Planning 2-Stage Accelerated Life Tests LC Tang (董润楨), Ph.D Department of Industrial and Systems Engineering National University of Singapore Overview • Planning a sequential Accelerated Life Test (ALT) • Motivation of using an Auxiliary Stress (AS) • An integrated planning framework for sequential ALT with an AS • Numerical illustrations A Constant-Stress ALT Time Probability distributions Maximum Test Duration Life-stress relationship Use Stress Low Stress Mid Stress High Stress Stress Level A Scale-Accelerated Weibull Lifetime Model • Standardization of stress x x s s sk s0 sk 1 s0 : use stress sk : the highest stress • Weibull lifetime distribution at any stress log T - SEV • A scale-accelerated failure time model 0 1 x , 1 0 is a constant independent of stress Motivations of Sequential ALT Planning • ALT planning based on the Maximum Likelihood theory Step 1: Step 2: Step 3: Specify ALT model parameter values Minimize the asymptotic variance of ML estimator Evaluate the plan using simulations • Locally optimal for specified model parameters 0 , 1, • Problems: – There often exists a high margin of specification error – Developed plans are usually sensitive to the specified value A Framework of Sequential ALT Planning Information Planning information e.g. test duration, specified parameter values, etc. Information on the slope parameter 1 Planning Procedure Plan & Perform the test at the highest stress to quickly obtain failures Preliminary information on ( 0 , ) Planning information e.g. test duration, number of stress levels, sample sizes, etc. • • Plan the tests at lower stress levels Tang, L.C. and Liu, X. (2010) “Planning for Sequential Accelerated Life Tests”, Journal of Quality Technology, 42, 103-118. Liu, X. and Tang, L.C. (2009) “A Sequential Constant-Stress Accelerated Life Testing Scheme and Its Bayesian Inference”, Quality and Reliability Engineering International, 25, 91-109. Part I Planning Sequential Constant-Stress Accelerated Life Tests Sample Size at the Highest Stress Level Specify the values of 0 (or H ) and the censoring time cH the expected number of failures RH Sample Size: RH p nH Page 8 p 1 exp cH / exp( H ) 1/ Inference at the Highest Stress Level Time in log-scale θH l θH ; DH θH H , Stress 0 High Low 1 Use Inference at the Highest Stress Level ˆ ) θˆ H | y H ~ N (θˆ H , Σ H where θˆ H arg max (θ H ) ˆ Iˆ 1 Σ H Generalized MLE Covariance matrix H Iˆ H [ 2l (θ H ; D H ) / θ H 2 ]θ H θˆ H Observed information Page 10 Construction of Prior Distributions k , i , for xi 0,1 H , Information on the value of 1 i k 1 xi is a constant Construction of Priors at Low Stresses for any i 1,..., H 1, there exists a one-one transformation θi (θH ) with non-vanishing Jacobian θH / θi ,such that 1 (θi ) ( i 1 xi , ) 1 1 ~U ( 1 , 1 ) ( i 1 xi ) dF ( 1 ) i ( i ˆ H ) 2 erf ( i ) erf ( i ) exp ˆ 23/ 2 ( var(ˆ H ))1/ 2 (i i ) 2 var( ) H where i ˆ H 1 xi , i ˆ H 1 xi , i ˆ H 1 xi , cov(ˆ H , ˆ H ) / var1/ 2 ( ˆ H ) var1/ 2 (ˆ H ) i var1/ 2 (ˆ H ) i var1/ 2 (ˆ H ) ( i ˆ H ) var1/ 2 ( ˆ H ) (2 var( ˆ H ) var(ˆ H )(1 2 ))1/ 2 i i var1/ 2 (ˆ H ) i var1/ 2 (ˆ H ) ( i ˆ H ) var1/ 2 ( ˆ H ) (2 var( ˆ H ) var(ˆ H )(1 2 ))1/ 2 i Page 12 Illustration of the Sequential ALT Plan & Run the test at the highest stress Time in log-scale Deduction of Prior Distributions Stress 0 High Page 13 Low 1 Use Pre-Posterior Analysis & Optimization The Bayesian Optimization Criterion Given the information obtained under the highest stress, the optimum sample allocation and stress combinations