Naval Postgraduate School Monterey, California: Thesis

NAVAL POSTGRADUATE SCHOOL
Monterey, California
THESIS
RELIABILITY ANALYSIS OF THE 4.5 ROLLER BEARING
by
Cole Muller
June 2003
Thesis Advisor: David H. Olwell

Second Reader: Samuel E. Buttrey
Approved for public release; distribution is unlimited.

THIS PAGE INTENTIONALLY LEFT BLANK
REPORT DOCUMENTATION PAGE Form Approved OMB No. 0704-0188
Public reporting burden for this collection of information is estimated to average 1 hour per response, including the time for
reviewing instruction, searching existing data sources, gathering and maintaining the data needed, and completing and
reviewing the collection of information. Send comments regarding this burden estimate or any other aspect of this collection of
information, including suggestions for reducing this burden, to Washington headquarters Services, Directorate for Information
Operations and Reports, 1215 Jefferson Davis Highway, Suite 1204, Arlington, VA 22202-4302, and to the Office of
Management and Budget, Paperwork Reduction Project (0704-0188) Washington DC 20503.
1. AGENCY USE ONLY (Leave blank) 2. REPORT DATE 3. REPORT TYPE AND DATES COVERED
June 2003 Master’s Thesis
4. TITLE AND SUBTITLE: 5. FUNDING NUMBERS
Reliability Analysis of the 4.5 Roller Bearing
6. AUTHOR(S) Muller, Cole
7. PERFORMING ORGANIZATION NAME(S) AND ADDRESS(ES) 8. PERFORMING
Naval Postgraduate School ORGANIZATION REPORT
Monterey, CA 93943-5000 NUMBER
9. SPONSORING /MONITORING AGENCY NAME(S) AND ADDRESS(ES) 10. SPONSORING/MONITORING
N/A AGENCY REPORT NUMBER
11. SUPPLEMENTARY NOTES The views expressed in this thesis are those of the author and do not reflect the official
policy or position of the Department of Defense or the U.S. Government.
12a. DISTRIBUTION / AVAILABILITY STATEMENT 12b. DISTRIBUTION CODE
Approved for public release; distribution is unlimited.
13. ABSTRACT (maximum 200 words)
The J-52 engine used in the EA-6B Prowler has been found to have a faulty design which has
led to in-flight engine failures due to the degradation of the 4.5 roller bearing. Because of cost
constraints, the Navy developed a policy of maintaining rather than replacing the faulty engine
with a re-designed engine. With an increase in Prowler crashes related to the failure of this
bearing, the Navy has begun to re-evaluate this policy. This thesis analyzed the problem using
methods in reliability statistics to develop policy recommendations for the Navy. One method
analyzed the individual times to failure of the bearings and fit the data to a known distribution.
Using this distribution, we estimated lower confidence bounds for the time which 0.0001% of
the bearings are expected to fail, finding it was below fifty hours. Such calculations can be
used to form maintenance and replacement policies. Another approach analyzed oil samples
taken from the J-52 engine. The oil samples contain particles of different metals that compose
the 4.5 roller bearing. Linear regression, classification and regression trees, and discriminant
analysis were used to determine that molybdenum and vanadium levels are good indicators of
when a bearing is near failure.
14. SUBJECT TERMS 15. NUMBER OF

Reliability Statistics, Data Analysis PAGES
83
16. PRICE CODE

17. SECURITY 18. SECURITY 19. SECURITY 20. LIMITATION
CLASSIFICATION OF CLASSIFICATION OF THIS CLASSIFICATION OF OF ABSTRACT
REPORT PAGE ABSTRACT
Unclassified Unclassified Unclassified UL
NSN 7540-01-280-5500 Standard Form 298 (Rev. 2-89)
Prescribed by ANSI Std. 239-18
i
ii
Approved for public released; distribution is unlimited.
RELIABILITY ANALYSIS OF THE 4.5 ROLLER BEARING
Cole Muller
Ensign, United States Navy
B.S., United States Naval Academy, 2002
Submitted in partial fulfillment of the

requirements for the degree of
MASTER OF SCIENCE IN APPLIED SCIENCE (OPERATIONS RESEARCH)
from the
NAVAL POSTGRADUATE SCHOOL

June 2003
Author:
Cole Muller
Approved by:
David H. Olwell
Thesis Advisor
Samuel E. Buttrey
Second Reader
James N. Eagle
Chairman, Department of Operations Research
iii
iv
ABSTRACT
The J-52 engine used in the EA-6B Prowler has been found to have a faulty
design which has led to in-flight engine failures due to the degradation of the 4.5 roller
bearing. Because of cost constraints, the Navy developed a policy of maintaining rather
than replacing the faulty engine with a re-designed engine. With an increase in Prowler
crashes related to the failure of this bearing, the Navy has begun to re-evaluate this
policy. This thesis analyzed the problem using methods in reliability statistics to develop
policy recommendations for the Navy.
One method analyzed the individual times to failure of the bearings and fit the
data to a known distribution. Using this distribution, we estimated lower confidence
bounds for the time which 0.0001% of the bearings are expected to fail, finding it was
below fifty hours. Such calculations can be used to form maintenance and replacement
policies.
Another approach analyzed oil samples taken from the J-52 engine. The oil
samples contain particles of different metals that compose the 4.5 roller bearing. Linear
regression, classification and regression trees, and discriminant analysis were used to
determine that molybdenum and vanadium levels are good indicators of when a bearing
is near failure.
v
vi
TABLE OF CONTENTS
I. INTRODUCTION ........................................................................................................ 1
A. OVERVIEW.......................................................................................................... 1
B. BACKGROUND ................................................................................................... 2
C. PROBLEM & PURPOSE .................................................................................... 3
D. APPROACHES ..................................................................................................... 3
II. LIFE DATA ANALYSIS ............................................................................................ 5

A. APPROACH.......................................................................................................... 5
B. DATA ..................................................................................................................... 5
C. ASSUMPTIONS.................................................................................................... 6
D. DISTRIBUTIONS AND PLOTS ......................................................................... 6
E. SELECTING A DISTRIBUTION TO MODEL THE DATA ........................ 11
F. PARAMETER ESTIMATES AND CONFIDENCE BOUNDS ..................... 12
G. RESULTS FROM LIFE DATA ANALYSIS ................................................... 15
III. OIL FILTER ANALYSIS ...................................................................................... 17

A. TECHNIQUES.................................................................................................... 17
B. DATA ................................................................................................................... 17
C. SIMPLE LINEAR REGRESSION MODEL ................................................... 22
D. REGRESSION TREE MODEL ........................................................................ 30
E. CLASSIFICATION TREE MODELS.............................................................. 35
F. DISCRIMINANT ANALYSIS........................................................................... 37
IV. CONCLUSIONS AND RECOMMENDATIONS ................................................ 41

A. CONCLUSIONS ................................................................................................. 41
B. RECOMMENDATIONS.................................................................................... 41
C. FUTURE RESEARCH....................................................................................... 42
APPENDIX A. FAILURE-TIME DISTRIBUTION FUNCTIONS.......................... 43

APPENDIX B. COMMON LIFETIME DISTRIBUTIONS...................................... 45
APPENDIX C. PROBABILITY PLOTTING............................................................. 49
APPENDIX D. METHODS OF PARAMETER ESTIMATION ............................. 51
APPENDIX E. FISHER MATRIX BOUNDS ............................................................. 55
APPENDIX F. REGRESSION TREES ...................................................................... 57
APPENDIX G. CLASSIFICATION TREES ............................................................. 59
vii
APPENDIX H. DATA USED FOR LIFE DATA ANALYSIS ................................. 61
APPENDIX I. DATA COLLECTED FOR OIL FILTER ANALYSIS ................... 63
LIST OF REFERENCES ............................................................................................... 65
INITIAL DISTRIBUTION LIST .................................................................................. 67
viii
LIST OF FIGURES
Figure 1 - Failed Roller Bearing. 1

Figure 2 - Probability plot using the 3-paramter Weibull distribution to fit data. 7
Figure 3 - Failure rate using the 3-parameter Weibull distribution to fit data. 8
Figure 4 - Probability plot with the lognormal distribution modeling the data. 9
Figure 5 - Failure rate of the data when fitted to the lognormal distribution. 10
Figure 6 - Probability plot with the 2-parameter Weibull distribution to fit data. 11
Figure 7 - Probability plot of the lognormal distribution using MLE. 13
Figure 8 - Bearing Stage vs. Vanadium plot. 18
Figure 9 - Bearing Stage vs. Silver plot. 19
Figure 10 - Bearing Stage vs. Molybdenum plot. 19
Figure 11 - Bearing Stage vs. Iron plot. 20
Figure 12 - Bearing Stage vs. Vanadium plot with perfect classification. 21
Figure 13 - Bearing Stage vs. Molybdenum plot with perfect classification. 21
Figure 14 - Residuals of the model plotted against the standard normal quantiles. 23
Figure 15 - Histogram of the residuals from our simple linear model. 24
Figure 16 - Plot of the fitted values vs. the residuals of our initial model. 24
Figure 17 – Residual plot with 2-way interactions vs. standard normal quantiles. 26
Figure 18 - Histogram of residuals of model that includes two-way interactions. 27
Figure 19 - Plot of the fitted values vs. residuals of our model with interactions. 27
Figure 20 - Histogram of residuals of model with V, Mo and V:Mo. 28
Figure 21 – Residual plot with V, Mo and V:Mo vs. standard normal quantiles. 29
Figure 22 - Plot of the fitted values vs. residuals for model with V, Mo and V:Mo. 29
Figure 23 - Initial regression tree. 31
Figure 24 - Deviance vs. Size. 33
Figure 25 - Pruned regression tree. 34
Figure 26 - Classification tree. 36
Figure 27 - Histograms of the LCF values. 38
ix
x
LIST OF TABLES
Table 1- Table of coefficients for initial linear model. 23

