864 IEEE TRANSACTIONS ON RELIABILITY, VOL. 60, NO.

4, DECEMBER 2011

Mahalanobis-Taguchi System as a Multi-Sensor

Based Decision Making Prognostics Tool for
Centrifugal Pump Failures
Ahmet Soylemezoglu, Sarangapani Jagannathan, Senior Member, IEEE, and Can Saygin

Abstract—A novel Mahalanobis Taguchi System (MTS) based FN false negatives

fault detection, isolation, and prognostics scheme is presented. The
proposed scheme fuses data from multiple sensors into a single FP false positives
system level performance metric using Mahalanobis Distance
(MD), and generates fault clusters based on MD values. MD LES large eddy simulation
thresholds derived from the clustering analysis are used for fault MD Mahalanobis distance
detection and isolation. When a fault is detected, the prognostics
scheme, which monitors the progression of the MD values over MTGS Mahalanobis-Taguchi Gram-Schmidt technique
time, is initiated. Then, using a linear approximation, time to
failure is estimated. The performance of the scheme has been MTS Mahalanobis-Taguchi system
validated via experiments performed on a mono-block centrifugal MTS Mahalanobis space
water pump testbed. The pump has been instrumented with
vibration, pressure, temperature, and flow sensors; and exper- OA orthogonal arrays
iments involving healthy and various types of faulty operating
conditions have been performed. The experiments show that the PIV particle image velocimetry
proposed approach renders satisfactory results for centrifugal
RANS Reynolds averaged Navier-Stokes
water pump fault detection, isolation, and prognostics. Overall,
the proposed solution provides a reliable multivariate analysis and RS relative sensitivity
real-time decision making tool that 1) fuses data from multiple
sensors into a single system level performance metric; 2) extends S/N signal to noise ratio
MTS by providing a single tool for fault detection, isolation, and
prognosis, eliminating the need to develop each separately; and TN true negatives
3) offers a systematic way to determine the key parameters, thus TP true positives
reducing analysis overhead. In addition, the MTS-based scheme
is process independent, and can easily be implemented on wireless TTF time to failure
motes1, and deployed for real-time monitoring, diagnostics, and
prognostics in a wide variety of industrial environments. WD Wigner distribution
Index Terms—Centrifugal pump, fault isolation and prognostics,
Mahalanobis distance-based fault detection, Mahalanobis Taguchi
system, real-time decision making.
NOTATION
th characteristic in the th observation
ACRONYMS
number of observations
aOFAT adaptive one-factor-at-a-time
standard deviation of the th characteristic
CFD computational fluid dynamics
normalized value of the th characteristic in the th
observation
Manuscript received April 11, 2010; revised January 07, 2011; accepted April standard deviation of the normalized values
04, 2011. Date of publication October 14, 2011; date of current version De-
cember 02, 2011. Associate Editor: L. Walls. correlation matrix
A. Soylemezoglu is with the Department of Engineering Management and
Systems Engineering, Missouri University of Science and Technology, Rolla, inverse of the correlation matrix
MO 65401 USA (e-mail: soylemez@mst.edu).
S. Jagannathan is with the Department of Computer and Electrical Engi- Mahalanobis Distance for the th observation
neering, Missouri University of Science and Technology, Rolla, MO 65409 USA
(e-mail: sarangap@mst.edu). number of characteristics
C. Saygin is with the Department of Mechanical Engineering, University of
Texas San Antonio, San Antonio, TX 78249 USA (e-mail: can.saygin@utsa. signal-to-noise ratio for the th run of the OA
edu).
Color versions of one or more of the figures in this paper are available online number of abnormalities under consideration
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TR.2011.2170255
E expected value
1A mote, also known as a sensor node, is a node in a wireless sensor network. cov covariance

0018-9529/$26.00 © 2011 IEEE

SOYLEMEZOGLU et al.: MTS AS A MULTISENSOR BASED DECISION MAKING PROGNOSTICS TOOL 865

Fig. 1. Cause-effect diagram for seal failure.

Fig. 2. Cause-effect diagram for impeller failure.

I. INTRODUCTION Based on the literature survey regarding centrifugal pumps

presented below, failures of the seal and the impeller are found

C ENTRIFUGAL pumps are extensively used in a wide va-

riety of process industries, such as chemical processing
plants, food service, electric utility, and automotive companies.
to be the predominant failure modes. Figs. 1 and 2 show the
cause-effect diagram for the seal, and impeller failures in
centrifugal pumps, respectively. In general, researchers have
Centrifugal pumps are often critical components in the entire approached the centrifugal pump performance monitoring and
production and process chain [1]. fault diagnosis problem in two ways: 1) utilizing qualitative
or quantitative model based techniques, and 2) using process
A. Pump Failures history based models. Table I summarizes some of the various
According to a report prepared under the Specific Actions for types of failure modes, and the sensors that can be utilized to
Vigorous Energy Efficiency program of the European Commis- identify these failures; whereas Table II shows the detection
sion, pumps annually consume 117 Terawatt-hours of electricity techniques.
in Europe alone [2]. Unexpected failure of centrifugal pumps In their paper, Byskov et al. [3] investigate the flow field in a
can become costly due to the interruption in processes, and fur- centrifugal pump impeller at design load and quarter-load using
ther costs are incurred in terms of repair and replacement. In large eddy simulation (LES). They compare velocities predicted
addition, the failure of a pump can also cause potential damage by the LES, and steady-state Reynolds averaged Navier-Stokes
in the equipment elsewhere in the process chain. In a centrifugal (RANS) simulations based on the Baldwin-Lomax, and Chien
pump, bearings, seals, and impellers are critical components that k-e turbulence models with experimental data obtained from
directly affect the desired performance of the pump. “Normal particle image velocimetry (PIV). They report that using LES
operation” of a pump is referred to a pump running without for analyzing the flow field in centrifugal pumps provides an
any defects or impending failures, while “abnormal operation” improved insight into the basic fluid dynamics with a satisfac-
refers to a pump running with defects that will eventually lead tory accuracy compared to experiments.
to failures.” Coutier-Delgosha et al. [4] study a pump with a two-dimen-
Bearing, seal, and impeller defects, as well as cavitation, sional curvature blade geometry in cavitating, and non-cav-
can result in problems such as abnormal noise, leakage, drop itating conditions using different experimental techniques,
in hydraulic performance, structural vibration, and damage on and a three-dimensional numerical model. They compare
the pump components by corrosion and pitting. Therefore, a experimental and numerical results under various flow condi-
variety of techniques that strive to maximize the availability tions. Their results show the ability of the developed model
of pumps have been developed and presented in the literature. to simulate complex three dimensional cavitation in rotating
866 IEEE TRANSACTIONS ON RELIABILITY, VOL. 60, NO. 4, DECEMBER 2011

