Identifying maintenance needs with machine learning: a case study in railways

Cyber-physical systems, particularly those with extended service lives such as railways, often necessitate significant investment in maintenance activities encompassing repairs, upgrades, or inspections. These decisions are generally based on fixed schedules, or informed by the judgment of experienced maintenance staff. To improve this process, predictive maintenance (PdM) has emerged as a viable solution to anticipate maintenance needs and preempt system failures. With data-driven PdM, maintenance needs are identified through machine learning (ML) solutions that monitor the system logs and recommend interventions before a failure occurs. This paper presents preliminary findings from a case study concerning the development of a ML system for PdM in railways. We present the current maintenance process, the existing logging platform, and our strategy for leveraging log data to support PdM. Our preliminary results are promising. However, they show that, although the log dataset spans three years and three railway vehicles, in some cases the log data alone are insufficient for accurately inferring maintenance requirements. To address the problem, we discuss the necessity of employing synthetic data generation methods and rule-based, knowledge-driven strategies.