use the indices
We
tests.
=
i
=
j
1,2,. ..,m to
introduce
represent the n defect
assume, without loss of generality, that the case base
we can
possible defect types. Otherwise,
all
and
1,2,. ..,n
we
introduce a
dummy defect
type called
"1"
"unidentified" (which includes "no trouble found"), and a fictitious focused test with a
outcome
if
the board has an unidentified defect, and a "0"
defect types (we perform this fictitious test only
the set of
partial signature for
combinations)
the expected deterministic
unambiguously identify
to
all
other
known
However,
the
defects
is
known
each defect must be accurate, and we must have enough information
must known
the analyzer
for
known candidate
We permit incomplete knowledge of defect signatures.
empty).
(i.e.,
when
outcome
the true defect.
outcomes
We
for
enough
assume
also
defect-test
that the test
observations and measurements are error-free. These assumptions ensure that the
diagnosis process
set containing
is
accurate, and always terminates "properly" with the candidate defect
only the true defect.
At any intermediate
for the tests
defects
iteration of the sequential search process,
we have performed
are consistent with the observed outcomes)
can distinguish between the remaining defacts); hence,
outcomes completely capture our current knowledge about
state vector representing this current
knowledge.
We
specify the
m is the number of initially available tests):
three values:
=
Xj
=
1 if
test j's
state vector
X^
if test j
was previously performed and
outcome was
has elements
" 1
x^'
",
=
and
N
X:
N
=
for all
j
if test j
=
state
X.
Initially,
D(X^) =
(
l,2,...,n) is the set
-19-
all
previously observed
We let X denote the
state X as the following
tiiej
its
element
x.
can take
observed outcome was "0",
initial
Let D(X) and T(X) denote,
and available
of
and available
has not yet been applied. The
l,2,...,m.
respectively, the index sets of candidate defects
tiie
(i.e.,
die board.
m-vector (where
X:
observed outcomes
thus far determine the remaining candidate defects
whose expected outcomes
tests (that
tiie
tests
corresponding
possible defects, and
to
T(X^
=
any
{
the set of all available tests. Starting with this initial state, the iterations of
1,2,. ..,m} is
moving from one
the diagnosis process correspond to
terminal state
X
whose
intermediate state X,
and update the
1),
we
state
select and apply a
vector (change
The expected number of tests
p
j
>
this section,
we
will
contains defect
i.
j
)
reach a
contains only one defect. At any
€ T(X), observe
N
from
X:
and the
assume
denote the prior probability
test
we
state to the next until
its
outcome O: (=
or
to O;).
complete the diagnosis of a defective board depends on
to
the probability of different defects,
remainder of
D(X
candidate defect set
testing policy that
that the
during the previous month; alternatively,
set P|
board contains only one defect. Let
equal to the relative frequency of defect
we might
probability based on qualitative information
the
of the diagnosis process) that the board
(at the start
For instance, we might
we choose. For
(e.g.,
i
subjectively estimate this prior
new
operators on the
problems
line,
at
the vendors' facility). Since, by our definitions and assumptions, the n defect types are
mutually exclusive and collectively exhaustive, the prior probabilities
sum
to
1, i.e.,
n
Ip°
1=1
As we apply
tests
=
(4.1)
1.
'
and eliminate defects from the candidate defect
set,
we must update
the
defect probabilities to reflect the test outcomes. Let Pi(X) represent the conditional
probability that the board contains defect
defect
i
that
probability,
i
given that
we
are currently in state
D(X) must have zero
X Any
.
does not belong to the candidate defect
set
and the conditional probabilities
remaining defects must sum
for all
conditional
to
1.
Thus,
p.(X)
=
ifi««
D(X), and
1
=
I
p^)
heD(X)
h
p°/{
•
At any intermediate (non-terminal)
next
test
only depends on the current
remaining
must choose
tests until the
the test
state; the
j
at
minimum
cost-to-go
each subsequent
state,
if
we
e T(X)
are currentiy at state X, the optimal test
that
minimizes the expected number of
diagnosis process terminates. These feaaires suggest die
C*(X) from
we
state
optimal
X
test
sequencing problem.
as the expected
select the next test "optimally".
the index of the "optimal" test to apply
this
(4.2)
exact sequence of previous tests that led to
following dynamic programming formulation for the
the
ie D(X).
stage of the diagnosis process, our choice of the
the current state is irrelevant Therefore,
selection policy
if
if
we
are in state X.
test.
-20
We define
number of remaining
tests
if,
Let j*(X) e T(X) represent
Let us describe
how
to identify
For each
perform
test
test
j
€ T(X),
j
Applying
.
