Highly accurate positioning is an important prerequisite for automatic train driving. In terms of the problems that the existing machine learning is used for train positioning, such as the insufficient theoretical basis for feature selection and difficulty in determining the proper structure of model, which lead to the unstable and inaccurate data about train positioning. A new positioning method about urban rail train is proposed based on an ensemble deep belief network (DBN). This method firstly preprocesses the original dataset, then uses the Pearson coefficient to filter the features, finally utilizes the quantum particle swarm algorithm (QPSO) to optimize the structure of the DBN-based learner. Comparing the proposed QPSO-DBN model with the ensemble model about the classical machine learning methods and the traditional optimized algorithms, respectively, the positioning accuracy of the train is further improved. Finally, the superiority of the proposed model is verified by simulation experiments.