A stochastic track maintenance scheduling model based on deep reinforcement learning approaches

A data-driven railway track maintenance scheduling framework based on a stochastic track deterioration model and deep reinforcement learning approaches is proposed. Various track conditions such as track geometry and the support capacity of the infrastructure are considered in estimating the track deterioration rate and the track quality index obtained thereby is used to predict the state of each track segment. Further, the framework incorporates additional field-specific constraints including the number of tampings and the latest maintenance time of ballasted track are also introduced to account for the field conditions as accurately as possible. From these conditions, the optimal maintenance action for each track segment is determined based on the combined constraints of cost and ride comfort. In the present study, two reinforcement learning (RL) models, namely the Duel Deep Q Network (DuDQN) and Asynchronous Advantage Actor Critic (A3C) models, were employed to establish a decision support system of track maintenance, and the models' advantages and disadvantages were compared. Field application of the models was conducted based on field maintenance data, and the DuDQN model was found to be more suitable in our case. The optimal number of tampings before renewal was determined from the maintenance costs and field conditions, and the cost effect of ride comfort was investigated using the proposed deep RL model. Finally, possible improvements to the models were explored and are briefly outlined herein.