A diagnostic framework with a novel simulation data augmentation method for rail damages based on transfer learning

The ultrasonic guide wave (UGW) has good application prospects in steel rail damage diagnosis, but the features of the rail damage implied in the UGW are complex. Deep learning enables an end-to-end approach to fault diagnosis. Nevertheless, a large amount of diversity data is needed for training, whereas the ultrasonic wave guide signals of simulation and repeated experiments lack diversity. Therefore, in this paper, a diagnostic framework based on simulation and transfer learning for rail damage is developed to tackle the problems mentioned above. The proposed framework is based on deep learning with a simulation pretraining strategy to build convolutional neural network (CNN) models through parameter fine-tuning for damage diagnosis. Specifically, for the problem that the simulation data lacks diversity, a damage mechanism-based data diversity augmentation method is proposed; this obtains the diagnostic high-value simulation data including supporting features, and expanded the diversity of the simulation data. Adopting the proposed method of data augmentation and transfer learning (TL), a diagnostic model for rail damage utilizing augmented UGW signals is constructed. The finite element simulation data of UGW with damages at different locations and depths of rails are augmented to achieve the pretraining of CNN models, and the model transfer is performed with the experimental data of rails. Ultimately, through comparative studies it can be concluded that (1) The TL diagnostic framework makes full use of the finite element simulation data to realize the model pretraining. (2) The proposed data augmentation method realizes the diversity expansion of simulation data containing supporting features and ensures the efficient application of simulation data in model pretraining.