EARLY FAULT SEVERITY PREDICTION IN DIGITAL-TWIN–ENABLED INDUSTRIAL SYSTEMS USING IMBALANCE-AWARE MACHINE LEARNING AND TEMPORAL FEATURE ENGINEERING

Continuous industrial telemetry in high dimensions generated by digital twin (DT) platforms offers opportunities for large-scale fault and predictive maintenance (PdM). However, the real-life applications of PdM remain very challenging. In this paper, we investigate the potential of early fault severity prediction using a five-year, hourly dataset of DT-derived data from January 2019 to January 2024. The data includes 38 variables, including measurements such as vibration, temperature, pressure, acoustic signals, operational load, maintenance and repair history, anomaly scores, fault probability estimates, and a multiclass fault diagnosis label of No Fault or Critical Fault. The research presented in this paper follows an end-to-end analytical approach, using a set of machine-learning classifiers, then optimizing them using class-weighted classifiers, probability calibration, and a cost-sensitive decision policy, all aligned with the strategic goals of PdM. The results of the experiments conducted in this research show that the baseline and temporally enhanced models achieve ≈0.69 accuracy. When class-weighted learning is introduced, it increases fault recall to 0.059 and macro-F1 to 0.451, while threshold optimization achieves perfect fault recall (1.00) with a corresponding fault F1-score of 0.468. The main findings of this work show that the joint effects of temporal modelling, imbalance-aware learning, and calibrated decision thresholds govern early severity prediction performance. Our research is expected to yield a reproducible, deployment-oriented framework that improves minority-class detection and enables effective maintenance decisions for DT-enabling Industry 4.0/5.0 applications. JEL Classification Codes: O14, O32, L42.