~

Computersind.EngngVol.28, No. 4, pp. 793-81I, 1995

ElsevierScienceLtd. Printedin GreatBritain

Pergamon

MODERN

0360-8352(95)00012-7

COMPUTER-AIDED

MANUFACTURING

EQUIPMENT
AND

MAINTENANCE
AND

SYSTEMS:

OF
REVIEW

PERSPECTIVE

JAY LEE
National Science Foundation, 4201 Wilson Blvd, Arlington, VA 22230, U.S.A.

(Received 1 February 1995)

Abstract--This

paper examines various methods in modern computer-aided maintenance including

machine monitoring, fault detection and fault diagnostics. A perspectiveon proactive maintenance by
monitoring the degradation of manufacturing equipment and systems is presented and illustrated. If the
behavior of manufacturing equipment and systemscan be monitored and measured adaptively then an
early warning of possible faults can be generated. By doing this, maintenancepersonnelcan perform early
diagnostics and part replacementduring regular maintenancehours. The paper also addresses the research
needs based on the industrial perspective. The author believes that the development of in-process
monitoring of machine degradation and faults is one of the most important research tasks for increasing
machine uptime and improving production quality.

INTRODUCTION
A major problem confronting American manufacturers today is how to economically achieve high
quality in the goods which they manufacture. There are three ways to achieve consistent
high-quality parts: (1) 100% postproduction inspection; (2) in-process machine monitoring and
control; (3) zero-defects (Poka-Yoki) design [1] for machines and processes. Postproduction
inspection to ensure high-quality products is not economically feasible. If done manually, it
takes far more labor to inspect the output of a machine than it does to operate the machine.
If done automatically, it requires specialized, automated, inspection equipment which is unique
to the product and which can cost far more than the production equipment itself. In both
cases, the inspection may miss critical defects and in the manual case may result in excessive reject
rates.
Zero-defects design (Poka-Yoki in Japanese) is a mistake-proofing technique employed by many
Japanese manufacturers to eliminate the defects early in the design stage. Many Poka-Yoki ideas
are developed and implemented by line operators working with engineering staff, tooling or
machine specialists. The zero-defects design concept is based on human engineering to simplify the
manufacturing process prior to implementing automation. The result is high quality, high morale,
and greater participation of operators in improving manufacturing processes, products, and the
organization as a whole. Although it sounds easy, it is very difficult to implement zero-defects based
design in the U.S. due to an unsympathetic organizational culture. American manufacturers seem
to be interested in retrofitting new manufacturing and processing technologies, such as intelligent
sensing, automation, and intelligent control systems. The problem is that expensive machines and
processes are already built and installed in existing production lines. However, it is essential to
improve existing conditions if the U.S. is to be competitive in the global economy [2].
In the in-process monitoring and control approach, a process is set up to produce high-quality
parts, then is monitored to determine whether the process or the machine is no longer operating
correctly. Statistical process control (SPC) is a popular method for monitoring the trend of process
parameters such as dimension tolerance in assembling automobile bodies. However, this technique
is being used to monitor the controllable process parameters, such as flow rate, temperature,
force and conveyer speed, etc. [3-7]. Currently, no proven technique has been employed to
monitor the degradation of a machine and its process in order to adaptively prevent faults or failure
[2, 8, 9].
793

794

Jay Lee

Importance
Equipment reliability and maintenance drastically affect the three key elements of competitiveness~quality, cost, and production lead time. Well-maintained machines hold tolerances better,
help reduce scrap and rework, and raise consistency and quality of the part. They increase uptime
and yields of good parts, thereby cutting total production costs, and also can shorten lead times
by reducing downtime and the need for retooling [2]. The recent rush to embrace computer-integrated manufacturing (CIM) has further increased the use of relatively unknown and untested
technology. Much of the controller software has been written with the assumption that all
equipment works properly and logically, with high reliability. In a CIM system, a fault can cause
machine and production downtime and create serious productivity losses.
Typically, when a production system goes down, only a small fraction of the downtime is spent
repairing the machine that causes the failure. The greatest part of downtime is spent locating the
source of the problem [9]. Difficulty in identifying the causes of system failures has been attributed
to several factors, including system complexity and lack of adequate troubleshooting tools [10]. An
essential prerequisite for the development of an automated supervision system is a process that can
detect faults early. If U.S. machine tool utilization, for example, increased from 50 to 100% (based
on unlimited mean-time-between-failure), investment in equipment could be cut in half [2].
Generally speaking, the operational performance of components, machines and processes can
be divided into four states: the normal operation state, the degraded state, maintenance state and
the failure state [19]. Figure 1 shows these states. The degraded states can be defined at a gross
level or at a detailed level. At a gross level, a component is described as degraded whenever a
deterioration occurs but does not cause loss of its function. For example, component degradation
occurs whenever corrective maintenance is required, such as a loose belt, a worn brush of a motor,
dust on a photosensor, etc., but the components have not failed. Detailed degradation is associated
with a given range of characteristics of the components, or performances of the component, such
as detailed degraded states for a car battery indicator and temperature indicator, etc. The advantage
of defining more detailed degradation states is that we can accurately predict impacts on the failure
of the component. When aging occurs, the component and machine generally progress through a
series of degradation states before failure occurs. If a degradation condition can be measured and
detected, then a proactive and corrective maintenance activities can be performed before a worse
degradation condition or failures occur.
A fault is defined as the inability of a system or machine to perform its expected function. A
fault manifests itself as a deviation of the system or machine behavior from its specified behavior.
Typically, a fault could develop during any stage of degradation. For example, a dust-covered
photosensor has a slower response time, which causes an operational delay. A worn gear could
introduce a position error for a positioning table or robot arm. Generally speaking, there are three

MAINTENANCE
STATE

I-

OPERATION
STATE

FAILURE
STATE

\',,,i ./,,,,,o,,,.
HUMAN

ERROR

Fig. 1. Definitionof machine performancestates.

Modern computer-aided maintenance

795

types of faults in a manufacturing system: equipment faults, quality faults, and task faults
(introduced by human errors).
Fault detection is the identification of a component or set of components in the system which
caused the fault. Fault diagnosis requires collecting and processing large amounts of often
incomplete data to determine the nature and severity of equipment and system malfunctions.
Ideally, fault diagnosis results in an accurate, unambiguous identification of the component or
components responsible for the anomalous behavior.
A maintenance state is the stage of maintenance performed either before the failure (preventive)
or after the failure (reactive).
A failure state is caused by degraded components, faults, or human errors. Generally, there are
two kinds of failures: (1) soft failure: a failure that can be corrected shortly after it occurs (minutes
on average); and (2) hard failure: a failure that causes the machine to be down for a long period
of time (hours on average).

