Measure Phase
Purpose
Measure establishes a trustworthy baseline: can you actually see the process? If the measurement system is noisy, every later conclusion is built on sand.
Concepts
Data types. Continuous (time, weight, length, temperature) and discrete/attribute (pass/fail, defect counts, categories). Continuous data carries more information per sample, so attribute data usually needs far larger samples for the same confidence.
Measurement scales. Nominal (labels), ordinal (ranked, uneven intervals), interval (equal intervals, arbitrary zero), ratio (equal intervals, true zero). Only ratio supports statements like "twice as much."
MSA and Gage R&R. Repeatability is the same appraiser measuring the same part twice (equipment/precision); reproducibility is different appraisers getting different results (appraiser effect). Rule of thumb for %GRR: under 10% acceptable, 10–30% marginal, over 30% unacceptable - fix the system before collecting data.
Operational definitions. Every metric needs a written definition of what is counted, where, when, and by whom, or two shifts will record two different numbers.
Baseline capability. Cp = (USL − LSL) / 6σ measures spread only; Cpk = min[(USL − mean), (mean − LSL)] / 3σ measures spread and centering. A centered process has Cp = Cpk. Pp/Ppk use overall sigma; Cp/Cpk use within-subgroup sigma.
Worked example
Spec limits 4 to 10, mean 7, within-subgroup σ = 0.8. Cp = (10 − 4) / (6 × 0.8) = 6 / 4.8 = 1.25. Cpk = min(10−7, 7−4) / (3 × 0.8) = 3 / 2.4 = 1.25. Centered, so the two agree.
Traps
Treating accuracy and precision as the same thing, quoting Cp when the process is off-center, and collecting attribute data with tiny samples.