The maritime industry heavily relies on vessel maintenance to ensure the operational integrity and safety, as it is responsible for transporting more than 80% of global trade. Despite the industry’s strong need for efficient maintenance techniques, there has been a noticeable gap in research regarding the use of data-driven methods to enhance vessel reliability. This study seeks to fill this gap, by examining the feasibility of deploying deep learning models to predict the Remaining Useful Life (RUL) of Heavy Fuel Oil (HFO) purification systems, taking into account also the challenges of the limited computational resources available on maritime vessels, as well as the substantial costs associated with implementing such models. Towards this direction, the impact of various optimization techniques (early stopping and pruning) on three state-of-the-art models (Long Short-Term Memory Network, Convolutional Neural Network, Autoencoders) was evaluated using operational vessel data provided by Laskaridis Shipping Co. Ltd., demonstrating the feasibility of deploying predictive maintenance (PdM) systems in real-world edge-constrained marine settings, potentially transforming maintenance practices and reducing operational costs.