Explainable machine learning with SMOTE for predictive maintenance of ship engines

Abstract

Predictive maintenance is essential for ensuring the operational reliability of ship engines by identifying abnormal conditions before they result in costly failures or downtime. This study introduces an advanced framework for predictive maintenance by integrating explainable machine learning techniques, the Synthetic Minority Over-sampling Technique (SMOTE), and Bayesian optimization for hyperparameter tuning. Initially, machine learning models, including Decision Trees, k-Nearest Neighbors (kNN), Linear Discriminant Analysis, Support Vector Machines, Bagged Trees, Boosted Trees, and Artificial Neural Networks, were trained on a highly imbalanced dataset, which led to poor performance on minority class predictions. To mitigate this limitation, SMOTE was applied to balance the training data, enhancing the models’ ability to predict minority class instances. Bayesian optimization was then employed to fine-tune the models’ hyperparameters, resulting in significant improvements in classification metrics. The results, based on a publicly available dataset, demonstrate that the Optimizable Ensemble and Optimizable kNN models achieved superior discrimination performance, with AUCs exceeding 98%. To improve model transparency, SHapley Additive exPlanations (SHAP) plots were utilized to interpret the contributions of individual features to model predictions. Among the predictors, Engine RPM and Fuel Pressure emerged as the most influential, as evidenced by their broad SHAP value distributions and high magnitudes, underscoring their critical role in assessing the operational conditions of ship engines.