Six Sigma & Statistical Process Control — Engineering Reference
1. At a glance
Six Sigma is a data-driven, defect-reduction methodology that treats every business process as a statistical system whose output variation must be characterised, attacked at root cause, and held in a state of statistical control. The numerical goal — 6σ from the spec mean to the nearest spec limit — translates (with Motorola’s 1.5σ long-term drift allowance) to 3.4 DPMO (defects per million opportunities) or 99.99966 % yield.
- Birth: Motorola 1986, formalised by Bill Smith and Mikel Harry; trademark filed 1993.
- Productisation: Jack Welch made it the operating system of GE in 1996; cited US$10 B in savings over the next five years in the 2000 GE annual report.
- Frameworks:
- DMAIC (Define-Measure-Analyze-Improve-Control) — improve existing process.
- DMADV (V = Validate, sometimes “Verify”) — design new product or process to Six Sigma capability (a.k.a. DFSS, Design For Six Sigma).
- Lean Six Sigma (LSS) merges the Toyota Production System’s waste-elimination toolkit (see
[[Engineering/lean-manufacturing]]) with Motorola’s statistical toolkit; ~80 % of Fortune 500 Six Sigma programs are now LSS. - Output: a portfolio of chartered projects, each producing a quantified financial benefit signed off by Finance.
2. Why it matters
Statistical rigor moves quality from craft (“the line is running well today”) to science (“σ_short-term = 0.10 mm, C_pk = 1.5, in-control”). ASQ estimates ≥ US$1 T cumulative savings across global Fortune 1000 companies by 2025. Applications now reach far beyond discrete manufacturing:
- Services: call-centre handle time, error rates, abandoned-call ratio.
- Healthcare: medication error rate, ED wait times, surgical site infections.
- Finance: trade-settlement defects, AML false-positive rate, loan-origination cycle.
- Software: defect density, escaped-defect rate, MTTR on production incidents.
- Pharma / regulated: ICH Q9 quality risk management, CPV (continued process verification) under FDA Process Validation Stage 3.
It is the quantitative complement to lean’s qualitative waste removal: lean cuts time and motion, Six Sigma cuts variation. The two together are why Toyota, Motorola, and GE Aviation became operational benchmarks. The discipline has waned in marketing prominence since ~2015 (replaced rhetorically by agile, DevOps, data science), but persists as a hard requirement in automotive (IATF 16949), aerospace (AS9100), medical devices (ISO 13485), and pharma (FDA cGMP).
3. First principles
3.1 Variation is the enemy
Every measurable output has variation; the job is to (a) characterise it, (b) decide if it is common cause (inherent to the system) or special cause (assignable), and (c) act accordingly. Acting on common cause as if it were special — Deming’s tampering — increases variance.
3.2 σ-level → DPMO mapping (Motorola 1.5σ shift convention)
| σ-level (short-term) | Long-term Z (with 1.5σ shift) | DPMO | Yield |
|---|---|---|---|
| 1 σ | -0.5 | 691 462 | 30.85 % |
| 2 σ | 0.5 | 308 538 | 69.15 % |
| 3 σ | 1.5 | 66 807 | 93.32 % |
| 4 σ | 2.5 | 6 210 | 99.379 % |
| 5 σ | 3.5 | 233 | 99.9767 % |
| 6 σ | 4.5 | 3.4 | 99.99966 % |
| 7 σ | 5.5 | 0.019 | 99.9999981 % |
Without the 1.5σ shift, true two-sided 6σ would be 2 × Φ(-6) ≈ 0.002 DPMO. The shift accounts for empirical long-term mean drift Motorola observed across hundreds of processes.
