MIT OpenCourseWare http://ocw.mit.edu 2.830J / 6.780J / ESD.63J Control of Manufacturing Processes (SMA 6303) Spring 2008 For information about citing these materials or our Terms of Use, visit: http://ocw.mit.edu/terms. Control of Manufacturing Processes Subject 2.830/6.780/ESD.63 Spring 2008 Lecture #8 Process Capability & Alternative SPC Methods March 4, 2008 Manufacturing 1 Agenda • Control Chart Review – hypothesis tests: α, β and n – control charts: α, β, n, and average run length (ARL) • Process Capability • Advanced Control Chart Concepts Manufacturing 2 Average Run Length • How often will the data exceed the ±3σ limits if Δμx = 0? Prob(x > μ x + 3σ x ) + Pr ob(x < μ x − 3σ x ) = 3 / 1000 0.45 0.4 0.35 0.3 0.25 0.2 0.15 0.1 0.05 -4 Manufacturing -3 −3σ 0 -2 -1 μ0 1 2 3 +3σ 4 3 Detecting Mean Shifts: Chart Sensitivity • Consider a real shift of Δμx: 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 Sample Number • How many samples before we can expect to detect the shift on the xbar chart? Manufacturing 4 Average Run Length • How often will the data exceed the ±3σ limits if Δμx = +1σ? Prob(x > μ x + 2σ x ) + Prob(x < μ x − 4σ x ) = 0.023 + 0.001 = 24 / 1000 0.45 0.45 Actual 0.4 0.4 0.35 0.35 Assumed Distribution 0.3 0.3 Distribution 0.25 0.25 -4 Manufacturing −3σ -4 -3 0.2 0.2 0.15 0.15 0.1 Δμ 0.1 0.05 0.05 0 -3 -2 -2 -1 μ -1 0 pe 0 01 12 23 +3σ 34 4 5 Definition • Average Run Length (arl): Number of runs (or samples) before we can expect a limit to be exceeded = 1/pe – for Δμ = 0 – for Δμ = 1σ arl = 3/1000 = 333 samples arl = 24/1000 = 42 samples Even with a mean shift as large as 1σ, it could take 42 samples before we know it!!! Manufacturing 6 Effect of Sample Size n on ARL • Assume the same Δμ = 1σ – Note that Δμ is an absolute value • If we increase n, the Variance of xbar σx decreases: σx = n • So our ± 3σ limits move closer together Manufacturing 7 ARL Example Original Distribution -4 −3σ -3 0.450.45 0.45 0.40.4 0.4 0.350.35 0.35 0.30.3 0.3 0.250.25 0.25 0.20.2 0.2 0.150.15 0.15 0.10.1 Δμ 0.1 0.050.05 0.05 00 -2 -4 new limits -3 -1 -4 -2 −3σ∗ -3 -1 μ 0-2 New Distribution pe 0 1-1 1 20 31 +3σ∗ 422 3 34 +3σ 4 same absolute shift As n increases pe increases so ARL decreases Manufacturing 8 Another Use of the Statistical Process Model: The Manufacturing -Design Interface • We now have an empirical model of the process 0.45 0.4 0.35 0.3 How “good” is the process? 0.25 0.2 0.15 0.1 0.05 Is it capable of producing what we need? Manufacturing 0 -4 -3 −3σ -2 -1 0 μ 1 2 3 4 +3σ 9 Process Capability • Assume Process is In-control • Described fully by xbar and s • Compare to Design Specifications – Tolerances – Quality Loss Manufacturing 10 Design Specifications • Tolerances: Upper and Lower Limits Characteristic Dimension Target Lower x* Specification Limit Upper Specification Limit LSL USL Manufacturing 11 Design Specifications • Quality Loss: Penalty for Any Deviation from Target QLF = L*(x-x*)2 How to Calibrate? x*=target Manufacturing 12 Use of Tolerances: Process Capability • Define Process using a