Interpretable Machine Learning for NOx Emission Anomaly Detection in Marine Diesel Engines Using Real Operational Data

Abstract

Nitrogen oxide (NOx) emissions from marine diesel engines remain a major environmental concern in the maritime industry, requiring effective monitoring strategies under actual operating conditions. This study proposes a Random Forest-based approach for detecting anomalous operating conditions associated with NOx emissions using operational data collected from an actual merchant vessel during an 11-day voyage. Following data preprocessing and threshold-based operational screening, a total of 347,587 valid observations were retained for model development. The Random Forest classifier was evaluated using Accuracy, Precision, Recall, F1-score, and ROC-AUC. The proposed model achieved an Accuracy of 96.87%, Precision of 97.35%, Recall of 93.18%, F1-score of 95.22%, and a ROC-AUC of 99.39%, demonstrating reliable classification performance for distinguishing normal and anomalous operating conditions. Feature importance analysis revealed that operational parameters contributed differently to the classification process, with ship speed and fuel system parameters exhibiting the greatest influence. An engineering interpretation of these findings indicates that the identified parameters are consistent with established combustion mechanisms associated with thermal NOx formation in marine diesel engines. The proposed approach extends the application of machine learning beyond anomaly detection by providing an interpretable framework that links data-driven classification results with engineering knowledge, offering potential support for onboard condition monitoring and operational decision-making in NOx emission management.