Accelerated Testing Background

Accelerated Testing Obtaining Reliability Information Quickly

Todays manufacturers face strong pressure to: Improve productivity, product eld reliability, and overall quality using new technology. Develop newer, higher technology products in record time. Implies increased need for up-front testing of materials, components and systems. Accelerated tests provide timely information for product design and development. Users must be aware of potential pitfalls

William Q. Meeker Department of Statistics and Center for Nondestructive Evaluation Iowa State University Ames, IA 50011

Overview What is Reliability? Dierent kinds of accelerated tests R(t) = 1 F (t) The probability that a system, vehicle, machine, device, and so on, will perform its intended function under encountered operating conditions, for a specied period of time. Quality over time A powerful marketing tool An engineering discipline requiring support from Physics and chemistry Statistics Example 1Evaluation of an insulating structure Example 2New-technology microelectronic logic device Accelerated Degradation Tests Importance of physics of failure and physical/chemical models (and sensitivity analysis) Example 3Microelectronic RF amplier device Connecting with the eld Example 4Appliance eld reliability Areas for further research

Some Applications of Accelerated Tests

Breakdown Times in Minutes of a Mylar-Polyurethane Insulating Structure (from Kalkanis and Rosso 1989)

Assess component or material reliability or durability. Make design decisions to improve reliability or lower cost Verify predictions produced with physical models (e.g. FEM)
Minutes

10

10

System test to simulate eld-use at accelerated conditions. Predict product eld performance. Identify and x potential failure modes at system/subsystem level (HALT and STRIFE tests). Screening (100% or audit) testing of manufactured product (e.g. ESS and burn-in).

10

10

10

10

-1

100

150

200 kV/mm

250

300

350 400

Inverse Power Relationship-Lognormal Model

Plot of Inverse Power Relationship-Lognormal Model Fitted to the Mylar-Polyurethane Data (also Showing 361.4kV Data Omitted from the ML Estimation)

The inverse power relationship-lognormal model is Pr[T t; volt] = nor where

10
3

log(t)

10 10

= 0 + 1x, and x = log(Voltage Stress). assumed to be constant.

Minutes

10 10 10 10

90% 50% 10%

-1

50

100 kV/mm

200

500

Lognormal Probability Plot of the Inverse Power Relationship-Lognormal Model Fitted to the Mylar-Polyurethane Data

Methods of Acceleration Three fundamentally dierent methods of accelerating a reliability test:

.99 .98 .95 .9 .8

Proportion Failing

Increase the use-rate of the product (e.g., test a toaster 400 times/day). Higher use rate reduces test time. Use elevated temperature or humidity to increase rate of failure-causing chemical/physical process. Increase stress (e.g., voltage or pressure) to make degrading units fail more quickly. Use a physical/chemical (preferable) or empirical model relating degradation or lifetime at use conditions.
10
5

.7 .6 .5 .4 .3 .2 .1 .05 .02 .01 10

0

219.0
1

157.1

122.4
2

100.3
3

50 kV/mm
4

10

10

10

10

Minutes

10

Interval ALT Data for a New-Technology IC Device Tests run at 150, 175, 200, 250, and 300C.

New-Technology Integrated Circuit Device ALT Data

10

Developers interested in estimating activation energy of the suspected failure mode and the long-life reliability.
Hours

10

10

Failures had been found only at the two higher temperatures. After early failures at 250 and 300C, there was some concern that no failures would be observed at 175C before decision time. Thus the 200C test was started later than the others.

10

10

x x x

x x x

10

100

150

200 Degrees C

250

300

350

11

12

The Arrhenius-Lognormal Regression Model Elevated Temperature Acceleration of Chemical Reaction Rates The Arrhenius model Reaction Rate, R(temp), is
Ea Ea 11605 = 0 exp kB (temp C + 273.15) temp K where temp K = temp C+273.15 is temperature in Kelvin and R(temp) = 0 exp

The Arrhenius-lognormal regression model is Pr[T t; temp] = nor where log(t)

kB = 1/11605 is Boltzmanns constant in units of electron volts per K. The reaction activation energy, Ea, and 0 are characteristics of the product or material being tested. The reaction rate Acceleration Factor is
AF(temp, tempU , Ea ) = R(temp) R(tempU ) 11605 11605 tempU K temp K

= 0 + 1x, x = 11605/(temp K) = 11605/(temp C + 273.15) and 1 = Ea is the activation energy is constant

= exp Ea

When temp > tempU , AF (temp, tempU , Ea) > 1.

13

14

Arrhenius Plot Showing ALT Data and the Arrhenius-Lognormal Model ML Estimation Results for the New-Technology IC Device.

