You are on page 1of 7

Accelerated Testing Background

Accelerated Testing Obtaining Reliability Information Quickly

• Today’s manufacturers face strong pressure to: Improve productivity, product field 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

1

2

Overview What is Reliability? • Different 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 specified period of time. • Quality over time • A powerful marketing tool • An engineering discipline requiring support from Physics and chemistry Statistics • Example 1—Evaluation of an insulating structure • Example 2—New-technology microelectronic logic device • Accelerated Degradation Tests • Importance of physics of failure and physical/chemical models (and sensitivity analysis) • Example 3—Microelectronic RF amplifier device • Connecting with the field • Example 4—Appliance field reliability • Areas for further research

3

4

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

4

10

3

• • • • • •

• System test to simulate field-use at accelerated conditions. • Predict product field performance. • Identify and fix 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

2

• • • • • • •

• • • • • •

• • • • • • • • • • • • •

10

1

10

0

10

-1

100

150

200 kV/mm

250

300

350 400

5

6

0 1 157. Higher use rate reduces test time. there was some concern that no failures would be observed at 175◦C before decision time.7 . and 300◦C.2 . • σ assumed to be constant. • Thus the 200◦C test was started later than the others.4 .3 3 50 kV/mm 4 10 10 10 10 Minutes 9 10 Interval ALT Data for a New-Technology IC Device • Tests run at 150. Minutes 10 10 10 10 2 • • • • • • • • • • • • • • • • • • • • • • • • • • • 1 90% 50% 10% 0 -1 50 100 kV/mm 200 500 7 8 Lognormal Probability Plot of the Inverse Power Relationship-Lognormal Model Fitted to the Mylar-Polyurethane Data Methods of Acceleration Three fundamentally different methods of accelerating a reliability test: . Use a physical/chemical (preferable) or empirical model relating degradation or lifetime at use conditions.1 122. • Increase stress (e. New-Technology Integrated Circuit Device ALT Data 10 7 • Developers interested in estimating activation energy of the suspected failure mode and the long-life reliability. 200.4 2 100. • After early failures at 250 and 300◦C.9 .g. volt] = Φnor where 10 3 log(t) − µ σ 10 10 5 4 • • • • • • • µ = β0 + β1x.95 . Hours 10 6 10 5 • Failures had been found only at the two higher temperatures.98 . 10 5 .1 .02 .Inverse Power Relationship-Lognormal Model Plot of Inverse Power Relationship-Lognormal Model Fitted to the Mylar-Polyurethane Data (also Showing 361. test a toaster 400 times/day). 10 4 10 3 x x x x x x 10 2 100 150 200 Degrees C 250 300 350 11 12 .. 175.g.5 .4kV Data Omitted from the ML Estimation) The inverse power relationship-lognormal model is Pr[T ≤ t.. • Use elevated temperature or humidity to increase rate of failure-causing chemical/physical process.99 .01 10 0 219.3 . voltage or pressure) to make degrading units fail more quickly.8 Proportion Failing • Increase the use-rate of the product (e.05 . 250.6 . and • x = log(Voltage Stress).

15) temp K where temp K = temp ◦C+273.6 . masking another! Proportion Failing • Masked failure modes may be the first one to show up in the field.005 . and γ0 are characteristics of the product or material being tested. R(temp). Ea ) = R(temp) R(tempU ) 11605 11605 − tempU K temp K • µ = β0 + β1x. The reaction activation energy.2 .002 50% 10% 1% .02 . 13 14 Arrhenius Plot Showing ALT Data and the Arrhenius-Lognormal Model ML Estimation Results for the New-Technology IC Device. Ea.95 .05 .The Arrhenius-Lognormal Regression Model Elevated Temperature Acceleration of Chemical Reaction Rates • The Arrhenius model Reaction Rate.02 Hours 10 5 10 4 Proportion Failing 10 3 x x x x x x .1 . temp] = Φnor where log(t) − µ σ kB = 1/11605 is Boltzmann’s constant in units of electron volts per K.3 .01 .4 .15 is temperature in Kelvin and R(temp) = γ0 exp The Arrhenius-lognormal regression model is Pr[T ≤ t.6 .9 .9 . • x = 11605/(temp K) = 11605/(temp ◦C + 273.7 . is −Ea −Ea × 11605 = γ0 exp kB (temp ◦C + 273.7 .95 10 7 .002 .2 . Lognormal Probability Plot Showing the Arrhenius-Lognormal Model ML Estimation Results for the New-Technology IC Device .0001 10 300 Deg C 2 • Accelerated test may focus on one known failure mode. 250 10 3 200 10 4 175 150 10 5 100 10 6 10 7 Hours 17 18 .8 Pitfall 4: Masked Failure Mode .0005 .05 .5 .8 .0005 .1 . • The reaction rate Acceleration Factor is AF(temp. tempU . tempU .01 .5 .4 . • Masked failure modes could dominate in the field. Ea) > 1.0001 300 Deg C 10 2 10 2 250 10 3 200 10 4 175 150 10 5 100 10 6 100 150 200 250 300 350 10 7 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 = .3 .15) • and β1 = Ea is the activation energy • σ is constant = exp Ea • When temp > tempU . AF (temp.005 .8 10 6 .

