A Practical MTBF Estimate for PCB Design Considering

Component and Non-component Failures

Tongdan Jin, Teradyne Inc.
Peng Wang, United Technologies Research Center
Qiyu Huang, Shanghai Jiaotong University, China
Key Words: MTBF, Mean, Variance, Non-Component Failure Rate

SUMMARY & CONCLUSIONS design errors, poor solder joints, software bugs and process-
related problems. For some products, non-component failures
Accurate reliability prediction for MTBF (mean-time- even dominate failure rates in the product early time. The
between-failures) is always desirable and anticipated before competitive electronic market on the other hand requires the
the new product is ramped up for customer shipment. In new product to be released to customers in the shortest time
reality it is often difficult to obtain the accurate MTBF frame, leaving no time for design engineers to eliminate all
estimate for a new product due to the lack of the testing time potential failures in the pilot line, especially those non-
in pilot line and limited field failure data. In this paper, a component problems (i.e. software bug). Therefore the actual
practical reliability prediction model is presented for field PCBs exhibit low reliability performance, which is in
predicting the MTBF of PCB (printed-circuit-board) in the contrast to the expected high MTBF that is estimated only
design phase. Unlike conventional reliability prediction from component failure rates.
models, which usually focus on parts failure rates, the method
presented here not only incorporates component failures but Failures Breakdown by Root-Cause Catagory
also non-component failures which include design, software, 45%
manufacturing and process issues. Component failure rates 40% PCB-A
are computed using either historical data or the nominal failure 35% PCB-B
30%
rates together with operating conditions such as temperature PCB-C
25%
Qty

and electrical derating. Triangular distributions are used to 20%

model non-component failure rates due to design errors, 15%
software bugs, manufacturing and handling problems. Finally, 10%
5%
the confidence intervals for the new product MTBF are
0%
obtained based on six-sigma criteria. The method was applied
Manufacuring

Software
Design
Update
Component

Process

Others
to the design of a DC/analog instrument board that is used in
the semiconductor testing equipment.

1. INTRODUCTION
Figure 1 PCB failures breakdown by root-cause category
Design for reliability in electronics products has been
Figure 1 presents the failure data from three different
accomplished with the introduction of reliability prediction
PCB product lines. Data is breakdown by root-cause
tools in 1960s. Two most widely used tools are Mil-HDBK-
categories: component, design, manufacturing, process,
217 and Bellcore/Telcordia TR-332. Based on these two
software and others. Data are collected from field returns
standards, various commercial software applications were
during one year time frame since first customer shipment. An
implemented to facilitate the estimation of the product
interesting observation is component failures, though the
reliability. Including the Mil-HDBK and Bellcore standards,
largest bar in the pareto chart, contribute less than 45% of
most conventional MTBF prediction methods are developed
failures for all three types of products. For PCB-B component
based on component failure rates and the bill-of-materials
failures count for only 25% of field failures. The second
(BOM) of the product. In this paper the word PCB and
largest failures is Others (i.e. No-Failure-Found or NFF),
product are used interchangeably, so do the component and
which is 25% for all field returns on average. NFF is such
part.
kind of failure that happened at the customer site, but the
MTBF estimated from component failure rates is quite
failure can not be duplicated at the factory repair center.
optimistic in general. In field operation a PCB could fail due
Figure 1 indicates that ignoring the non-component failures,
to defective components and non-component issues or both.
the resulting prediction would be twice or quadruple higher
This is particularly true for the new product during the early
than the actual MTBF in the field.
introduction phase. These non-component failures include
As the component fabrication process continues to