Present practice
Today, many factories are still performing maintenance on equipment in a reactive, or
breakdown, mode, since traditional process monitoring systems can detect machine or process
faults only when they occur. Reactive maintenance is expensive because of extensive unplanned
downtime and damage to machinery [2, 9]. In high-performance systems one often cannot tolerate
significant degradation in performance during normal system operation. This had led many U.S.
manufacturers to look to suppliers to smarter equipment that will ease the need for strong technical
support. To avoid machine downtime, some companies have used preventive and predictive
maintenance approaches by either using historical maintenance data or by sensing machine
conditions. However, there are no maintenance data available for many newly developed machines
or advanced manufacturing systems. Adding sensors to monitor machine conditions will also
increase machine complexity and will require more highly trained personnel [11]. Therefore, some
manufacturers have designed fault detection systems that take system redundancy into account,
for example, a system containing several back-up systems.
REVIEW OF RESEARCHWORK
Figure 2 shows the conventional machine performance monitoring system paradigm which
includes sensing, process fault detection and diagnostics. Typically, the sensing parameters are
sensors, actuators, and time. For example, tool wear can be monitored by using the cutting force,
accelerometer, torque, temperature, spindle motor current, acoustic emission (AE) signals and the
actual cutting time.
In many manufacturing systems, on-line fault detection is often based on methods using
threshold value, logic, state-table, time-window and simulation (use of process inputs to drive a
simulation of acceptable system behavior in parallel with the actual system operation). The off-line
fault detection and fault diagnostics are normally performed by gathering the data from the sensors
and the symptoms from the system, then the problem is diagnosed through signature analysis.
Today, neural networks and expert systems are the two most popular approaches in fault
diagnostics. The following sections summarize the related research work in monitoring, fault
detection and fault diagnostics.

Methods For Detecting Faults

The conventional fault detection methods are: (1) modelling and parameter estimation; and (2)
discrete observation. The following subsections discuss these methods.

Modelling and parameter estimation approach

Measurable input signals and output signals can be directly used to monitor changes in a process.
The limit and trend checking technique is the most straightforward method in fault detection. The
limits are set such that a large enough distance to the appearance of damage is retained. The limit
check can also be applied on the trend of the signal. To predict the signal, mathematical models

796

Jay Lee

OBSERVED SYMPTOMS
(OFF-LINE)

CHANGE
OPERA'ON
STRATEGY

MONITORING
(ON-LINE)

PROCESS

t
SENSING

PROCESSTYPE:

PARAMETERS:

* DISCRETE
CONTINUOUS

SENSORS
ACTUATORS
TIME

FAULT
DETECTION

I _ IFAULT
I - ~ DIAGNOSTICS
I

STOP
OPERATION
AND REPAIR

FAULT
DETECTION
METHODS:
1. CONTINUOUSPROCESS

DIAGNOSTICS
METHOOS:
(BASEDON SIGNATURE
ANALYSIS)

FAULT
MANAGEMENT

* PARAMETER
* EXPERTSYSTEM
ESTIMATIONMETHOD * NEURAL NETWORK
2. DISCRETEPROCESS
* THRESHOLD
* LOGIC
STATE-TABLE
* TIME WINDOW
* SIMULATION

Fig. 2. The conventional fault monitoring, detection and diagnostic process.

of deterministic signals, or stochastic signals, or a deterministic process and a stochastic signal

model need to be used. On-line fault detection and diagnostics have been explored by using
real-time knowledge-based reasoning, parameter estimation and dynamic data system (DDS)
modelling approaches [27]. If process faults are indicated by internal, nonmeasurable process state
variables, attempts can be made to reconstruct or estimate these state variables from the measurable
ones by using a known process model [28]. Figure 3 shows the generalized structure of fault
detection methods based on process models and nonmeasurable parameters. A survey of design
methods for failure detection in dynamic systems can be found in Willsky's [29] work. These
methods are useful if models can be developed for a manufacturing system. However in many cases,
models are not available or are very difficult to develop due to the lack of process information.
Discrete observation approach

Fault detection through modelling and estimates of parameters is used for continuous systems.
However, a discrete state system or event-based system needs a discrete dynamic description. The
inputs and outputs of these systems are seen as sequences of events over time or states over time.
Most process control for manufacturing applications is for discrete event-based control: a
programmable logic controller (PLC) is a typical example. This controller can detect faults based
on the ladder diagram. Detecting and diagnosing faults through a discrete interface offers many
challenges. The frame of observation provides indications of system operation only at certain points
in time. These indications of system operation are infrequent, and there may be delays in the
information about a fault. Furthermore, because of the sequential nature of the discrete system,
the sensor only indicates the status of the operations. The correlation among the sensors and
actuators is not provided. Due to this reason, many manufacturing systems use a watchdog timer
to clock the operation and check the operation pattern. Since the system depends on the time and
sensor, the operation has to be stopped when a fault is detected. Therefore, it is very difficult to
perform fault-tolerant control and management. The following summarizes the conventional fault
detection methods for discrete processes.
Finite automata state machine. In this method, a particular sequence (or set of sequences) of
actuator and sensor signals is monitored by using state followers. The state machine uses a finite
state language to track the sequences of observations based on a graph of the operation. The
real-time control system (RCS) developed by Barbera et al. [30] is based on task decomposition
and a state table to use an "if-then" type of monitoring command in detecting faults in a

Modem computer-aided maintenance

797

MACHINE OR
PROCESS

NORMAL
PROCESS
MODEL

FAULT
PROCESS
MODEL

PROCESS
OBSERVATION

STATE
~ ) ( PARAMETER
ESTIMATION, .ESTIMATION

GENERATIONOF CHANGES
AND ERRORSIGNALS

FAULT
DETECTION

FAULT
SIGNATURE
FAULT DECISION
FAULTTIME

FAULT
LOCATION / ( FAULT
SIZE

[ CAUSE
FAULT

FAULT DIAGNOSTICS
Fig. 3. Generalized structure of fault detection methods based on the process model.

manufacturing system. The premise of this approach is that you need a complete understanding
of all situations in a process is needed and each condition has to be listed as an "if and then"
statement in a state table.
PLC-based logic diagram. Most flexible manufacturing systems use PLC because it is userfriendly. However, for a sophisticated manufacturing system, the PLC has a limited capability
because of the inflexibility of the programming system. Recently, many PLC manufacturers, such
as GE Fanuc and Siemens have developed a sequential function chart (SFC) to assist the operator
in diagnostics [31]. An SFC describes a machine's sequential cycle with the steps and conditions
necessary for the transition from one step to the next. Because of its graphical nature, an SFC is
much easier to read than a program listing, so that production personnel unskilled in reading PLC
programs may well be able to identify machine or process faults. A binary fault file is written to
the display when a step timer expires or under certain conditions; each fault bit writes a different
device name or short message on the screen. A PLC-based diagnostics system developed for
Cummins Engine Co. Ltd. to detect faults in a manufacturing cells is discussed by Bititci and Ross
[32]. The detection system was implemented on a PLC in parallel to the system controller, rather
than within it. PLC approach is limited since it can only detect fault when it occurs, it is impossible
to measure the degradation of a device such as a sensor, to provide the operator with a warning
signal.
Simulation. A simulation approach to fault detection uses process inputs to drive a simulation
of acceptable system behavior in parallel with the actual system operation. Many researchers have

