Intelligent identification of ballastless track subgrade settlement based on vehicle-rail vibration data

Monitoring uneven subgrade settlement in ballastless track systems remains challenging due to the limited accuracy and discontinuous nature of conventional detection methods. In this study, we propose a novel deep learning approach based on a convolutional neural network and long short-term memory (CNN-LSTM) model to accurately identify uneven subgrade settlement by analyzing vehicle-rail dynamic response data. First, a vehicle-track-subgrade coupled model simulates vibration responses under differential settlements. Through sensitivity analysis of vehicle and track dynamics, we identify carbody vertical acceleration, nodding angular velocity, and rail displacement as optimal CNN-LSTM inputs. By leveraging the convolutional neural network (CNN)'s capability to extract spatial features and the long short-term memory (LSTM)'s strength in capturing temporal dependencies, the hybrid network effectively models the relationships between dynamic indicators and subgrade settlement. The results indicate that combined vehicle-rail responses enhances identification, particularly for track-subgrade deformation mismatches. The CNN-LSTM model achieves a detection accuracy of 99.26%, outperforming four benchmark models-backpropagation (BP) neural network, radial basis function (RBF) network, CNN, and LSTM-which validates its robustness and practical effectiveness. This research provides both theoretical insights and practical guidelines for intelligent monitoring of subgrade settlement in ballastless track systems.