Integration of statistical process control and machine learning methods in quality assurance engineering practice on highly variable production lines

Abstract

The article addresses the challenge of ensuring product quality on high-variability production lines under increasing complexity of cyber-physical manufacturing systems and the limitations of classical statistical process control methods. The aim of the study is to develop and experimentally validate a hybrid approach that integrates SPC tools and machine learning methods for proactive quality management in the production of carbon-fiber-and-epoxy-based prepregs. The object of the study is a real aerospace-grade production line, for which a high-frequency dataset (45 parameters, over 2 TB) was collected over 12 months, including both process parameters and final quality indicators with defect labels. At the first stage, traditional SPC tools (XmR charts, Cp and Cpk indices) were applied to assess process stability and identify critical parameters; it was shown that several variables, in particular binder viscosity and carbon-fiber tension, exhibit low reproducibility and serve as major sources of variability. At the second stage, data preprocessing and analysis were performed using machine learning methods: outliers were removed using Isolation Forest, normalization was carried out, and a random forest model was used to rank the 45 parameters by importance, which revealed non-obvious effects of shop-floor humidity and calender-roll temperature. To predict defects, logistic regression, SVM, and XGBoost models were trained and compared, with gradient boosting demonstrating the best performance (F1-score 0.959 with high precision and recall), confirming the pronounced nonlinearity of the process under study. Additionally, the DBSCAN algorithm made it possible to identify hidden clusters of operating regimes associated with both quality degradation and measurement-equipment deterioration, opening opportunities for predictive maintenance. Based on the obtained results, a three-level quality-management architecture is proposed, in which SPC ensures basic monitoring, predictive ML models provide proactive defect-risk control, and cluster analysis performs the function of detecting new anomalous states, forming a data-driven foundation for implementing the Industry 4.0 concept on high-variability production lines.