Deep Learning for Intelligent Train Driving with Augmented BLSTM

Abstract

As one of the most important means of transportation, the railway trains in modern society are eagerly approaching automatic driving due to their congenital advantages on operating environments compare to, e.g., road traffics. However, considering the factors of energy-efficiency, punctuality, as well as safety issues, the derivation of railway intelligent train driving process is challenging due to the high dimensionality, nonlinearity, complex constraints, and time-varying characteristics. The time-sequential train control sequences bears similarities to the text streams, which can be studied using Bidirectional LSTM (BLSTM) related methods. In this paper, we propose a dexterously augmented BLSTM model named BLSTM+, which considers both the sequential properties of the time-series and the peculiarities of each control process, for deep learning of control decision making process in such typical industrial control problems. The driving records of experienced human drivers are employed to train the BLSTM+ model. A reinforcement updating mechanism on the BLSTM+ model is also developed to further improve the performance by continuously mining the best-ranked driving records. Experimental results show that an energy saving of over 10% could be achieved under punctuality requirements when compared with the average level of human drivers.