Deep Neural Networks Applied to the Dynamic Helper System in a GPGPU

Abstract

This paper presents a study about the use of two deep neural networks (MLP and LSTM) in the Dynamic Helper System (DHS), which is a system that only exists in the EFC railway. The DHS assists a fully loaded train in sections with slopes and climbs. Therefore, this work shows that it is possible to predict the next action of the DHS using a classification algorithm, specifically, deep neural networks. The training and testing dataset is a real one, and the training task was performed using PyCUDA, Keras, and TensorFlow in a General Purpose GPU. Results using a real-world dataset indicate that deep neural networks can reach an accuracy above 99% and precision about 81% on predicting the next action of the DHS.