A New Deep Learning Strategy to Predict Time-Series of Marine Diesel Engines

Maintenance of marine engines can radically change thanks to Artificial Intelligence. Deep Learning (DL) algorithms may overcome the limitations of the still widely used preventive maintenance, which is based on statistics of component failures. DL allows prediction of events such as faults, anomali...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:IEEE transactions on industry applications S. 1 - 11
Hauptverfasser: Maione, Francesco, Lino, Paolo, Giannino, Giuseppe, Maione, Guido
Format: Journal Article
Sprache:Englisch
Veröffentlicht: IEEE 2025
Schlagworte:
autoen-coder
deep learning
Diesel engines
Engines
genetic algorithm
Long Short-Term Memory
Maintenance
Marine engine
Marine vehicles
Monitoring
Prediction algorithms
predictive maintenance
Prognostics and health management
Sensors
supervised learning
Training
unsupervised learning
ISSN:0093-9994, 1939-9367
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung:Maintenance of marine engines can radically change thanks to Artificial Intelligence. Deep Learning (DL) algorithms may overcome the limitations of the still widely used preventive maintenance, which is based on statistics of component failures. DL allows prediction of events such as faults, anomalies, and failures by using the state of health of components, such that maintenance depends on the real conditions, and costs and idle time are limited, while the life of components increases. However, the DL approaches developed so far for marine engines have two major weaknesses: they require a huge amount of historical data and do not predict the future time-series behavior of monitored variables. Then, this work proposes an online procedure that takes data from the field to train a DL algorithm predicting the deviation of critical variables from the expected healthy behaviors. This deviation indicates the performance degradation due to failures. The prediction is improved by cyclic update of data, such that the training dataset is not frozen. The strategy is tested on simulated data by comparing three different Long ShortTerm Memory (LSTM) schemes. It is remarked that prediction of the time-series patterns of engine variables can provide in advance accurate information on failure occurrence. Moreover, failure indicators were determined by combining the synthetic data generated by simulation with the knowledge of experts. This result helps robustness to uncertainties and variations due to the different designs and real operations of real of marine engines.
ISSN:0093-9994
1939-9367
DOI:10.1109/TIA.2025.3608691