The Health Monitoring Model of Locomotive Diesel Engine System Based on Dynamic Graphs and DGCRN

To solve the problems of the complex and time-varying dependencies among multivariate sensor signals of locomotive diesel engine. This paper presents a health monitoring model for the locomotive diesel engine system based on a Dynamic Graph Convolutional Recurrent Network (DGCRN), this model integrates a dynamic graph generation module with a predefined static graph, enabling more effective spatiotemporal representation learning. A curriculum learning strategy is further adopted to improve training efficiency and long-sequence prediction performance. Experiments conducted on monitoring data from 138 sensors demonstrate that the proposed model achieves superior prediction accuracy for key parameters. Compared with using only static graphs, the dynamic graph enhanced model yields lower MAE, MAPE, and RMSE, and its prediction curves more closely follow real sensor trajectories. Specifically, the dynamic graph-based model reduces the MAE by 21.9%, MAPE by 21.4%, and RMSE by 17.2% compared to the only static graph model. These results confirm the effectiveness of incorporating dynamic graph structures and validate the DGCRN as a promising approach for health monitoring and performance prediction of complex diesel engine systems.