0 173 Degrees C 5. requires statistical methods not yet widely available.g.Possible results for a typical temperature-accelerated failure mode on an IC device Unmasked Failure Mode with Lower Activation Energy 10 6 10 6 10 5 10 5 Hours Hours 10 4 10 4 Mode 2 10 3 10 3 10% 10 2 10% 10 2 Mode 1 10% 10 1 10 40 60 80 Degrees C 100 120 140 1 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. (Reduction to failure-time data loses information) 133 Degrees C 1..0 • Useful reliability inferences even with 0 failures. (Modern computing capabilities should help here) • Degradation level may not correlate well with failure.5 • More justification and credibility for extrapolation. 83 Degrees C 0. 10. (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 80◦C • Degradation data may be difficult or impossible to obtain (e. taking apart a motor to measure wear). • Substantial measurement error can diminish the information in degradation data. destructive measurements)..g. Percent Increase in Operating Current 0 0 5 10 15 1000 2000 Hours 3000 4000 23 24 .0 Percent Increase • Can be more informative than time-to-failure data. • Analyses more complicated. • Obtaining degradation data may have an effect on future product degradation (e.

.02 . dt k1 > 0.. • Understanding the relationship between the laboratory test results and product field reliability will provide stronger basis for using future laboratory tests to predict field performance. (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.01 .6 . 195◦C.5 .6 -0.7 .0 -0. The Arrhenius model describing the effect that temperature has on the rate of a simple first-order chemical reaction is 4000 0 k1 = γ0 exp −Ea kB × (temp + 273. • Find a model (transfer function) to relate laboratory test to field use.g. 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 field failures. • Distribution of environmental conditions (e. Seek understanding of reasons for lack of agreement.0 -1. 0 10000 20000 Cycles 30000 40000 50000 29 30 .9 .8 -1. 237 Degrees C 195 150 Degrees C 80 Degrees C Proportion Failing .2 • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • -1.98 .4 .005 .1 .99 . stress spectra distributions). and 237◦C (Use conditions 80◦C) Arrhenius Model Temperature Effect on Chemical Degradation A1 k1 -A2 and the rate equations for this reaction are 0.95 . • Distribution of use-rates in actual use.4 • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • -0. cycling rate).g.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: .4 • • • • • • • • • • • • • • • • • • • • • 150 Degrees C • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • 195 Degrees C • • • • • • • • • • • • • • • • • • • • • • • • • • • • • 237 Degrees C • • • • • • 1000 2000 3000 • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • dA1 = −k1A1 dt Solving these gives and dA2 = k1A1. Same failure mechanisms operating in laboratory tests? Same factors (environmental noises) exciting the failure mechanisms? Identify laboratory/field discrepancies to improve test procedures.8 • Effects of acceleration (e. in some situations.05 . inferred from the available data.Device-B Power Drop Accelerated Degradation Test Results at 150◦C.2 • • • • • • • • • • • • • Power drop in dB -0.001 10^1 10^2 10^3 10^4 10^5 These factors may be given or.2 .3 .

