Abstract
The digital transformation of industries within the framework of Industry 4.0 has accelerated the adoption of Cyber-Physical Systems (CPS), which integrate physical processes with computational and communication technologies. While these systems offer numerous advantages in monitoring and control, their complexity creates challenges in ensuring reliability and resilience. To address these challenges, Prognostics and Health Management (PHM) frameworks use modern sensor technologies and machine learning (ML) to anticipate faults, extend asset life, and improve operational efficiency. This paper presents a comprehensive review of ML methodologies applied to PHM within the domain of CPS, while exploring key advancements, existing challenges, and future prospects. A novel taxonomy is proposed to classify existing research based on hardware and software fault types, PHM stages, data characteristics, ML techniques, and performance metrics, with the aim of guiding researchers and practitioners in selecting appropriate ML methods for end-to-end PHM tasks spanning data collection, preprocessing, model development, and validation. The review also identifies emerging trends, including the growing adoption of deep learning techniques such as transformers and large language models. Additionally, it underscores the need for more holistic approaches that address the deeper integration of physical and cyber systems, the complexity of cascading fault scenarios, and the persistent gap between academic research and industrial applications. The findings of this study provide a valuable foundation for advancing PHM strategies and supporting their effective implementation within the evolving landscape of CPS.