Automated OCEL Transformation for Real-Time Conformance in Complex Manufacturing

Object-Centric Process Mining (OCPM) offers a powerful framework for analyzing multi-object interactions in complex manufacturing systems; however, its practical deployment remains constrained by the labor-intensive transformation of heterogeneous production data into Object-Centric Event Logs (OCELs). This study addresses that bottleneck by developing and empirically validating an automated transformation pipeline capable of converting raw industrial event streams into conformance-ready OCELs with 97.5\% role-assignment accuracy. Unlike static mapping approaches, the proposed architecture integrates a multi-criteria decision-making (MCDM) layer to systematically evaluate and select classifiers and process discovery algorithms under competing performance criteria, including precision, robustness, latency, and model fitness. An uncertainty-aware active learning mechanism further reduces manual intervention by requesting expert validation only when entropy thresholds exceed predefined bounds. Experimental validation was conducted using a large-scale dataset from HewSaw timber manufacturing operations comprising 54,976 events across 2,000 production cases. Results indicate that GA-optimized XGBoost provides superior balance across accuracy and operational throughput, while Inductive Miner achieves the highest fitness (96.7\%) and generalization (90.1\%) for process discovery. Furthermore, integrating object-centric process features with traditional operational indicators improves real-time conformance prediction accuracy from 75.6\% to 92.4\% (AUC = 0.934), demonstrating the predictive value of multi-object contextual modeling. With average transformation latencies of approximately 1.27 ms per event, the pipeline satisfies real-time deployment constraints. The findings suggest that systematic algorithm selection combined with object-centric feature integration enables scalable, high-fidelity conformance monitoring in complex manufacturing environments, effectively reducing reliance on manual data preparation