Multi-Task Deep Learning Methods for Determining Railway Major Technical Standards

Railway major technical standards (RMTSs) are prerequisites for subsequent railway design processes, which have fundamental influences on controlling railway transport capacity and investment. However, the traditional manual decision process is laborious and time-consuming. It also has difficulties in quantifying the interrelations among different standard parameters. In this paper, we propose a so-called voting election framework for efficient RMTS decision-making, in which three stages, namely constituency partition, constituency election and ballot summary, are developed. Then, for constituency election, four types of multi-task neural networks, namely, cascaded, overall parallel, grouped parallel and hybrid multi-task neural networks, are designed to identify the interrelations among six kinds of RMTSs. Besides, hyperparameter fine-tuning experiments are conducted to optimize the performance of the proposed networks. Finally, based on a dataset comprising realistic railway cases, it is found that the grouped parallel multi-task neural network performs the best by achieving an average decision accuracy of 86.9% compared to manual decision results and by outperforming the previous single-task neural network method by 10.7% in average decision accuracy.