C(X j) denote
let
test
gives two possible outcomes O. =
j
different next state (obtained by changing
respectively).
Each of these two next
from
x^
states has
can, therefore, express C(X,j) in terms of these
purpose,
T(X)
let
at state
us define k(X,Oj), for Oj =
X
results in
outcome
or
1,
and
their
represent the respective subsets of defects in
current state
X
j
and the
additional information,
indeterminate
is
are 0,
test
j;
we omit
we assume
that
S
=
•
•'
16
probabilities,
=
1
X
outcome when
it
(these subsets
and j for
j
e
test
in
depend on
simplicity).
Z
P:(X) + 0.5 *
the
Without
j is
unknown
or
is:
(4.3)
P:(X).
i€D!uDU
outcome
for test
j
Z
+ 0.5*
X is:
applied at state
(4.4)
Pi(X).
leDIuDU
we can compute
the cosi-to-go C(X,j) as follows:
test j
)
denote, respectively, the next states (with
e T(X)
or
is
I.
The
first
number of tests
that are performed),
cosi-to-go from the two possible next states.
cost-to-go
C*(X) from
test that
X
applied at state
is
minimum
is
1,
DU, and DI
+rt(X,Oj=0)C*(Xu(Oj=0}) + 7r(X,Oj=l)C*(Xu{Oj=l)),
outcome O: of
function counts the
X
or
to
D(X) whose expected outcomes (recorded
term
state
X
as the
achieves this
C*(X)
=
j*(X)
=
a terminal state,
i.e.,
additional tests; hence, the
minimum
Min
{
is
test
last
j
(recall that
or
then compute the
j
e T(X)
),
1)
on the
our cost
two terms contain
the optimal decision j*(X),
the
minimum
j
e
i.e.,
and
(4.6)
(4.7)
c(X,j).
D(X) contains exactly one
=0
We
=
(4.5)
value of C(X,j) over aU available tests
C(X,j):
argmin
T(X)
J 6
minimum
C*(X)
minimum
and the
x.
the constant I)
(i.e.,
right-hand side of equation (4.5) represents the "cost" of applying
If
to a
depends on the conditional
expected outcomes. Let DO, DI,
DI
where Xu{O;=0) and Xu{Oj=l
T(X); the
X
as the probability that applying test
any defect whose outcome for
Zp,(X)
=
j:(X,0,= 1)
the
each leading
in state
we
if
minimum cost-to-go. We
subsequent minimum costs-to-go. For this
leDO
Similarly, the likelihood of a "1"
when
X
equally likely to give a "0" or "1" outcome. Thus, 7r(X,0;= 0), the
K(X,O:=0)
C(X,j)
N
1,
state
an associated
arguments
the
or
current value
unknown, and indeterminate
1,
probability that test j will give a "0"
Using these
its
O;. This probability
probabilities of the defects in D(X),
the case base) for test
from the current
the cost-to- go
defect, then
we do
not require any
cost-to-go
if
-21
ID(X)I =
1,
(4.8)
where
iSI
number of elements
represents the
dynamic programming recursive equations
Equations (4.6) and (4.8) are the
in set S.
to
compute
minimum
the
cost-to-go and the
optimal decision for every state X.
Dynamic Programming Procedure:
•
Initialization:
For every terminal
Initialize:
•
X,
state
Labeled_states <—
C*(X) =
set
set
of
all
0.
terminal states.
Iterative step:
For every
state
X
whose next
belong
states all
to Labeled_states,
use equation (4.5) to compute C(X,j) for every
test j
e T(X);
use equation (4.6) to compute C*(X);
equation (5) gives the optimal decision j*(X)
add
Repeat
At every
X
state
X;
to Labeled_states;
iterative step until all states are labeled.
iterative step of this solution procedure,
whose next
at state
states
Xu{Oi=0) and Xu{0.=
l
},
we can
find at least
for every next test
j
one "unlabeled"
€ T(X), are
(otherwise, the state transition diagram must contain a directed cycle, which
The dynamic programming formulation can
(e.g.,
minimize the expected
precedence constraints
easily
total testing "cost"
(e.g., test j
must precede
is
all
state
labeled
impossible).
accomodate other objective functions
when
tests
test
due
j'
have different costs) as well as
to technological or operational
constraints) and probabilistic test outcomes.
We can represent the optimal testing policy
algorithm as a "look-up" table;
found by the dynamic programming
this table specifies the best
next test j*(X) for every possible
I,
state
X. Let us
illustrate
how
represent the state at the end of the k
not a terminal
test j*
state.
e T(K^) for
the state vector
by
At
setting the
x^'-^
apply
we
this test,
=
is
use the look-up table to identify the optimal next
and observe
element equal
j
and
iteration of the diagnosis process,
iteration (k+l),
state x'',
X
suppose X
the analyzer might use this table during diagnosis. Let
to the
its
outcome
0.».
observed outcome,
\\
ifj^tj*, and
Oj.
ifj=j*.
