Unseen rail damage diagnosis framework using mechanism-embedded generative network

Abstract Rail damage can pose a tremendous hazard for high-speed trains, making damage diagnosis critical in the field of engineering. Currently, deep learning enables an end-to-end approach for rail damage diagnosis. However, the training and test data in real applications are often out of distribution, and the test data may even represent fault categories that were previously unseen. To address this situation, an unseen damage diagnosis framework (UDDF) that effectively embeds the mechanism damage features from the simulation signals of all possible damage categories has been proposed. In particular, the mechanism-embedded generative adversarial networks in the UDDF utilize a hierarchical embedding technique to ensure the stability of the mechanism embedding process. In addition, a k -means clustering discriminator uses an unsupervised method to guarantee the minimum intra-category sample spacing of the generated unseen categories. After the generation of all types of damage categories, the generated and existing original data are included as a new dataset for the training of a diagnostic model. The trained diagnostic model can perform classification tasks without acquiring all types of damage signals in real situations. Finally, the effectiveness of our proposed diagnostic framework is validated through comparative and ablation studies on a dataset that contains finite element simulation and experimental data of ultrasonic guided wave signals with damage at different locations and depths of rails.