3.3 Capability and performance indices
- Process capability (short-term, within-subgroup σ̂ from R̄/d₂ or s̄/c₄):
- C_p = (USL − LSL) / (6σ̂)
- C_pk = min[(μ̂ − LSL)/(3σ̂), (USL − μ̂)/(3σ̂)]
- Process performance (long-term, overall s):
- P_p = (USL − LSL) / (6s)
- P_pk = min[(x̄ − LSL)/(3s), (USL − x̄)/(3s)]
- Interpretation thresholds (AIAG / industry consensus):
- C_pk < 1.00 — incapable
- 1.00–1.33 — marginal (1.33 is the IATF 16949 minimum for new processes)
- 1.33–1.67 — capable
- ≥ 1.67 — highly capable
- ≥ 2.00 — Six-Sigma capable (matches C_p = 2.0 i.e. spec = ±6σ)
- C_p assumes centred + normal; C_pk penalises off-centring; C_pm (Taguchi) penalises deviation from target T.
3.4 Control limits vs. specification limits
Control limits = voice of the process (±3σ̂ of the plotted statistic). Spec limits = voice of the customer (engineering tolerance). Confusing them is the most common rookie mistake — control limits move with the process, spec limits do not.
3.5 Distributional assumptions
x̄–R, x̄–s, and capability formulas assume the underlying distribution is approximately normal. Non-normal data needs (a) a transform (Box-Cox 1964 or Johnson family), (b) a non-normal capability metric (ISO 22514-4 percentile method: C_pk = min[(x̄ − LSL)/(x̄ − x_0.00135), …]), or (c) a non-parametric chart.
4. DMAIC — phase-by-phase
| Phase | Goal | Core tools | Deliverable |
|---|---|---|---|
| Define | Frame the problem in $ and customer impact | Project charter, SIPOC, VOC, CTQ tree, Kano model, stakeholder map | Signed charter with Y = f(X) hypothesis |
| Measure | Quantify current state | VSM (lean overlap), data-collection plan, operational definitions, Gauge R&R / MSA, baseline C_pk | Validated measurement system, baseline σ-level |
| Analyze | Find root causes | Pareto, Ishikawa, 5-Whys, FMEA, regression, hypothesis tests (t, ANOVA, χ², proportion), graphical EDA | Statistically confirmed vital few X’s |
| Improve | Develop, pilot, prove fixes | DOE (full / fractional / RSM), pilot study, kaizen burst, SMED, poka-yoke, 5S, kanban (lean tools) | Optimised setpoints, validated improvement, pilot evidence |
| Control | Hold the gains | SPC, control plan, mistake-proofing, training, audit cadence, response plan | Control plan, sustained C_pk ≥ target |
Typical Black-Belt project: 3–6 months elapsed, 60–80 % of one BB’s time, US$50 k–1 M hard-savings. Master Black Belt mentors 5–10 BBs and ratifies financial claims with Finance.
5. Belts and roles
| Role | Training | Typical commitment | Certification (representative) |
|---|---|---|---|
| White Belt | 4–8 h | Awareness only | Internal |
| Yellow Belt | 1–2 weeks | Team member, SME | ASQ CSSYB, IASSC ICYB |
| Green Belt | 2–4 weeks | Leads small project | ASQ CSSGB, IASSC ICGB |
| Black Belt | 4–5 weeks | Full-time on projects | ASQ CSSBB, IASSC ICBB, SSGI BB |
| Master Black Belt | + 1–2 yr OJT | Mentors BBs, trains | ASQ CMBB (≥ 5 yr BB experience required) |
| Champion / Sponsor | Exec briefing | Removes blockers | None — VP/Director level |
Certification bodies: ASQ (American Society for Quality, most recognised), IASSC (International Association for Six Sigma Certification, exam-only), CSSC (Council for Six Sigma Certification), SSGI (Six Sigma Global Institute). University-affiliated programs: Villanova, Purdue, Cal Poly. ASQ exams require documented project portfolio for BB and MBB.
6. Statistical Process Control (SPC) charts
Walter Shewhart, Bell Labs, invented the control chart on 16 May 1924 — the founding artifact of SPC. The ±3σ limits balance Type I (false alarm, ~0.27 % per point) against Type II (miss) error.