1 Probability .95 . i=1 • Then the failure probability as a function of time is FT (t.02 .02 .0005 .Appliance Use-Rate Distribution (discretized lognormal distribution) Example Use-Rate Model • Life of a component in cycles of use.001 .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 .5 .8 .05 0. ηR.8 .10 k 0.1 .2 .001 .05 .7 .15 log(c) − µ σ • Actual use-rate has a distribution given by the proportion of users πi (i = 1.002 .01 .002 .00 Thu May 10 22:24:23 CDT 2001 Test Cycles per Week 35 36 .00005 .1 .5 Fraction Failing . σ) and µi = µ − 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 . .20 1.02 .05 0.2 .0005 .05 .7 . k) that use the appliance at constant rate Ri.999 .6 .005 .00001 2 5 10 20 50 100 200 500 1000 2000 5000 Laboratory Field Field Variability Lognormal f (r. .003 . .4 . . . . .01 .6 .4 .7 .00 5.5 . θ ) = 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 0.2 .3 .01 .01 0.9 .05 .3 .9 .00 20.0002 . .005 .98 . has a distribution FC (c) = P (C ≤ c) = Φ 0.3 .0 πi Φ i=1 log (t) − µi σ where θ = (µ1. µk .005 .95 . σR ) Density for the Wear Failure Mode (unloaded cycles relative to field 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 0.9 . where k πi = 1.

etc.60. W. 1995. environmental considerations. 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. 1993 Proceedings Annual Reliability and Maintainability Symposium. IEEE Transactions on Reliability R-47. • Meeker. Balakrishnan.98.A. • Users of Accelerated Testing must beware of pitfalls 40 References References • D. Robust function for attaining high reliability at low cost.84 centime= 183. Tseng. 10 3 Days 10 2 • Consider including some tests at the use conditions.Q. 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.5 quantile)= 0. Hamada. and Data Analyses. NY: Unipub/Kraus International Publications. Statistical Tools for the Rapid Development & Evaluation of High-Reliability Products. M. • Physical comparison of lab and filed 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 effectively with all issues Product/reliability/design engineers to identify productuse profiles. L. 114-118. M. John Wiley and Sons. and Hamada. Quinlan. Statisticians to help with stochastic modeling. (1998a). H. 1990. • T. editor.5 quantile)= 9. Using degradation data from a factorial experiment to improve fluorescent lamp reliability. pp 183-191. W.183. Inc.Q. New York: John Wiley & Sons. Taguchi. and Escobar.Planning Accelerated Tests Simulation of a Proposed Accelerated Life Test Plan • Most basic ideas of traditional DOE still hold 10 5 Temp= 78.A. Chiao. 1987. as much as possible the amount of extrapolation used. Accelerated Testing: Statistical Models. and to help quantify uncertainty in results. 363-369.4632 Average( 0. W. • L. Nelson. and Escobar. Boca Raton: CRC Press. potential failure modes or weaknesses that need to be evaluated. 10 1 10% Results based on 500 simulations Lines shown for 50 simulations 10 0 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 field environment • Bayesian methods for analysis and planning (especially adaptive test plans) • Accelerated degradation test planning • Degradation analysis and planning with coarse (e. System of Experimental Design. Condra. Reliability Improvement with Design of Experiments. Pitfalls of Accelerated Testing.7266 SD(Ea)= 0. 0. (1998b). Test Plans. 187-198.7330. • M. Using statistically designed experiments to improve reliability and to achieve robust reliability. Inc.7265.5116 Average(Ea)= 0.08594 • Limit. Byrne. ordered categorical and censored) data. S. IEEE Transactions on Reliability R-44.183 parameters= -16. Inc. 1993. Hamada.1 quantile)= 8. Analysis of experiments for reliability improvement and robust reliability. • M. (1995). . 0. • Use simulation to investigate properties of alternative ALT plans.138 SD( 0. N. W. New York: Marcel Dekker. Statistical Methods for Reliability Data. (1995).120 n= 155. 1993. White Plains. • Meeker. 1995 June. Hamada. in Recent Advances in Life-Testing and Reliability. IEEE Transactions on Reliability R-44. fit models. L.g. 42 .014 SD( 0. Journal of Quality Technology. C. J. • G.Q. plan tests. • W. 41 • Meeker.6000 Log time quantiles at 50 Degrees C Average( 0.1 quantile)= 0.