You are on page 1of 7

Tools for Analysis of Accelerated Life and

Degradation Test Data


Keywords: Accelerated life testing, physics of failure, probabilistic, accelerated degradation testing

Reuel Smith and Mohammad Modarres


University of Maryland
Center for Risk and Reliability
College Park, MD
smithrc@umd.edu
modarres@umd.edu

Abstract—This paper discusses educational tools, resources available software including R [4], MATLAB [5], WinBUGS
and experience in teaching physics-of-failure (PoF) modeling and [6], and OpenBUGS [7].
data analysis methods used in accelerated life testing (ALT) and
accelerated degradation testing (ADT). The concepts introduced The outline for this paper is as follows: Section II briefly
in this paper include several probabilistic and highly technical covers the topics that are part of an in-depth accelerated testing
approaches such as the PoF and mechanistic approaches to program curriculum as well as models and tools that have been
reliability prediction and assessments as well as advanced made available to analyze them followed by three case studies
Bayesian analysis methods used in ALT and ADT to analyze life and the analysis methods used for evaluation. Section III
data obtained through various methods of constant and variable concludes the discussion with final thoughts and more
stress testing efforts. Other resources include codes and information about available resources.
algorithms from the readily available computational tools R,
MATLAB, and WinBUGS for solving advanced model and life
II. OVERVIEW AND TOOLS FOR ALT AND ADT ANALYSES
estimation methods related to PoF, ALT and ADT analyses. This
paper provides an overview of these resources and several The educational concept, including the analytical methods,
example problems. models, and analysis tools, can be divided into three principle
parts. Part one introduces the rudimentary concepts of the
probabilistic physics of failure (or PPoF) such as the standard
I. INTRODUCTION mechanistic models of failure mechanisms such as the stress-
Risk and reliability tools and methods are always evolving strength models, the damage endurance models, and the
in terms of effectiveness, usefulness, and realism due to the performance-requirements models. All three of which are used
incorporation of data and information from real sources and of to define the nature of failure and/or degradation mechanisms.
modeling of failure phenomena [1]. Modeling of failure Examples of the models introduce mechanistically and
phenomena often relies on the physics of failure (PoF) probabilistically include fatigue-damage, wear-damage,
approach [1]. The PoF modeling uses the accelerated life corrosion-damage, and creep-damage.
testing (ALT) [2, 3] approach to assess parameters of the
Part two covers all kinds of accelerated testing data analysis
models by accelerating mechanistic damages such as fatigue
methods. The bulk of this part focuses on ALT and ADT as
and wear failures. A similar application that relates life to the
well as the concepts of life-stress modeling, approaches to
amount of mechanistic damage is accelerated degradation
maximum likelihood estimation (MLE) for life-stress model
testing (ADT). Reliability engineers should be cognizant and
parameter estimation, Bayesian approaches to ALT/ADT
recognize the importance of these practices and therefore need
analysis, and step-stress analysis [8] approach and models.
to include them in their reliability engineering traning and
Part two concludes with formal quantitative and best practices
education [1].
test planning procedures for ALT or ADT.
Understanding the subject areas of PoF, ALT, and ADT
Part three focuses on applications of the physics-based
begin with the definition of each concept which includes the
models in reliability engineering and prognosis and health
overall concepts, objectives, and framework in relation to the
management (PHM). Here engineering students are exposed to
probabilistic PoF and mechanistic failure models [1]. Further
various case studies from field examples of reliability
understanding of these methods is reinforced through data
engineering that blend the topics and concepts that were
analysis methodologies which aid in the evaluation of these
learned in the first two parts and help develop ideas for
models. Abundant examples of these methodologies [1] give
individual and group projects that is also part of the required
the trainees and students the necessary practice in applying and
learning process.
understanding the techniques that have been covered. In
addition, several algorithms and codes that have been Multiple examples that apply the theoretical topics are
developed for problem solving [1] make use of readily covered, including access to a collection of algorithms and
codes that can be used to understand how some of the

978-1-5090-1880-2/16/$31.00 ©2016 IEEE


presented examples can be solved computationally. On the
practical side, examples of test facilities for traditional ADT
and ALT as well as HALT and HASS are covered [9].
The following are three specific examples that demonstrate
the type of problems that would be discussed in accelerated
testing education and the tools used to solve them. They focus
on the three rudimentary topics in accelerated testing: MLE,
ALT, and ADT.

