Anomaly detection in engine vibration data

Abstract

This thesis attempted to detect and predict anomalies in marine vessel engines autonomously. Detecting anomalies and predicting future ones in engines could save considerable amounts of time and money; this is due to the potential of marine vessels encountering engine anomalies while at sea, which could be costly and, in worst-case scenarios, might endanger the health of the passengers or the crew. Therefore, it is crucial to detect anomalies and conduct the necessary repairs when the marine vessel is in the harbor. The used dataset is obtained from the cruise ferry Wasaline and contains engine vibration data from the four main engines. Vibrational acceleration from one engine is used to create a health indicator of each time the engine was operating, which is used to train an exponential degradation model. The effect of external variables that might affect the health indicator was investigated, but no clear results were found. The exponential degradation model then attempts to predict the remaining useful life of the engine. The resulting model performed poorly; however, a part of the poor performance can be attributed to the data lacking any label or indicator of engine health.