Predicting marine engine behavior under variable ambient temperature using neural networks

The use of machine learning in marine engine optimization, control, and predictive maintenance has become increasingly valuable under variable environmental conditions. Predictive maintenance, supported by anomaly detection algorithms, enables early identification of potential failures and can reduce downtime and costs in practice. Environmental conditions strongly influence marine engine performance, yet this aspect is rarely addressed in the literature. This study investigates the association between ambient temperature and engine behavior using two artificial neural network (ANN) models trained on experimental data from a large-bore single-cylinder marine diesel engine. The experiments covered ambient temperature spanning 5–45 °C under non–climate-controlled, field-like conditions. The models were trained and evaluated with a 50%/25%/25% split at the sample (crank-angle) level, focusing on identifying temperature-associated patterns within the observed operating domain rather than extrapolating to unseen operating or thermal conditions. The first model (pressure-based) employs in-cylinder pressure signals; the second (accelerometer-based) uses a scalar magnitude from tri-axial accelerometers with minimal preprocessing (≈10° CA moving average) as a low-cost alternative. Both models retrieve a slight decrease in delivered fuel flow at fixed injection commands as temperature increases, in line with the literature on the temperature dependence of fuel viscosity/atomization and injection hydraulics. The accelerometer-based ANN captures this trend with lower precision, requiring more conservative alert thresholds in practice. The pressure-based ANN also allows for testing other temperature-related trends, such as ignition timing advance at higher temperatures, consistent with expected combustion behavior.