TABLE I
SUMMARY OF THE VARIOUS FAILURE MODES AND SENSORS USED FOR DETECTION

TABLE II
TECHNIQUES USED FOR FAULT DETECTION

machinery, and the associated effects of pump performance, itation conditions. They use a quasi-three dimensional analysis
which can be utilized in future pump designs. method to model a 7-blade pump impeller across a wide range of
Hirschi et al. [5] propose a three-dimensional numerical flow coefficients and cavitation numbers. Their results show that
method which can predict the cavitation behavior of centrifugal performance trends associated with off-design flow and blade-
pumps. The proposed method, which allows performance cavitation, including breakdowns, compare qualitatively well
drop prediction, consists of assuming the cavity interface with experimental results.
is a free surface boundary of the computation domain, and Byington et al. [8] develop an inline intelligent pump mon-
computing the single phase flow. The unknown shape of the itor for critical hydraulic pumps and motors used in aircraft sys-
interface is determined using an iterative procedure matching tems. They acquire pressure, flow, and temperature data from
the cavity surface to a constant pressure boundary. Using the the pump system, and reduce the raw data by processing and
proposed method, the main cavitation characteristics of the extracting salient features using high and low bandwidth anal-
pump impeller can be predicted, and it compares well with ysis. Their overall approach includes performance modeling,
Navier-Stokes computations. signal processing and feature extraction, feature level fusion, au-
Langthjem and Olhoff develop a computationally simple, fast tomated classification using neuro-fuzzy classifiers, and knowl-
numerical model to calculate the flow induced noise in a two- edge fusion for estimating degradation through the collection
dimensional centrifugal pump [6]. They use a discrete vortex of in-line pump data. They demonstrate the success of their ap-
model to simulate the flow within the pump, and their results proach by running five different pumps to failure.
show that the proposed method is one of the simplest methods Dong et al. [9] investigate the effect of the modification of
of capturing the essential features of rotational flow. Medvitz et impeller geometry on unsteady flow, pressure fluctuations, and
al. [7] use a multi-phase computational fluid dynamics (CFD) noise in a centrifugal pump. They utilize PIV, pressure, and
method to analyse centrifugal pump flow under developing cav- noise measurements after performing various modifications to
SOYLEMEZOGLU et al.: MTS AS A MULTISENSOR BASED DECISION MAKING PROGNOSTICS TOOL 867

the impeller, and try to determine the overall effects on the pump based on a series of standard fault pattern pairings between fault
performance. Koo and Kim [10] develop a Wigner Distribution symptoms and faults. The fuzzy neural network model is trained
(WD) based vibration monitoring system to analyse vibration to memorize these symptom-fault pairs. Because fuzzy neural
signals collected from nuclear reactor coolant pumps. They use networks adopt bi-directional association, they make use of in-
a rotor kit to simulate abnormal operation conditions such as formation from both the fault symptoms and the faults, which
bearing rubbing, shaft bending, and shaft misalignment; and de- improve the recognition rates. They verify their results through
velop a neural network based online diagnostic method which experiments performed on water pump sets of an oil plant, and
analyses the vibration data by utilizing Fourier Transforms, and report a well distinguished ability to perform diagnosis of rotary
a special software developed with the WD. They achieve a clas- machines.
sification accuracy of 81.25% using their proposed method. The As it can be seen from the literature survey, most authors
study shows that the WD makes it easy to analyse the vibration report success in condition monitoring and fault diagnosis
signals, and it helps operators better grasp the cause of vibra- of pumps using various quantitative model-based, qualitative
tions in the pump. model-based, and process history based methods. However,
Lebold et al. [11] develop a non-intrusive torsional vibration to the best of our knowledge, a method which can also suc-
method to monitor and track small changes in crack growth in cessfully predict the remaining useful life of a pump or its
the shaft of a reactor coolant pump. They grow fatigue cracks in components in addition to fault diagnosis is not presented.
a laboratory scale shaft under a controlled process. The shaft is In this paper, a Mahalanobis-Taguchi System (MTS) based
equipped with a photosensitive optical encoder, and the encoder prognostics tool for centrifugal pumps is developed.
data are processed with a customized torsional vibration algo-
rithm to produce a torsional vibration spectrum. They show that B. MTS-Based Methods
the torsional vibration measurement method has the ability to de- MTS has been widely used in various diagnostic applications
tect and track natural frequency shifts, which potentially allows that deal with data classification. In MTS, Mahalanobis Dis-
online diagnostics and prevention of shaft failure due to cracks. tance (MD) is used to identify the degree of abnormality of data,
Perovic et al. [12] develop a fuzzy logic system to con- whereas the Taguchi methods, which utilize orthogonal arrays
trol cavitation, blockage, and impeller damage in centrifugal (OA), and signal-to-noise ratio (S/N), are used to evaluate the
pumps. They establish fault signatures from the pump motor accuracy of predictions, and to optimize the system [16]. MD,
current data by relating spectral features to individual faults introduced by P.C. Mahalanobis in 1936 [17], is a multivariate
using fuzzy logic inference. They then build a fuzzy logic generalized measure used to determine the distance of a data
system for the final diagnostic decision making. They test their point to the mean of a group. MD is measured in terms of the
method by performing experiments on two centrifugal pumps, standard deviations from the mean of the samples, and provides
and state that the analysis of fault signatures using fuzzy logic a statistical measure of how well the unknown data set matches
inference achieves good success in pump diagnostics. with the ideal one. The advantage of the MD is that it is sensitive
Sakhtivel et al. [13] perform vibration-based fault diagnosis to the intervariable changes in the reference data. Therefore, it
of a mono-block centrifugal pump. They study normal opera- has traditionally been used to classify observations into different
tion of the pump against various faults, including bearing fault, groups and diagnoses [18].
impeller fault, seal fault, and cavitation. They measure vibra- The benefits of MTS as a pattern recognition and data classi-
tions at the pump inlet using a piezo-electric accelerometer, and fication tool can be summarized as follows.
use clear acrylic pipes at the inlet and outlet of the pump to • It is a robust methodology, insensitive to variations in mul-
see the cavitation. They extract standard statistical features such tidimensional systems.
as standard deviation, kurtosis, skewness, and sample variance • It can handle many different types of data sets and effec-
from the vibration data, and use a C4.5 decision tree algorithm tively consolidate the data into a useful metric.
to classify the features into the operating conditions mentioned • Implementation of MTS requires limited knowledge of sta-
above. They report good experimental results which are appli- tistics.
cable to practical applications of fault diagnosis of mono-block • It relies typically on simple arithmetic, contextual knowl-
centrifugal pumps. edge, and intuition.
In a later paper, Sakhtivel et al. [14] present the use of • Its success has been demonstrated in various practical ap-
decision trees and rough sets to generate classification rules plications.
from statistical features extracted from vibration signals under In their paper, Chinnam et al. [19] use ten features derived
normal operation and faulty operating conditions of centrifugal from two degradation signals (thrust force, and torque) in a
pumps. They build a fuzzy classifier using decision trees and drilling operation to determine the condition of the drilling tool.
rough set rules using experimental data. The performance of The authors utilize MD to generate a measurement scale, and
the decision tree and the rough set are compared. They also then they use the Taguchi methods to reduce the number of fea-
evaluate the classification accuracy of a principal component tures down to six. They use the 99th percentile of the MDs from
analysis based decision tree-fuzzy system. The results show the normal group as a threshold, which shows a superior perfor-
that the performance of the decision tree-fuzzy hybrid system mance when compared with the any of the individual features.
performs better than the rough set-fuzzy hybrid system. Wang et al. [20] used the iris data, which is often used in sta-
Zhang et al. [15] propose a fuzzy neural network model for tistical textbooks, and credit card data from a Taiwanese bank,
fault diagnosis of rotary machines. The fault diagnosis system is to display the effectiveness of MTS in data classification. They
868 IEEE TRANSACTIONS ON RELIABILITY, VOL. 60, NO. 4, DECEMBER 2011

