Improve
Generate, select, pilot, prove
Improve converts validated causes into countermeasures. Generate broadly (brainstorming, benchmarking, TRIZ), screen with a solution-selection or Pugh matrix weighted to CTQs and feasibility, then pilot before full rollout.
Design of experiments
DOE is the efficient way to learn about multiple factors at once. Screening designs (fractional factorials, Plackett-Burman) at resolution III or IV isolate the vital few; full factorials and response surface methods (central composite, Box-Behnken) model curvature and locate optimum settings. Randomize the run order, block known nuisance variables, and replicate to estimate error. Understand main effects, interactions, aliasing/confounding, and the difference between statistical and practical significance. Robust design uses orthogonal arrays and signal-to-noise ratios to reduce sensitivity to noise factors.
Lean and risk
Match tools to the waste observed: 5S, standard work, SMED, kanban/pull, TPM, cellular flow, kaizen events. Poka-yoke prevents errors at the source - design out first, then automate, then warn. FMEA prioritizes failure modes using RPN = severity × occurrence × detection, and severity is best reduced by design change, occurrence by prevention, detection by better controls.
Traps
Changing several factors at once and calling it a DOE, choosing a solution before the cause is validated, and implementing with no pilot and no plan for control.
Rule of thumb: a two-level full factorial with k factors needs 2^k runs; resolution IV means main effects are clear of two-factor interactions but those interactions are confounded with each other.