604
Authorized licensed use limited to: University of Pisa. Downloaded on January 15,2024 at 11:39:55 UTC from IEEE Xplore. Restrictions apply.
1-4244-0008-2/06/$20.00 (C) 2006 IEEE
improve, component failure rates have steadily declined over the expected target value, a more accurate, yet practical,
the years to the point where non-component failure sources reliability estimate should be developed such that it not only
become dominant failure rates for a PCB. Figure 2 presents incorporates component failures but also considers non-
the MTBF run chart over the time for PCB-A as given in component issues due to design, manufacturing, software and
Figure 1. The reliability target for this product is 40,000 hours process etc. Very few papers have been published to link the
MTBF. Based on the conventional prediction model, the PCB failures with design, manufacturing, software and
target could be reached in week 41 if only component failure process issues. Gullo[2] and Johnson and Gullo[3] described
rates are considered. The actual MTBF is the smooth line that an in-service reliability prediction tool HIRAP developed by
is lower than the prediction curve. This product in fact takes Honeywell engineers. Using a top-down approach, HIRAP
57 weeks to reach the target MTBF after removing key non- breaks failures into seven categories using similarity
component failure modes through the corrective actions on coefficients between the predecessor products and the new
design, software, manufacturing, and handling. design. Categories 1-5 consist of broad component types that
historically have demonstrated to have similar failure rates.
120,000
MTBF Run Chart for a PCB Product
Categories 6 and 7 are used to address process, manufacturing
component failures only and design errors. If historical failure data are well maintained
100,000
comp, design. Software failures and the new product development is evolutionary, not
all failures revolutionary, HIRAP is a convenient and accurate tool to
80,000
predict the new product failure rate or MTBF in design phase.
MTBF (hours)

MTBF hits the target

60,000 in week 41?
Orman et al [1] developed a simulation-based reliability
prediction model for product conceptual design with unknown
40,000
component failure rates. Unlike HIRAP, the model is
20,000
implemented based on bottom-up approach and uses triangular
distributions to model parameters for components with
0 unknown failure rates. Like most conventional approach, the
1
3
5
7
9
11
13
15
17
19
21
23
25
27
29
31
33
35
37
39
41
43
45
47
49
51
53
55
57

model does not incorporate the non-component failures into

Time in week
the reliability estimate. Therefore, the simulation tends to
Figure 2 Actual MTBF vs. Predictions yield an optimistic system MTBF. Jin and Su [5] observed
that components of the same type operating on the same board
Both figures 1 and 2 have indicated that accurate usually have different temperature and electrical stressing due
reliability prediction is essential to customer satisfactions. It to different cooling efficiency and functional requirements.
also helps the product manufacturer to manage the engineering They proposed a stochastic reliability prediction model to
resources to implement corrective actions after the first estimate electrical device failure rates that explicitly
customer shipment. Optimistic MTBF prediction could incorporate the variations of temperature and electrical
potentially mislead design engineers to withdraw their stressing. The confidence intervals for the lower bound of
recourses early from reliability monitoring and correction PCB failure rate are also derived based on six-sigma criteria.
implementation process. The paper aims at the development Similarly to [1], their model does not address the failures
of a practical MTBF prediction method that effectively caused by non-component issues. This paper extends the
incorporates component, design, software, manufacturing and research work previously done in [5]. It aims to establish a
process failures. Armed with the accurate MTBF estimation practical estimation tool that incorporates uncertainties of
for the new product, product developers can wisely leverage component and non-component failures for new design.
the engineering resource to implement the corrective actions
and remove potential failure modes before the design team 3. NEW MTBF ESTIMATE
works on new projects.
The paper consists of five sections: the introduction A good reliability data tracking system is essential for the
section indicates the gap between the actual product MTBF accurate reliability prediction. To create and maintain an
and the estimation using conventional predictions. In section accurate reliability database, it usually requires dedicated
two a brief literature review will be presented to describe the personals to track and record the failure data reported from
current research on MTBF predictions. In Section 3 a bottom- repair centers. Due to the dynamic movement of employees in
up MTBF prediction method is proposed by incorporating today’s companies, not every company like Honeywell [2, 3]
both component and non-component failure rates. In section 4 has the capability to maintain a consistent database, on which
an application of the MTBF prediction using the new model is the HIRAP is developed, that systematically tracks and
demonstrated on the design of an analog wave generating records product failures due to design, manufacturing and field
board. The paper concludes with a summary that highlights services. In semiconductor testing equipment industry, boards
the importance of non-component failure impact on new PCB manufacturing and equipment repair activities continue to
product reliability estimation. move offshore driven by the competitive labor cost. This
could create extra challenges to collect repair data related to
2. LITERATURE REVIEW PCB components and non-components failures due to
communication or language barriers.
To close the gap between the actual product MTBF and The new MTBF estimate is developed based on