Normal Distribution • Superimpose x*, LSL and USL • Evaluate Expected Performance 0.45 0.4 0.35 0.3 0.25 0.2 0.15 0.1 LSL -4 -3 −3σ Manufacturing 0.05 x* 0 -2 -1 μ0 1 USL 2 3 +3σ 4 13 Process Capability • Definitions (USL − LSL) tolerance range Cp = = 6σ 99.97% confidence range • Compares ranges only • No effect of a mean shift Manufacturing 14 Process Capability: Cpk C pk ⎛ (USL − μ ) (LSL − μ) ⎞ = min ⎝ , ⎠ 3σ 3σ = Minimum of the normalized deviation from the mean • Compares effect of offsets Manufacturing 15 Cp = 1; Cpk = 1 0.45 0.4 0.35 0.3 0.25 0.2 0.15 0.1 0.05 0 -4 Manufacturing -3 -2 -1 0 1 2 3 4 16 Cp = 1; Cpk = 0 0.45 0.4 0.35 0.3 0.25 0.2 0.15 0.1 0.05 0 -4 -3 -2 Manufacturing -1 0 1 2 3 4 17 Cp = 2; Cpk = 1 0.45 0.4 0.35 0.3 0.25 0.2 0.15 0.1 0.05 0 -4 Manufacturing -3 -2 -1 0 1 2 3 4 18 Cp = 2; Cpk = 2 0.45 0.4 0.35 0.3 0.25 0.2 0.15 0.1 0.05 0 -4 Manufacturing -3 -2 -1 0 1 2 3 4 19 Effect of Changes • In Design Specs • In Process Mean • In Process Variance • What are good values of Cp and Cpk? Manufacturing 20 Cpk Table Manufacturing Cpk z P<LS or P>USL 1 3 1E-03 1.33 4 3E-05 1.67 5 3E-07 2 6 1E-09 21 The “6 Sigma” problem P(x > 6σ) = 18.8x10-10 Cp=2 0.45 Cpk=2 0.4 0.35 0.3 0.25 0.2 0.15 0.1 0.05 -4 LSL −3σ∗ -3 -2 0 -1 0 1 +3σ∗ 2 3 4 USL 6σ Manufacturing 22 The 6 σ problem: Mean Shifts P(x>4σ) = Cp=2 31.6x10-6 0.45 Cpk=4/3 0.4 Even with a mean shift of 2σ we have only 32 ppm out of spec 0.35 0.3 0.25 0.2 0.15 0.1 0.05 0 -4 -3 -2 -1 0 1 LSL 2 3 4 USL 4σ Manufacturing 23 Capability from the Quality Loss Function QLF = L(x) =k*(x-x*)2 x* Given L(x) and p(x) what is E{L(x)}? Manufacturing 24 Expected Quality Loss E{L(x)} = E[k(x − x*) 2 ] = k [E(x ) − 2 E(xx*) + E(x * )] 2 2 = kσ + k( μx − x*) 2 x Penalizes Variation Manufacturing 2 Penalizes Deviation 25 Process Capability • • • • • The reality (the process statistics) The requirements (the design specs) Cp - a measure of variance vs. tolerance Cpk - a measure of variance from target Expected Loss - an overall measure of goodness Manufacturing 26 Xbar Chart Recap • xbar - S (or R) charts – plot of sequential sample statistics – compare to assumptions • normal • stationary • Interpretation – hypothesis tests on μ and σ – confidence intervals – “randomness” • Application – Real-time decision making Manufacturing 27 Real-Time 1.8 1.6 1.4 1.2 1 0.8 0.6 0.4 0.2 0 1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 Sample Number Manufacturing 28 Beyond Xbar • Good Points – Simple and “transparent” – Enforces Assumptions • Normality (via Central Limit) • Independent (via long sampling times) • Limitations – n>1 to get Xbar and S – ARL is typically large • Not very sensitive to small changes – Slow time response Manufacturing 29 Beyond Xbar • What if n=1? – Have a Lot of Data – Want Fast Response to Changes • How to Compute Control Chart Statistics? – Running Chart and Running Variance? – Running Average and Running Variance? – Running Average with Forgetting Factor • How to Increase Sensitivity to Small, Persistent Mean Shift? – Integrate the Error Manufacturing 