One Class Classification Based Anomaly Detection for Marine Engines

Abstract

Although extensive research has been undertaken to detect faults in marine diesel engines, significant challenges still remain such as the need for non-invasive monitoring methods and the need to obtain rare and expensive datasets of multiple faults from which machine learning algorithms can be trained upon. This paper presents a method that uses non-invasive engine monitoring methods (vibration sensors) and doesn’t require training on faulty data. Significantly, the one class classification algorithms used were tested on a very large number (12) of actual diesel engine faults chosen by diesel engine experts and maritime engineers, which is rare in this field. The results show that by learning on only easily obtainable healthy data samples, all of these faults, including big end bearing wear and ‘top end’ cylinder leakage, can be detected with very minimal false positives (best balanced error rate of 0.15%) regardless of engine load. These results were achieved on a test engine and the method was also applied to an operational vehicle/passenger ferry engine where it was able to detect a fault on one of the cylinders that was confirmed by the vessel’s engineering staff. Additionally, it was also able to confirm that a sensor fault occurred. Significantly it highlights how the ‘healthiness’ of an engine can be assessed and monitored over time, whereby any changes in this health score can be noted and appropriate action taken during scheduled maintenance periods before a serious fault develops.