A unified predictive explainable causal framework for real-time quality control in smart manufacturing

Manufacturing quality decisions increasingly rely on sensor data, yet translating predictive insights into reliable operational actions remains difficult. This study develops a Unified Predictive–Explainable–Causal Loop (UPCL) that connects quality prediction, model interpretation, process monitoring, feasibility-aware intervention analysis, and causal evaluation within a common decision-support structure. The framework is examined using a reported dataset of 400 production batches from two industrial sites in Qatar, alongside a supplementary analysis of 50 available batch records. The original evaluation reports a ROC-AUC of 0.714, suggesting moderate discrimination between conforming and non-conforming batches. In the supplementary analysis, five-fold cross-validation yields a ROC-AUC of 0.732, although classification at the conventional 0.50 threshold fails to identify any non-conforming batches. This finding highlights the importance of evaluating decision thresholds rather than relying on discrimination alone. Model explanations are further examined for their consistency under resampling, while cross-site comparisons reveal limitations in transferring predictive relationships between different production environments. Feasibility-constrained simulations identify operating conditions associated with lower defect rates, improved first-pass yield, and reduced carbon intensity; these outcomes are interpreted as model-based scenarios rather than verified intervention effects. The study also incorporates a difference-in-differences design to examine changes in quality indicators following implementation, although the original causal estimates require further verification. Overall, the findings illustrate both the potential and the practical limitations of integrating predictive analytics, explainability, and operational evaluation into manufacturing quality management. The proposed framework provides a structured basis for more transparent quality-control decisions while emphasizing the need for reliable data, site-specific validation, and careful assessment before operational deployment.