605
Authorized licensed use limited to: University of Pisa. Downloaded on January 15,2024 at 11:39:55 UTC from IEEE Xplore. Restrictions apply.
component and non-component failure rates. For components derating factor. The temperature factor for electronic
used in PCB, some have known failure rates as they can be component is usually modeled by Arrhenius equation as
derived from historical field data. Others may not have follows,
explicit failure rates because they are new and used for the Ea 1 1
( − )
first time in the design. Unlike component failures, failure π (T ) = π T = e k T0 T
(4)
rates of non-components are more project-specific. For where
instance, design problem is related to the design experience of Ea= activation energy (eV)
engineers and the complexity of the product. In k = Boltzman constant (8.62×10-5 eV/K)
manufacturing solder joint defects are correlated with number T0 = reference temperature (313K)
of solder joins on the PCB and the profile of the reflow T = operating ambient temperature (K)
temperature. The larger the number of solder joins on the Similarly, the electrical stressing or derating factor πE can
PCB, the more likely the poor solders could happen. In the be estimated by the following equation as,
following paragraphs, failure rates for components and non-
components are derived based on the availability of historical
π ( p ) = π E = e m ( p − p0 ) (5)
data, component suppliers’ inputs, temperature and stressing where,
variations, and root-cause categories. p = actual applied electrical stressing percentage
p0= reference stressing and equal to 50%
3.1 Components with Historical Data m = fitting parameter
Equation 5 is used to model the electrical factor given in
Let λ denotes the failure rate for a particular component Telcordia TR-332 [6]. The fitting value m ranges from 0.006
type used in the product. If historical failure data such as to 0.059. For details users can refer to [6]. It is very common
cumulative hours and the number of defects are available, the that same types of components are repeatedly used in one
failure rate can be directly estimated by board. For example, a 15V ceramic captor can be used in
total faults multiple places within a board. When the ambient
λ= (1)
cumulative hours temperatures of the components doe not vary more than a few
For example, assume a type of 5V relay is used in an degrees, a single temperature reading T or the average
existing PCB product, say PCB-A. Each PCB-A uses 20 temperature is adequate to be substituted into equation 4 for
relays and total of 150 PCBs are installed in the field. In the estimating πT. In reality, ambient temperatures may vary in
past one year, assuming seven boards have been returned from the range of 10 to 30 degrees among the same type of
field, fours boards were diagnosed as bad relays with one components [5]. Then the average temperature is not
defective relay on each board (other three boards failed in no- sufficient to be used to estimate πT. To quantify the
components). Also assuming that each board operates 8760 uncertainty of the component failure rate caused by
hours a year, then the failure rate for this type of relay denoted temperature variation, πT should be treated as a random value.
as λr can be estimated as The mean and the variance of πT can be estimated as follows
4 when the temperature distribution is known,
λr = +∞ Ea 1 1
8760 × 20 × 150 ( − )
E[π T ] = ∫e
k T0 x
f T ( x)dx (6)
=1.5×10-7 faults/hour (2) 0
For a new product under design, say PCB-B, there are 30 +∞ 2 Ea 1 1
( − )
pieces of 5V relays used in the board, then the failure rate for E[π T2 ] = ∫ e k T0 x
f T ( x)dx (7)
this type of relay can be treated as 1.5×10-7 faults/hour if the 0
operating conditions are similar to predecessor products. Var (π T ) = E[π T2 ] − (E[π T )
2
(8)
Typical lifetime for Semiconductor test equipment lasts only
five to seven years due to technology obsoleteness, thus the where f T ( x) = probability density function (pdf) of T.
component failure rate can be treated as time-independent
before they enter the wear-out phase. ASIC Temperature Distribution
14 0.08
3.2 Components without Historical Data 12 histogram 0.07
pdf 0.06
If components are new with few failure data available, 10
0.05
Quantity

