Machine Learning and Artificial Intelligence Enabled Failure Prediction for Maritime Propulsion Systems
Authors
Lytje-Dorfman, Emil ; Køpke, Jacob Vitfell
Term
4. semester
Education
Publication year
2025
Submitted on
2025-06-04
Pages
118
Abstract
This thesis examines how machine learning (ML) and artificial intelligence (AI) can predict failures in maritime propulsion systems. It focuses on two critical failure types observed in MAN Energy Solutions’ engines: (1) accumulator membrane failure in sequential two‑stroke engines and (2) pipe damage during methanol fuel changeovers in dual‑fuel systems. The aim is predictive maintenance—spotting early warning signs so action can be taken before breakdowns occur. The study uses time‑series sensor data from operating vessels and controlled test engines. A feature engineering pipeline was built to turn many raw measurements into meaningful indicators and to handle challenges such as high dimensionality, redundant signals, and class imbalance between normal operation and rare failures. Feature selection produced interpretable, task‑relevant subsets that reveal system behavior leading up to failures. Several ML/AI models were evaluated, including Random Forests, 1D Convolutional Neural Networks (1D CNNs), and Long Short‑Term Memory (LSTM) networks. They were tested on two tasks: (1) binary classification of whether a failure is imminent and (2) estimating remaining useful life (RUL). The experiments varied how much history the model reads (reading window) and how far ahead it predicts (prediction window). To address the scarcity of failure cases, data augmentation techniques (SMOTE and DTW‑SMOTE) were used to synthesize additional training examples. Findings show that failures during methanol changeovers are highly predictable: Random Forest and 1D CNN models achieved 91% accuracy for changeover failure classification and an R2 of 0.826 for RUL prediction. Accumulator membrane failure was harder to predict due to limited failure data and lower signal resolution; the best accuracy was 71% with a 1D CNN. These results highlight the importance of choosing the right model, setting task‑specific window sizes, and keeping features interpretable in predictive maintenance. The thesis concludes that ML and AI can be used to predict pipe damage during methanol changeovers while providing insights into system behavior through feature engineering. Predicting accumulator membrane failure in sequential two‑stroke engines is more challenging, and further work is needed to develop robust, high‑performing models.
Dette speciale undersøger, hvordan maskinlæring (ML) og kunstig intelligens (AI) kan forudsige fejl i skibsmotorers fremdriftssystemer. Fokus er på to kritiske fejl, der er observeret i MAN Energy Solutions’ motorer: 1) akkumulatormembranfejl i sekventielle totaktsmotorer og 2) rørskader under skift til/fra metanol i dual‑fuel systemer. Målet er prædiktivt vedligehold – at opdage tegn på kommende fejl, så man kan handle i tide. Arbejdet bygger på tidsseriedata fra sensorer på operative skibe og kontrollerede testmotorer. En databehandlingspipeline (feature engineering) blev udviklet for at omsætte mange rå målinger til meningsfulde indikatorer og håndtere udfordringer som høj dimensionalitet, overlappende information og skæv fordeling mellem normal drift og sjældne fejl. Gennem feature selection er der udvalgt forståelige og opgave-relevante indikatorer, som giver indblik i, hvad der sker i systemet op til en fejl. Flere ML/AI‑modeller blev afprøvet, herunder Random Forest, 1D Convolutional Neural Networks (1D CNN) og Long Short‑Term Memory‑netværk (LSTM). De blev evalueret på to opgaver: 1) at afgøre om en fejl er på vej (ja/nej), og 2) at estimere resterende levetid (Remaining Useful Life, RUL). Der blev eksperimenteret med, hvor lang en historik modellen læser (læsevindue), og hvor langt frem i tiden den skal forudsige (forudsigelsesvindue). For at håndtere få fejltilfælde blev der brugt dataudvidelsesteknikker (SMOTE og DTW‑SMOTE) til syntetisk at supplere træningsdata. Resultaterne viser, at fejl under metanolskift er meget forudsigelige: Random Forest og 1D CNN opnår 91 % nøjagtighed i klassifikation af skiftrelaterede fejl og en R2‑værdi på 0,826 for RUL‑estimatet. For akkumulatormembranfejl er opgaven vanskeligere på grund af få fejlcases og lavere signalopløsning; her gav 1D CNN den bedste nøjagtighed på 71 %. Fundene understreger vigtigheden af korrekt modelvalg, passende valg af læse- og forudsigelsesvinduer samt tolkbare indikatorer i prædiktivt vedligehold. Konklusionen er, at ML og AI kan bruges til at forudsige rørskader under metanolskift og samtidig give nyttig indsigt i systemadfærd via feature engineering. For akkumulatormembranfejl i sekventielle totaktsmotorer er der behov for yderligere arbejde og data for at opnå mere robuste og højtydende modeller.
[This apstract has been rewritten with the help of AI based on the project's original abstract]
Keywords