Machine Learning for Predictive Maintenance of Industrial Systems: A Review of Methods, Benchmarks, Challenges and Future Directions
Shamsuddeen Mohammed Sada
Department of Software Engineering, Federal University of Allied Health Sciences, Enugu, Nigeria.
Nnenna Harmony Nwobodo-Nzeribe
Department of Computer Engineering, Enugu State University of Science and Technology, Enugu, Nigeria.
Adesegun Nurudeen Osijirin *
Division of General Studies and Digital Literacy, Federal University of Allied Health Sciences, Enugu, Nigeria.
*Author to whom correspondence should be addressed.
Abstract
Aims: To critically review how machine learning is applied to predictive maintenance of industrial assets, to compare the main method families and benchmark practices, and to identify the barriers that limit industrial adoption and the priorities for future research.
Study Design: Narrative (non-systematic) review.
Methodology: Peer-reviewed journal articles, conference papers and established review studies published between 2006 and 2026 were identified from major scientific databases and publisher platforms using combinations of the terms predictive maintenance, condition monitoring, fault diagnosis, prognostics, remaining useful life and machine learning. Seminal and highly cited works, benchmark-dataset papers and recent advances were prioritised. The evidence was synthesised by maintenance task, method family, dataset, evaluation metric and deployment issue.
Results: Machine learning in predictive maintenance has progressed from feature-engineered shallow classifiers to deep convolutional, recurrent and attention-based networks that learn directly from raw sensor signals, and, more recently, to transfer learning, federated learning, physics-informed hybrid models and large foundation models. Deep models report strong accuracy on public benchmarks for bearing fault diagnosis and turbofan remaining-useful-life estimation, and hybrid physics and deep-learning models have extended prognostic horizons while requiring less training data. However, most evidence comes from a small number of laboratory or simulated datasets; performance under domain shift, scarce failure data, uncertainty quantification, interpretability and the link between prediction and maintenance decisions remain weakly addressed.
Conclusion: Machine learning is a mature enabler of predictive maintenance, but industrial value depends less on marginal benchmark gains than on robust, explainable, uncertainty-aware and decision-oriented models validated on field data. Standardised evaluation protocols, open industrial datasets, hybrid knowledge-and-data models and cost-benefit evidence are the main priorities for future work.
Keywords: Predictive maintenance, machine learning, condition monitoring, fault diagnosis, remaining useful life, prognostics and health management, deep learning, transfer learning, federated learning, physics-informed modelling