Recurrent Neural Networks for predicting ship motor temperatures - aiming to help prevent motor overheating.

Abstract

For electric propulsion motors installed on marine vessels, the prevention of overheating is usually based on resistor temperature detector sensors. The monitoring system raises an alarm if the detected temperatures reach a pre-defined safety limit. Field experience reveals, however, that damage to the motors may have already occurred before the alarm is triggered due to the delay in the heat transfer process. This thesis aims at developing a data-driven approach to detect the real-time anomalies of motor temperatures, based on extensive field data provided by ABB, a Swedish–Swiss multinational technology corporation. This study consists of two parts. First, a novel algorithm is presented to enable the use of the enormous dataset with the available computing resources and time. Second, a data-driven approach is developed to predict the motor temperatures under a normal navigating state, based on convolutional neural network and long short-term memory models. The anomalies of motor temperature can then be identified by comparing the predicted normal temperatures and the ones detected by sensors. Compared to a baseline linear model, the developed approach provides an improvement of approximately 73.94% with respect to the mean squared error. The current study also investigates the performance of the developed approach in multiple simulated scenarios that can be of practical interest. The thorough evaluations have suggested a substantial potential of the approach with respect to its practicality, generalization, and precision.