| Abstract: |
Manufacturing plants need maintenance policies that protect equipment reliability without sacrificing throughput, quality, or cost. This review examines past work on artificial intelligence (AI)-based predictive maintenance (PdM) and synthesizes evidence for a manufacturing framework connecting condition monitoring, failure diagnosis, remaining useful life (RUL), maintenance decisions, and production scheduling. A structured evidence synthesis covers foundational prognostics and health management, machine learning, deep learning, industrial IoT, digital twins, knowledge-based reasoning, and studies linking predictions to maintenance actions. The synthesis finds a consistent technical pattern: sensing and analytics are increasingly capable of detecting abnormal states and estimating degradation, yet evidence for sustained plant-level reliability and productivity gains remains less mature. Benchmark accuracy often dominates reporting, while calibration, lead time, false-alarm burden, cross-machine transfer, technician response, and production effects are reported inconsistently. Because equipment, datasets, failure definitions, horizons, and outcome measures differ, a pooled numerical effect estimate would be misleading; the paper therefore presents a transparent, paragraph-form descriptive meta-analysis rather than claiming a statistical effect-size meta-analysis. On this basis, it proposes a human-supervised, uncertainty-aware framework that fuses sensor, process, maintenance, and production data; deploys edge and cloud analytics; converts risk estimates into cost- and schedule-aware work orders; and learns from completed interventions. Reliability and productivity are treated as joint outcomes measured through availability, mean time between failures, downtime, OEE, throughput, quality, and lifecycle cost. The paper identifies evidence gaps and proposes a validation ladder for future operational evaluation. Its contribution is an integrative synthesis and deployable architecture, not a new empirical performance claim. |