Table 2 - Coefficients of linear model with interactions. 25
Table 3 - Coefficients that model bearing stage by V, Mo and V:Mo. 28
Table 4 - Linear discriminant coefficients. 38
xi
xii
ACKNOWLEDGMENTS
The completion of this thesis signifies a major milestone in my continued

education. That being said, without the help of certain individuals, this work would not
have been possible.
Foremost, I would like to thank my thesis advisor, Dave Olwell, for all the help he
has been to me. I also thank my second reader Sam Buttrey for his assistance, in
particular, the S-Plus help he has given me. Next, I thank NAVAIR for providing the
necessary data to make this thesis work possible. Lastly, I would like to thank Lynn
Yates for answering the many questions I had about the data.
xiii
xiv
I. INTRODUCTION
A. OVERVIEW
The EA-6B Prowler is the only operational electronic attack aircraft flown by the
U.S. military today. It has a unique mission that is necessary to the success of most
military operations. The Prowler conducts Electronic Attack (EA) missions in support of
U.S. and coalition air operations that are vital to our national security interests. It
possesses a unique mission capability that is in continual high demand to support
worldwide military operations and is the only aircraft in the U.S. inventory that is
dedicated to the suppression of enemy air defenses (Pitts, 2002).
The Prowler design is about thirty years old and there is a need to replace it with
an airframe that incorporates today’s advanced technology. Until a replacement aircraft
is built, however, the 122 Prowlers in the U.S. inventory are the only EA-capable aircraft
(Pitts, 2002).
Figure 1 - A failed roller bearing.
1
In the late 1980’s, it was found that the J-52 engine, the engine used in the
Prowler, had a faulty design. Oil flow was insufficient through the engine, and
consequently there were problems within the engine. The 4.5 bearing was the part that
most often failed because of the poor engine design. After a few in-flight failures of the
bearing, the Navy ordered a study to be done in order to determine how to correct the
problem. The study recommended that the engine be redesigned, but due to cost
constraints, the Navy decided to replace the bearings whenever the aircraft came in for a
major inspection. The cost of a bearing is around two thousand dollars, while a new
engine design could be well into the millions of dollars (Barber, 2002).
B. BACKGROUND
In November 2001, two Prowlers crashed within a week, and both crashes were
most likely caused by the failure of the 4.5 roller bearing (Selinger, 2002). In the
Whidbey Island crash, the 4.5 roller bearing in the “right engine failed, touching off a
chain reaction that ultimately destroyed both engines. Other parts in the turbine section in
which the bearing was housed ‘were liberated’ and rocketed into the left engine, severing
fuel and hydraulic lines along the way, an examination of the engines revealed” (Barber,
2002).
Immediately following these two crashes, the Navy temporarily grounded its fleet
of EA-6B Prowlers, but eventually relaxed this restriction, mainly due to the necessity to
have Electronic Attack aircraft available in Operation Enduring Freedom. The Navy has
begun to procure a new generation of Electronic Attack aircraft, but until that time, a new
replacement policy must be put into place regarding this bearing.
The Prowler is a twin-engine aircraft, and with approximately 122 Prowlers in
service for the Navy, there are nearly 250 Pratt and Whitney J-52 engines, not counting
spare engines, with the 4.5 bearing operational in the fleet today (Pitts, 2002).
2
C. PROBLEM & PURPOSE
This thesis looks at two distinct approaches determining the reliability of the 4.5
bearing. We want to model the reliable life of the bearing, as well as identify bearings
that are likely to fail shortly. Reliability is defined as the probability that a system will
perform its intended function for a specified period of time. In this thesis, reliability is of
more interest than availability, which is the fraction of time that a system is available for
use (Meeker and Escobar, p.2). Determining the availability of the aircraft is not as
important, for this issue, than the reliability. We want to decrease, if not eliminate, in-
flight failures, so reliability is the natural way to analyze this situation.
Reliability, in this case, is the probability that the engine bearing will not fail in
flight. We are not as concerned with other failures within the engine, but want to look
specifically at the 4.5 bearing and the effect it has on the reliability of the aircraft. The
life scale that we will use is not actual age, but instead hours in which the engine was
operating, or engine hours.
D. APPROACHES
The first of the two approaches examined is a standard life data analysis. A life
data analysis is done when the collected data is the time to failure of many identical units,
including suspensions for those units not yet failed. This data is fit to a distribution
model to get failure rates, quantiles, and probabilities. This type of analysis can produce
estimates of the probability of a failure before a specified time, the hazard function at a
specified time, and also the proportion of units that will fail in a specified time.
It is important to use the correct distributional model, or results from the model
can be greatly different than those that actually take place. Many times extrapolation is
necessary to gain useful information from the data. For example, if a test has run for 400
hours and what is desired is the proportion failing at 900 hours, using the wrong
distributional model can very well lead to incorrect conclusions. Similarly, if we have
data on 250 engines but want to calculate the reliable life associated with 99.9999%
reliability, we will extrapolate to a point well before the first observed failure.
3
The times to failure, given in engine operating hours, that are attributed to the 4.5
bearing were used to fit the data to a distribution. This allows us to predict probability of
failure of the bearing over a time interval with a given confidence level. This in turn
allows us to evaluate and suggest crude replacement policies.
Life data analysis gives us some good insight into the overall problem, but there
are limitations. Since the data is binary (failed or suspended bearings), this type of
analysis did not use other available information. This analysis recommends that the Navy
should replace each bearing very frequently to maintain a certain reliability level, but this
interval is impractical. Therefore, an effective policy based on this type of analysis
would be a cost-ineffective policy to adopt.
The other approach used regression techniques from data analysis such as simple
linear regression, classification and regression trees (CART), and discriminant analysis.
This process explored the levels of metallic materials found in the filter of the engine.
The idea was that an increase in certain levels of metals and other elements in the filter
could be used to predict failure of a bearing. The elements that make up the bearing that
are used in this approach are silver, molybdenum, iron, and vanadium.
The oil filter analysis program is done by “evaluating the residue the filter
collects” in order to predict the failure of a component, in this case the 4.5 roller bearing.
When the bearing fails, the particles “that are generated are too large to be picked up in a
spectroanalysis . . . are picked up in the filter and can be evaluated using filter analysis”
(Oil Lab, 2002).
Data on hand that was used for this approach is the percentage of different metals
found in the oil filter, along with the cumulative mass of each metal within the filter. The
collected data also indicates which oil samples come from engines that have failed
bearings so a classification can be made.
The results from these approaches are used to create recommendations to the
Navy to influence policy so that the risk of in-flight bearing failures is reduced. The
recommendations identify appropriate approaches discussed in this paper to pursue and
why.
4
II. LIFE DATA ANALYSIS
A. APPROACH
The first approach used was a standard life data analysis. A life data analysis is
done when the collected data is the time to failure of many identical units, including
suspensions for those units not yet failed. The idea of this approach is to develop a model
that fits a distribution to the data, which is described below. Once we found an
acceptable distribution for our data, we then estimated the parameters of the distribution
so that we could generate probability plots and failure-time distribution functions, which
are described more thoroughly in Appendix A, and also estimate confidence intervals to
aid in our analysis.
B. DATA
The data we collected is a mixture of two different types of data. The first type is
known as complete data. Complete data is data that has exact values for the time to
failure of the bearing. Right-censored data is data that has not failed by a certain time,
which is the case for the vast majority of the data that we have. Both types of data were
used for the life data analysis (Meeker and Escobar, p.34).
Listed in Appendix H is the data we used for this analysis. There are sixty-six
data entries, with eleven failures and fifty-five suspensions. This is considered a small
sample size because of the number of observed failures is low.
To analyze this type of data properly, we must understand the methods of analysis
and parameter estimation appropriate to censored data. These methods are similar to
those methods used on complete data, but are modified to fit the needs of censored data.
For example, in a complete data set, it is easy to calculate the mean of the data. We
simply sum the data entries and divide by the number of entries to get this desired mean.
However, when dealing with censored data, we need to adjust for the interval of
uncertainty that comes with each data point that is censored. To take the mean of
censored data, we must account for the data points that did not fail by the end of the data
5
collection period. Therefore, the probability of failure, also known as unreliability, needs
to be adjusted. These methods are discussed further in Appendices C and D.
C. ASSUMPTIONS
We expect the 4.5 bearing of the J-52 engine to have an increasing failure rate.
As the bearing gets worn down, it should be more susceptible to failure. There is the
possibility of infant mortality, in which case the failure rate function would have a
‘bathtub-shaped’ look. We consider the data to be well past the time of infant mortality,
so we can disregard this possibility for the failure rate.
We are primarily concerned with the distribution of the life of the bearing during
early life. We considered distributions that have strictly increasing failure rates during
early life, but included distributions, such as the lognormal, that fail the strictly increasing
failure rate test over the whole domain.
D. DISTRIBUTIONS AND PLOTS

We used a commercial software package, Weibull++, to help with the analysis.
Weibull++ is able to take the data and fit distributions to this data and also estimate the
parameters of these distributions. It also has a function that tests the goodness of fit of
each distribution, and this function then recommends the distributions that are the best fit
for the data. The goodness of fit tests are a weighted score of Anderson-Darling,
Kolmogorov-Smirnov, and Maximum Likelihood tests.
Properties of common lifetime distributions are given in Appendix B. For our
data, the first distribution that Weibull++ recommends to us is the two-parameter
exponential distribution. One of the characteristics of the exponential distribution is that
it is a memory-less distribution, i.e. it has a constant failure rate. However, one of our
assumptions was an increasing failure rate of the bearing, at least during the portion of
time that we explored. Because of this, we eliminated the exponential distribution from
consideration.
The next best-fit distribution is the three-parameter Weibull distribution. We took
a look at the probability plot, seen in Figure 2, and it seemed like a decent fit. Appendix
6
C describes the idea of probability plotting in greater detail. Looking at the failure rate
function in Figure 3, we saw that the PDF is zero before approximately 580 hours. This
would make it hard to analyze the data and come up with a one-sided confidence bound
given a desired reliability level. Given this, we preferred to look at other distributions
that fit the data fairly well, yet allowed these desired one-sided confidence bounds.
Probability - Weibull
99.00
Weibull
j52
90.00 W3 MLE - SRM MED