compare the performance of MTS with discriminant and step- ratio, angle-over-torque ratio, total angle turned, and work done.
wise discriminant analysis, and conclude that MTS shows better Then, the signature is processed using the MTS-based approach
performance. Cudney et al. [21] apply MTS to study the rela- to detect the grip-length, all of which occurs in real-time as
tionship between available vehicle level performance data for the fastening process is completed. The proposed approach is
vehicle ride, and the corresponding consumer satisfaction rat- also applied to detect the presence of re-used fasteners, which
ings for the purpose of improving customer-driven quality. They is another quality concern because some material properties and
analyse six vehicle ride parameters from sixty-seven datasets physical conditions of bolts and nuts can change if they are
using the MTS, determining the key parameters. The authors reused several times.
are able to show that the MTS methodology can be used to eval- Dasgupta [28] develop a unified framework for achieving
uate consumer satisfaction, and manufacturers can utilize this process control and improvement during the implementation
knowledge to make the appropriate design decisions regarding phase of Six Sigma. The proposed framework takes advantage
attributes that are delivered in a product. of the adaptability of MTS as a classification, variable selection,
Cudney et al. [22] further investigate the performance and monitoring tool. The framework is explained using a simu-
data and customer satisfaction relationship using the Ma- lated example. Srinivasaraghavan and Allada [29] also propose
halanobis-Taguchi Gram-Schmidt (MTGS) technique, and an MTGS based methodology to support the implementation of
showed improved results. Foster et al. [23] develop an alterna- lean processes. Their methodology assists contemporary lean
tive search procedure to be used within the MTS. The adaptive assessment tools through providing a measure of leanness by
One-Factor-AT-a-Time (aOFAT) procedure replaces the or- benchmarking exemplary lean industries along with sugges-
thogonal arrays traditionally utilized in MTS, and features are tions for improvements based on cost considerations.
individually added or removed from the classification system MTS is a data driven method that relies on multivariate
depending on their impact on the overall signal-to-noise ratio. statistical analysis. Therefore, it is important to compare MTS
The experimentation shows that the aOFAT procedure renders to well-established multivariate statistical techniques such as
greater improvements over the median with the same or fewer principal component analysis, discriminant analysis, cluster
design alternatives being explored, and also exhibits a good analysis, canonical component analysis, and multidimensional
ability to generalize to new instances after training. scaling, as well as artificial neural networks. For an in-depth
MTS has also been used as a process control and improve- comparison of these techniques to MTS, please refer to Kim et
ment tool to aid manufacturing systems. Asada [24] use MTS al. [30], Cudney et al. [31], Jugulum [32], Woodall et al. [33],
to forecast the yield of a wafer production system, which is af- and Su and Hsiao [34].
fected by variability of the electrical properties, and environ- MTS was developed by Taguchi as a diagnostic technique
mental dust. MDs are used to compare the relationship between using multivariate data. The MD introduced by Mahalanobis
wafer yield and distance. Hayashi et al. [25] develop an MD in 1936 takes the correlation structure of a system into account
based method to maximize productivity in a semi-conductor [34]. Therefore, MTS-based clustering (i.e., classification) is not
manufacturing facility. Their method consists of real-time mon- a new technique. However, the main contribution of this study
itoring of the ratio of work-in-process inventory to number of is the combination of MTS-based clustering with prognostics
in-process wafers, which they call flow factor , using MD values and its application to pump failures, which has not been studied
to identify the out-of-specification processes and tools, and take in the literature. By combining these methods, the benefits of
the necessary actions. MTS have been expanded over to the prognostics field. More
Mohan et al. [26] propose an MTS-based diagnostic and root specifically, this paper 1) uses MTS to fuse multi-sensor infor-
cause analysis scheme for online monitoring of grip-length of mation into a single system performance metric; 2) builds upon
pull type fasteners. Their scheme utilizes data collected from an MTS by introducing the use of MD-based fault clusters for fault
aerospace industry pull-type fastening tool instrumented with isolation; 3) utilizes the progression of MD values over time to
various sensors. Sensor data are analysed using MTS, and the estimate the remaining useful life of components (i.e., time to
quality of the current operation is communicated back to the failure); and 4) applies this proof-of-concept model to the do-
user via a wireless network. The authors were able to obtain main of centrifugal pump monitoring, diagnostics, and prognos-
100% data collection on each fastener, as opposed to other statis- tics area.
tical process control techniques which rely on sampling. In [27] The MTS approach is a diagnostic and forecasting technique
is presented an MTS-based methodology which detects grip for multivariate data. MTS establishes a classifier by con-
length of bolted joints in real-time during fastening. Grip length structing a continuous measurement scale rather than directly
is the length of the unthreaded portion of a bolt shaft. When the learning from the training set. To verify the robustness of MTS,
total thickness of joining members is greater than the grip length Su and Hsiao [34] compare MTS with several popular classi-
of the bolt, it is called under-grip, which compromises the struc- fication techniques, including stepwise discriminate analysis,
tural integrity of a joint. A pneumatic, hand-held, rotary-type decision tree analysis, back-propagation neural networks, and
tool for bolted joints is integrated with a torque sensor and an support vector machines. The results indicate that MTS is the
optical encoder to obtain torque-angle signatures. The proposed most robust technique to deal with the classification problem.
approach reads in various characteristics from the torque-angle
process signature to infer about the quality of a bolted joint. The C. Robustness of MTS-Based Methods
process signature data are used to calculate the mean and stan- To evaluate classifiers, two metrics are primarily used [34].
dard deviation of various factors, including torque-over-angle The sensitivity metric is defined as the accuracy on the posi-
SOYLEMEZOGLU et al.: MTS AS A MULTISENSOR BASED DECISION MAKING PROGNOSTICS TOOL 869