for tests under lower stresses are chosen to minimize the pre-posterior expectation of the posterior variance of certain life percentile under use stress over the specified range of β1 Min C (ξ ) = E1 {var( y p (1))} = E1 {c var(θ0 )cT } c [1, log( log(1 p))] Page 14 Problem Formulation Design Matrix X 1 x1 1 xH 1 1 xH T E1 (var( y p ( x1 ))) Λ E1 (var( y p ( xH 1 ))) var( y p (0)) Min E1 (var( y p (1))) 1( XT Λ 1X) 11T s.t. x {( x1 , x2 ,..., xH 1 ) π {(1 , 2 ,..., H ) H 1 H : 0 xi 1} and xH 0 : i i 1 and 0 i 1} Page 15 Pre-Posterior Analysis 1 E1 (var( y p ( xi ))) 1 1 Σ i I θi I 1 1 cΣi cd 1 i 1 where 2 log l θi 2 log l θi I θi E ; y f y dy 2 2 θi θi 2 log θ i i I θi 2 Page 16 Information contained in the prior density Information expected to obtain at stress level i Adhesive Bond Test (Meeker and Escobar 1998) • Total Sample Size: 300 • Total Testing Duration: 6 months =183days • Standardized Testing Region: 0 xH x xU 1 • Assumptions: log(T ) ~ SEV , log A Activation energy, Ea 1 Boltzmann constant, k B 8.6171105 T (Arrhenius) 0 1 x, is a constant Page 17 0 log A Ea k B 1 sH , 1 Ea k B 1 ( s0 sH ) Planning at the Highest Temp Planning information: H 4.72 0.6 RH 15 cH 60 50 samples are needed Page 18 RH nH p Posterior Density Simulated Failure times: 33.3, 48.4, 39.3, 58.8, 47.4, 60.0, 33.6, 19.4, 38.0, 28.6, 60.0, 53.2, 17.7, 25.4, 44.5, 34.6, 16.9, 60.0, 31.7, 60.0 ,49.2, 60.0, 10.953, 60.0, 18.8, 3.3, 1.4, 17.3, 46.8, 40.9, 60.0, 28.4, 60.0, 4.2, 21.9, 49.6, 20.6, 60.0, 46.6, 6.4, 25.2, 60.0, 13.6, 29.5, 60.0, 60.0, 31.3, 29.4, 54.3, 34.0 Page 19 Normal Approximation ˆ ) θˆ H | data ~ N (θˆ H , Σ H ˆ Iˆ 1 and Iˆ [ 2l (θ ) / θ 2 ] where Σ H H H H H θ H θˆ H Page 20 Planning of an ALT with 2 Stress Levels Sample size nL 300 nH 250 100 Test duration cL 183 cH 123 Posterior density at xH ˆ θ H | y H ~ N θˆ H , Σ H Specified range of 1 1 3.84,5.12 (i.e.Ea 0.6, 0.8) E1 var y p 1 in log-scale Planning Information: 10 1 0.1 0.01 0.1 0.2 0.3 0.4 High 0.5 0.6 xL 0.7 0.8 0.9 1 Low Page 21 Effects of the pre-specified slope parameter Suppose we raise the expectation of the product reliability Ea 0.6,0.8 Ea 0.6,0.9 i.e. 1 3.84,5.12 1 3.84,5.76 E1 var y p 1 in log-scale Effect: 100 10 Beta1 ranges from 3.84 to 5.76 1 0.1 Run the test under a higher stress to produce more failures Beta1 ranges from 3.84 to 5.12 0.01 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 High xL Low Page 22 Plan an ALT with 3 stress levels Planning Information: Min E1 (var( y0.1 (1)); xL , ) nL 250, cL 123 H , , 1 3.84,5.12 s.t. nL M (1 ) p ( xL ) RL Additional constraints: nL M p ( xM ) RM nL 250 (1 ) 0 x H x L 2 xM 1 nM 250 xM for 0 1 0 1 xL 2 Minimum number of failure RL and RM where p( xL ) (1 exp( exp( ))) ( , )d d ) (1 exp( exp( ))) ( , )d d L 0 p( xM L L L L 0 M M M M Page 23 M The feasible region Page 24 Interior Penalty Function Method Page 25 Page 26 Inference from Test Results Simulated failure times (assume 0 4, 1 4, 4) Stress Sample Test Expected Simulated Observed Level Size Duration Failures Failure Times Failures 15 33.3, 48.4, 39.3, 58.8, 47.4, 60.0, 33.6, 19.4, 38.0, 28.6, 60.0, 53.2, 17.7, 25.4, 44.5, 34.6, 16.9, 60.0, 31.7, 60.0 ,49.2, 60.0, 10.953, 60.0, 18.8, 3.3, 1.4, 17.3, 46.8, 40.9, 60.0, 28.4, 60.0, 4.2, 21.9, 49.6, 20.6, 60.0, 46.6, 6.4, 25.2, 60.0, 13.6, 29.5, 60.0, 60.0, 31.3, 29.4, 54.3, 34.0 38 High xH 0 Mid xM 0.39 Low xL 0.78 50 20 230 60 123 123 5 46.1 62.5 86.2 98.9 101.7 123 (×224) 5 5 22.8 44.8 59.1 84.4 87.7 105.2 123 (×224) 6 Page 27 Inference • Results obtained under the high stress H , 0.0112 0.0003 ~ N 3.87, 0.65 , 0.0003 0.0086 • Results obtained under the mid and low stress M , Increasing Decreasing 0.0156 0.0016 ~ N 5.28, 0.594 , 0.0016 0.0080 L , 0.0377 0.0060 ~ N 7.24, 0.664 , 0.0060 0.0042 Simulation Study Planning information: Total Sample Size: 300 Total Test Duration: 183 Pre-specified ALT model parameters: 9 scenarios are considered *For sequential plans: We set the expected number of failures at the high stress level at 15 within 60 days *For each simulation scenario: a. both sequential and non-sequential plans are generated; b. failure data are generate according to the optimum plans; c. 10th percentile are use stress are estimated; d. repeat b and c for 100 times, and move to another scenario Page 29 Simulation Design Table - k %: the specified value is k% lower than the true value +k %: the specified value is k% higher than the true value (0): the specified value is the true value Scenarios Pre-specified Pre-specified Pre-specified 1 Pre-specified 1 (non-sequential) (sequential) 0 1 (0) (0) (0) - 20 % ~ + 20 % 2 - 25 % - 25 % - 20 % - 20 % ~ + 20 % 3 - 25 % - 25 % + 20 % - 20 % ~ + 20 % 4 - 25 % + 25 % - 20 % - 20 % ~ + 20 % 5 - 25 % + 25 % + 20 % - 20 % ~ + 20 % 6 + 25 % - 25 % - 20 % - 20 % ~ + 20 % 7 + 25 % - 25 % + 20 % - 20 % ~ + 20 % 8 + 25 % + 25 % - 20 % - 20 % ~ + 20 % 9 + 25 % + 25 % + 20 % - 20 % ~ + 20 % Page 30 Simulation Results Page 31 Precision 1. Sequential plans yields more precise estimation Variance 2. Sequential plans gives a conservative sense of statistical precision: Sample variance > Asymptotic variance 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 Sample variance (non-sequential plan) Asymptotic variance (sequential plan) Sample variance (sequential plan) 0 Page 32 Asymptotic variance (non-sequential plan) 1 2 3 4 5 6 7 Simulation scenarios 8 9 10 Effect of Parameter Mis-specification on Precision Sequential Plans Non-sequential Plans Effect on the expected variance Effect on the observed variance 0.270 0.1945 0.016 -0.2075 0.044 0.043 -0.007 -0.1180 For sequential plan: 0.083 0.0655 Since 0.035 0.0185 1. -0.031 -0.009 Model parameters 0 and are estimated at stage one; For non-sequential plan: 2. An interval value of 1 is used Results are sensitive to the specified model parameters 0 and 1 . Hence, the plan robustness to the misspecification of model parameters Page 33 has been enhanced 0 1 0 * 1 0 * 1 * 0 * 1 * Effect on the expected variance Effect on the observed variance 0 -0.053 -0.038 (0, 0.0001) (- 0.0001,0) 0 * (- 0.0001,0) (- 0.0001,0) Robustness Define the Relative Error (RE) as: sample variance - asymptotic variance asymptotic variance 3. Sequential plans is more robust to mis-specification of model parameters 2.5 RE (non-sequential plan) 2 RE 1.5 RE 1 (sequential plan) 0.5 0 0 Page 34 1 2 3 4 5 6 7 Simulation scenarios 8 9 10 Effect of Parameter Mis-specification on the Relative Error (RE) Non-sequential Plans Sequential Plans Effect 0 1 0 * 1 0 * 1 * 0 * 1 * Effect 0 -0.7684 0 * -0.7187 -0.2367 0.0011 (0, 0.0001) (0, 0.0001) 0.4905 0.1334 For sequential plan: 0.1201 Since -0.0532 1. Model parameters 0 and are estimated at stage one; For non-sequential plan: 2. An interval value of 1 is used RE is sensitive to the pre-specified model parameters 0 and 1 . Hence, the plan robustness to the misspecification of model parameters has been enhanced Page 35 Simulation Results 4. Sequential plans reduce the degree of extropolation; 5. Sequential plans are