A. MLE ALT Example


Consider a set of complete failures data obtained from an
ALT temperature acceleration test listed in Table I,

TABLE I. TIME TO FAILURES OBSERVED IN ACCELERATED LIFE TEST


(ALT)

Number of Recorded Failure Times


Temperature (°K)
Units Tested (hours)

248, 456, 528, 731, 813, 965,


406 8
972, 1528

436 6 164, 176, 289, 319, 386, 459

466 6 92, 105, 155, 184, 219, 235

Fig. 1. Probability plotting of accelerated life tests at three stress levels


Assume that it has already been shown that the scatter in times
to failure data listed in Table I can be best modeled by the From the probability plot of the data in Fig. 1 as well as from
Weibull distribution (although the appropriateness of this the Weibull fit, the parameters for each stress level are found
assumption can also be tested), and that the most appropriate as,
life-stress relationship for this particular ALT data is the
Arrhenius model (1),  α = 899, β = 1.9 at T=406 °K
 α = 342, β = 2.5 at T=436 °K
Ea  α = 188, β = 2.7 at T=466 °K
 L  A exp  
KT
with a mean  of 2.46. Assuming the Weibull parameter α to
where L is the life of the product, A is a constant to be be the L of interest, the Arrhenius model parameters are
determined, T is the temperature (in Kelvin), K is obtained through a least-squares linear regression. The
Boltzmann’s constant (note that 1 / K = 11605 eV-1°K), and Arrhenius relationship is simplified to a linear expression by
Ea is the activation energy in eV (taken as an unknown taking the natural log of both sides of (1).
parameter). Assuming that the use-level temperature is 353 °K,
estimate life at this use level using the plotting method and the Ea 1
MLE and compare the results.  ln L  ln A   
K T
Plotting the accelerated data typically is done by
calculating the percent failure F t  by way of the Kimball which is shown in Fig.2,
plotting position,

i  0.375
 F t    
n  0.25

where n is the number of units tested at a certain stress level


and i is the unit that fails at time ti given a certain stress level.
The Kimball plotting position produces the following:
5201
 L  0.0024 exp  
T

and the MLE based use-life is L = 11,272 hours. The MLE


estimate is more formal and more accurate and the results
shown in Table II, underline the approximate nature of the
plotting methods.

TABLE II. SUMMARY OF PARAMETER AND USE LIFE RESULTS

Method β A Ea Use Life (hours)

Plotting 2.46 0.0043 0.43 5,404

MLE 2.69 0.0024 0.45 11,272

Fig. 2. Extrapolation for Arrehnius life-stress with Weibull time to failure The results obtained above are all obtained from downlodable
MATLAB and R routines discussed in [1].
Therefore, at the use-stress (353 °K), the parameters are  =
B. Bayesian ADT Example
0.0043 (hrs.), Eˆ a / K = 4959 (°K), and Êa = 0.43 (eV). Thus
Consider an accelerated degradation test of the relative
the life-stress relationship would be, luminosity (proportion of initial luminosity) for LEDs [10].
Suppose the standard (“use”) operating temperature for LEDs
is 20 °C. The LED failure is defined as occurring when the
4959
 L  0.0043 exp   LED relative luminosity (i.e., performance) drops to 0.5 or
T lower, i.e., 50% of initial luminosity. The accelerated
degradation test that involves testing 25 units each at 25 °C, 65
and the “use-life” is L = 5,404 hours by using the “use” °C, and 105 °C, and the data is completely documented in the
temperature. original reference [10], but presented graphically in Fig. 3.
Compare these results to those obtained by using MLE where
the log-likelihood  for a Weibull-Arrhenius model (with all
failed data) is written as,

