nd J I A N J U N S H I
2
1
Department of Systems and Industrial Engineering, The University of Arizona, P.O. Box 210020,
Tucson, Arizona 85721-0020
2
Department of Industrial and Operations Engineering, The University of Michigan, Ann Arbor,
MI 48109-2117
Received August 1999 and accepted April 2000
In this paper, a new methodology is presented for developing a diagnostic system using waveform
signals with limited or with no prior fault information. The key issues studied in this paper are
automatic fault detection, optimal feature extraction, optimal feature subset selection, and diagnostic
performance assessment. By using this methodology, a diagnostic system can be developed and its
performance is continuously improved as the knowledge of process faults is automatically
accumulated during production. As a real example, the tonnage signal analysis for stamping process
monitoring isprovided to demonstrate the implementation of thismethodology.
Keywords: Automatic feature extraction, Haar transform, waveform signals, process monitoring,
fault diagnosis
1. Introduction
Monitoring and diagnostic systems have played an
important role in modern manufacturing process
control. Many intelligent or knowledge-based systems
have been successfully developed for different
application domains. However, the development of
such a system normally requires suf®cient prior
knowledge or fault condition data, which ishard to
satisfy in manufacturing systems. This is especially
true for a new product or process launch. Thus, the
motivation of the research presented in this paper is to
address this important issue by developing a metho-
dology for monitoring and diagnostic system
development with limited or with no prior fault
information. Waveform signals are used as the
essential information for the diagnostic system
development.
Waveform signals represent a class of analog or
digital signals over time, which normally can be
measured using in-process sensors in a manufacturing
process. It has broad potential applications, such as
tonnage signals in stamping, torque signals in tapping,
and force signals in welding, which are shown in Fig.
1 (a)±(c), respectively. In general, those waveform
signals contain rich information that can be related to
both product quality and process variables. The
characteristics of those waveform signals studied in
this paper are summarized as follows:
*
Non-stationary.
*
Working cycle-based signals, meaning that each
cycle of a waveform signal covers a complete
cycle of an operation.
*
Segmental signals, meaning that different seg-
ments represent different process stages, which
may have different potential process faults.
*
Localized time and frequency componentsin
different segments.
*
In-process automatic sensing. There are two basic approaches in diagnostic system
development by using waveform signals: a model-
based approach and a feature-based approach. In the
model-based approach, observations are considered as
a time-ordered stochastic process. The critical
concern of using this approach is to have an
appropriate process model which is sensitive to
process faults but robust to process noises (Deibert
and Hol¯ing, 1992). In addition, fault modelsor fault
signal characteristics need to be known before making
fault detection. However, these types of information
are normally not available at the beginning of the
production due to the complex relationship between
waveform signals and the associated manufacturing
process. Thus, a model-based approach is generally
not effective when waveform signals are used for
diagnostic system development.
A feature-based approach is more suitable to a
complex process where waveform signals are used for
process diagnosis. In such a system, features are
considered as random variables or as a random set.
Feature extraction and feature subset selection are
critical steps to reduce the number of attributes or data
dimension considered in the decision-making step
(Kharin, 1992). The conventional procedure to
develop a feature-based diagnostic system is shown
in Fig. 2. The essential requirement, or precondition,
to use this conventional procedure is that suf®cient
historical fault data or prior fault knowledge are
available before developing a diagnostic system. In
many applications, this precondition is not satis®ed
especially during the new machine or process launch.
Therefore, it isa challenging problem to develop a
feature-based diagnostic system with limited or with
no prior fault information.
In thispaper, we will propose an automatic feature
extraction methodology for the development of a
feature-based diagnostic system using waveform
signals. In this methodology, the wavelet analysis is
used as a basic tool for feature extraction. The wavelet
analysis is selected for this research due to its
multiresolution nature, its localized properties in
both time and frequency domains, its fast algorithms
ready for an on-line implementation, and itsef®cient
data compression for feature extraction. The Haar
transform is selected in the paper because it has an
explicit geometrical interpretation for a detected
change of a Haar coef®cient. These interpretations
can be easily associated with the pro®le change of a
waveform signal due to process faults.
