A methodology for data-driven diagnosis of marine two-stroke dual-fuel engines

As maritime industry moves towards a carbon neutral future, ship systems are gradually becoming more sophisticated and complex. Apart from decarbonisation, digitalisation of those systems is crucial to improve their energy efficiency, enhance safety and reduce their environmental footprint. Data driven diagnostics as well as prognostics and health management (PHM) digital tools are expected to contribute significantly to the safe and cost-effective operation of ships. The development of these tools requires datasets representing wide operating envelopes, which however, are not readily available in the shipping industry. This study aims to develop a novel health diagnosis methodology for two-stroke marine engines, with the objective to identify the engine components’ health condition using acquired performance parameters. The methodology consists of the following three phases: (a) data correction and conditioning that employs advanced data analysis methods to remove noise and outliers from historical engine datasets leading to reference distributions of key performance parameters; (b) development and calibration of a physics-based digital twin by extending an existing thermodynamic model and incorporating the degradation patterns of selected engine components; this digital twin will then be employed to derive simulated datasets representing the engine operation across a wider range of conditions than covered by the available measurements. (c) development of data-driven models based on three methods; namely, artificial neural networks (NN) of multilayer perceptron (MLP) type and k-Nearest Neighbor (kNN) and support vector machines (SVM) which proved to be most effective in previous studies, to predict the health indicators of engine components. This data-driven model will be trained and validated based on the synthetic datasets, whereas it will be tested by providing as input non-supervised datasets. The proposed methodology will be demonstrated through case studies considering the fuel injector degradation of a marine two-stroke dual-fuel engine, for which extensive measured datasets are available. The expected results include the data-driven model layout to provide the highest accuracy. Proving this methodology’s effectiveness will facilitate its implementation as part of a shipboard diagnosis system, as well as the development of prognostics and health management tools, which are required for smart/intelligent ship operation.