798

Jay Lee

used simulation models, such as a Petri net model [33]. The model represents the processes of the
system and acts as a state observer which listens to interprocess messages and sensors in the
manufacturing process. Sequencing is incorporated in the structure of the Petri nets and timing is
incorporated by setting upper and lower limits on firing times of each transition. Detection is
accomplished by simulating the model along with the system. Barbera [34] at Advanced Technology
& Research Corp. (Laurel, MD) has developed an integrated control, design, and simulation tool
set for an automated material handling system for mail processing. This tool is based on real-time
control system (RCS) methodology. The timing, operation sequence, I/O status and control action
are coded at each operation state under a hierarchical programming structure. The work was
developed on a Silicon Graphics IRIS workstation and has been demonstrated with many
automation applications. This approach is very effective if the simulation model can be continuously updated and enhanced since manufacturing environment and conditions change frequently.
A robust simulation model should be able to learn and modify itself adaptively. In many cases,
the threshold values of fault detection algorithm also have to be re-calibrated from time to time
due to the variation of the manufacturing process.
Event-based time-window approach. Event-based control is a discrete-event form of control logic
in which the controller expects to receive confirming sensory responses to its control commands
within preset time windows that are determined by a discrete-event model of the system. The
conventional approach in signature recognition for fault detection has several degrees of difficulties.
It will become very difficult when the signature changes due to the condition changes, such as
temperature, pressure, spring adjustment etc. Figure 4 shows the level of identification for different
behaviors of the component.
The event-based controller starts in a check state, and stays in a wait phase for the minimum
allowed processing time t. A sensory input received during this period indicates an error since the
(expected) sensory response arrived too early. Once the time period has elapsed without external
Levels of identification

1.achan,oofstae

easy
I

2. the direction of the change of state

lJl

3. the current level of the change in

state

4. from simple to complex state

change

it unique from the other state

changes

Fig. 4. Degree of difficulty in signature recognition for fault detection.

diff cult

Modern computer-aided maintenance

799

interruption the model changes to a window phase, and stays in this phase for a duration specified
by the time window. A sensory input is expected to arrive during the period. If the input is received
and it tests valid, the appropriate next control command is issued. The check state is updated, and
the model changes back to a wait state for another appropriate duration. If the test fails an error
is reported. Finally, if the window period has elapsed without receiving the expected sensory input,
another error will be processed. The event-based controller moves through its check states in
concert with the received inputs, as long as all input signals arrive within their expected time
windows. In comparison with the conventional sampled-data logic, event-based controllers demand
less precision from their sensors. These sensors can have threshold-like characteristics. The burden
of the precision is placed on the time windows rather than on the sensors.
Event-based controllers are less sensitive to the received sensory input, and they can therefore
be made more robust. In the sampled-data control paradigm, sensors must be able to measure the
actual analog value of the sensed signal, and report it back to the controller. At the sampling time
t, the sensed signal S should be larger than S(min) but smaller than S(max). The controller will
thus compare the actual measured value of the sensed signal S with the two thresholds S(min) and
S(max). The threshold values may be functions of time. In the event-based control paradigm, a
simple threshold sensor is used which records when the sensed signal crosses the threshold in either
direction. The actual time t of the threshold crossing is then compared with the time window limits
t(min) and t(max) which are fixed values than can be predetermined.
Another important advantage of event-based control is that the error messages that are issued
by the controller contain information that can be directly used for diagnostic purposes. With the
event-based control paradigm, the control process can be readily interfaced with rule-based
symbolic reasoning logic in an advanced intelligent automation environment.
Methods for Diagnosing Faults

Most of the fault diagnostics research work is based on signature analysis of the process of
component. To diagnose a fault from these recorded, sampled and analyzed data, expert systems
and neural networks are the two most popular approaches to searching for the faults. The research
using these approaches is surveyed and summarized below.
Neural networks approach

Automated fault detection and a diagnostic systems based on neural networks have been
implemented for jet and rocket engines by Dietz et al. [35]. In this approach, fault diagnosis is
performed by mapping or associating patterns of input data to patterns representing fault
conditions. The expert system, called the integrated expert system (IES), employs neural networks
to analyze patterns of engine behavior, and create a set of hypotheses regarding anomalies of engine
behavior. The hypotheses generated by the neural network model are then given to the
knowledge-based inference model to perform reasoning on the basis of engine design knowledge,
heuristic knowledge and fault and repair history.
Greenwood and Stevenson from Netrologic, Inc. [36] have estimated tool wear using neural
networks. Uhrig and Guo [37] have used the backpropagation neural network to identify transient
operating conditions in nuclear power plants. Franklin et al. [14] from GTE have used the
ADALINE and the NADALINE, a one-layer neural network, to monitor fluorescent bulb
manufacturing. Two approaches are being compared by Franklin et al. in their research. One
employs standard statistical procedures to find correlations between sensor measurement and
quality. The sensor data from the production line are collected over a period of time and
correlations are made off-line at frequent intervals using analysis, such as linear regression. The
second approach is to estimate the correlations incrementally, as the data are collected on-line and
in real time. The estimates are updated incrementally using connectionist learning procedures.
HNC, Inc. (San Diego, CA) introduced a product called KnowledgeNet, which is an integrated
expert system and neural network approach which eliminate the "black box" approach to neural
network applications by providing explanations of network results. Neural networks often
complement expert systems. In practice, the best domain expert cannot think of every potential
problem, especially as circumstances change over time. For this reason, expert systems are typically

800

Jay Lee

useful over a limited range of problems and circumstances. The inability to generalize or adapt
to change has limited the success of expert systems. Neural networks, on the other hand, have
shown the ability to learn without explicit programming or knowledge engineering. Neural
networks learn by being presented with training sets consisting of input patterns and correct
interpretations of those patterns. The neural networks organize themselves in order to minimise
the difference between their outputs and the "ideal" outputs provided in the learning set.
These two technologies have very different strengths and weakness, suggesting the hybrid
systems taking advantage of the two technologies complementary abilities would be extremely
valuable.
Ford Motors has developed automotive control system diagnostics using neural networks for
rapid pattern classification of large data sets [13]. The data for training the systems were
obtained by introducing 26 different faults into the power plants (shorted, plug, open plug,
plugged injector, broken manifold pressure sensors, etc.) and observing the engine operation at
a fast idle. A backpropagation classification algorithm was used for fault recognition. However,
real-time fault detection and diagnostics have not been implemented due to the limitation of
the real-time data acquisition and learning.
Netrologic Inc. has also developed a neural network-based fault detection and diagnostic
tool for space systems in NASA. Feedforward neural networks were used to recognise the
signature of a valve in the ATLAS rocket (which contains some 150 valves of various types)
[12]. The data were collected in the form of electric current transient signatures from ATLAS
rocket valves installed on a pneumatic test bench. The data were then used to train and test
two neural networks: one was trained to distinguish among signatures for three separate types
of valves during a valve-open state change (rising current), and the other was trained to
distinguish among signatures for the same types of valves during a valve-closed state change
(falling current). The results show that the signature of the valve is very sensitive to the spring
adjustment and the working conditions (temperature, pressure, etc.). The problem of handling
failure during transients was not addressed. The test did not include faults, pressure variations,
multiple events interference, and electromagnetic interference effectsl
Expert systems approach
Expert systems have been used extensively to diagnose manufacturing equipment. General
Motors has developed and implemented an expert system called "Charley" to track machine
conditions using vibration signatures from the machines [38]. Love and Simaan [39] integrated
statistical process control principles and knowledge of the process to automatically arrive at a
comprehensive method to detect and diagnose out-of-control conditions in a manufacturing
process. This approach consists of capturing data from the process and passing selected signals
from it through a two-level decision-making system. Holloway and Krogh [40] proposed a
behavioral model approach for fault detection and diagnosis in automated manufacturing
systems. Their model provides the basis for on-line fault detection by generating expected
system response signals which are compared on-line, in real-time, with the actual sensor signals
from the system. Fault diagnosis is accomplished by maintaining a current set of operational
assumptions which identify the system components which could cause deviations from the
expected behavior. Jones et al. [41] proposed qualitative and quantitative approaches to diagnose plant faults. Jowers and Painter [42] developed a symbolic diagnostics method using an
object-oriented programming technique. Yue and Lirov [43] from AT&T proposed a semantic
controller approach for a computerized strategy choice aiding tool architecture for metadiagnosis. Arjunan [44] proposed a method to diagnose multiple faults, especially simultaneous occurrence of multiple faults. Pattipati and Alexandridis [45] integrated information theory and
heuristic and/or graph search methods to subdue the computational explosion of the optimal
test sequencing problem. Su and Govindaraj [46] performed experiments on a training simulator for large dynamic system diagnosis. Narayanan and Viswanadham [47] developed a
methodology to merge graph propagation. The model contains a hierarchical representation of
available knowledge about the characteristics of fault propagation within the system. Each level
in the hierarchy contains one or more structures which together represent a view of the system
under a particular granularity. The granularity of view increases with levels.

