An Overview of the 217Plus

TM
System Reliability Assessment Methodology
In the Third Quarter 2006 edition of the RIAC
Journal, we introduced the Handbook of
217Plus
TM
Reliability Prediction Models that
the RIAC has published to provide insight into
the methodology and models that make up the
217Plus
TM
approach to system reliability as-
sessment [Reference 1]. We briefy introduced
the primary factors that form the basis of the
methodology:
1. Whether information exists on a predecessor
system
2. The amount of empirical reliability data that
is available for that system
3. Whether the reliability analyst chooses to assess
the processes used in system development
Figure 1 provides an overview of the 217Plus
TM

approach to failure rate estimation that is based
on the above three factors. Note that, for the pur-
poses of our discussion, system applies to the
highest level defnition of the item defned within
217Plus
TM
. A system, therefore, can be a true
system, a product, an equipment, an assembly, a
subassembly, i.e., any level of complexity that the
user wishes to defne.
If a system to be analyzed using 217Plus
TM
is an
evolution of a predecessor system (i.e., an earlier,
but similar, confguration to the new design), then
a prediction can be performed on both the pred-
ecessor system and the new system. The results
of these two predicted system failure rates form
the basis of a ratio that can be used to modify
the observed failure rate of the predecessor sys-
tem. The result of this predecessor analysis is
1

in Figure 1.
If enough empirical data (feld, test or both) is
available on the new system to be analyzed, it can
be combined with the 217Plus
TM
predicted failure
rate of the new system using a Bayesian approach
to form the best failure rate estimate possible.
As the quantity of empirical data increases, the
failure rate using the Bayesian combination will
be increasingly dominated by the empirical data.
The result of this Bayesian combination is pre-
sented as
2
in Figure 1.
The minimum amount of analysis required for
a 217Plus
TM
reliability prediction is the summa-
tion of component estimated failure rates, plus
other data that may be available to the analyst.
The current twelve component models used by
217Plus
TM
are included in the Handbook, and
will be introduced in more detail in future edi-
tions of the RIAC Journal. The result of the com-
ponent-based prediction is represented by
IA, new

in Figure 1. This predicted value can be further
modifed within 217Plus
TM
through the applica-
tion of the optional Process Grade Analysis, or
other modifcations to default environmental
stress or operational profles. These modifca-
tions are refected in the failure rate represented
by
predicted, new
in Figure 1.
The rest of this article will discuss each ele-
ment of the 217Plus
TM
methodology presented in
Figure 1 in more detail.
Note that the 217Plus
TM
methodology calculates
failure rates in terms of failures per million calen-
dar hours, not operating hours. Therefore, user
inputs for feld data or user-defned failure rates
need to be converted to a calendar hour basis
prior to incorporating them into a 217Plus
TM
reli-
ability prediction. The conversion factors are:
Calendar hours = Operating hours / Duty cycle
Operating hours = Calendar hours x Duty cycle

IA, predecessor

IA,predecessor
represents the initial failure rate as-
sessment of the predecessor system. This is the
sum of the predicted component failure rates,
David Nicholls, Reliability Information Analysis Center
Figure 1. 217Plus
TM
Approach to Failure Rate Estimation [Reference 2]
PAGE 21
An Overview of the 217Plus
TM
System Reliability Assessment Methodology
and uses the twelve 217Plus
TM
component failure
rate models, data from the RIAC Nonelectronic
(NPRD) and Electronic (EPRD) Part Databases,
or user-defned data on components from other
sources.

observed, predecessor

observed, predecessor
is the observed failure rate of the
predecessor system, and represents the point es-
timate of the failure rate, which is equal to the
number of observed failures divided by the cumu-
lative number of operating hours.
Optional Data
Optional data is used to enhance the predicted
failure rate by factoring in more detailed infor-
mation pertaining to environmental stresses,
operating profle factors, and Process Grades.
217Plus
TM
contains default values for the envi-
ronmental stresses and operational profle, but in
the event that actual values of these parameters
are known, either through analysis or measure-
ment, they should be used instead of the de-
faults. The application of Process Grades within
217Plus
TM
is also optional, allowing the user the
option of evaluating the specifc processes used
in the development and sustainment of a system.
If the process grades are not used, default values
are provided for each process (failure cause), so
that the user can evaluate any or all of the proc-
esses. The use of the Process Grade option of
217Plus
TM
is included in the Handbook, and will
be addressed in more detail in a future edition of
the RIAC Journal.

predicted, predecessor

predicted, predecessor
is the predicted failure rate of the
predecessor system after combining the initial as-
sessment (
IA,predecessor
) with the Optional Data, if
used.