TABLE III
ROBUSTNESS OF MTS

tive examples (true positives/(true positives false negatives)), a MD-based fault clustering method to classify faults into
whereas specificity is the accuracy on the negative examples different categories. Using these clusters, fault detection and
(true negatives/(true negatives false positives)). Sensitivity root cause analysis is performed. The scheme also utilizes the
is the accuracy rate on the nonconforming items. Specificity progression of the MD values over time to facilitate prognosis
is the accuracy on conforming items. In the Sensitivity equa- of time to failure for components. The details of the scheme are
tion, true positives (TP) represents nonconforming items diag- presented in the rest of the section.
nosed as nonconforming items, and false negatives (FN) rep-
resents nonconforming items diagnosed as conforming items. A. MTS
In the specificity equation, true negatives (TN) represents con- MTS starts with data collection on normal operational ob-
forming items diagnosed as conforming items, and false pos- servations. Then, MD is calculated using certain characteristics
itives (FP) represents conforming items diagnosed as noncon- to investigate whether MD has the ability to differentiate the
forming items. normal operation group from an abnormal operation group. If
A third metric is called the relative sensitivity (RS), which is MD cannot detect the normal operation group using those par-
sensitivity divided by specificity. The RS can be used to judge ticular characteristics, then a new combination of characteristics
whether or not a classifier has the balanced ability to predict the need to be explored. When the right set of characteristics are
positive (nonconforming) items, and the negative (conforming) found, Taguchi methods are employed to evaluate the contribu-
items. The target RS value is one, which means the classifi- tion of each characteristic. If possible, dimensionality is reduced
cation method has balanced compatibility between classifying by eliminating those characteristics that do not add value to the
nonconforming observations and conforming items. If the RS analysis. MTS consists of four stages [16], [18], as follows.
is much lower or higher than 1, it indicates that the classifier Stage 1: Construction of Mahalanobis Space: The first step
has some bias on classifying [34], [35]. The accuracy rate is an- in MTS is the construction of the Mahalanobis Space (MS)
other metric showing the overall capability of the proposed MTS as the reference. To construct MS, the variables that represent
method; the higher the accuracy, the better the classification ca- “normal” operation conditions are defined and sampled. In this
pability [35]. study, new bearings that are properly lubricated represent the
In this section, a comparative study is presented to demon- normal operation case. The steps for the construction of MS are
strate the robustness of the MTS approach based on [26], [27]. In outlined below.
[26], an MTS-based diagnostic and root cause analysis scheme 1) Calculate the mean for each characteristic in the normal
for monitoring grip length of pull-type fasteners is presented. operation data set as
This approach is implemented on a fastening tool integrated
with a strain gage, a linear variable differential transformer, a
pressure sensor, and a mote for wireless communication. In [27], (1)
another MTS-based methodology is presented; it detects the
grip length of bolted joints in real-time during fastening. Grip 2) Then, calculate the standard deviation for each character-
length is the length of the unthreaded portion of a bolt shaft. istic
When the total thickness of joining members is greater than the
grip length of the bolt, it is called under-grip, which is a non-
conformity compromising the structural integrity of a joint. A
(2)
pneumatic, hand-held, rotary-type tool for bolted joints is in-
tegrated with a torque sensor and an optical encoder to obtain
torque-angle signatures. The experiments show that the MTS 3) Next, normalize each characteristic, form the normalized
approach is effective in determining the nonconformities. data matrix , and take its transpose
The experimental data from each study [26], [27] have been
used to calculate the robustness metrics, as shown in Table III.
The results agree with the findings of Su and Hsiao [34] that (3)
MTS is extremely robust to deal with such diagnostics and prog-
nostics problems. 4) Then, verify that the mean of the normalized data is zero.

II. METHODOLOGY
In this paper, a MTS-based fault detection, isolation, and
prognostics scheme is presented. The proposed scheme utilizes (4)
870 IEEE TRANSACTIONS ON RELIABILITY, VOL. 60, NO. 4, DECEMBER 2011

Fig. 3. MD-based fault clustering. Fig. 4. Illustration of angle based fault isolation.

5) Verify that the standard deviation of the normalized data is

one.

(5)

6) Form the correlation matrix for normalized data. Cal-

culate the matrix elements as

(6)
Fig. 5. Linearly aligned fault clusters.
7) Calculate the inverse of the correlation matrix .
8) Finally, calculate MD as
can take. For the abnormal operation cases, MD values are cal-
(7) culated using the combination of the characteristics dictated by
the OA, and the larger-the-better signal-to-noise ratio is calcu-
Stage 2: Validation of MS: To validate MS, observations that lated as follows [16], [18].
fall outside the normal operation group are identified, and the
corresponding MD values are calculated. The characteristics of
(8)
the abnormal operation group are normalized using the mean
and the standard deviation of the corresponding characteristics
in the normal operation group. The correlation matrix corre- Then, an average signal-to-noise ratio at level-1 and level-2
sponding to the normal operation group is used to compute the of each characteristic is obtained, and the gain in signal-to-noise
MDs of the abnormal operation cases. If MS has been con- ratio values for each characteristic is calculated as [16], [18]
structed using the appropriate characteristics, then the MDs of
the abnormal operation group should have higher values than
that of the normal operation group. (9)
Stage 3: Identification of the Useful Characteristics: In this The characteristics with positive gain are selected for the de-
stage, Orthogonal Arrays (OA), and Signal-to-Noise (SN) ratios tection of abnormalities, and the rest are discarded.
are used to select the critical variables. OA is used to estimate Stage 4: Decision Making: In the final stage, the application
the effects of features, and the effect of interactions, by mini- under investigation is monitored via data collection using the
mizing the number of experiments. In MTS, each variable is as- MS. MDs are calculated; and if MD 1, it is concluded that
signed to one column, and is set with two levels; that is, level 1 is the application is displaying abnormal operation behavior, and
corresponding to the presence of a variable, and level 2 is corre- appropriate corrective actions need to be taken. If MD 1, then
sponding to the absence of a variable. In other words, the second the conditions are normal operation.
phase is Taguchi’s robust engineering, which is applied to de-
termine the important variables, and then optimize the system. B. MD-Based Diagnostics
An orthogonal array is a table that lists the combination of A fault is a change from the normal operation of a system.
characteristics, which enables testing the effects of the pres- Fault detection is the first step in diagnostics, which indicates
ence or absence of a characteristic. The size of the OA is de- the occurrence of a fault in a monitored system. Diagnostics,
termined by the number of characteristics, and the levels they which is also referred to as fault isolation, aims to identify the
SOYLEMEZOGLU et al.: MTS AS A MULTISENSOR BASED DECISION MAKING PROGNOSTICS TOOL 871

Fig. 6. Centrifugal water pump testbed.

