On Predicting Marine Engine Measurements with Synthetic Data in Scarce Dataset

The scarcity of high-quality maritime datasets poses a significant challenge for machine learning (ML) applications in marine engineering, particularly in scenarios where real-world data collection is limited or impractical. This study investigates the effectiveness of synthetic data generation and cross-modeling in predicting operational metrics of LNG carrier engines. A total of 38 real-world data points were collected from port and starboard engines, focusing on four target outputs: mechanical efficiency, fuel consumption, load, and effective power. CopulaGAN, a hybrid generative model combining statistical copulas and generative adversarial networks, was employed to produce synthetic datasets. These were used to train multilayer perceptron (MLP) regression models, which were optimized via grid search and validated through five-fold cross-validation. The results show that synthetic data can yield accurate models, with mean absolute percentage errors (MAPE) below 2% in most cases. The combined synthetic datasets consistently outperformed those generated from single-engine data. Cross-modeling was partially successful, as models trained on starboard data generalized well to port data but not vice versa. The engine load variable remained challenging to predict due to its narrow and low-range distribution. Overall, the study highlights synthetic data as a viable solution for enhancing the performance of ML models in data-scarce maritime applications.