F=11 / S=55
CB[FM]@95.00%
1-Sided-U [T1]
50.00
Unreliability, F(t)
10.00
5.00
1.00
10.00 100.00 1000.00 10000.00
Time, (t)
β=1.1484, η=2878.4851, γ=580.7850
Figure 2 – This is the probability plot using the three-parameter Weibull distribution to fit our data.
The curved line on the right is the data fit to the model on the original scale. The straight line is the
data fit to the model after the translation parameter has been subtracted. Above the straight line is a
one-sided upper confidence bound at the 95% level for the failure probability. The best-fit line is not
a great fit to the data, which is not uncommon, as MLE estimates for small samples such as this are
known to be biased. Suspensions are noted on the horizontal axis.
7
Unreliability vs Time Plot
1.00
Weibull
j52
W3 MLE - SRM MED

F=11 / S=55
CB[FM]@95.00%
1-Sided-U [T1]
0.80
Unreliability, F(t)=1-R(t)
0.60
0.40
0.20
0
0 400.00 800.00 1200.00 1600.00 2000.00
Time, (t)
β=1.1484, η=2878.4851, γ=580.7850
Figure 3 – This is the failure rate using the three-parameter Weibull distribution to fit the data. The
failure rate is conditional on the translation parameter, which is fixed. We can see as a result that
the function is zero before approximately 580 flight hours. Also drawn on this plot is the upper
confidence bound. Because of this, it is hard to determine a one-sided confidence bound with a given
reliability level, and thus we want to look at another distribution that would allow us to do so. We
also want to at least admit the possibility of early failures, so a translation parameter is not desired.
We then looked at the next recommended distribution, which is the lognormal

distribution. The probability plot in Figure 4 looked slightly worse than that from the
three-parameter Weibull distribution, although the difference was minor. The failure rate
function was increasing, at least in the time interval that we explored. In Figure 5, it can
be seen that the failure rate increased until approximately 2800 hours, which is well past
the scope of this analysis. We were only interested in what occurs in the first couple
hundred hours, so we disregarded the fact that the failure rate will eventually decrease.
8
Probability - Lognormal
99.00
Lognormal
j52
L2 MLE - SRM MED

F=11 / S=55
CB[FM]@95.00%
1-Sided-U [T1]
Unreliability, F(t)
50.00
10.00
5.00
1.00
100.00 1000.00 10000.00
Time, (t)
µ=7.8072, σ=0.6647
Figure 4 - This is the probability plot with the lognormal distribution modeling the data. This looks
to be a better fit than in Figure 2. Once again, we also have included the 95% upper confidence
bound on unreliability.
9
Unreliability vs Time Plot
1.00
Lognormal
j52
L2 MLE - SRM MED

F=11 / S=55
CB[FM]@95.00%
1-Sided-U [T1]
0.80
Unreliability
0.60
0.40
0.20
0
100.00 2080.00 4060.00 6040.00 8020.00 10000.00
Time, (t)
µ=7.8072, σ=0.6647
Figure 5 - This is the failure rate of the data when fitted to the lognormal distribution. Despite being
a non-monotonic function, it is an increasing function up until a certain point. Our expected time to
failure is well before this point, so we can use the lognormal distribution to fit our data.
The last distribution that we explored was the two-parameter Weibull distribution.
Even though Weibull++ ranks this distribution lower than the lognormal, we wanted to
take a look at the probability plot of the data fit to this distribution, seen in Figure 6. We
can see from this plot that this distribution is a bad fit, especially in the lower tail of
interest, and for this reason we omitted this distribution from our analysis.
10
Probability - Weibull
99.00
Weibull
j52
90.00 W2 MLE - SRM MED

F=11 / S=55
CB[FM]@95.00%
1-Sided-U [T1]
50.00
Unreliability, F(t)
10.00
5.00
1.00
100.00 1000.00 10000.00
Time, (t)
β=2.1732, η=2995.2324
Figure 6 – This is the probability plot modeling the data with the two-parameter Weibull
distribution. We again have a 95% one-sided upper confidence bound. The best-fit line is clearly not
a good fit.
E. SELECTING A DISTRIBUTION TO MODEL THE DATA

We decided that the exponential distribution and the two-parameter Weibull
distribution were not good fits to model our data based on our assumptions. The three-
parameter Weibull distribution seemed to represent our data a bit better than those of the
lognormal distribution; however it did not allow for easy calculations of the one-sided
confidence bound for small probabilities of failure. Also, there is a history of modeling
bearing life with the lognormal distribution (Lawless, p.228). Because of this, we chose
the lognormal distribution to model our data.
11
F. PARAMETER ESTIMATES AND CONFIDENCE BOUNDS
Now that we selected a distribution, we needed to estimate the parameters of this
distribution. We let Weibull++ calculate them, using either Rank Regression (RRX) or
Maximum Likelihood Estimators (MLE). For more on these two techniques and the
differences between them, refer to Appendix D. Since we have censored data with a
sufficiently large sample size, we used the MLE method to estimate the desired
parameters. Using this method, we got the following parameters for our model: µ=
7.8072, σ = 0.6647.
We set a reliability level that is high enough so that the risk of failure is very
small. Due to the catastrophic nature of just one failure, we set the reliability level to be
99.9999%. Once the probability of failure is greater than 0.0001%, we would
recommend that the bearing be replaced. Using the probability plot generated from
Weibull++, seen in Figure 7, after approximately 100 engine hours, the point estimate of
the reliability level has dropped to 99.9999%, our given threshold level.
12
Probability - Lognormal
99.00
Lognormal
j52
L2 MLE - SRM MED

F=11 / S=55
CB[FM]@95.00%
1-Sided-U [T1]
50.00
Unreliability, F(t)
10.00
5.00
1.00
0.50
0.10
0.05
0.01
5.00E-3
1.00E-3
5.00E-4
1.00E-4
10.00 100.00 1000.00 10000.00
Time, (t)
µ=7.8072, σ=0.6647
Figure 7 – This is the probability plot of the lognormal distribution with the estimated parameters
using MLE. This plot shows the expected time of failure for 0.0001% of all aircraft to be 104.3
hours, with a 95% upper confidence bound (Fisher-Matrix) of 42.8 flight hours. Once again, the
circles represent the failures and the triangles represent the suspensions.
We also calculated the exact value of the expected engine hours until 0.0001% of
the bearings failed using the MLE method. Using the process described earlier on
MLE’s, we obtained 104.3 engine hours as the expected time where 0.0001% of the
bearings should fail.
We also determined a confidence interval for this time. Since we had a small
amount of data, the time 104.3 can vary quite a bit from sample to sample. We sought a
confidence interval that is 95% confident that the probability of failure is less than
0.0001%. A discussion on confidence bounds, to gain a better understanding of how they
are derived, follows.
Confidence bounds estimate the precision of an estimator. The estimator for us
was the number of engine hours by which 0.0001% of the engines had failed. Since we
could not get the true reliability value because of the data being only a sample of J-52
13
engines, we could only estimate this value using different probabilistic methods. Our
parameter estimates will change slightly from sample to sample, thus changing the
reliability value that we estimated. Confidence bounds produce an interval that should
contain the reliability value for a certain percentage of intervals generated from this
sampling scheme, in our case, 95%.
Confidence bounds can be one or two-sided. Two-sided bounds have both an
upper and lower bound to the interval, while one-sided bounds have either an upper or
lower limit, but not both. We can further break one-sided bounds down into two types.
We can have either one-sided upper confidence bounds or one-sided lower confidence
bounds. One-sided upper confidence bounds have an upper bound on the interval and a
lower bound being the start time, in most cases zero. One-sided lower confidence
bounds, thus, have a lower bound on the interval and an upper bound that goes out to
infinity. In our case, we did not care what happened at later time periods, and therefore
had no utility for an upper confidence bound. We gained more useful information if we
chose to use the one-sided lower confidence bound to construct our interval for the
estimator of the engine hours it takes for 0.0001% of the engines to fail.
There are two different approaches to constructing confidence bounds that we
considered for our model. They were Fisher Matrix (FM) confidence bounds and
Likelihood Ratio (LR) bounds. Fisher Matrix bounds use the assumption of
asymptotically normal MLE estimates of the parameters, and therefore need a sufficiently
large sample size. Likelihood Ratio bounds do not depend on asymptotic normality as
strongly as doe Fisher Matrix bounds, and thus work better than FM when the sample
sizes are smaller. We discuss each of these methods in greater detail.
To construct confidence bounds using FM, we need to know the mean and the
variance of the function that we are examining. Since we are inquiring about the cdf, or
unreliability function, of the lognormal distribution at 0.0001% unreliability, we need the
mean and variance of this function in order to use Fisher Matrix bounds.
The mean can be calculated easily enough using MLE’s and their invariance
properties, and we have already done so earlier. Now we need the variance of the
function. The variance of a function depends on the variance of each of the parameters
14
estimated in the function, as well as the covariance between them. Appendix E gives a
more in-depth look at this idea.
For our data, the Fisher Matrix method gave us a 95% one-sided lower confidence
bound of 42.8 engine hours. Thus, the Fisher Matrix method suggested that a 99.9999%
reliability threshold with 95% confidence was 42.8 engine hours.
The Likelihood Ratio bound method is based upon the following equation:
 L (θ ) 
 ≥ χ α ,k ,
(1.1) −2 ln  2
 L (θˆ ) 
where L (θ ) is the likelihood function with unknown parameter(s) θ , L θˆ () is the
likelihood function calculated with our estimated MLE parameters, χ 2α ,k is the chi-
()
squared test statistic with probability α and k degrees of freedom. L θˆ is calculated
using the MLE method, and since we know the parameter estimates, the only unknown
term in the equation is L (θ ) . We can then solve for this term (Meeker and Escobar,
p.185).
Since we chose the lognormal distribution to model the data, we have two
parameters that can vary. Thus, a set of values will satisfy the equation. Numerical
methods are then used to find this set. For our data, using the Likelihood Ratio method to
calculate our 95% one-sided lower confidence bound, we got a value of 31.1 engine
hours for the lower confidence bound. Thus, a 99.9999% reliability threshold with 95%
confidence was 31.1 engine hours, according to this method.
A more conservative approach would be to use the method that gets the lower of
the two confidence bounds, which would be using the Likelihood Ratio bound method.
Both methods, however, suggest replacing the bearing too frequently for practical
purposes.
G. RESULTS FROM LIFE DATA ANALYSIS

