Anomaly Detection and Breakdown Diagnosis for Condition Monitoring of Marine Engines

Marine vessels are complex interconnected systems and maintaining the health of individual components of the system increases uptime, boosts efficiency and safety of vessel operations. With recent transition from preventive to predictive maintenance of engineering assets and the prevalence of IoT sensors embedded within these assets, effective condition monitoring of engineering assets in marine vessels is now a reality. This paper aims at developing a solution for condition monitoring and diagnosis of potential breakdowns in the main engines of marine vessels using sensor data. Specifically, we analyze irregularly sampled multi-variate time series data originating from multiple sensors onboard the vessel engine to develop an autoencoder-based anomaly detection model for effective condition monitoring of the engines. In addition to the anomaly detection model, we devise a hierarchical framework to diagnose the potential cause of breakdown. The model is trained on data obtained from engines of two vessels. We train the model using historical time-series data corresponding to the vessel’s condition and operational profile over a month. The model is validated using historical time-series data collected over a year of vessel operations. The performance study of our model demonstrates its ability to predict breakdowns in advance with an average F-1 score of 85.3%, and average 11-14 days in advance of the actual reported breakdown dates. This proposed solution can be a promising tool for the condition monitoring and diagnostics of marine vessel engines.