AI-AUGMENTED SIX SIGMA FOR QUALITY 4.0: A DATA-DRIVEN WORKFLOW FOR DEFECT PREDICTION AND PROCESS IMPROVEMENT

Abstract

Purpose: One of the most important management breakthroughs of the late 20th century is Six Sigma. It is a holistic mindset for attaining process and organizational excellence rather than just a statistical methodology. Intelligent, data-driven quality improvement is made possible by Quality 4.0's integration with Industry 4.0 technology and artificial intelligence. This study develops and empirically evaluates an artificial intelligence-augmented Lean Six Sigma framework for Quality 4.0 by integrating Define-Measure-Analyze-Improve-Control with statistical quality analysis, leakage-controlled machine learning, explainable artificial intelligence, probability calibration and decision-threshold analysis.

Design/methodology/approach: The full University of California, Irvine Artificial Intelligence for Industry 2020 Predictive Maintenance dataset (10,000 observations; 339 failures) was analyzed. Identifier and failure-mode columns were excluded from predictive inputs. Welch tests with Hedges’ g and bootstrap 95% confidence intervals compared process variables between failure and normal operation. A multivariable logistic model estimated adjusted odds ratios. Logistic regression, random forest and Extreme Gradient Boosting were evaluated by repeated stratified 5-fold cross-validation repeated three times. area under the receiver operating characteristic curve, area under the precision-recall curve, precision, recall, harmonic mean of precision and recall, specificity and balanced accuracy were reported. Out-of-fold probabilities supported calibration and threshold analysis, and Tree-based Shapley additive explanations were used for model interpretation.

Findings: Failures represented 3.39% of observations. Extreme Gradient Boosting provided the strongest discrimination (area under the receiver operating characteristic curve 0.977; area under the precision-recall curve 0.788) and recall (0.826), while random forest produced the highest harmonic mean of precision and recall (0.707) and better Brier performance. Torque, tool wear and rotational speed were the leading Shapley additive explanation drivers. Threshold analysis demonstrated an explicit trade-off between failure capture and false alarms.

Practical implications: Define-Measure-Analyze-Improve-Control is strengthened by linking Measure to data quality and baselines, Analyze to inferential statistics and explainable artificial intelligence, Improve to threshold/action policies, and Control to calibration, statistical process control and drift monitoring. Originality/value – The paper reports original full-dataset, leakage-controlled results rather than relying on externally published benchmark scores.