In-cylinder pressure prediction for marine engines using machine learning

Patil, Chaitanya and Theotokatos, Gerasimos and Milioulis, Konstantinos (2023) In-cylinder pressure prediction for marine engines using machine learning. In: The 8th International Symposium on Ship Operations, Management and Economics, 2023-03-07 - 2023-03-08, Eugenides Foundation Auditorium. (https://doi.org/10.5957/SOME-2023-014)

[thumbnail of Patil-etal-SOME-2023-In-cylinder-pressure-realtime-prediction-for-marine-engine]
Preview
 Text. Filename: Patil_etal_SOME_2023_In_cylinder_pressure_realtime_prediction_for_marine_engine.pdf
Accepted Author Manuscript
License: Strathprints license 1.0
Download (1MB)| Preview

Abstract

First principle Digital Twins (DT) for marine engines are widely used to estimate in-cylinder pressure, which is a key parameter informing health of ship power plants. However, development and application of DT faces barriers, as they require exhaustive calibration and high computational power, which render their implementation for shipboard systems challenging. This study aims at developing a data-driven DT of low computational cost for predicting instantaneous pressure. Two different approaches using Artificial Neural Networks (ANN) with distinct input parameters are assessed. The first predicts in-cylinder pressure as a function of the phase angle, whereas the second predicts the discrete Fourier coefficients (FC) corresponding to the in-cylinder pressure variations. The case study of a conventional medium speed four-stroke diesel marine engine is employed, for which the first principle DT based on a thermodynamic, zero-dimensional approach was setup and calibrated against shop trials measurements. The DT is subsequently employed to generate data for training and validating developed ANNs. The derived results demonstrate that the second approach exhibits mean square errors within ±2% and requires the lowest computations cost, rendering it appropriate for marine engines DTs. Sensitivity analysis results verify the amount of training data and number of Fourier coefficients required to achieve adequate accuracy.

CORE (COnnecting REpositories)