  
 1   
 1  
      
  
M
      

ti ti
N i ln    exp    
i 1  A exp E a  A exp a
E
   A exp a
E
 
 KTi  KT    KTi  
 i
 
 

where M is the total number of failures from all stress levels


and N i is the total number of failures at time t i . The most
direct MLE evaluation is to take the derivative of the
likelihood with respect to the parameters being evaluated and
set them equal to zero.
Fig. 3. LED luminosity data for three stress levels [10]

   The LED degradation data model follows the form, where


  0;  0; 0  an expression for the true degradation of luminosity at time t
 A E a and temperature T (in degrees Kelvin) is:
Solving all three equations (6) simultaneously using one of
1
the tools provided produces ˆ = 2.69, Â = 0.0024, Eˆ a / K = 
2
   1 1  
5,201 (°K), and Êa = 0.45 (eV). Thus the MLE based life-  Yijk  1   1 t ijk exp   3 11605      ijk 

   u
T T 
i  
stress relationship is given as,
where t ijk is the k-th time of the j-th LED unit at the i-th
temperature level.  ijk is the normally distributed (0,σε2),
model error (Note that 1 and  2 are positive.) Tu is the use
temperature of 293.15 °K. Assume that the true degradation
follows the Arrhenius relationship. Regarding the prior
distributions for the parameters 1 ,  2 ,  3 ,and   we use
the non-informative prior in forms of the uniform distribution
UNIF(0,100). The objective now would be to estimate the
posterior mean and the upper and lower values of the
degradation model parameters including the MTTF estimate of
the LED at standard operating temperature.
Bayesian parameter estimation is done by way of a
MATLAB [5] routine developed as part of the educational
resources provided [1], which performs Markov-Chain Monte
Carlo (MCMC) simulation on the Bayesian Equation (9).

  
 
l D | π0 
 

 π1  | D     Fig. 4. The extrapolated MTTF of each LED unit at 50% luminosity on a

   

  D |  π 0  d
life-stress relationship plot

and the MTTF at use-stress is calculated to be 121,975 hours.


 

where the posterior 1  | D is a function of the likelihood [1].
   
 
l D |  and prior  0  , D is the known data, and  is a C. Step-Stress Example[11]
vector of the set of parameters being updated. From the Consider a step-stress test of cable insulation, where a step-
parameter estimates, the posterior parameter distributions for stress test can conceptually be illustrated in Fig. 5.
the entire data is obtained and given in Table III.

TABLE III. POSTERIOR DEGRADATION PARAMETER DISTRIBUTIONS

Posterior Mean
Lower Bound Upper Bound
Parameter

β1 9.7 x 10-6 0.027 0.062

β2 0.29 0.61 1.0


Fig. 5. A conceptual step-stress test
β3 0.0047 0.57 1.3
Given a known nominal stress S 0 , a number n of units are
σε 0.0071 0.091 0.20 initially tested at a low stress level S1 until the stress changing
point t i when the stress is raised to S 2 . Observed failure
times are denoted as t *j and censored units are denoted as
The MTTF estimate at operating temperature is obtained from
the Bayesian posterior for each LED unit, where the following t *j  . The test continues until a predetermined censoring time
times-to-failure can be obtained given that failure occurs at
50% initial luminosity. These are shown in the following t n at S n . The example step-stress test was run to estimate life
stress-life relationship on Fig. 4. at a design stress of 400 volts/mm. The data used are listed in
Table IV.
TABLE IV. STEP-STRESS TEST DATA ON CABLE 1 (* BEFORE STEP 1, TABLE V. PRIOR STEP-STRESS TEST DATA ON CABLE 2 (* BEFORE STEP
EACH SPECIMEN WAS HELD 10 MINUTES EACH AT 5, 10, 15, AND 20 KV; + 1, EACH SPECIMEN WAS HELD 10 MINUTES EACH AT 5, 10, 15, AND 20 KV; +
DENOTES A RUN TIME WITHOUT FAILURE) DENOTES A RUN TIME WITHOUT FAILURE)

