Development of an AI-Based Predictive Maintenance and Fault Diagnosis System for Marine Diesel Engines

Accurate classification of fault types in marine diesel engines is critical for ensuring operational reliability and optimizing maintenance efficiency. In this paper, we introduce a hybrid deep-learning framework for multi-class fault diagnosis based on time-series sensor data. The proposed architecture synergistically combines Convolutional Neural Networks for local feature extraction with Bidirectional Long Short-Term Memory layers to capture temporal dependencies and incorporates the Time2Vec encoding scheme for explicit modeling of the time dimension. Sensor signals are segmented via a three-step sliding-window approach and grouped according to operational conditions. Performance is evaluated using accuracy, precision, recall and F1-score metrics. A comparative analysis of multiple configurations identifies the Time2Vec-CNN-BiLSTM model as the optimal architecture, achieving an accuracy of $\mathbf{9 6 \%}$. These results demonstrate the framework's robustness in time-series fault diagnosis and highlight its potential to advance predictive maintenance practices in marine diesel engine systems.