Augmented reality (AR) is increasingly adopted in industrial environments as a tool for improving employee training and supporting complex assembly operations. The purpose of this study was to investigate the design, implementation, and strategic potential of AR-based work instructions using Microsoft HoloLens 2 in a real manufacturing environment. The study proposes and applies an integrated evaluation framework combining direct observation, performance evaluation, semi-structured interviews, quantitative SWOT analysis, PDCA-based process assessment, and economic cost analysis to assess AR-based training in a real manufacturing environment. AR training was implemented through Microsoft Dynamics 365 Guides for a standardized assembly procedure and evaluated with respect to training efficiency, user interaction, and feasibility of deployment. The results indicate improved task guidance consistency and descriptive performance indicators, suggesting enhanced training support under real production conditions. The SWOT analysis identified a favorable SO strategic position, highlighting strong internal capabilities and promising external opportunities for further deployment. The cost analysis shows that AR-based training becomes economically advantageous when applied to a larger number of trainees, despite high initial investment costs. Overall, the study demonstrates that AR-based training, when evaluated through a structured strategic and economic framework, represents a promising and strategically advantageous approach for industrial education, provided that ergonomic challenges, user adaptation, and financial constraints are systematically addressed.
Operator-dependent dimensional measurements can produce inconsistent categorical quality decisions, particularly near classification boundaries. This study presents an integrated framework combining measurement-system analysis, inter-operator agreement, analytical-threshold sensitivity, and machine learning to characterize decision instability. Six operators each measured the width and height of 25 Charpy V-notch specimens three times using one digital micrometer (900 measurements). Four exploratory symmetric thresholds around 10 mm generated 1200 operator–specimen–dimension–threshold decisions; disagreement was defined relative to the consensus of the other five operators, not certified dimensional ground truth. Overall, 114 decisions (9.5%) disagreed with consensus, with higher disagreement for height (93/600; 15.5%) than width (21/600; 3.5%). Under nested specimen-grouped cross-validation, the extended Random Forest model achieved AUROC 0.908 (95% specimen-cluster bootstrap confidence interval, 0.878–0.940) and PR-AUC 0.565 (0.448–0.668); a transparent boundary-distance baseline achieved 0.889 and 0.445, respectively. Feature-ablation analysis indicated additional predictive information beyond categorical decision and boundary distance, although target-related dependence remained. Lowering the model-output cutoff from 0.50 to 0.15 increased recall from 0.439 to 0.851 at the cost of more verification warnings. Findings support exploratory risk-based screening, while small specimen numbers, shared instrumentation, and lack of independent reference dimensions limit claims about error detection and industrial transferability.