We
then update
i.e.,
(4.9)
Eliminating the candidate defects that have an expected (known) outcome different from O.*
gives the new, smaller candidate defect set D(X'''^b corresponding to the
If D(X'''^^)
repeat
tiie
contains only one defect, the diagnosis process
above process for the new
to create the decisition table.
By
state.
is
new
state
complete. Otherwise,
The dynamic program can be solved
storing this table in a
"*"
test selection functions.
.
we
"off-line"
computer database, the case-based
diagnosis support system can automate the look-up and
-22
X
Although the dynamic programming procedure
line,
it
is
computationally intensive for complex circuits that have a large number of
available tests. Furthermore,
We,
probabilities Pj change.
next test
easy to state and can be solved off-
is
at
we must
re-solve the problem every time the prior
therefore, consider
some
on-line heuristic rules to select the
each iteration of the diagnosis process.
4.3 Heuristic Decision
Rules
Easiest to implement are on-line rules that dynamically select the next test
at
each stage
(instead of referring to a look-up table) based on simple computations using information
about the current
Intuitively, the
state.
on-hne rule must select a
We list below
the candidate defect set.
three out of
many
test that
can rapidly prune
possible rules to achieve this
objective.
Highest failure
rule:
This rule selects the
least
test
with the highest probability of failure, where
probable outcome of a
optimal strategy for our example
test that
in
Section
4.
1
Recall that
test
patterned after the
select the "focused"
proves the defect by
failing).
permit indeterminate outcomes as well as incomplete knowledge of
we
defects
regardless of the
test's
test
unknown
with
the available test with the largest
we
We
might enhance
prefer a test for which the
unknown
or indeterminate outcomes
remain
in the
candidate set
actual outcome. Hence, to rapidly prune the candidate defect
the remaining candidate defects).
probabilities (e.g.,
A
relatively ineffective since these defects
is
set, this rule selects
defects with
is
define the
rule:
certain (deterministic) test outcomes.
many
rule
where we always
proves the most likely candidate defect (the
Most known outcomes
for
mode. This
test as its "failure"
we
sum
or indeterminate outcomes
is
number of known outcomes
this rule
(for
by incorporating defect
of probabilities of
all
candidate
the smallest).
Smallest expected remaining defects rule:
To rapidly home
in
smallest expected
outcomes O: =
the true defect, this rule selects the available test
number of remaining
and
remaining defects
unknown
on
if
1,
we
defects (which
is
the
with the
j
sum, over the two
of the probability of obtaining outcome O. times the number of
observe outcome 0:). Observe that a
test that
has
many
or indeterminate outcomes (see previous rule) will likely have a large
of remaining defects for either outcome. Again,
we
can modify
this rule to
number
account for
the "weight" or probability of the remaining defects.
We might enhance these myopic rules to incorporate
"look-ahead" features,
consider the likely states two or more iterations hence. Alternatively,
to
we might consider
"static" rules that priortize the tests before staning the diagnosis process; at
23
i.e.,
every stage, the
available test with the highest priority
applied next. Unlike the dynamic programming
is
procedure, these heuristic rules do not guarantee optimality,
might require more
on average than
tests
i.e., in
the long run, they
the optimal policy. Their relative effectiveness
depends on the nature and probability distnbution of various defects, the
(e.g.,
focussed versus joint
To smdy
tests),
and the
level of
knowledge regarding
the relative effectiveness of heuristic rules,
extensive simulation study.
The study applied
test characteristics
test
outcomes.
Semmelbauer [1992] conducted an
selected heuristic rules to
many random
problem instances with varying characteristics such as the number of possible defect
the
number of
results),
available tests, the "density" of the case base
and the shape of the defect probability distribution
commonly occuring
corresponds to a few
many
defects and
(i.e.,
the fraction of
(e.g., a
skewed
all
defects).
diagnosis environment, the simulation generates several problem instances
heuristic rule.
number of
tests
tests for
each
rule.
distribution
For each
(i.e.,
defective
required to complete the diagnosis using each
The average number of tests over
expected number of
test
"rare" defects, while an
uniform distribution implies approximately equal likelihood for
boards) and records the
types,
known
A
these
random
instances estimates the
naive rule that randomly selects the next
each stage provides a benchmark for comparison.