equation 1 can not be directly used to estimate the new 8

0.04
pdf

component failure rate. Component suppliers usually provide 6

0.03
the nominal failure rate λ0 that represents the component 4 0.02
reliability when it operates at 40oC with 50% electrical
2 0.01
derating. In real applications operation conditions are often
0 0
harsher than the nominal condition. The actual component
<65 [65, 70)[70, 75)[75, 80)[80, 85)[85, 90) >90
failure rate λ can be modified by following equation,
λ = λ 0π T π E (3) Degree in Celsius

Where πT is temperature factor and πE is electrical Figure 3. ASIC Temperature Distribution

606
Authorized licensed use limited to: University of Pisa. Downloaded on January 15,2024 at 11:39:55 UTC from IEEE Xplore. Restrictions apply.
 Temperature distribution for components can be measured non - component faults
using thermometers by probing the wire close to the λ= (15)
cumulative hours
component surface. Boards need to be warmed up and operate
in normal conditions as if used in the field. Multiple readings Taking the same example in section 3.1, if three failing
are recommended for each component to remove the outlier boards returned from field are root-caused as cold solder joins,
data. Figure 3 plots the temperature profile for a type of then the manufacturing failure rate denoted as λm is
ASICs (application specific IC) measured from one PCB. 3
λm = =2.283×10-6 faults/hour (16)
This chart is originally presented in [5] and reproduced here. 8760 × 150
It shows that device temperature varies from 65 oC to 90 oC. By the same token, equation 15 can be used to compute
Components operating at above 80oC are expected to have other types of non-component failure rates such as design
much higher failure rates than those operating at the lower errors, software bugs, process issues. Different from
temperature. The temperature profile can be approximated as a component failures which duplicate the failure modes among
normal distribution with mean of 76.3oC and standard different product lines; non-component failures usually do not
deviation of 5.7 oC. have such high degrees of duplications. For example, most
Similarly, for the same type of components used in a design errors and software bugs may be unique to one product.
PCB, electrical derating p may not be the same for all these In other words, failures occurred on predecessor products may
components. It may vary from one component to another not resume in the new design. Therefore the useful dataset for
depending on the circuitry requirements and functionality. It non-component failures is scare when applied to the new
is therefore more reasonable to treat p as a random variable design analysis. The triangular distribution is often used as a
whose value is uniformly distributed within an interval. The subjective modeling of a population for which there is only
probability density function for p can be described as follows, limited failure data or no data available [7]. The triangular
 1 distribution for λ is represented with a to denote the smallest
 u≤x≤v possible value of the failure rate, b denotes the largest possible
f p ( x) =  u − v (9)
 0 otherwise value of the failure rate, and c as the mode to denote the most
likely value. For example, suppose the min is a, the max is b
Here u and v represent the minimum and the maximum
and the mode is c. The probability density function of the
derating percentages. The mean and the variance of πE can be
unknown failure rate λ is given in Figure 4.
estimated by
e m ( v − p0 ) − e m (u − p0 )
v
E[π E ] = ∫ e m ( x − p0 ) f P ( x)dx = (10) h
u
m( v − u )
g(λ)

