Vol 77 No 4 77402

Abstract

To address the challenges of inaccurate exhaust gas temperature (EGT) prediction under varying load conditions and the difficulty in engineering implementation within intelligent marine engine rooms, this study proposes a comprehensive solution. The load (L), turbocharger speed (N), and turbocharger exhaust outlet temperature (T1) are selected as core input parameters using Spearman’s rank correlation analysis, reducing the number of parameters by approximately 60 % compared to multi-parameter models. The 0-80 % load range is equally divided into eight intervals of 10 % each, with a dedicated Light Gradient Boosting Machine (LightGBM) predictor developed for each interval. This approach achieves a test set Mean Absolute Error (MAE) of 1.8584, Mean Squared Error (MSE) of 9.7828 and coefficient of determination (R²) of 0.9866, significantly outperforming a full-interval model. And based on the data characteristics of each subinterval, a dynamic threshold warning system with three-level logic is developed, which achieves a fault identification rate of over 92 % for simulated faults. Furthermore, a four-layer Machine Learning Operations (MLOps) architecture is implemented, with the model containerized into a 480 megabytes (MB) image and deployed via a local automated pipeline suitable for network-limited, low-power edge environments. System resource usage remains below 650 % Central Processing Unit (CPU) and 1900 mebibytes (MiB) memory. Validation through six-hour offline closed loop operation and a 180-day simulated aging test confirms the solution’s robustness and practical potential for real-ship intelligent engine room health management.