6.1 Variables charts (continuous data)
| Chart | Subgroup n | Plot | When to use |
|---|---|---|---|
| x̄ + R | 2–10 | mean, range | Most common; n small |
| x̄ + s | > 10 | mean, std dev | n large; s is more efficient than R |
| I-MR | 1 | individual, MR | Low-volume, slow processes, destructive tests |
| EWMA | any | exp. weighted mean | Small persistent shifts (0.5–1.5 σ) |
| CUSUM | any | cumulative sum | Small shifts; faster than Shewhart at < 1.5σ |
| Hotelling T² | any | multivariate D² | ≥ 2 correlated CTQs |
6.2 Attribute charts (count / proportion data)
| Chart | Statistic | n | Distribution |
|---|---|---|---|
| p | proportion defective | variable | Binomial |
| np | count defective | constant | Binomial |
| c | count of defects per unit | constant | Poisson |
| u | defects per unit | variable | Poisson |
6.3 Pattern rules (special-cause signals)
Western Electric Handbook (1956) core four rules:
- One point outside ±3σ.
- Two of three consecutive points beyond ±2σ (same side).
- Four of five consecutive points beyond ±1σ (same side).
- Eight consecutive points on the same side of the centerline.
Nelson rules (1984) extend to eight tests including trend (6 increasing/decreasing), oscillation (14 alternating), and stratification (15 within ±1σ). Minitab and JMP implement Nelson by default. Each added rule raises the false-alarm rate — combining all eight gives Type I ~1 % per point; pick a subset, don’t enable everything.
6.4 Out-of-control response
Every chart needs a documented OCAP (Out-of-Control Action Plan): who is alerted, who investigates, what is paused, where the disposition is logged. Without OCAP the chart is wallpaper.
7. Design of Experiments (DOE)
DOE is the multivariate, statistically efficient alternative to OFAT (one-factor-at-a-time), which misses interactions and wastes runs. Fisher (1925, agricultural plots) is the origin; Box, Hunter, and Hunter (1978/2005) the modern engineering reference.
| Design | Runs (k factors) | What it gives you | When |
|---|---|---|---|
| Full factorial 2^k | 2^k | All main effects + all interactions | k ≤ 5; characterisation |
| Fractional factorial 2^(k-p) | 2^(k-p) | Aliased subset; resolution III/IV/V | Screening at k ≥ 5 |
| Plackett-Burman | 8, 12, 16, 20… | Main effects only (Res III) | Screening many factors cheaply |
| 3^k full / Box-Behnken | 3^k / 13–27 | Quadratic effects, no axial points | Modest curvature, no extreme corners |
| Central Composite (CCD) | 2^k + 2k + n_c | Full second-order RSM | Optimisation near the optimum |
| Taguchi L_n orthogonal arrays | per L_n table | Robust design via S/N ratio | Noise-factor handling, manufacturing |
| Mixture (Scheffé) | varies | Components Σ = 1 (formulation) | Alloys, blends, pharma formulations |
| D-optimal / I-optimal | custom | Min | X’X |
Resolution (Roman numeral) tells what is aliased: III aliases main with 2-fi; IV aliases 2-fi with 2-fi; V is clear up through 2-fi with 3-fi. Production engineering rarely needs Resolution > V.
Sequential strategy: screening (Res III) → fold-over or augment to Res IV → characterisation (full or Res V) → RSM around the optimum → confirmation runs (always; 3–5 replicates at the optimum to verify the model’s prediction).
Randomisation, replication, blocking — Fisher’s three principles. Skipping randomisation lets time confound your factor effects (oven drift, tool wear, ambient temperature). Block on known nuisance variables (operator, day, batch of raw material).
8. Worked examples (with units)
Example A — Process capability of a machined diameter
- Drawing: Ø10.0 mm, USL = 10.5 mm, LSL = 9.5 mm.
- Measurements from a 25-subgroup x̄ + R chart, n = 5:
- x̄̄ = 10.05 mm, R̄ = 0.232 mm, d₂(n=5) = 2.326 → σ̂ = R̄/d₂ = 0.0998 mm ≈ 0.10 mm.
