Ballastless track arching recognition based on one-dimensional residual convolutional neural network and vehicle response

Abstract

In the long-term service process, high-temperature load will lead to ballastless track arching. The ballastless track arching will reduce the structural integrity, affect the smoothness of the line, and even endanger the safety of driving. Existing detection methods are costly, complex and not suitable for large-scale. This study proposes a low-cost, simple detection method based on one-dimensional residual convolutional neural network and vehicle response to recognize the arching. The vehicle responses in the simulation model are used as inputs to compare the recognition accuracies of different algorithms for the arching, and the vehicle response features extracted by the optimal algorithm are visualized in 3D using t-Distributed Stochastic Neighbor Embedding. The results show that the vehicle vertical acceleration is sensitive to the arching amplitude, and the influence of the arching on the vehicle vertical acceleration ranges from about 40 m, and, the vehicle vertical acceleration in the range of 40 m is chosen as the input. The Network 2–2 containing two residual blocks 2 has the highest recognition accuracy. The accuracy at different vehicle speeds ranged from 96.5 % to 97.5 %, and the average accuracy at a single arching magnitude ranged from 95.0 % to 100 %. The original vehicle response features under different arching magnitudes are very complexly distributed in 3D space with more overlapping. After Network 2–2 processing, the vehicle response features corresponding to different arching magnitudes are basically separated in 3D space, and only very few features overlap. It shows that Network 2–2 has strong robustness and vehicle response feature extraction capability.