Decay detection of a marine gas turbine with contaminated data based on isolation forest approach

Machine learning is an effective way to realise the condition monitoring of marine machinery. However, it is challenging to realise this purpose based on supervised learning in practice due to the lack of labelled data. To overcome this problem, we propose to use isolation forest to realise the decay detection of a marine gas turbine with normal data. Besides, we consider the impact of data contamination for the first time compared with previous literatures. We also experiment with the same datasets with support vector data description (SVDD) as a comparison. The results show that the isolation forest is very suitable for the decay detection of the marine gas turbine, and it shows a significant advantage over support vector data description in the tolerance to contaminated data. The dataset we experiment with is from a real-data validated numerical simulator developed for a Frigate's propulsion plant.

Automated summary

This study proposes using isolation forest for decay detection of marine gas turbines and considers the impact of data contamination on results. Through experimentation and comparison, it is demonstrated that isolation forest has a significant advantage over support vector data description in tolerance to contaminated data.