Vehicle System Dynamics

Coupler instability remains a critical challenge for long-marshaled heavy-haul trains, degrading coupler-buffer performance and even triggering severe safety incidents. Existing studies mainly improve coupler stability through parameter adjustment, yet neglect subsystem coupling effects and dynamics performance constraints. To address this, this study proposes a novel framework integrating deep learning and multi-objective optimization to enhance heavy-haul coupler stability. First, a coupler stability database is constructed based on a validated three-dimensional multi-body dynamics (MBD) heavy-haul train model. Then, a deep learning surrogate model (DBCM-CEDNet) combining the dual-branch convolutional and multi-layer perceptron module with the cascaded encoder-decoder network is trained under the joint supervision of data loss and physical constraint, which accurately captures the nonlinear mapping relationship between coupler-buffer, suspension system parameters and dynamics performance. Finally, DBCM-CEDNet is employed as the fitness function for the improved non-dominated sorting genetic algorithm III (NSGA-III) to efficiently generate Pareto solution sets, and an optimal parameter matching scheme is selected via minimum-distance-based posterior strategy. The optimized scheme effectively suppresses coupler lateral instability, with dynamics indices within safety limits. Overall, the proposed method navigates the complex parameter design space, reveals multi-system parameter interactions, and identifies optimal coupler stability configurations with lower computational costs than MBD simulations.