especially robust to mis-specification of the intercept parameters (scenarios 6-9) due to the preliminary test under the high stress Optimum low stress (non-sequential plan) 120 Temperature 100 80 60 40 20 Use stress Optimum low stress (sequential plan) 0 0 Page 36 1 2 3 4 5 6 7 Simulation scenarios 8 9 10 Effect of Parameter Mis-specification on the Optimum Low Stress level Non-sequential Plans Sequential Plans Effect 0 1 0 * 1 0 * 1 * 0 * 1 * Effect 0 12.25 0 * -13.25 3.25 -5 (- 0.0001,0) (- 0.0001,0) -1.25 3.25 For sequential plan: 1.75 Since 0.75 1. Model parameters 0 and are estimated at stage one; For non-sequential plan: 2. An interval value of 1 is used RE is sensitive to the pre-specified model parameters 0 and 1 . Hence, the plan robustness to the misspecification of model parameters has been enhanced Page 37 Comparison with 4:2:1 Plan ASR c: |Ase Ase when model parameters are correctly specified| Ase when model parameters are correctly specified Test duration at the highest stress level Extension from 2-Stage Planning to a Full Sequential Planning • • • • • 2-Stage Planning Prior distributions for all low stresses are constructed simultaneously (all-at-one) Tests at all low stresses are planned simultaneously Full Sequential Planning Only the prior distribution for one low stress is constructed Only the test at one low stresses are planned More tests at low stresses are planned iteratively The basic framework still works ! Part II Planning Sequential Constant-Stress Accelerated Life Tests with Stepwise Loaded Auxiliary Stress Liu X and Tang LC (2010), “Planning sequential constant-stress accelerated life tests with stepwise loaded auxiliary acceleration factor”, Journal of Statistical Planning and Inference, 140, 1968-1985. Motivations of an Auxiliary Stress • Testing more units near the use condition is intuitively appealing, because more testing is being done closer to the use condition • Failures are elusive at low stress levels for highly reliable testing items – the lowest stress level is forced to be elevated, resulting in high, sometimes intolerable, degree of extrapolation in estimating product reliability at use stress Illustration Low degree of extrapolation with zero failure Time high degree of extrapolation with more failures Maximum Test Duration Use Stress Candidate low stress 1 Candidate low stress 2 High Stress Stress Level Auxiliary Stress • An Auxiliary Stress (AS), with roughly known effect on product life, is introduced to further amplify the failure probability at low stress levels • Examples of possible AS: – In the reliability test of micro relays operating at difference levels of silicone vapor (ppm), the usage rate (Hz) might be used as an auxiliary factor (Yang 2005). – In the temperature-accelerated life test, the humidity level controlled in the testing chamber might be used as an AS (Livingston 2000). – Dimension of testing samples (Bai and Yun 1996) • Joseph and Wu (2004) and Jeng et al. (2008) proposed a method known as the Failure Amplification Method (FAMe) for the Design of Experiments. – FAMe was developed for system optimization while ALT is used for reliability estimation at user condition through extrapolation. Model Extension • Standardization of the level of AS h (v vuse ) (vmax vuse )1 vuse : use stress vmax : the highest stress • The extended testing region: [0,1]2 • A scale-accelerated failure time model 0 1 x 2 h is a constant independent of stress Examples: Hallberg-Peck model Higher usage rate model (Yang 2005) An