A Deep Transfer Learning-Based Open Scenario Diagnostic Framework for Rail Damage Using Ultrasound Guided Waves

With the widespread application of ultrasonic guided wave (UGW) in the diagnosis of structural damage of steel, the intelligent diagnosis based on data driven for rail damages has been developed. However, the more realistic scenario is usually open, where unseen (or unknown) classes can emerge inevitably, which greatly hinders the diagnosis. Zero-shot learning (ZSL) has emerged in recent years to equip the deep model with the ability to recognize unseen categories. Therefore, this article proposes a deep transfer learning-based open-set diagnostic framework (OSDF) for rail damages. The framework uses simulated UGW signals of full categories to deal with the diagnosis of the unseen category in the open scenario. In the proposed framework, a dual-path convolution with attention mechanism (DPA)-based feature extractor is proposed to learn the damage features in the envelope signals of UGW and the sensitive information of damage in the original UGW signal. Meanwhile, to improve the accuracy of transfer diagnosis, a multidomain adversarial transfer learning (MATL) method is constructed. The effectiveness and advantages of the proposed method are further verified by performing unseen category diagnosis experiments, including six different damage types and comparing with other models. Diagnostic accuracy was validated for the overall sample and the unseen sample. The results of the two sets of comparison experiments show that the proposed method achieves diagnostic accuracies as high as 90.73% and 93.68% for the unseen category scenarios, respectively, which is a significant improvement over the 62.69% and 75.02% of the utilized baseline diagnostic method, and its essential reasons are illustrated in the mechanism analysis.