Fig. 7. Erosion damage on the impeller surface.

root cause of a fault [36]. In this study, an MD-based data clus-
tering technique is used to classify centrifugal pump failure data
into different fault groups. The technique takes advantage of the
fact that MD is calculated using the inverse of the correlation
matrix (6). The correlation matrix consists of the correlation co-
efficients, which is a normalized measure of the strength of the
linear relationship between variables as

Fig. 8. OA analysis gains for filter clog failure.

(10)
C. MD-Based Prognostics
The correlation between the MDs can be visualized by plot- Prognostics involves a) detection of an abnormal operation
ting the current MD value against the previous MD value. In condition or fault, b) identifying the root cause or fault isolation,
this case, signatures obtained from faulty operating regimes are and c) estimating the remaining useful life or time to failure. In
grouped into individual clusters as shown in Fig. 3. the prognostics case, the MD values are calculated using a pre-
After clustering faults based on their MD ranges, a threshold determined time window as the process continues. Fault detec-
for each fault cluster is determined. In this application, a tion occurs once the MDs leave the normal operation regime,
threshold includes a lower and a higher MD value, which iden- and the tracking of the MD trend is initiated. By observing the
tifies a specific fault in that range. Among various techniques, direction of the transition of the MD trend between the normal
such as probabilistic thresholding method [34] and bisection operation case and one of the fault clusters, the possible root
method [35], a threshold can be simply calculated as a range at cause is identified. This comparison is accomplished by calcu-
three times the standard deviation above and below the average lating the angle between the unknown point U and the mean
MD value of a fault cluster. Consequently, the type of the fault MD value of each of the known fault clusters, assuming the fault
is isolated by comparing the average MD value of the process progression will follow a roughly linear trajectory (Fig. 4). In
with the thresholds set for various faulty conditions. The pro- this case, the smaller the angle, the more likely that the unknown
cedures for determining a threshold is an important aspect in point is progressing towards one of the fault clusters.
carrying out a diagnosis and prognosis process effectively. A Once the MD data enters the root cause failure cluster, prog-
useful threshold has a significant impact on the accuracy of nosis of the time to failure is initiated via linear approximation.
the decision making process; incorrect thresholds may cause First, the slope of the progression of the MD value is generated
numerous false alarms, leading to excessive downtime. An using the MD value of the current, and the previous time win-
effective threshold can enhance the diagnostic and forecasting dows. Then, using the slope, and the current MD value, the time
ability of MTS [16], [18]. Therefore, it is necessary to deter- to failure is determined at each instant of time (11). The failure
mine the value of the threshold in such a way that the total loss threshold is selected by the designer when the performance of
is minimized. the machine is unsatisfactory.
Because the main objective of this study is to introduce a
proof of concept MTS model to demonstrate its prognostics (11)
ability for detecting pump failures, and determining the time to
failure, this paper uses the mean and three times the stan- In some cases, the fault clusters may lie on a linear path. The
dard deviation approach to set thresholds, which is explained in linear alignment of the fault clusters would make the classifi-
Section IV-C. cation of the fault progression difficult. For example, consider
872 IEEE TRANSACTIONS ON RELIABILITY, VOL. 60, NO. 4, DECEMBER 2011

Fig. 9. OA analysis gains for impeller failure.

Fig. 10. OA analysis gains for seal failure. Fig. 12. MD based fault clusters using only lateral acceleration.

The water pump testbed was operated for 150 hours to collect
the signature for normal operating conditions. After the initial
data collection was complete, three types of pump failures, one
at a time, were seeded to accelerate the testing. In one of the
tests, the seal was mechanically defaced to create an artificial
seal failure. The pump was operated with the damaged seal, and
sensor signatures were collected. During the experiment, a drop
in the inlet pressure was noticed.
In another experiment, impeller failure was investigated. The
failure was seeded by intermittent dry running of the pump,
which caused erosion on the impeller surface. Because the ex-
periments were taking a long time, the blades on the impeller
were mechanically filed down to further accelerate the testing.
The experiment was carried out until a large change from the
normal operation sensor signature was noticed. As the damage
Fig. 11. MD based fault clusters using only the outlet pressure. on the impeller progressed, a corresponding drop in flow values
was observed. The accelerometer signatures were also observed
to deviate from the normal operation case. Fig. 7 shows the
the fault clusters in Fig. 5, and assume that the MD value of the damage on the impeller surface at the end of the experiment.
current time window is being marked in the plot with an “x”. In the final experiment, the filter clog failure was studied. The
As the process continues, the MD values leave the normal oper- pump was operated while contaminants such as sand were added
ation condition, following the direction of the arrow; but it is not to the pump’s water reservoir. Because the filter is located right
clear whether the actual fault is progressing towards fault-1 or before the main body of the pump, a clog in the filter resulted in
fault-2. In this case, a new set of MD values is calculated using a drop in the inlet and outlet pressures, and in the outlet flow.
one of the faulty data sets as a basis. Consequently, the MS is
shifted, and the identification of the MD progression becomes IV. EXPERIMENTAL RESULTS
straightforward.
A. Identification of Key Parameters
III. DESIGN OF EXPERIMENTS The construction of the MS requires the selection of an ini-
To facilitate the development of the proposed multi-sensor tial set of parameters [18]. In some cases, raw sensor data may
based MTS prognostics tool, a centrifugal water pump testbed provide sufficient information to differentiate between various
has been constructed at the Missouri University of Science and failure modes. When the raw sensor data are not sufficient, it
Technology. The 1/2 HP centrifugal pump has been instru- is customary to search for higher level features that can be ex-
mented with sensors tracted by further processing sensor data e.g., the bearing char-
• at the inlet, with pressure, and temperature; and acteristic frequencies obtained from vibration data for rolling
• at the outlet, with pressure, temperature, and flow. element bearings. In this study, the raw sensor measurements
In addition to these sensors, two accelerometers were in- are used as the input parameters for the MTS because they pro-
stalled in the pump casing to measure the lateral, and vertical vide adequate information. The input parameters are
vibration. Fig. 6 shows a schematic representation of the testbed • lateral acceleration,
setup. • vertical acceleration,
SOYLEMEZOGLU et al.: MTS AS A MULTISENSOR BASED DECISION MAKING PROGNOSTICS TOOL 873

Fig. 13. Magnified view of Fig. 12.

TABLE IV
MEAN AND STANDARD DEVIATION OF MDS Fig. 15. Magnified view of MD based fault clusters.