- C_p = (10.5 − 9.5)/(6 × 0.10) = 1.0/0.6 = 1.67 (capable potential).
- C_pk = min[(10.05 − 9.5)/0.30, (10.5 − 10.05)/0.30] = min[1.83, 1.50] = 1.50.
- The off-centring (μ̂ = 10.05 vs target 10.0) costs 0.17 in C_pk. Centring would restore C_pk = C_p = 1.67.
- Short-term Z to nearest spec = 3 × C_pk = 4.5σ. Long-term DPMO (with 1.5σ shift) ≈ Φ(-3.0) × 10⁶ ≈ 1 350 DPMO. Without shift, ≈ Φ(-4.5) × 10⁶ ≈ 3.4 DPMO.
Example B — Variable Gauge R&R (ANOVA method)
Study: 10 parts × 3 operators × 3 trials = 90 measurements, randomised order.
ANOVA variance decomposition:
- σ²_repeatability = MS_error
- σ²_reproducibility = (MS_operator − MS_op×part)/(n_parts × n_trials)
- σ²_interaction = (MS_op×part − MS_error)/n_trials
- σ²_part = (MS_part − MS_op×part)/(n_operators × n_trials)
- σ²_GRR = σ²_repeat + σ²_reprod (+ σ²_interaction if significant)
- σ²_total = σ²_GRR + σ²_part
%GR&R = √(σ²_GRR) / √(σ²_total) × 100 % (study variation) or × 5.15 σ_GRR / Tolerance (tolerance-based).
AIAG MSA 4th ed acceptance:
- < 10 % — gauge acceptable.
- 10–30 % — conditionally acceptable (cost / criticality dependent).
- ≥ 30 % — unacceptable, reject.
Number of Distinct Categories (NDC) = ⌊1.41 × σ_part / σ_GRR⌋. Must be ≥ 5 to resolve part-to-part variation.
Worked numbers:
- σ̂_repeat = 0.020 mm, σ̂_reprod = 0.010 mm → σ_GRR = √(0.020² + 0.010²) = 0.0224 mm.
- σ̂_part = 0.080 mm → σ_total = √(0.0224² + 0.080²) = 0.0831 mm.
- %GR&R = 0.0224 / 0.0831 = 26.9 % → conditionally acceptable.
- NDC = ⌊1.41 × 0.080 / 0.0224⌋ = ⌊5.04⌋ = 5 → just adequate.
Example C — 2³ full factorial DOE on injection-mould scrap rate
Factors: A = melt temperature (220 / 250 °C), B = hold pressure (60 / 90 MPa), C = cool time (10 / 20 s). Response y = scrap rate (% defective parts per 100 cycles). 8 runs, 2 replicates → 16 cycles, fully randomised.
Effect estimate = (mean at ”+”) − (mean at ”−”). ANOVA F-test: MS_effect / MSE, df = 1, df_error = 8 (for 2³ × 2 reps).
| Effect | Estimate (%) | F | p-value |
|---|---|---|---|
| A | −2.10 | 18.4 | 0.003 |
| B | −1.30 | 7.1 | 0.029 |
| C | −0.40 | 0.7 | 0.43 |
| A·B | +0.85 | 3.0 | 0.12 |
| A·C | −0.10 | 0.04 | 0.84 |
| B·C | +0.05 | 0.01 | 0.92 |
| A·B·C | +0.05 | 0.01 | 0.92 |
Significant at α = 0.05: A and B. Reduced model:
ŷ = 4.20 − 1.05·x_A − 0.65·x_B (x in coded ±1 units)
Predicted optimum (A = +1, B = +1, C = ‑1 since insignificant): ŷ = 4.20 − 1.05 − 0.65 = 2.50 % scrap. Confirmation runs: 5 cycles at the optimum gave x̄ = 2.62 %, s = 0.30 %; 95 % CI [2.25 %, 2.99 %] contains the prediction → model validated.