e 2 m ( v − p0 ) − e 2 m ( u − p0 )
v
E[π E2 ] = ∫ e 2 m ( x − p0 ) f P ( x)dx = (11)
u
2 m (v − u )
Var[π E ] = E[π E2 ] − (E[π E ])
2
(12) λ
If the component derating profile is better to be fit by a c b
other distributions, fp(x) in equations (10) and (11) can be
adjusted accordingly. Component derating percentages can be Figure 4 Triangular Distribution for Failure Rate λ
obtained from design engineers after the product’s BOM is The mathematical expression for g(λ) in Figure 4 is
determined. Through the paper, the uniform distribution is represented as follows:
used because it requires no specific information for p while
 2
still captures the derating variations. Finally the mean and  (λ − a ) a ≤ λ ≤ c
variance of the component failure rate can be computed as  (c − a)(b − a )
E[λ ] = λ 0 E[π T ]E[π E ] (13)  2
g (λ ) =  (λ − b) c < λ ≤ b (17)
var(λ ) = λ20 ( E[π T2 ]E[π E2 ] − ( E[π T ]) 2 ( E[π E ]) 2 ) (14)  (c − b)(b − a )

 0 otherwise
3.3 Non-component Failure Rate Estimate 
Parameters for the triangular distribution can be derived
Non-component failure rates usually are more difficult to from the dataset which it is intended to describe or model.
estimate than component failure rates. Non-component failure Provided the dataset does not contain any anomalous points,
rates are determined by various factors, some of them are the minimum and maximum can be obtained by sorting the
human-related. For example, the design and software failure values in ascending order and selecting the first and last points
rates are correlated with the design experience and the as a and b. Then the mode can be estimated by
similarity to the predecessor product. Manufacturing failure
rate is more likely associated with the complexity of the board c=3 λ -b-a (18)
and production experiences etc. The definition of non- Here λ is the sample mean for the dataset or the average
component failure rate λ is similar to that of the component failure rate. For example, to estimate the manufacturing
failure rate, which is failure rate of new product PCB-D, we find three predecessor
products that are close to PCB-D in terms of component

607
Authorized licensed use limited to: University of Pisa. Downloaded on January 15,2024 at 11:39:55 UTC from IEEE Xplore. Restrictions apply.
usage, manufacturing process, and total defects opportunity Sigma criteria can be used to calculate the confidence levels
per board. Historical data shows that manufacturing failure for MTBF. Based on Six-Sigma criterion, the maximum
rates for these three products are: 1.2×10-6, 1.4×10-6 and 2.4 achievable MTBF with 99.7% confidence can be estimated by,
×10-6. Then the sample mean λ =(1.2×10-6+1.4×10-6 +2.3 ×10- 1
Pr{λ PCB ≤ } ≥ 99.7% (24)
6
)/3=1.6×10-6. The min a =1.2×10-6, the max b =2.4 ×10-6, and MTBF
the mode c =1.3×10-6. The mean and the variance of the Quite often the confidence levels of the board MTBF are
failure rate λ can be obtained as, desirable before the volume manufacturing. The lower bound
b of MTBF can be improved by either using lower failure rate
E[λ ] = ∫ xg ( x)dx (19) components, reducing the variability of failure rates
a
b (temperature and derating controls), or enhancing the design,
E[λ2 ] = ∫ x 2 g ( x)dx (20) software testing, manufacturing and process.
a