We
test at
summarize some important
conclusions from these simulations (for more details, see Semmelbauer [1992]):
•
Principled test sequencing can give significant savings:
most known
rules (such as the
the
•
number of
tests
performed) relative
(i.e.,
to
Problem
size
probabilities
random
test
30%
(in
sequencing.
as the density of the case base increases.
A
fewer unknown and indeterminate outcomes) eliminates more
defects at each stage since
•
relatively simple heuristic
outcomes rule) gave savings of approximately
The number of tests performed decreases
denser case base
Even
it
contains more information.
(number of defect types) and the skewness of
do not appear
to
the distribution of defect
have a significant impact on the relative performance of
different heuristics.
Next,
we
discuss
some
lessons
we
learned by implementing a prototype of the case-based
circuit diagnosis system.
5.
5.1
Prototype Implementation and Future Directions
Features of
CDSS
prototype
We implemented and tested a prototype of the Circuit Diagnosis Support System
(CDSS)
in a
production environment to diagnose an acaial pnnted circuit board.
The
primary purpose of the implementation was to prove the concept, and understand human
issues associated with introducing
computer support
-24
tools in a predominantly
manual
The prototype was
diagnosis environment.
computers
at the
Omnis
written in
5 for Macintosh
mice, and keyboards as input devices. The
Omnis program communicates
the facility. This central database stores
machine serving
base, p)ermitting multi-user access via a local network.
via an
(ORACLE) on
(Standard Query Language) interface with a relational database
UNIX
SE
technician stations; these terminals use touch-screens, bar-code readers,
SQL
the central
and manipulates the case
The prototype system was
not
integrated with the factory information system, and used only a simple decision rule (the
most known outcomes
rule) to select the next test at
each stage. Also, the prototype did not
incorporate the signature augmentation feature described in Section
[1992] describes the prototype implementation
in
3.
Semmelbauer
greater detail.
Observations and feedback from users over a five-week period led to several system
improvements. The
Figure
4, that
final version
of the system employs a user-friendly interface, shown
displays matching cases (the candidate defect
outcomes, and permits technicians
One
the candidate defect set.
to assess the usefulness
set),
and previous
tests
in
and
of each available test in pruning
of the important system features
is
an user-specified option to
override the automatic test selection procedure; using this option, technicians can choose
the next test manually
approach,
its
from the
list
of available
tests.
The
simplicity of the case-based
transparent logic and operation, and the manual override option greatly
facilitated the introduction
The system was
tested
and acceptance of the system
in the
existing environment.
on a double-sided board containing approximately 400
components, half analog (radio frequency) and half digital. Only part of the board was
supported by the
Over
CDSS;
the test circuit contains approximately
the course of five weeks,
70 analog components.
467 boards were diagnosed using the prototype system;
these boards had 87 different defect types (involving 54 components,
multiple defect types).
We estimate that the average
from about 25 minutes
p>er
with
CDSS
is
assistance. (This estimate
is
based on an informal observation for random
cases to be searched. In practice,
higher
when
we found
weeks of case
if
potential disadvantage of the case-based
the case base contains a very large
that the
number of
response time was not adversely
addition. For situations that require a significantly
number of cases, we might consider
limiting the
number of cases
that are stored;
the case base reaches this limit, an old case (with the smallest current probability of
occurence) must be discarded
in
One
the degradation in performance
affected even after five
time to diagnose each board dropped
board for purely manual diagnosis to approximately 10 minutes
boards rather than a rigorous time study.)
approach
some of which had
response time per
the case base
if
in
test iteration
order to add a
new
case. This strategy trades off the gain
against the additional effon needed to re-leam and update
an old defect occurs again.
25-
To
evaluate the system's learning
each week of the
during a week that
Figure 5 shows
period.
trial
is
how
technicians
felt
Coverage
is
fifth
defined as the proportion of defects occuring
initially
contained cases corresponding
were most Ukely or
significant (based
week, coverage increased
to
85%
90%.
rate
5.2
knowledge can be quite inadequate
makes
it
17%
relies
on
its
management
support.
To
dunng
the first
test circuit;
week of
The CDSS's
rapid learning
dynamic electronics assembly environment.
exploiting
potential requires
its full
rapidly develop full defect coverage, the system
regular use by experienced technicians as they detect and diagnose
many
as with
to use the
system because they beUeve that they have the
requires
new
defects.
knowledge-based systems, experienced technicians are less likely
However,
phenomenon
Figure 5
of the defect types, suggesting that
for this application.
especially attractive in today's
As
their experience).
Interestingly,
Managing the diagnosis function
Although the CDSS is effective and easy to use,
appropriate
on
its
to defects that the
coverage was achieved for the
operation, the system could only diagnose about
initial
weeks of
the system's coverage increased during the first five
shows, after 3 weeks of using the CDSS,
by the
the coverage of the system during
correctly diagnosed using the current information in the case base.
