The Train Maintenance Scheduling Problem (TMSP) is a real-world problem that aims at complete maintenance tasks of trains by scheduling their activities on a service site. Common methods of constructing optimal solutions to this problem are difficult as the problem consists of several highly-related sub-problems. Currently, NS is using a lo- cal search algorithm to provide solutions for the problem. However, it has several deficiencies such as solution randomness and lacking flexibility for rescheduling. In this research, we investigated the applicability of sequential decision making and supervised learning for solving TMSP. First, we formulate the TMSP problem with a reactive sequential mechanism and define the state and action space. Next, we design a feature representation for states and come up with the best kind of neural network structure through comparisons. Then, we conduct experiments to compare several search strategies with the trained network as the heuristic and find the best one. Fi- nally, we evaluate the solvability of our system and conclude that our approach has a certain capability for solving small-scale problems.