9. Lean Six Sigma combined
LSS overlays DMAIC structure on lean’s TIMWOODS waste taxonomy (Transport-Inventory-Motion-Waiting-Overproduction-Overprocessing-Defects-Skills; see [[Engineering/lean-manufacturing]]):
- Measure phase incorporates Value-Stream Mapping; VA (value-add) vs NVA (non-value-add) time analysis tags every step.
- Analyze phase applies hypothesis testing to confirm which lean wastes are causal, not just visible.
- Improve phase alternates kaizen events (rapid, qualitative) with DOE (slower, quantitative).
- Control phase uses SPC, daily-management visual boards, and lean’s tier-1/2/3 management cadence.
Track record:
- GE Aviation — engine assembly cycle time reduced ~40 % with first-pass yield improved (Welch’s 2000 letter to shareholders).
- Bank of America — transactional defects reduced ~88 % over 2001–2005 (HBR case).
- Caterpillar — adopted “6 Sigma” division-wide 2001; reported >US$2 B benefits by 2003.
- Toyota — does not badge its work as Six Sigma; the same statistical tools (Genchi Genbutsu + jidoka) predate it inside TPS.
Caveats: transactional / service metrics have heavier non-normality, smaller samples, and harder operational definitions than discrete manufacturing. Don’t force x̄–R on call-handle time; use I-MR with a log transform or use a survival model.
10. MSA — Measurement System Analysis
You cannot improve what you cannot measure repeatably. AIAG MSA Reference Manual 4th ed (2010) (5th ed in development as AIAG-VDA harmonised) defines:
| Property | What it measures | Study type |
|---|---|---|
| Resolution | Smallest detectable increment | Rule of 10 vs tolerance |
| Bias | Mean error vs reference standard | Type 1 study; t-test |
| Linearity | Bias as a function of measurand magnitude | Regression across range |
| Stability | Drift over time | X̄ + R chart of a standard |
| Repeatability (EV) | Same operator, same part, repeated reads | GR&R / Type 2 |
| Reproducibility (AV) | Different operators, same part | GR&R / Type 2 |
| Interaction | Operator × part | ANOVA GR&R |
| Attribute agreement | For pass/fail or category gauges | Kappa (Cohen, Fleiss) |
Type 1 — single operator, single part, ≥ 50 reads — bias + repeatability only; gauge-acceptance gate before full GR&R. Type 2 — full crossed GR&R, ANOVA or X̄-R methods. Type 3 — like Type 2 but excluding operator (automated gauge). Attribute Agreement Analysis — for go/no-go gauges or human visual inspection; report Kappa (within, between, vs reference). Kappa ≥ 0.75 generally acceptable.
11. Edge cases and gotchas
- Non-normal data: x̄–R limits still robust by CLT for n ≥ 4; individuals charts are not robust. Transform (Box-Cox 1964 λ, Johnson SB/SU/SL) or switch to non-normal capability (ISO 22514-4 percentile method).
- Small samples: Shewhart limits unreliable with < 25 subgroups; use Bayesian methods or wide initial limits then tighten.
- Autocorrelated data (chemical process, continuous flow, slow sensors): traditional SPC false-alarms because successive points are not independent. Fit ARIMA, plot residuals on a Shewhart chart, or use EWMA with a re-tuned λ.
- Multivariate processes with correlated CTQs: running n univariate charts inflates Type I (1 − (1 − 0.0027)^n); use Hotelling T² for in-control monitoring and MEWMA for shift detection.
- Short-run / low-volume: standardised x̄–R (Z transform per part number), DNOM (deviation from nominal), or QS-9000 short-run procedures.
- Tampering (Deming’s funnel experiment, Out of the Crisis 1986): adjusting setpoint after every measurement amplifies variance from σ² to 2σ². Rule 1 (leave it alone) wins for common-cause variation.
- 1.5 σ shift (Motorola convention): contentious; some authors (Wheeler) argue it conflates short- and long-term capability. Quote both: “Z_short = 6.0, Z_long with 1.5σ shift = 4.5, DPMO = 3.4.”