B. Analysis on the Number and Types of Sensors Required

In an ideal case, one would like to use the fewest possible

sensors to perform diagnostics and prognostics on a system. The
implementation of multiple sensors may become costly, or may
altogether be unfeasible due to the mechanical limitations of the
system. To determine the number and type of sensors required
for identifying different centrifugal pump failures, a series of
analyses using only the data from one sensor at a time were
performed for the filter clog, seal, and impeller failure cases.
1) Pressure Sensor Installed on the Pump Outlet: The outlet
pressure of a centrifugal pump is one of the main performance
characteristics. Therefore, it would be ideal to be able to deter-
mine the health state of the entire pump components by mea-
suring only the outlet pressure. To determine the adequacy of
using only a pressure sensor on the outlet of the pump, MD fault
clusters are formed as mentioned in Section II-B. Fig. 11 shows
the fault clusters.
As it can be seen from Fig. 11, the normal operation, and the
impeller failure cases can clearly be identified using the outlet
pressure. However, the seal failure, and the filter clog failure
fault clusters overlap with each other.
2) Accelerometer Installed on the Pump Casing in the Lateral
Axis: Another good indicator of the health state of a centrifugal
Fig. 14. MD based fault clusters using all sensors. pump is the amount of vibration that it generates. In a normal op-
eration case, the pump should operate within a specified range of
vibrations; therefore, the amount of vibration experienced on the
• inlet pressure, pump could be a good indicator of potential problems. Fig. 12
• outlet pressure, and depicts the MD based fault clusters using the data obtained from
• outlet flow. the lateral accelerometer on the pump casing. Fig. 13 shows a
The orthogonal array is utilized for the S/N analysis. zoomed-in plot of failure clusters.
Using the larger-the-better type signal to noise ratios, and the As it can be seen from Figs. 12 and 13, the impeller failure
overall gains, all of the parameters mentioned above are selected can clearly be identified using only the lateral acceleration in-
as the final input parameters for further analysis because none formation. However, the normal operation behavior, seal failure,
of the inputs could be eliminated. Figs. 8–10 show the results and filter clog failure conditions completely overlap with each
for the OA analysis for the experiments. other, and cannot be separated.
874 IEEE TRANSACTIONS ON RELIABILITY, VOL. 60, NO. 4, DECEMBER 2011

Fig. 16. Fault detection and isolation scheme.

TABLE V each fault type, the thresholds for each fault cluster are identi-
FAULT DETECTION AND ISOLATION RESULTS fied. Fig. 14 shows the fault clusters that are used in the final
analysis. Notice that there are no transition points between the
fault clusters in Fig. 14. This result is because the experiments
were performed in an accelerated manner. Fig. 15 shows a mag-
nified view of the normal operation, filter clog, and the impeller
failures.
Fig. 16 displays the fault detection and isolation scheme for
the centrifugal pump. The detection starts with the loading and
acquisition of test data. Then, the mean of the MD values for
the current time window is calculated. If the mean MD is larger
than that of the threshold set for the normal operation case, a
fault is detected. Next, the mean MD value is compared with
the thresholds of the other fault cases. If the mean MD value
stays within the thresholds of a cluster for a user determined
time window, then the fault type is isolated.
Table V shows the fault detection and isolation results that
were obtained using different combinations of thresholds for the
C. Fault Detection and Isolation Using Multiple Sensors fault clusters. The results show a high success rate in correct
detection and isolation of centrifugal pump faults. In the case
The analyses in Sections 3.B.1 and 3.B.2 show that using a where the filter clog failure threshold is selected as 570, and the
single type of sensor is not sufficient for successful fault isola- seal failure threshold is selected as 562, the proposed scheme
tion on a centrifugal pump. Therefore, multiple types of sensors misses the detection. This result is due to the slight overlapping
are required. Using all of the sensors mentioned in Section 3.A, between the fault clusters, as can be seen in Fig. 15.
MD values for all of the experiments are calculated. Table IV Collecting operational data from the pump testbed used for
shows the observed minimum, and maximum values of MDs, this study has been an extremely time-consuming process. The
as well as the standard deviations for each fault type. Using the data in this paper were obtained over a period of 6 months via
mean and three times the standard deviation of the MDs for accelerated degradation testing. It was not possible to obtain
SOYLEMEZOGLU et al.: MTS AS A MULTISENSOR BASED DECISION MAKING PROGNOSTICS TOOL 875

using the filter clog case as a basis for comparison. This result re-
aligns the fault clusters as seen in Fig. 17, where the normal op-
erating condition cluster now sits between the filter clog failure,
and the seal failure clusters. The realignment improves the prob-
ability of correctly identifying the root cause of a fault because
the progression of a filter clog fault and a seal fault would now
proceed in opposite directions. By selecting the seal failure and
the impeller failure cases as the normal operation case in MD
calculations, two other alignments of the fault clusters are ob-
tained, which covers all of the possible fault progression sce-
narios.
Fig. 18 shows the MD-value based prognostic scheme. The
proposed scheme exploits the idea of using each of the fault
cases as a reference in the generation of fault clusters. At
, test data are acquired, and the MD values are calculated using
the normal operation case as the basis for comparison. Fault
detection occurs when the MD of the current process crosses
the predetermined threshold of the normal operation case. If a
fault is detected, a flag is set to initiate the parallel calculation of
MD values with the faulty cases as the references, and the time
is advanced to the next time window. Then, by using the four
references, the direction of progression of the fault to its fault
cluster is determined, and the root cause identified.
Fig. 19 shows the time evolution of MD values for the seal
failure experiment. Fig. 20 shows the time-to-failure (TTF) es-
timation plot for the same experiment using a thirty minute
time window using (11). The failure threshold value selected
for the experiment is 827. The estimation is triggered as soon as
a particular operation enters a fault cluster, and stays within the
cluster for a period of two time windows to minimize false fault
classification alarms.
Fig. 21 shows the time evolution of MD values for the filter
Fig. 17. Realigned fault clusters using filter clog as the normal operation case. clog failure experiment, while Fig. 22 shows the time to failure
estimation for the same experiment using a 30 minute time
window, and a failure threshold value of 582.
different data sets by enforcing failures on the pump. How-
ever, to demonstrate the robustness of MTS-based clustering
and prognostics, fifteen sets of data collected on bearing faults V. CONCLUSION
are used [37]. The data set includes cage faults, inner race de-
fects, outer race defects, and normal operation conditions (i.e., In this paper, a novel multi-sensor based MTS fault detection,
bearings without defects). This experiment [37] shows that the isolation, and prognostics scheme is presented. The proposed
MTS-based approach deployed on the bearing was effective in scheme utilizes MD based fault clustering, and the progression
forecasting the defects. of MD values for diagnostics and prognostics. The performance
of the scheme was validated through experiments performed on
D. Prognostics a centrifugal water pump testbed.
Advantages of the proposed approach can be summarized as
As described in Section II-B, determining the direction of ac- follows.
tual progression of a fault is difficult if the fault clusters are • The proposed approach fuses all of the pertinent informa-
aligned linearly (Figs. 14 and 15). In general, transition data are tion obtained from a multi-sensor environment into a single
utilized to obtain the direction of fault progression, and to iden- performance metric, which is the MD.
tify the root cause by finding the fault cluster. Unfortunately, • The proposed approach can function equally well in both
for the experiments performed in this study, due to the nature of the data and feature domains. In the case when the raw data
accelerated fault progression, the transition data are not avail- collected from sensors is adequate to differentiate between
able. If transition data were available, it could have displayed a various system health states, the data can directly be used
gradual progression from the healthy state to the faulty state in as inputs to the system. If the raw data are not sufficient,
a nonlinear fashion, which could be utilized for prognostics. then higher level features can be extracted from the data
To determine the actual direction of the progression of a fault, by further analysis, and these features can be utilized as
and to identify the root cause, a new set of MDs are calculated inputs.
876 IEEE TRANSACTIONS ON RELIABILITY, VOL. 60, NO. 4, DECEMBER 2011

Fig. 18. MD value based prognostic scheme.

