From anomaly detection to root cause analysis: a novel maritime maintenance framework using digital twin synthetic dataset

This paper presents a novel approach in the context of condition-based maintenance (CBM) in the maritime sector, leveraging unsupervised learning through Long Short-Term Memory (LSTM) networks. As ships transition toward unmanned operations, the traditional approach of scheduled maintenance becomes inadequate. CBM leverages advanced sensor networks, real-time monitoring systems, and predictive analytics to enable proactive maintenance strategies without human presence onboard. The research carried out in this paper aims to develop a methodology that combines anomaly detection with feature contribution analysis to identify the presence of anomalies and determine which system parameter signals are most responsible for anomalous behaviour. The approach is validated using a marine gearbox digital twin that generates synthetic operational data. In particular, the adopted model focuses on kinematic parameters such as torque and temperature. Moreover, the digital twin enables the simulation of both step changes and continuous degradation patterns to emulate an anomaly. A key innovation lies in the integration of contribution analysis within the LSTM framework, providing deeper insights into fault progression. Moreover, by using synthetic data with known anomaly patterns, the methodology demonstrates the ability to correctly identify the features most responsible for reconstruction error. This capability is particularly significant for real-world maritime applications, where the critical parameters driving system degradation are typically unknown beforehand and must be discovered through analysis. The framework addresses the common challenge of scarce run-to-failure data in maritime maintenance by adopting a fully data-driven, unsupervised approach where failure characteristics are not pre-defined. The methodology encompasses data generation through digital twin simulation, data processing, Health Index (HI) construction, and predictive analysis. The results validate the effectiveness of this integrated approach in both identifying potential failures and their root causes, enabling more informed maintenance planning in maritime applications. Indeed, the validation approach utilizes synthetic data generated through a digital twin to address the scarcity of real-world run-to-failure data in maritime applications.