Similar to component failure rate estimation using

historical data as shown in section 3.1, non-component failure 4. ILLUSTRATIVE EXAMPLE
rate distribution is estimated from the failure data of
predecessor products that are similar in complexity and In this section, the new MTBF estimation method is
functionality to the new design. applied to the design of a new PCB product. The board
consists of 54 component types with total of 759 individual
3.4 PCB Failure Rate Estimate parts. Seven new component types with total of 119 parts are
The total failure rate for the PCB is the summation of the first used in the board. Since no historical failure data is
failure rates of components and the non-components. available, failure rates for these new components will be
k l extrapolated based on nominal failure rates and the actual
λ PCB = λ d + λ s + λ m + λ p + λ o + ∑ ni λi + ∑n λ i i (21) operating conditions. Table 1 lists operating parameters for
i =1 i = k +1
seven new components: temperatures and electrical stressing.
where The activation energy Ea and fitting parameter m can be
λd = failure rate of design errors obtained from component suppliers or found in technical
λs = failure rate of software bugs handbooks such as [6]. In industry FIT is often used instead
λm = failure rate of manufacturing of failure rate. FIT is defined as number of failures in 109
λp = failure rate of process hours. For the 1MB EPROM with FIT of 77.4, the failure rate
λo = failure rate of other issues is 77.4/109=7.74×10-8 faults per hour.
k= total number of new component types The other 47 component types have been used on
l= total number of component types in the PCB predecessor products. Therefore failure rates can be directly
λi = failure rates for component type i computed using equation 1. For confidential reasons, actual
ni= quantity for ith type of component used in the PCB names of component types are replaced by numerical labels as
As indicated in sections 3.1 and 3.2, failure rates for seen in column “Component Type” in Table 2.
existing component types and new component types are Similarly non-component parameters are estimated from
estimated differently. Failure rate λi (for i=1, 2, … k) predecessor products. Reliability engineers can review the
represents new component types; and λi (for i=k+1, k+2, … l) existing product lines and select board types that have similar
represent existing component types. complexity and functionality compared to the new design,
It is noticed that λd, λs, λm, λp, λo and λi (for i=1, 2, … k) from which the distribution parameters for design, software,
are stochastic values as described in sections 3.2 and 3.3, so manufacturing, and process are estimated. Table 3 presents
λPCB is also a random variable. Failure rates for existing the non-component failure rates for predecessor products from
component types are treated as deterministic values and which triangular distribution parameters (a, b, and c) for non-
therefore the variance for the failure rate is zero. Though the component failure rates are obtained and listed in the table.
exact distribution for λPCB is difficult to obtain, the mean and Based on parameters given in Tables 1-3, we are able to
the variance for λPCB can be estimated as follows, compute the failure rate and associated confidence levels for
the new PCB using equations 22-24. The detailed computation
µ = E[λ PCB ] = E[λ d ] + E[λ s ] + E[λm ] + E[λ p ]
is automated using Matlab codes. The mean and the variance
k l (22) of λPCB of the new PCB is obtained as µ=1.08×10-5 and
+ E[ λ o ] + ∑ n i E[ λ i ] + ∑n λ
i =1 i = k +1
i i
σ2=3.3×10-12 as given in Table 4 Substituting µ and σ2 into
σ 2 = var(λPCB ) = var(λd ) + var(λs ) + var(λm ) + var(λ p ) equation 24, the lower MTBF bound for the new design is
k (23) obtained as 63,000 hours with 99.7% confidence level.
+ var(λo ) + ∑ ni2 var(λi ) If the target MTBF is 65,000 hours, the MTBF of the new
i =1
design needs to be further improved by either selecting more
Equation 21 shows that λPCB is the summation of
reliable components or reducing the non-component failure
independent failure rate estimates for components and non-
rates. Assuming the design team decide to remove all issues
components. It is also easy to show λd, λs, λm, λp, λo and λi
within the root-cause bucket called “Others” by implementing
(i=1, 2, … k) are uniformly bounded, then based on central
design-for-test (DFT), this will ideally eliminate all “Others”
limit theorem λPCB tends to be normally distributed with mean
failures and makes E[λo]=0 and var(λo)=0. Re-calculation
µ and variance σ2. Because λPCB is normally distributed, Six-

608
Authorized licensed use limited to: University of Pisa. Downloaded on January 15,2024 at 11:39:55 UTC from IEEE Xplore. Restrictions apply.
gives E[λPCB]=1.01×10-5 and var(λPCB)=2.88×10-12. After we new products, reliability must be designed in early phase of
substitute the new mean and variance into equation 24 again, it prototype stage when the product is still in the concept
yields the lower MTBF bound as 67,600 hours which exceeds development stage. Both the component and non-component
the target by 2,600 hours. In practice, instead of attacking one risks need to be addressed and appropriately mitigated by
bucket issue, design engineers usually work with cross- identifying and eliminating high-risk items. This might create
functional teams resolve multiple issues covering the whole extra work load through design for reliability in pilot line and
failure spectrum. reliability testing, but the stake of the final reward is high
when we finally ramp-up and ship high reliability products to
5. CONCLUSIONS our customers. This means tremendous cost savings and
makes the new product to be more competitive than our
In semiconductor testing equipment industry, fast-to- competitors. Meanwhile, engineering staff can be transferred
market is essential for gaining the market share while high to new product design without the need of allocating extra
reliability keeps the product competitive from long-term resources to monitor and improve the reliability of existing
perspective. To maintain the momentum of competitiveness of products.