Fig. 19. Time evolution of MD values for the seal failure experiment.
Fig. 21. Time evolution of MD values for the filter clog failure experiment.

Fig. 20. Time to failure estimation for the seal failure experiment.
Fig. 22. Time to failure estimation for the filter clog failure experiment.

• The process provides a unique solution for centrifugal

pump fault detection, isolation, and prognostics, elimi- • The process is application independent. It can be applied
nating the need for developing a tool for each separately. to a wide variety of multi-variate problems provided that
SOYLEMEZOGLU et al.: MTS AS A MULTISENSOR BASED DECISION MAKING PROGNOSTICS TOOL 877

a priori normal operation and abnormal operation data are [18] G. Taguchi and R. Jugulum, The Mahalanobis-Taguchi Strategy - A
available. Pattern Technology System. New York: John Wiley & Sons, 2002.
[19] R. B. Chinnam, B. Rai, and N. Singh, “Tool-condition monitoring from
• The process provides a methodical way to identify the fea- degradation signals using Mahalanobis Taguchi System,” in Robust En-
tures critical for the process, reducing the dimensionality gineering, ASI’s 20th Annual Symposium, 2004, presented at the.
of the problem. Therefore, the analysis overhead is reduced [20] H. Wang, C. Chiu, and C. Su, “Data classification using Mahalanobis
Taguchi System,” Journal of Chinese Institute of Industrial Engineers,
because efforts can be concentrated on the key features. vol. 21, pp. 606–618, 2004.
• Because the proposed scheme is not computationally ex- [21] E. A. Cudney, K. Paryani, and K. M. Ragsdell, “Applying the Maha-
haustive, it can easily be implemented onto wireless motes, lanobis-Taguchi System to vehicle ride,” Journal of Industrial and Sys-
tems Engineering, vol. 1, pp. 251–259, 2007.
and deployed to perform online monitoring (i.e., real-time) [22] E. A. Cudney, R. Jugulum, and K. Paryani, “Forecasting consumer sat-
and decision making in a variety of industrial environ- isfaction for vehicle ride using a multivariate measurement system,”
ments. International Journal of Industrial and Systems Engineering, vol. 4,
pp. 683–696, 2009.
The analysis and results presented in this paper, especially [23] C. R. Foster, R. Jugulum, and D. D. Frey, “Evaluating an adaptive One-
the thresholds, are valid for the particular experimental setup Factor-At-a-Time search procedure within the Mahalanobis-Taguchi
and operating conditions. However, the proposed approach can System,” International Journal of Industrial and Systems Engineering,
vol. 4, pp. 600–614, 2009.
be applied to other processes. [24] M. Asada, “Wafer yield prediction by the Mahalanobis-Taguchi
system,” in Statistical Methodology, IEEE International Workshop on,
REFERENCES 2001, pp. 25–28.
[1] E. Lihovd, T. I. Johannessen, C. Steinebach, and M. Rasmussen, “Intel- [25] S. Hayashi, Y. Tanaka, and E. Kodama, “A new manufacturing control
ligent diagnosis and maintenance management,” Journal of Intelligent system using Mahalanobis distance for maximizing productivity,”
Manufacturing, vol. 9, pp. 523–537, 1998. Semiconductor Manufacturing, IEEE Trans., vol. 15, pp. 442–446,
[2] H. Falkner and D. Reeves, Study on improving the energy efficiency of 2002.
pump ETSU, AEAT PLC. SAVE Study, U.K., 2001. [26] D. Mohan, C. Saygin, and J. Sarangapani, “Real-time detection of grip
[3] R. K. Byskov, C. B. Jacobsen, and N. Pedersen, “Flow in a centrifugal length deviation during pull-type fastening: a Mahalanobis–Taguchi
pump impeller at design and off-design conditions—Part II: Large eddy System (MTS)-based approach,” The International Journal of Ad-
simulations,” Journal of Fluids Engineering, vol. 125, pp. 73–83, 2003. vanced Manufacturing Technology, vol. 39, pp. 995–1008, 2008.
[4] O. Coutier-Delgosha, R. Fortes-Patella, J. L. Reboud, M. Hofmann, and [27] C. Saygin, D. Mohan, and J. Sarangapani, “Real-time detection of
B. Stoffel, “Experimental and numerical studies in a centrifugal pump grip length during gastening of bolted joints: a Mahalanobis-Taguchi
with two-dimensional curved blades in cavitating condition,” Journal System (MTS)-based approach,” Journal of Intelligent Manufacturing,
of Fluids Engineering, vol. 125, pp. 970–978, 2003. vol. 21, pp. 377–392, 2010.
[5] R. Hirschi, P. Dupont, F. Avellan, J. N. Favre, J. F. Guelich, and E. [28] T. Dasgupta, “Integrating the improvement and the control phase of Six
Parkinson, “Centrifugal pump performance drop due to leading edge Sigma for categorical responses through application of Mahalanobis-
cavitation: Numerical predictions compared with model tests,” Journal Taguchi System MTS,” International Journal of Industrial and Systems
of Fluids Engineering, vol. 120, pp. 705–711, 1998. Engineering, vol. 4, pp. 615–630, 2009.
[6] M. A. Langthjem and N. Olhoff, “A numerical study of flow-induced [29] J. Srinivasaraghavan and V. Allada, “Application of mahalanobis dis-
noise in a two-dimensional centrifugal pump. Part I. Hydrodynamics,” tance as a lean assessment metric,” The International Journal of Ad-
Journal of Fluids and Structures, vol. 19, pp. 349–368, 2004. vanced Manufacturing Technology, vol. 29, pp. 1159–1168, 2006.
[7] R. B. Medvitz, R. F. Kunz, D. A. Boger, J. W. Lindau, A. M. Yocum, [30] S. B. Kim, K.-L. Tsui, T. Sukchotrat, and V. C. P. Chen, “A com-
and L. L. Pauley, “Performance analysis of cavitating flow in cen- parison study and discussion of the Mahalanobis-Taguchi System,” In-
trifugal pumps using multiphase CFD,” Journal of Fluids Engineering, ternational Journal of Industrial and Systems Engineering, vol. 4, pp.
vol. 124, pp. 377–383, 2002. 631–644, 2009.
[8] C. S. Byington, M. Watson, D. Edwards, and B. Dunkin, “In-line health [31] E. A. Cudney, D. Drain, K. Paryani, and S. Naresh, “A comparison of
monitoring system for hydraulic pumps and motors,” in Aerospace the Mahalanobis-Taguchi System to a standard statistical method for
Conference, 2003, pp. 3279–3287. defect detection,” Journal of Industrial and Systems Engineering, vol.
[9] R. Dong, S. Chu, and J. Katz, “Effect of modification to tongue and im- 2, pp. 250–258, 2009.
peller geometry on unsteady flow, pressure fluctuations, and noise in a [32] R. Jugulum, “Comparison between Mahalanobis-Taguchi System and
centrifugal pump,” Journal of Turbomachinery, vol. 119, pp. 506–515, Artificial Neural Networks,” Journal of Quality Engineering Forum,
1997. vol. 10, pp. 60–73, 2002.
[10] I. S. Koo and W. W. Kim, “The development of reactor coolant pump [33] W. H. Woodall, R. Koudelik, K.-L. Tsui, S. B. Kim, Z. G.
vibration monitoring and a diagnostic system in the nuclear power Stoumbos, and C. P. Carvounis, “A review and analysis of the
plant,” ISA Transactions, vol. 39, pp. 309–316, 2000. Mahalanobis-Taguchi System,” Technometrics, vol. 45, pp. 1–15,
[11] M. S. Lebold, K. Maynard, K. Reichard, M. Trethewey, J. Hasker, C. 2003.
Lissenden, and D. Dobbins, “A non-intrusive technique for on-line [34] C. T. Su and Y. H. Hsiao, “An Evaluation of the robustness of MTS
shaft crack detection and tracking,” in Aerospace Conference, 2005, for imbalanced data,” IEEE Trans. Knowledge and Data Engineering,
pp. 1–11. vol. 19, pp. 1321–1332, 2007.
[12] S. Perovic, P. J. Unsworth, and E. H. Higham, “Fuzzy logic system to [35] T. Yang and Y.-T. Cheng, “The use of Mahalanobis-Taguchi System
detect pump faults from motor current spectra,” in Industry Applica- to improve flip-chip bumping height inspection efficiency,” Microelec-
tions Conference, 2001. Thirty-Sixth IAS Annual Meeting. Conference tronics Reliability, vol. 50, pp. 407–414, 2010.
Record of the 2001 IEEE, vol. 1, pp. 274–280. [36] R. Isermann, Fault Diagnosis Systems: An Introduction from Fault De-
[13] N. R. Sakthivel, V. Sugumaran, and S. Babudevasenapati, “Vibration tection to Fault Tolerance. Berlin, Germany: Springer, 2005.
based fault diagnosis of monoblock centrifugal pump using decision [37] A. Soylemezoglu, “Sensor data-based decision making,” Ph.D. Disser-
tree,” Expert Systems with Applications, vol. 37, pp. 4040–4049, 2010. tation, Engineering Management and Systems Engineering, Missouri
[14] N. R. Sakthivel, V. Sugumaran, and B. B. Nair, “Comparison of deci- University of Science and Technology, Rolla, MO, 2010.
sion tree-fuzzy and rough set-fuzzy methods for fault categorization of
mono-block centrifugal pump,” Mechanical Systems and Signal Pro- Ahmet Soylemezoglu received his B.S. degree in Mechanical Engineering from
cessing, vol. 24, pp. 1187–1906, 2010. the Eastern Mediterranean University in Famagusta, Cyprus in 2000. He earned
[15] S. Zhang, T. Asakura, X. Xu, and B. Xu, “Fault diagnosis system for his M.S. degrees in Engineering Management in 2001 and in Manufacturing in
rotary machine based on fuzzy neural networks,” JSME, vol. 46, pp. 2004. He recently earned his Ph.D. in Engineering Management with an em-
1035–1041, 2003. phasis on Manufacturing Engineering at the Missouri University of Science
[16] G. Taguchi, S. Chowdury, and Y. Wu, The Mahalanobis Taguchi and Technology (formerly University of Missouri-Rolla) in 2010. Dr. Soyleme-
System. New York: McGraw Hill, 2001. zoglu’s research interests are discrete event systems, flexible manufacturing sys-
[17] P. C. Mahalanobis, “On the generalized distance in statistics,” in Pro- tems, shop floor control, computer integrated manufacturing, preventive main-
ceedings, National Institute of Science of India, 1936, vol. 2, pp. 49–55. tenance, and shop floor prognostics.
878 IEEE TRANSACTIONS ON RELIABILITY, VOL. 60, NO. 4, DECEMBER 2011