The case base
operation.
we measured
rate,
some fundamental changes
in the
least to learn.
way
Overcoming
this
the diagnosis function is
managed.
Technicians have traditionally been managed as hourly production workers despite their
important role in
tiie
quality
improvement loop and
tiieir
high
skill level.
Emphasizing
their
contributions to defect prevention and process improvement, rather than viewing diagnosis
as their sole function,
can produce rapid progress towards worid-class manufacturing
standards. This shift in emphasis
from
fault finding to proactive process
improvement
requires changing the technicians' performance evaluation metrics, providing appropriate
incentives,
sharing.
initial
1.
and creating an environment
that
is
conducive
to learning
and information
Based on our experience with the prototype system, we suggest
steps to
implement these changes and exploit
the following
the technicians' skills:
Change the technician's job description and expectations: Technicians' first priority
must be defect prevention and process improvement. They are among the most skilled
workers on the line, and are often the fu-st to identify defects. Thus, they are better
suited than the design or process engineers (who are typically less concerned with dayto-day operations) to respond quickly to production problems.
2.
Measure quality as well as quantity: Emphasizing the quality of diagnosis and the
consequent improvements and learning can create a productive culture. Traditional
productivity metrics such as the number of boards diagnosed per month often increase
the variability in diagnosis times by prompting technicians to work on the "easy" boards
first.
-26
3.
Introduce technical supervision through master technician. Traditionally, technicians
report to the production or line supervisor who often does not have detailed technical
knowledge about circuit behavior. Appointing a master-technician to supervise the
diagnosi^repair op)erarion and provide consulting help to the technicians can greatly
improve productivity, especially in an environment containing multiple assembly lines.
4.
Interact closely with product development. Closer cooperation with product
development might be encouraged, for instance, through a rotation program or a career
path firom technician to development engineer.
We
believe that electronics companies will be naturally driven towards these changes in
order to cope with increasing product diversity and small-lot production.
Future directions
5.3
Our prototype implementation can be
features that
tests to
we described
in
Sections 3 and
have more than two outcomes,
expUcit feature,
(iii)
enhanced
easily
(ii)
4.
to incorporate several
These enhancements include:
(i)
system
permitting
providing defect validation and verification as an
incorporating the automatic case base updating mechanisms
(e.g.,
signature augmentation), (iv) integrating fully with existing factory control and information
systems, (v) collecting actual
test
measurements (continuous values) and automatically
inferring the corresponding test outcomes, (vi) permitting
(vii)
more than one
minimizing the expected testing cost per board when
tests
have different costs,
accounting for precedence constraints and sequence-dependency
incorporating
more sophisticated
test
(i.e.,
among
(viii)
tests, (ix)
sequencing methods based on defect probabilities,
permitting multiple signatures for the
measurements
defect per board,
same
defect,
and
(xi)
(x)
adjusting for error-prone
permitting tests with false positive and false negative outcomes; see.
for example, Nachlas et
al.
[1990]). These enhancements are possible even with current
hardware, software, and development tools. The case-based approach offers two very
promising possibilities
automated probing
in the future
—
integrating with a model-based approach, and adding
capabilities.
We have argued that,
while the model-based approach offers the promise of seamless
and instantaneous design-diagnosis integration, the method
applications because of
approach
is to
combine
its
is
not suitable for real-time
extensive computational requirements.
One very promising
the model-based and case-based approaches in a hybrid system that
captures the automatic reasoning capabilities of diagnostic models with the quick response
of a case base that stores only the most important and relevant defects. The model-based
approach can complement the case-based system
new
test is
added
to the case base,
we can
in
two ways:
(i)
whenever
a
new
use the circuit model to verify the correctness of
the partial signature (specified by the analyzer or obtained by recording the test
which
led to the
new
defect)
and even augment
base with model-based predictions of
test
defect or
it (i.e.,
replace
outcomes); and,
-27-
(ii)
unknown values
when
outcomes
in the
the case-based
case
approach terminates with an empty candidate defect
new
hypothesis regarding the
defect,
set,
the circuit
and propose appropriate
model can generate
tests to isolate
it.