Integrated Framework of Sequential ALT Planning with an Auxiliary Stress Planning Information e.g. Sample size; Test duration; Specified model parameters Step 1: Plan and perform the life test at the highest stress level Step 2: Compute the number of failures at low stresses Is an AS needed? No yes Is an AS available? yes Step 3b: Plan the tests at low stresses with an AS i.e. optimize sample allocation, stress combinations, and the loading profile of AS No Step 3a: Plan the tests at low stresses without an AS i.e. optimize sample allocation, and stress combinations Step 1 Planning & Inference at the Highest Temperature Level ALT for Electronic Controller Experiment Target: To demontrate the 10% life quantile at use condition exceeds 2 years Stress Factor: Temperature (other factors, such as humidity, voltage, etc are set to use level) Planning information and Assumptions: 1). 120 sample units and 75 days are available. 2). The use temperature is 450 C 318K The highest temperature allowed in the test is 850 C 358K 3). Failure time T follows Weibull distribution F t log t 4). is a constant, independent of temperature; follows Arrhenius stress-life relationship i log A Activation energy, Ea 1 0 1 si Boltzmann constant, k B Ti where 0 log A 1 Ea si 11605 / Ti Test Planning at the Highest Stress Planning Inputs: target number of failures: rk 1/ k Rk exp ck exp parameter values: k ( 0 ), confidence level: censoring time: ck Planning Output: nk rk 1 C i 0 i nk nk i 1 R R k k 1 i Risk of see less failures than expected (Binomial Bogey test, Yang 2007) Testing Output: ˆ H ( ˆ0 ) or ˆ Results Planning information: k 7.5 0.677 rk 6 ck 720hr 0.9 44 samples are needed Data Obtained at the Highest Stress Weibull Probability Plot for Observed Failure Data Time-to-failure (hours) 79.559 210.47 400.56 491.41 673.98 109.4 204.7 425.32 117.15 328.99 720×29 Note: This is just a particular run 590.03 138.94 149.95 643.31 351.87 0 -0.5 -1 -1.5 -2 -2.5 R² = 0.8995 1.7 2.2 2.7 3.2 Statistical Inference at the Highest Stress Posterior distribution derived from a constant prior : θk ; y l θk ; y y j k 1 y j k ck k exp j log exp 1 exp j j 1 where θk ( k , ); 0 if the data is censored, otherwise 1 nk Normal Approximation to the Posterior distribution (Berger 1985) ˆ ) N (θˆ ,[ Iˆ ]1 ) θ k y ~ N (θˆ k , Σ k k k where 2 ˆI = l (θ k ; y ) k θ k 2 θ (observed Fisher information at θˆ k ) k θˆ k 0.0529 ˆ Iˆ 1 0.1142 θˆ k [7.35, 0.90] Σ k k symmetric 0.0489 Illustration • The quality of the approximation needs to be checked e.g. Kolmogorov-Smirnov (K-S) test (Martz et.al 1988, Technometrics). • The posterior normality needs to be checked e.g. Kass and Slate 1994 Ann. Statist. ). Step 2 Computation of the Expected Number of Failures at Low Stress Levels Construction of Prior Distributions k , i , for xi 0,1 H , Information on the value of 1 i k 1 xi is a constant Density Function of the Constructed Prior 1 i , ( i 1 xi , ) 1 ( i 1 xi ) ( 1 )d 1 i ( ˆ ) 2 1 3/ 2 exp erf ( i ) erf ( i ) 1/ 2 2 ( var(ˆ )) (i i ) 2 var(ˆ ) i 1,..., k 1 where ( 1 ) is a uniform distribution defined on an interval [1 ,1 ] erf is the error function given by the definite integral erf ( z ) 2 1/ 2 z 0 e t dt i ˆ k 1 xi i ˆ k 1 xi , i ˆ k 1 xi , cov( ˆ k , ˆ ) / var1/ 2 ( ˆ k ) var1/ 2 (ˆ ) i var1/ 2 (ˆ ) i var1/ 2 (ˆ ) ( ˆ ) var1/ 2 ( ˆ k ) (2 var( ˆ k ) var(ˆ )(1 2 ))1/ 2 i i var1/ 2 (ˆ ) i var1/ 2 (ˆ ) ( ˆ ) var1/ 2 ( ˆ k ) (2 var( ˆ k ) var(ˆ )(1 2 ))1/ 2 i 2 Illustration of the Constructed Priors at 65⁰C and 45⁰C Let