Table 1 Parameters for Computing Failure Rates of New Components

Input Data Output Data

i Component Name ni FIT Ea µT σT u v m µπ T

σπ T
µπ E
σπ E

1 1MB EPROM 14 37.4 1.4 43.7 1.7 60 80 0.024 1.90 0.53 1.63 1.04
2 0.2W 1% Resistor 32 3.33 1.2 47.6 8.3 50 70 0.019 5.08 7.11 1.22 0.53
3 FPGA 26 6.1 0.7 56.2 4.1 60 85 0.024 3.74 1.16 1.74 1.18
4 Switch Transistor 26 5.81 1.2 51.8 2.8 60 75 0.024 5.37 2.03 1.53 0.91
5 Op Amp 3 9.00 1.2 46.6 5.2 55 75 0.024 3.17 2.46 1.45 0.83
6 Clock Driver 14 0.823 1.2 65.1 1.9 65 80 0.024 27.86 6.51 1.73 1.13
7 Switch Diode 4 11.25 1.0 63.2 6.4 60 80 0.029 15.74 11.08 1.81 1.25

Table 2 Parameters for Existing Component Failure Rates

i Comt type ni FIT i Comp Type ni FIT i Comp Type ni FIT
8 comp 1 16 130 24 comp 17 1 3.643 40 comp 33 2 0.823
9 comp 2 10 5.23 25 comp 18 1 3.643 41 comp 34 1 1.029
10 comp 3 10 10.00 26 comp 19 1 3.643 42 comp 35 1 1.00
11 comp 4 10 5.23 27 comp 20 1 3.33 43 comp 36 1 1.00
12 comp 5 1 10.00 28 comp 21 4 0.823 44 comp 37 1 0.823
13 comp 6 9 1.029 29 comp 22 4 0.823 45 comp 38 1 0.823
14 comp 7 1 9.00 30 comp 23 4 0.823 46 comp 39 1 0.823
15 comp 8 9 0.823 31 comp 24 4 0.823 47 comp 40 1 0.823
16 comp 9 2 3.643 32 comp 25 3 0.823 48 comp 41 1 0.823
17 comp 10 2 3.643 33 comp 26 3 0.823 49 comp 42 1 0.823
18 comp 11 8 0.823 34 comp 27 2 1.029 50 comp 43 1 0.823
19 comp 12 8 0.823 35 comp 28 2 1.00 51 comp 44 252 0.00005
20 comp 13 6 1.029 36 comp 29 2 0.823 52 comp 45 24 0.0004
21 comp 14 7 0.823 37 comp 30 2 0.823 53 comp 46 177 0.00005
22 comp 15 6 0.823 38 comp 31 2 0.823 54 comp 47 26 0.0002
23 comp 16 6 0.823 39 comp 32 2 0.823

Table 3 Triangular Distribution Parameters for Non-Component Failure Rates

Input Data (FIT) Output Data (faults/hour)
PCB-A PCB-B PCB-C a b c
Design Errors 148.5 423.6 855.0 1.49E-7 8.55E-7 4.24E-7
Software Bug 29.7 145.3 95.0 2.97E-8 1.45E-7 9.50E-8
Manufacturing 207.9 242.1 285.0 2.08E-7 2.85E-7 2.42E-7
Process 118.8 302.6 665.0 1.19E-7 6.65E-7 3.03E-7
Others 475.3 641.5 807.5 4.75E-7 8.08E-7 6.42E-7

