This paper introduces a novel approach centred on unsupervised learning, specifically leveraging state-of-the-art recurrent neural networks (RNNs), particularly Long Short-Term Memory (LSTM). The study aligns with the domain of predictive maintenance (PdM) and time series analysis for evaluating the health status of devices. The proposed methodology seeks to enhance current PdM practices by integrating machine learning (ML) into the conventional statistical framework. ML techniques are increasingly applied in the industrial sector, demonstrating their capacity to capture complex correlations that may elude human operators. Indeed, the management of scheduled maintenance in the naval sector is a complex challenge, primarily due to the diversity of devices and systems on board ships. These systems vary in nature, complexity, and criticality, making it challenging to adopt a standardized maintenance strategy. Solutions based on Condition-Based Maintenance (CBM) aim to perform maintenance based on the actual operational conditions of a system, rather than following fixed schedules. This approach relies on the use of sensors, continuous monitoring, and advanced diagnostics to assess the health status of components and predict failure times. However, despite advancements in this direction, there are still significant challenges to address. One of the main challenges is the heterogeneity of systems on board ships. Furthermore, approaching the predictive maintenance task for onboard ship equipment, a key challenge emerges when attempting to estimate the Remaining Useful Life (RUL) of a component given the scarcity of run-to-failure data. For such a reason, the developed model adopts a fully data-driven approach, where the failure characteristics (run-to-failure data) of the equipment are not pre-defined. Moreover, the framework encompasses data acquisition, data preprocessing, Health Index (HI) construction, and the prediction of remaining useful life grounded in the examination of a documented failure case related to maritime equipment. Indeed, the framework has been tested on data corresponding to an actual equipment failure, enabling a direct comparison with its application to data from normally functioning equipment without anomalies. The presented unsupervised approach serves a dual purpose: enabling real-time detection of potential failures and facilitating trend analysis of the device's health index. This dual functionality proves valuable in estimating the projected failure period.