Application of Machine Learning Algorithms in Predictive Maintenance for Reducing Production Downtime: A Systematic Review

The costs associated with unplanned equipment downtime in the capital-intensive industries such as heavy manufacturing can have a devastating financial effect. Research shows there is a median cost per incident of unplanned equipment downtime greater than USD ۱۲۵,۰۰۰, and global estimates are that Fortune Global ۵۰۰ industrials experience total annual losses of approximately USD ۱.۵ trillion. As a result, the development of Predictive Maintenance (PdM) utilizing Machine Learning (ML) and Deep Learning (DL) algorithms is recognized as one of the most significant advances enabling industries to reduce unplanned equipment downtime by predicting failures prior to their occurrence. This systematic review synthesizes peer-reviewed research published on ML algorithms being applied to PdM in order to reduce unplanned downtime during the time frame of January ۲۰۲۲ through March ۲۰۲۶. The systematic review process was guided by the PRISMA ۲۰۲۰ framework, using structured searches of Scopus, Web of Science, IEEE Xplore, ScienceDirect and SpringerLink. The initial searches resulted in ۱,۸۴۷ results, and after de-duplication and two-stage screening (Cohen's к= ۰.۸۱ between independent reviewers), as well as quality appraisal using the CASP method, the final sample included ۶۲ high-quality primary research papers for synthesis. Results indicate the evolution of the PdM field from traditional supervised learning algorithms, such as Random Forest, SVM, and XGBoost, to blending deep hybrid architectures of CNN, LSTM, and Transformer-based attention mechanisms primarily for the purposes of estimating Remaining Useful Life (RUL) of equipment. The PdM field is being shaped by three primary frontiers: i) Explainable Artificial Intelligence (XAI) to aid in the establishment of human-confidence levels in the predictions of the algorithm; ii) Federated Learning systems to allow for private cross-plant modeling training; and iii) Digital Twin and AI fusions for prescriptive maintenance in closed-loop scenarios. The quantitative synthesis indicated that mature ML-based PdM systems demonstrated median reductions in unplanned equipment downtime of ۳۰% to ۵۰%, a reduction in maintenance costs of ۱۸% to ۲۵% and a median payback period of ۸ to ۱۴ months; however, this systematic review also noted the significant deployment failure rate (۳۰-۴۰%) and the capital expenditure range on average ($۲۵۰k to $۲m) that are rarely addressed by the optimistic publications within the field. Furthermore, a novel conceptual framework titled the DEPTH Framework (i.e., Data, Edge, Prediction, Trust, Human-in-the-Loop) was described, and validated through a case-study walkthrough in order to aid future implementation within the industrial setting.

کلیدواژه ها:

Predictive Maintenance ، Machine Learning ، Deep Learning ، Production Downtime ، Industry ۴.۰ ، Industry ۵.۰ ، Remaining Useful Life ، Systematic Review ، PRISMA ، Digital Twin ، Explainable AI