A Systematic Review of AI-Enabled Predictive Quality Control in Advanced Metal Manufacturing Systems

Abstract

This study addresses the persistent problem that quality control in advanced metal manufacturing remains largely reactive, while defect formation is nonlinear, high-dimensional, and costly to detect late, creating scrap, rework, warranty risk, and inspection bottlenecks. The purpose was to quantify and synthesize how AI-enabled predictive quality control (PQC) performs and how deployment-ready it is across manufacturing settings, using a quantitative cross-sectional, case-based design in which each eligible publication is treated as a comparable “case.” The sample comprised 52 peer-reviewed cases (2005–2023) spanning industrial and laboratory environments, including enterprise production cases (real factory datasets) and digitally integrated settings where analytics are typically deployed through cloud and enterprise workflows. Key variables included (1) model family (deep learning, classical ML, hybrid), (2) data modality (machine vision, sensor time-series, NDT-based data), (3) validation rigor (internal vs external), (4) performance metrics (F1, AUC, precision, recall), and (5) implementation readiness (1–5 rubric). The analysis plan used structured coding, frequency distributions, cross-tabulation by context and modality, and comparative summary statistics (means, SD). Headline findings show that deep learning dominated 61.5% (32/52) of cases versus 34.6% (18/52) classical ML and 3.8% (2/52) hybrid approaches; machine vision was the most common input (44.2%, 23/52), followed by sensor time-series (36.5%, 19/52) and NDT-driven data (19.2%, 10/52). Multi-modal fusion appeared in 30.8% (16/52) of cases and 75.0% (12/16) reported performance gains over single-modality baselines. Reporting rigor was uneven: while 86.5% (45/52) reported standard predictive metrics, only 21.2% (11/52) reported external validation. Industrial enterprise datasets showed lower average performance (mean F1 = 0.84, SD = 0.07) than lab datasets (mean F1 = 0.91, SD = 0.05), yet more often reported operational impact (63.2% vs 21.7%). Implications are that PQC benefits most when embedded in enterprise-quality data ecosystems, prioritized for drift-aware validation, and paired with explainability and monitoring to support auditability, risk-based inspection, and scalable deployment.