Lognormal Probability Plot Showing the Arrhenius-Lognormal Model ML Estimation Results for the New-Technology IC Device

.95

10

.9 .8

10

.7 .6 .5 .4 .3 .2 .1 .05 .02

Hours

10

10

Proportion Failing

10

x x x

x x x

.01 .005 .002 50% 10% 1% .0005 .0001 300 Deg C 10

2

10

250 10
3

200 10
4

175

150 10
5

100 10
6

100

150

200

250

300

350

10

Degrees C on Arrhenius scale

15

Hours

16

Lognormal Probability Plot Showing the Arrhenius-Lognormal Model ML Estimation Results for the New-Technology IC Device with Given Ea = .8 Pitfall 4: Masked Failure Mode
.95 .9 .8 .7 .6 .5 .4 .3 .2 .1 .05 .02 .01 .005 .002 .0005 .0001 10 300 Deg C
2

Accelerated test may focus on one known failure mode, masking another!

Proportion Failing

Masked failure modes may be the rst one to show up in the eld.

Masked failure modes could dominate in the eld.

250 10
3

200 10
4

175

150 10
5

100 10
6

10

Hours

17

18

Possible results for a typical temperature-accelerated failure mode on an IC device

Unmasked Failure Mode with Lower Activation Energy

10

10

10

10

Hours

Hours

10

10

4
Mode 2

10

10

10%

10

2
10%

10

Mode 1 10%

10

10 40 60 80 Degrees C 100 120 140

40

60

80 Degrees C

100

120

140

19

20

Percent Increase in Resistance Over Time for Carbon-Film Resistors (Shiomi and Yanagisawa 1979)

Advantages of Using Degradation Data Instead of Time-to-Failure Data

Degradation is natural response for some tests.

10.0 173 Degrees C

5.0

Percent Increase

Can be more informative than time-to-failure data. (Reduction to failure-time data loses information)

133 Degrees C 1.0

Useful reliability inferences even with 0 failures.

83 Degrees C

0.5

More justication and credibility for extrapolation. (Modeling closer to physics-of-failure)

0 2000 4000 6000 8000 10000

Hours
21 22

Limitations of Degradation Data

Percent Increase in Operating Current for GaAs Lasers Tested at 80C

Degradation data may be dicult or impossible to obtain (e.g., destructive measurements). Obtaining degradation data may have an eect on future product degradation (e.g., taking apart a motor to measure wear). Substantial measurement error can diminish the information in degradation data. Analyses more complicated; requires statistical methods not yet widely available. (Modern computing capabilities should help here) Degradation level may not correlate well with failure.
Percent Increase in Operating Current

0
0

10

15

1000

2000 Hours

3000

4000

23

24

Device-B Power Drop Accelerated Degradation Test Results at 150C, 195C, and 237C (Use conditions 80C)

Arrhenius Model Temperature Eect on Chemical Degradation

A1 k1 -A2
and the rate equations for this reaction are

0.0

-0.2

Power drop in dB

-0.4

-0.6

-0.8

-1.0

-1.2

-1.4

150 Degrees C 195 Degrees C 237 Degrees C 1000 2000 3000

dA1 = k1A1 dt Solving these gives

and

dA2 = k1A1, dt

k1 > 0.

(1)

A1(t) = A1(0) exp(k1t) A2(t) = A2(0) + A1(0)[1 exp(k1t)]

where A1(0) and A2(0) are initial conditions. The Arrhenius model describing the eect that temperature has on the rate of a simple rst-order chemical reaction is
4000

k1 = 0 exp

Ea kB (temp + 273.15)
26

Hours
25

Lognormal-Arrhenius Model Fit to the Device-B Time-to-Failure Data with Degradation Model Estimates

What Do Accelerated Test Results Tell Us About Field Reliability? Need information on:

.99 .98 .95 .9 .8

Eects of acceleration (e.g., cycling rate). Distribution of use-rates in actual use. Distribution of environmental conditions (e.g., stress spectra distributions).
237 Degrees C 195 150 Degrees C 80 Degrees C

Proportion Failing

.7 .6 .5 .4 .3 .2 .1 .05 .02 .01 .005 .001 10^1 10^2 10^3 10^4 10^5

These factors may be given or, in some situations, inferred from the available data.

Hours

27

28

Establish a Transfer Function Relating Laboratory Tests and Field Performance

Component-A Laboratory Test Cycles to Failure

Carefully compare laboratory tests results and eld failures. Same failure mechanisms operating in laboratory tests? Same factors (environmental noises) exciting the failure mechanisms? Identify laboratory/eld discrepancies to improve test procedures. Seek understanding of reasons for lack of agreement. Find a model (transfer function) to relate laboratory test to eld use. Understanding the relationship between the laboratory test results and product eld reliability will provide stronger basis for using future laboratory tests to predict eld performance.

10000

20000 Cycles

30000

40000

50000

29

30

Appliance Use-Rate Distribution (discretized lognormal distribution)

Example Use-Rate Model

Life of a component in cycles of use, has a distribution FC (c) = P (C c) =

0.15

log(c)

Actual use-rate has a distribution given by the proportion of users i (i = 1, . . . , k) that use the appliance at constant rate Ri, where k i = 1. i=1 Then the failure probability as a function of time is FT (t; ) = P (T t) =
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 Relative Frequency of Appliance Uses per Week

0.05

0.10

0.0

i
i=1

log (t) i

where = (1, . . . , k , ) and i = log(Ri).