Modern computer-aided maintenance

801

For a complex decision-making environment, Jacob and Pirkul [48] proposed a network of
knowledge-based systems. Ishida et al. [49] proposed a topological approach based on graph theory
for large-scale system diagnosis. Pazzani [50] illustrated an approach to failure-driven learning.
When a failure occurs, the diagnostic system uses the learning approach to identify which
information from the device models should be incorporated into new rules. There are several
popular diagnostics reasoning strategies. Structured reasoning is the simplest approach; decisions
are based solely on go/no go information and can be very efficient [51]. Qualitative reasoning uses
domain-specific criteria to translate the numerical data into an initial qualitative description [52].
Deep model reasoning is used to run a model of the equipment being diagnosed under various fault
assumptions [53, 54]. Fink and Lush [55] have illustrated how to couple deep models with
experimental knowledge.
In general, much research has been done in the area of fault diagnostics and fault detection. Little
research has been done in the area of in-process machine degradation monitoring for manufacturing systems. Some related research work in machine or system degradation and its impact to
proactive maintenance are reviewed and summarized as follows.

PERSPECTIVE ON PROACTIVE MAINTENANCE THROUGH DEGRADATION MONITORING--A NEW

THINKING PARADIGM

The proactive maintenance approach is to use integrated, investigative and corrective practices
to significantly extend machinery life. The goal of proactive maintenance is to eliminate failures
of equipment forever. If the behavior of a machine can be monitored adaptively, then an early
warning of possible faults can be generated. By doing this, maintenance personnel can perform
early diagnostics and part replacement during regular daily maintenance hours. Therefore, the
mean-time-between-failure can be extended to an unlimited length. Figure 5 shows the concept of
the proposed proactive maintenance implementation technique. When aging or wear occurs, the
component and machine generally progress through a series of degradation states before failure
or fault occurs. If a degradation can be measured and detected, then a proactive and corrective
maintenance activities can be performed before a worse degradation condition is reached that could
cause component failure and machine downtime.

I PROACTIVEMAINTENANCETECHNIQUE1
MONITOR THE DEGRADATIONOF MACHINE AND
SCHEDULE MAINTENANCEWORK DURING THE DAILY

REGULARMAINTENANCE
HOURS

DEGREE
OFFAILURE

PROBABILITY
OF FAILURE
-iI

MEAN-TIME-.E EEN-FAILURE I/

(MTBF)

DEGRADATION

I I
....

I i!

I . . . . . . .

lI

IlL I

D.~-.

~/

w,,ow.... /

_
BREAK
IN PERIOD - ' ~

',

I
I
"~
I
,

I
"tBEHAVIOR

MONITORING

RANGE t . . . .

-~I
I

TIME
Fig. 5. Proactive maintenance concept diagram.

FAILURE
LINE

802

Jay Lee

Methods for monitoring degradation

Degradation is defined as a decreased performance level when a parameter is measured over
in a specific deviation. Typically, a component can go through a series of degradation states
before it fails. There are two types of components in a machine--active components (e.g.
motors, sensors and valves), and passive components (e.g. structures and bearings). Subudhi
and Taylor [20] at Brookhaven National Laboratory tested electric motors to monitor age-related
degradation in nuclear power plants. The plug reverse cycle test method was used to induce
accelerated aging in the motor subcomponents. The motor winding temperature, running current,
and bearing temperature were continuously monitored. Test results showed that winding
temperature is a good indicator for monitoring degradation of insulation. A high bearing
temperature or vibration excursion is a clear indication of bearing wear, misalignment or their aging
mechanisms.
Hoopingarner and Burns [21] at Pacific Northwest Laboratory studied in ways to monitor aging
of diesel generators in nuclear plants. Haynes [22-24] at Oak Ridge National Laboratory has also
conducted degradation research on motor-operated valves. His work monitored the valve stem
position (using LVDT), springback displacement, actuator output torque, motor current and
time of actuation of all control switches. The research led to the conclusion that the single
most informative measurement parameter was also the one most easily acquired: the motor
current.
Samanta et al. [25] at Brookhaven National Laboratory developed a degradation model using
a Markov approach to analyze historic component degradation rates and failure rate data for
nuclear power plants. This research showed that degradation can be modeled to identify aging
effects. Analysis of their research shows that besides data on components failure, significant
information exists on component degraded conditions, including times at which such degraded
conditions are observed, and also values of observed parameters that indicate degradation. Often,
a component reliability record contains much more information on degradation than failure. The
concept behind Samanta's work was to use degradation data from the past maintenance record
to develop component reliability characteristics and to understand aging effects. This research
treated degradation as a discrete rather than a continuous process.
A surface layer activation (SLA) technique was developed by Blatchley and Sioshansi [26] to
monitor wear, erosion and corrosion of parts from outside a running machine. Gamma rays
penetrate the material surrounding the activated spot; therefore, the amount of material worn from
the part can be reliably measured by monitoring the intensity of ~-rays from an outside location.
This method is widely used for diagnostic monitoring of critical machine parts to check wear and
corrosion. Moving engine parts and mechanical power train components have been monitored by
SLA techniques. Most recently, Lee and Kramer [57] proposed a methodology in using the CMAC
(Cerebellar Model for Articulation Controller) neural networks to monitor the machine behavior.
A Pattern Discrimination Model (PDM) was used to measure the performance degradation
quantitatively. Figure 6 shows the system concept of the proposed methodology, an integrated
learning, monitoring and recognition technique, for a proactive maintenance system. The Cerebellar Model Articulation Controller (CMAC) is used to adaptively learn machine behavior and a
Pattern Discrimination Model (PDM) based on the CMAC is used to monitor, recognize and
quantify behavioral changes. The PDM serves as a watchdog to monitor the behavior of the
machine by using a confidence value that represents the conditional probability of degradation. The
fault then can be detected by comparing the confidence value of the PDM output with a threshold
confidence value.
The new machine behavioral learning, degradation monitoring and fault detection technique was
tested and implemented in monitoring the position of a stepping motor encoder and degradation
in the accuracy of a robot. To examine the developed CMAC-PDM and to investigate its capability
for monitoring degradation and detecting faults, several test patterns were generated by setting the
faults and degradations through different positions of the input encoder value. Seven different
degradation test patterns were generated by adjusting the backlash screws of the robot arms and
wrist. The experiments were conducted in the Robotics Laboratory at the University of Maryland--College Park. The following sections examine the capability and sensitivity of the developed
technique in monitoring degradation and detecting faults.

