Data-driven predictions for the health of engine bearings

Abstract

The main bearings which support the crankshaft of the main engines on the Ocean-going Patrol Vessels (OPV) of the Royal Netherlands Navy (RNLN), are critical components. Failure of those bearings results in an impact on the capabilities and deployability of those ships, and the risk of structural damage to the engine. By applying data analysis, these failures can be prevented more efficiently than the current maintenance strategies do. By predicting the bearing temperature, which is an indicator of its health, deviant behaviour can be detected, which corresponds to a damaged bearing. The bearing temperatures are predicted using regression techniques, including linear regression, recursive regression, and regression adjusted variables. Despite the complexity of the engine, which has many interdependent components, and the dynamic operational behaviour of the naval ships, the models compared in this the- sis are able to predict failures. The available data consists of five years of operational data from the HNLMS Groningen, in which two bearing failures have occurred. By com- bining regression techniques, feature selection methods, and monitoring charts, 33 failure prediction models are constructed and compared. To make sure those models are deployed efficiently at the RNLN, they contain interpretable and transparent techniques. This is necessary for naval personnel that make time-critical maintenance decisions, to trust in the maintenance decision support. The best performing models are constructed by combining recursive regression, which retrains the temperature prediction model for every observation, with forward feature selection, a straightforward selection method, and charts based on the weighted average, which can detect small and persistent trends.