609
Authorized licensed use limited to: University of Pisa. Downloaded on January 15,2024 at 11:39:55 UTC from IEEE Xplore. Restrictions apply.
 Table 4 Mean and Variance of Failure Rates for Components, Non-Components and New PCB
Components Non-component New PCB
Root Case New Existing
Category Comps comps Design Software Mfg Process Others PCB
mean 6.52E-6 2.42E-6 4.76E-7 9.00E-8 2.45E-7 3.62E-7 6.41E-7 1.08E-5
variance 2.42E-12 0 2.47E-13 8.66E-15 6.07E-14 1.44E-13 4.16E-13 3.30E-12

ACKNOWLEDGEMNETS Rutgers University-New Brunswick. He holds a Ph.D. from

Industrial Engineering and MS in Electrical and Computer
The authors would like to thank Teradyne GCS engineers
Engineering. Both are from Rutgers University in 2001. His
for their providing reliability data to develop this paper.
recent publications appeared in IEEE Transaction on
Reliability and IIE Transaction. His research interests include
REFERENCES: reliability estimations and optimizations for microlectronics.

1. Stephen Ormon, Richard Cassady, Allen Greenwood, “A

Peng Wang, PhD
Simulation-Based Reliability Prediction Models for
United Technologies Research Center
Conceptual Design”, pp 433-426, RAMS 2001.
411 Silver Lane, MS129-57
2. Lou Gullo, “In-service Reliability Assessment and Top-
East Hartford, CT 06108 USA
down Approach Provides Alternative Reliability
Prediction Method”, The 1999 Annual Reliability and
[email protected]
Maintainability Symposium (RAMS) Proceedings.
3. Bruce Johnson, Lou Gullo, “Improvements in Reliability
Peng Wang is a Research Engineer at United Technologies
Assessment and Prediction Methods”, pp181-187, RAMS
Research Center. Before he earned his PhD in Industrial and
2000.
Systems Engineering from Rutgers University in 2002, he also
4. W. J. FABRYCKY, B. S. Blanchard, Life-Cycle Cost and
received his MS in Statistics and MS in Industrial Engineering
Economic Analysis, Prentice Hall, 1991.
from Rutgers. He received his BS in Mechanical Engineering
5. T Jin, P. Su, “Minimize System Reliability Variability
from Northwestern Polytechnic University, China. His
Based on Six-Sigma Criteria Considering Component
research interests are in the areas of reliability modeling based
Operational Uncertainties”, RAMS 2005.
on physics of failure, reliability prediction, repairable system
6. Bellcore, Reliability Prediction Procedure for Electronic
reliability analysis and trend test, and accelerated life testing.
Equipment, SR-332, Issue 5, Dec 1995.
7. http://www.brightonwebs.co.uk/distributions/triangular.as
Qiyu Huang, PhD
p
School of Microelectronics
Shanghai Jiaotong University
BIOGRPGIES
1954 HuaShan Road 15th Floor
Shanghai 200030, China
Tongdan Jin, PhD.
Teradyne Inc.
[email protected]
300 River Park Drive, SM:3001-14
North Reading, MA 01864 USA
Dr. Qiyu Huang joined the School of Microelectronics,
Shanghai Jiaotong University in China as an assistant
[email protected]
professor in 2004. Huang got his Ph.D. and M.S. from the
Department of Electrical and Computer Engineering,
Tongdan Jin currently works at Teradyne Inc. as a Reliability University of Virginia in 2000 and 2003, respectively. He
Engineer. His duties include reliability modeling, prediction, received his B.S. in Physics from Peking University, Beijing,
failure modes and root cause analysis, and corrective action China. His research interest includes microfabrication,
implementations for semiconductor testing systems. Prior to nanoscale structures and devices, and reliability related issues
joining Teradyne, he was a research and teaching assistant at

610
Authorized licensed use limited to: University of Pisa. Downloaded on January 15,2024 at 11:39:55 UTC from IEEE Xplore. Restrictions apply.