We can
also apply the model-based system off-line to initialize the case base. Note that these
schemes apply
model only
the circuit
to resolve exceptions; hence, they
do not degrade
the
average response time of the system during regular diagnosis. The hybrid approach,
therefore, offers the advantage of a robust case-based system that permits user interaction
and provides rapid on-line response, with the enhanced model-based capabilities
quickly to changes
in
product design (using model-based reasoning).
of the diagnosis function into
development of
its
the case-based
to react
This decomposition
short-term and long-term components permits independent
and model-based subsystems; for example, we might
first
develop the case-based system using currendy available hardware and software, preparing
the infrastructure for a future implementation of the model-based
improves
state-of-the-art
to
The second promising
to the
CDSS. As
testing
will
the
the approach practicable.
direction concerns the addition of automated probing capabilities
printed circuit assemblies
and probing
signal generators,
make
component when
become
become smaller and more dense, automatic
a necessity.
Most
test
equipment, such as power supplies,
and frequency analyzers, are already controllable by a computer;
however, currently most diagnosis operations use manual probing. With the emergence of
new
robotic technologies with precise and swift
movement
capabilities (with vision
feedback control), automated flexible probing becomes feasible
A case-based
system controlling the automatic
circuit
in
manufacturing
and
settings.
probing and measurement equipment
can reduce diagnosis time by performing routine and repetitive diagnosis functions without
human
intervention; analyzers
would
still
be required to troubleshoot
new
defects and
resolve inconclusive diagnoses.
In conclusion, this paper has propxssed a viable approach to reduce diagnosis time and
effort,
and provide rapid feedback for quality improvement.
based approach
is its ability to
A
novel feature of our case-
incorporate principled test selection methods. Although
developed the dynamic programming formulation and heuristic
we
test selection rules in the
context of off-line diagnosis, the underlying principles also apply to in-line (in-circuit and
functional) testing.
Computer
assisted diagnosis
repair strategy.
is,
of course, just one element of the overall testing and
We have alluded to the many other decisions and tradeoffs to consider in
formulating an effective testing strategy. These decisions include:
•
how
to incorporate diagnosis
and
testability issues in the
product design process?
Williams and Parker [1983] present a comprehensive survey of issues and methods
relating to design for testability;
•
where
to locate in-line test stations,
and what
-28
tests to
perform
at
each station?
Pappu and Graves [1992] discuss
this issue in a
generic production line setting, and
Chevalier [1992] considers alternative in-line testing strategies for printed circuit board
assembly operations;
•
whether
to
perform detailed, diagnostic
As we noted
times and
in
Section
2, this
tests in-line or off-line?
decision depends on the tradeoff between higher cycle
capacity requirements on the main assembly line versus greater off-line
test
effort;
•
how
to "discretize" the
For instance, the
test
continous outcomes of tests?
measurements
in
analog circuits are continuous values; the
engineer must specify ranges of values that correspond
the test.
When
proposes a method to
set the
outcomes for
to,
test
'
"PASS" or "FAIL
say,
measurements are error-prone. Chevalier [1992]
the
ranges based on the relative costs and probabilities of false
acceptance and rejection;
•
what
tests to include in the test set for off-line
As we
discussed in Section 4 the choice of joint versus focused
expected number of iterations
•
when
diagnosis?
to
implicitly
assumed
the board can be repaired. For
(including the cost
of work
impacts the
complete the diagnosis;
to stop the testing process for a
Our discussion
tests
board?
that the defect
some
identified so that
defects, the cost to diagnose and repair the board
might exceed the value of the board.
in process, etc.)
boards contain multiple defects, a
must be conclusively
common
policy
is
to discard
any board
that
When
goes
through the diagnosis-repair loop more than a cenain prespecified number of times.
Conversely, analyzers might sometimes terminate the diagnosis process prematurely
(before conclusively proving a defect), and
recommend
repairing the
most
likely cause
of the problem; and,
•
how
Ou
to screen boards that enter diagnosis?
and Wein [1992] and Longtin
et al. [1992] discuss screening strategies to discard
semiconductor wafers with low yield
in
order to avoid wasteful usage of the subsequent
detailed wafer-probing capacity. Similar screening strategies might alleviate diagnosis
bottlenecks in printed circuit board assembly.
As
these issues illustrate, testing and diagnosis in printed circuit board assembly operations
many important and challenging research
as we have defined it, focusses on finding
continues to offer
opportunities. Finally, circuit
board diagnosis,
the source of the
that the
its
board can be repaired. For process improvement,
root cause,
i.e.,
the
we need
problem so
to trace the
component vendor, board design, or processing
problem
to
step responsible for
introducing the defect; proactive process control requires a further step, namely,
anticipating and preventing defects.
Given
massive volumes of data generated by the
the
placement, reflow, inspection, testing, diagnosis, and repair stations in today's high
volume
electronics assembly environment,
we
require automated systems that
29
systematically monitor and screen the signals, recognize patterns, identify the root causes,
and
initiate
process control and improvement activities.