Modern computer-aided maintenance

Event-Based
Process
Control

803

tt

Time

Sensors

\
( Machine ~
[
or
=

s2Sl
\
| A1

~~ Acquisition
Learning

Degreeof ]

|ehav!oml
Pattern Recognition

Degradation

Monitoring

iP

Fault

Detection

Actuators
Pattern
Dl=crlm!naUon

CerebellarModal
ArticulationController

M ~ I (PDM)

(CMAC)

TABLE LOOK-UP

NEURAL NETWORKS
Fig. 6. System concept of machine degradation monitoring and fault detection using neural networks
approach.

EXPERIMENT I--ENCODER POSITION MONITORING AND FAULT DETECTION

The encoder is one of the key elements in a motion control system to provide position feedback
for motor or positioning tables. For high-precision motion control systems, such as chip inspection
and electronics assembly, the assurance of encoder feedback is critical. Tests were conducted to
investigate the feasibility of monitoring degradation and faults by using the developed technique.

Experiment setup
In Fig. 7, a stepping motor with an encoder was powered through an HP power amplifier and
controlled by a PC. The purpose of this experiment was to examine the developed technique and
investigate the relationship among several CMAC parameters, such as address array size,
confidence value (performance indicator) and convergence.
There are 13 samples and 14 inputs (one motor speed and 13 position patterns) for CMAC
training. Table 1 shows that the encoder has the same input patterns for 13 different rotating speeds.

HP POWERAMPLIFIER

//PPI%G:OTO

//~./~, '~

PTICALENCODER

TRANSISTORS (X4) - . k ~ ~

Pc

OPTO-ISOLATOR

Fig. 7. Experimental setup for encoder position monitoring and fault detection.

Jay Lee

804

Table I. Data used for training

14 Inputs for CMAC training
CMAC

Motor
speed
3.5165
4,0110
4,5055
5,0000
5.4945
6,0440
6,4835
6.9780
7.5275
8.0220
8.5165
9.0110
9.5055

Desired
output
(assigned value)

Encoder position signal pattern

(13 patterns)
1,0
1.0
1,0
1.0
1.0
1,0
1,0
1,0
1,0
1.0
1.0
1,0
1,0

0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0

1.0
1.0
1.0
1.0
1.0
1.0
1.0
1.0
1.0
1.0
1.0
1.0
1.0

0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0

1.0
1.0
1.0
1.0
1.0
1.0
1.0
1.0
1.0
1.0
1.0
1.0
1.0

0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0

1.0
1.0
1.0
1.0
1.0
1.0
1.0
1.0
1.0
1.0
1.0
1.0
1.0

0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0

1.0
1.0
1.0
1.0
1.0
1.0
1.0
1.0
1.0
1.0
1.0
1.0
1.0

0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0

1.0
1.0
1.0
1.0
1.0
1.0
1.0
1.0
1.0
1.0
1.0
1.0
1.0

0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0

1.0
1.0
1.0
1.0
1.0
1.0
1.0
1.0
1.0
1.0
1.0
1.0
1.0

70,000.000
80,000.000
90,000,000
100,000.000
I10,000.000
120,000.000
130,000.000
140,000.000
150,000.000
160,000,000
170,000.000
180,000.000
190,000.000

The encoder output pattern should be the same at different rotating speeds. This set of data was
used to train the CMAC. Once the behavior of the encoder is trained, the CMAC learns the
behavioral patterns and maps the patterns in the weight table. During training, the PDM always
has the correct pattern and thus the confidence value is 100%. During the testing, however, if there
is a fault or degradation, the pattern may be changed. The confidence value would also be changed
based on the degree of the degradation or fault conditions. A lower value indicates a higher
probability of fault or degradation.
Table 2. Testing data fault detectoin and degradation recognition
Fault pattern
Data set A

1
2
3
4
5
6
7
8
9
10

5.0000
5.0000
5.0000
5.0000
5.0000
5.0000
5.0000
5.0000
5.0000
5.0000

1.0
1.0
1.0
1.0
1.0
1.0
1.0
1.0
1.0
0.0

1.0
0.0
0.0
0.0
0.0
0.0
1.0
1.0
1.0
1.0

1.0
1.0
1.0
1.0
1.0
1.0
1.0
1.0
1.0
0.0

0.0
1.0
0.0
0.0
0.0
0.0
0.0
0.0
1.0
1.0

Data set B

II
12
13
14
15
16
17
18
19
20

5.0000
5.0000
5.0000
5.0000
5.0000
5.0000
5.0000
5.0000
5.0000
5.0000

1.0
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1

0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0

1.0
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1

0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0

1.0
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1

0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0

1.0
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1

Data set C

21
22
23
24
25
26
27
28
29
30

5.0000
5.0000
5.0000
5.0000
5.0000
5.0000
5.0000
5.0000
5.0000
5.0000

0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0

0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0

1.0
1.0
1.0
1.0
1.0
1.0
1.0
1.0
1.0
1.0

0.0
0.0
0,0
0.0
0,0
0,0
0.0
0,0
0,0
0.0

1.0
1.0
1.0
1.0
1.0
1.0
1.0
1.0
1.0
1.0

0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0

Data set D

31
32
33
34
35
36
37
38
39
40

5.0000
5.0000
5,0000
5.0000
5.0000
5.0000
5.0000
5.0000
5.0000
5.0000

1.0
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.0

0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1

1,0
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1

0,1
0,1
0,1
0.1
0,1
0.1
0.1
0,1
0.1
0,1

1.0
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1

0.1
0.1
0.1
0.1
0.1
0.1
0. I
0.1
0.1
0.1

1.0
1.0
1.0
1.0
1.0
1.0
1.0
1.0
1.0
0.0

0.0
0.0
1.0
0.0
0.0
0.0
1.0
1.0
1.0
1.0

1.0
1.0
1.0
1.0
1.0
1.0
1.0
1.0
1.0
0.0

0.0
0.0
0.0
1.0
0.0
0.0
0.0
0.0
1.0
1.0

1.0
1.0
1.0
1.0
1.0
1.0
1.0
1.0
1.0
0.0

0.0
0.0
0.0
0.0
1.0
0.0
0.0
1.0
1.0
1.0

1.0
1.0
1.0
1.0
1.0
1.0
1.0
1.0
1.0
0.0

0.0
0.0
0.0
0.0
0.0
1.0
0.0
0.0
1.0
1.0

1.0
1.0
1.0
1.0
1.0
1.0
1.0
1.0
1.0
0.0

0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0

1.0
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1

0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0

1.0
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0. I

0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0

1.0
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1

1.0
1.0
1.0
1.0
1.0
1.0
1.0
1.0
1.0
1.0

0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0

1.0
1.0
1.0
1.0
1.0
1.0
1.0
1.0
1.0
1.0

0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0

1.0
1.0
1.0
1.0
1.0
1.0
1.0
1.0
1.0
1.0

0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0

1.0
1.0
1.0
1.0
1.0
1.0
1.0
1.0
1.0
1.0

1.0
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1

0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1

1.0
0.9
0,8
0.7
0.6
0.5
0.4
0.3
0.2
0.1

0.1
0.1
0.1
0,1
0.1
0.1
0.1
0.1
0.1
0.1

1.0
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1

0.1
0.1
0.1
0,1
0.1
0.1
0.1
0.1
0.1
0.1

1.0
0.9
0.8
0,7
0.6
0.5
0.4
0.3
0.2
0.1

Degradation patterns

805

M o d e r n computer-aided maintenance
Table 3. Testing results--confidence value for different address table sizes
Address table size
20

