# Proactive fault prediction in marine diesel engines using multivariate machine learning

## TL;DR

This paper presents a proactive maintenance strategy for marine diesel engines using machine learning to predict and prevent faults before they occur.

## Contribution

The study introduces a multivariate ConvLSTM and random forest model for accurate fault prediction in marine diesel engines.

## Key findings

- The ConvLSTM model reduced RMSE by 15.4453% compared to decision tree regression models.
- The random forest classifier achieved an accuracy of 82.168% in fault prediction.
- The approach uses metrics like cylinder temperatures and vibrations to anticipate engine faults.

## Abstract

Ocean shipping is the backbone of international trade contributing to global economic growth. Consequently, ensuring that ships operate in an energy-efficient manner is crucial to a more sustainable global transportation. Engine failures in these contexts can lead to severe consequences including compromised safety, operational disruptions, and substantial economic losses ranging between 10% and 30% of total operating costs due to unscheduled maintenance. The proposed research integrates marine diesel engines diagnostics with machine learning (ML) algorithms to develop an advanced proactive maintenance strategy to anticipate engine performance trends and proactively identify potential faults before they escalate. Employing an experimental approach on a 4-stroke diesel engine, the controlled simulations were conducted to replicate various failure scenarios to collect data and capture crucial metrics such as temperatures across cylinders, vibrations along axes, and fluctuations in cooling water temperatures. The data were analysed using advanced ML algorithms aimed at enhancing the accuracy and reliability of future fault prediction, by employing a multivariate convolutional long short-term memory (ConvLSTM) model tailored for time series analysis, and a classification model using a random forest (RF) classifier. As a result, the ConvLSTM model decreased the RMSE by 15.4453% compared to decision tree regression models, while the RF classifier achieved an accuracy of 82.168%.