Acknowledgments
We are very grateful to Brian Santoro, Thomas
:
Babin, and Dennis Miller for initiating
and providing constant support and encouragement. We thank the analyzers,
especially Zakiuddin Mashood and Karl Browder, and members of the DOC team for
sharing their insights on circuit diagnosis.
project
30-
this
Table
1
:
Complete Case base
for
5-component,
serial
network
Figure
1
1:
Cumulative probabiiity of various defects
00% T
CO
CM
Defect
Number
-T
(in
in
to
deer. prob. order)
r~-
CO
Figure 2: Flow Chart for Case-based
Initialize
Case Base
CDSS
to
include most common
anticipated defects and their
signatures
I
Initialize
Candidate Defect Set (CDS)
New
Available Test Set (ATS)
to
Board
be diagnosed
I
Select Next Test
from ATS
I
Apply Next Test
and observe outcome
i.
Update CDS: eliminate all defects
whose expected outcome does not
match observed outcome
I
Update ATS: remove latest test +
all tests whose outcomes are
the same for all remaining defects
I
No
CDS
contains >
1
defect?
D
( ATS empty?
Yes
I
f
Add new
tests
j
Add new case
(defect)
+ any additional tests
Perform
defect verification
remaining
(apply
tests)
signature augmentation
Figure
3:
Input
terminal
Comp
2
J
1
Probe point
for
Defect
Test
i
i
Test 1
implies
Proh>e point
for
Test 2
component
i
is
Series network example
»
Comp 3
—j ^
Comp
probe point
for
Probe point
Test 3
defective, for
i
=
for
Test 4
1,2,. ..,5
measures the voltage between components
i
and
(i+1), for
i
=
1,2,3,
and
4.
Figure 4:
r
Sample screens from CDSS prototype
n
1—
Figure
5:
Cumulative coverage and number of cases added
to
CDSS
3
Week
References
Bennetts, R.G., 1981. Introduction to Digital Board Testing Crane, Russak
.
&
Co.
and G. J. Lieberman, 1984. "Inspection Policies for Fault Location".
Operations Research, Vol. 32, No. 3, pp. 566-574.
Butler, D. A.,
Chevalier, P., 1992. "Optimal Inspection for Circuit Board Assembly", Part II of Ph.D.
thesis, Operations Research Center, Massachusetts Institute of Technology, Cambridge,
Massachusetts.
Davis, R., 1984. "Diagnostic Reasoning Based on Structure and Behavior," Anificial
Intelligence Vol. 24, pp. 347-410.
.
Davis, R., and W. C. Hamscher, 1988. "Model-Based Reasoning: Troubleshooting," A.I.
Memo No. 1059, Artificial Intelligence Laboratory, Massachusetts Institute of
Technology, Cambridge, Massachusetts.
J., and B.C. Williams, 1987. "Diagnosing Multiple Faults," Artificial
Intelligence Vol. 32, pp.97- 130.
de Kleer,
.
Frank, W., and B. Sauve, 1989. "Expert Systems for Manufacturing and Testing
Applications," Electrical Communication Volume 63, Number 2.
.
Garcia-Rubio, M., 1988. "Key Problems and Solutions in Electronics Testing,"
Productivity and Ouality Improvements in Electronics Assembly J. A. Edosomwan and
A. Ballakur (eds.), McGraw Hill Book Company.
.
Gensereth, M.R., 1984. "The Use of Design Descriptions in Automated Diagnosis,
Artificial Intelligence Vol. 24, pp. 41 1-436.
.
W.C, 1988. "Model-Based Troubleshooting of Digital Systems," A.I.
Technical Report No. 1074, Artificial Intelligence Laboratory, Massachusetts Institute of
Technology, Cambridge, Massachusetts.
Hamscher,
Harmon, P., and D. King, 1985. Artificial
Wiley and Sons, Inc., New York.
Intelligence in Business: Expert
Systems John
.
Henneman, and Rouse, 1984. "Measures of human problem solving performance in fault
diagnosis tasks," IEEE Transactions on Systems. Man. Cybernetics SMC 14, pp. 99.
112.
—
Dynamic
and H. Tokumaru, 1990. "A framework for case-based diagnosis
Frame and Reasoning on the Representation," Transactions of the Societv of Instrument
Ishida, Y.,
and Control Engineers Japan.
.
Kritz,
J.,
and H. Sugaya, 1986. "Knowledge-Based Testing and Diagnosis of Analog
FTCS Digest of Papers 16th Annual International Symposium on
Circuit Boards,"
.
Fault-Tolerant Computing Systems, pp. 378-383.
Lee, C.N., and P. Liu, 1988. "Case-based Reasoning for Robotic Assembly Cell
Diagnosis," Proceedings of the SPIE
The International Societv for Optical
241-246.