The Navy currently employs a policy that replaces the bearings at major
inspection intervals, which vary but can be more than one thousand engine hours apart,
which is drastically high compared with the results from our model (Barber, 2002). The
15
point estimate of the probability of such a failure by one thousand engine hours is 0.088.
This can be seen in Figure 7, although it is difficult to read from the plot. Given this, and
that five of the sixty-six bearings failed before this time, we can see that a better policy
should be implemented. Replacing the bearings every thirty or forty hours, though, might
not be the best policy either. The cost-effectiveness of this policy would not be very
high. If we take a look at the data, the first failure occurred after 646 engine hours. This
suggests that the bearings more than likely can be used a lot longer than 31.1 hours by
accepting slightly more risk. This is a shortcoming of our 99.9999% reliability
requirement. Replacing bearings every thirty or forty hours discards a lot of useful life in
each bearing without gaining much more reliability. Considering the high cost of both
replacing the bearing and maintenance, this does not seem to be a very efficient policy.
In light of this, we want to come up with a better model to more accurately determine
when a bearing will fail.
16
III. OIL FILTER ANALYSIS
A. TECHNIQUES
The second approach used techniques from data analysis: simple linear regression
models, classification and regression trees (CART), and discriminant analysis. These
processes used the levels of metallic materials found in the filter of the engine as a
predictor of failure. The idea was that an increase in certain levels of metals and other
elements in the filter could be used to predict failure of a bearing. The main elements
that make up the 4.5 bearing are molybdenum (4%), iron (90%), and vanadium(1%); the
cage that holds the bearing is made of silver. From our analysis, we determined the
levels of each of the metals found in the oil filter that are the best indicators of a failed
bearing.
B. DATA
As the engine is used, friction between the bearings and the shaft cause particles
of the metallic bearings to break off from the bearing. These particles get mixed in with
the oil in the oil filter. The data that has been collected is a sample of the debris found in
the oil filter. The sample is placed on a patch that is one square centimeter in area and
the mass of each metal found in this sample is recorded. The units of the level of each
metal found in the sample are grams per square centimeter. The cumulative amount of
each metal is the data that is used in the analysis.
Each time the oil filter is changed, the contents of the filter are analyzed for the
presence of the aforementioned metallic substances. The data we shall use is updated
every time the filter for a specific engine is changed, in order to represent the additional
amounts of the different metals found in the single oil filter. We are interested in the
cumulative build-up of the metals inside the filter because this could represent the overall
life span of the bearing, and therefore could be the best predictor available as to the
health of the individual bearing.
There are twenty-one sample engines from which the data was collected. The
levels of silver, molybdenum, iron, and vanadium are recorded for each engine. Also
17
recorded is the health of each bearing. The health of the bearing is broken down into a
integer value between zero and five. Zero indicates that the bearing was brand new,
while a five indicated that the bearing had completely failed. When the bearing begins to
skid and wear on the rollers, it is said that the bearing stage is one. A bearing stage of
two indicates there is noticeable wear on the bearing rollers. Stage three is classified as
when the cage of the bearing begins to crack. When the cage incurs severe wear, the
bearing stage is four. We can interpolate between the stages to come up with a
continuous function for bearing stage. Figures 8-11 below present the plots of the level
of each metal versus the bearing stage.
4
Bearing Stage
3
2
1
0.0 0.002 0.004 0.006 0.008 0.010 0.012
Vanadium
Figure 8 – This is the Bearing Stage vs. Vanadium plot. There seems to be some positive correlation
between the two. We found ρ = 0.855.
18
4
Bearing Stage
3
2
1
0.0 0.5 1.0 1.5
Silver
Figure 9 – This is the Bearing Stage vs. Silver plot. It is not clear if there is any correlation between
the two. We found ρ = 0.359.
4
Bearing Stage
3
2
1
0.0 0.02 0.04 0.06 0.08
Molybdenum
Figure 10 – This is the Bearing Stage vs. Molybdenum plot. Again, it looks like there is some positive
correlation between the two. We found ρ = 0.854.
19
4
Bearing Stage
3
2
1
0.2 0.4 0.6
Iron
Figure 11 – This is the Bearing Stage vs. Iron plot. There may be a slight positive correlation
between the two, but it is not quite clear. We found ρ = 0.304.
Figure 8 shows that the level of vanadium and bearing stage are positively
correlated. In general, as the level of vanadium found in the oil filter sample rises, so
does the bearing stage. This is also true of molybdenum, seen in Figure 10. Iron and
silver do not show this trend. In fact, these two metals do not show any obvious trend
that would help predict the stage of the bearing. We can also see in Figures 12 and 13
that the levels of vanadium and molybdenum in this sample can be classified perfectly,
meaning that a naïve model by inspection would yield no misclassification of any of the
bearings.
20
4
Bearing Stage
3
2
1
0.0 0.002 0.004 0.006 0.008 0.010 0.012

Vanadium
Figure 12 – This is the Vanadium vs. Bearing Stage plot with perfect classification. If we classify
bearings with a stage of 2.5 or greater as bearings in need of replacement, then we can say that if the
level of vanadium found in the oil filter is greater than approximately 0.0055 g/cm2, the bearing
should be replaced.
4
Bearing Stage
3
2
1
0.0 0.02 0.04 0.06 0.08
Molybdenum
Figure 13 - This is the Molybdenum vs. Bearing Stage plot with perfect classification. If we classify
bearings with a stage of 2.5 or greater as bearings in need of replacement, then we can say that if the
level of molybdenum found in the oil filter is greater than approximately 0.0041 g/cm2, the bearing
should be replaced.
21
C. SIMPLE LINEAR REGRESSION MODEL
First, we created a simple linear regression model to predict the bearing stage
given the levels of vanadium, iron, molybdenum, and silver found in the oil filter. The
regression model equation is of the form
(1.2) E[Y ] = b0 + b1 x1 + ... + bk −1 xk −1 ,
where k is the number of input variables that will be used to predict the response variable
(Hamilton, p.66). To perform this linear regression, we assumed that the expected value
of the response variable given the input variables was linear. For our data, we had four
input variables. Therefore, the model that we constructed was of the form
(1.3) yˆ = b0 + bV xV + bFe xFe + bMo xMo + bAg xAg ,
where bo is the estimate for the intercept and bk is the estimate for the slope of the k-th
input variable.
For each of our data entries, we used the input variables to generate an expected
response variable. We then took the difference between this expected value and the
observed value. Using all our data entries, we calculated the sum of the squares of this
difference, which is known as RSS.
n 2
(1.4) RSS = ∑ ( yi − yˆi )

i =1
The coefficients that minimized the RSS are the ones we chose for our model (Devore,
p.498). We used S-Plus to generate these coefficients for our model.
By creating a linear model in S-Plus, we got the coefficients seen in Table 1. We
see that the p-values of each individual element are high, suggesting that the removal of
any one of these elements will not hurt the model. The R2 of the model is 0.7516.
22
Value Std. Error t value Pr(>|t|)
(Intercept) 1.0 0.32 3.15 0.0062
V 186.36 211.20 0.8824 0.3906
Fe -0.71 1.60 -0.4447 0.6625
Mo 19.97 32.13 0.6215 0.5430
Ag 0.41 0.67 0.6152 0.5471
Table 1- Table of coefficients for initial linear model. If the model is accepted, we assume that the
errors are normally distributed.
We checked the residuals for normality. Figure 14 shows a plot of the residuals against
the standard normal quantiles. The plot does not look too bad, except the lower tail
seems a bit skewed, so we decided that the residuals did not look to be normally
distributed.
1.0
0.5
0.0
Residuals
-0.5
-1.0
-2 -1 0 1 2
Quantiles of Standard Normal
Figure 14 - Residuals of the model plotted against the standard normal quantiles. The fitted line is
the line that the residuals should fall on if they are normally distributed.
23
8
6
4
2
0
-1.5 -1.0 -0.5 0.0 0.5 1.0 1.5
Residuals
Figure 15 - This is a histogram of the residuals from our simple linear model. The residuals are not
quite symmetric, and therefore do not look to be normally distributed.
1.0
0.5
0.0
Residuals
-0.5
-1.0
1 2 3 4
Fitted : V + Fe + Mo + Ag
Figure 16 - This is a plot of the fitted values vs. the residuals of our initial model.
24
Figure 15 is a histogram of the residuals, and from here the residuals also do not look to
be from the normal distribution. Figure 16 is a plot of the fitted values against the
residuals. This plot does not seem to be suggesting that our model is a good fit. After
looking at various transformations of the data, there were no candidates whose residuals
looked a lot better than this model, however.
We continued with this model and added two-way interactions into it. After we
added the interactions, we ran the stepAIC function in S-Plus. The stepAIC function
automates the stepwise addition of terms from our model that significantly decreases the
residual sum of squares (Venables and Ripley, p.186). Table 2 shows the coefficients of
the terms that remained in our model. The R2 of this model is 0.8598, which is
significantly higher than other models we looked at.

