Dynamic track stabilization (DTS) is essential for maintaining ballast beds and improving track stability. However, accurately identifying the “stable/unstable” states of the ballast bed during operation remains challenging, and traditional offline inspections can cause significant delays. To address this, this study proposes a ballast bed state recognition method utilizing a residual neural network (ResNet). It uses sleeper lateral acceleration signals as input, preprocesses them with Finite Impulse Response (FIR) band-pass filtering, and creates time–frequency image features through the Transient Extraction Transform (TET). These features are then combined with ballast compaction and stabilization parameters. The model employs a shared encoder and uses a combined weighted multi-task loss function. This loss function merges the classification cross-entropy with the settlement and lateral resistance regression losses, and applies dynamic weighting based on the GradNorm technique. This approach balances multi-task training and enhances the model’s robustness under different operational conditions. A total of 864 samples (605/173/86 for training/validation/test) are used for model training and evaluation. Results show that the ResNet18 model achieves 96.71% accuracy on the validation set, and the binary classification accuracy on the test set exceeds 98%, as shown by the confusion matrix. Compared to DenseNet201, training time is reduced by 77.26% under the same conditions, leading to a 340.9% increase in training efficiency. A field engineering case further demonstrates that the proposed model can accurately identify ballast bed states during DTS, with recognition results consistent with engineering criteria, indicating strong potential for practical application.
扫码关注我们
求助内容：
应助结果提醒方式：
京公网安备 11010802042870号