In the contemporary academic and industrial contexts, predictive maintenance has emerged as a pivotal concern, particularly in the era of artificial intelligence's integration into maintenance operations. This technique promises notable benefits, including cost savings, increased component remaining useful life, enhanced system availability, and the capability to monitor operations in real time. There are three primary categories of predictive maintenance: the data-driven approach, the model-based method and the hybrid model. In railway maintenance play a pivotal role to avoid line interruption and ensure high rolling stock availability. This article provides a comprehensive review underscoring the main methods utilized in recent years thorough a deep analysis of 599 relevant articles obtained from scientific databases. PRISMA method was used to scan relevant studies in the last two decades in three popular databases: Scopus, Web of science and Science Direct. Accordingly, the study presents a survey of predictive maintenance frameworks, approaches and algorithms, in which data driven approach using machine and deep learning techniques were widely employed. Also, it shed light on the major rail equipment that was the focus of the researches. To harness the complementary strengths of both data-driven and model-based, we have proposed a novel hybrid framework that synergistically integrates real-time analytics with theoretical modeling. In conclusion, limitations and challenges were introduced with potential improvements and perspectives for future work.