(Intercept) 0.64 0.59 1.08 0.3019
V -1497.62 819.27 -1.83 0.0948
Fe 4.97 3.44 1.44 0.1766
Mo 248.73 115.33 2.16 0.0540
Ag -4.49 2.52 -1.78 0.1028
V:Fe 4126.47 2222.65 1.86 0.0903
V:Mo 4396.29 2935.63 1.50 0.1624
V:Ag 549.03 377.26 1.46 0.1735
Fe:Mo -803.90 345.49 -2.33 0.0401
Fe:Ag 8.34 3.78 2.21 0.0493
Table 2 – These are the coefficients of linear model with interactions. Some terms have high p-
values, suggesting that we can remove that term and the model would not be affected much.
In Figure 17, we see the plot of the residuals of our model with interactions
against the standard normal quantiles. Figure 18 is the histogram of our model with
interactions. These plots show some improvement from the previous model with respect
to the residuals being normally distributed. We removed some of the terms with high p-
values individually, and although the R2 value of each of the new models did not drop
25
significantly, the residuals looked less normal than the residuals from this model. Figure
19 is the fitted values versus residuals plot. This plot suggests to us that this model may
be better than our previous one.
1.0
0.5
Residuals
0.0
-0.5
-1.0
-2 -1 0 1 2
Figure 17 – This is a plot of the residuals of the model with two-way interactions against the standard
normal quantiles. Residuals that are perfectly normal would all fall on the line, but the residuals
from our model are close to the line in most cases, therefore we conclude the residuals are normally
distributed.
26
10
8
6
4
2
0
-1.5 -1.0 -0.5 0.0 0.5 1.0 1.5
Residuals
Figure 18 – This is a histogram of the residuals of the model that includes two-way interactions. The
shape of this histogram is somewhat close to that of the normal distribution. Given the small sample
size, we shall assume that the residuals are normally distributed.
2
1
Residuals
0
-1
1 2 3 4
Fitted : (V + Fe + Mo + Ag)^2
Figure 19 - This is the plot of the fitted values vs. residuals of our model with interactions. There
seems to be one extreme outlier, although after looking at the data point, it is not immediately clear
why this point is so extreme.
27
From these models, we saw an emphasis was placed on the presence of vanadium
and molybdenum in the oil filter when determining the bearing stage, so we wanted to
construct a model that simply modeled the bearing stage as a function of the presence of
these two metals together. The coefficients of such a model are seen in Table 3. This
model has an R2 of 0.7674.

(Intercept) 1.20 0.34 3.55 0.0025
V 71.02 178.10 0.40 0.6950
Mo 11.00 24.34 0.45 0.6569
V:Mo 2483.82 1972.94 1.26 0.2251
Table 3 – This is the table of coefficients of the linear model that models bearing stage by simply the
level of vanadium, molybdenum, and the interaction between the two.
8
6
4
2
0
-1.5 -1.0 -0.5 0.0 0.5 1.0
Residuals
Figure 20- Histogram of Residuals of model with only vanadium, molybdenum and the interaction
between the two.
28
0.5
0.0
Residuals
-0.5
-1.0
-2 -1 0 1 2
Figure 21 - Residuals of model with vanadium, molybdenum, and the interaction between the two.
0.5
0.0
Residuals
-0.5
-1.0
2 3 4 5
Fitted : V + Mo + V:Mo
Figure 22 - This is the fitted values versus residuals plot for our model that included just V, Mo, and
the interaction between the two.
29
Figures 20-22 show the plots of the residuals, which do not look as normally
distributed as the residuals from the previous model. This last model showed that the
levels of vanadium and molybdenum found together in the oil filter are a good indicator
of the health of the associated bearing.
Overall, the best model is probably the second one, although there are many terms
chosen by stepAIC for inclusion in the model that had high p-values. However, the plots
of the residuals look to be closer to the normal distribution than those of any other model,
and the R2 value of this model is also significantly higher than the other models. We
looked at other regression tools to construct some more models in order to validate this
result as best we can. Also, Figures 8 and 10 indicate that the response is probably
piecewise linear, which makes the use of standard regression models problematic.
D. REGRESSION TREE MODEL
Our second approach to analyze our data was to use regression trees. Regression
trees use least squares regression to develop a stepwise tree structure. In regression, there
is a response variable, y, and independent variable, x. We use the x-values to construct a
predictive model. These models can be used to accurately predict the response variable
for future x-values. They also can be used to see the relationships between the x and y
variables. Refer to Appendix F for more details on regression trees.
We analyzed the oil filter data using regression trees. The data has five ordered
variables, one being the bearing stage and the other four being the level of the
corresponding metal found in the filter analysis. Since we constructed a model that
predicts the bearing stage based on the levels of the different metals found in the filter,
our independent variables were the levels of the four metals, and our response variable
was the bearing stage.
We used the computer program S-Plus to help with the construction of our
models. The tree function in S-Plus was used to create our first regression tree. We
declared the response variable to be the bearing stage and all the other variables as our
30
inputs. S-Plus did the calculations of the deviance and constructed a tree that reduced the
total deviance of the model. In Figure 23, we see the initial tree that S-Plus built based
on our data.
V<0.00544194
|
Ag<0.250423
4.117
Ag<0.163453
1.180
1.780 1.560
Figure 23 – Initial regression tree with the intermediate nodes labeled with the appropriate split and
the terminal nodes labeled with the appropriate predicted bearing stage. This model has four
terminal nodes. The ‘yes’ branch is to the right and the ‘no’ branch is to the left. Accordingly, the
mean bearing stage for an oil sample with V < 0.00544194 and Ag > 0.250423 is 1.180.
From this tree, we see that the first split was determined by the amount of
vanadium that is found in the oil filter. This model said that if the total mass of vanadium
found in the filter is less than 0.00544194 g/cm2, then we should follow the branch to the
left of the node, which gets us to another intermediate node based on the level of silver in
the filter. If the level of vanadium is greater than 0.00544194 g/cm2, then we branch off
to the right of the first node, which leads to a terminal node. The expected bearing stage
31
at the terminal node is 4.117, which is the average of all the bearing stages from the data
entries that have a vanadium level greater than 0.00544194 g/cm2.
There are four terminal nodes in this model, with a calculated deviance of 0.2445.
We see that this model incorporates only the levels of vanadium and silver when
determining the predicted bearing stage. This suggests that the levels of molybdenum
and iron found in the oil filter have little to no effect on the stage of the bearing or are
confounded with other predictors. If we look at the tree, we notice that all of the terminal
nodes that descend from the split where the vanadium level is less than 0.00544194 g/cm2
have predicted bearing stage values less than two. This suggests that the biggest
indicator of bearing health is the amount of vanadium that is found in the oil filter.
We now turn our attention to pruning this model. Since we split each node until
splitting made no difference, we have run the risk of over-fitting our model to the data.
To counteract this, we must prune the tree backwards to get a good balance of descriptive
and predictive power. This technique creates a nested sequence of sub-trees, from which
the best-sized tree is chosen (Venables and Ripley, p.327).
A tree is determined to be of best size when the deviation is the smallest. In S-
Plus, we ran a cross-validation function to determine what size tree gives us the smallest
deviance. We used cross-validation to suggest to us the best size of our tree without
over-fitting the model to the data. Looking at the deviances of the different sizes, we
chose the best size to be two. Figure 24 shows us a picture of the deviance compared to
the size of the tree.
32
29.00 0.80 0.12 -Inf
40
30
deviance
20
10
1.0 1.5 2.0 2.5 3.0 3.5 4.0

size
Figure 24 – There is a large drop off in deviance when the size of the model moves from one to two.
For sizes larger than two, the deviance does not significantly drop, leading to the best size chosen to
be two.
From this picture, we see that the deviance does not decrease noticeably for size
values of greater than two. Since we wanted the simplest model possible, we chose the
smallest size among all the deviances that are close, which gives us the best size of two
for our model.
Then we wanted to prune our tree appropriately, so we took this recommendation
and ran the prune function, which pruned our model down to a tree with the desired size.
Figure 25 shows us our new model, which is a tree of size two.
33
V<0.00544194
|
1.507 4.117
Figure 25 – The pruned regression tree that is of size two. This tree indicates that the only split that
really matters is the split on the vanadium levels.
This tree splits the original node with the same variable as our previous tree. This
is because all pruning did was remove splits in nodes that are too far down the tree. This
new model agrees with what was stated earlier, that the best model simply looks at the
level of vanadium in the oil filter. When this level goes above a certain threshold, the
predicted stage of the bearing then changes.
The regression tree that we have constructed as our model to predict the stage of
the bearing by the level of certain metals found in the oil filter is a very simple one. It
suggests that vanadium may be used as a lone indicator of the health of the bearing, as
suggested in Figure 12. This model almost seems a bit too simple, so we will take a third
approach to analyzing this oil filter data to see if we can gain similar results.
34
E. CLASSIFICATION TREE MODELS
When analyzing the oil filter data, we did not care as much about what the exact
bearing stage was, but instead all we really wanted to know was whether or not the
bearing had failed. To do so, we needed to modify the data that we used when we
constructed our regression tree model. Instead of an ordered response variable for our
bearing stage, we classified the bearing stage into two distinct types: failed bearings and
non-failed bearings.
We used another tree-based model to analyze our data: the classification tree. The
idea is to select splits in each node so that the subsets created by the split are more pure
than the original node. Classification trees are similar to regression trees, except the
response variable is categorical instead of a continuous numerical value. Appendix G has
more on this topic.
Again, we used S-Plus to construct our classification trees. We also used many of
the same functions in S-Plus used to build regression trees. Before we began building our
model, though, we first converted the data into a data type that we could use. We decided
on a way to label whether a bearing has failed or not. We chose a cutoff value such that
all bearings with bearing stage values higher than this cutoff were said to be failed
bearings. Based on the levels of perfect classification observed in Figures 12 and 13, we
chose the cutoff of the bearing stage to be 2.5, which is after the stage when there is
noticeable wear on the bearing and before the cage has cracked. All bearings in our data
that have bearing stages that are less than 2.5 were said to have not ‘failed’, while those
with bearing stage values of 2.5 or greater were considered to have failed. With the
proper data type, we built our classification model.
35
V<0.00544194
|
FALSE TRUE
Figure 26 - Classification tree with the intermediate nodes labeled with the appropriate split and the
terminal nodes labeled with the predicted response. This model has three terminal nodes.
Figure 26 shows the classification tree that was generated in S-Plus from our data
set. At the terminal nodes, the response tells whether the bearing is projected to be a
failed one. True responses indicate that the bearing has failed, while false indicates
otherwise. This model says that the level of vanadium is plays a large part in telling
whether a bearing has failed or not. It suggests that the level of vanadium found in the oil
filter of a failed bearing is at least 0.00544194 g/cm2, which is the same cutoff level used
in our regression tree model.
With our classification tree, we had a misclassification rate of zero. This follows
from Figure 12. The model we constructed had perfect classification of our sample,
suggesting that we had a very good model. We did not have a need to prune the model
any further, as we were already guaranteed to have minimal misclassification (zero) with
a small model size (two terminal nodes).
The results from both the regression and classification tree models are very
similar, mainly because the processes used to construct them are similar. It is more
intuitive to use the classification tree model because this model gives us a clearer answer.
With the regression tree model, the response variable was the expected bearing stage, but
with the classification model, the output was simply whether the bearing is projected to
have failed or not. This is the question we are trying to answer, so the classification tree
36
model makes more sense for us to use. Regardless, though, both models give strikingly
similar results.
F. DISCRIMINANT ANALYSIS
The last regression technique we explored was discriminant analysis.
Discriminant analysis is another classification technique that tries to find a set of
coefficients that defines a function that separates groups of variables maximally. This
function is known as a Linear Classification Function (LCF). This function can be
written LCF = wV
1 1 + ... + wkVk , where w is the discriminant coefficient, V is the set of
variables, and k is the number of variables.