30

40

50

60

70

80

90

100

Data set A

I
2
3
4
5
6
7
8
9
10

0
5
5
5
0
5
0
10
0
0

36.67
33.33
33.33
36.67
36.67
36.67
40
36.67
33.33
36.67

55
52.5
50
50
55
52.5
50
57.5
50
50

70
62
62
64
60
60
62
60
62
62

68.33
68.33
73.33
68.33
68.33
71.67
70
66.67
73.33
68.33

78.57
72.86
71.43
74.29
74.29
75.71
75.71
71.43
72.86
75.71

77.5
75
78.75
78.75
78.75
75
76.25
77.5
76.25
77.5

82.22
78.89
77.78
82.22
78.89
78.89
80
80
81.11
77.78

84
83
83
82
81
83
80
83
83
81

Data set B

I1
12
13
14
15
16
17
18
19
20

100
85
80
65
60
50
40
30
20
10

100
90
86.67
76.67
73.33
66.67
60
53.33
46.67
40

100
92.5
90
82.5
80
75
70
65
60
55

100
94
92
86
84
80
76
72
70
66

100
95
93.33
88.33
88.33
86.67
83.33
80
76.67
73.33

100
95.71
94.29
90
88.57
85.71
82.86
80
77.14
74.29

100
96.25
95
91.25
90
87.5
85
82.5
80
77.5

100
96.67
95.56
92.22
92.22
90
87.78
85.56
83.33
81.11

100
97
96
93
92
90
88
86
85
83

Data set C

21
22
23
24
25
26
27
28
29
30

85
80
65
60
50
40
30
20
10
0

90
86.67
76.67
73.33
70
66.67
60
53.33
46.67
40

92.5
90
82.5
80
77.5
72.5
67.5
62.5
57.5
52.5

94
92
86
84
80
78
76
72
70
66

96.67
95
90
88.33
85
81.67
78.33
75
73.33
71.67

97.14
95.71
91.43
90
87.14
84.29
81.43
78.57
75.71
72.86

97.5
96.25
92.5
91.25
88.75
86.25
85
82.5
80
77.5

96.67
95.56
92.22
91.11
88.89
87.78
85.56
83.33
81.11
78.89

97
96
93
92
90
88
86
84
83
81

Data set D

31
32
33
34
35
36
37
38
39
40

90
75
70
55
50
40
30
20
10
0

93.33
83.33
80
70
66.67
60
53.33
46.67
40
33.33

95
87.5
85
77.5
75
70
65
60
55
50

96
90
88
82
80
76
72
68
66
62

98.33
93.33
91.67
86.67
86.67
85
81.67
78.33
75
70

97.14
92.86
91.43
87.14
85.71
82.86
80
77.14
74.29
71.43

97.5
93.75
92.5
88.75
87.5
85
82.5
80
77.5
75

98.89
95.56
94.44
91.11
91.11
88.89
86.67
84.44
82.22
78.89

98
95
94
91
90
88
86
84
83
81

Test results

To examine the developed CMAC-PDM for degradation and fault detection, several fault
conditions were created by setting the fault through different positions of the input encoder value.
Table 2 shows the test fault patterns and degradation patterns. In this table, a "0" represents a
fault. Values between "1" and "0" represent degraded conditions. Table 3 shows the confidence
values for each fault and degradation condition when a different address array sizes were used in
the CMAC. It shows that address array sizes 20 and 30 can better detect fault and degradation
by providing lower confidence values.
EXPERIMENT

2--MEASUREMENT

OF ROBOT DEGRADATION

The developed model was tested and implemented in monitoring degradation in the accuracy
of a robot (PUMA 560). Position accuracy, position repeatability and path straightness are often
evaluated and stated in statistical terms. This allows robot manufacturers and machine tool
manufacturers to predict the positioning behavior of various moving elements to a stated
confidence level, and provides the customer with an assurance of the machine's performance.
Typically, poor position accuracy and path straightness are symptoms of the degraded performance
which may be due to mechanical wear, backlash of the gear train, vibration and thermal effects.
In this experiment, the position accuracy and path straightness were used as evaluation criteria for
measuring robot performance.
To examine the developed CMAC-PDM and to investigate its capability for analyzing and
monitoring the performance degradation of a robots, seven different degradation test patterns were

806

Jay Lee

560

Fig. 8. Robot position accuracy degradation measurement using laser interferometer.

generated by adjusting the backlash screws of the robot arms and wrist. Figure 8 shows the
experimental setup. The HP 5518A (two-frequency He-Ne laser) interferometer was used to
measure the robot performance (position accuracy and path straightness). To cause degradation
conditions, the backlash screws were adjusted for the arms and the wrist of the robot. Once
backlash was introduced, the robot path straightness and position accuracy were measured by the
laser interferometer. Before adjusting the backlash of the robot arm, a set of path straightness and
position accuracy data was generated to train the CMAC by programming the robot to follow the
taught path pattern.
A degraded test pattern was generated by introducing different levels of backlash adjustment.
Seven adjustments were made in this experiment, with each level of adjustment representing a
different degree of degradation (see Fig. 9). These sets of patterns were used to test the
CMAC-PDM and to investigate its capability in monitoring degradation. Figure 10 shows the
confidence plot for each level of adjustment. The graph indicates that an array sizes of 20 clearly
shows the measured variation in the accuracy. For address array size 20, the test was repeated for
different weight table sizes. Figure 10 shows the confidence plot for these seven levels of
degradation.
RESEARCH NEEDS
The National Research Council's report, The Competitive Edge: Research Priorities of U.S.
Manufacturing [2] lists five research topics most important for manufacturing competitiveness. They
are intelligent manufacturing control, equipment reliability and maintenance, manufacturing of and
with advanced engineered materials, the product realization process, and manufacturing skills
improvement. In regard to equipment reliability and maintenance, it states that there are two
categories: people-related needs and equipment-related needs. Equipment-related needs are
grouped under the categories of performance measurement, tools and techniques and methodologies for equipment reliability and maintenance. The most important research in performance
measurement is in on-line monitoring of equipment and processes.
The author believes that the development of in-process monitoring of machine degradation and
faults is one of the most important research tasks for increasing machine uptime and improving
production quality. One of the major problems in monitoring processes and machines is that the
degradation of the sensors, actuators and other components cannot be monitored using conventional techniques. The conventional methods of fault detection in manufacturing processes only
permit recognition when limit values of measurable output signals (primarily on sensors) have
already been transgressed. Most of the fault detection schemes are based on the logic diagram, state
table and rule-based event. A fault is detected when it occurs. Therefore, the operation has to be

0.001

0.0027

0.004

0.0015
0.002

0.126

0.08

0.12

0.087

0.11

0.074

4.54

0.002

0.003

0.0024

0.003

0.002

0.125

ADJ. 2

ADJ. 3

ADJ. 4

ADJ. 5

ADJ. 6

ADJ. 7

3.9

0.12

4.3
0.199

4.4

0.125

4.1

0.12

48

0.005
0.0001
0.008
0.002
0.05

0.001

0.05

0.001

0.068

0.05
0.0014
0.122
0.0035

0.08

0.1

0.003

0.05

0.0009

8
0.02

0.0002
0.002

0.045

0.0002

0.06

0.0016

o
0.02

0.04

0.07

0.001

0.04

0.001

0.06

0.026

0.06

0.002

0.09

0.013

0.0001
0.06

0.0018

0.0176

0.0005

0.06

0.001

0.03

0.004

0.001

0.019

0.0005

0.05

Straightness

0.085

Position
accuracy

0.0024

Straightness

Travel distance 3

0.04

Position
accuracy

Travel distance 2

Fig. 9. T e s t i n g d a t a sets for different b a c k l a s h a d j u s t m e n t c o n d i t i o n s .