Voltage Pattern Specimen Data Voltage Pattern Specimen Data


Total Total
Final Time to Thickness Final Time to Thickness
Step* Kilovolts Hold (min) Step* Kilovolts Hold (min)
Step Failure (mm) Step Failure (mm)
(min) (min)
1 26 15 5 102 27 1 26 15 5 114 27
15 5 113 27 15 6 119 27
2 28.5 15 5 113 27 2 28.5 15 6 116 27
15 6 117 27
3 31 60 6 370+ 29.5 3 31
60 6 345 29.5 60 7 423 29.5
4 33.4 60 6 345+ 28 4 33.4 60 7 423+ 28

5 36 240 6 1249 29 5 36 240 6 1401.5 29


240 6 1333 29 240 6 1261.7 29
6 38.5 240 6 1333+ 29 6 38.5 240 6 1419.4 29
240 5 1096.9 29 240 6 1364.2 29
7 41 240 6 1250.8 30 7 41
240 5 1097.9 29 960 2 1398.1 30
8 43.5 8 43.5 960 3 2691.7 30
960 3 2460.9 30 960 3 2000 30
9 46 960 3 2460.9+ 30 9 46 960 3 2000+ 30
960 3 2700.4 30 960 3 2149.6 30
10 48.5 960 4 2923.9 30 10 48.5 960 2 1429.7 30
960 2 1160 30 960 2 1502.7 30
960 3 1962.9 30 960 2 1440.9 30
960 1 363.9+ 30 960 4 3349.1 30
960 1 898.4+ 30 960 4 3349.1+ 30
960 5 4142.1 30

For this example the most appropriate life-stress


Each specimen was first held for 10 minutes each at 5kV, relationship for this ALT is the inverse power law (IPL) shown
10kV, 15kV, and 20kV before it went into step 1 at 26kV. In in Equation (10),
steps 1 through 10, a specimen has the same hold time at each
voltage (15 minutes, 1 hour, 4 hours, or 16 hours). Table IV
1
shows the step number and the total time on test when a  L  
specimen failed and its insulation thickness (used to calculate AVi p
its stress as the voltage divided by the thickness). Consider the
case where the tested cable 1 was a newer version of an older Where V is the stress given in volts/mm and A and p are
cable (i.e., cable 2) tested before. In this case we can use the
MLE result of the test data from cable 2 as the prior unknown parameters. Like the previous example, Bayesian
information in a Bayesian updating process. Assume that the estimation is performed on (9) by way of the same MATLAB
cable 2 test data are the ones shown in Table V. routine provided in [1] where the likelihood is defined as,

   1  F t 
n cij rik
 l f i tc*i1 , tc*i 2 ,, tc*ij * * *
i ri1 , t ri 2 , , t rik
. i 1 j 1 k 1
where, n is number of steps; cij is number of failure data TABLE VII. POSTERIOR DISTRIBUTION RESULTS OF CABLE 2

points at step i ; t c*ij is time of j-th failure data point at step i ;


Parameter Lower Bound Upper Bound
rik is number of right censored data points at step i ; is t r*ik
time of k-th right censored data point at step i . Assuming that A 3.70 x 10-9 5.68 x 10-5
the failure time is subject to a Weibull distribution at each
stress level, Fi and f i respectively stand for the corresponding p 1.12 x 10-3 1.82
cdf (12) and pdf (13) for the failure data points at step Si,
β 4.68 x 10-3 12.10

  
Fi t c*ij , Vi  1  exp  
 t cij  t i 1   i 1 AVi

*
   
p 

Again the results obtained above are all obtained from
downloadable MATLAB or OpenBUGS routines provided in
[1].


f i tc*ij ,Vi   AV t
i
p *
cij  
 ti 1   i 1 AVi p
 1
exp    
p 
 tcij  ti 1   i 1 AVi 

