DEVELOPMENT OF AN INTELLIGENT PREDICTIVE MAINTENANCE FRAMEWORK FOR INDUSTRIAL EQUIPMENT USING ARTIFICIAL INTELLIGENCE AND DIGITAL TWIN TECHNOLOGY

Abstract

Vibration-based machine-learning classifiers have become a common substitute for fixed-interval preventive maintenance, yet they remain fundamentally pattern-matching tools: they flag deviations from historical normal behavior without reference to the physical degradation mechanism producing the deviation, which limits both the lead time and the precision of the failure predictions they generate. This study proposes a Digital Twin-Coupled Anomaly-to-Remaining-Useful-Life (DT-ARUL) framework, in which a physics-based digital twin of the monitored equipment generates an expected sensor response under current operating conditions, the residual between this expected response and the measured IoT sensor stream is treated as a physically interpretable degradation signal, and a data-driven correction model maps the accumulated residual trajectory onto a probabilistic remaining-useful-life estimate that is continuously updated through Bayesian filtering as new measurements arrive. A dynamic maintenance-threshold optimization layer subsequently converts this RUL distribution into a maintenance-triggering decision by minimizing the expected total cost of premature replacement against the expected cost of in-service failure. The framework was evaluated on a representative rotating-equipment case (centrifugal pump and induction-motor assembly) instrumented with vibration, temperature, and current sensors, compared against a fixed-interval baseline and a standalone vibration-classifier approach. The proposed framework extended mean prediction lead time from 4.2 to 11.6 days, reduced the false-alarm rate from 18.4% to 6.3%, and lowered annual maintenance cost by 44% relative to baseline, with a five-year net present value more than double that of the standalone classifier approach.