A Drift-Conscious Machine Learning Governance Framework for Phase Specific Model Selection in Lean Six Sigma DMAIC

Abstract

Lean Six Sigma projects increasingly adopt machine learning for prediction, diagnosis, and improvement support, but model selection is often treated as a one-time accuracy exercise. This creates governance risk because DEFINE, MEASURE, ANALYZE, IMPROVE, and CONTROL require different evidence standards. This study addresses this limitation by developing a drift-conscious, phase-adaptive machine learning governance framework for DMAIC-based model selection. An industrial batch-level dataset was segmented by DMAIC phase: DEFINE (n = 10), MEASURE (n = 6), ANALYZE (n = 6), IMPROVE (n = 4), and CONTROL (n = 4). All critical variables recorded 0.00 missingness. Candidate models were evaluated using phase-specific metrics: Spearman ranking stability for DEFINE, Performance Degradation Index for MEASURE, cross-validation variability for ANALYZE, desirability-gain stability for IMPROVE, and PSI with rolling RMSE for CONTROL. Results confirmed that no single model dominated all phases. In DEFINE, gradient boosting produced stable driver prioritisation with Spearman ρₛ = 0.95, identifying drying time, moisture, compression speed, compression force, and feeder rate as the leading drivers. In MEASURE, gradient boosting showed the strongest robustness, with mean PDI = 0.56 and worst-case PDI = 1.35, outperforming Ridge and Linear Regression under perturbation. In ANALYZE, Ridge and Lasso delivered superior diagnostic stability, with CV SD values of 0.29 and 0.28, respectively. In IMPROVE, mean gains clustered around 0.08, but reversal rates ranged from 0.20 to 0.34. In CONTROL, PSI increased from 0.02 to 0.68 across 30 windows, validating drift-triggered reassessment.