*

   III. CLOSING REMARKS
A brief overview of an accelerated testing program and its
and  i 1 stands for the equivalent time at step i-1 depicted by, publically free resources have been presented and three
examples were provided to underline the breadths of these
educational / training resources. It was shown that a broad set
n
 Vi 1  of accelerated testing theory, methods, models, tools, and
  i 1  t i 1  t i  2   i  2   
  applications would be necessary in order to facilitate the
 Vi  understanding of accelerated testing in the engineering
professioni.
Thus in this case there are three parameters that should be
updated: A , p , and  . Bayesian estimation is initially REFERENCES
performed on the prior data in Table V assuming non-
informative uniform priors for the parameters UNIF(0,100). [1] M. Modarres, M. Amiri and C. Jackson, Probabilistic Physics of Failure
This computes the prior parameter distributions (also uniformly Approach to Reliability: Modeling, Accelerated Testing, Prognosis and
distributed) given in Table VI. Reliability Assessment, College Park, MD: Center for Risk and
Reliability A.J. Clark School of Engineering, 2015. (PDF version is
available for download from http://crr.umd.edu/node/156)
TABLE VI. PRIOR DISTRIBUTION RESULTS OF CABLE 2 [2] W. Nelson, Accelerated Testing - Statistical Models, Test Plans, and
Data Analysis, Hoboken, NJ: John Wiley and Sons, 1990.
Parameter Lower Bound Upper Bound [3] W. Nelson, Accelerated Testing - Statistical Models, Test Plans, and
Data Analysis, New York: John Wiley and Sons, 2004.
[4] R. C. Team, R-Code v3.3.0, 2016.
A 4.38 x 10-10 5.69 x 10-5 [5] MathWorks, MATLAB R2016a, 2016.
[6] T. B. Project, WinBUGS v1.4.3, 2007.
p 1.37 x 10-4 2.78 [7] O. Foundation, OpenBUGS v3.2.3, 2014.
[8] W. Zhao, A. Mettas, X. Zhao, P. Vassiliou and E. A. Elsayed,
"Generalized step stress accelerated life model [LED case study],"
β 1.74 x 10-4 12.10 Business of Electronic Product Reliability and Liability, 2004
International Conference on, pp. 19-25, 2004.
[9] W. Staszewski, C. Boller and G. R. Tomlinson, Health Monitoring of
Aerospace Structures: Smart Sensor Technologies and Signal
The prior information in Table VI is then used for the next Processing, West Sussex, UK: John Wiley & Sons, 2004.
phase which obtains the posterior for the parameters using the [10] M. S. Hamada, A. Wilson, S. Reese and H. Martz, "Chapter 8: Using
data from Table IV as the evidence (new data). The results for Degradation Data to Assess Reliability," in Bayesian Reliability, New
the posterior for the parameters are as follows. York, Springer Science+Business Media, 2008, pp. 271-317.
[11] W. Nelson, "Accelerated Life Testing Step-Stress Models and Data
Analyses," IEEE Transactions of Reliability, Vols. R-29, no. 2, pp. 103 -
108, 1980.
BIOGRAPHY School of Engineering, University of Maryland, College Park. His
Author 1 Reuel Smith is a PhD level Reliability Engineering graduate research areas are probabilistic risk assessment and management,
student at the A.J. Clark School of Engineering at the University of uncertainty analysis and physics of failure degradation modeling.
Maryland College Park. His current research is in the area of fatigue Professor Modarres has over 350 papers in archival journals and
crack propagation, detection, and modeling. He has an M.S. degree in proceedings of conferences, including several books in various areas of
both Aerospace Engineering and Mechanical Engineering from the risk and reliability engineering. He is a University of Maryland
University of Maryland College Park. Distinguished Scholar. Professor Modarres received his M.S. and PhD
in Nuclear Engineering from MIT and M.S. in Mechanical Engineering
also from MIT.
Author 2 Mohammad Modarres is the Nicole Y. Kim Eminent Professor of
Engineering and Director Center for Risk and Reliability, A.J. Clark

i
Example of an accelerated testing program and toolset is publically available for download at www.crr.umd.edu [1]

You might also like