0.001

0.07

Straightness

0.0019

Position
accuracy

Travel distance 1

ADJ. 1

Backlash
adjustment
condition

808

Jay Lee
120 -

W e i g h t table held c o n s t a n t at 1024

100

~. 80
~ 6o
~

40
v

"~

20

Test n u m b e r
Fig. 10. Test set vs confidence.

stopped to diagnose the problem. The machine also has to be stopped or reset when the fault is
detected. It is very difficult to perform fault-tolerant process control if the behavior of the machine
or process is not monitored and learned continuously. Today's industrial controller has limited
capabilities in monitoring the degraded machines and processes in a computer-integrated environment. It is also very difficult to justify the use of high-tech maintenance technology such as
intelligent sensing and remote diagnosis to monitor and detect faults because of high cost and high
degree of complexity of these approaches [14-18].
In addition to these problems, several real-life problems also needed to be considered:
(I) Can an economical system be developed?
Artificial intelligence systems and neural networks have traditionally required large, expensive
computers and long development times. The development of a classifier, such as backpropagation,

CMAC NEURAL NETWORK

CONTROLLER

COMMAND

~~

EHAVIORAL LEARNING

PATTERN
DISCRIMINATION MODEL
(PDM)
RECOGNITION &
MONITQR[N~
MEASURE
CONFIDENCEVALUE
[C~ AT tl, t2, ~,...
IF
.'v(tl )>cv(t2),cv(t3)>..

DEGRADATION

ACTUATORS
MACHINE
SENSORS

IF
CV(I) < CV(threshold)

STATISTICAL
PROCESS
CONTROL(SPC)

EXTERNAL SENSORS

HUMAN OBSERVA~ONS

Fig. 11. Machine degradation and fault detection technique.

FAULT
DETECTION

Modern computer-aided maintenance

809

is very time-consuming and expensive because of the knowledge required to code the software. This
aspect of the problem involves the number of examples of faults and perhaps the number of
different conditions which must be included in the training set in order to obtain satisfactory
performance in testing. Normally, it is difficult to justify this approach to fault detection because
it is expensive to develop and not easily understood. Therefore, a simpler neural network approach
such as a learning function based on a table lookup method and lower development cost are the
most desirable features for implementing machine degradation monitoring and fault detection in
a practical manufacturing environment. Figure 11 shows a system concept of machine degradation
monitoring and fault detection using a Cerebellar Model Articulation Controller (CMAC) neural
networks [57] which has been described previously. In this approach, CMAC is a real-time data
acquisition and learning tool for the behavior of the actuators and sensors and PDM is a watchdog
in monitoring the behavior pattern by comparing the current performance confidence value to a
previous one. The unique part of this approach is the use of CMAC to learn the good or normal
pattern under different working conditions rather than to learn the bad or wrong pattern. For many
new machines, the bad or wrong condition is unpredictable. Therefore the implementation of
proactive maintenance could be quite simple and economical.
(2) How can humans be integrated into such a system in a way that is easy to extend and which applies
to specific production situations?

Workers and union problems must be considered. A human-based technology implementation

approach will encourage the workers to participate in corporate automation efforts, It becomes
apparent, from experience, that the knowledge gap in the organization is so great that the
technology for intelligent monitoring will have to be integrated seamlessly and invisibly into the
machines themselves. Most of the current research work in intelligent manufacturing is trying to
eliminate the workers from the control loop. In fact, the success of any computer-integrated
manufacturing is determined by the continuous involvement of the workers, not computers. Figure
12 shows an approach in integrating machine learning and human reasoning for fault detection
and process monitoring in a proactive maintenance environment. In this approach, the operation

FAULT
"~ _ /'
STOP
ELIMINATION ~ O P E R A T I O N

.........

TOLERANT
CONTROL

i
_

PROCESS

FAULT

SENSING

SENSORS ~

CMAC

ACTUATORS

BEHAVIORAL

TIME

LEARNN
I G/
RECOGNT
IO
IN

~i~ii!iiiiiiiiiiiiiil
REASONING

REASONING

INTEGRATED
REASONING

~i~!ili !i i i~ili!il

INTELLIGENT MONITORING
AGENT (Watchdog)

Fig. 12. Fault detection and process monitoring using integrated learning and reasoning methodology.

810

Jay Lee

is first reasoned by a worker to determine the operatability of the machine. If an additional

assistance is needed, a rule-based computer reasoning could be utilized. The author has experienced
that the workers in the production line have more creativities if their value has been considered
in an operation system, such as a computer-integrated manufacturing system.

(3) How can the technology be successfully applied to the wide variety of production machines found
in a typical factory?
There are a variety of machines in the manufacturing factory and controller which range from
relays to the latest in microprocessor technology. The technology should be implemented in a
modular fashion so that the interface with machine can be integrated by using a building block
approach.

SUMMARY

This paper reviewed the research work in machine performance monitoring under the traditional
operation paradigm. A proactive maintenance approach is also examined to extend the operation
life of machine to an unlimited length by monitoring the degradation process of the machine or
manufacturing system. Finally, the research needs for machine performance monitoring was
presented.
REFERENCES
1. S. Shingo. POKA-YOKI (Zero-Defects Design). Productivity Press, Cambridge, MA (1989).
2. National Research Council. The Competitiveness Edge: Research Priorities for U.S. Manufacturing. National Academy
Press (1990).
3. S. T. Scott. A three-phase model for quality. Qual. Mag. Jan, 2-4 (1991).
4. H. Park. Assessing machine performance. Am. Mach. Jan, 39-42 (1992).
5. D. Shainin and P. Shainin. OK isn't not enough. Qual. Mag. Feb, 19-22 (1990).
6. E. P. Papadakis. A computer-automated statistical process control method with timely response. J. Engng Costs
Product. Econ. 18, 301-310 (1990).
7. M. Craig. Predicting and optimizing assembly variation. Qual. Mag. Jan, 16-18 (1991).
8. Equipment Should Improve Through Use. Am. Mach. Rept, Sept, 81-100 (1991).
9. G. Mendelbaum and R. Mizuno. Directions in maintenance. Mainten. Technol. Mag. Jan, 45-48 (1992).
10. J. G. Wohl. Cognitive capability versus system complexity in electronic maintenance. IEEE Trans. Syst. Man. Cybernet.
SMCI3 624-626 0983).
11. R. J. Meltzer. Sensor reliability of MTBF. Sensor Mag. Jan (1992).
12. Netrologic, Inc. Report on Space Transportation Analysis and Intelligent Space System, NASA SBIR NAS9-17995
(1990).
13. K. Marko. Automotive control system diagnostics using neural network for rapid pattern classification of large data
sets. Proc. Int. Neural Net Society Meeting, Washington, DC, pp. 13-16 (1989).
14. J. A. Franklin, R. S. Sutton and C. W. Anderson. Application of connectionist learning methods to manufacturing
process monitoring. Proc. IEEE 3rd Int. Syrup. on Intelligent Control, pp. 709-712. (1988).
15. Maintenance in the 90's--A Plan to Support the USPS Goal of Automation and Mechanization. USPS Report,
Washington, DC (1990).
16. E. F. Pardue, K. R. Piety and R. Moore. Element of reliability-based machinery maintenance. Sound Vibrat. May, 14-20
(1992).
17. J. C. Fitch. Contaminant monitoring: the overlooked predictive maintenance. Mainten. Technol. Mag. Jnn, 41-46
(1992).
18. J. Lee. Review of computer-aided predictive maintenance technology for machinery, USPS Technology Resource Dept.
Technical Report (1991).
19. Z. Kohavi. Switching and Finite Automata Theory. McGraw-Hill, New York (1978).
20. M. Subudhi and J. H. Taylor. Testing of electric motors for monitoring age related degradation. Nucl. Technol. 84,
5~61 (1989).
21. C. R. Hoopingamer and J. Burns. The Role of Monitoring Diesel Generator Aging. Pacific Northwest Lab. Report For
DOE, DE-AC06-76RLO-1830 (1992).
22. H. D. Haynes. Assessment of diagnostic method for determining degradation of motor-operated valves. Proceedings
of Aging Research Conf. (1992).
23. H. D. Haynes. Check valves: Oak Ridge's New Diagnostics. A S M E Mech. Engng May, 64--69 (1991).
24. H. D. Haynes. Condition monitoring of machinery using motor current signal analysis. Sound Vibrat. Sept, 14-25
(1989).
25. P. K. Samanta, W. E. Vesely, M. Subudhi and F. Hsu. Degradation Modelling with Application to Aging and
Maintenance Effectiveness Evaluation. BNL-Nureh-52252, U.S. Nuclear Regulatory Commission Report, Washington,
DC (1991).
26. C. Blatchley and P. Sioshansi. Monitoring wear with gamma rays. Mach. Des. Oct, 99-102 (1990).