Discriminant analysis is used in order to find common groupings of variables. A
threshold, which is used to classify objects into groups, is determined. If the LCF is
greater than or equal to the threshold level, the object is classified into one group. If the
LCF is less than the threshold, then it is in the other group (Groth, p.34-35).
Once again, we used S-Plus to aid with constructing the model. After creating a
model through the lda function, we looked at the coefficients of the linear discriminant
generated for our model. These coefficients are those of the LCF. We have our w’s for
our function. The input variables, Vk, are used to generate the response variable to
determine which group each data entry is classified into. We also obtain the threshold
value from S-Plus.
We shall omit the precise mathematics behind linear discriminant analysis, but for
a more complete overview of the topic, consult Chapter 11 in Modern Applied Statistics
with S-Plus by W.N. Venables and B.D. Ripley.
The data we used for this approach was the same used in the classification tree
approach. We split the stage of the bearings into two groups. A bearing stage of 2.5 or
greater suggested that the bearing was in need of replacement. Those bearings less than
2.5 were said to not need replacement. The linear discriminant coefficients for our model
are shown in Table 4.
37
LD1
V 194.6005467
Fe 0.7540444
Mo 50.0668160
Ag -0.1222041
Table 4 - Linear discriminant coefficients
Therefore, the LCF is the following: LCF = 194.60VV + 0.75VFe + 50.07VMo − 0.12VAg .
2.5
2.0
1.5
1.0
0.5
0.0
-2 0 2
group FALSE
2.5
2.0
1.5
1.0
0.5
0.0
-2 0 2
group TRUE
Figure 27 - These histograms show what LCF values each data entry has. The threshold value has
been normalized to zero to make it easy to see which side of the threshold each data entry is on. This
threshold once again gives us perfect classification.
In Figure 27, we can see the LCF values for each data entry. Each value has been
adjusted so that the threshold value is zero. This makes it easy to see which side of the
threshold a data entry is predicted to be on. Once again, we have perfect classification of
our sample from the model. We can clearly see that the coefficients for vanadium and
molybdenum are the dominant terms in the equation, which leads us to conclude that the
38
amounts of vanadium and molybdenum found in the oil filter are the key to determining
whether or not the bearing should be replaced. This result is consistent with the other
models that we have constructed.
39
40
IV. CONCLUSIONS AND RECOMMENDATIONS
A. CONCLUSIONS
The current Navy policy does not give an acceptable reliability level for the
bearing (currently 91.2%, according to our model), so we used life data analysis to find a
replacement policy that ensures that 99.9999% of the bearings will not fail with 95%
confidence. This analysis gave us a lower confidence bound of less than fifty hours,
which is not a practical policy either. This policy would cost too much and not be very
efficient due to much useful life of each bearing being discarded. Therefore, a policy
should be based on some other type of analysis. The analysis that we have used is the oil
filter analysis. Every model that we constructed greatly depended on the level of
vanadium found in the oil filter sample. Some models also suggested that when
molybdenum is present with vanadium, this dependence was even greater. We have
therefore determined from this analysis that the level of vanadium found in the oil filter is
a key indicator of bearing failure, with molybdenum being a secondary factor when
vanadium is already present. The regression models seem to be too crude due to the lack
of a larger sample size and non-linear behavior seen in Figures 8-11, so we do not
suggest exact levels to use to monitor the health of the bearings. Nonetheless, these
models have been helpful in identifying the indicators of a failed bearing.
B. RECOMMENDATIONS
Recommendations for policy changes are to use the results from the life data
analysis to obtain an acceptable interval of time in which to gather oil filter samples. To
implement an effective policy, we round our result of 42.8 engine hours down to forty
engines hours as our time between samples. When the cumulative levels of vanadium get
too large in the oil filter, we suggest replacing the degraded bearing. The cutoff that our
model suggests is 0.00544194 g/cm2. Despite this being a result of a small sample size,
we recommend a cutoff at this point until a larger sample size can be collected. We also
41
recommend that data on the 4.5 bearing continue to be collected and the health of the
bearings monitored.
C. FUTURE RESEARCH
When future data is collected, it would be helpful to record the age of the bearing
as well as the levels of the metals found via the oil filter analysis. With this extra piece of
data, we could combine the two approaches that we have done to come up with a more
precise model that could be used to determine failed bearings based upon both the levels
of the metals found in the oil filter and the age of the bearing. This would allow us to
estimate the margin of safety in the oil filter approach.
42
APPENDIX A. FAILURE-TIME DISTRIBUTION FUNCTIONS
This appendix gives an overview of the failure-time distribution functions used in the analysis.
The probability density function, or PDF, of a distribution completely specifies

the probability distribution of a continuous random variable. The PDF is denoted as f(t).
The area under the PDF curve is always equal to one. The cumulative density function,
F(t), is the total area under the PDF curve up to the point in time, t. Thus, we can
represent the CDF in Equation A.1.
t
(A.1) F ( t ) = ∫ f ( s ) ds
0
The CDF therefore represents the probability of failure in the interval [0,t].
The reliability function is simply the probability of a non-failure over the interval
[0,t]. We know that the CDF is the probability of failure over the same interval, and we
now call this the unreliability function. Equation A.2 shows the reliability function.
(A.2) R (t ) = 1 − F (t )
This is also known as the survival function.