Ea 0.8,1.2, i.e 1 ~ Uniform0.8,1.2 Uncertainty over becomes larger for lower testing temperature Expected Number of Failures at Low Stress In order to see 5 failures, the temperature is almost on the middle of the testing region !! Another Point of View: Prior Information v.s Information To Be Obtained det I θi det Ii 2l (θi ) 2 log (θi ) where I θi =E and I θi = 2 θi 2 θi Information to be obtained by performing a test at stress level i Little Information obtained from low temp “Information” contained in the prior knowledge Step 3 Planning at the Lower Temperature Level With Auxiliary Stress •The choice of AS •The loading of AS •The integration of AS in test planning The Choice of AS Assumpotions: 1). The effect of AS is well understood 2). The failure mode does NOT change after an AS is introduced Auxiliary Stress: Humidity Hallberg-Peck Model (Livingston, 2000): RH j 0 1s p log RH 0 RH 0 : use humidity level, 60% RH : humidity level in test ( 90%) The Choice of Loading Profile for AS ConstantStress Loading Step-Stress Loading A 2-step step-stress loading profile is preferred due to the following reasons: • The initial loading will not be too harsh • The stress can be dynamically monitored given a target time compression factor (only amplify the failure as needed) • The verification of the effect of the selected AS is possible Setting a Target Acceleration Factor equivalent test duration, ci(e) Time Compression Factor: i actual test duration, ci LCEM Cumulative Exposure Model (Yeo and Tang 1999, Tang 2003) A Bayesian Planning Problem Min E1 (var( y p (1))) 1( XT Λ 1X) 11T s.t. target time compression i , for i 1,..., k 1 ( x1 , x2 ,..., xk 1 ) [0,1]k 1 Stress levels (1 , 2 ,..., k 1 ) [0,1]k 1 : i 1 i 1 k Sample allocation (h1 , h2 ,..., hk 1 ) [0,1]k 1 Initial level of AS ( 1 , 2 ,..., k 1 ) [0, c]k 1 Stress changing time for AS k 1 where T 1 1 1 X x x x k 1 2 0 E1 (var( y p ( x1 ))) 0 E1 (var( y p ( x2 ))) Λ 0 0 0 0 0 0 var( y p ( xk )) 0 Planning Results Planning Information: Sample size n1 120 n2 76 Test duration c1 1800 720 1080 Posterior density at xH θ 2 ~ N θˆ 2 , Iˆ 1 θ2 1 ~ Uniform 0.8,1.2 p3 Maximum RH = 90% Use RH = 60% 3 Testing Condition Temp (C) RH (%) Use 45 60 Low 53 High 85 Testing Duration Sample Size See Profile 1080hrs 76 60 720hrs 44 Humidity Loading Profile at Low Temperature Low Humidity Level: 60% High Humidity Level: 90% Holding Time: 170.5 hrs Expected Failures: Interval [0, 170.5] : No failure Interval [170.5,1080]: 5 failures Interval [1080, ): 71 censored Illustration: ALT without AS Relative Humidity Point B: (53, 60%) Failure Probability < 0.01 Point A: (85, 60%) Failure Probability = 0.32 60% Temperature 53 85 Illustration: ALT with AS Relative Humidity Point D: (53, 90%) Failure Probability = 0.08 90% Point C: (53, 60%) Failure Probability < 0.01 Point A: (85, 60%) Failure Probability = 0.32 60% Temperature 53 85 Sensitivity of the Optimal Plan to p RHT: Relative change of low humidity holding time RT/RH Relative change of low humidity/low temperature RSD: Relative change of Asymptotic SD Sensitivity of the Plan to the Activation Energy RHT: Relative change of low humidity holding time RT/RH Relative change of low humidity/low temperature RSD: Relative change of Asymptotic SD Evaluation of the Developed ALT Plan References of Part II • Liu X and Tang LC (2010), “Planning sequential constantstress accelerated life tests with stepwise loaded auxiliary acceleration factor”, Journal of Statistical Planning and Inference, 140, 1968-1985.