30 Chart Design: n=1 Designs - Running Averages • Sensitivity: Ability to detect small changes (e.g. mean shifts) • Time Response: Ability to Catch Changes Quickly • Noise Rejection?: Higher Variance Manufacturing 31 Xbar “Filtering” 1.4 1.2 1 0.8 0.6 Run Data Xbar n=4 0.4 0.2 0 1 6 11 16 21 26 31 36 41 46 51 56 61 66 71 76 81 86 91 96 -0.2 -0.4Manufacturing 32 Filtering • Reduced Peaks • Hides intermediate data • Reduces the “frequency content” of the output Manufacturing 33 Independence and Correlation • Independence: Current output does not depend on prior • Correlation: Measure of Independence – e.g. auto correlation function Rxx (τ ) = E[x(t)x(t + τ )] Manufacturing 34 Correlation Rxx (τ ) = E[x(t)x(t + τ )] For a linear 1st order system For an uncorrelated τ~ 1 sec: process 1 0.8 0.6 0.4 0.2 0 -4 Manufacturing -3 Tmin -2 -1 0 1 2 3 Tmax 4 35 Sampling: Frequency and Distribution of Samples 1 1 0.8 0.8 0.8 0.6 0.6 0.6 0.4 0.4 0.4 0.2 0.2 0.2 1 0 -4 -3 Tmin -2 0 0 -3 -2 012341 2 31 4 0-1 -4-1-3 -2 -1-4 T Tmin Tm Tmax 2 3 Tmax 4 SAMPLE TIME Manufacturing 36 Correlation and Sampling Correlated Samples Uncorrelated Samples Correlation Time (e.g.) • Taking samples beyond correlation time guarantees independence Manufacturing 37 Sampling and Averaging • Sampling Frequency Affects – Time Response – Correlation • Averaging – Filters Data – Slows Response Manufacturing 38 Alternative Charts: Running Averages • More averages/Data • Can use run data alone and average for S only • Can use to improve resolution of mean shift j+n xRj n measurements at sample j Manufacturing SR j 2 1 = ∑ xi n i=j Running Average 1 j +n 2 = (x − x ) ∑ Rj Running Variance n − 1 i= j i 39 Specific Case: Weighted Averages y j = a1 x j −1 + a2 x j −2 + a3 x j −3 + ... • How should we weight measurements?? – All equally? (as with Running Average) – Based on how recent? • e.g. Most recent are more relevant than less recent? Manufacturing 40 Consider an Exponential Weighted Average 0.25 Define a weighting function Wt − i = r (1 − r ) 0.2 i 0.15 Exponential Weights 0.1 0.05 0 1 Manufacturing 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 41 Exponentially Weighted Moving Average: (EWMA) Ai = rxi + (1 − r)Ai −1 Recursive EWMA ⎛ σ x2 ⎞ ⎛ r ⎞ 2t σA = ⎜ ( ) 1 − 1 − r ⎟⎝ ⎝ n ⎠ 2 − r⎠ [ ] σA = UCL, LCL = x ± 3σ A Manufacturing time σx ⎛ r ⎞ 2 n ⎝ 2 − r⎠ for large t 42 Effect of r on σ multiplier 0.9 0.8 plot of (r/(2-r)) vs. r 0.7 0.6 0.5 0.4 wider control limits 0.3 0.2 0.1 0 0 Manufacturing 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 r 43 SO WHAT? • The variance will be less than with xbar, σA = σx n ⎛ r ⎞ ⎝ 2 − r⎠ = σx ⎛ r ⎞ ⎝ 2 − r⎠ • n=1 case is valid • If r=1 we have “unfiltered” data – Run data stays run data – Sequential averages remain • If r<<1 we get long weighting and long delays – “Stronger” filter; longer response time Manufacturing 44 EWMA vs. Xbar 1 0.9 r=0.3 0.8 Δμ = 0.5 σ 0.7 xbar EWMA UCL EWMA LCL EWMA grand mean UCL LCL 0.6 0.5 0.4 0.3 0.2 0.1 0 0 Manufacturing 50 100 150 200 250 300 45 Mean Shift Sensitivity EWMA and