Sarangapani Jagannathan received his Ph.D. degree in Electrical Engineering Can (John) Saygin is an associate professor of mechanical engineering, and a
from the University of Texas, Arlington in 1994. Since September 2001, he has research investigator in the Center for Advanced Manufacturing and Lean Sys-
been at the Missouri University of Science and Technology (former University tems (CAMLS) at the University of Texas at San Antonio (UTSA). He is also
of Missouri-Rolla) where he is currently a Rutledge-Emerson Distinguished the director of the Interactive Technology Experience Center (iTEC), and the
Professor and Site Director for the NSF Industry/University Cooperative Re- director of the Manufacturing Systems and Automation (MSA) Laboratory at
search Center on Intelligent Maintenance Systems. He has coauthored over 90 UTSA. He received his BS (’89), MS (’92), and PhD (’97) degrees in Mechan-
peer reviewed journal articles, most of them in IEEE Trans. with his students, ical Engineering with emphasis on manufacturing engineering from the Middle
180 refereed IEEE conference articles, several book chapters, and three books East Technical University, Ankara in Turkey. As the author or co-author of
entitled “Neural network control of robot manipulators and nonlinear systems” over 90 technical papers, Dr. Saygin’s research interests include network-cen-
published by Taylor & Francis of London in 1999, “Discrete-time neural net- tric manufacturing systems, web-based manufacturing system control, modeling
work control of nonlinear discrete-time systems” CRC Press in April 2006, and and analysis of automated manufacturing systems, flexible manufacturing sys-
“Wireless Ad Hoc and Sensor Networks: Performance, Protocols and Control” tems, and integration of process planning with scheduling and control. He is a
CRC Press, April 2007. He holds 18 patents with several pending. He so far member of the Society of Manufacturing Engineers (SME), Institute of Indus-
supervised the completion of 14 doctoral students, and 26 M.S. students. His trial Engineers (IIE), American Society of Mechanical Engineers (ASME), and
research interests include adaptive and neural network control, computer/com- American Society for Engineering Education (ASEE).
munication/sensor networks, prognostics, and autonomous systems/robotics.