IA,new

IA,new
represents the initial assessment of the
new system. This is calculated as the sum of the
predicted component failure rates, and uses the
217Plus
TM
component failure rate models, data
from the RIAC NPRD and EPRD databases, and
other data that may be available to the analyst.
A reliability prediction performed in accordance
with this method is the minimum level of analysis
that will result in a predicted reliability value.
Applying the Optional Data can further refne
this value.

predicted, new

predicted, new
is the predicted failure rate of the new
system after combining the initial assessment with
the Optional Data, if used. If the Optional Data
is not used, then
predicted, new
is equal to
IA,new
.

1
is the failure rate estimate of the new system
after the predicted failure rate of the new system
(
predicted, new
) is combined with the predicted and
observed information from the predecessor system
(
predicted, predecessor
and
observed, predecessor
, respectively).
The equation that translates the failure rate of the
predecessor system to the new system is:
The values of
predicted,new
and
predicted,predecessor
are ob-
tained using the component reliability prediction
procedures, equations and data previously de-
scribed. The ratio
observed,predecessor
/
predicted,predecessor

accounts for the differences in the predicted and
observed failure rates of the predecessor system.
This ratio inherently accounts for the differences
in the systems that are accounted for in the com-
ponent reliability prediction methodology.
This methodology can be used when the new sys-
tem is an evolutionary extension of predecessor
designs. If similar processes are used to design
and manufacture a new system, and the same re-
liability prediction processes and data are used,
then there is every reason to believe that the
predicted/observed ratio of the new system will
be similar to that observed on the predecessor
system.
This methodology implicitly assumes that there
is enough operating time and failures on which to
base a value of
observed,predecessor
. For this purpose,
the observance of failures is critical to derive a
point estimate of the failure rate (i.e., failures di-
David Nicholls, Reliability Information Analysis Center
continued on page 22
vided by hours). A single-sided confdence level
estimate of the failure rate should not be used.
a
i
a
i
represents the number of failures for the i
th
set
of data on the new system.
b
i
b
i
is the cumulative number of operating hours for
the ith set of data on the new system.
AF
i
AF
i
is the acceleration factor between the condi-
tions of test or feld data on a new system and the
conditions under which the predicted failure rate
is desired. If the data is from feld applications in
the same environment for which the prediction is
desired, the AF value will be one. If the data is
from accelerated test data or from feld data in a
different environment, then the AF value needs
to be determined. If the applied stresses are
higher than the anticipated feld use environment
of the new system, AF will be a value greater than
one. The acceleration factor can be determined
by performing a reliability prediction at both the
test and use conditions, but AF can only be deter-
mined in this manner if the reliability prediction
model is capable of discerning the effects of the
accelerating stress(es) of the test. As an example,
consider a life test in which a system was exposed
to a temperature higher than what it would be
exposed to in feld-deployed conditions. In this
case, the AF can be calculated as follows:
where,

T1
= the predicted failure rate at the test con-
ditions obtained by performing a prediction
of the system at the test conditions

T2
= the predicted failure rate at the use condi-
tions obtained by performing a prediction of
the system at the use conditions
THE JOURNAL OF THE RELIABILITY INFORMATION ANALYSIS CENTER FOURTH QUARTER - 2006
An Overview of the 217Plus
TM
System Reliability Assessment Methodology
continued from page 21
b
i

b
i
is the effective cumulative number of hours of
the test or feld data used. If the tests were per-
formed at accelerated conditions, the equivalent
number of hours needs to be converted to the
conditions of interest, as follows:
a
o
a
o
is the effective number of failures associated
with the predicted failure rate. If unknown, use
0.5. In the event that predicted and observed
data is available on enough predecessor systems,
this value can be tailored. This tailoring method
will be discussed shortly.