Wavelet analysis has been widely used in image
and speech processing for decades. Much research has
been focused on the data shrinkage and signal noise
®ltering (Coifman and Yale, 1992; Donoho and
Johnstone, 1994). The wavelet transform used as a
feature extraction method hasrecently received more
and more attention for process monitoring in
manufacturing processes, such as drill condition
monitoring (Tansel et al., 1993), face milling failure
Fig. 1. Waveform signals measured in the different production
processes.
Fig. 2. Procedures for a feature-based fault diagnosis system
development.
258
Jin and Shi detection (Kasashima et al., 1995), and spalling
detection on ball bearings(Mori et al., 1996). In
these applications, the feature extraction method or
wavelet coef®cient (or a function of them) selection is
mostly based on engineering knowledge, or based on
the use of conventional trial-and-error approaches
when suf®cient prior fault data are available. The
relevant coef®cients associated with characteristic
frequencies of a dynamic system are usually selected
asfeatures. However, for a complex manufacturing
process, there is no suf®cient prior knowledge to
describe the complex relationship between waveform
signals and process faults. Thus, it is very dif®cult to
pre-determine the process characteristic frequencies,
and the trial-and-error approach cannot be imple-
mented.
Thispaper pres
entsa new methodology for
developing a diagnostic system with limited or with
no prior fault information. In thismethodology, a
monitoring decision for detecting process faults is
made ®rst based on normal production condition.
Then, the detected fault is further classi®ed, and the
knowledge of process faults is continuously accumu-
lated during the use of this diagnostic system in
production. When a new process fault is found, the
optimal feature subset is adaptively updated to include
the new characteristics of the newly detected fault.
Thus, the diagnostic capability could be continuously
improved asnew fault data are automatically
accumulated and classi®ed.
Ascan be seen, the proposed methodology isto
emphasize how to continuously improve diagnostic
ability through machine learning. The major research
issues discussed in the paper show the common
research problems of applying arti®cial intelligence to
machine learning for the process monitoring and fault
diagnosis purpose. An adaptive supervisory learning
method isused in thispaper for feature extraction and
fault classi®cation. The capability of adaptive
learning of process fault knowledge during routine
manufacturing production can be seen from the
following two aspects: (1) The increase of the fault
samples in the known fault clusters can ®ne-tune the
parameters of an existing classi®er to improve its
diagnostic accuracy; and (2) The addition of the newly
identi®ed fault patternscan enhance itsdiagnostic
ability for more process fault diagosis. More
discussions on these features will be given in
Section 3.
The outline of this paper is listed as follows. An
overview of the new methodology isgiven in Section
2. Then, Section 3 providesdetailed res
earch
proceduresfor the methodology development. After
that, a real example of stamping processes is provided
in Section 4. Finally, a summary and information
about future work are given in Section 5.
2. Methodology overview
The novel idea of thispaper isto develop a
methodology for developing a diagnostic system by
using waveform signals with limited or with no prior
fault information. The basic principles and generic
proceduresare shown in Fig. 3. In thismethodology,
there isno requirement of prior knowledge of process
faults. Thus, in-process information assessment, fault
classi®cation, and fault knowledge accumulation are
very critical. The essential idea of this method is that
more and more knowledge about process faults can be
accumulated during the use of the process monitoring
system. That knowledge is then used to further
improve the diagnostic performance. For this purpose,
in-process information assessment needs to be
addressed in step 1 to judge whether the current
measurement re¯ects a normal working condition, a
known fault, or a new fault. If the current measure-
ment represents a normal working condition or a
known fault condition, it will be added into the cluster
Fig. 3. The framework of diagnostic system development with
adaptive learning.
Automatic feature extraction
259 of the normal working conditionsor the known fault
conditions respectively. The inclusion of this new
sample will increase the sample size in th