31

32

Predicted Field Reliability of Component-A as a Weighted Average of Weibull Distributions

Predicted Field Reliability of Component-A as a Weighted Average of Lognormal Distributions

.999 .98 .9 .7 .5 .3 .2

.95 .9 .8 .7 .6 .5

Fraction Failing

.1

Probability

.4 .3 .2 .1 .05 .02 .01 .005 .002 .001

.05 .02 .01 .005 .003 .001

.0005 .0002 .0001 50 100 200 500 1000 2000 5000

50

100

200

500 Weeks of Service

1000

2000

5000

Weeks of Service

Mon Apr 10 14:01:32 CDT 2000

Fri Mar 23 21:27:13 CST 2001

33

34

Fitted Use-Rate Model for the Wear Failure Mode

Lab: subset AccWear Appliance B Wear Failure Mode ALT data Field: subset Field Appliance B Wear Failure Mode
.95 .9 .8 .7 .6 .5 .4 .3 .2 .1 .05 .02 .01 .005 .002 .0005 .00005 .00001 2 5 10 20 50 100 200 500 1000 2000 5000 Laboratory Field

Field Variability Lognormal f (r; R, R ) Density for the Wear Failure Mode (unloaded cycles relative to eld days of use)

Lab: subset AccWear Appliance B Wear Failure Mode ALT data Field: subset Field Appliance B Wear Failure Mode

Fraction Failing

Lab time: Test Cycle

Field time: Weeks

Thu May 10 22:23:56 CDT 2001

0.01

0.05

0.20

1.00

5.00

20.00

Thu May 10 22:24:23 CDT 2001

Test Cycles per Week

35 36

Planning Accelerated Tests

Simulation of a Proposed Accelerated Life Test Plan

Most basic ideas of traditional DOE still hold

10
5

Temp= 78,98,120 n= 155,60,84 centime= 183,183,183 parameters= -16.7330, 0.7265, 0.6000

Log time quantiles at 50 Degrees C Average( 0.1 quantile)= 8.014 SD( 0.1 quantile)= 0.4632 Average( 0.5 quantile)= 9.138 SD( 0.5 quantile)= 0.5116 Average(Ea)= 0.7266 SD(Ea)= 0.08594

Limit, as much as possible the amount of extrapolation used.

10
4

In censored accelerated life tests (failure time is response) allocate more test units to low acceleration factor level than high acceleration factor levels.

10

Days

10

Consider including some tests at the use conditions. Use simulation to investigate properties of alternative ALT plans.

10

1
10% Results based on 500 simulations Lines shown for 50 simulations

10

40

60

80

100

120

140

160

Degrees C

37

38

Areas for Further Research

Concluding Remarks

Physical/statistical models for failure acceleration Methods for sensitivity analysis when empirical models must be used Prediction of service life in complicated environments Physical/statistical models the eld environment Bayesian methods for analysis and planning (especially adaptive test plans) Accelerated degradation test planning Degradation analysis and planning with coarse (e.g. ordered categorical and censored) data. Physical comparison of lab and led failures to validate testing methods
39

Accelerated Testing can be valuable tool when used carefully There is no magic in Accelerated Testing Cross-disciplinary teams are needed to deal eectively with all issues Product/reliability/design engineers to identify productuse proles, environmental considerations, potential failure modes or weaknesses that need to be evaluated, etc. Experts in materials and the chemistry/physics of failure to help in the understanding of an suggest/develop appropriate models for acceleration of particular failure modes. Statisticians to help with stochastic modeling, plan tests, t models, and to help quantify uncertainty in results. Users of Accelerated Testing must beware of pitfalls
40

References References D. Byrne, J. Quinlan, Robust function for attaining high reliability at low cost, 1993 Proceedings Annual Reliability and Maintainability Symposium, 1993, pp 183-191. L. W. Condra, Reliability Improvement with Design of Experiments, 1993, New York: Marcel Dekker, Inc. M. Hamada, Using statistically designed experiments to improve reliability and to achieve robust reliability, IEEE Transactions on Reliability R-44, 1995 June. M. Hamada, Analysis of experiments for reliability improvement and robust reliability, in Recent Advances in Life-Testing and Reliability, 1995, N. Balakrishnan, editor, Boca Raton: CRC Press. Meeker, W.Q. and Hamada, M. (1995), Statistical Tools for the Rapid Development & Evaluation of High-Reliability Products, IEEE Transactions on Reliability R-44, 187-198.
41

Meeker, W.Q. and Escobar, L.A. (1998a), Statistical Methods for Reliability Data. John Wiley and Sons, Inc. Meeker, W.Q. and Escobar, L.A. (1998b), Pitfalls of Accelerated Testing. , IEEE Transactions on Reliability R-47, 114-118. W. Nelson, Accelerated Testing: Statistical Models, Test Plans, and Data Analyses, 1990, New York: John Wiley & Sons, Inc. G. Taguchi, System of Experimental Design, 1987; White Plains, NY: Unipub/Kraus International Publications. T. S. Tseng, M. Hamada, C. H. Chiao, (1995), Using degradation data from a factorial experiment to improve uorescent lamp reliability, Journal of Quality Technology, 363-369.

42