A multi-domain machine learning framework for intelligent condition monitoring of marine diesel engines

Abstract Accurate and early monitoring of injector faults in marine diesel engines is essential for maintaining operational efficiency and preventing unplanned downtime. This study develops a novel multi-domain machine learning framework based on wavelet transform (WT), principal component analysis (PCA), and dynamic independent component analysis (DICA). The framework extracts complementary statistical and spectral indicators from vibration signals, encompassing time, frequency, and time–frequency-domain representations. The multi-domain WT-based PCA–DICA framework ensures effective dimensionality reduction and separation of correlated sources, enabling the identification of hidden features that are difficult to capture through single-domain analysis. A diverse set of machine learning classifiers, including support vector machines, naïve Bayes, k-nearest neighbors, logistic regression, AdaBoost, and linear discriminant analysis, is systematically evaluated using a five-fold cross-validation approach. The proposed approach, tested and validated on the Sulzer 6AL20/24 test engine, achieves classification accuracies and AUC values exceeding 98%, demonstrating robustness to noise, class imbalance, and limited fault data. The proposed framework demonstrates robustness against noisy and unbalanced datasets through data augmentation techniques (slicing and random shuffling) and systematic feature concentration analysis. This multi-domain, data-driven framework provides a reliable, scalable approach to injector fault diagnosis and can be extended to other types of faults and engine configurations. The results provide valuable insights into condition-based maintenance and intelligent marine-engine health-monitoring systems.