The final function relating to a given distribution that we wish to explore in life
data analysis is the hazard rate, or more commonly referred to as the failure rate. The
failure rate is the probability of failure at time t in the next ∆t of time, given that the
system has not failed before that time. The failure rate equation is given in Equation A.3.
P(T ∈ (t , t + ∆t ))
(A.3) h(t ) = lim
∆t → 0 ∆t ⋅ R (t )
The relation of the failure rate to the PDF and CDF is shown in Equation A.4
(Meeker and Escobar, p.28).
f (t )
(A.4) h (t ) =
R (t )
43
44
APPENDIX B. COMMON LIFETIME DISTRIBUTIONS
This appendix gives a brief overview of a few common lifetime distributions.
Weibull Distribution
The Weibull distribution is a reliability distribution commonly used in life data
analysis. It tends to be a good model for times-to-failure of both electronic and
mechanical equipment. This distribution can have up to three parameters. If the location
parameter is assumed to be zero, then the two-parameter Weibull distribution results.
Different behaviors can be modeled by the Weibull distribution, depending on the
values of these parameters. The shape parameter, denoted as β, can affect the
characteristics of the shape of the PDF curve, reliability, and failure rate. The shape
parameter is also called the slope parameter, since it gives the slope of the CDF when
plotted on probability paper. Changing the values of β can have distinctively different
effects on the distribution properties. The property that we shall explore in detail is the
hazard function.
It can be shown that when 0 < β < 1, the failure rate is a monotonic, decreasing
function. When β = 1, the failure rate is constant for all values of t. This is because when
β = 1, the Weibull distribution reduces to the memory-less distribution, the exponential
distribution. When β > 1, we can show that the failure rate is a monotonic, increasing
function (Devore, p.179-180). Equations B.1-B.4 show the reliability function, CDF,
PDF, and hazard function explicitly for the Weibull distribution.
β
t
− 
η 
(B.1) R (t ) = e
β
t
− 
η 
(B.2) F (t ) = 1 − e
β
β −1 t
β t − 
η 
(B.3) f (t ) =   e
η η 
β −1
β t
(B.4) h(t ) =  
η η 
45
Exponential Distribution
The exponential distribution has a location parameter and may have a translation
parameter. It is usually considered to be the easiest distribution to work with. Because it
is so easy to manipulate, it is often misused. It is used to represent systems that are
assumed to have a constant failure rate, as seen in the special case of the Weibull
distribution with β = 1. The failure rate for the exponential distribution is λ, which is a
constant. The two-parameter exponential distribution has scale parameter, 1/λ, or η, and
location parameter, γ. The effect of γ is that the distribution is simply shifted along the x-
axis. For positive values of γ, this implies that failures cannot occur before time t = γ
(Devore, p.174-175). The CDF, PDF, reliability function, and hazard function can all be
obtained from those of the Weibull distribution with β = 1.
Normal Distribution
The normal distribution is the most used distribution and is occasionally used for
reliability analysis and times-to-failure of electronic and mechanical systems. There are
two parameters in the normal distribution that need to be estimated, namely, the mean, µ,
of the normal times to failure and the standard deviation, σ, of the times to failure. It
should be noted that to use the normal distribution with life data, we must be careful to
only consider it when the mean is relatively high and the standard deviation small in
comparison. This is due to the PDF of the normal distribution extending to negative
infinity, leading to negative times-to-failure, which usually does not make much sense.
Since the normal distribution has much utility in the modeling of life data, we can justify
this with a high mean compared to the standard deviation. The hazard function of the
normal distribution is a monotonically increasing function (Devore, p.158-162).
Equations B.5-B.7 give the CDF, PDF and hazard function for the normal distribution.
t
(B.5) F (t ) = ∫
−∞
f ( s )ds
46
2
1  t −µ 
1 − 
σ 
(B.6) f (t ) = e 2
σ 2π
f (t )
(B.7) h(t ) =
1 − F (t )
Lognormal Distribution
The lognormal distribution can be often used to model reliability data, cycles-to-
failure in fatigue, loading variables in probabilistic design and material strengths. This
distribution is used when the natural logarithms of the times-to-failure are normally
distributed. Like the normal distribution, the lognormal distribution has two parameters;
µ = E[ln(t )] , which is the location parameter, and σ = Var[ln(t )] , which is the scale
parameter. This distribution is similar to the normal distribution in many ways, although
we must explore the hazard function a bit to see that it is increasing monotonically only
for a while, where it then begins to monotonically decrease out towards infinity (Devore,
p.181-182). The CDF, PDF, and hazard function for the lognormal distribution are
similar to those from the normal distribution and are shown in Equations B.8-B.10. The
CDF and PDF of the normal distribution are here denoted Fn(t) and fn(t).
(B.8) F (t ) = Fn ⋅ ln(t )
1
(B.9) f (t ) = f n ⋅ ln(t )
σ
f (t )
(B.10) h(t ) =
1 − F (t )
47
48
APPENDIX C. PROBABILITY PLOTTING
This appendix gives a more detailed explanation of probability plotting.
The essence of probability plotting is that if the plot is based on the correct
distribution, then plotting sample points should result in a nearly-straight line (Devore,
p.186). Probability plotting tries to linearize the cumulative density function (CDF) of
the distribution. For the Weibull distributions, we use the logarithms of the times to
failure as the inputs, or our x-values, and we must have some method to obtain our y-
values, or median ranks value. This method is the median ranks method. From complex
calculations, the plots can then be drawn. Refer to Statistical Methods for Reliability
Data written by W. Meeker and L. Escobar for further detail.
49
50
APPENDIX D. METHODS OF PARAMETER ESTIMATION
This appendix gives the details of both RRX and MLE methods of parameter estimation.
To fit a line to the data to estimate the parameters of the distribution, we could use
a least squares method, known as rank regression on X (RRX), to perform our linear
regression. This method takes the sum of the squares of the horizontal difference
between the actual value, x, and the x-value that lies on the regression line at point y. The
regression line that minimizes this sum is considered to fit best and therefore is the line
that we shall choose.
To develop the probability of failure on complete data, we simply rank the failure
times in order of occurrence relative to each other. Since all data points are known, this
can be done. But with censored data, we cannot always be certain which event will
happen next. If a data point was suspended at a time before another data point failed,
there is ambiguity as to which data point was the first to fail. We must modify our way
of ranking the failures. To do so, we use a weighting scheme. We find out how many
different scenarios can occur with the suspension occurring first, and then the same with
the failure occurring first. Then we calculate the mean order number (MON), which is
represented in Equation D.1.
(D.1) MON =
( ca + db )
c+d
The variable a is the position of the failure if the suspension happened first, c is the
number of different scenarios that can occur if the suspension occurs first, b is the
position of the failure if the failure happens first, and d is the number of different
scenarios that can occur from this position. This mean order number is then the value
assigned to the failure. After doing this method to each failure, we have a mean order
number for every failure entry. We now can proceed as we would with complete data,
generating our y-values and knowing our x-values exactly to come up with our
probability plots to obtain the parameters of the chosen distribution (ReliaSoft, p.39, 53-
55).
51
While RRX is a popular method of analysis of censored life data, there is a
problem with it. Since the mean order number is simply the position of the failed data
points relative to other failures, there is no compensation for how spread-out the failures
are from each other. To rectify this, we explore the method of Maximum Likelihood
Estimators, commonly referred to as MLE’s.
First, we shall explain the method of MLE’s for complete data in order to get a
better grasp at how MLE’s will handle censored data better. The underlying idea of
MLE’s is to get the most likely value of the parameters of the chosen distribution that
best describes the data. Say we have a probability density function (PDF) that is as
follows:
(D.2) f ( x;θ1 ,...,θ k ) ,
where x is a continuous random variable and there are k unknown parameters to be
estimated. The likelihood function seen in Equation D.3 is the product of each f(xi),
where i is an element of the set of all x-values, or failure times in our case.
(D.3) L = Π f ( xi ;θ1 ,...,θ k )
i
Taking the partial derivatives of the natural log of the likelihood function with respect to
the parameters and setting these equal to zero allows us to solve the system of k equations
simultaneously to obtain our estimated θ values for the k parameters to be estimated.
Under regularity conditions, MLE’s converge to the correct values as the sample
size increases, making MLE’s a very efficient and accurate method of parameter
estimation for large sample sizes. These regularity conditions, however, do not hold
when using a threshold parameter (Meeker and Escobar, p.622). By the Central Limit
Theorem, the large sample size also allows us to assume the distribution of the estimates
to be normal, allowing us to use the Fisher Matrix confidence bounds that will be
explained later. MLE’s also deal much better with right-censored data, and we now shall
explore the method of using MLE’s with censored data (Devore, p.268).
The likelihood function for MLE analysis of data with censored data needs to
account for not just the failures, but also the suspensions as well. We use the same
52
technique described above, but we now add another term to the equation to account for
each suspension. Thus, we get a likelihood function as follows:
(D.4) L = Π f ( xi ,θ1 ,...,θ k ) ⋅ Π [1 − F ( x j ;θ1 ,...,θ k )] ,
i∈I j∈J
where I is the set of complete data points (i.e. failure times) and J is the set of suspended
data points (Meeker and Escobar, p.174-176).
RRX and MLE methods both assume that a distribution is already known, but life
data does not usually tell the analyst what distribution it follows, if any, so the analyst
must use a variety of factors in order to decide on a distribution and estimate its
parameters. Some of the aspects of distributions to consider include the probability
density function, cumulative density function, reliability and unreliability functions, the
mean life function, commonly referred to as the mean-time-to-failure (MTTF), and the
hazard function, which is also known as the failure rate.
53
54
APPENDIX E. FISHER MATRIX BOUNDS
This appendix gives a few more details about Fisher Matrix Bounds.
Using the log-likelihood function, Λ, obtained in the MLE method, we can
construct the Hessian matrix of the function, given in Equation E.1 (ReliaSoft, p.68).
 ∂2Λ ∂ 2Λ 
 ∂θ 2 ∂θ1 ∂θ 2 
F = 2
1
(E.1)
 ∂Λ ∂ 2Λ 
 
 ∂θ 2 ∂θ1 ∂θ 22 
Using the estimated values of the parameters, we can invert F to come up with the
covariance matrix, where we can then get the variance of each of the parameters and the
covariance between them, seen in Equation E.2.
−1
 ∂ 2Λ ∂2Λ 
 ∧
ˆ
∧
  ∂θ1 ∂θ 2 
 Var (θ1 ) Cov(θˆ1 ,θˆ2 )   ∂θ12
(E.2) =
 ∧ ˆ ˆ ∧
ˆ   ∂ 2Λ ∂2Λ 
Cov(θ1 ,θ 2 ) Var (θ 2 )   
 ∂θ 2 ∂θ1 ∂θ 22 
We then can go back and use these values to obtain the variance of the function using the
delta method. Now that we have the expected value and the variance, we can go ahead
and use the assumption of normality to construct our one-sided lower confidence bound
(Devore, p.270-271).
55
56
APPENDIX F. REGRESSION TREES
This appendix gives a more detailed account of regression trees.
Regression trees are models that take input variables and use them to predict the
response variable. To use the x-values as a predictor, we need to define how to measure
the accuracy of this predictor. One way to do so is to take the average error,
N
1
(F.1)
N
∑| y
n =1
n − d ( xn ) | .
yn is the response value for the n-th data entry, and d(xn) is the predicted yn value from the
model. This is known as least absolute deviation regression. Of course, we also have the
more traditional measure of accuracy in regression, which is the average squared error,
N
1
(F.2)
N
∑(y
n =1
n − d ( xn )) 2 .
This is commonly called least squares regression. We shall use the method of least
squares to define our measure of accuracy (Breiman, et al, p.222).
The model consists of a sequence of binary splits, where each split results in two
more nodes in the model. There are two types of nodes, terminal and intermediate.
Intermediate nodes are nodes that branch further down, depending on certain criteria
described later. Terminal nodes are nodes where the predicted response variable has been
determined, and this value is constant. At each intermediate node, one of the input
variables determines which branch of the tree to follow. The predicted y-value, or output,
of the model at a terminal node is simply the average of all the y-values of the data entries
at that specific node. Each intermediate node has a selected input variable associated
with it and the split depends solely on the value of a single variable. A cutoff level is
determined for this input variable, and all data entries that have values less than the cutoff
level are branched on one side of the tree, and the values greater than the cutoff level
branch off along the other side, which creates two new nodes further down the tree. The
tree continues to branch at each intermediate node until all branches reach a terminal
node.
57
At each node we are really asking a question with a binary response. The
question is whether xi ≤ c, where c is defined as the cutoff value and xi corresponds to the
i-th variable, which is where the node split is based. If the answer to the question is yes,
then we follow the branch to the left. If the response is no, the branch to the right is
followed.
There are three necessary elements needed when determining what the tree should
look like in our model. The first is a way to select the split at intermediate nodes. Within
each node, the error is calculated and the split that reduces the overall error the greatest is
selected. Therefore, the regression tree simply looks to maximize the decrease in the
error by splitting nodes as necessary.
The second element needed is a rule for determining when a node is terminal.
Since the model is seeking to minimize the error, splitting at a node occurs when the error
is significantly decreased. Thus, if splitting a node does not result in a significant
decrease in the error, then no split occurs and we say that the node is terminal. There are
other criteria used by S-Plus that we do not get into here.
The last element is a rule to assign the y-values to each terminal node. We have
already stated that the y-value for each terminal node is the average of all the y-values for
all data points at that specific node, which yields a constant value. This value is the value
that minimizes the within node squared error.
The resulting tree of nodes forms the model that we use to make future
predictions. We take the input data and run it down the tree, following the branches that
our data point satisfies. When we reach a terminal node, we take the expected y-value for
that node to be the predicted y-value for our data point (Breiman, et al, p.228-232).
58
APPENDIX G. CLASSIFICATION TREES
This appendix gives a more detailed account of classification trees.
As it was with regression trees, the construction of classification trees is based on

