Computer Experiments to Predict Propagation of Variation An Aircraft Engine Blade Assembly Case Study Presented at the Fall Technical Conference King of Prussia, PA October 2002 By David Rumpf and GE Aircraft Engines (781) 594-5508 fax (781) 594-0954 1000 Western Ave. Lynn, MA 01910 USA David.Rumpf@ae.ge.com William J. Welch Department of Statistics and Actuarial Science (519) 888-4567 x5545 fax (519) 746-1875 University of Waterloo Waterloo, Ontario N2L 3G1 Canada wjwelch@uwaterloo.ca Special thanks to Robert Shankland, GEAE Engineering for his patience and expertise in running the analytic computer stress model. David Rumpf, Statistician GE Aircraft Engines Page 1 Blade Assembly Stress Study X1=interference fit Y1a,b= notch mean and alternating stress X2=CP rabbet interference fit Y2a,b = rabbet fillet mean and alternating stress X3=CP drop a function of three drops David Rumpf, Statistician Goals: 1) A DFSS (design for six sigma) design which meets LCF life requirements 2) Tolerance requirements for X1, X2, X3 What was available: • A computer model which evaluates stress for any specific set of X1, X2 and X3 values. Run time ~ 1 hour • Two stress points, Y1 and Y2 as shown. • Two outcomes for each location, mean stress and alternating stress. • Alternating is bigger driver for part low cycle fatigue life. What was needed: • Non-linear transfer functions for Y1a,b and Y2a,b versus X1, X2 and X3 which could be used for multiple Monte Carlo models, ~1000 iterations each, for stress versus tolerances on X1, X2 and X3. GE Aircraft Engines Page 2 Six Sigma, Producibility and Robust Design “sigma level” Typical Historical Situation Process Tolerance meets Customer Need 2 xbar +/- 3 s only +/- 2 s fit within tol Quality plan relies on inspection. Expect to have rework, scrap and MRB activity. Reaching six sigma goal requires combined Manufacturing/Design Effort “sigma level” Combined 6 Engineering, Manufacturing 6 s Goal Improved Manufacturing Process Robust Design allows Wider Tolerance to meet Customer Need xbar +/- 3 s +/- 6 s will fit within tol Quality plan focus on parameter control and process monitoring. First time yield 100%. David Rumpf, Statistician GE Aircraft Engines Page 3 Statistical/Design of Experiment Opportunities Manufacturing Process Improvement: • Screening designs • Factorial designs • Leveraging • EVOP • Quality Improvement metrics • Review by Vice-president Engineering, meeting Customer Needs and improving producibility: • Quality Function Deployment • Voice of the Customer • Robust Design: • Screening Focus of this presentation • Factorial • Response Surface Designs • Producibility Scorecards • Review by Vice-president David Rumpf, Statistician GE Aircraft Engines Page 4 Robust Design Y = Stress or Useful Life X’s are parameters which impact Y which could include: • Part Key Characteristic values • Environment • Mating part Key Characteristic values • Customer usage pattern X’s are typically a combination of controllable and noise (uncontrollable) factors Y = f(XC , XN) Goals: • Target Y • Minimize sY, that is, variation in Y David Rumpf, Statistician GE Aircraft Engines Page 5 Why Robust Design Statistical/Engineering method for product/process improvement (Taguchi’s idea) Two types of factor, control (Xc) and hard to control (Xn or noise) • Control factor levels can change target • Hard to control factors have variation during normal process or usage Robust design: Set Xc to take advantage of non-linearity in Y = f(Xn) • Design space is typically non-linear Non-linear Response 14 Response 12 10 8 6 4 2 0 1 David Rumpf, Statistician 2 3 4 Xn GE Aircraft Engines 5 6 7 Page 6 Wu and Hamada recommend a two step process Obtain Transfer Functions: Ybar = f1(XC) sY = f2 (XC) Typically one finds different sets of X’s in the two transfer functions If Target is goal: • Minimize variation in Y, the harder objective • Minimize Ybar distance from target If maximum or minimum is the goal: • Optimize Ybar • Minimize variation in Y An alternative approach is non-linear optimization of Z where Z = |Target – Ybar|/ sY Experiments: Planning, Analysis and Parameter Design Optimization, CFJ Wu and M Hamada, Wiley 2000 David Rumpf, Statistician GE Aircraft Engines Page 7 Statistical Issues with Analytic Models Designing a new part: • Typically done analytically • Often a complex, time consuming process to obtain a result for a single set of parameter values • Examples include finite element analysis models, system models, etc • Leads to serious optimization and simulation issues Recommended approach: • Run designed experiment, typically Response Surface, to capture non-linear effects • Use RSM transfer function for • Optimization • Simulation to estimate effect of variation in X’s on Y Statistical Problem: • Analytic models have no random variation, always the same answer for a set of X values • RSM assumes normally