2
is the best estimate of the new system failure
rate after using all available data and information.
As much empirical data as possible should be
used in the assessment. This is done by math-
ematically combining
1
with empirical data.
Bayesian techniques are used for this purpose.
This technique accounts for the quantity of data
by weighting large amounts of data more heavily
than small amounts.
1
forms the prior distribu-
tion, comprised of a
0
and a
o
/
1
. If empirical data
(i.e., test or feld data) is available for the system
under analysis, it is combined with
1
based on
the following equation:

2
is the best estimate of the failure rate, and a
o
is
the equivalent number of failures of the prior
distribution corresponding to the reliability pre-
diction. For these calculations, 0.5 should be
used unless a tailored value can be derived. An
example of this tailoring is provided in the next
section. a
o
/
1
is the equivalent number of hours
associated with
1
, and a
1
through a
n
are the
number of failures experienced in each source
of empirical data. There may be n different
sources of data available (for example, each of the
n sources corresponds to individual tests or feld
data from the population of systems). b
1
through
b
n
is the equivalent number of cumulative oper-
ating hours experienced for each individual data
source. These values must be converted to equiv-
alent hours by accounting for any accelerating ef-
fects between the use conditions.
Tailoring the Bayesian
Constant, a
o
, in
2
This section discusses tailoring of the a
o
value
used in the Bayesian equations. The value of a
o

is proportional to the degree of weighting given to
the predicted value (
1
). The constant a
0
is cho-
sen such that the uncertainty in the failure rate
estimate, as calculated with the Chi-square distri-
bution, equates to the observed uncertainty. The
default value of 0.5 to be used in the equation is
based on the observed/predicted ratio from a wide
variety of systems, applications, industries, etc.
As such, there are many noise factors contribut-
ing to the variability in this ratio. However, if the
user of the 217Plus
TM
methodology has enough
data on which to derive a tailored value of a
0
, it
should be derived and used. While the default
value of 0.5 represents the large degree of uncer-
tainty inherent when a diverse data set is used, a
typical 217Plus
TM
user will generally be analyzing
systems with a much more narrow focus, in terms
of system type, environment, operating profle,
etc. As such, with enough data, the value of a
0

can be increased.
To estimate the value of a
o
that should be used, a
distribution of the following metric is calculated
for all systems for which both predicted and ob-
served data is available:
Figure 2. Comparison of Observed Uncertainty with the Uncertainty Calculated Using the Chi-Square Distribution
The lognormal distribution will generally ft this
metric well, but others (i.e., Weibull) can also be
used. The cumulative value of this distribution is
then plotted. Next, the failure rate multipliers, as
determined from the Chi-square distribution, are
calculated and plotted. This Chi-square distribu-
tion should be determined and plotted for various
numbers of failures to ensure that the distribu-
tion of observed/predicted failure rate ratios falls
between the Chi-square values. In most cases,
one, two and three failures should be suffcient.
Next, the plots are compared to determine which
Chi-square distribution most closely matches
the observed uncertainty values. The number
of failures associated with that distribution then
becomes the value of a
0
. Figure 2 illustrates an
example for which this analysis was performed.
As can be seen from Figure 2, the observed un-
certainty does not precisely match the Chi-square
calculated uncertainty for any of the one, two or
three failures used in this analysis. This is likely
due to the fact that the population of systems on
which this analysis is based is not homogene-
ous, as assumed by the Chi-square calculation.
However, the confdence levels of interest are
generally in the range of 60 to 90 percent. In
this range, the Chi-square calculated uncertainty
with 2 failures most closely approximates the ob-
served uncertainty. Therefore, in this example, an
a
0
value of 2 was used.

The uncertainties represented by the distribution
of observed/predicted failure rates are typical of
what can be expected when historical data on
predecessor systems are collected and analyzed
to improve the reliability prediction process. For
example, using this example, one can be 80% cer-
tain that the actual failure rate for a system or
product will be less than 2.2 times the predicted
value.
Next Issue
The next edition of the RIAC Journal (1st
Quarter 2007) will present an introduction to the
217Plus
TM
component failure rate models.
References
1. The RIAC Handbook of 217Plus
TM
Reliability
Prediction Models, Journal of the Reliability
Information Analysis Center, Third Quarter 2006,
available for PDF download from the RIAC at
http://theRIAC.org
2. Denson, W.K., Handbook of 217Plus
TM

Reliability Prediction Models, Reliability
Information Analysis Center (RIAC), 26 May
2006, ISBN 1-933904-02-X