the same three principles. Instead of an expected y-value, though, the output is the
predicted class within the response variable to which the data point most likely belongs.
The probability associated with each class is calculated and the class with the highest
probability is selected.
The process of selecting node splits is different only because we wish to minimize
a different impurity function. Terminal nodes are determined when the impurity does not
decrease with a split in that node, just as we labeled nodes as terminal in regression trees
when the error ceased to decrease.
In a classification tree, we assume the responses to be multinomial. A
multinomial response is one where each observation has a probability associated with it
of resulting in each of the outcomes. These probabilities are denoted pi for the ith
outcome. Thus, for n trials, the probability of seeing ni outcomes of type i is proportional
to Π pi ni . This is the likelihood function and taking the logarithm of this function gives
i
us the log-likelihood, the quantity to be minimized in the classification tree. Again, we

go through the tree and ask a question at each node. If the variable associated with the
node is ordered, then the question remains the same as with regression trees: is xi ≤ c? If
the node split is determined by a categorical variable, then the question to be asked is
whether or not xi is an element of a determined subset of the responses to that variable.
We follow the appropriate branch depending on the answer to the question at that node.
When we reach a terminal node, we look at the assigned categorical response. This is the
predicted response for our data point (Breiman, et al, p. 27-36).
59
60
APPENDIX H. DATA USED FOR LIFE DATA ANALYSIS
Life Condition Life Condition

1085 S 1795 S
100 S 1500 S
1890 F 1628 F
1390 S 1145 S
759 S 152 S
1380 S 246 S
971 S 61 S
861 S 966 S
1165 S 462 S
997 S 437 S
1079 S 887 S
1152 S 1199 S
977 S 159 S
424 S 1022 S
3428 S 763 S
2087 S 555 S
1297 S 646 F
727 S 2238 S
820 S 2294 S
1388 F 897 F
663 S 1153 S
810 S 1427 S
2892 S 80 S
951 F 2153 S
1167 F 767 S
853 S 711 F
546 S 911 S
1203 F 736 S
2181 F 85 S
917 S 1042 S
1070 S 2871 S
799 S 719 S
1231 S 750 F
61
62
APPENDIX I. DATA COLLECTED FOR OIL FILTER ANALYSIS
Bearing Stage V Mass G/CM2 Fe Mass G/CM2 Mo Mass G/CM2 Ag Mass G/CM2
2.1 0.0004 0.3750 0.0071 0.2353
1.3 0.0001 0.7365 0.0146 1.0639
2.0 0.0052 0.1775 0.0383 0.1534
0.5 0.0000 0.3938 0.0251 0.2644
1.4 0.0000 0.1067 0.0031 0.1735
4.2 0.0077 0.2655 0.0544 0.4189
1.0 0.0044 0.2877 0.0348 0.2011
4.5 0.0119 0.3192 0.0836 0.2866
1.3 0.0004 0.0518 0.0052 0.0591
2.0 0.0046 0.2858 0.0365 0.6711
2.0 0.0019 0.0898 0.0164 0.1157
4.7 0.0082 0.6378 0.0800 1.8727
1.9 0.0016 0.1607 0.0178 0.2364
2.0 0.0022 0.0937 0.0230 0.1186
1.0 0.0036 0.1865 0.0235 0.3373
1.1 0.0008 0.0702 0.0097 0.2874
1.6 0.0000 0.0540 0.0032 0.0337
4.0 0.0068 0.2401 0.0592 0.1096
1.4 0.0012 0.1437 0.0146 0.1866
3.0 0.0057 0.4156 0.0439 0.6974
4.3 0.0076 0.2780 0.0573 0.2288
63
64
LIST OF REFERENCES
Barber, M., “Navy Knew of Faulty Engine Part Years Ago,” Seattle Post-Intelligencer,
03 Aug 2002.
Breiman, L., Freidman, J., Olshen, R. and Stone, C., Classification and Regression Trees,
Wadsworth International Group, 1984.
Devore, J., Probability and Statistics for Engineering and the Sciences, Duxbury, 1995.
Groth, Robert, Data Mining: A Hands-On Approach for Business Professionals, Prentice
Hall, 1998.
Hamilton, Lawrence C., Regression with Graphics, Duxbury Press, 1992.
Lawless, J.F., Statistical Models and Methods for Lifetime Data, Wiley, 1982.
Meeker, W. and Escobar, L., Statistical Methods for Reliability, Wiley, 1998.
Oil Lab Analysis, Inc., “Filter Analysis”, Oct 2002.

<http://www.oillab.com/index.htm>, Jun 2003.
Pitts, J., Larsen, R. and Kirk, M., “Letter to Appropriations Committee on EWWG
Priorities”, 25 Apr 2002.
< http://www.house.gov/pitts/initiatives/ew/020425ew-appropspriorities.htm>, Jun 2003.
ReliaSoft Corporation, Life Data Analysis Reference, ReliaSoft Publishing, 2000.
Selinger, M., “Prowler Engine Problems Seen Underscoring Need For New Aircraft,”
Aerospace Daily, 2002.
Venables, W.N., and Ripley, B.D., Modern Applied Statistics with S-Plus, Springer,
1999.
65
66
INITIAL DISTRIBUTION LIST
1. Defense Technical Information Center

Fort Belvoir, Virginia
2. Dudley Knox Library

Naval Postgraduate School
3. Dr. David Olwell

Operations Research Department (Code OR/OL)
4. Dr. Samuel Buttrey

Operations Research Department (Code OR)
5. Lynnwood Yates
Orange Park, FL
6. Craig Paylor
RCM Program Division Manager
Naval Air Depot
Cherry Point, North Carolina
7. ENS Cole Muller

Hatboro, Pennsylvania
67

Naval Postgraduate School Monterey, California: Thesis

Uploaded by

Copyright:

Available Formats

Naval Postgraduate School Monterey, California: Thesis

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Naval Postgraduate School Monterey, California: Thesis

Uploaded by

Copyright:

Available Formats

NAVAL POSTGRADUATE SCHOOL

RELIABILITY ANALYSIS OF THE 4.5 ROLLER BEARING

Thesis Advisor: David H. Olwell

Approved for public release; distribution is unlimited.

14. SUBJECT TERMS 15. NUMBER OF

16. PRICE CODE

RELIABILITY ANALYSIS OF THE 4.5 ROLLER BEARING

Submitted in partial fulfillment of the

MASTER OF SCIENCE IN APPLIED SCIENCE (OPERATIONS RESEARCH)

NAVAL POSTGRADUATE SCHOOL

II. LIFE DATA ANALYSIS ............................................................................................ 5

III. OIL FILTER ANALYSIS ...................................................................................... 17

IV. CONCLUSIONS AND RECOMMENDATIONS ................................................ 41

APPENDIX A. FAILURE-TIME DISTRIBUTION FUNCTIONS.......................... 43

LIST OF REFERENCES ............................................................................................... 65

INITIAL DISTRIBUTION LIST .................................................................................. 67

Figure 1 - Failed Roller Bearing. 1

Table 1- Table of coefficients for initial linear model. 23

The completion of this thesis signifies a major milestone in my continued

Figure 1 - A failed roller bearing.

D. DISTRIBUTIONS AND PLOTS

90.00 W3 MLE - SRM MED

β=1.1484, η=2878.4851, γ=580.7850

W3 MLE - SRM MED

β=1.1484, η=2878.4851, γ=580.7850

We then looked at the next recommended distribution, which is the lognormal

L2 MLE - SRM MED

L2 MLE - SRM MED

90.00 W2 MLE - SRM MED

E. SELECTING A DISTRIBUTION TO MODEL THE DATA

L2 MLE - SRM MED

where L (θ ) is the likelihood function with unknown parameter(s) θ , L θˆ () is the

G. RESULTS FROM LIFE DATA ANALYSIS

0.0 0.002 0.004 0.006 0.008 0.010 0.012

0.0 0.5 1.0 1.5

0.0 0.02 0.04 0.06 0.08

0.2 0.4 0.6

0.0 0.002 0.004 0.006 0.008 0.010 0.012

0.0 0.02 0.04 0.06 0.08

(1.4) RSS = ∑ ( yi − yˆi )

Quantiles of Standard Normal

-1.5 -1.0 -0.5 0.0 0.5 1.0 1.5

Value Std. Error t value Pr(>|t|)

Quantiles of Standard Normal

-1.5 -1.0 -0.5 0.0 0.5 1.0 1.5

Value Std. Error t value Pr(>|t|)

-1.5 -1.0 -0.5 0.0 0.5 1.0

Quantiles of Standard Normal

D. REGRESSION TREE MODEL

1.0 1.5 2.0 2.5 3.0 3.5 4.0

variables, and k is the number of variables.

Table 4 - Linear discriminant coefficients

The probability density function, or PDF, of a distribution completely specifies

This is also known as the survival function.

This appendix gives a brief overview of a few common lifetime distributions.

obtained from those of the Weibull distribution with β = 1.

This appendix gives a more detailed explanation of probability plotting.

Using the log-likelihood function, Λ, obtained in the MLE method, we can