Predictive Maintenance of Marine Engines Using a Long Short-Term Memory Algorithm

The main problems in the implementation of predictive maintenance of marine engines are choosing the variables to monitor and designing an approach that is flexible enough and applicable to different assets. This paper proposes a procedure that tackles both the problems. First, it identifies key indicators of the state of health of the engine. Secondly, it develops an online learning by a Long Short-Therm Memory (LSTM) predictive algorithm that can be implemented after turning on the engine. The algorithm continuously updates its knowledge of the aging law of components to have more reliable information on maintenance. Simulation results show the effectiveness of the LSTM algorithm to predict failures using data from an exhaust manifold temperature sensor.