K-means clustering and Deep Q-Network enhanced fault diagnosis for ZPW-2000A jointless track circuits

Abstract

To address the demands of China’s rapidly expanding railway networks and the development of high-speed rail technology, this study proposes an intelligent fault diagnosis method based on K-means clustering combined with an improved Deep Q-Network (DQN) to enhance the diagnostic accuracy of ZPW-2000A jointless track circuits, which are challenged by data sparsity and imbalanced fault distribution. First, a simulation model of the ZPW-2000A jointless track circuit is constructed by integrating historical fault data with a four-terminal network based on transmission line theory. Synthetic unlabeled data are generated through random fault scenario simulations at fixed time steps and fused with field-measured data to establish a comprehensive training dataset. Second, K-means clustering is applied for preliminary fault data partitioning, providing meaningful state representations for subsequent diagnosis. Finally, fault identification is implemented through the DQN framework, in which a prioritized experience replay strategy is incorporated to mitigate the negative impact of sparse and imbalanced data. Experimental results show that the proposed method achieves an average diagnostic accuracy of 99.98% across 11 typical fault categories, surpassing SGD-SVM (95.55%), Random Forest (95.71%), and CNN (97.45%) by 4.43%, 4.27%, and 2.53%, respectively. Furthermore, the proposed approach demonstrates superior robustness under imbalanced datasets and improved generalization in real-world fault scenarios. These findings confirm that the integration of K-means clustering with an improved DQN effectively alleviates data sparsity and significantly enhances the diagnostic performance of ZPW-2000A jointless track circuits.