Tiny Machine Learning for Maintenance Applications: A Systematic Literature Review and Optimization Perspectives

Abstract

Industry 4.0 has revolutionized maintenance strategies and has brought predictive maintenance (PdM), powered by machine learning, into the limelight as a key technology for operational efficiency. However, the conventional cloud-based ML technology does not come without challenges such as communication delays, operational costs, and energy consumption. Tiny Machine Learning (TinyML) is an optimized implementation for machine learning computation on constrained devices distributed at the edge of the network. This paper contributes a full systematic literature review on TinyML applications in maintenance, with a focus on existing methods and potential optimizations. We reviewed 38 original articles from the Scopus, ScienceDirect, and Web of Science using the PRISMA technique. As a result, CNNs are the leader in this domain, with work largely focused on the vibrational and visual data to detect faults in rotation machineries, solar panels, and infrastructure systems. The optimal challenges are the model size, power efficiency, and environmental adaptability. We discuss the open research issues such as the online learning capability, automatic TinyML framework, multisensor integration, and distributed learning strategy. This comprehensive review offers researchers and industrial users an overview of TinyML to develop autonomous, robust, and energy-efficient maintenance systems.