Engineering SPIE Vol.1008,
—
.
Rl-
Longtin, M., L. M. Wein, and R. Welsch, 1992. "Sequential Screening in Semiconductor
Manufacturing, II: Exploiting Spatial Dependence", Working paper, Sloan School of
Management, Massachusetts Institute of Technology, Cambridge, Massachusetts.
Maunder, CM., and R.E. Tulloss, 1990. The Test Access Port and Boundary-Scan
Architecture IEEE Computer Society Press Tutorial.
.
Moore, J. C, W. B. Richmond, and A. B. Whinston, 1988. "A Decision Theoretic
Approach to File Search", Computer Science in Economics and Management. Vol.
I
pp.
,
3-19.
J. C, and A. B. Whinston, 1986. "A Model of Decision-making with Sequential
Information- Acquisition", Decision Support Systems, Vol. 2, pp. 285-307.
Moore,
J. A., S. R. Loney, and B. A. Binney, 1990. "Diagnostic-Strategy Selection for
Series Systems ', IEEE Transactions on Reliability, Vol. 39, No. 3, pp. 273-279.
Nachlas,
Newstead, M.A., and R. Pettipher, 1986. "Knowledge Acquisition for Expert Systems,"
Electrical
Communication Volume 60.
.
pp. 115-121.
Noble, P.J.W., 1989. Printed Circuit Board Assembly Open University Press (Robotics
.
Series)
J., and L. M. Wein, 1992. "Sequential Screening in Semiconductor Manufacturing,
Exploiting Lx)t-to-lot Variability", Working paper, Sloan School of Management,
Massachusetts Institute of Technology, Cambridge, Massachusetts.
Ou,
I:
Pappu, S., and S. C. Graves, 1992. "A Dual-ascent Algorithm for Determining the
Optimal Testing Strategy", unpublished manuscript, Massachusetts Institute of
Technology, Cambridge, Massachusetts.
Preist,
C, and
Circuits,"
B. Welham, 1990. "Modelling Bridge Faults for Diagnosis in Electronic
Working Notes of First International Workshop on Principles of Diagnosis
(unpublished), Stanford University, Stanford, California.
Reiter, R., 1987.
"A Theory of Diagnosis from
First Principles," Artificial Intelligence .
Vol. 32, pp. 57-95.
Rissland, E.L., and D.B. Skalak, 1988. "Case-based reasoning in a rule-governed
domain," Proceedings - Fifth Conference on Artificial Intelligence Applications. IEEE.
Rouse, W.B., and N.M. Morris, 1985. "Review and Evaluation of Empirical Research
Troubleshooting," Journal of t he Human Factors Societv Vol. 27, No. 5.
in
.
Rouse, W.B. and R.M. Hunt, 1986. "Human Problem Solving in Fault Diagnosis Tasks."
Research Note 86-33, U.S.Army Research Institute for the Behavioral and Social
Sciences.
Semmelbauer, T., 1992. "A Case-based Approach for Diagnosing Failed Printed Circuit
Boards in Manufacturing", Masters' thesis, Massachusetts Institute of Technology,
Cambridge, Massachusetts.
Soederberg, E. M., 1992. "Using Defect Data to Characterize Assembly Process
Capabilities and Constraints for Printed Circuit
thesis,
Board Design for Assembly", Masters'
Massachusetts Institute of Technology, Cambridge, Massachusetts.
-R2-
Tong, D.W., E. Walther, and K.C. Zalondek, 1989. "Diagnosing an Analog Feedback
System Using Model-Based Reasoning," Artificial Intelligent Systems in Government.
Proceedings of the IEEE.
Memo
Trice, A., and R. Davis, 1989. "Consensus Knowledge Acquisition," A.I.
No.
1 183, Artificial Intelligence Laboratory, Massachusetts Institute of Technology,
Cambridge, Massachusetts.
Wenstrup, D. J., 1991. "Material Identification and Tracking in Manufacturing", Masters'
thesis, Massachusetts Institute of Technology, Cambridge, Massachusetts.
Williams, T. W., and K. P. Parker, 1983. "Design for Testability-A Survey". Proceedings
of the IEEE, Vol. 7 1 No. 1 pp. 98- 1 1 2.
,
,
Model Based Reasoning to Diagnosis of Faults in
Navigation Equipment," Master's Thesis, Air Force Institute of Technology,
Wright-Patterson Air Force Base, Ohio.
Yost, R.E., 1989. "Application of
Intertial
-R3-
/
{J
u
J o
Date Due
AU6.
t
6
1993
MIT LIBRARIES
'
mil nil nil
3
'
111'
II
loao ^077^b^3
3