Modem computer-aided maintenance

811

27. J. Lee. Adaptive control tool monitoring in machining. SME Technical Paper MR86-131, Conference Proceeding of
Advanced Machining Technology for Cell and FMS, 5 (1986).
28. R. Isermann. Process fault detection based on modelling and estimation methods. Proc. Int. Federal. of Automatic
Control, pp. 387-403 (1984).
29. A. S. Willsky. A survey of design method for failure detection in dynamic system. Automatica 12, 601-611 (1976).
30. A. J. Barbera, M. L. Fitzgerald and J. S. Albus. Concepts for a real time sensory interactive control system architecture.
Proc. 15th Southeastern Syrup. on System Theory, April (1982).
31. W. Fairey. A fault finding hierarchy for PLC-controlled equipment. Programmable Control Magazine, pp. 44-45 (1991).
32. U. S. Bitite and A. Ross. PLC based diagnostics system for FMS. Proc. of 7th Int. Conf. on FMS, pp. 201-211 (1988).
33. R. Valette, J. Cardeso and D. Dubois. Monitoring manufacturing by means of Petri-nets. Proc. IEEE Conf. On
Intelligent Control 0989).
34. A. J. Barbera. Computer Simulation, Design and Control Tool. Technical Notes, Advanced Technology Research Corp.
(1990).
35. W. E. Dietz, E. I. Keich and M. All. Jet and rocket engine fault diagnosis in real time. J. Neut. Network Comput. Sum,
5-19 (1989).
36. F. Stevenson and D. Greenwood. Tool Wear Estimation Using Neural Networks. Netrologic, Inc. Report, San Diego,
CA.
37. R. E. Uhrig and Z. Guo. Use of neural networks to identify transient operating conditions in nuclear plants. Proc.
SPIE, Vol. 1095, Application of Artifical Intelligence VII, pp. 851-856 (1989).
38. R. Marczewski. An expert system for machine tool diagnosis. Proc. 2nd Engng Soc. of Detroit Conf. on Expert System,
Detroit, pp. 161-172 0988).
39. P. Love and M. Simaan. A knowledge-based approach for detection and diagnosis out-of-control events in
manufacturing processes. Proc. IEEE 3rd Int. Conf. in Intelligent Control, pp. 736-741 (1989).
40. L. E. Holloway and B. H. Krogh. Fault detection and diagnosis in manufacturing system: a behavioral model approach.
Proc. the 2nd Int. Conf. on CIM, Troy, NY, pp. 252-259 (1990).
4l. A. H. Jones, B. Porter and R. N. Fripp. Qualitative and quantitative approach to the diagnosis of plant faults. Proc.
IEEE 3rd Int. Syrup. on Intelligent Control, pp. 87-92 (1989).
42. S. Jowers and J. H. Painter. Symbolic diagnosis for intelligent control. Proc. IEEE 3rd Int. Syrup. on Intelligent Control,
pp. 280-290 (1989).
43. O. Yue and Y. Lirov. Metadiagnosis. Proc. IEEE 3rd Int. Syrup. lntell., pp. 303-306 (1989).
44. M. M. Arjunan. Diagnosing multiple faults in intelligent control and automated systems. Proc. IEEE 3rd Int. Syrup.
Imell. Control, pp. 93-97 (1989).
45. K. R. Pattipati and M. Alexandrisis. Application of heuristic and information theory to sequential fault diagnosis. Proc.
IEEE 3rd Int. Syrup. On Intell. Control, pp. 291-296 (1989).
46. Y. L. Su and T. Govindaraj. Fault diagnosis in a large dynamic system: experiments on a training simulator. IEEE
Trans. On Syst. Man. Cybernet. SMC-16 (1986).
47. N. H. Narayanan and N. Viswanadham. A methodology for knowledge acquisition and reasoning in failure analysis
of system. 1EEE Trans. Syst. Man. Cybernet. SMC-17 (1987).
48. V. Jacob and H. Pirkul. A framework for networked knowledge-based system. 1EEE Trans. On Syst. Man. Cybernet.
20, 1990.
49. Y. Ishida, N. Adachi and H. Tokumaru. A topological approach to failure diagnosis of large scale system. IEEE Trans.
On Syst. Man. Cybernet. SMC-15, 129-141 (1985).
50. M. J. Pazzani. Failure-driven learning for fault diagnosis heuristics. IEEE Trans. Syst. Man. Cybernet. SMC-17,
380 394 0987).
51. R. Milne. Strategies for diagnosis. IEEE Trans. Syst. Man. Cybernet. SMC-17, 333-339 (1987).
52. K. D. Forbus. Interpreting observations of physical systems. IEEE Trans. Syst. Man. Cybernet. SMC-17, 350-359
0987).
53. W. C. Yoon and J. M. Hammer. Deep-reasoning fault diagnosis: an aid and a model. IEEE Trans. Syst. Man. Cybernet.
18 (1988).
54. S. Post and A. Sage. An overview of automated reasoning. IEEE Trans, Syst. Man. Cybernet. 20, 202-224 (1990).
55. J. C. Lush and P. K. Fink. Expert system and diagnostic expertise in the mechanical and electrical domains. IEEE Trans.
Syst. Man. Cybernet. SMC-17, 340-349 (1987).
56. J. Lee. In process machine degradation monitoring and fault detection using a neural networks approach. D.Sc.
Dissertation, George Washington University (1992).
57. J. Lee and B. M. Kramer. Analysis of machine degradation using a neural networks approach. J. Manufact. Syst. 12
(1993).