Xbar comparison 1.2 xbar 1 EWMA UCL EWMA 0.8 LCL EWMA 0.6 Grand Mean 3/6/03 UCL 0.4 LCL Mean shift = .5 σ 0.2 n=5 r=0.1 0 1 5 Manufacturing 9 13 17 21 25 29 33 37 41 45 49 46 Effect of r 1 0.9 xbar 0.8 EWMA 0.7 UCL EWMA 0.6 LCL EWMA 0.5 Grand Mean 0.4 UCL 0.3 LCL 0.2 0.1 r=0.3 0 1 3 Manufacturing 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 47 Small Mean Shifts • What if Δμx is small wrt σx ? • But it is “persistent” • How could we detect? – ARL for xbar would be too large Manufacturing 48 Another Approach: Cumulative Sums • Add up deviations from mean – A Discrete Time Integrator j C j = ∑ (x i − x) i=1 • Since E{x-μ}=0 this sum should stay near zero • Any bias in x will show as a trend Manufacturing 49 Mean Shift Sensitivity: CUSUM 8 7 t Ci = ∑ (x i − x ) 6 i =1 5 4 Mean shift = 1σ 3 Slope cause by mean shift Δμ 2 1 49 47 45 43 41 39 37 35 33 31 29 27 25 23 21 19 17 15 13 11 9 7 5 3 1 0 -1 Manufacturing 50 Control Limits for CUSUM • Significance of Slope Changes? – Detecting Mean Shifts • Use of v-mask – Slope Test with Deadband Upper decision line ⎛1 − β⎞ d = ln⎜ ⎟ δ ⎝ α ⎠ 2 δ= θ Δx σx ⎛ Δx ⎞ θ = tan ⎜ ⎟ ⎝ 2k ⎠ −1 d Lower decision line Manufacturing where k = horizontal scale factor for plot 51 Use of Mask 8 7 6 5 θ=tan-1(Δμ/2k) k=4:1; Δμ=0.25 (1σ) tan(θ) = 0.5 as plotted 4 3 2 1 49 47 45 43 41 39 37 35 33 31 29 27 25 23 21 19 17 15 13 11 9 7 5 3 1 0 -1 Manufacturing 52 An Alternative • Define the Normalized Statistic Zi = Xi − μ x σx • And the CUSUM statistic Which has an expected mean of 0 and variance of 1 t Si = ∑Z i =1 t i Which has an expected mean of 0 and variance of 1 Chart with Centerline =0 and Limits = ±3 Manufacturing 53 Example for Mean Shift = 1σ 5 Normalized CUSUM 4 3 2 Mean Shift = 1 47 45 43 41 39 37 35 33 31 29 27 25 23 21 19 17 15 13 11 9 7 5 3 0 1 1 σ -1 Manufacturing 54 Tabular CUSUM • Create Threshold Variables: Ci = max[0, xi − ( μ 0 + K ) + Ci −1 ] Accumulates + + Ci = max[0,( μ 0 − K ) − xi + − deviations Ci −1 ] from the mean − K= threshold or slack value for accumulation Δμ K= 2 typical Δμ = mean shift to detect H : alarm level (typically 5σ) Manufacturing 55 Threshold Plot 6 μ 0.495 σ 0.170 k=δμ/2 0.049 h=5σ 0.848 5 C+ C- 4 C+ CH threshold 3 2 1 0 1 3 Manufacturing 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 56 Alternative Charts Summary • Noisy Data Need Some Filtering • Sampling Strategy Can Guarantee Independence • Linear Discrete Filters have Been Proposed – EWMA – Running Integrator • Choice Depends on Nature of Process Manufacturing 57 Summary of SPC • Consider Process a Random Process – Can never predict precise value • Model with P(x) or p(x) – Assume p(x,t) = p(x) • Shewhart Hypothesis – In-control = purely random output • Normal, independent stationary • “The best you can do!” – Not in-control • Non-random behavior • Source can be found and eliminated Manufacturing 58 The SPC Hypothesis 0.9 0.8 0.7 Process 0.6 Y 0.5 0.4 0.3 0.2 0.1 0 1 2 3 4 5 6 7 8 9 10 11 Sample Number 12 13 14 15 16 17 18 19 20 In-Control ... ... p(y) Not In-Control Manufacturing 59