- C_p with bad C_pk: high C_p means potential capability; off-centre process can still ship defects. Always pair (C_p, C_pk) or (P_p, P_pk).
- Rational subgrouping (Shewhart 1931): subgroup such that within-subgroup variation captures only common cause. Pulling 5 consecutive parts then 5 again 1 h later mixes mean shifts into σ̂, inflating control limits and hiding special cause.
- DOE without randomisation: time-order trends (tool wear, ambient temperature) confound factor effects. Always randomise; block if randomisation is impractical.
- DOE without replication: estimates exist but residual df = 0 → no F-test possible. Use centre points or normal-probability plot of effects (Daniel 1959).
- Six-Sigma fatigue (~2010–2015): big-bang corporate programs lost momentum; replaced rhetorically by agile, DevOps, lean startup, AI/ML. Discipline persists in regulated industries (auto, aero, med-device, pharma).
- AI-augmented SPC: vendors like Falkonry, Seeq, Aveva PI Vision (formerly OSIsoft), Splunk Industrial overlay anomaly-detection ML on traditional SPC streams; useful for high-cardinality multivariate processes but does not replace fundamentals.
- “Six Sigma killed innovation at 3M” (BusinessWeek 2007 critique of McNerney): real risk — DMAIC optimises existing processes; DMADV / DFSS is needed when invention is the goal.
12. Tools and software
| Tier | Tool | Strength | Notes |
|---|---|---|---|
| Statistical (commercial) | Minitab | De-facto standard; taught in 90 % of BB programs | Per-user licence ~US$1.7 k/yr |
| JMP (SAS) | Interactive graphics, strong DOE platform | Popular in semiconductors, chemicals | |
| Design-Expert (Stat-Ease) | DOE-focused; best RSM workflow | Smaller install base | |
| Statistica / TIBCO Spotfire | Enterprise analytics with SPC modules | Often bundled | |
| Open-source | R: qcc, qicharts2, SixSigma, fitdistrplus, DoE.base, rsm, AlgDesign | All major SPC, MSA, DOE | Free; integrates with knitr/Quarto |
| Python: pyspc, doepy, statsmodels, scipy.stats, pingouin | Notebook workflow | Lighter SPC ecosystem than R | |
| Spreadsheet | Excel + Analysis ToolPak / SigmaXL add-in | Baseline; OK for charts, weak for DOE | Audit risk; not for regulated work |
| Industrial / OT | AVEVA PI (was OSIsoft), GE Proficy SmartSignal, Honeywell Experion PKS Quality, SAP MII Quality, Splunk Industrial Asset Intelligence, Falkonry, Seeq | SPC at MES / historian level | Connect to PLCs ([[Engineering/realtime-embedded]]) |
| Training | Villanova, Purdue COE, MoreSteam, GoLeanSixSigma, Lean Sensei, ASQ section courses | BB / MBB curricula | US$1 k–10 k/seat |
| Certification | ASQ (CSSGB / CSSBB / CMBB), IASSC, CSSC, SSGI | Recognised credentials | ASQ requires project portfolio for BB+ |
13. Cross-references
[[Engineering/reliability-engineering]]— companion: hazard rate, MTBF, Weibull, FMEA share statistical language with SPC.[[Engineering/lean-manufacturing]]— companion: waste taxonomy, VSM, kanban, SMED feed the LSS Improve phase.[[Engineering/ergonomics-human-factors]]— process design and operator variability.[[Engineering/system-identification]]— regression and ANOVA overlap with DOE model fitting.[[Engineering/fatigue-analysis]]— material variability and reliability of structural components.[[Engineering/realtime-embedded]]— data acquisition for SPC at the line.[[Engineering/supply-chain-management]]— supplier quality and PPAP/IATF 16949 gates.[[Engineering/chemical-process-fundamentals]]— DOE heritage in pharma and process chemistry.[[Languages/Tier3/scientific]]— modern statistical workflow (R, Python, notebooks).
