Digital Twins for Marine Engine-Room Systems: A Classification Framework for Training and Predictive Maintenance

Abstract

Within the current context and with the changes that have taken place in recent years, the maritime sector is facing a set of transformations that may lead to the adoption of complex cyber-physical systems. These developments facilitate high operational reliability for vessel’s critical elements, including both the power generation plant and the safety of navigation systems, while simultaneously enhancing energy efficiency and environmental sustainability. Until now, traditional maintenance models (corrective, preventive, predictive and static predictive) have been the main resource. These models face several limitations, the statistical sense being the foremost. These operations also require experienced personnel; yet, a shortage of specialized personnel combined with generational turnover creates significant recruitment bottlenecks. Additional challenges include increasing technological complexity and the growing demand for continuous availability on ships’ operation with minimal downtime.The methodology follows an exploratory and verification-oriented approach, combining a review of the technical literature and up-to-date information on the state of the art in this domain with the development of functional digital architectures. These architectures integrate data acquisition and communication, simulation of outcomes, dynamic modelling, and predictive analysis based on machine learning, together with the use of realistic physical systems. Preliminary analyses suggest that implementing a modular architecture makes it possible to supervise the installation at different levels of complexity. Moreover, detailed models are already showing the ability to detect minor deviations in machinery performance. This should make possible to identify not only the root causes of failures that will occur in the installation if no countermeasures are taken, well before they become irreversible, but also to estimate the remaining time before such failures, and to support decisions that prevent unplanned shutdowns. Further on, this technology has the potential to complement the existing training software.The results obtained and the data analysed provide indications of the viability of the modular approach, although they also show that it significantly increases the complexity. It is, however, a technology rooted in solid foundations, such as the physical twins historically used by some space agencies. The findings likewise reveal major challenges, particularly in managing on time the large volumes of data generated on board and in addressing potential vulnerabilities to cyberattacks on these data. It has been observed that the effectiveness of these tools is closely tied to the correct modelling of the twin and to the availability of high-quality, on time data that are properly labelled, segmented and updated at an appropriate frequency. This last aspect, related to temporal resolution and synchronization frequency, is still not always explicitly considered in digital twin classifications, although it is essential for practical implementation. For this reason, this paper proposes a classification scheme to support the selection of the appropriate type of model according to their purpose, required fidelity, data demand and synchronization needs. The technology is promising, but its implementation at plant level entails a degree of complexity that calls for progressive development and rigorous validation across multiple operational scenarios.