Evaluating the Performance of Generative Adversarial Network, Synthetic Minority Oversampling Technique, and Adaptive Synthetic Sampling in Marine Diesel Engine Fault Diagnosis using Vibration Data
Abstract
Machine learning-driven fault diagnosis of marine diesel engines is a major research focus, challenged by scarce, insufficient fault data. This study explores vibration data under five conditions: normal operation, single-cylinder ignition failure, piston ring wear, nozzle blockage, and exhaust valve leakage. It employs GAN (Generative Adversarial Network), SMOTE (Synthetic Minority Oversampling Technique), and ADASYN (Adaptive Synthetic Sampling) to augment data, addressing imbalance. Principal Component Analysis (PCA) with Cosine Similarity (CS) and Pearson Correlation Coefficient (PCC) qualitatively and quantitatively analyzed augmentation effects. Classification performance was evaluated using Support Vector Machine (SVM), Decision Tree, and Random Forest via accuracy, precision, recall, and F1-score. Experiments showed GAN, by enhancing sample diversity and realism, outperformed SMOTE and ADASYN, increasing average CS by 0.04-0.08 and PCC by 0.05-0.09 across faults. Compared to the original dataset, GAN-enhanced data improved the three classifiers' accuracy by ~3.3%, 7.6%, and 5.8%, with the most significant gains in other metrics, supporting real-world marine diesel fault diagnosis.