14. Citations
Canonical textbooks:
- Montgomery, D.C. Introduction to Statistical Quality Control, 8th ed., Wiley, 2019.
- Montgomery, D.C. Design and Analysis of Experiments, 10th ed., Wiley, 2019.
- Box, G.E.P., Hunter, J.S., Hunter, W.G. Statistics for Experimenters, 2nd ed., Wiley, 2005.
- Pyzdek, T., Keller, P. The Six Sigma Handbook, 5th ed., McGraw-Hill, 2018.
- George, M.L. Lean Six Sigma: Combining Six Sigma with Lean Speed, McGraw-Hill, 2002.
- Harry, M., Schroeder, R. Six Sigma: The Breakthrough Management Strategy, Doubleday, 2000.
Foundational papers:
- Shewhart, W.A. memorandum, Bell Labs, 16 May 1924 (control chart origin); Economic Control of Quality of Manufactured Product, Van Nostrand, 1931.
- Western Electric Co. Statistical Quality Control Handbook, 1956 (WE rules).
- Nelson, L.S. “The Shewhart Control Chart — Tests for Special Causes.” J. Quality Technology 16(4), 1984.
- Box, G.E.P., Wilson, K.B. “On the Experimental Attainment of Optimum Conditions.” J.R. Stat. Soc. B 13(1), 1951 (RSM).
- Box, G.E.P., Cox, D.R. “An Analysis of Transformations.” J.R. Stat. Soc. B 26(2), 1964.
- Hotelling, H. “Multivariate Quality Control” in Techniques of Statistical Analysis, McGraw-Hill, 1947.
- Taguchi, G. Introduction to Quality Engineering, Asian Productivity Organization, 1986.
- Plackett, R.L., Burman, J.P. “The Design of Optimum Multifactorial Experiments.” Biometrika 33(4), 1946.
Standards:
- AIAG SPC Reference Manual, 2nd ed., 2005.
- AIAG MSA Reference Manual, 4th ed., 2010 (5th ed AIAG-VDA harmonised, in development).
- ISO 13053-1:2011 Quantitative methods in process improvement — Six Sigma — DMAIC methodology.
- ISO 13053-2:2011 — Tools and techniques.
- ISO 22514-1:2014 / -2:2017 / -4:2016 Statistical methods in process management — Capability and performance.
- ISO 7870-1/-2/-3/-4 (control charts).
- ISO 3534-1/-2/-3 (statistics — vocabulary and symbols).
- IATF 16949:2016 (automotive QMS, mandates SPC + MSA + PPAP).
- AS9100D:2016 (aerospace QMS).
- ISO 13485:2016 (medical-device QMS).
- FDA Guidance for Industry — Process Validation: General Principles and Practices, Jan 2011 (Stage 3 CPV).
Practitioner / historical:
- Deming, W.E. Out of the Crisis, MIT CAES, 1986 (System of Profound Knowledge, funnel experiment).
- Juran, J.M. Juran’s Quality Handbook, 7th ed., McGraw-Hill, 2017.
- Crosby, P.B. Quality Is Free, McGraw-Hill, 1979.
- Wheeler, D.J. Understanding Statistical Process Control, 3rd ed., SPC Press, 2010 (critic of the 1.5σ shift).
- Welch, J. GE Annual Report, 1996 and 2000 (Six Sigma as GE operating system).
- Smith, B. internal Motorola white paper, 1986 (origin of “Six Sigma” terminology).
- Hahn, G.J., Hill, W.J., Hoerl, R.W., Zinkgraf, S.A. “The Impact of Six Sigma Improvement.” The American Statistician 53(3), 1999.
- BusinessWeek, “At 3M, A Struggle Between Efficiency And Creativity,” 11 Jun 2007 (cited critique).
Tier 2 deep-dive reference. Maintained under the Engineering library schema ([[Engineering/_schema]]). Companion notes: [[Engineering/reliability-engineering]], [[Engineering/lean-manufacturing]].