distributed error in residuals from model fit • Residuals from analytic model are entirely lack of fit. David Rumpf, Statistician GE Aircraft Engines Page 8 Case Study was a Learning Process for the GEAE Author Initial approach: (Note: All results coded for proprietary reasons) • Full Factorial with center-point, 9 computer runs, ~ 9 hours run time • Y’s = life required log transformation • Choose to use Y’s = stress • Interactions and curvature were significant, see Y1a graph below Centerpoint5 05 -0 .0 Interaction Plot (data means) for notch 0 .0 04 5 -0 .0 07 - 0 .0 03 1 -0 .0 0 30 RetArm 20 0.0045 -0.0055 10 30 CP Rab 20 -0.003 -0.007 10 CP Drop 30 20 0 -0.01 David Rumpf, Statistician Largest interaction and curvature effects Results led team to an RSM design in 3 factors. 10 GE Aircraft Engines Page 9 RSM Design • Face centered central composite design, illustrated for two factors below • We ran 15 runs, no repeated center-points since computer model has no random variation Factor B RSM analysis requires 9+ runs for a 2 factor design, 15+ runs for a 3 factor design. Factor A David Rumpf, Statistician GE Aircraft Engines Page 10 RSM Results Analysis plan and results: • Chose most parsimonious model via backwards selection based on p values Response Y1a Y1b Y2a Y2b Main 3 3 3 3 Two-way 1 1 0 1 Quadratic 1 2 1 0 R-sq adjusted 98.2 95.0 98.7 94.3 Residual versus Standard Order Run Number Concerns: 1) High residuals, especially for Rabbet stress, cause concern. 2) R-sq not as high as desired 5 Y1a Y1b Y2a Y2b 4 Questions: 1) Does this transfer function fit well enough for engineering need? 2) Is there a better way to fit analytic/computer model results Residual 3 2 1 0 -1 -2 -3 0 5 10 15 Run-Num David Rumpf, Statistician GE Aircraft Engines Page 11 Enter Professor William Welch and a Space Filling Design GEAE author looked for help. Jeff Wu suggested Professor W. Welch of the University of Waterloo New Experimental Plan: • Space filling design instead of DOE or RSM design • Recommended for computer/analytic experiments • Multiple levels to provide better estimate for non-linear and interaction effects • 33 runs for 3 factors • Doubled the number of runs • Spaces levels for each factor at 1/32nd of the range • Plots below show experimental grid, 2 factors at a time • Computer experiment was run for the 33 sets of conditions (~30 hours of run time) 0.005 0.000 -0.003 0.000 CP Drop CP Rabbet RetArm -0.004 -0.005 -0.005 -0.006 -0.005 -0.010 -0.007 -0.007 -0.006 -0.005 -0.004 CP Rabbet David Rumpf, Statistician -0.003 -0.005 -0.010 -0.005 0.000 0.000 0.005 RetArm CP Drop GE Aircraft Engines Page 12 Approximating Random Variable Function Model Treat Deterministic output Y(x) as a realization of a random function (stochastic process) Y(x) = Ybar(x) + Z(x) Sacks et al, Statistical Science, 1989 Intuition: • Model correlation between Z(x) and Z(x’) for any two input vectors x and x’ • x close to x’ – correlation large • x far from x’ – correlation small • Leads to a distribution of Y(x) at any x given the Y’s at the design points Perform diagnostic tests on model • Accuracy of prediction? Standard error of prediction? David Rumpf, Statistician GE Aircraft Engines Page 13 Accuracy comparison (e.g., Notch-Alt or Y1b) Error = Y – fitted Y Fitted Y is leave-one-out cross validated (take observation out and predict it) RMSE = 1/n sum (Y – fitted Y)2 -------------------------------------------------Approximating Cross-Validated Model RMSE --------------------------------------------------Polynomial – 2nd degree 0.71 Polynomial – 3rd degree 1.04 Random function 0.48 3rd degree polynomial fits even worse than 2nd degree! David Rumpf, Statistician GE Aircraft Engines Page 14 Diagnostic Checking of Random-Function Model Accuracy assessment: Plot Y versus fitted Y Standard error (se) assessment: Plot (Y – fitted Y) / se(fitted Y) versus fitted Y (Fitted Y is leave-one-out cross validation) David Rumpf, Statistician GE Aircraft Engines Page 15 Visualization of Input-Output Relationships e.g. Y1b as a function of RetArm Other two inputs (CPRabbet and CPDrop) averaged out David Rumpf, Statistician GE Aircraft Engines Page 16 Propagating Variation Through the Random Function Model CPRabbet, CPDrop, RetArm have independent N(mu, sigma) distributions e.g., set mu = center of range Sample CPRabbet, CPDrop, RetArm and pass through model to get a distribution of e.g., Y1b values David Rumpf, Statistician GE Aircraft Engines Page 17 Conclusions RSM approach: • Good starting point • Will work fine for Ybar and simple underlying Physics Space Filling Design: • Allows us to model responses with very nonlinear underlying Physics Random-function model: • Provides valid standard errors of prediction • Can adapt to nonlinearities in data • Fast, so can quickly propagate variation inputs => outputs via Monte Carlo David Rumpf, Statistician GE Aircraft Engines Page 18